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 visitAND(SDNode *N);
249 SDValue visitANDLike(SDValue N0, SDValue N1, SDNode *LocReference);
250 SDValue visitOR(SDNode *N);
251 SDValue visitORLike(SDValue N0, SDValue N1, SDNode *LocReference);
252 SDValue visitXOR(SDNode *N);
253 SDValue SimplifyVBinOp(SDNode *N);
254 SDValue visitSHL(SDNode *N);
255 SDValue visitSRA(SDNode *N);
256 SDValue visitSRL(SDNode *N);
257 SDValue visitRotate(SDNode *N);
258 SDValue visitBSWAP(SDNode *N);
259 SDValue visitCTLZ(SDNode *N);
260 SDValue visitCTLZ_ZERO_UNDEF(SDNode *N);
261 SDValue visitCTTZ(SDNode *N);
262 SDValue visitCTTZ_ZERO_UNDEF(SDNode *N);
263 SDValue visitCTPOP(SDNode *N);
264 SDValue visitSELECT(SDNode *N);
265 SDValue visitVSELECT(SDNode *N);
266 SDValue visitSELECT_CC(SDNode *N);
267 SDValue visitSETCC(SDNode *N);
268 SDValue visitSIGN_EXTEND(SDNode *N);
269 SDValue visitZERO_EXTEND(SDNode *N);
270 SDValue visitANY_EXTEND(SDNode *N);
271 SDValue visitSIGN_EXTEND_INREG(SDNode *N);
272 SDValue visitSIGN_EXTEND_VECTOR_INREG(SDNode *N);
273 SDValue visitTRUNCATE(SDNode *N);
274 SDValue visitBITCAST(SDNode *N);
275 SDValue visitBUILD_PAIR(SDNode *N);
276 SDValue visitFADD(SDNode *N);
277 SDValue visitFSUB(SDNode *N);
278 SDValue visitFMUL(SDNode *N);
279 SDValue visitFMA(SDNode *N);
280 SDValue visitFDIV(SDNode *N);
281 SDValue visitFREM(SDNode *N);
282 SDValue visitFSQRT(SDNode *N);
283 SDValue visitFCOPYSIGN(SDNode *N);
284 SDValue visitSINT_TO_FP(SDNode *N);
285 SDValue visitUINT_TO_FP(SDNode *N);
286 SDValue visitFP_TO_SINT(SDNode *N);
287 SDValue visitFP_TO_UINT(SDNode *N);
288 SDValue visitFP_ROUND(SDNode *N);
289 SDValue visitFP_ROUND_INREG(SDNode *N);
290 SDValue visitFP_EXTEND(SDNode *N);
291 SDValue visitFNEG(SDNode *N);
292 SDValue visitFABS(SDNode *N);
293 SDValue visitFCEIL(SDNode *N);
294 SDValue visitFTRUNC(SDNode *N);
295 SDValue visitFFLOOR(SDNode *N);
296 SDValue visitFMINNUM(SDNode *N);
297 SDValue visitFMAXNUM(SDNode *N);
298 SDValue visitBRCOND(SDNode *N);
299 SDValue visitBR_CC(SDNode *N);
300 SDValue visitLOAD(SDNode *N);
301 SDValue visitSTORE(SDNode *N);
302 SDValue visitINSERT_VECTOR_ELT(SDNode *N);
303 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
304 SDValue visitBUILD_VECTOR(SDNode *N);
305 SDValue visitCONCAT_VECTORS(SDNode *N);
306 SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
307 SDValue visitVECTOR_SHUFFLE(SDNode *N);
308 SDValue visitSCALAR_TO_VECTOR(SDNode *N);
309 SDValue visitINSERT_SUBVECTOR(SDNode *N);
310 SDValue visitMLOAD(SDNode *N);
311 SDValue visitMSTORE(SDNode *N);
312 SDValue visitMGATHER(SDNode *N);
313 SDValue visitMSCATTER(SDNode *N);
314 SDValue visitFP_TO_FP16(SDNode *N);
316 SDValue visitFADDForFMACombine(SDNode *N);
317 SDValue visitFSUBForFMACombine(SDNode *N);
319 SDValue XformToShuffleWithZero(SDNode *N);
320 SDValue ReassociateOps(unsigned Opc, SDLoc DL, SDValue LHS, SDValue RHS);
322 SDValue visitShiftByConstant(SDNode *N, ConstantSDNode *Amt);
324 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
325 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
326 SDValue SimplifySelect(SDLoc DL, SDValue N0, SDValue N1, SDValue N2);
327 SDValue SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1, SDValue N2,
328 SDValue N3, ISD::CondCode CC,
329 bool NotExtCompare = false);
330 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
331 SDLoc DL, bool foldBooleans = true);
333 bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
335 bool isOneUseSetCC(SDValue N) const;
337 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
339 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
340 SDValue CombineExtLoad(SDNode *N);
341 SDValue combineRepeatedFPDivisors(SDNode *N);
342 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
343 SDValue BuildSDIV(SDNode *N);
344 SDValue BuildSDIVPow2(SDNode *N);
345 SDValue BuildUDIV(SDNode *N);
346 SDValue BuildReciprocalEstimate(SDValue Op);
347 SDValue BuildRsqrtEstimate(SDValue Op);
348 SDValue BuildRsqrtNROneConst(SDValue Op, SDValue Est, unsigned Iterations);
349 SDValue BuildRsqrtNRTwoConst(SDValue Op, SDValue Est, unsigned Iterations);
350 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
351 bool DemandHighBits = true);
352 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
353 SDNode *MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg,
354 SDValue InnerPos, SDValue InnerNeg,
355 unsigned PosOpcode, unsigned NegOpcode,
357 SDNode *MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL);
358 SDValue ReduceLoadWidth(SDNode *N);
359 SDValue ReduceLoadOpStoreWidth(SDNode *N);
360 SDValue TransformFPLoadStorePair(SDNode *N);
361 SDValue reduceBuildVecExtToExtBuildVec(SDNode *N);
362 SDValue reduceBuildVecConvertToConvertBuildVec(SDNode *N);
364 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
366 /// Walk up chain skipping non-aliasing memory nodes,
367 /// looking for aliasing nodes and adding them to the Aliases vector.
368 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
369 SmallVectorImpl<SDValue> &Aliases);
371 /// Return true if there is any possibility that the two addresses overlap.
372 bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const;
374 /// Walk up chain skipping non-aliasing memory nodes, looking for a better
375 /// chain (aliasing node.)
376 SDValue FindBetterChain(SDNode *N, SDValue Chain);
378 /// Holds a pointer to an LSBaseSDNode as well as information on where it
379 /// is located in a sequence of memory operations connected by a chain.
381 MemOpLink (LSBaseSDNode *N, int64_t Offset, unsigned Seq):
382 MemNode(N), OffsetFromBase(Offset), SequenceNum(Seq) { }
383 // Ptr to the mem node.
384 LSBaseSDNode *MemNode;
385 // Offset from the base ptr.
386 int64_t OffsetFromBase;
387 // What is the sequence number of this mem node.
388 // Lowest mem operand in the DAG starts at zero.
389 unsigned SequenceNum;
392 /// This is a helper function for MergeStoresOfConstantsOrVecElts. Returns a
393 /// constant build_vector of the stored constant values in Stores.
394 SDValue getMergedConstantVectorStore(SelectionDAG &DAG,
396 ArrayRef<MemOpLink> Stores,
399 /// This is a helper function for MergeConsecutiveStores. When the source
400 /// elements of the consecutive stores are all constants or all extracted
401 /// vector elements, try to merge them into one larger store.
402 /// \return True if a merged store was created.
403 bool MergeStoresOfConstantsOrVecElts(SmallVectorImpl<MemOpLink> &StoreNodes,
404 EVT MemVT, unsigned NumElem,
405 bool IsConstantSrc, bool UseVector);
407 /// This is a helper function for MergeConsecutiveStores.
408 /// Stores that may be merged are placed in StoreNodes.
409 /// Loads that may alias with those stores are placed in AliasLoadNodes.
410 void getStoreMergeAndAliasCandidates(
411 StoreSDNode* St, SmallVectorImpl<MemOpLink> &StoreNodes,
412 SmallVectorImpl<LSBaseSDNode*> &AliasLoadNodes);
414 /// Merge consecutive store operations into a wide store.
415 /// This optimization uses wide integers or vectors when possible.
416 /// \return True if some memory operations were changed.
417 bool MergeConsecutiveStores(StoreSDNode *N);
419 /// \brief Try to transform a truncation where C is a constant:
420 /// (trunc (and X, C)) -> (and (trunc X), (trunc C))
422 /// \p N needs to be a truncation and its first operand an AND. Other
423 /// requirements are checked by the function (e.g. that trunc is
424 /// single-use) and if missed an empty SDValue is returned.
425 SDValue distributeTruncateThroughAnd(SDNode *N);
428 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
429 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
430 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {
431 auto *F = DAG.getMachineFunction().getFunction();
432 ForCodeSize = F->hasFnAttribute(Attribute::OptimizeForSize) ||
433 F->hasFnAttribute(Attribute::MinSize);
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);
1346 case ISD::AND: return visitAND(N);
1347 case ISD::OR: return visitOR(N);
1348 case ISD::XOR: return visitXOR(N);
1349 case ISD::SHL: return visitSHL(N);
1350 case ISD::SRA: return visitSRA(N);
1351 case ISD::SRL: return visitSRL(N);
1353 case ISD::ROTL: return visitRotate(N);
1354 case ISD::BSWAP: return visitBSWAP(N);
1355 case ISD::CTLZ: return visitCTLZ(N);
1356 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1357 case ISD::CTTZ: return visitCTTZ(N);
1358 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1359 case ISD::CTPOP: return visitCTPOP(N);
1360 case ISD::SELECT: return visitSELECT(N);
1361 case ISD::VSELECT: return visitVSELECT(N);
1362 case ISD::SELECT_CC: return visitSELECT_CC(N);
1363 case ISD::SETCC: return visitSETCC(N);
1364 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1365 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1366 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1367 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1368 case ISD::SIGN_EXTEND_VECTOR_INREG: return visitSIGN_EXTEND_VECTOR_INREG(N);
1369 case ISD::TRUNCATE: return visitTRUNCATE(N);
1370 case ISD::BITCAST: return visitBITCAST(N);
1371 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1372 case ISD::FADD: return visitFADD(N);
1373 case ISD::FSUB: return visitFSUB(N);
1374 case ISD::FMUL: return visitFMUL(N);
1375 case ISD::FMA: return visitFMA(N);
1376 case ISD::FDIV: return visitFDIV(N);
1377 case ISD::FREM: return visitFREM(N);
1378 case ISD::FSQRT: return visitFSQRT(N);
1379 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1380 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1381 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1382 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1383 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1384 case ISD::FP_ROUND: return visitFP_ROUND(N);
1385 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1386 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1387 case ISD::FNEG: return visitFNEG(N);
1388 case ISD::FABS: return visitFABS(N);
1389 case ISD::FFLOOR: return visitFFLOOR(N);
1390 case ISD::FMINNUM: return visitFMINNUM(N);
1391 case ISD::FMAXNUM: return visitFMAXNUM(N);
1392 case ISD::FCEIL: return visitFCEIL(N);
1393 case ISD::FTRUNC: return visitFTRUNC(N);
1394 case ISD::BRCOND: return visitBRCOND(N);
1395 case ISD::BR_CC: return visitBR_CC(N);
1396 case ISD::LOAD: return visitLOAD(N);
1397 case ISD::STORE: return visitSTORE(N);
1398 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1399 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1400 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1401 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1402 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1403 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1404 case ISD::SCALAR_TO_VECTOR: return visitSCALAR_TO_VECTOR(N);
1405 case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N);
1406 case ISD::MGATHER: return visitMGATHER(N);
1407 case ISD::MLOAD: return visitMLOAD(N);
1408 case ISD::MSCATTER: return visitMSCATTER(N);
1409 case ISD::MSTORE: return visitMSTORE(N);
1410 case ISD::FP_TO_FP16: return visitFP_TO_FP16(N);
1415 SDValue DAGCombiner::combine(SDNode *N) {
1416 SDValue RV = visit(N);
1418 // If nothing happened, try a target-specific DAG combine.
1419 if (!RV.getNode()) {
1420 assert(N->getOpcode() != ISD::DELETED_NODE &&
1421 "Node was deleted but visit returned NULL!");
1423 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1424 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1426 // Expose the DAG combiner to the target combiner impls.
1427 TargetLowering::DAGCombinerInfo
1428 DagCombineInfo(DAG, Level, false, this);
1430 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1434 // If nothing happened still, try promoting the operation.
1435 if (!RV.getNode()) {
1436 switch (N->getOpcode()) {
1444 RV = PromoteIntBinOp(SDValue(N, 0));
1449 RV = PromoteIntShiftOp(SDValue(N, 0));
1451 case ISD::SIGN_EXTEND:
1452 case ISD::ZERO_EXTEND:
1453 case ISD::ANY_EXTEND:
1454 RV = PromoteExtend(SDValue(N, 0));
1457 if (PromoteLoad(SDValue(N, 0)))
1463 // If N is a commutative binary node, try commuting it to enable more
1465 if (!RV.getNode() && SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1466 N->getNumValues() == 1) {
1467 SDValue N0 = N->getOperand(0);
1468 SDValue N1 = N->getOperand(1);
1470 // Constant operands are canonicalized to RHS.
1471 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1472 SDValue Ops[] = {N1, N0};
1474 if (const auto *BinNode = dyn_cast<BinaryWithFlagsSDNode>(N)) {
1475 CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops,
1478 CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops);
1481 return SDValue(CSENode, 0);
1488 /// Given a node, return its input chain if it has one, otherwise return a null
1490 static SDValue getInputChainForNode(SDNode *N) {
1491 if (unsigned NumOps = N->getNumOperands()) {
1492 if (N->getOperand(0).getValueType() == MVT::Other)
1493 return N->getOperand(0);
1494 if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1495 return N->getOperand(NumOps-1);
1496 for (unsigned i = 1; i < NumOps-1; ++i)
1497 if (N->getOperand(i).getValueType() == MVT::Other)
1498 return N->getOperand(i);
1503 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1504 // If N has two operands, where one has an input chain equal to the other,
1505 // the 'other' chain is redundant.
1506 if (N->getNumOperands() == 2) {
1507 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1508 return N->getOperand(0);
1509 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1510 return N->getOperand(1);
1513 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1514 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1515 SmallPtrSet<SDNode*, 16> SeenOps;
1516 bool Changed = false; // If we should replace this token factor.
1518 // Start out with this token factor.
1521 // Iterate through token factors. The TFs grows when new token factors are
1523 for (unsigned i = 0; i < TFs.size(); ++i) {
1524 SDNode *TF = TFs[i];
1526 // Check each of the operands.
1527 for (const SDValue &Op : TF->op_values()) {
1529 switch (Op.getOpcode()) {
1530 case ISD::EntryToken:
1531 // Entry tokens don't need to be added to the list. They are
1536 case ISD::TokenFactor:
1537 if (Op.hasOneUse() &&
1538 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1539 // Queue up for processing.
1540 TFs.push_back(Op.getNode());
1541 // Clean up in case the token factor is removed.
1542 AddToWorklist(Op.getNode());
1549 // Only add if it isn't already in the list.
1550 if (SeenOps.insert(Op.getNode()).second)
1561 // If we've changed things around then replace token factor.
1564 // The entry token is the only possible outcome.
1565 Result = DAG.getEntryNode();
1567 // New and improved token factor.
1568 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Ops);
1571 // Add users to worklist if AA is enabled, since it may introduce
1572 // a lot of new chained token factors while removing memory deps.
1573 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
1574 : DAG.getSubtarget().useAA();
1575 return CombineTo(N, Result, UseAA /*add to worklist*/);
1581 /// MERGE_VALUES can always be eliminated.
1582 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1583 WorklistRemover DeadNodes(*this);
1584 // Replacing results may cause a different MERGE_VALUES to suddenly
1585 // be CSE'd with N, and carry its uses with it. Iterate until no
1586 // uses remain, to ensure that the node can be safely deleted.
1587 // First add the users of this node to the work list so that they
1588 // can be tried again once they have new operands.
1589 AddUsersToWorklist(N);
1591 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1592 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1593 } while (!N->use_empty());
1594 deleteAndRecombine(N);
1595 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1598 static bool isNullConstant(SDValue V) {
1599 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(V);
1600 return Const != nullptr && Const->isNullValue();
1603 static bool isNullFPConstant(SDValue V) {
1604 ConstantFPSDNode *Const = dyn_cast<ConstantFPSDNode>(V);
1605 return Const != nullptr && Const->isZero() && !Const->isNegative();
1608 static bool isAllOnesConstant(SDValue V) {
1609 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(V);
1610 return Const != nullptr && Const->isAllOnesValue();
1613 static bool isOneConstant(SDValue V) {
1614 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(V);
1615 return Const != nullptr && Const->isOne();
1618 /// If \p N is a ContantSDNode with isOpaque() == false return it casted to a
1619 /// ContantSDNode pointer else nullptr.
1620 static ConstantSDNode *getAsNonOpaqueConstant(SDValue N) {
1621 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N);
1622 return Const != nullptr && !Const->isOpaque() ? Const : nullptr;
1625 SDValue DAGCombiner::visitADD(SDNode *N) {
1626 SDValue N0 = N->getOperand(0);
1627 SDValue N1 = N->getOperand(1);
1628 EVT VT = N0.getValueType();
1631 if (VT.isVector()) {
1632 if (SDValue FoldedVOp = SimplifyVBinOp(N))
1635 // fold (add x, 0) -> x, vector edition
1636 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1638 if (ISD::isBuildVectorAllZeros(N0.getNode()))
1642 // fold (add x, undef) -> undef
1643 if (N0.getOpcode() == ISD::UNDEF)
1645 if (N1.getOpcode() == ISD::UNDEF)
1647 // fold (add c1, c2) -> c1+c2
1648 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
1649 ConstantSDNode *N1C = getAsNonOpaqueConstant(N1);
1651 return DAG.FoldConstantArithmetic(ISD::ADD, SDLoc(N), VT, N0C, N1C);
1652 // canonicalize constant to RHS
1653 if (isConstantIntBuildVectorOrConstantInt(N0) &&
1654 !isConstantIntBuildVectorOrConstantInt(N1))
1655 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, N0);
1656 // fold (add x, 0) -> x
1657 if (isNullConstant(N1))
1659 // fold (add Sym, c) -> Sym+c
1660 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1661 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1662 GA->getOpcode() == ISD::GlobalAddress)
1663 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1665 (uint64_t)N1C->getSExtValue());
1666 // fold ((c1-A)+c2) -> (c1+c2)-A
1667 if (N1C && N0.getOpcode() == ISD::SUB)
1668 if (ConstantSDNode *N0C = getAsNonOpaqueConstant(N0.getOperand(0))) {
1670 return DAG.getNode(ISD::SUB, DL, VT,
1671 DAG.getConstant(N1C->getAPIntValue()+
1672 N0C->getAPIntValue(), DL, VT),
1676 if (SDValue RADD = ReassociateOps(ISD::ADD, SDLoc(N), N0, N1))
1678 // fold ((0-A) + B) -> B-A
1679 if (N0.getOpcode() == ISD::SUB && isNullConstant(N0.getOperand(0)))
1680 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, N0.getOperand(1));
1681 // fold (A + (0-B)) -> A-B
1682 if (N1.getOpcode() == ISD::SUB && isNullConstant(N1.getOperand(0)))
1683 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1.getOperand(1));
1684 // fold (A+(B-A)) -> B
1685 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1686 return N1.getOperand(0);
1687 // fold ((B-A)+A) -> B
1688 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1689 return N0.getOperand(0);
1690 // fold (A+(B-(A+C))) to (B-C)
1691 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1692 N0 == N1.getOperand(1).getOperand(0))
1693 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1694 N1.getOperand(1).getOperand(1));
1695 // fold (A+(B-(C+A))) to (B-C)
1696 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1697 N0 == N1.getOperand(1).getOperand(1))
1698 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1699 N1.getOperand(1).getOperand(0));
1700 // fold (A+((B-A)+or-C)) to (B+or-C)
1701 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1702 N1.getOperand(0).getOpcode() == ISD::SUB &&
1703 N0 == N1.getOperand(0).getOperand(1))
1704 return DAG.getNode(N1.getOpcode(), SDLoc(N), VT,
1705 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1707 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1708 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1709 SDValue N00 = N0.getOperand(0);
1710 SDValue N01 = N0.getOperand(1);
1711 SDValue N10 = N1.getOperand(0);
1712 SDValue N11 = N1.getOperand(1);
1714 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1715 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1716 DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10),
1717 DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11));
1720 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1721 return SDValue(N, 0);
1723 // fold (a+b) -> (a|b) iff a and b share no bits.
1724 if (VT.isInteger() && !VT.isVector()) {
1725 APInt LHSZero, LHSOne;
1726 APInt RHSZero, RHSOne;
1727 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1729 if (LHSZero.getBoolValue()) {
1730 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1732 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1733 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1734 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero){
1735 if (!LegalOperations || TLI.isOperationLegal(ISD::OR, VT))
1736 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1);
1741 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1742 if (N1.getOpcode() == ISD::SHL && N1.getOperand(0).getOpcode() == ISD::SUB &&
1743 isNullConstant(N1.getOperand(0).getOperand(0)))
1744 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0,
1745 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1746 N1.getOperand(0).getOperand(1),
1748 if (N0.getOpcode() == ISD::SHL && N0.getOperand(0).getOpcode() == ISD::SUB &&
1749 isNullConstant(N0.getOperand(0).getOperand(0)))
1750 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1,
1751 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1752 N0.getOperand(0).getOperand(1),
1755 if (N1.getOpcode() == ISD::AND) {
1756 SDValue AndOp0 = N1.getOperand(0);
1757 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1758 unsigned DestBits = VT.getScalarType().getSizeInBits();
1760 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1761 // and similar xforms where the inner op is either ~0 or 0.
1762 if (NumSignBits == DestBits && isOneConstant(N1->getOperand(1))) {
1764 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1768 // add (sext i1), X -> sub X, (zext i1)
1769 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1770 N0.getOperand(0).getValueType() == MVT::i1 &&
1771 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1773 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1774 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1777 // add X, (sextinreg Y i1) -> sub X, (and Y 1)
1778 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
1779 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
1780 if (TN->getVT() == MVT::i1) {
1782 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
1783 DAG.getConstant(1, DL, VT));
1784 return DAG.getNode(ISD::SUB, DL, VT, N0, ZExt);
1791 SDValue DAGCombiner::visitADDC(SDNode *N) {
1792 SDValue N0 = N->getOperand(0);
1793 SDValue N1 = N->getOperand(1);
1794 EVT VT = N0.getValueType();
1796 // If the flag result is dead, turn this into an ADD.
1797 if (!N->hasAnyUseOfValue(1))
1798 return CombineTo(N, DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N1),
1799 DAG.getNode(ISD::CARRY_FALSE,
1800 SDLoc(N), MVT::Glue));
1802 // canonicalize constant to RHS.
1803 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1804 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1806 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N1, N0);
1808 // fold (addc x, 0) -> x + no carry out
1809 if (isNullConstant(N1))
1810 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1811 SDLoc(N), MVT::Glue));
1813 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1814 APInt LHSZero, LHSOne;
1815 APInt RHSZero, RHSOne;
1816 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1818 if (LHSZero.getBoolValue()) {
1819 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1821 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1822 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1823 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1824 return CombineTo(N, DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1),
1825 DAG.getNode(ISD::CARRY_FALSE,
1826 SDLoc(N), MVT::Glue));
1832 SDValue DAGCombiner::visitADDE(SDNode *N) {
1833 SDValue N0 = N->getOperand(0);
1834 SDValue N1 = N->getOperand(1);
1835 SDValue CarryIn = N->getOperand(2);
1837 // canonicalize constant to RHS
1838 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1839 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1841 return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(),
1844 // fold (adde x, y, false) -> (addc x, y)
1845 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1846 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1);
1851 // Since it may not be valid to emit a fold to zero for vector initializers
1852 // check if we can before folding.
1853 static SDValue tryFoldToZero(SDLoc DL, const TargetLowering &TLI, EVT VT,
1855 bool LegalOperations, bool LegalTypes) {
1857 return DAG.getConstant(0, DL, VT);
1858 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
1859 return DAG.getConstant(0, DL, VT);
1863 SDValue DAGCombiner::visitSUB(SDNode *N) {
1864 SDValue N0 = N->getOperand(0);
1865 SDValue N1 = N->getOperand(1);
1866 EVT VT = N0.getValueType();
1869 if (VT.isVector()) {
1870 if (SDValue FoldedVOp = SimplifyVBinOp(N))
1873 // fold (sub x, 0) -> x, vector edition
1874 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1878 // fold (sub x, x) -> 0
1879 // FIXME: Refactor this and xor and other similar operations together.
1881 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
1882 // fold (sub c1, c2) -> c1-c2
1883 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
1884 ConstantSDNode *N1C = getAsNonOpaqueConstant(N1);
1886 return DAG.FoldConstantArithmetic(ISD::SUB, SDLoc(N), VT, N0C, N1C);
1887 // fold (sub x, c) -> (add x, -c)
1890 return DAG.getNode(ISD::ADD, DL, VT, N0,
1891 DAG.getConstant(-N1C->getAPIntValue(), DL, VT));
1893 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1894 if (isAllOnesConstant(N0))
1895 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
1896 // fold A-(A-B) -> B
1897 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1898 return N1.getOperand(1);
1899 // fold (A+B)-A -> B
1900 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1901 return N0.getOperand(1);
1902 // fold (A+B)-B -> A
1903 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1904 return N0.getOperand(0);
1905 // fold C2-(A+C1) -> (C2-C1)-A
1906 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? nullptr :
1907 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1908 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1910 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(),
1912 return DAG.getNode(ISD::SUB, DL, VT, NewC,
1915 // fold ((A+(B+or-C))-B) -> A+or-C
1916 if (N0.getOpcode() == ISD::ADD &&
1917 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1918 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1919 N0.getOperand(1).getOperand(0) == N1)
1920 return DAG.getNode(N0.getOperand(1).getOpcode(), SDLoc(N), VT,
1921 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1922 // fold ((A+(C+B))-B) -> A+C
1923 if (N0.getOpcode() == ISD::ADD &&
1924 N0.getOperand(1).getOpcode() == ISD::ADD &&
1925 N0.getOperand(1).getOperand(1) == N1)
1926 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1927 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1928 // fold ((A-(B-C))-C) -> A-B
1929 if (N0.getOpcode() == ISD::SUB &&
1930 N0.getOperand(1).getOpcode() == ISD::SUB &&
1931 N0.getOperand(1).getOperand(1) == N1)
1932 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1933 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1935 // If either operand of a sub is undef, the result is undef
1936 if (N0.getOpcode() == ISD::UNDEF)
1938 if (N1.getOpcode() == ISD::UNDEF)
1941 // If the relocation model supports it, consider symbol offsets.
1942 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1943 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1944 // fold (sub Sym, c) -> Sym-c
1945 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1946 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1948 (uint64_t)N1C->getSExtValue());
1949 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1950 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1951 if (GA->getGlobal() == GB->getGlobal())
1952 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1956 // sub X, (sextinreg Y i1) -> add X, (and Y 1)
1957 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
1958 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
1959 if (TN->getVT() == MVT::i1) {
1961 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
1962 DAG.getConstant(1, DL, VT));
1963 return DAG.getNode(ISD::ADD, DL, VT, N0, ZExt);
1970 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1971 SDValue N0 = N->getOperand(0);
1972 SDValue N1 = N->getOperand(1);
1973 EVT VT = N0.getValueType();
1975 // If the flag result is dead, turn this into an SUB.
1976 if (!N->hasAnyUseOfValue(1))
1977 return CombineTo(N, DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1),
1978 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1981 // fold (subc x, x) -> 0 + no borrow
1984 return CombineTo(N, DAG.getConstant(0, DL, VT),
1985 DAG.getNode(ISD::CARRY_FALSE, DL,
1989 // fold (subc x, 0) -> x + no borrow
1990 if (isNullConstant(N1))
1991 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1994 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1995 if (isAllOnesConstant(N0))
1996 return CombineTo(N, DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0),
1997 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
2003 SDValue DAGCombiner::visitSUBE(SDNode *N) {
2004 SDValue N0 = N->getOperand(0);
2005 SDValue N1 = N->getOperand(1);
2006 SDValue CarryIn = N->getOperand(2);
2008 // fold (sube x, y, false) -> (subc x, y)
2009 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
2010 return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1);
2015 SDValue DAGCombiner::visitMUL(SDNode *N) {
2016 SDValue N0 = N->getOperand(0);
2017 SDValue N1 = N->getOperand(1);
2018 EVT VT = N0.getValueType();
2020 // fold (mul x, undef) -> 0
2021 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2022 return DAG.getConstant(0, SDLoc(N), VT);
2024 bool N0IsConst = false;
2025 bool N1IsConst = false;
2026 bool N1IsOpaqueConst = false;
2027 bool N0IsOpaqueConst = false;
2028 APInt ConstValue0, ConstValue1;
2030 if (VT.isVector()) {
2031 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2034 N0IsConst = isConstantSplatVector(N0.getNode(), ConstValue0);
2035 N1IsConst = isConstantSplatVector(N1.getNode(), ConstValue1);
2037 N0IsConst = isa<ConstantSDNode>(N0);
2039 ConstValue0 = cast<ConstantSDNode>(N0)->getAPIntValue();
2040 N0IsOpaqueConst = cast<ConstantSDNode>(N0)->isOpaque();
2042 N1IsConst = isa<ConstantSDNode>(N1);
2044 ConstValue1 = cast<ConstantSDNode>(N1)->getAPIntValue();
2045 N1IsOpaqueConst = cast<ConstantSDNode>(N1)->isOpaque();
2049 // fold (mul c1, c2) -> c1*c2
2050 if (N0IsConst && N1IsConst && !N0IsOpaqueConst && !N1IsOpaqueConst)
2051 return DAG.FoldConstantArithmetic(ISD::MUL, SDLoc(N), VT,
2052 N0.getNode(), N1.getNode());
2054 // canonicalize constant to RHS (vector doesn't have to splat)
2055 if (isConstantIntBuildVectorOrConstantInt(N0) &&
2056 !isConstantIntBuildVectorOrConstantInt(N1))
2057 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0);
2058 // fold (mul x, 0) -> 0
2059 if (N1IsConst && ConstValue1 == 0)
2061 // We require a splat of the entire scalar bit width for non-contiguous
2064 ConstValue1.getBitWidth() == VT.getScalarType().getSizeInBits();
2065 // fold (mul x, 1) -> x
2066 if (N1IsConst && ConstValue1 == 1 && IsFullSplat)
2068 // fold (mul x, -1) -> 0-x
2069 if (N1IsConst && ConstValue1.isAllOnesValue()) {
2071 return DAG.getNode(ISD::SUB, DL, VT,
2072 DAG.getConstant(0, DL, VT), N0);
2074 // fold (mul x, (1 << c)) -> x << c
2075 if (N1IsConst && !N1IsOpaqueConst && ConstValue1.isPowerOf2() &&
2078 return DAG.getNode(ISD::SHL, DL, VT, N0,
2079 DAG.getConstant(ConstValue1.logBase2(), DL,
2080 getShiftAmountTy(N0.getValueType())));
2082 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
2083 if (N1IsConst && !N1IsOpaqueConst && (-ConstValue1).isPowerOf2() &&
2085 unsigned Log2Val = (-ConstValue1).logBase2();
2087 // FIXME: If the input is something that is easily negated (e.g. a
2088 // single-use add), we should put the negate there.
2089 return DAG.getNode(ISD::SUB, DL, VT,
2090 DAG.getConstant(0, DL, VT),
2091 DAG.getNode(ISD::SHL, DL, VT, N0,
2092 DAG.getConstant(Log2Val, DL,
2093 getShiftAmountTy(N0.getValueType()))));
2097 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
2098 if (N1IsConst && N0.getOpcode() == ISD::SHL &&
2099 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2100 isa<ConstantSDNode>(N0.getOperand(1)))) {
2101 SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT,
2102 N1, N0.getOperand(1));
2103 AddToWorklist(C3.getNode());
2104 return DAG.getNode(ISD::MUL, SDLoc(N), VT,
2105 N0.getOperand(0), C3);
2108 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
2111 SDValue Sh(nullptr,0), Y(nullptr,0);
2112 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
2113 if (N0.getOpcode() == ISD::SHL &&
2114 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2115 isa<ConstantSDNode>(N0.getOperand(1))) &&
2116 N0.getNode()->hasOneUse()) {
2118 } else if (N1.getOpcode() == ISD::SHL &&
2119 isa<ConstantSDNode>(N1.getOperand(1)) &&
2120 N1.getNode()->hasOneUse()) {
2125 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2126 Sh.getOperand(0), Y);
2127 return DAG.getNode(ISD::SHL, SDLoc(N), VT,
2128 Mul, Sh.getOperand(1));
2132 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
2133 if (N1IsConst && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
2134 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2135 isa<ConstantSDNode>(N0.getOperand(1))))
2136 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
2137 DAG.getNode(ISD::MUL, SDLoc(N0), VT,
2138 N0.getOperand(0), N1),
2139 DAG.getNode(ISD::MUL, SDLoc(N1), VT,
2140 N0.getOperand(1), N1));
2143 if (SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1))
2149 SDValue DAGCombiner::visitSDIV(SDNode *N) {
2150 SDValue N0 = N->getOperand(0);
2151 SDValue N1 = N->getOperand(1);
2152 EVT VT = N->getValueType(0);
2156 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2159 // fold (sdiv c1, c2) -> c1/c2
2160 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2161 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2162 if (N0C && N1C && !N0C->isOpaque() && !N1C->isOpaque())
2163 return DAG.FoldConstantArithmetic(ISD::SDIV, SDLoc(N), VT, N0C, N1C);
2164 // fold (sdiv X, 1) -> X
2165 if (N1C && N1C->isOne())
2167 // fold (sdiv X, -1) -> 0-X
2168 if (N1C && N1C->isAllOnesValue()) {
2170 return DAG.getNode(ISD::SUB, DL, VT,
2171 DAG.getConstant(0, DL, VT), N0);
2173 // If we know the sign bits of both operands are zero, strength reduce to a
2174 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
2175 if (!VT.isVector()) {
2176 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2177 return DAG.getNode(ISD::UDIV, SDLoc(N), N1.getValueType(),
2181 // fold (sdiv X, pow2) -> simple ops after legalize
2182 // FIXME: We check for the exact bit here because the generic lowering gives
2183 // better results in that case. The target-specific lowering should learn how
2184 // to handle exact sdivs efficiently.
2185 if (N1C && !N1C->isNullValue() && !N1C->isOpaque() &&
2186 !cast<BinaryWithFlagsSDNode>(N)->Flags.hasExact() &&
2187 (N1C->getAPIntValue().isPowerOf2() ||
2188 (-N1C->getAPIntValue()).isPowerOf2())) {
2189 // If dividing by powers of two is cheap, then don't perform the following
2191 if (TLI.isPow2SDivCheap())
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.
2231 if (N1C && !TLI.isIntDivCheap())
2232 if (SDValue Op = BuildSDIV(N))
2236 if (N0.getOpcode() == ISD::UNDEF)
2237 return DAG.getConstant(0, SDLoc(N), VT);
2238 // X / undef -> undef
2239 if (N1.getOpcode() == ISD::UNDEF)
2245 SDValue DAGCombiner::visitUDIV(SDNode *N) {
2246 SDValue N0 = N->getOperand(0);
2247 SDValue N1 = N->getOperand(1);
2248 EVT VT = N->getValueType(0);
2252 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2255 // fold (udiv c1, c2) -> c1/c2
2256 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2257 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2259 if (SDValue Folded = DAG.FoldConstantArithmetic(ISD::UDIV, SDLoc(N), VT,
2262 // fold (udiv x, (1 << c)) -> x >>u c
2263 if (N1C && !N1C->isOpaque() && N1C->getAPIntValue().isPowerOf2()) {
2265 return DAG.getNode(ISD::SRL, DL, VT, N0,
2266 DAG.getConstant(N1C->getAPIntValue().logBase2(), DL,
2267 getShiftAmountTy(N0.getValueType())));
2269 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
2270 if (N1.getOpcode() == ISD::SHL) {
2271 if (ConstantSDNode *SHC = getAsNonOpaqueConstant(N1.getOperand(0))) {
2272 if (SHC->getAPIntValue().isPowerOf2()) {
2273 EVT ADDVT = N1.getOperand(1).getValueType();
2275 SDValue Add = DAG.getNode(ISD::ADD, DL, ADDVT,
2277 DAG.getConstant(SHC->getAPIntValue()
2280 AddToWorklist(Add.getNode());
2281 return DAG.getNode(ISD::SRL, DL, VT, N0, Add);
2285 // fold (udiv x, c) -> alternate
2286 if (N1C && !TLI.isIntDivCheap())
2287 if (SDValue Op = BuildUDIV(N))
2291 if (N0.getOpcode() == ISD::UNDEF)
2292 return DAG.getConstant(0, SDLoc(N), VT);
2293 // X / undef -> undef
2294 if (N1.getOpcode() == ISD::UNDEF)
2300 SDValue DAGCombiner::visitSREM(SDNode *N) {
2301 SDValue N0 = N->getOperand(0);
2302 SDValue N1 = N->getOperand(1);
2303 EVT VT = N->getValueType(0);
2305 // fold (srem c1, c2) -> c1%c2
2306 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2307 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2309 if (SDValue Folded = DAG.FoldConstantArithmetic(ISD::SREM, SDLoc(N), VT,
2312 // If we know the sign bits of both operands are zero, strength reduce to a
2313 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
2314 if (!VT.isVector()) {
2315 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2316 return DAG.getNode(ISD::UREM, SDLoc(N), VT, N0, N1);
2319 // If X/C can be simplified by the division-by-constant logic, lower
2320 // X%C to the equivalent of X-X/C*C.
2321 if (N1C && !N1C->isNullValue()) {
2322 SDValue Div = DAG.getNode(ISD::SDIV, SDLoc(N), VT, N0, N1);
2323 AddToWorklist(Div.getNode());
2324 SDValue OptimizedDiv = combine(Div.getNode());
2325 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2326 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2328 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2329 AddToWorklist(Mul.getNode());
2335 if (N0.getOpcode() == ISD::UNDEF)
2336 return DAG.getConstant(0, SDLoc(N), VT);
2337 // X % undef -> undef
2338 if (N1.getOpcode() == ISD::UNDEF)
2344 SDValue DAGCombiner::visitUREM(SDNode *N) {
2345 SDValue N0 = N->getOperand(0);
2346 SDValue N1 = N->getOperand(1);
2347 EVT VT = N->getValueType(0);
2349 // fold (urem c1, c2) -> c1%c2
2350 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2351 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2353 if (SDValue Folded = DAG.FoldConstantArithmetic(ISD::UREM, SDLoc(N), VT,
2356 // fold (urem x, pow2) -> (and x, pow2-1)
2357 if (N1C && !N1C->isNullValue() && !N1C->isOpaque() &&
2358 N1C->getAPIntValue().isPowerOf2()) {
2360 return DAG.getNode(ISD::AND, DL, VT, N0,
2361 DAG.getConstant(N1C->getAPIntValue() - 1, DL, VT));
2363 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2364 if (N1.getOpcode() == ISD::SHL) {
2365 if (ConstantSDNode *SHC = getAsNonOpaqueConstant(N1.getOperand(0))) {
2366 if (SHC->getAPIntValue().isPowerOf2()) {
2369 DAG.getNode(ISD::ADD, DL, VT, N1,
2370 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL,
2372 AddToWorklist(Add.getNode());
2373 return DAG.getNode(ISD::AND, DL, VT, N0, Add);
2378 // If X/C can be simplified by the division-by-constant logic, lower
2379 // X%C to the equivalent of X-X/C*C.
2380 if (N1C && !N1C->isNullValue()) {
2381 SDValue Div = DAG.getNode(ISD::UDIV, SDLoc(N), VT, N0, N1);
2382 AddToWorklist(Div.getNode());
2383 SDValue OptimizedDiv = combine(Div.getNode());
2384 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2385 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2387 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2388 AddToWorklist(Mul.getNode());
2394 if (N0.getOpcode() == ISD::UNDEF)
2395 return DAG.getConstant(0, SDLoc(N), VT);
2396 // X % undef -> undef
2397 if (N1.getOpcode() == ISD::UNDEF)
2403 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2404 SDValue N0 = N->getOperand(0);
2405 SDValue N1 = N->getOperand(1);
2406 EVT VT = N->getValueType(0);
2409 // fold (mulhs x, 0) -> 0
2410 if (isNullConstant(N1))
2412 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2413 if (isOneConstant(N1)) {
2415 return DAG.getNode(ISD::SRA, DL, N0.getValueType(), N0,
2416 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2418 getShiftAmountTy(N0.getValueType())));
2420 // fold (mulhs x, undef) -> 0
2421 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2422 return DAG.getConstant(0, SDLoc(N), VT);
2424 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2426 if (VT.isSimple() && !VT.isVector()) {
2427 MVT Simple = VT.getSimpleVT();
2428 unsigned SimpleSize = Simple.getSizeInBits();
2429 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2430 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2431 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2432 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2433 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2434 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2435 DAG.getConstant(SimpleSize, DL,
2436 getShiftAmountTy(N1.getValueType())));
2437 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2444 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2445 SDValue N0 = N->getOperand(0);
2446 SDValue N1 = N->getOperand(1);
2447 EVT VT = N->getValueType(0);
2450 // fold (mulhu x, 0) -> 0
2451 if (isNullConstant(N1))
2453 // fold (mulhu x, 1) -> 0
2454 if (isOneConstant(N1))
2455 return DAG.getConstant(0, DL, N0.getValueType());
2456 // fold (mulhu x, undef) -> 0
2457 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2458 return DAG.getConstant(0, DL, VT);
2460 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2462 if (VT.isSimple() && !VT.isVector()) {
2463 MVT Simple = VT.getSimpleVT();
2464 unsigned SimpleSize = Simple.getSizeInBits();
2465 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2466 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2467 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2468 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2469 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2470 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2471 DAG.getConstant(SimpleSize, DL,
2472 getShiftAmountTy(N1.getValueType())));
2473 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2480 /// Perform optimizations common to nodes that compute two values. LoOp and HiOp
2481 /// give the opcodes for the two computations that are being performed. Return
2482 /// true if a simplification was made.
2483 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2485 // If the high half is not needed, just compute the low half.
2486 bool HiExists = N->hasAnyUseOfValue(1);
2488 (!LegalOperations ||
2489 TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) {
2490 SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2491 return CombineTo(N, Res, Res);
2494 // If the low half is not needed, just compute the high half.
2495 bool LoExists = N->hasAnyUseOfValue(0);
2497 (!LegalOperations ||
2498 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2499 SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2500 return CombineTo(N, Res, Res);
2503 // If both halves are used, return as it is.
2504 if (LoExists && HiExists)
2507 // If the two computed results can be simplified separately, separate them.
2509 SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2510 AddToWorklist(Lo.getNode());
2511 SDValue LoOpt = combine(Lo.getNode());
2512 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2513 (!LegalOperations ||
2514 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2515 return CombineTo(N, LoOpt, LoOpt);
2519 SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2520 AddToWorklist(Hi.getNode());
2521 SDValue HiOpt = combine(Hi.getNode());
2522 if (HiOpt.getNode() && HiOpt != Hi &&
2523 (!LegalOperations ||
2524 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2525 return CombineTo(N, HiOpt, HiOpt);
2531 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2532 if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS))
2535 EVT VT = N->getValueType(0);
2538 // If the type is twice as wide is legal, transform the mulhu to a wider
2539 // multiply plus a shift.
2540 if (VT.isSimple() && !VT.isVector()) {
2541 MVT Simple = VT.getSimpleVT();
2542 unsigned SimpleSize = Simple.getSizeInBits();
2543 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2544 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2545 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2546 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2547 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2548 // Compute the high part as N1.
2549 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2550 DAG.getConstant(SimpleSize, DL,
2551 getShiftAmountTy(Lo.getValueType())));
2552 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2553 // Compute the low part as N0.
2554 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2555 return CombineTo(N, Lo, Hi);
2562 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2563 if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU))
2566 EVT VT = N->getValueType(0);
2569 // If the type is twice as wide is legal, transform the mulhu to a wider
2570 // multiply plus a shift.
2571 if (VT.isSimple() && !VT.isVector()) {
2572 MVT Simple = VT.getSimpleVT();
2573 unsigned SimpleSize = Simple.getSizeInBits();
2574 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2575 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2576 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2577 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2578 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2579 // Compute the high part as N1.
2580 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2581 DAG.getConstant(SimpleSize, DL,
2582 getShiftAmountTy(Lo.getValueType())));
2583 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2584 // Compute the low part as N0.
2585 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2586 return CombineTo(N, Lo, Hi);
2593 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2594 // (smulo x, 2) -> (saddo x, x)
2595 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2596 if (C2->getAPIntValue() == 2)
2597 return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(),
2598 N->getOperand(0), N->getOperand(0));
2603 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2604 // (umulo x, 2) -> (uaddo x, x)
2605 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2606 if (C2->getAPIntValue() == 2)
2607 return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(),
2608 N->getOperand(0), N->getOperand(0));
2613 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2614 if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM))
2620 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2621 if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM))
2627 /// If this is a binary operator with two operands of the same opcode, try to
2629 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2630 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2631 EVT VT = N0.getValueType();
2632 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2634 // Bail early if none of these transforms apply.
2635 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2637 // For each of OP in AND/OR/XOR:
2638 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2639 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2640 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2641 // fold (OP (bswap x), (bswap y)) -> (bswap (OP x, y))
2642 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2644 // do not sink logical op inside of a vector extend, since it may combine
2646 EVT Op0VT = N0.getOperand(0).getValueType();
2647 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2648 N0.getOpcode() == ISD::SIGN_EXTEND ||
2649 N0.getOpcode() == ISD::BSWAP ||
2650 // Avoid infinite looping with PromoteIntBinOp.
2651 (N0.getOpcode() == ISD::ANY_EXTEND &&
2652 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2653 (N0.getOpcode() == ISD::TRUNCATE &&
2654 (!TLI.isZExtFree(VT, Op0VT) ||
2655 !TLI.isTruncateFree(Op0VT, VT)) &&
2656 TLI.isTypeLegal(Op0VT))) &&
2658 Op0VT == N1.getOperand(0).getValueType() &&
2659 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2660 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2661 N0.getOperand(0).getValueType(),
2662 N0.getOperand(0), N1.getOperand(0));
2663 AddToWorklist(ORNode.getNode());
2664 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode);
2667 // For each of OP in SHL/SRL/SRA/AND...
2668 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2669 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2670 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2671 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2672 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2673 N0.getOperand(1) == N1.getOperand(1)) {
2674 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2675 N0.getOperand(0).getValueType(),
2676 N0.getOperand(0), N1.getOperand(0));
2677 AddToWorklist(ORNode.getNode());
2678 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
2679 ORNode, N0.getOperand(1));
2682 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2683 // Only perform this optimization after type legalization and before
2684 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2685 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2686 // we don't want to undo this promotion.
2687 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2689 if ((N0.getOpcode() == ISD::BITCAST ||
2690 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) &&
2691 Level == AfterLegalizeTypes) {
2692 SDValue In0 = N0.getOperand(0);
2693 SDValue In1 = N1.getOperand(0);
2694 EVT In0Ty = In0.getValueType();
2695 EVT In1Ty = In1.getValueType();
2697 // If both incoming values are integers, and the original types are the
2699 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2700 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1);
2701 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op);
2702 AddToWorklist(Op.getNode());
2707 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2708 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2709 // If both shuffles use the same mask, and both shuffle within a single
2710 // vector, then it is worthwhile to move the swizzle after the operation.
2711 // The type-legalizer generates this pattern when loading illegal
2712 // vector types from memory. In many cases this allows additional shuffle
2714 // There are other cases where moving the shuffle after the xor/and/or
2715 // is profitable even if shuffles don't perform a swizzle.
2716 // If both shuffles use the same mask, and both shuffles have the same first
2717 // or second operand, then it might still be profitable to move the shuffle
2718 // after the xor/and/or operation.
2719 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) {
2720 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2721 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2723 assert(N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() &&
2724 "Inputs to shuffles are not the same type");
2726 // Check that both shuffles use the same mask. The masks are known to be of
2727 // the same length because the result vector type is the same.
2728 // Check also that shuffles have only one use to avoid introducing extra
2730 if (SVN0->hasOneUse() && SVN1->hasOneUse() &&
2731 SVN0->getMask().equals(SVN1->getMask())) {
2732 SDValue ShOp = N0->getOperand(1);
2734 // Don't try to fold this node if it requires introducing a
2735 // build vector of all zeros that might be illegal at this stage.
2736 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2738 ShOp = DAG.getConstant(0, SDLoc(N), VT);
2743 // (AND (shuf (A, C), shuf (B, C)) -> shuf (AND (A, B), C)
2744 // (OR (shuf (A, C), shuf (B, C)) -> shuf (OR (A, B), C)
2745 // (XOR (shuf (A, C), shuf (B, C)) -> shuf (XOR (A, B), V_0)
2746 if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) {
2747 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2748 N0->getOperand(0), N1->getOperand(0));
2749 AddToWorklist(NewNode.getNode());
2750 return DAG.getVectorShuffle(VT, SDLoc(N), NewNode, ShOp,
2751 &SVN0->getMask()[0]);
2754 // Don't try to fold this node if it requires introducing a
2755 // build vector of all zeros that might be illegal at this stage.
2756 ShOp = N0->getOperand(0);
2757 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2759 ShOp = DAG.getConstant(0, SDLoc(N), VT);
2764 // (AND (shuf (C, A), shuf (C, B)) -> shuf (C, AND (A, B))
2765 // (OR (shuf (C, A), shuf (C, B)) -> shuf (C, OR (A, B))
2766 // (XOR (shuf (C, A), shuf (C, B)) -> shuf (V_0, XOR (A, B))
2767 if (N0->getOperand(0) == N1->getOperand(0) && ShOp.getNode()) {
2768 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2769 N0->getOperand(1), N1->getOperand(1));
2770 AddToWorklist(NewNode.getNode());
2771 return DAG.getVectorShuffle(VT, SDLoc(N), ShOp, NewNode,
2772 &SVN0->getMask()[0]);
2780 /// This contains all DAGCombine rules which reduce two values combined by
2781 /// an And operation to a single value. This makes them reusable in the context
2782 /// of visitSELECT(). Rules involving constants are not included as
2783 /// visitSELECT() already handles those cases.
2784 SDValue DAGCombiner::visitANDLike(SDValue N0, SDValue N1,
2785 SDNode *LocReference) {
2786 EVT VT = N1.getValueType();
2788 // fold (and x, undef) -> 0
2789 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2790 return DAG.getConstant(0, SDLoc(LocReference), VT);
2791 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2792 SDValue LL, LR, RL, RR, CC0, CC1;
2793 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2794 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2795 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2797 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2798 LL.getValueType().isInteger()) {
2799 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2800 if (isNullConstant(LR) && Op1 == ISD::SETEQ) {
2801 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2802 LR.getValueType(), LL, RL);
2803 AddToWorklist(ORNode.getNode());
2804 return DAG.getSetCC(SDLoc(LocReference), VT, ORNode, LR, Op1);
2806 if (isAllOnesConstant(LR)) {
2807 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2808 if (Op1 == ISD::SETEQ) {
2809 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(N0),
2810 LR.getValueType(), LL, RL);
2811 AddToWorklist(ANDNode.getNode());
2812 return DAG.getSetCC(SDLoc(LocReference), VT, ANDNode, LR, Op1);
2814 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2815 if (Op1 == ISD::SETGT) {
2816 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2817 LR.getValueType(), LL, RL);
2818 AddToWorklist(ORNode.getNode());
2819 return DAG.getSetCC(SDLoc(LocReference), VT, ORNode, LR, Op1);
2823 // Simplify (and (setne X, 0), (setne X, -1)) -> (setuge (add X, 1), 2)
2824 if (LL == RL && isa<ConstantSDNode>(LR) && isa<ConstantSDNode>(RR) &&
2825 Op0 == Op1 && LL.getValueType().isInteger() &&
2826 Op0 == ISD::SETNE && ((isNullConstant(LR) && isAllOnesConstant(RR)) ||
2827 (isAllOnesConstant(LR) && isNullConstant(RR)))) {
2829 SDValue ADDNode = DAG.getNode(ISD::ADD, DL, LL.getValueType(),
2830 LL, DAG.getConstant(1, DL,
2831 LL.getValueType()));
2832 AddToWorklist(ADDNode.getNode());
2833 return DAG.getSetCC(SDLoc(LocReference), VT, ADDNode,
2834 DAG.getConstant(2, DL, LL.getValueType()),
2837 // canonicalize equivalent to ll == rl
2838 if (LL == RR && LR == RL) {
2839 Op1 = ISD::getSetCCSwappedOperands(Op1);
2842 if (LL == RL && LR == RR) {
2843 bool isInteger = LL.getValueType().isInteger();
2844 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2845 if (Result != ISD::SETCC_INVALID &&
2846 (!LegalOperations ||
2847 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
2848 TLI.isOperationLegal(ISD::SETCC,
2849 getSetCCResultType(N0.getSimpleValueType())))))
2850 return DAG.getSetCC(SDLoc(LocReference), N0.getValueType(),
2855 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
2856 VT.getSizeInBits() <= 64) {
2857 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2858 APInt ADDC = ADDI->getAPIntValue();
2859 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2860 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
2861 // immediate for an add, but it is legal if its top c2 bits are set,
2862 // transform the ADD so the immediate doesn't need to be materialized
2864 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
2865 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
2866 SRLI->getZExtValue());
2867 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
2869 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2872 DAG.getNode(ISD::ADD, DL, VT,
2873 N0.getOperand(0), DAG.getConstant(ADDC, DL, VT));
2874 CombineTo(N0.getNode(), NewAdd);
2875 // Return N so it doesn't get rechecked!
2876 return SDValue(LocReference, 0);
2887 SDValue DAGCombiner::visitAND(SDNode *N) {
2888 SDValue N0 = N->getOperand(0);
2889 SDValue N1 = N->getOperand(1);
2890 EVT VT = N1.getValueType();
2893 if (VT.isVector()) {
2894 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2897 // fold (and x, 0) -> 0, vector edition
2898 if (ISD::isBuildVectorAllZeros(N0.getNode()))
2899 // do not return N0, because undef node may exist in N0
2900 return DAG.getConstant(
2901 APInt::getNullValue(
2902 N0.getValueType().getScalarType().getSizeInBits()),
2903 SDLoc(N), N0.getValueType());
2904 if (ISD::isBuildVectorAllZeros(N1.getNode()))
2905 // do not return N1, because undef node may exist in N1
2906 return DAG.getConstant(
2907 APInt::getNullValue(
2908 N1.getValueType().getScalarType().getSizeInBits()),
2909 SDLoc(N), N1.getValueType());
2911 // fold (and x, -1) -> x, vector edition
2912 if (ISD::isBuildVectorAllOnes(N0.getNode()))
2914 if (ISD::isBuildVectorAllOnes(N1.getNode()))
2918 // fold (and c1, c2) -> c1&c2
2919 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
2920 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2921 if (N0C && N1C && !N1C->isOpaque())
2922 return DAG.FoldConstantArithmetic(ISD::AND, SDLoc(N), VT, N0C, N1C);
2923 // canonicalize constant to RHS
2924 if (isConstantIntBuildVectorOrConstantInt(N0) &&
2925 !isConstantIntBuildVectorOrConstantInt(N1))
2926 return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0);
2927 // fold (and x, -1) -> x
2928 if (isAllOnesConstant(N1))
2930 // if (and x, c) is known to be zero, return 0
2931 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2932 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2933 APInt::getAllOnesValue(BitWidth)))
2934 return DAG.getConstant(0, SDLoc(N), VT);
2936 if (SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1))
2938 // fold (and (or x, C), D) -> D if (C & D) == D
2939 if (N1C && N0.getOpcode() == ISD::OR)
2940 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2941 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2943 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2944 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2945 SDValue N0Op0 = N0.getOperand(0);
2946 APInt Mask = ~N1C->getAPIntValue();
2947 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2948 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2949 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
2950 N0.getValueType(), N0Op0);
2952 // Replace uses of the AND with uses of the Zero extend node.
2955 // We actually want to replace all uses of the any_extend with the
2956 // zero_extend, to avoid duplicating things. This will later cause this
2957 // AND to be folded.
2958 CombineTo(N0.getNode(), Zext);
2959 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2962 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2963 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2964 // already be zero by virtue of the width of the base type of the load.
2966 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2968 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2969 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2970 N0.getOpcode() == ISD::LOAD) {
2971 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2972 N0 : N0.getOperand(0) );
2974 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2975 // This can be a pure constant or a vector splat, in which case we treat the
2976 // vector as a scalar and use the splat value.
2977 APInt Constant = APInt::getNullValue(1);
2978 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2979 Constant = C->getAPIntValue();
2980 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2981 APInt SplatValue, SplatUndef;
2982 unsigned SplatBitSize;
2984 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2985 SplatBitSize, HasAnyUndefs);
2987 // Undef bits can contribute to a possible optimisation if set, so
2989 SplatValue |= SplatUndef;
2991 // The splat value may be something like "0x00FFFFFF", which means 0 for
2992 // the first vector value and FF for the rest, repeating. We need a mask
2993 // that will apply equally to all members of the vector, so AND all the
2994 // lanes of the constant together.
2995 EVT VT = Vector->getValueType(0);
2996 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2998 // If the splat value has been compressed to a bitlength lower
2999 // than the size of the vector lane, we need to re-expand it to
3001 if (BitWidth > SplatBitSize)
3002 for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
3003 SplatBitSize < BitWidth;
3004 SplatBitSize = SplatBitSize * 2)
3005 SplatValue |= SplatValue.shl(SplatBitSize);
3007 // Make sure that variable 'Constant' is only set if 'SplatBitSize' is a
3008 // multiple of 'BitWidth'. Otherwise, we could propagate a wrong value.
3009 if (SplatBitSize % BitWidth == 0) {
3010 Constant = APInt::getAllOnesValue(BitWidth);
3011 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
3012 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
3017 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
3018 // actually legal and isn't going to get expanded, else this is a false
3020 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
3021 Load->getValueType(0),
3022 Load->getMemoryVT());
3024 // Resize the constant to the same size as the original memory access before
3025 // extension. If it is still the AllOnesValue then this AND is completely
3028 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
3031 switch (Load->getExtensionType()) {
3032 default: B = false; break;
3033 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
3035 case ISD::NON_EXTLOAD: B = true; break;
3038 if (B && Constant.isAllOnesValue()) {
3039 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
3040 // preserve semantics once we get rid of the AND.
3041 SDValue NewLoad(Load, 0);
3042 if (Load->getExtensionType() == ISD::EXTLOAD) {
3043 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
3044 Load->getValueType(0), SDLoc(Load),
3045 Load->getChain(), Load->getBasePtr(),
3046 Load->getOffset(), Load->getMemoryVT(),
3047 Load->getMemOperand());
3048 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
3049 if (Load->getNumValues() == 3) {
3050 // PRE/POST_INC loads have 3 values.
3051 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
3052 NewLoad.getValue(2) };
3053 CombineTo(Load, To, 3, true);
3055 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
3059 // Fold the AND away, taking care not to fold to the old load node if we
3061 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
3063 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3067 // fold (and (load x), 255) -> (zextload x, i8)
3068 // fold (and (extload x, i16), 255) -> (zextload x, i8)
3069 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
3070 if (N1C && (N0.getOpcode() == ISD::LOAD ||
3071 (N0.getOpcode() == ISD::ANY_EXTEND &&
3072 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
3073 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
3074 LoadSDNode *LN0 = HasAnyExt
3075 ? cast<LoadSDNode>(N0.getOperand(0))
3076 : cast<LoadSDNode>(N0);
3077 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
3078 LN0->isUnindexed() && N0.hasOneUse() && SDValue(LN0, 0).hasOneUse()) {
3079 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
3080 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
3081 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
3082 EVT LoadedVT = LN0->getMemoryVT();
3083 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
3085 if (ExtVT == LoadedVT &&
3086 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy,
3090 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
3091 LN0->getChain(), LN0->getBasePtr(), ExtVT,
3092 LN0->getMemOperand());
3094 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
3095 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3098 // Do not change the width of a volatile load.
3099 // Do not generate loads of non-round integer types since these can
3100 // be expensive (and would be wrong if the type is not byte sized).
3101 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
3102 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy,
3104 EVT PtrType = LN0->getOperand(1).getValueType();
3106 unsigned Alignment = LN0->getAlignment();
3107 SDValue NewPtr = LN0->getBasePtr();
3109 // For big endian targets, we need to add an offset to the pointer
3110 // to load the correct bytes. For little endian systems, we merely
3111 // need to read fewer bytes from the same pointer.
3112 if (DAG.getDataLayout().isBigEndian()) {
3113 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
3114 unsigned EVTStoreBytes = ExtVT.getStoreSize();
3115 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
3117 NewPtr = DAG.getNode(ISD::ADD, DL, PtrType,
3118 NewPtr, DAG.getConstant(PtrOff, DL, PtrType));
3119 Alignment = MinAlign(Alignment, PtrOff);
3122 AddToWorklist(NewPtr.getNode());
3125 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
3126 LN0->getChain(), NewPtr,
3127 LN0->getPointerInfo(),
3128 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
3129 LN0->isInvariant(), Alignment, LN0->getAAInfo());
3131 CombineTo(LN0, Load, Load.getValue(1));
3132 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3138 if (SDValue Combined = visitANDLike(N0, N1, N))
3141 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
3142 if (N0.getOpcode() == N1.getOpcode())
3143 if (SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N))
3146 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
3147 // fold (and (sra)) -> (and (srl)) when possible.
3148 if (!VT.isVector() &&
3149 SimplifyDemandedBits(SDValue(N, 0)))
3150 return SDValue(N, 0);
3152 // fold (zext_inreg (extload x)) -> (zextload x)
3153 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
3154 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3155 EVT MemVT = LN0->getMemoryVT();
3156 // If we zero all the possible extended bits, then we can turn this into
3157 // a zextload if we are running before legalize or the operation is legal.
3158 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
3159 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
3160 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
3161 ((!LegalOperations && !LN0->isVolatile()) ||
3162 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) {
3163 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
3164 LN0->getChain(), LN0->getBasePtr(),
3165 MemVT, LN0->getMemOperand());
3167 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
3168 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3171 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
3172 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
3174 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3175 EVT MemVT = LN0->getMemoryVT();
3176 // If we zero all the possible extended bits, then we can turn this into
3177 // a zextload if we are running before legalize or the operation is legal.
3178 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
3179 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
3180 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
3181 ((!LegalOperations && !LN0->isVolatile()) ||
3182 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) {
3183 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
3184 LN0->getChain(), LN0->getBasePtr(),
3185 MemVT, LN0->getMemOperand());
3187 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
3188 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3191 // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const)
3192 if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) {
3193 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
3194 N0.getOperand(1), false);
3195 if (BSwap.getNode())
3202 /// Match (a >> 8) | (a << 8) as (bswap a) >> 16.
3203 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
3204 bool DemandHighBits) {
3205 if (!LegalOperations)
3208 EVT VT = N->getValueType(0);
3209 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
3211 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3214 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
3215 bool LookPassAnd0 = false;
3216 bool LookPassAnd1 = false;
3217 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
3219 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
3221 if (N0.getOpcode() == ISD::AND) {
3222 if (!N0.getNode()->hasOneUse())
3224 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3225 if (!N01C || N01C->getZExtValue() != 0xFF00)
3227 N0 = N0.getOperand(0);
3228 LookPassAnd0 = true;
3231 if (N1.getOpcode() == ISD::AND) {
3232 if (!N1.getNode()->hasOneUse())
3234 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3235 if (!N11C || N11C->getZExtValue() != 0xFF)
3237 N1 = N1.getOperand(0);
3238 LookPassAnd1 = true;
3241 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
3243 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
3245 if (!N0.getNode()->hasOneUse() ||
3246 !N1.getNode()->hasOneUse())
3249 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3250 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3253 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
3256 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
3257 SDValue N00 = N0->getOperand(0);
3258 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
3259 if (!N00.getNode()->hasOneUse())
3261 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
3262 if (!N001C || N001C->getZExtValue() != 0xFF)
3264 N00 = N00.getOperand(0);
3265 LookPassAnd0 = true;
3268 SDValue N10 = N1->getOperand(0);
3269 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
3270 if (!N10.getNode()->hasOneUse())
3272 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
3273 if (!N101C || N101C->getZExtValue() != 0xFF00)
3275 N10 = N10.getOperand(0);
3276 LookPassAnd1 = true;
3282 // Make sure everything beyond the low halfword gets set to zero since the SRL
3283 // 16 will clear the top bits.
3284 unsigned OpSizeInBits = VT.getSizeInBits();
3285 if (DemandHighBits && OpSizeInBits > 16) {
3286 // If the left-shift isn't masked out then the only way this is a bswap is
3287 // if all bits beyond the low 8 are 0. In that case the entire pattern
3288 // reduces to a left shift anyway: leave it for other parts of the combiner.
3292 // However, if the right shift isn't masked out then it might be because
3293 // it's not needed. See if we can spot that too.
3294 if (!LookPassAnd1 &&
3295 !DAG.MaskedValueIsZero(
3296 N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16)))
3300 SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00);
3301 if (OpSizeInBits > 16) {
3303 Res = DAG.getNode(ISD::SRL, DL, VT, Res,
3304 DAG.getConstant(OpSizeInBits - 16, DL,
3305 getShiftAmountTy(VT)));
3310 /// Return true if the specified node is an element that makes up a 32-bit
3311 /// packed halfword byteswap.
3312 /// ((x & 0x000000ff) << 8) |
3313 /// ((x & 0x0000ff00) >> 8) |
3314 /// ((x & 0x00ff0000) << 8) |
3315 /// ((x & 0xff000000) >> 8)
3316 static bool isBSwapHWordElement(SDValue N, MutableArrayRef<SDNode *> Parts) {
3317 if (!N.getNode()->hasOneUse())
3320 unsigned Opc = N.getOpcode();
3321 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
3324 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3329 switch (N1C->getZExtValue()) {
3332 case 0xFF: Num = 0; break;
3333 case 0xFF00: Num = 1; break;
3334 case 0xFF0000: Num = 2; break;
3335 case 0xFF000000: Num = 3; break;
3338 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
3339 SDValue N0 = N.getOperand(0);
3340 if (Opc == ISD::AND) {
3341 if (Num == 0 || Num == 2) {
3343 // (x >> 8) & 0xff0000
3344 if (N0.getOpcode() != ISD::SRL)
3346 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3347 if (!C || C->getZExtValue() != 8)
3350 // (x << 8) & 0xff00
3351 // (x << 8) & 0xff000000
3352 if (N0.getOpcode() != ISD::SHL)
3354 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3355 if (!C || C->getZExtValue() != 8)
3358 } else if (Opc == ISD::SHL) {
3360 // (x & 0xff0000) << 8
3361 if (Num != 0 && Num != 2)
3363 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3364 if (!C || C->getZExtValue() != 8)
3366 } else { // Opc == ISD::SRL
3367 // (x & 0xff00) >> 8
3368 // (x & 0xff000000) >> 8
3369 if (Num != 1 && Num != 3)
3371 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3372 if (!C || C->getZExtValue() != 8)
3379 Parts[Num] = N0.getOperand(0).getNode();
3383 /// Match a 32-bit packed halfword bswap. That is
3384 /// ((x & 0x000000ff) << 8) |
3385 /// ((x & 0x0000ff00) >> 8) |
3386 /// ((x & 0x00ff0000) << 8) |
3387 /// ((x & 0xff000000) >> 8)
3388 /// => (rotl (bswap x), 16)
3389 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
3390 if (!LegalOperations)
3393 EVT VT = N->getValueType(0);
3396 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3400 // (or (or (and), (and)), (or (and), (and)))
3401 // (or (or (or (and), (and)), (and)), (and))
3402 if (N0.getOpcode() != ISD::OR)
3404 SDValue N00 = N0.getOperand(0);
3405 SDValue N01 = N0.getOperand(1);
3406 SDNode *Parts[4] = {};
3408 if (N1.getOpcode() == ISD::OR &&
3409 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) {
3410 // (or (or (and), (and)), (or (and), (and)))
3411 SDValue N000 = N00.getOperand(0);
3412 if (!isBSwapHWordElement(N000, Parts))
3415 SDValue N001 = N00.getOperand(1);
3416 if (!isBSwapHWordElement(N001, Parts))
3418 SDValue N010 = N01.getOperand(0);
3419 if (!isBSwapHWordElement(N010, Parts))
3421 SDValue N011 = N01.getOperand(1);
3422 if (!isBSwapHWordElement(N011, Parts))
3425 // (or (or (or (and), (and)), (and)), (and))
3426 if (!isBSwapHWordElement(N1, Parts))
3428 if (!isBSwapHWordElement(N01, Parts))
3430 if (N00.getOpcode() != ISD::OR)
3432 SDValue N000 = N00.getOperand(0);
3433 if (!isBSwapHWordElement(N000, Parts))
3435 SDValue N001 = N00.getOperand(1);
3436 if (!isBSwapHWordElement(N001, Parts))
3440 // Make sure the parts are all coming from the same node.
3441 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
3445 SDValue BSwap = DAG.getNode(ISD::BSWAP, DL, VT,
3446 SDValue(Parts[0], 0));
3448 // Result of the bswap should be rotated by 16. If it's not legal, then
3449 // do (x << 16) | (x >> 16).
3450 SDValue ShAmt = DAG.getConstant(16, DL, getShiftAmountTy(VT));
3451 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3452 return DAG.getNode(ISD::ROTL, DL, VT, BSwap, ShAmt);
3453 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3454 return DAG.getNode(ISD::ROTR, DL, VT, BSwap, ShAmt);
3455 return DAG.getNode(ISD::OR, DL, VT,
3456 DAG.getNode(ISD::SHL, DL, VT, BSwap, ShAmt),
3457 DAG.getNode(ISD::SRL, DL, VT, BSwap, ShAmt));
3460 /// This contains all DAGCombine rules which reduce two values combined by
3461 /// an Or operation to a single value \see visitANDLike().
3462 SDValue DAGCombiner::visitORLike(SDValue N0, SDValue N1, SDNode *LocReference) {
3463 EVT VT = N1.getValueType();
3464 // fold (or x, undef) -> -1
3465 if (!LegalOperations &&
3466 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3467 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3468 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()),
3469 SDLoc(LocReference), VT);
3471 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3472 SDValue LL, LR, RL, RR, CC0, CC1;
3473 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3474 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3475 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3477 if (LR == RR && Op0 == Op1 && LL.getValueType().isInteger()) {
3478 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3479 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3480 if (isNullConstant(LR) && (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3481 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(LR),
3482 LR.getValueType(), LL, RL);
3483 AddToWorklist(ORNode.getNode());
3484 return DAG.getSetCC(SDLoc(LocReference), VT, ORNode, LR, Op1);
3486 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3487 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3488 if (isAllOnesConstant(LR) && (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3489 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(LR),
3490 LR.getValueType(), LL, RL);
3491 AddToWorklist(ANDNode.getNode());
3492 return DAG.getSetCC(SDLoc(LocReference), VT, ANDNode, LR, Op1);
3495 // canonicalize equivalent to ll == rl
3496 if (LL == RR && LR == RL) {
3497 Op1 = ISD::getSetCCSwappedOperands(Op1);
3500 if (LL == RL && LR == RR) {
3501 bool isInteger = LL.getValueType().isInteger();
3502 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3503 if (Result != ISD::SETCC_INVALID &&
3504 (!LegalOperations ||
3505 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
3506 TLI.isOperationLegal(ISD::SETCC,
3507 getSetCCResultType(N0.getValueType())))))
3508 return DAG.getSetCC(SDLoc(LocReference), N0.getValueType(),
3513 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3514 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == ISD::AND &&
3515 // Don't increase # computations.
3516 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3517 // We can only do this xform if we know that bits from X that are set in C2
3518 // but not in C1 are already zero. Likewise for Y.
3519 if (const ConstantSDNode *N0O1C =
3520 getAsNonOpaqueConstant(N0.getOperand(1))) {
3521 if (const ConstantSDNode *N1O1C =
3522 getAsNonOpaqueConstant(N1.getOperand(1))) {
3523 // We can only do this xform if we know that bits from X that are set in
3524 // C2 but not in C1 are already zero. Likewise for Y.
3525 const APInt &LHSMask = N0O1C->getAPIntValue();
3526 const APInt &RHSMask = N1O1C->getAPIntValue();
3528 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3529 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3530 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3531 N0.getOperand(0), N1.getOperand(0));
3532 SDLoc DL(LocReference);
3533 return DAG.getNode(ISD::AND, DL, VT, X,
3534 DAG.getConstant(LHSMask | RHSMask, DL, VT));
3540 // (or (and X, M), (and X, N)) -> (and X, (or M, N))
3541 if (N0.getOpcode() == ISD::AND &&
3542 N1.getOpcode() == ISD::AND &&
3543 N0.getOperand(0) == N1.getOperand(0) &&
3544 // Don't increase # computations.
3545 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3546 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3547 N0.getOperand(1), N1.getOperand(1));
3548 return DAG.getNode(ISD::AND, SDLoc(LocReference), VT, N0.getOperand(0), X);
3554 SDValue DAGCombiner::visitOR(SDNode *N) {
3555 SDValue N0 = N->getOperand(0);
3556 SDValue N1 = N->getOperand(1);
3557 EVT VT = N1.getValueType();
3560 if (VT.isVector()) {
3561 if (SDValue FoldedVOp = SimplifyVBinOp(N))
3564 // fold (or x, 0) -> x, vector edition
3565 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3567 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3570 // fold (or x, -1) -> -1, vector edition
3571 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3572 // do not return N0, because undef node may exist in N0
3573 return DAG.getConstant(
3574 APInt::getAllOnesValue(
3575 N0.getValueType().getScalarType().getSizeInBits()),
3576 SDLoc(N), N0.getValueType());
3577 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3578 // do not return N1, because undef node may exist in N1
3579 return DAG.getConstant(
3580 APInt::getAllOnesValue(
3581 N1.getValueType().getScalarType().getSizeInBits()),
3582 SDLoc(N), N1.getValueType());
3584 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask1)
3585 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf B, A, Mask2)
3586 // Do this only if the resulting shuffle is legal.
3587 if (isa<ShuffleVectorSDNode>(N0) &&
3588 isa<ShuffleVectorSDNode>(N1) &&
3589 // Avoid folding a node with illegal type.
3590 TLI.isTypeLegal(VT) &&
3591 N0->getOperand(1) == N1->getOperand(1) &&
3592 ISD::isBuildVectorAllZeros(N0.getOperand(1).getNode())) {
3593 bool CanFold = true;
3594 unsigned NumElts = VT.getVectorNumElements();
3595 const ShuffleVectorSDNode *SV0 = cast<ShuffleVectorSDNode>(N0);
3596 const ShuffleVectorSDNode *SV1 = cast<ShuffleVectorSDNode>(N1);
3597 // We construct two shuffle masks:
3598 // - Mask1 is a shuffle mask for a shuffle with N0 as the first operand
3599 // and N1 as the second operand.
3600 // - Mask2 is a shuffle mask for a shuffle with N1 as the first operand
3601 // and N0 as the second operand.
3602 // We do this because OR is commutable and therefore there might be
3603 // two ways to fold this node into a shuffle.
3604 SmallVector<int,4> Mask1;
3605 SmallVector<int,4> Mask2;
3607 for (unsigned i = 0; i != NumElts && CanFold; ++i) {
3608 int M0 = SV0->getMaskElt(i);
3609 int M1 = SV1->getMaskElt(i);
3611 // Both shuffle indexes are undef. Propagate Undef.
3612 if (M0 < 0 && M1 < 0) {
3613 Mask1.push_back(M0);
3614 Mask2.push_back(M0);
3618 if (M0 < 0 || M1 < 0 ||
3619 (M0 < (int)NumElts && M1 < (int)NumElts) ||
3620 (M0 >= (int)NumElts && M1 >= (int)NumElts)) {
3625 Mask1.push_back(M0 < (int)NumElts ? M0 : M1 + NumElts);
3626 Mask2.push_back(M1 < (int)NumElts ? M1 : M0 + NumElts);
3630 // Fold this sequence only if the resulting shuffle is 'legal'.
3631 if (TLI.isShuffleMaskLegal(Mask1, VT))
3632 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0),
3633 N1->getOperand(0), &Mask1[0]);
3634 if (TLI.isShuffleMaskLegal(Mask2, VT))
3635 return DAG.getVectorShuffle(VT, SDLoc(N), N1->getOperand(0),
3636 N0->getOperand(0), &Mask2[0]);
3641 // fold (or c1, c2) -> c1|c2
3642 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
3643 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3644 if (N0C && N1C && !N1C->isOpaque())
3645 return DAG.FoldConstantArithmetic(ISD::OR, SDLoc(N), VT, N0C, N1C);
3646 // canonicalize constant to RHS
3647 if (isConstantIntBuildVectorOrConstantInt(N0) &&
3648 !isConstantIntBuildVectorOrConstantInt(N1))
3649 return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0);
3650 // fold (or x, 0) -> x
3651 if (isNullConstant(N1))
3653 // fold (or x, -1) -> -1
3654 if (isAllOnesConstant(N1))
3656 // fold (or x, c) -> c iff (x & ~c) == 0
3657 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3660 if (SDValue Combined = visitORLike(N0, N1, N))
3663 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3664 if (SDValue BSwap = MatchBSwapHWord(N, N0, N1))
3666 if (SDValue BSwap = MatchBSwapHWordLow(N, N0, N1))
3670 if (SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1))
3672 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3673 // iff (c1 & c2) == 0.
3674 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3675 isa<ConstantSDNode>(N0.getOperand(1))) {
3676 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3677 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) {
3678 if (SDValue COR = DAG.FoldConstantArithmetic(ISD::OR, SDLoc(N1), VT,
3681 ISD::AND, SDLoc(N), VT,
3682 DAG.getNode(ISD::OR, SDLoc(N0), VT, N0.getOperand(0), N1), COR);
3686 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3687 if (N0.getOpcode() == N1.getOpcode())
3688 if (SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N))
3691 // See if this is some rotate idiom.
3692 if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N)))
3693 return SDValue(Rot, 0);
3695 // Simplify the operands using demanded-bits information.
3696 if (!VT.isVector() &&
3697 SimplifyDemandedBits(SDValue(N, 0)))
3698 return SDValue(N, 0);
3703 /// Match "(X shl/srl V1) & V2" where V2 may not be present.
3704 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3705 if (Op.getOpcode() == ISD::AND) {
3706 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3707 Mask = Op.getOperand(1);
3708 Op = Op.getOperand(0);
3714 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3722 // Return true if we can prove that, whenever Neg and Pos are both in the
3723 // range [0, OpSize), Neg == (Pos == 0 ? 0 : OpSize - Pos). This means that
3724 // for two opposing shifts shift1 and shift2 and a value X with OpBits bits:
3726 // (or (shift1 X, Neg), (shift2 X, Pos))
3728 // reduces to a rotate in direction shift2 by Pos or (equivalently) a rotate
3729 // in direction shift1 by Neg. The range [0, OpSize) means that we only need
3730 // to consider shift amounts with defined behavior.
3731 static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned OpSize) {
3732 // If OpSize is a power of 2 then:
3734 // (a) (Pos == 0 ? 0 : OpSize - Pos) == (OpSize - Pos) & (OpSize - 1)
3735 // (b) Neg == Neg & (OpSize - 1) whenever Neg is in [0, OpSize).
3737 // So if OpSize is a power of 2 and Neg is (and Neg', OpSize-1), we check
3738 // for the stronger condition:
3740 // Neg & (OpSize - 1) == (OpSize - Pos) & (OpSize - 1) [A]
3742 // for all Neg and Pos. Since Neg & (OpSize - 1) == Neg' & (OpSize - 1)
3743 // we can just replace Neg with Neg' for the rest of the function.
3745 // In other cases we check for the even stronger condition:
3747 // Neg == OpSize - Pos [B]
3749 // for all Neg and Pos. Note that the (or ...) then invokes undefined
3750 // behavior if Pos == 0 (and consequently Neg == OpSize).
3752 // We could actually use [A] whenever OpSize is a power of 2, but the
3753 // only extra cases that it would match are those uninteresting ones
3754 // where Neg and Pos are never in range at the same time. E.g. for
3755 // OpSize == 32, using [A] would allow a Neg of the form (sub 64, Pos)
3756 // as well as (sub 32, Pos), but:
3758 // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos))
3760 // always invokes undefined behavior for 32-bit X.
3762 // Below, Mask == OpSize - 1 when using [A] and is all-ones otherwise.
3763 unsigned MaskLoBits = 0;
3764 if (Neg.getOpcode() == ISD::AND &&
3765 isPowerOf2_64(OpSize) &&
3766 Neg.getOperand(1).getOpcode() == ISD::Constant &&
3767 cast<ConstantSDNode>(Neg.getOperand(1))->getAPIntValue() == OpSize - 1) {
3768 Neg = Neg.getOperand(0);
3769 MaskLoBits = Log2_64(OpSize);
3772 // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1.
3773 if (Neg.getOpcode() != ISD::SUB)
3775 ConstantSDNode *NegC = dyn_cast<ConstantSDNode>(Neg.getOperand(0));
3778 SDValue NegOp1 = Neg.getOperand(1);
3780 // On the RHS of [A], if Pos is Pos' & (OpSize - 1), just replace Pos with
3781 // Pos'. The truncation is redundant for the purpose of the equality.
3783 Pos.getOpcode() == ISD::AND &&
3784 Pos.getOperand(1).getOpcode() == ISD::Constant &&
3785 cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() == OpSize - 1)
3786 Pos = Pos.getOperand(0);
3788 // The condition we need is now:
3790 // (NegC - NegOp1) & Mask == (OpSize - Pos) & Mask
3792 // If NegOp1 == Pos then we need:
3794 // OpSize & Mask == NegC & Mask
3796 // (because "x & Mask" is a truncation and distributes through subtraction).
3799 Width = NegC->getAPIntValue();
3800 // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC.
3801 // Then the condition we want to prove becomes:
3803 // (NegC - NegOp1) & Mask == (OpSize - (NegOp1 + PosC)) & Mask
3805 // which, again because "x & Mask" is a truncation, becomes:
3807 // NegC & Mask == (OpSize - PosC) & Mask
3808 // OpSize & Mask == (NegC + PosC) & Mask
3809 else if (Pos.getOpcode() == ISD::ADD &&
3810 Pos.getOperand(0) == NegOp1 &&
3811 Pos.getOperand(1).getOpcode() == ISD::Constant)
3812 Width = (cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() +
3813 NegC->getAPIntValue());
3817 // Now we just need to check that OpSize & Mask == Width & Mask.
3819 // Opsize & Mask is 0 since Mask is Opsize - 1.
3820 return Width.getLoBits(MaskLoBits) == 0;
3821 return Width == OpSize;
3824 // A subroutine of MatchRotate used once we have found an OR of two opposite
3825 // shifts of Shifted. If Neg == <operand size> - Pos then the OR reduces
3826 // to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the
3827 // former being preferred if supported. InnerPos and InnerNeg are Pos and
3828 // Neg with outer conversions stripped away.
3829 SDNode *DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos,
3830 SDValue Neg, SDValue InnerPos,
3831 SDValue InnerNeg, unsigned PosOpcode,
3832 unsigned NegOpcode, SDLoc DL) {
3833 // fold (or (shl x, (*ext y)),
3834 // (srl x, (*ext (sub 32, y)))) ->
3835 // (rotl x, y) or (rotr x, (sub 32, y))
3837 // fold (or (shl x, (*ext (sub 32, y))),
3838 // (srl x, (*ext y))) ->
3839 // (rotr x, y) or (rotl x, (sub 32, y))
3840 EVT VT = Shifted.getValueType();
3841 if (matchRotateSub(InnerPos, InnerNeg, VT.getSizeInBits())) {
3842 bool HasPos = TLI.isOperationLegalOrCustom(PosOpcode, VT);
3843 return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted,
3844 HasPos ? Pos : Neg).getNode();
3850 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3851 // idioms for rotate, and if the target supports rotation instructions, generate
3853 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL) {
3854 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3855 EVT VT = LHS.getValueType();
3856 if (!TLI.isTypeLegal(VT)) return nullptr;
3858 // The target must have at least one rotate flavor.
3859 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3860 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3861 if (!HasROTL && !HasROTR) return nullptr;
3863 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3864 SDValue LHSShift; // The shift.
3865 SDValue LHSMask; // AND value if any.
3866 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3867 return nullptr; // Not part of a rotate.
3869 SDValue RHSShift; // The shift.
3870 SDValue RHSMask; // AND value if any.
3871 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3872 return nullptr; // Not part of a rotate.
3874 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3875 return nullptr; // Not shifting the same value.
3877 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3878 return nullptr; // Shifts must disagree.
3880 // Canonicalize shl to left side in a shl/srl pair.
3881 if (RHSShift.getOpcode() == ISD::SHL) {
3882 std::swap(LHS, RHS);
3883 std::swap(LHSShift, RHSShift);
3884 std::swap(LHSMask , RHSMask );
3887 unsigned OpSizeInBits = VT.getSizeInBits();
3888 SDValue LHSShiftArg = LHSShift.getOperand(0);
3889 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3890 SDValue RHSShiftArg = RHSShift.getOperand(0);
3891 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3893 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3894 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3895 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3896 RHSShiftAmt.getOpcode() == ISD::Constant) {
3897 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3898 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3899 if ((LShVal + RShVal) != OpSizeInBits)
3902 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3903 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
3905 // If there is an AND of either shifted operand, apply it to the result.
3906 if (LHSMask.getNode() || RHSMask.getNode()) {
3907 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3909 if (LHSMask.getNode()) {
3910 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3911 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3913 if (RHSMask.getNode()) {
3914 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3915 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3918 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, DL, VT));
3921 return Rot.getNode();
3924 // If there is a mask here, and we have a variable shift, we can't be sure
3925 // that we're masking out the right stuff.
3926 if (LHSMask.getNode() || RHSMask.getNode())
3929 // If the shift amount is sign/zext/any-extended just peel it off.
3930 SDValue LExtOp0 = LHSShiftAmt;
3931 SDValue RExtOp0 = RHSShiftAmt;
3932 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3933 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3934 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3935 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3936 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3937 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3938 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3939 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3940 LExtOp0 = LHSShiftAmt.getOperand(0);
3941 RExtOp0 = RHSShiftAmt.getOperand(0);
3944 SDNode *TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt,
3945 LExtOp0, RExtOp0, ISD::ROTL, ISD::ROTR, DL);
3949 SDNode *TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt,
3950 RExtOp0, LExtOp0, ISD::ROTR, ISD::ROTL, DL);
3957 SDValue DAGCombiner::visitXOR(SDNode *N) {
3958 SDValue N0 = N->getOperand(0);
3959 SDValue N1 = N->getOperand(1);
3960 EVT VT = N0.getValueType();
3963 if (VT.isVector()) {
3964 if (SDValue FoldedVOp = SimplifyVBinOp(N))
3967 // fold (xor x, 0) -> x, vector edition
3968 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3970 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3974 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3975 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3976 return DAG.getConstant(0, SDLoc(N), VT);
3977 // fold (xor x, undef) -> undef
3978 if (N0.getOpcode() == ISD::UNDEF)
3980 if (N1.getOpcode() == ISD::UNDEF)
3982 // fold (xor c1, c2) -> c1^c2
3983 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
3984 ConstantSDNode *N1C = getAsNonOpaqueConstant(N1);
3986 return DAG.FoldConstantArithmetic(ISD::XOR, SDLoc(N), VT, N0C, N1C);
3987 // canonicalize constant to RHS
3988 if (isConstantIntBuildVectorOrConstantInt(N0) &&
3989 !isConstantIntBuildVectorOrConstantInt(N1))
3990 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
3991 // fold (xor x, 0) -> x
3992 if (isNullConstant(N1))
3995 if (SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1))
3998 // fold !(x cc y) -> (x !cc y)
3999 SDValue LHS, RHS, CC;
4000 if (TLI.isConstTrueVal(N1.getNode()) && isSetCCEquivalent(N0, LHS, RHS, CC)) {
4001 bool isInt = LHS.getValueType().isInteger();
4002 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
4005 if (!LegalOperations ||
4006 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) {
4007 switch (N0.getOpcode()) {
4009 llvm_unreachable("Unhandled SetCC Equivalent!");
4011 return DAG.getSetCC(SDLoc(N), VT, LHS, RHS, NotCC);
4012 case ISD::SELECT_CC:
4013 return DAG.getSelectCC(SDLoc(N), LHS, RHS, N0.getOperand(2),
4014 N0.getOperand(3), NotCC);
4019 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
4020 if (isOneConstant(N1) && N0.getOpcode() == ISD::ZERO_EXTEND &&
4021 N0.getNode()->hasOneUse() &&
4022 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
4023 SDValue V = N0.getOperand(0);
4025 V = DAG.getNode(ISD::XOR, DL, V.getValueType(), V,
4026 DAG.getConstant(1, DL, V.getValueType()));
4027 AddToWorklist(V.getNode());
4028 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V);
4031 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
4032 if (isOneConstant(N1) && VT == MVT::i1 &&
4033 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
4034 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
4035 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
4036 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
4037 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
4038 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
4039 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
4040 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
4043 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
4044 if (isAllOnesConstant(N1) &&
4045 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
4046 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
4047 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
4048 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
4049 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
4050 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
4051 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
4052 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
4055 // fold (xor (and x, y), y) -> (and (not x), y)
4056 if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
4057 N0->getOperand(1) == N1) {
4058 SDValue X = N0->getOperand(0);
4059 SDValue NotX = DAG.getNOT(SDLoc(X), X, VT);
4060 AddToWorklist(NotX.getNode());
4061 return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1);
4063 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
4064 if (N1C && N0.getOpcode() == ISD::XOR) {
4065 if (const ConstantSDNode *N00C = getAsNonOpaqueConstant(N0.getOperand(0))) {
4067 return DAG.getNode(ISD::XOR, DL, VT, N0.getOperand(1),
4068 DAG.getConstant(N1C->getAPIntValue() ^
4069 N00C->getAPIntValue(), DL, VT));
4071 if (const ConstantSDNode *N01C = getAsNonOpaqueConstant(N0.getOperand(1))) {
4073 return DAG.getNode(ISD::XOR, DL, VT, N0.getOperand(0),
4074 DAG.getConstant(N1C->getAPIntValue() ^
4075 N01C->getAPIntValue(), DL, VT));
4078 // fold (xor x, x) -> 0
4080 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
4082 // fold (xor (shl 1, x), -1) -> (rotl ~1, x)
4083 // Here is a concrete example of this equivalence:
4085 // i16 shl == 1 << 14 == 16384 == 0b0100000000000000
4086 // i16 xor == ~(1 << 14) == 49151 == 0b1011111111111111
4090 // i16 ~1 == 0b1111111111111110
4091 // i16 rol(~1, 14) == 0b1011111111111111
4093 // Some additional tips to help conceptualize this transform:
4094 // - Try to see the operation as placing a single zero in a value of all ones.
4095 // - There exists no value for x which would allow the result to contain zero.
4096 // - Values of x larger than the bitwidth are undefined and do not require a
4097 // consistent result.
4098 // - Pushing the zero left requires shifting one bits in from the right.
4099 // A rotate left of ~1 is a nice way of achieving the desired result.
4100 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT) && N0.getOpcode() == ISD::SHL
4101 && isAllOnesConstant(N1) && isOneConstant(N0.getOperand(0))) {
4103 return DAG.getNode(ISD::ROTL, DL, VT, DAG.getConstant(~1, DL, VT),
4107 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
4108 if (N0.getOpcode() == N1.getOpcode())
4109 if (SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N))
4112 // Simplify the expression using non-local knowledge.
4113 if (!VT.isVector() &&
4114 SimplifyDemandedBits(SDValue(N, 0)))
4115 return SDValue(N, 0);
4120 /// Handle transforms common to the three shifts, when the shift amount is a
4122 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, ConstantSDNode *Amt) {
4123 SDNode *LHS = N->getOperand(0).getNode();
4124 if (!LHS->hasOneUse()) return SDValue();
4126 // We want to pull some binops through shifts, so that we have (and (shift))
4127 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
4128 // thing happens with address calculations, so it's important to canonicalize
4130 bool HighBitSet = false; // Can we transform this if the high bit is set?
4132 switch (LHS->getOpcode()) {
4133 default: return SDValue();
4136 HighBitSet = false; // We can only transform sra if the high bit is clear.
4139 HighBitSet = true; // We can only transform sra if the high bit is set.
4142 if (N->getOpcode() != ISD::SHL)
4143 return SDValue(); // only shl(add) not sr[al](add).
4144 HighBitSet = false; // We can only transform sra if the high bit is clear.
4148 // We require the RHS of the binop to be a constant and not opaque as well.
4149 ConstantSDNode *BinOpCst = getAsNonOpaqueConstant(LHS->getOperand(1));
4150 if (!BinOpCst) return SDValue();
4152 // FIXME: disable this unless the input to the binop is a shift by a constant.
4153 // If it is not a shift, it pessimizes some common cases like:
4155 // void foo(int *X, int i) { X[i & 1235] = 1; }
4156 // int bar(int *X, int i) { return X[i & 255]; }
4157 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
4158 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
4159 BinOpLHSVal->getOpcode() != ISD::SRA &&
4160 BinOpLHSVal->getOpcode() != ISD::SRL) ||
4161 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
4164 EVT VT = N->getValueType(0);
4166 // If this is a signed shift right, and the high bit is modified by the
4167 // logical operation, do not perform the transformation. The highBitSet
4168 // boolean indicates the value of the high bit of the constant which would
4169 // cause it to be modified for this operation.
4170 if (N->getOpcode() == ISD::SRA) {
4171 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
4172 if (BinOpRHSSignSet != HighBitSet)
4176 if (!TLI.isDesirableToCommuteWithShift(LHS))
4179 // Fold the constants, shifting the binop RHS by the shift amount.
4180 SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)),
4182 LHS->getOperand(1), N->getOperand(1));
4183 assert(isa<ConstantSDNode>(NewRHS) && "Folding was not successful!");
4185 // Create the new shift.
4186 SDValue NewShift = DAG.getNode(N->getOpcode(),
4187 SDLoc(LHS->getOperand(0)),
4188 VT, LHS->getOperand(0), N->getOperand(1));
4190 // Create the new binop.
4191 return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS);
4194 SDValue DAGCombiner::distributeTruncateThroughAnd(SDNode *N) {
4195 assert(N->getOpcode() == ISD::TRUNCATE);
4196 assert(N->getOperand(0).getOpcode() == ISD::AND);
4198 // (truncate:TruncVT (and N00, N01C)) -> (and (truncate:TruncVT N00), TruncC)
4199 if (N->hasOneUse() && N->getOperand(0).hasOneUse()) {
4200 SDValue N01 = N->getOperand(0).getOperand(1);
4202 if (ConstantSDNode *N01C = isConstOrConstSplat(N01)) {
4203 if (!N01C->isOpaque()) {
4204 EVT TruncVT = N->getValueType(0);
4205 SDValue N00 = N->getOperand(0).getOperand(0);
4206 APInt TruncC = N01C->getAPIntValue();
4207 TruncC = TruncC.trunc(TruncVT.getScalarSizeInBits());
4210 return DAG.getNode(ISD::AND, DL, TruncVT,
4211 DAG.getNode(ISD::TRUNCATE, DL, TruncVT, N00),
4212 DAG.getConstant(TruncC, DL, TruncVT));
4220 SDValue DAGCombiner::visitRotate(SDNode *N) {
4221 // fold (rot* x, (trunc (and y, c))) -> (rot* x, (and (trunc y), (trunc c))).
4222 if (N->getOperand(1).getOpcode() == ISD::TRUNCATE &&
4223 N->getOperand(1).getOperand(0).getOpcode() == ISD::AND) {
4224 SDValue NewOp1 = distributeTruncateThroughAnd(N->getOperand(1).getNode());
4225 if (NewOp1.getNode())
4226 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
4227 N->getOperand(0), NewOp1);
4232 SDValue DAGCombiner::visitSHL(SDNode *N) {
4233 SDValue N0 = N->getOperand(0);
4234 SDValue N1 = N->getOperand(1);
4235 EVT VT = N0.getValueType();
4236 unsigned OpSizeInBits = VT.getScalarSizeInBits();
4239 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4240 if (VT.isVector()) {
4241 if (SDValue FoldedVOp = SimplifyVBinOp(N))
4244 BuildVectorSDNode *N1CV = dyn_cast<BuildVectorSDNode>(N1);
4245 // If setcc produces all-one true value then:
4246 // (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<<N1CV)
4247 if (N1CV && N1CV->isConstant()) {
4248 if (N0.getOpcode() == ISD::AND) {
4249 SDValue N00 = N0->getOperand(0);
4250 SDValue N01 = N0->getOperand(1);
4251 BuildVectorSDNode *N01CV = dyn_cast<BuildVectorSDNode>(N01);
4253 if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC &&
4254 TLI.getBooleanContents(N00.getOperand(0).getValueType()) ==
4255 TargetLowering::ZeroOrNegativeOneBooleanContent) {
4256 if (SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, SDLoc(N), VT,
4258 return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C);
4261 N1C = isConstOrConstSplat(N1);
4266 // fold (shl c1, c2) -> c1<<c2
4267 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
4268 if (N0C && N1C && !N1C->isOpaque())
4269 return DAG.FoldConstantArithmetic(ISD::SHL, SDLoc(N), VT, N0C, N1C);
4270 // fold (shl 0, x) -> 0
4271 if (isNullConstant(N0))
4273 // fold (shl x, c >= size(x)) -> undef
4274 if (N1C && N1C->getAPIntValue().uge(OpSizeInBits))
4275 return DAG.getUNDEF(VT);
4276 // fold (shl x, 0) -> x
4277 if (N1C && N1C->isNullValue())
4279 // fold (shl undef, x) -> 0
4280 if (N0.getOpcode() == ISD::UNDEF)
4281 return DAG.getConstant(0, SDLoc(N), VT);
4282 // if (shl x, c) is known to be zero, return 0
4283 if (DAG.MaskedValueIsZero(SDValue(N, 0),
4284 APInt::getAllOnesValue(OpSizeInBits)))
4285 return DAG.getConstant(0, SDLoc(N), VT);
4286 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
4287 if (N1.getOpcode() == ISD::TRUNCATE &&
4288 N1.getOperand(0).getOpcode() == ISD::AND) {
4289 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4290 if (NewOp1.getNode())
4291 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, NewOp1);
4294 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4295 return SDValue(N, 0);
4297 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
4298 if (N1C && N0.getOpcode() == ISD::SHL) {
4299 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4300 uint64_t c1 = N0C1->getZExtValue();
4301 uint64_t c2 = N1C->getZExtValue();
4303 if (c1 + c2 >= OpSizeInBits)
4304 return DAG.getConstant(0, DL, VT);
4305 return DAG.getNode(ISD::SHL, DL, VT, N0.getOperand(0),
4306 DAG.getConstant(c1 + c2, DL, N1.getValueType()));
4310 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
4311 // For this to be valid, the second form must not preserve any of the bits
4312 // that are shifted out by the inner shift in the first form. This means
4313 // the outer shift size must be >= the number of bits added by the ext.
4314 // As a corollary, we don't care what kind of ext it is.
4315 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
4316 N0.getOpcode() == ISD::ANY_EXTEND ||
4317 N0.getOpcode() == ISD::SIGN_EXTEND) &&
4318 N0.getOperand(0).getOpcode() == ISD::SHL) {
4319 SDValue N0Op0 = N0.getOperand(0);
4320 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4321 uint64_t c1 = N0Op0C1->getZExtValue();
4322 uint64_t c2 = N1C->getZExtValue();
4323 EVT InnerShiftVT = N0Op0.getValueType();
4324 uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits();
4325 if (c2 >= OpSizeInBits - InnerShiftSize) {
4327 if (c1 + c2 >= OpSizeInBits)
4328 return DAG.getConstant(0, DL, VT);
4329 return DAG.getNode(ISD::SHL, DL, VT,
4330 DAG.getNode(N0.getOpcode(), DL, VT,
4331 N0Op0->getOperand(0)),
4332 DAG.getConstant(c1 + c2, DL, N1.getValueType()));
4337 // fold (shl (zext (srl x, C)), C) -> (zext (shl (srl x, C), C))
4338 // Only fold this if the inner zext has no other uses to avoid increasing
4339 // the total number of instructions.
4340 if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() &&
4341 N0.getOperand(0).getOpcode() == ISD::SRL) {
4342 SDValue N0Op0 = N0.getOperand(0);
4343 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4344 uint64_t c1 = N0Op0C1->getZExtValue();
4345 if (c1 < VT.getScalarSizeInBits()) {
4346 uint64_t c2 = N1C->getZExtValue();
4348 SDValue NewOp0 = N0.getOperand(0);
4349 EVT CountVT = NewOp0.getOperand(1).getValueType();
4351 SDValue NewSHL = DAG.getNode(ISD::SHL, DL, NewOp0.getValueType(),
4353 DAG.getConstant(c2, DL, CountVT));
4354 AddToWorklist(NewSHL.getNode());
4355 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
4361 // fold (shl (sr[la] exact X, C1), C2) -> (shl X, (C2-C1)) if C1 <= C2
4362 // fold (shl (sr[la] exact X, C1), C2) -> (sr[la] X, (C2-C1)) if C1 > C2
4363 if (N1C && (N0.getOpcode() == ISD::SRL || N0.getOpcode() == ISD::SRA) &&
4364 cast<BinaryWithFlagsSDNode>(N0)->Flags.hasExact()) {
4365 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4366 uint64_t C1 = N0C1->getZExtValue();
4367 uint64_t C2 = N1C->getZExtValue();
4370 return DAG.getNode(ISD::SHL, DL, VT, N0.getOperand(0),
4371 DAG.getConstant(C2 - C1, DL, N1.getValueType()));
4372 return DAG.getNode(N0.getOpcode(), DL, VT, N0.getOperand(0),
4373 DAG.getConstant(C1 - C2, DL, N1.getValueType()));
4377 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
4378 // (and (srl x, (sub c1, c2), MASK)
4379 // Only fold this if the inner shift has no other uses -- if it does, folding
4380 // this will increase the total number of instructions.
4381 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
4382 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4383 uint64_t c1 = N0C1->getZExtValue();
4384 if (c1 < OpSizeInBits) {
4385 uint64_t c2 = N1C->getZExtValue();
4386 APInt Mask = APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - c1);
4389 Mask = Mask.shl(c2 - c1);
4391 Shift = DAG.getNode(ISD::SHL, DL, VT, N0.getOperand(0),
4392 DAG.getConstant(c2 - c1, DL, N1.getValueType()));
4394 Mask = Mask.lshr(c1 - c2);
4396 Shift = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4397 DAG.getConstant(c1 - c2, DL, N1.getValueType()));
4400 return DAG.getNode(ISD::AND, DL, VT, Shift,
4401 DAG.getConstant(Mask, DL, VT));
4405 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
4406 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
4407 unsigned BitSize = VT.getScalarSizeInBits();
4409 SDValue HiBitsMask =
4410 DAG.getConstant(APInt::getHighBitsSet(BitSize,
4411 BitSize - N1C->getZExtValue()),
4413 return DAG.getNode(ISD::AND, DL, VT, N0.getOperand(0),
4417 // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
4418 // Variant of version done on multiply, except mul by a power of 2 is turned
4421 if (N1C && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
4422 (isa<ConstantSDNode>(N0.getOperand(1)) ||
4423 isConstantSplatVector(N0.getOperand(1).getNode(), Val))) {
4424 SDValue Shl0 = DAG.getNode(ISD::SHL, SDLoc(N0), VT, N0.getOperand(0), N1);
4425 SDValue Shl1 = DAG.getNode(ISD::SHL, SDLoc(N1), VT, N0.getOperand(1), N1);
4426 return DAG.getNode(ISD::ADD, SDLoc(N), VT, Shl0, Shl1);
4429 if (N1C && !N1C->isOpaque())
4430 if (SDValue NewSHL = visitShiftByConstant(N, N1C))
4436 SDValue DAGCombiner::visitSRA(SDNode *N) {
4437 SDValue N0 = N->getOperand(0);
4438 SDValue N1 = N->getOperand(1);
4439 EVT VT = N0.getValueType();
4440 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4443 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4444 if (VT.isVector()) {
4445 if (SDValue FoldedVOp = SimplifyVBinOp(N))
4448 N1C = isConstOrConstSplat(N1);
4451 // fold (sra c1, c2) -> (sra c1, c2)
4452 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
4453 if (N0C && N1C && !N1C->isOpaque())
4454 return DAG.FoldConstantArithmetic(ISD::SRA, SDLoc(N), VT, N0C, N1C);
4455 // fold (sra 0, x) -> 0
4456 if (isNullConstant(N0))
4458 // fold (sra -1, x) -> -1
4459 if (isAllOnesConstant(N0))
4461 // fold (sra x, (setge c, size(x))) -> undef
4462 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4463 return DAG.getUNDEF(VT);
4464 // fold (sra x, 0) -> x
4465 if (N1C && N1C->isNullValue())
4467 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
4469 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
4470 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
4471 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
4473 ExtVT = EVT::getVectorVT(*DAG.getContext(),
4474 ExtVT, VT.getVectorNumElements());
4475 if ((!LegalOperations ||
4476 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
4477 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
4478 N0.getOperand(0), DAG.getValueType(ExtVT));
4481 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
4482 if (N1C && N0.getOpcode() == ISD::SRA) {
4483 if (ConstantSDNode *C1 = isConstOrConstSplat(N0.getOperand(1))) {
4484 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
4485 if (Sum >= OpSizeInBits)
4486 Sum = OpSizeInBits - 1;
4488 return DAG.getNode(ISD::SRA, DL, VT, N0.getOperand(0),
4489 DAG.getConstant(Sum, DL, N1.getValueType()));
4493 // fold (sra (shl X, m), (sub result_size, n))
4494 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
4495 // result_size - n != m.
4496 // If truncate is free for the target sext(shl) is likely to result in better
4498 if (N0.getOpcode() == ISD::SHL && N1C) {
4499 // Get the two constanst of the shifts, CN0 = m, CN = n.
4500 const ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1));
4502 LLVMContext &Ctx = *DAG.getContext();
4503 // Determine what the truncate's result bitsize and type would be.
4504 EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - N1C->getZExtValue());
4507 TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorNumElements());
4509 // Determine the residual right-shift amount.
4510 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
4512 // If the shift is not a no-op (in which case this should be just a sign
4513 // extend already), the truncated to type is legal, sign_extend is legal
4514 // on that type, and the truncate to that type is both legal and free,
4515 // perform the transform.
4516 if ((ShiftAmt > 0) &&
4517 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
4518 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
4519 TLI.isTruncateFree(VT, TruncVT)) {
4522 SDValue Amt = DAG.getConstant(ShiftAmt, DL,
4523 getShiftAmountTy(N0.getOperand(0).getValueType()));
4524 SDValue Shift = DAG.getNode(ISD::SRL, DL, VT,
4525 N0.getOperand(0), Amt);
4526 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, TruncVT,
4528 return DAG.getNode(ISD::SIGN_EXTEND, DL,
4529 N->getValueType(0), Trunc);
4534 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
4535 if (N1.getOpcode() == ISD::TRUNCATE &&
4536 N1.getOperand(0).getOpcode() == ISD::AND) {
4537 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4538 if (NewOp1.getNode())
4539 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, NewOp1);
4542 // fold (sra (trunc (srl x, c1)), c2) -> (trunc (sra x, c1 + c2))
4543 // if c1 is equal to the number of bits the trunc removes
4544 if (N0.getOpcode() == ISD::TRUNCATE &&
4545 (N0.getOperand(0).getOpcode() == ISD::SRL ||
4546 N0.getOperand(0).getOpcode() == ISD::SRA) &&
4547 N0.getOperand(0).hasOneUse() &&
4548 N0.getOperand(0).getOperand(1).hasOneUse() &&
4550 SDValue N0Op0 = N0.getOperand(0);
4551 if (ConstantSDNode *LargeShift = isConstOrConstSplat(N0Op0.getOperand(1))) {
4552 unsigned LargeShiftVal = LargeShift->getZExtValue();
4553 EVT LargeVT = N0Op0.getValueType();
4555 if (LargeVT.getScalarSizeInBits() - OpSizeInBits == LargeShiftVal) {
4558 DAG.getConstant(LargeShiftVal + N1C->getZExtValue(), DL,
4559 getShiftAmountTy(N0Op0.getOperand(0).getValueType()));
4560 SDValue SRA = DAG.getNode(ISD::SRA, DL, LargeVT,
4561 N0Op0.getOperand(0), Amt);
4562 return DAG.getNode(ISD::TRUNCATE, DL, VT, SRA);
4567 // Simplify, based on bits shifted out of the LHS.
4568 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4569 return SDValue(N, 0);
4572 // If the sign bit is known to be zero, switch this to a SRL.
4573 if (DAG.SignBitIsZero(N0))
4574 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1);
4576 if (N1C && !N1C->isOpaque())
4577 if (SDValue NewSRA = visitShiftByConstant(N, N1C))
4583 SDValue DAGCombiner::visitSRL(SDNode *N) {
4584 SDValue N0 = N->getOperand(0);
4585 SDValue N1 = N->getOperand(1);
4586 EVT VT = N0.getValueType();
4587 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4590 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4591 if (VT.isVector()) {
4592 if (SDValue FoldedVOp = SimplifyVBinOp(N))
4595 N1C = isConstOrConstSplat(N1);
4598 // fold (srl c1, c2) -> c1 >>u c2
4599 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
4600 if (N0C && N1C && !N1C->isOpaque())
4601 return DAG.FoldConstantArithmetic(ISD::SRL, SDLoc(N), VT, N0C, N1C);
4602 // fold (srl 0, x) -> 0
4603 if (isNullConstant(N0))
4605 // fold (srl x, c >= size(x)) -> undef
4606 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4607 return DAG.getUNDEF(VT);
4608 // fold (srl x, 0) -> x
4609 if (N1C && N1C->isNullValue())
4611 // if (srl x, c) is known to be zero, return 0
4612 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
4613 APInt::getAllOnesValue(OpSizeInBits)))
4614 return DAG.getConstant(0, SDLoc(N), VT);
4616 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
4617 if (N1C && N0.getOpcode() == ISD::SRL) {
4618 if (ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1))) {
4619 uint64_t c1 = N01C->getZExtValue();
4620 uint64_t c2 = N1C->getZExtValue();
4622 if (c1 + c2 >= OpSizeInBits)
4623 return DAG.getConstant(0, DL, VT);
4624 return DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4625 DAG.getConstant(c1 + c2, DL, N1.getValueType()));
4629 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
4630 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
4631 N0.getOperand(0).getOpcode() == ISD::SRL &&
4632 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
4634 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
4635 uint64_t c2 = N1C->getZExtValue();
4636 EVT InnerShiftVT = N0.getOperand(0).getValueType();
4637 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
4638 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
4639 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
4640 if (c1 + OpSizeInBits == InnerShiftSize) {
4642 if (c1 + c2 >= InnerShiftSize)
4643 return DAG.getConstant(0, DL, VT);
4644 return DAG.getNode(ISD::TRUNCATE, DL, VT,
4645 DAG.getNode(ISD::SRL, DL, InnerShiftVT,
4646 N0.getOperand(0)->getOperand(0),
4647 DAG.getConstant(c1 + c2, DL,
4652 // fold (srl (shl x, c), c) -> (and x, cst2)
4653 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1) {
4654 unsigned BitSize = N0.getScalarValueSizeInBits();
4655 if (BitSize <= 64) {
4656 uint64_t ShAmt = N1C->getZExtValue() + 64 - BitSize;
4658 return DAG.getNode(ISD::AND, DL, VT, N0.getOperand(0),
4659 DAG.getConstant(~0ULL >> ShAmt, DL, VT));
4663 // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask)
4664 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
4665 // Shifting in all undef bits?
4666 EVT SmallVT = N0.getOperand(0).getValueType();
4667 unsigned BitSize = SmallVT.getScalarSizeInBits();
4668 if (N1C->getZExtValue() >= BitSize)
4669 return DAG.getUNDEF(VT);
4671 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
4672 uint64_t ShiftAmt = N1C->getZExtValue();
4674 SDValue SmallShift = DAG.getNode(ISD::SRL, DL0, SmallVT,
4676 DAG.getConstant(ShiftAmt, DL0,
4677 getShiftAmountTy(SmallVT)));
4678 AddToWorklist(SmallShift.getNode());
4679 APInt Mask = APInt::getAllOnesValue(OpSizeInBits).lshr(ShiftAmt);
4681 return DAG.getNode(ISD::AND, DL, VT,
4682 DAG.getNode(ISD::ANY_EXTEND, DL, VT, SmallShift),
4683 DAG.getConstant(Mask, DL, VT));
4687 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
4688 // bit, which is unmodified by sra.
4689 if (N1C && N1C->getZExtValue() + 1 == OpSizeInBits) {
4690 if (N0.getOpcode() == ISD::SRA)
4691 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1);
4694 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
4695 if (N1C && N0.getOpcode() == ISD::CTLZ &&
4696 N1C->getAPIntValue() == Log2_32(OpSizeInBits)) {
4697 APInt KnownZero, KnownOne;
4698 DAG.computeKnownBits(N0.getOperand(0), KnownZero, KnownOne);
4700 // If any of the input bits are KnownOne, then the input couldn't be all
4701 // zeros, thus the result of the srl will always be zero.
4702 if (KnownOne.getBoolValue()) return DAG.getConstant(0, SDLoc(N0), VT);
4704 // If all of the bits input the to ctlz node are known to be zero, then
4705 // the result of the ctlz is "32" and the result of the shift is one.
4706 APInt UnknownBits = ~KnownZero;
4707 if (UnknownBits == 0) return DAG.getConstant(1, SDLoc(N0), VT);
4709 // Otherwise, check to see if there is exactly one bit input to the ctlz.
4710 if ((UnknownBits & (UnknownBits - 1)) == 0) {
4711 // Okay, we know that only that the single bit specified by UnknownBits
4712 // could be set on input to the CTLZ node. If this bit is set, the SRL
4713 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
4714 // to an SRL/XOR pair, which is likely to simplify more.
4715 unsigned ShAmt = UnknownBits.countTrailingZeros();
4716 SDValue Op = N0.getOperand(0);
4720 Op = DAG.getNode(ISD::SRL, DL, VT, Op,
4721 DAG.getConstant(ShAmt, DL,
4722 getShiftAmountTy(Op.getValueType())));
4723 AddToWorklist(Op.getNode());
4727 return DAG.getNode(ISD::XOR, DL, VT,
4728 Op, DAG.getConstant(1, DL, VT));
4732 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
4733 if (N1.getOpcode() == ISD::TRUNCATE &&
4734 N1.getOperand(0).getOpcode() == ISD::AND) {
4735 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4736 if (NewOp1.getNode())
4737 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, NewOp1);
4740 // fold operands of srl based on knowledge that the low bits are not
4742 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4743 return SDValue(N, 0);
4745 if (N1C && !N1C->isOpaque()) {
4746 SDValue NewSRL = visitShiftByConstant(N, N1C);
4747 if (NewSRL.getNode())
4751 // Attempt to convert a srl of a load into a narrower zero-extending load.
4752 SDValue NarrowLoad = ReduceLoadWidth(N);
4753 if (NarrowLoad.getNode())
4756 // Here is a common situation. We want to optimize:
4759 // %b = and i32 %a, 2
4760 // %c = srl i32 %b, 1
4761 // brcond i32 %c ...
4767 // %c = setcc eq %b, 0
4770 // However when after the source operand of SRL is optimized into AND, the SRL
4771 // itself may not be optimized further. Look for it and add the BRCOND into
4773 if (N->hasOneUse()) {
4774 SDNode *Use = *N->use_begin();
4775 if (Use->getOpcode() == ISD::BRCOND)
4777 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
4778 // Also look pass the truncate.
4779 Use = *Use->use_begin();
4780 if (Use->getOpcode() == ISD::BRCOND)
4788 SDValue DAGCombiner::visitBSWAP(SDNode *N) {
4789 SDValue N0 = N->getOperand(0);
4790 EVT VT = N->getValueType(0);
4792 // fold (bswap c1) -> c2
4793 if (isConstantIntBuildVectorOrConstantInt(N0))
4794 return DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N0);
4795 // fold (bswap (bswap x)) -> x
4796 if (N0.getOpcode() == ISD::BSWAP)
4797 return N0->getOperand(0);
4801 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
4802 SDValue N0 = N->getOperand(0);
4803 EVT VT = N->getValueType(0);
4805 // fold (ctlz c1) -> c2
4806 if (isConstantIntBuildVectorOrConstantInt(N0))
4807 return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0);
4811 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4812 SDValue N0 = N->getOperand(0);
4813 EVT VT = N->getValueType(0);
4815 // fold (ctlz_zero_undef c1) -> c2
4816 if (isConstantIntBuildVectorOrConstantInt(N0))
4817 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4821 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4822 SDValue N0 = N->getOperand(0);
4823 EVT VT = N->getValueType(0);
4825 // fold (cttz c1) -> c2
4826 if (isConstantIntBuildVectorOrConstantInt(N0))
4827 return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0);
4831 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4832 SDValue N0 = N->getOperand(0);
4833 EVT VT = N->getValueType(0);
4835 // fold (cttz_zero_undef c1) -> c2
4836 if (isConstantIntBuildVectorOrConstantInt(N0))
4837 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4841 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4842 SDValue N0 = N->getOperand(0);
4843 EVT VT = N->getValueType(0);
4845 // fold (ctpop c1) -> c2
4846 if (isConstantIntBuildVectorOrConstantInt(N0))
4847 return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0);
4852 /// \brief Generate Min/Max node
4853 static SDValue combineMinNumMaxNum(SDLoc DL, EVT VT, SDValue LHS, SDValue RHS,
4854 SDValue True, SDValue False,
4855 ISD::CondCode CC, const TargetLowering &TLI,
4856 SelectionDAG &DAG) {
4857 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
4867 unsigned Opcode = (LHS == True) ? ISD::FMINNUM : ISD::FMAXNUM;
4868 if (TLI.isOperationLegal(Opcode, VT))
4869 return DAG.getNode(Opcode, DL, VT, LHS, RHS);
4878 unsigned Opcode = (LHS == True) ? ISD::FMAXNUM : ISD::FMINNUM;
4879 if (TLI.isOperationLegal(Opcode, VT))
4880 return DAG.getNode(Opcode, DL, VT, LHS, RHS);
4888 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4889 SDValue N0 = N->getOperand(0);
4890 SDValue N1 = N->getOperand(1);
4891 SDValue N2 = N->getOperand(2);
4892 EVT VT = N->getValueType(0);
4893 EVT VT0 = N0.getValueType();
4895 // fold (select C, X, X) -> X
4898 if (const ConstantSDNode *N0C = dyn_cast<const ConstantSDNode>(N0)) {
4899 // fold (select true, X, Y) -> X
4900 // fold (select false, X, Y) -> Y
4901 return !N0C->isNullValue() ? N1 : N2;
4903 // fold (select C, 1, X) -> (or C, X)
4904 if (VT == MVT::i1 && isOneConstant(N1))
4905 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4906 // fold (select C, 0, 1) -> (xor C, 1)
4907 // We can't do this reliably if integer based booleans have different contents
4908 // to floating point based booleans. This is because we can't tell whether we
4909 // have an integer-based boolean or a floating-point-based boolean unless we
4910 // can find the SETCC that produced it and inspect its operands. This is
4911 // fairly easy if C is the SETCC node, but it can potentially be
4912 // undiscoverable (or not reasonably discoverable). For example, it could be
4913 // in another basic block or it could require searching a complicated
4915 if (VT.isInteger() &&
4916 (VT0 == MVT::i1 || (VT0.isInteger() &&
4917 TLI.getBooleanContents(false, false) ==
4918 TLI.getBooleanContents(false, true) &&
4919 TLI.getBooleanContents(false, false) ==
4920 TargetLowering::ZeroOrOneBooleanContent)) &&
4921 isNullConstant(N1) && isOneConstant(N2)) {
4925 return DAG.getNode(ISD::XOR, DL, VT0,
4926 N0, DAG.getConstant(1, DL, VT0));
4929 XORNode = DAG.getNode(ISD::XOR, DL0, VT0,
4930 N0, DAG.getConstant(1, DL0, VT0));
4931 AddToWorklist(XORNode.getNode());
4933 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, XORNode);
4934 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, XORNode);
4936 // fold (select C, 0, X) -> (and (not C), X)
4937 if (VT == VT0 && VT == MVT::i1 && isNullConstant(N1)) {
4938 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4939 AddToWorklist(NOTNode.getNode());
4940 return DAG.getNode(ISD::AND, SDLoc(N), VT, NOTNode, N2);
4942 // fold (select C, X, 1) -> (or (not C), X)
4943 if (VT == VT0 && VT == MVT::i1 && isOneConstant(N2)) {
4944 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4945 AddToWorklist(NOTNode.getNode());
4946 return DAG.getNode(ISD::OR, SDLoc(N), VT, NOTNode, N1);
4948 // fold (select C, X, 0) -> (and C, X)
4949 if (VT == MVT::i1 && isNullConstant(N2))
4950 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4951 // fold (select X, X, Y) -> (or X, Y)
4952 // fold (select X, 1, Y) -> (or X, Y)
4953 if (VT == MVT::i1 && (N0 == N1 || isOneConstant(N1)))
4954 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4955 // fold (select X, Y, X) -> (and X, Y)
4956 // fold (select X, Y, 0) -> (and X, Y)
4957 if (VT == MVT::i1 && (N0 == N2 || isNullConstant(N2)))
4958 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4960 // If we can fold this based on the true/false value, do so.
4961 if (SimplifySelectOps(N, N1, N2))
4962 return SDValue(N, 0); // Don't revisit N.
4964 // fold selects based on a setcc into other things, such as min/max/abs
4965 if (N0.getOpcode() == ISD::SETCC) {
4966 // select x, y (fcmp lt x, y) -> fminnum x, y
4967 // select x, y (fcmp gt x, y) -> fmaxnum x, y
4969 // This is OK if we don't care about what happens if either operand is a
4973 // FIXME: Instead of testing for UnsafeFPMath, this should be checking for
4974 // no signed zeros as well as no nans.
4975 const TargetOptions &Options = DAG.getTarget().Options;
4976 if (Options.UnsafeFPMath &&
4977 VT.isFloatingPoint() && N0.hasOneUse() &&
4978 DAG.isKnownNeverNaN(N1) && DAG.isKnownNeverNaN(N2)) {
4979 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
4982 combineMinNumMaxNum(SDLoc(N), VT, N0.getOperand(0), N0.getOperand(1),
4983 N1, N2, CC, TLI, DAG);
4988 if ((!LegalOperations &&
4989 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT)) ||
4990 TLI.isOperationLegal(ISD::SELECT_CC, VT))
4991 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT,
4992 N0.getOperand(0), N0.getOperand(1),
4993 N1, N2, N0.getOperand(2));
4994 return SimplifySelect(SDLoc(N), N0, N1, N2);
4997 if (VT0 == MVT::i1) {
4998 if (TLI.shouldNormalizeToSelectSequence(*DAG.getContext(), VT)) {
4999 // select (and Cond0, Cond1), X, Y
5000 // -> select Cond0, (select Cond1, X, Y), Y
5001 if (N0->getOpcode() == ISD::AND && N0->hasOneUse()) {
5002 SDValue Cond0 = N0->getOperand(0);
5003 SDValue Cond1 = N0->getOperand(1);
5004 SDValue InnerSelect = DAG.getNode(ISD::SELECT, SDLoc(N),
5005 N1.getValueType(), Cond1, N1, N2);
5006 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Cond0,
5009 // select (or Cond0, Cond1), X, Y -> select Cond0, X, (select Cond1, X, Y)
5010 if (N0->getOpcode() == ISD::OR && 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 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Cond0, N1,
5020 // select Cond0, (select Cond1, X, Y), Y -> select (and Cond0, Cond1), X, Y
5021 if (N1->getOpcode() == ISD::SELECT) {
5022 SDValue N1_0 = N1->getOperand(0);
5023 SDValue N1_1 = N1->getOperand(1);
5024 SDValue N1_2 = N1->getOperand(2);
5025 if (N1_2 == N2 && N0.getValueType() == N1_0.getValueType()) {
5026 // Create the actual and node if we can generate good code for it.
5027 if (!TLI.shouldNormalizeToSelectSequence(*DAG.getContext(), VT)) {
5028 SDValue And = DAG.getNode(ISD::AND, SDLoc(N), N0.getValueType(),
5030 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), And,
5033 // Otherwise see if we can optimize the "and" to a better pattern.
5034 if (SDValue Combined = visitANDLike(N0, N1_0, N))
5035 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Combined,
5039 // select Cond0, X, (select Cond1, X, Y) -> select (or Cond0, Cond1), X, Y
5040 if (N2->getOpcode() == ISD::SELECT) {
5041 SDValue N2_0 = N2->getOperand(0);
5042 SDValue N2_1 = N2->getOperand(1);
5043 SDValue N2_2 = N2->getOperand(2);
5044 if (N2_1 == N1 && N0.getValueType() == N2_0.getValueType()) {
5045 // Create the actual or node if we can generate good code for it.
5046 if (!TLI.shouldNormalizeToSelectSequence(*DAG.getContext(), VT)) {
5047 SDValue Or = DAG.getNode(ISD::OR, SDLoc(N), N0.getValueType(),
5049 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Or,
5052 // Otherwise see if we can optimize to a better pattern.
5053 if (SDValue Combined = visitORLike(N0, N2_0, N))
5054 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Combined,
5064 std::pair<SDValue, SDValue> SplitVSETCC(const SDNode *N, SelectionDAG &DAG) {
5067 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
5069 // Split the inputs.
5070 SDValue Lo, Hi, LL, LH, RL, RH;
5071 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
5072 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
5074 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));
5075 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));
5077 return std::make_pair(Lo, Hi);
5080 // This function assumes all the vselect's arguments are CONCAT_VECTOR
5081 // nodes and that the condition is a BV of ConstantSDNodes (or undefs).
5082 static SDValue ConvertSelectToConcatVector(SDNode *N, SelectionDAG &DAG) {
5084 SDValue Cond = N->getOperand(0);
5085 SDValue LHS = N->getOperand(1);
5086 SDValue RHS = N->getOperand(2);
5087 EVT VT = N->getValueType(0);
5088 int NumElems = VT.getVectorNumElements();
5089 assert(LHS.getOpcode() == ISD::CONCAT_VECTORS &&
5090 RHS.getOpcode() == ISD::CONCAT_VECTORS &&
5091 Cond.getOpcode() == ISD::BUILD_VECTOR);
5093 // CONCAT_VECTOR can take an arbitrary number of arguments. We only care about
5094 // binary ones here.
5095 if (LHS->getNumOperands() != 2 || RHS->getNumOperands() != 2)
5098 // We're sure we have an even number of elements due to the
5099 // concat_vectors we have as arguments to vselect.
5100 // Skip BV elements until we find one that's not an UNDEF
5101 // After we find an UNDEF element, keep looping until we get to half the
5102 // length of the BV and see if all the non-undef nodes are the same.
5103 ConstantSDNode *BottomHalf = nullptr;
5104 for (int i = 0; i < NumElems / 2; ++i) {
5105 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
5108 if (BottomHalf == nullptr)
5109 BottomHalf = cast<ConstantSDNode>(Cond.getOperand(i));
5110 else if (Cond->getOperand(i).getNode() != BottomHalf)
5114 // Do the same for the second half of the BuildVector
5115 ConstantSDNode *TopHalf = nullptr;
5116 for (int i = NumElems / 2; i < NumElems; ++i) {
5117 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
5120 if (TopHalf == nullptr)
5121 TopHalf = cast<ConstantSDNode>(Cond.getOperand(i));
5122 else if (Cond->getOperand(i).getNode() != TopHalf)
5126 assert(TopHalf && BottomHalf &&
5127 "One half of the selector was all UNDEFs and the other was all the "
5128 "same value. This should have been addressed before this function.");
5130 ISD::CONCAT_VECTORS, dl, VT,
5131 BottomHalf->isNullValue() ? RHS->getOperand(0) : LHS->getOperand(0),
5132 TopHalf->isNullValue() ? RHS->getOperand(1) : LHS->getOperand(1));
5135 SDValue DAGCombiner::visitMSCATTER(SDNode *N) {
5137 if (Level >= AfterLegalizeTypes)
5140 MaskedScatterSDNode *MSC = cast<MaskedScatterSDNode>(N);
5141 SDValue Mask = MSC->getMask();
5142 SDValue Data = MSC->getValue();
5145 // If the MSCATTER data type requires splitting and the mask is provided by a
5146 // SETCC, then split both nodes and its operands before legalization. This
5147 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5148 // and enables future optimizations (e.g. min/max pattern matching on X86).
5149 if (Mask.getOpcode() != ISD::SETCC)
5152 // Check if any splitting is required.
5153 if (TLI.getTypeAction(*DAG.getContext(), Data.getValueType()) !=
5154 TargetLowering::TypeSplitVector)
5156 SDValue MaskLo, MaskHi, Lo, Hi;
5157 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
5160 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MSC->getValueType(0));
5162 SDValue Chain = MSC->getChain();
5164 EVT MemoryVT = MSC->getMemoryVT();
5165 unsigned Alignment = MSC->getOriginalAlignment();
5167 EVT LoMemVT, HiMemVT;
5168 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
5170 SDValue DataLo, DataHi;
5171 std::tie(DataLo, DataHi) = DAG.SplitVector(Data, DL);
5173 SDValue BasePtr = MSC->getBasePtr();
5174 SDValue IndexLo, IndexHi;
5175 std::tie(IndexLo, IndexHi) = DAG.SplitVector(MSC->getIndex(), DL);
5177 MachineMemOperand *MMO = DAG.getMachineFunction().
5178 getMachineMemOperand(MSC->getPointerInfo(),
5179 MachineMemOperand::MOStore, LoMemVT.getStoreSize(),
5180 Alignment, MSC->getAAInfo(), MSC->getRanges());
5182 SDValue OpsLo[] = { Chain, DataLo, MaskLo, BasePtr, IndexLo };
5183 Lo = DAG.getMaskedScatter(DAG.getVTList(MVT::Other), DataLo.getValueType(),
5186 SDValue OpsHi[] = {Chain, DataHi, MaskHi, BasePtr, IndexHi};
5187 Hi = DAG.getMaskedScatter(DAG.getVTList(MVT::Other), DataHi.getValueType(),
5190 AddToWorklist(Lo.getNode());
5191 AddToWorklist(Hi.getNode());
5193 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);
5196 SDValue DAGCombiner::visitMSTORE(SDNode *N) {
5198 if (Level >= AfterLegalizeTypes)
5201 MaskedStoreSDNode *MST = dyn_cast<MaskedStoreSDNode>(N);
5202 SDValue Mask = MST->getMask();
5203 SDValue Data = MST->getValue();
5206 // If the MSTORE data type requires splitting and the mask is provided by a
5207 // SETCC, then split both nodes and its operands before legalization. This
5208 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5209 // and enables future optimizations (e.g. min/max pattern matching on X86).
5210 if (Mask.getOpcode() == ISD::SETCC) {
5212 // Check if any splitting is required.
5213 if (TLI.getTypeAction(*DAG.getContext(), Data.getValueType()) !=
5214 TargetLowering::TypeSplitVector)
5217 SDValue MaskLo, MaskHi, Lo, Hi;
5218 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
5221 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MST->getValueType(0));
5223 SDValue Chain = MST->getChain();
5224 SDValue Ptr = MST->getBasePtr();
5226 EVT MemoryVT = MST->getMemoryVT();
5227 unsigned Alignment = MST->getOriginalAlignment();
5229 // if Alignment is equal to the vector size,
5230 // take the half of it for the second part
5231 unsigned SecondHalfAlignment =
5232 (Alignment == Data->getValueType(0).getSizeInBits()/8) ?
5233 Alignment/2 : Alignment;
5235 EVT LoMemVT, HiMemVT;
5236 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
5238 SDValue DataLo, DataHi;
5239 std::tie(DataLo, DataHi) = DAG.SplitVector(Data, DL);
5241 MachineMemOperand *MMO = DAG.getMachineFunction().
5242 getMachineMemOperand(MST->getPointerInfo(),
5243 MachineMemOperand::MOStore, LoMemVT.getStoreSize(),
5244 Alignment, MST->getAAInfo(), MST->getRanges());
5246 Lo = DAG.getMaskedStore(Chain, DL, DataLo, Ptr, MaskLo, LoMemVT, MMO,
5247 MST->isTruncatingStore());
5249 unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
5250 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5251 DAG.getConstant(IncrementSize, DL, Ptr.getValueType()));
5253 MMO = DAG.getMachineFunction().
5254 getMachineMemOperand(MST->getPointerInfo(),
5255 MachineMemOperand::MOStore, HiMemVT.getStoreSize(),
5256 SecondHalfAlignment, MST->getAAInfo(),
5259 Hi = DAG.getMaskedStore(Chain, DL, DataHi, Ptr, MaskHi, HiMemVT, MMO,
5260 MST->isTruncatingStore());
5262 AddToWorklist(Lo.getNode());
5263 AddToWorklist(Hi.getNode());
5265 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);
5270 SDValue DAGCombiner::visitMGATHER(SDNode *N) {
5272 if (Level >= AfterLegalizeTypes)
5275 MaskedGatherSDNode *MGT = dyn_cast<MaskedGatherSDNode>(N);
5276 SDValue Mask = MGT->getMask();
5279 // If the MGATHER result requires splitting and the mask is provided by a
5280 // SETCC, then split both nodes and its operands before legalization. This
5281 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5282 // and enables future optimizations (e.g. min/max pattern matching on X86).
5284 if (Mask.getOpcode() != ISD::SETCC)
5287 EVT VT = N->getValueType(0);
5289 // Check if any splitting is required.
5290 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
5291 TargetLowering::TypeSplitVector)
5294 SDValue MaskLo, MaskHi, Lo, Hi;
5295 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
5297 SDValue Src0 = MGT->getValue();
5298 SDValue Src0Lo, Src0Hi;
5299 std::tie(Src0Lo, Src0Hi) = DAG.SplitVector(Src0, DL);
5302 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
5304 SDValue Chain = MGT->getChain();
5305 EVT MemoryVT = MGT->getMemoryVT();
5306 unsigned Alignment = MGT->getOriginalAlignment();
5308 EVT LoMemVT, HiMemVT;
5309 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
5311 SDValue BasePtr = MGT->getBasePtr();
5312 SDValue Index = MGT->getIndex();
5313 SDValue IndexLo, IndexHi;
5314 std::tie(IndexLo, IndexHi) = DAG.SplitVector(Index, DL);
5316 MachineMemOperand *MMO = DAG.getMachineFunction().
5317 getMachineMemOperand(MGT->getPointerInfo(),
5318 MachineMemOperand::MOLoad, LoMemVT.getStoreSize(),
5319 Alignment, MGT->getAAInfo(), MGT->getRanges());
5321 SDValue OpsLo[] = { Chain, Src0Lo, MaskLo, BasePtr, IndexLo };
5322 Lo = DAG.getMaskedGather(DAG.getVTList(LoVT, MVT::Other), LoVT, DL, OpsLo,
5325 SDValue OpsHi[] = {Chain, Src0Hi, MaskHi, BasePtr, IndexHi};
5326 Hi = DAG.getMaskedGather(DAG.getVTList(HiVT, MVT::Other), HiVT, DL, OpsHi,
5329 AddToWorklist(Lo.getNode());
5330 AddToWorklist(Hi.getNode());
5332 // Build a factor node to remember that this load is independent of the
5334 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo.getValue(1),
5337 // Legalized the chain result - switch anything that used the old chain to
5339 DAG.ReplaceAllUsesOfValueWith(SDValue(MGT, 1), Chain);
5341 SDValue GatherRes = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
5343 SDValue RetOps[] = { GatherRes, Chain };
5344 return DAG.getMergeValues(RetOps, DL);
5347 SDValue DAGCombiner::visitMLOAD(SDNode *N) {
5349 if (Level >= AfterLegalizeTypes)
5352 MaskedLoadSDNode *MLD = dyn_cast<MaskedLoadSDNode>(N);
5353 SDValue Mask = MLD->getMask();
5356 // If the MLOAD result requires splitting and the mask is provided by a
5357 // SETCC, then split both nodes and its operands before legalization. This
5358 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5359 // and enables future optimizations (e.g. min/max pattern matching on X86).
5361 if (Mask.getOpcode() == ISD::SETCC) {
5362 EVT VT = N->getValueType(0);
5364 // Check if any splitting is required.
5365 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
5366 TargetLowering::TypeSplitVector)
5369 SDValue MaskLo, MaskHi, Lo, Hi;
5370 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
5372 SDValue Src0 = MLD->getSrc0();
5373 SDValue Src0Lo, Src0Hi;
5374 std::tie(Src0Lo, Src0Hi) = DAG.SplitVector(Src0, DL);
5377 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MLD->getValueType(0));
5379 SDValue Chain = MLD->getChain();
5380 SDValue Ptr = MLD->getBasePtr();
5381 EVT MemoryVT = MLD->getMemoryVT();
5382 unsigned Alignment = MLD->getOriginalAlignment();
5384 // if Alignment is equal to the vector size,
5385 // take the half of it for the second part
5386 unsigned SecondHalfAlignment =
5387 (Alignment == MLD->getValueType(0).getSizeInBits()/8) ?
5388 Alignment/2 : Alignment;
5390 EVT LoMemVT, HiMemVT;
5391 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
5393 MachineMemOperand *MMO = DAG.getMachineFunction().
5394 getMachineMemOperand(MLD->getPointerInfo(),
5395 MachineMemOperand::MOLoad, LoMemVT.getStoreSize(),
5396 Alignment, MLD->getAAInfo(), MLD->getRanges());
5398 Lo = DAG.getMaskedLoad(LoVT, DL, Chain, Ptr, MaskLo, Src0Lo, LoMemVT, MMO,
5401 unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
5402 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5403 DAG.getConstant(IncrementSize, DL, Ptr.getValueType()));
5405 MMO = DAG.getMachineFunction().
5406 getMachineMemOperand(MLD->getPointerInfo(),
5407 MachineMemOperand::MOLoad, HiMemVT.getStoreSize(),
5408 SecondHalfAlignment, MLD->getAAInfo(), MLD->getRanges());
5410 Hi = DAG.getMaskedLoad(HiVT, DL, Chain, Ptr, MaskHi, Src0Hi, HiMemVT, MMO,
5413 AddToWorklist(Lo.getNode());
5414 AddToWorklist(Hi.getNode());
5416 // Build a factor node to remember that this load is independent of the
5418 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo.getValue(1),
5421 // Legalized the chain result - switch anything that used the old chain to
5423 DAG.ReplaceAllUsesOfValueWith(SDValue(MLD, 1), Chain);
5425 SDValue LoadRes = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
5427 SDValue RetOps[] = { LoadRes, Chain };
5428 return DAG.getMergeValues(RetOps, DL);
5433 SDValue DAGCombiner::visitVSELECT(SDNode *N) {
5434 SDValue N0 = N->getOperand(0);
5435 SDValue N1 = N->getOperand(1);
5436 SDValue N2 = N->getOperand(2);
5439 // Canonicalize integer abs.
5440 // vselect (setg[te] X, 0), X, -X ->
5441 // vselect (setgt X, -1), X, -X ->
5442 // vselect (setl[te] X, 0), -X, X ->
5443 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
5444 if (N0.getOpcode() == ISD::SETCC) {
5445 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
5446 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5448 bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
5450 if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
5451 (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) &&
5452 N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1))
5453 isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode());
5454 else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) &&
5455 N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1))
5456 isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode());
5459 EVT VT = LHS.getValueType();
5460 SDValue Shift = DAG.getNode(
5461 ISD::SRA, DL, VT, LHS,
5462 DAG.getConstant(VT.getScalarType().getSizeInBits() - 1, DL, VT));
5463 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift);
5464 AddToWorklist(Shift.getNode());
5465 AddToWorklist(Add.getNode());
5466 return DAG.getNode(ISD::XOR, DL, VT, Add, Shift);
5470 if (SimplifySelectOps(N, N1, N2))
5471 return SDValue(N, 0); // Don't revisit N.
5473 // If the VSELECT result requires splitting and the mask is provided by a
5474 // SETCC, then split both nodes and its operands before legalization. This
5475 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5476 // and enables future optimizations (e.g. min/max pattern matching on X86).
5477 if (N0.getOpcode() == ISD::SETCC) {
5478 EVT VT = N->getValueType(0);
5480 // Check if any splitting is required.
5481 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
5482 TargetLowering::TypeSplitVector)
5485 SDValue Lo, Hi, CCLo, CCHi, LL, LH, RL, RH;
5486 std::tie(CCLo, CCHi) = SplitVSETCC(N0.getNode(), DAG);
5487 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 1);
5488 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 2);
5490 Lo = DAG.getNode(N->getOpcode(), DL, LL.getValueType(), CCLo, LL, RL);
5491 Hi = DAG.getNode(N->getOpcode(), DL, LH.getValueType(), CCHi, LH, RH);
5493 // Add the new VSELECT nodes to the work list in case they need to be split
5495 AddToWorklist(Lo.getNode());
5496 AddToWorklist(Hi.getNode());
5498 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
5501 // Fold (vselect (build_vector all_ones), N1, N2) -> N1
5502 if (ISD::isBuildVectorAllOnes(N0.getNode()))
5504 // Fold (vselect (build_vector all_zeros), N1, N2) -> N2
5505 if (ISD::isBuildVectorAllZeros(N0.getNode()))
5508 // The ConvertSelectToConcatVector function is assuming both the above
5509 // checks for (vselect (build_vector all{ones,zeros) ...) have been made
5511 if (N1.getOpcode() == ISD::CONCAT_VECTORS &&
5512 N2.getOpcode() == ISD::CONCAT_VECTORS &&
5513 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
5514 SDValue CV = ConvertSelectToConcatVector(N, DAG);
5522 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
5523 SDValue N0 = N->getOperand(0);
5524 SDValue N1 = N->getOperand(1);
5525 SDValue N2 = N->getOperand(2);
5526 SDValue N3 = N->getOperand(3);
5527 SDValue N4 = N->getOperand(4);
5528 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
5530 // fold select_cc lhs, rhs, x, x, cc -> x
5534 // Determine if the condition we're dealing with is constant
5535 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
5536 N0, N1, CC, SDLoc(N), false);
5537 if (SCC.getNode()) {
5538 AddToWorklist(SCC.getNode());
5540 if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) {
5541 if (!SCCC->isNullValue())
5542 return N2; // cond always true -> true val
5544 return N3; // cond always false -> false val
5545 } else if (SCC->getOpcode() == ISD::UNDEF) {
5546 // When the condition is UNDEF, just return the first operand. This is
5547 // coherent the DAG creation, no setcc node is created in this case
5549 } else if (SCC.getOpcode() == ISD::SETCC) {
5550 // Fold to a simpler select_cc
5551 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(),
5552 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
5557 // If we can fold this based on the true/false value, do so.
5558 if (SimplifySelectOps(N, N2, N3))
5559 return SDValue(N, 0); // Don't revisit N.
5561 // fold select_cc into other things, such as min/max/abs
5562 return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC);
5565 SDValue DAGCombiner::visitSETCC(SDNode *N) {
5566 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
5567 cast<CondCodeSDNode>(N->getOperand(2))->get(),
5571 /// Try to fold a sext/zext/aext dag node into a ConstantSDNode or
5572 /// a build_vector of constants.
5573 /// This function is called by the DAGCombiner when visiting sext/zext/aext
5574 /// dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND).
5575 /// Vector extends are not folded if operations are legal; this is to
5576 /// avoid introducing illegal build_vector dag nodes.
5577 static SDNode *tryToFoldExtendOfConstant(SDNode *N, const TargetLowering &TLI,
5578 SelectionDAG &DAG, bool LegalTypes,
5579 bool LegalOperations) {
5580 unsigned Opcode = N->getOpcode();
5581 SDValue N0 = N->getOperand(0);
5582 EVT VT = N->getValueType(0);
5584 assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND ||
5585 Opcode == ISD::ANY_EXTEND || Opcode == ISD::SIGN_EXTEND_VECTOR_INREG)
5586 && "Expected EXTEND dag node in input!");
5588 // fold (sext c1) -> c1
5589 // fold (zext c1) -> c1
5590 // fold (aext c1) -> c1
5591 if (isa<ConstantSDNode>(N0))
5592 return DAG.getNode(Opcode, SDLoc(N), VT, N0).getNode();
5594 // fold (sext (build_vector AllConstants) -> (build_vector AllConstants)
5595 // fold (zext (build_vector AllConstants) -> (build_vector AllConstants)
5596 // fold (aext (build_vector AllConstants) -> (build_vector AllConstants)
5597 EVT SVT = VT.getScalarType();
5598 if (!(VT.isVector() &&
5599 (!LegalTypes || (!LegalOperations && TLI.isTypeLegal(SVT))) &&
5600 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())))
5603 // We can fold this node into a build_vector.
5604 unsigned VTBits = SVT.getSizeInBits();
5605 unsigned EVTBits = N0->getValueType(0).getScalarType().getSizeInBits();
5606 SmallVector<SDValue, 8> Elts;
5607 unsigned NumElts = VT.getVectorNumElements();
5610 for (unsigned i=0; i != NumElts; ++i) {
5611 SDValue Op = N0->getOperand(i);
5612 if (Op->getOpcode() == ISD::UNDEF) {
5613 Elts.push_back(DAG.getUNDEF(SVT));
5618 // Get the constant value and if needed trunc it to the size of the type.
5619 // Nodes like build_vector might have constants wider than the scalar type.
5620 APInt C = cast<ConstantSDNode>(Op)->getAPIntValue().zextOrTrunc(EVTBits);
5621 if (Opcode == ISD::SIGN_EXTEND || Opcode == ISD::SIGN_EXTEND_VECTOR_INREG)
5622 Elts.push_back(DAG.getConstant(C.sext(VTBits), DL, SVT));
5624 Elts.push_back(DAG.getConstant(C.zext(VTBits), DL, SVT));
5627 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Elts).getNode();
5630 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
5631 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
5632 // transformation. Returns true if extension are possible and the above
5633 // mentioned transformation is profitable.
5634 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
5636 SmallVectorImpl<SDNode *> &ExtendNodes,
5637 const TargetLowering &TLI) {
5638 bool HasCopyToRegUses = false;
5639 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
5640 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
5641 UE = N0.getNode()->use_end();
5646 if (UI.getUse().getResNo() != N0.getResNo())
5648 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
5649 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
5650 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
5651 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
5652 // Sign bits will be lost after a zext.
5655 for (unsigned i = 0; i != 2; ++i) {
5656 SDValue UseOp = User->getOperand(i);
5659 if (!isa<ConstantSDNode>(UseOp))
5664 ExtendNodes.push_back(User);
5667 // If truncates aren't free and there are users we can't
5668 // extend, it isn't worthwhile.
5671 // Remember if this value is live-out.
5672 if (User->getOpcode() == ISD::CopyToReg)
5673 HasCopyToRegUses = true;
5676 if (HasCopyToRegUses) {
5677 bool BothLiveOut = false;
5678 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
5680 SDUse &Use = UI.getUse();
5681 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
5687 // Both unextended and extended values are live out. There had better be
5688 // a good reason for the transformation.
5689 return ExtendNodes.size();
5694 void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
5695 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
5696 ISD::NodeType ExtType) {
5697 // Extend SetCC uses if necessary.
5698 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
5699 SDNode *SetCC = SetCCs[i];
5700 SmallVector<SDValue, 4> Ops;
5702 for (unsigned j = 0; j != 2; ++j) {
5703 SDValue SOp = SetCC->getOperand(j);
5705 Ops.push_back(ExtLoad);
5707 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
5710 Ops.push_back(SetCC->getOperand(2));
5711 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), Ops));
5715 // FIXME: Bring more similar combines here, common to sext/zext (maybe aext?).
5716 SDValue DAGCombiner::CombineExtLoad(SDNode *N) {
5717 SDValue N0 = N->getOperand(0);
5718 EVT DstVT = N->getValueType(0);
5719 EVT SrcVT = N0.getValueType();
5721 assert((N->getOpcode() == ISD::SIGN_EXTEND ||
5722 N->getOpcode() == ISD::ZERO_EXTEND) &&
5723 "Unexpected node type (not an extend)!");
5725 // fold (sext (load x)) to multiple smaller sextloads; same for zext.
5726 // For example, on a target with legal v4i32, but illegal v8i32, turn:
5727 // (v8i32 (sext (v8i16 (load x))))
5729 // (v8i32 (concat_vectors (v4i32 (sextload x)),
5730 // (v4i32 (sextload (x + 16)))))
5731 // Where uses of the original load, i.e.:
5733 // are replaced with:
5735 // (v8i32 (concat_vectors (v4i32 (sextload x)),
5736 // (v4i32 (sextload (x + 16)))))))
5738 // This combine is only applicable to illegal, but splittable, vectors.
5739 // All legal types, and illegal non-vector types, are handled elsewhere.
5740 // This combine is controlled by TargetLowering::isVectorLoadExtDesirable.
5742 if (N0->getOpcode() != ISD::LOAD)
5745 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5747 if (!ISD::isNON_EXTLoad(LN0) || !ISD::isUNINDEXEDLoad(LN0) ||
5748 !N0.hasOneUse() || LN0->isVolatile() || !DstVT.isVector() ||
5749 !DstVT.isPow2VectorType() || !TLI.isVectorLoadExtDesirable(SDValue(N, 0)))
5752 SmallVector<SDNode *, 4> SetCCs;
5753 if (!ExtendUsesToFormExtLoad(N, N0, N->getOpcode(), SetCCs, TLI))
5756 ISD::LoadExtType ExtType =
5757 N->getOpcode() == ISD::SIGN_EXTEND ? ISD::SEXTLOAD : ISD::ZEXTLOAD;
5759 // Try to split the vector types to get down to legal types.
5760 EVT SplitSrcVT = SrcVT;
5761 EVT SplitDstVT = DstVT;
5762 while (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT) &&
5763 SplitSrcVT.getVectorNumElements() > 1) {
5764 SplitDstVT = DAG.GetSplitDestVTs(SplitDstVT).first;
5765 SplitSrcVT = DAG.GetSplitDestVTs(SplitSrcVT).first;
5768 if (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT))
5772 const unsigned NumSplits =
5773 DstVT.getVectorNumElements() / SplitDstVT.getVectorNumElements();
5774 const unsigned Stride = SplitSrcVT.getStoreSize();
5775 SmallVector<SDValue, 4> Loads;
5776 SmallVector<SDValue, 4> Chains;
5778 SDValue BasePtr = LN0->getBasePtr();
5779 for (unsigned Idx = 0; Idx < NumSplits; Idx++) {
5780 const unsigned Offset = Idx * Stride;
5781 const unsigned Align = MinAlign(LN0->getAlignment(), Offset);
5783 SDValue SplitLoad = DAG.getExtLoad(
5784 ExtType, DL, SplitDstVT, LN0->getChain(), BasePtr,
5785 LN0->getPointerInfo().getWithOffset(Offset), SplitSrcVT,
5786 LN0->isVolatile(), LN0->isNonTemporal(), LN0->isInvariant(),
5787 Align, LN0->getAAInfo());
5789 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
5790 DAG.getConstant(Stride, DL, BasePtr.getValueType()));
5792 Loads.push_back(SplitLoad.getValue(0));
5793 Chains.push_back(SplitLoad.getValue(1));
5796 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
5797 SDValue NewValue = DAG.getNode(ISD::CONCAT_VECTORS, DL, DstVT, Loads);
5799 CombineTo(N, NewValue);
5801 // Replace uses of the original load (before extension)
5802 // with a truncate of the concatenated sextloaded vectors.
5804 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(), NewValue);
5805 CombineTo(N0.getNode(), Trunc, NewChain);
5806 ExtendSetCCUses(SetCCs, Trunc, NewValue, DL,
5807 (ISD::NodeType)N->getOpcode());
5808 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5811 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
5812 SDValue N0 = N->getOperand(0);
5813 EVT VT = N->getValueType(0);
5815 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5817 return SDValue(Res, 0);
5819 // fold (sext (sext x)) -> (sext x)
5820 // fold (sext (aext x)) -> (sext x)
5821 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5822 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT,
5825 if (N0.getOpcode() == ISD::TRUNCATE) {
5826 // fold (sext (truncate (load x))) -> (sext (smaller load x))
5827 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
5828 if (SDValue NarrowLoad = ReduceLoadWidth(N0.getNode())) {
5829 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5830 if (NarrowLoad.getNode() != N0.getNode()) {
5831 CombineTo(N0.getNode(), NarrowLoad);
5832 // CombineTo deleted the truncate, if needed, but not what's under it.
5835 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5838 // See if the value being truncated is already sign extended. If so, just
5839 // eliminate the trunc/sext pair.
5840 SDValue Op = N0.getOperand(0);
5841 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
5842 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
5843 unsigned DestBits = VT.getScalarType().getSizeInBits();
5844 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
5846 if (OpBits == DestBits) {
5847 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
5848 // bits, it is already ready.
5849 if (NumSignBits > DestBits-MidBits)
5851 } else if (OpBits < DestBits) {
5852 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
5853 // bits, just sext from i32.
5854 if (NumSignBits > OpBits-MidBits)
5855 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, Op);
5857 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
5858 // bits, just truncate to i32.
5859 if (NumSignBits > OpBits-MidBits)
5860 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5863 // fold (sext (truncate x)) -> (sextinreg x).
5864 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
5865 N0.getValueType())) {
5866 if (OpBits < DestBits)
5867 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op);
5868 else if (OpBits > DestBits)
5869 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op);
5870 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, Op,
5871 DAG.getValueType(N0.getValueType()));
5875 // fold (sext (load x)) -> (sext (truncate (sextload x)))
5876 // Only generate vector extloads when 1) they're legal, and 2) they are
5877 // deemed desirable by the target.
5878 if (ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5879 ((!LegalOperations && !VT.isVector() &&
5880 !cast<LoadSDNode>(N0)->isVolatile()) ||
5881 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, N0.getValueType()))) {
5882 bool DoXform = true;
5883 SmallVector<SDNode*, 4> SetCCs;
5884 if (!N0.hasOneUse())
5885 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
5887 DoXform &= TLI.isVectorLoadExtDesirable(SDValue(N, 0));
5889 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5890 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5892 LN0->getBasePtr(), N0.getValueType(),
5893 LN0->getMemOperand());
5894 CombineTo(N, ExtLoad);
5895 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5896 N0.getValueType(), ExtLoad);
5897 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5898 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5900 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5904 // fold (sext (load x)) to multiple smaller sextloads.
5905 // Only on illegal but splittable vectors.
5906 if (SDValue ExtLoad = CombineExtLoad(N))
5909 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
5910 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
5911 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5912 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5913 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5914 EVT MemVT = LN0->getMemoryVT();
5915 if ((!LegalOperations && !LN0->isVolatile()) ||
5916 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, MemVT)) {
5917 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5919 LN0->getBasePtr(), MemVT,
5920 LN0->getMemOperand());
5921 CombineTo(N, ExtLoad);
5922 CombineTo(N0.getNode(),
5923 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5924 N0.getValueType(), ExtLoad),
5925 ExtLoad.getValue(1));
5926 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5930 // fold (sext (and/or/xor (load x), cst)) ->
5931 // (and/or/xor (sextload x), (sext cst))
5932 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5933 N0.getOpcode() == ISD::XOR) &&
5934 isa<LoadSDNode>(N0.getOperand(0)) &&
5935 N0.getOperand(1).getOpcode() == ISD::Constant &&
5936 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, N0.getValueType()) &&
5937 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5938 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5939 if (LN0->getExtensionType() != ISD::ZEXTLOAD && LN0->isUnindexed()) {
5940 bool DoXform = true;
5941 SmallVector<SDNode*, 4> SetCCs;
5942 if (!N0.hasOneUse())
5943 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
5946 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN0), VT,
5947 LN0->getChain(), LN0->getBasePtr(),
5949 LN0->getMemOperand());
5950 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5951 Mask = Mask.sext(VT.getSizeInBits());
5953 SDValue And = DAG.getNode(N0.getOpcode(), DL, VT,
5954 ExtLoad, DAG.getConstant(Mask, DL, VT));
5955 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5956 SDLoc(N0.getOperand(0)),
5957 N0.getOperand(0).getValueType(), ExtLoad);
5959 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5960 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, DL,
5962 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5967 if (N0.getOpcode() == ISD::SETCC) {
5968 EVT N0VT = N0.getOperand(0).getValueType();
5969 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
5970 // Only do this before legalize for now.
5971 if (VT.isVector() && !LegalOperations &&
5972 TLI.getBooleanContents(N0VT) ==
5973 TargetLowering::ZeroOrNegativeOneBooleanContent) {
5974 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
5975 // of the same size as the compared operands. Only optimize sext(setcc())
5976 // if this is the case.
5977 EVT SVT = getSetCCResultType(N0VT);
5979 // We know that the # elements of the results is the same as the
5980 // # elements of the compare (and the # elements of the compare result
5981 // for that matter). Check to see that they are the same size. If so,
5982 // we know that the element size of the sext'd result matches the
5983 // element size of the compare operands.
5984 if (VT.getSizeInBits() == SVT.getSizeInBits())
5985 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5987 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5989 // If the desired elements are smaller or larger than the source
5990 // elements we can use a matching integer vector type and then
5991 // truncate/sign extend
5992 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5993 if (SVT == MatchingVectorType) {
5994 SDValue VsetCC = DAG.getSetCC(SDLoc(N), MatchingVectorType,
5995 N0.getOperand(0), N0.getOperand(1),
5996 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5997 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
6001 // sext(setcc x, y, cc) -> (select (setcc x, y, cc), -1, 0)
6002 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
6005 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), DL, VT);
6007 SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1),
6008 NegOne, DAG.getConstant(0, DL, VT),
6009 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
6010 if (SCC.getNode()) return SCC;
6012 if (!VT.isVector()) {
6013 EVT SetCCVT = getSetCCResultType(N0.getOperand(0).getValueType());
6014 if (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, SetCCVT)) {
6016 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
6017 SDValue SetCC = DAG.getSetCC(DL, SetCCVT,
6018 N0.getOperand(0), N0.getOperand(1), CC);
6019 return DAG.getSelect(DL, VT, SetCC,
6020 NegOne, DAG.getConstant(0, DL, VT));
6025 // fold (sext x) -> (zext x) if the sign bit is known zero.
6026 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
6027 DAG.SignBitIsZero(N0))
6028 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0);
6033 // isTruncateOf - If N is a truncate of some other value, return true, record
6034 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
6035 // This function computes KnownZero to avoid a duplicated call to
6036 // computeKnownBits in the caller.
6037 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
6040 if (N->getOpcode() == ISD::TRUNCATE) {
6041 Op = N->getOperand(0);
6042 DAG.computeKnownBits(Op, KnownZero, KnownOne);
6046 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
6047 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
6050 SDValue Op0 = N->getOperand(0);
6051 SDValue Op1 = N->getOperand(1);
6052 assert(Op0.getValueType() == Op1.getValueType());
6054 if (isNullConstant(Op0))
6056 else if (isNullConstant(Op1))
6061 DAG.computeKnownBits(Op, KnownZero, KnownOne);
6063 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
6069 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
6070 SDValue N0 = N->getOperand(0);
6071 EVT VT = N->getValueType(0);
6073 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
6075 return SDValue(Res, 0);
6077 // fold (zext (zext x)) -> (zext x)
6078 // fold (zext (aext x)) -> (zext x)
6079 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
6080 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT,
6083 // fold (zext (truncate x)) -> (zext x) or
6084 // (zext (truncate x)) -> (truncate x)
6085 // This is valid when the truncated bits of x are already zero.
6086 // FIXME: We should extend this to work for vectors too.
6089 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
6090 APInt TruncatedBits =
6091 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
6092 APInt(Op.getValueSizeInBits(), 0) :
6093 APInt::getBitsSet(Op.getValueSizeInBits(),
6094 N0.getValueSizeInBits(),
6095 std::min(Op.getValueSizeInBits(),
6096 VT.getSizeInBits()));
6097 if (TruncatedBits == (KnownZero & TruncatedBits)) {
6098 if (VT.bitsGT(Op.getValueType()))
6099 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, Op);
6100 if (VT.bitsLT(Op.getValueType()))
6101 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
6107 // fold (zext (truncate (load x))) -> (zext (smaller load x))
6108 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
6109 if (N0.getOpcode() == ISD::TRUNCATE) {
6110 if (SDValue NarrowLoad = ReduceLoadWidth(N0.getNode())) {
6111 SDNode* oye = N0.getNode()->getOperand(0).getNode();
6112 if (NarrowLoad.getNode() != N0.getNode()) {
6113 CombineTo(N0.getNode(), NarrowLoad);
6114 // CombineTo deleted the truncate, if needed, but not what's under it.
6117 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6121 // fold (zext (truncate x)) -> (and x, mask)
6122 if (N0.getOpcode() == ISD::TRUNCATE &&
6123 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
6125 // fold (zext (truncate (load x))) -> (zext (smaller load x))
6126 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
6127 if (SDValue NarrowLoad = ReduceLoadWidth(N0.getNode())) {
6128 SDNode* oye = N0.getNode()->getOperand(0).getNode();
6129 if (NarrowLoad.getNode() != N0.getNode()) {
6130 CombineTo(N0.getNode(), NarrowLoad);
6131 // CombineTo deleted the truncate, if needed, but not what's under it.
6134 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6137 SDValue Op = N0.getOperand(0);
6138 if (Op.getValueType().bitsLT(VT)) {
6139 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, Op);
6140 AddToWorklist(Op.getNode());
6141 } else if (Op.getValueType().bitsGT(VT)) {
6142 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
6143 AddToWorklist(Op.getNode());
6145 return DAG.getZeroExtendInReg(Op, SDLoc(N),
6146 N0.getValueType().getScalarType());
6149 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
6150 // if either of the casts is not free.
6151 if (N0.getOpcode() == ISD::AND &&
6152 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
6153 N0.getOperand(1).getOpcode() == ISD::Constant &&
6154 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
6155 N0.getValueType()) ||
6156 !TLI.isZExtFree(N0.getValueType(), VT))) {
6157 SDValue X = N0.getOperand(0).getOperand(0);
6158 if (X.getValueType().bitsLT(VT)) {
6159 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(X), VT, X);
6160 } else if (X.getValueType().bitsGT(VT)) {
6161 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
6163 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
6164 Mask = Mask.zext(VT.getSizeInBits());
6166 return DAG.getNode(ISD::AND, DL, VT,
6167 X, DAG.getConstant(Mask, DL, VT));
6170 // fold (zext (load x)) -> (zext (truncate (zextload x)))
6171 // Only generate vector extloads when 1) they're legal, and 2) they are
6172 // deemed desirable by the target.
6173 if (ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
6174 ((!LegalOperations && !VT.isVector() &&
6175 !cast<LoadSDNode>(N0)->isVolatile()) ||
6176 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, N0.getValueType()))) {
6177 bool DoXform = true;
6178 SmallVector<SDNode*, 4> SetCCs;
6179 if (!N0.hasOneUse())
6180 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
6182 DoXform &= TLI.isVectorLoadExtDesirable(SDValue(N, 0));
6184 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6185 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
6187 LN0->getBasePtr(), N0.getValueType(),
6188 LN0->getMemOperand());
6189 CombineTo(N, ExtLoad);
6190 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
6191 N0.getValueType(), ExtLoad);
6192 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
6194 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
6196 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6200 // fold (zext (load x)) to multiple smaller zextloads.
6201 // Only on illegal but splittable vectors.
6202 if (SDValue ExtLoad = CombineExtLoad(N))
6205 // fold (zext (and/or/xor (load x), cst)) ->
6206 // (and/or/xor (zextload x), (zext cst))
6207 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
6208 N0.getOpcode() == ISD::XOR) &&
6209 isa<LoadSDNode>(N0.getOperand(0)) &&
6210 N0.getOperand(1).getOpcode() == ISD::Constant &&
6211 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, N0.getValueType()) &&
6212 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
6213 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
6214 if (LN0->getExtensionType() != ISD::SEXTLOAD && LN0->isUnindexed()) {
6215 bool DoXform = true;
6216 SmallVector<SDNode*, 4> SetCCs;
6217 if (!N0.hasOneUse())
6218 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
6221 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), VT,
6222 LN0->getChain(), LN0->getBasePtr(),
6224 LN0->getMemOperand());
6225 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
6226 Mask = Mask.zext(VT.getSizeInBits());
6228 SDValue And = DAG.getNode(N0.getOpcode(), DL, VT,
6229 ExtLoad, DAG.getConstant(Mask, DL, VT));
6230 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
6231 SDLoc(N0.getOperand(0)),
6232 N0.getOperand(0).getValueType(), ExtLoad);
6234 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
6235 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, DL,
6237 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6242 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
6243 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
6244 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
6245 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
6246 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6247 EVT MemVT = LN0->getMemoryVT();
6248 if ((!LegalOperations && !LN0->isVolatile()) ||
6249 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT)) {
6250 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
6252 LN0->getBasePtr(), MemVT,
6253 LN0->getMemOperand());
6254 CombineTo(N, ExtLoad);
6255 CombineTo(N0.getNode(),
6256 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(),
6258 ExtLoad.getValue(1));
6259 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6263 if (N0.getOpcode() == ISD::SETCC) {
6264 if (!LegalOperations && VT.isVector() &&
6265 N0.getValueType().getVectorElementType() == MVT::i1) {
6266 EVT N0VT = N0.getOperand(0).getValueType();
6267 if (getSetCCResultType(N0VT) == N0.getValueType())
6270 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
6271 // Only do this before legalize for now.
6272 EVT EltVT = VT.getVectorElementType();
6274 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
6275 DAG.getConstant(1, DL, EltVT));
6276 if (VT.getSizeInBits() == N0VT.getSizeInBits())
6277 // We know that the # elements of the results is the same as the
6278 // # elements of the compare (and the # elements of the compare result
6279 // for that matter). Check to see that they are the same size. If so,
6280 // we know that the element size of the sext'd result matches the
6281 // element size of the compare operands.
6282 return DAG.getNode(ISD::AND, DL, VT,
6283 DAG.getSetCC(DL, VT, N0.getOperand(0),
6285 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
6286 DAG.getNode(ISD::BUILD_VECTOR, DL, VT,
6289 // If the desired elements are smaller or larger than the source
6290 // elements we can use a matching integer vector type and then
6291 // truncate/sign extend
6292 EVT MatchingElementType =
6293 EVT::getIntegerVT(*DAG.getContext(),
6294 N0VT.getScalarType().getSizeInBits());
6295 EVT MatchingVectorType =
6296 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
6297 N0VT.getVectorNumElements());
6299 DAG.getSetCC(DL, MatchingVectorType, N0.getOperand(0),
6301 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6302 return DAG.getNode(ISD::AND, DL, VT,
6303 DAG.getSExtOrTrunc(VsetCC, DL, VT),
6304 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, OneOps));
6307 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
6310 SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1),
6311 DAG.getConstant(1, DL, VT), DAG.getConstant(0, DL, VT),
6312 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
6313 if (SCC.getNode()) return SCC;
6316 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
6317 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
6318 isa<ConstantSDNode>(N0.getOperand(1)) &&
6319 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
6321 SDValue ShAmt = N0.getOperand(1);
6322 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
6323 if (N0.getOpcode() == ISD::SHL) {
6324 SDValue InnerZExt = N0.getOperand(0);
6325 // If the original shl may be shifting out bits, do not perform this
6327 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
6328 InnerZExt.getOperand(0).getValueType().getSizeInBits();
6329 if (ShAmtVal > KnownZeroBits)
6335 // Ensure that the shift amount is wide enough for the shifted value.
6336 if (VT.getSizeInBits() >= 256)
6337 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
6339 return DAG.getNode(N0.getOpcode(), DL, VT,
6340 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
6347 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
6348 SDValue N0 = N->getOperand(0);
6349 EVT VT = N->getValueType(0);
6351 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
6353 return SDValue(Res, 0);
6355 // fold (aext (aext x)) -> (aext x)
6356 // fold (aext (zext x)) -> (zext x)
6357 // fold (aext (sext x)) -> (sext x)
6358 if (N0.getOpcode() == ISD::ANY_EXTEND ||
6359 N0.getOpcode() == ISD::ZERO_EXTEND ||
6360 N0.getOpcode() == ISD::SIGN_EXTEND)
6361 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0));
6363 // fold (aext (truncate (load x))) -> (aext (smaller load x))
6364 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
6365 if (N0.getOpcode() == ISD::TRUNCATE) {
6366 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
6367 if (NarrowLoad.getNode()) {
6368 SDNode* oye = N0.getNode()->getOperand(0).getNode();
6369 if (NarrowLoad.getNode() != N0.getNode()) {
6370 CombineTo(N0.getNode(), NarrowLoad);
6371 // CombineTo deleted the truncate, if needed, but not what's under it.
6374 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6378 // fold (aext (truncate x))
6379 if (N0.getOpcode() == ISD::TRUNCATE) {
6380 SDValue TruncOp = N0.getOperand(0);
6381 if (TruncOp.getValueType() == VT)
6382 return TruncOp; // x iff x size == zext size.
6383 if (TruncOp.getValueType().bitsGT(VT))
6384 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, TruncOp);
6385 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, TruncOp);
6388 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
6389 // if the trunc is not free.
6390 if (N0.getOpcode() == ISD::AND &&
6391 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
6392 N0.getOperand(1).getOpcode() == ISD::Constant &&
6393 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
6394 N0.getValueType())) {
6395 SDValue X = N0.getOperand(0).getOperand(0);
6396 if (X.getValueType().bitsLT(VT)) {
6397 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, X);
6398 } else if (X.getValueType().bitsGT(VT)) {
6399 X = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, X);
6401 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
6402 Mask = Mask.zext(VT.getSizeInBits());
6404 return DAG.getNode(ISD::AND, DL, VT,
6405 X, DAG.getConstant(Mask, DL, VT));
6408 // fold (aext (load x)) -> (aext (truncate (extload x)))
6409 // None of the supported targets knows how to perform load and any_ext
6410 // on vectors in one instruction. We only perform this transformation on
6412 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
6413 ISD::isUNINDEXEDLoad(N0.getNode()) &&
6414 TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) {
6415 bool DoXform = true;
6416 SmallVector<SDNode*, 4> SetCCs;
6417 if (!N0.hasOneUse())
6418 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
6420 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6421 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
6423 LN0->getBasePtr(), N0.getValueType(),
6424 LN0->getMemOperand());
6425 CombineTo(N, ExtLoad);
6426 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
6427 N0.getValueType(), ExtLoad);
6428 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
6429 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
6431 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6435 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
6436 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
6437 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
6438 if (N0.getOpcode() == ISD::LOAD &&
6439 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
6441 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6442 ISD::LoadExtType ExtType = LN0->getExtensionType();
6443 EVT MemVT = LN0->getMemoryVT();
6444 if (!LegalOperations || TLI.isLoadExtLegal(ExtType, VT, MemVT)) {
6445 SDValue ExtLoad = DAG.getExtLoad(ExtType, SDLoc(N),
6446 VT, LN0->getChain(), LN0->getBasePtr(),
6447 MemVT, LN0->getMemOperand());
6448 CombineTo(N, ExtLoad);
6449 CombineTo(N0.getNode(),
6450 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
6451 N0.getValueType(), ExtLoad),
6452 ExtLoad.getValue(1));
6453 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6457 if (N0.getOpcode() == ISD::SETCC) {
6459 // aext(setcc) -> vsetcc
6460 // aext(setcc) -> truncate(vsetcc)
6461 // aext(setcc) -> aext(vsetcc)
6462 // Only do this before legalize for now.
6463 if (VT.isVector() && !LegalOperations) {
6464 EVT N0VT = N0.getOperand(0).getValueType();
6465 // We know that the # elements of the results is the same as the
6466 // # elements of the compare (and the # elements of the compare result
6467 // for that matter). Check to see that they are the same size. If so,
6468 // we know that the element size of the sext'd result matches the
6469 // element size of the compare operands.
6470 if (VT.getSizeInBits() == N0VT.getSizeInBits())
6471 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
6473 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6474 // If the desired elements are smaller or larger than the source
6475 // elements we can use a matching integer vector type and then
6476 // truncate/any extend
6478 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
6480 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
6482 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6483 return DAG.getAnyExtOrTrunc(VsetCC, SDLoc(N), VT);
6487 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
6490 SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1),
6491 DAG.getConstant(1, DL, VT), DAG.getConstant(0, DL, VT),
6492 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
6500 /// See if the specified operand can be simplified with the knowledge that only
6501 /// the bits specified by Mask are used. If so, return the simpler operand,
6502 /// otherwise return a null SDValue.
6503 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
6504 switch (V.getOpcode()) {
6506 case ISD::Constant: {
6507 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
6508 assert(CV && "Const value should be ConstSDNode.");
6509 const APInt &CVal = CV->getAPIntValue();
6510 APInt NewVal = CVal & Mask;
6512 return DAG.getConstant(NewVal, SDLoc(V), V.getValueType());
6517 // If the LHS or RHS don't contribute bits to the or, drop them.
6518 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
6519 return V.getOperand(1);
6520 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
6521 return V.getOperand(0);
6524 // Only look at single-use SRLs.
6525 if (!V.getNode()->hasOneUse())
6527 if (ConstantSDNode *RHSC = getAsNonOpaqueConstant(V.getOperand(1))) {
6528 // See if we can recursively simplify the LHS.
6529 unsigned Amt = RHSC->getZExtValue();
6531 // Watch out for shift count overflow though.
6532 if (Amt >= Mask.getBitWidth()) break;
6533 APInt NewMask = Mask << Amt;
6534 if (SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask))
6535 return DAG.getNode(ISD::SRL, SDLoc(V), V.getValueType(),
6536 SimplifyLHS, V.getOperand(1));
6542 /// If the result of a wider load is shifted to right of N bits and then
6543 /// truncated to a narrower type and where N is a multiple of number of bits of
6544 /// the narrower type, transform it to a narrower load from address + N / num of
6545 /// bits of new type. If the result is to be extended, also fold the extension
6546 /// to form a extending load.
6547 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
6548 unsigned Opc = N->getOpcode();
6550 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
6551 SDValue N0 = N->getOperand(0);
6552 EVT VT = N->getValueType(0);
6555 // This transformation isn't valid for vector loads.
6559 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
6561 if (Opc == ISD::SIGN_EXTEND_INREG) {
6562 ExtType = ISD::SEXTLOAD;
6563 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
6564 } else if (Opc == ISD::SRL) {
6565 // Another special-case: SRL is basically zero-extending a narrower value.
6566 ExtType = ISD::ZEXTLOAD;
6568 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
6569 if (!N01) return SDValue();
6570 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
6571 VT.getSizeInBits() - N01->getZExtValue());
6573 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, VT, ExtVT))
6576 unsigned EVTBits = ExtVT.getSizeInBits();
6578 // Do not generate loads of non-round integer types since these can
6579 // be expensive (and would be wrong if the type is not byte sized).
6580 if (!ExtVT.isRound())
6584 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
6585 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
6586 ShAmt = N01->getZExtValue();
6587 // Is the shift amount a multiple of size of VT?
6588 if ((ShAmt & (EVTBits-1)) == 0) {
6589 N0 = N0.getOperand(0);
6590 // Is the load width a multiple of size of VT?
6591 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
6595 // At this point, we must have a load or else we can't do the transform.
6596 if (!isa<LoadSDNode>(N0)) return SDValue();
6598 // Because a SRL must be assumed to *need* to zero-extend the high bits
6599 // (as opposed to anyext the high bits), we can't combine the zextload
6600 // lowering of SRL and an sextload.
6601 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
6604 // If the shift amount is larger than the input type then we're not
6605 // accessing any of the loaded bytes. If the load was a zextload/extload
6606 // then the result of the shift+trunc is zero/undef (handled elsewhere).
6607 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
6612 // If the load is shifted left (and the result isn't shifted back right),
6613 // we can fold the truncate through the shift.
6614 unsigned ShLeftAmt = 0;
6615 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
6616 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
6617 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
6618 ShLeftAmt = N01->getZExtValue();
6619 N0 = N0.getOperand(0);
6623 // If we haven't found a load, we can't narrow it. Don't transform one with
6624 // multiple uses, this would require adding a new load.
6625 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse())
6628 // Don't change the width of a volatile load.
6629 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6630 if (LN0->isVolatile())
6633 // Verify that we are actually reducing a load width here.
6634 if (LN0->getMemoryVT().getSizeInBits() < EVTBits)
6637 // For the transform to be legal, the load must produce only two values
6638 // (the value loaded and the chain). Don't transform a pre-increment
6639 // load, for example, which produces an extra value. Otherwise the
6640 // transformation is not equivalent, and the downstream logic to replace
6641 // uses gets things wrong.
6642 if (LN0->getNumValues() > 2)
6645 // If the load that we're shrinking is an extload and we're not just
6646 // discarding the extension we can't simply shrink the load. Bail.
6647 // TODO: It would be possible to merge the extensions in some cases.
6648 if (LN0->getExtensionType() != ISD::NON_EXTLOAD &&
6649 LN0->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits() + ShAmt)
6652 if (!TLI.shouldReduceLoadWidth(LN0, ExtType, ExtVT))
6655 EVT PtrType = N0.getOperand(1).getValueType();
6657 if (PtrType == MVT::Untyped || PtrType.isExtended())
6658 // It's not possible to generate a constant of extended or untyped type.
6661 // For big endian targets, we need to adjust the offset to the pointer to
6662 // load the correct bytes.
6663 if (DAG.getDataLayout().isBigEndian()) {
6664 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
6665 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
6666 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
6669 uint64_t PtrOff = ShAmt / 8;
6670 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
6672 SDValue NewPtr = DAG.getNode(ISD::ADD, DL,
6673 PtrType, LN0->getBasePtr(),
6674 DAG.getConstant(PtrOff, DL, PtrType));
6675 AddToWorklist(NewPtr.getNode());
6678 if (ExtType == ISD::NON_EXTLOAD)
6679 Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr,
6680 LN0->getPointerInfo().getWithOffset(PtrOff),
6681 LN0->isVolatile(), LN0->isNonTemporal(),
6682 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
6684 Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(),NewPtr,
6685 LN0->getPointerInfo().getWithOffset(PtrOff),
6686 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
6687 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
6689 // Replace the old load's chain with the new load's chain.
6690 WorklistRemover DeadNodes(*this);
6691 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
6693 // Shift the result left, if we've swallowed a left shift.
6694 SDValue Result = Load;
6695 if (ShLeftAmt != 0) {
6696 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
6697 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
6699 // If the shift amount is as large as the result size (but, presumably,
6700 // no larger than the source) then the useful bits of the result are
6701 // zero; we can't simply return the shortened shift, because the result
6702 // of that operation is undefined.
6704 if (ShLeftAmt >= VT.getSizeInBits())
6705 Result = DAG.getConstant(0, DL, VT);
6707 Result = DAG.getNode(ISD::SHL, DL, VT,
6708 Result, DAG.getConstant(ShLeftAmt, DL, ShImmTy));
6711 // Return the new loaded value.
6715 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
6716 SDValue N0 = N->getOperand(0);
6717 SDValue N1 = N->getOperand(1);
6718 EVT VT = N->getValueType(0);
6719 EVT EVT = cast<VTSDNode>(N1)->getVT();
6720 unsigned VTBits = VT.getScalarType().getSizeInBits();
6721 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
6723 // fold (sext_in_reg c1) -> c1
6724 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
6725 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1);
6727 // If the input is already sign extended, just drop the extension.
6728 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
6731 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
6732 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
6733 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT()))
6734 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
6735 N0.getOperand(0), N1);
6737 // fold (sext_in_reg (sext x)) -> (sext x)
6738 // fold (sext_in_reg (aext x)) -> (sext x)
6739 // if x is small enough.
6740 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
6741 SDValue N00 = N0.getOperand(0);
6742 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
6743 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
6744 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
6747 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
6748 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
6749 return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT);
6751 // fold operands of sext_in_reg based on knowledge that the top bits are not
6753 if (SimplifyDemandedBits(SDValue(N, 0)))
6754 return SDValue(N, 0);
6756 // fold (sext_in_reg (load x)) -> (smaller sextload x)
6757 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
6758 if (SDValue NarrowLoad = ReduceLoadWidth(N))
6761 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
6762 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
6763 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
6764 if (N0.getOpcode() == ISD::SRL) {
6765 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
6766 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
6767 // We can turn this into an SRA iff the input to the SRL is already sign
6769 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
6770 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
6771 return DAG.getNode(ISD::SRA, SDLoc(N), VT,
6772 N0.getOperand(0), N0.getOperand(1));
6776 // fold (sext_inreg (extload x)) -> (sextload x)
6777 if (ISD::isEXTLoad(N0.getNode()) &&
6778 ISD::isUNINDEXEDLoad(N0.getNode()) &&
6779 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
6780 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6781 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) {
6782 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6783 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
6785 LN0->getBasePtr(), EVT,
6786 LN0->getMemOperand());
6787 CombineTo(N, ExtLoad);
6788 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
6789 AddToWorklist(ExtLoad.getNode());
6790 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6792 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
6793 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
6795 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
6796 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6797 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) {
6798 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6799 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
6801 LN0->getBasePtr(), EVT,
6802 LN0->getMemOperand());
6803 CombineTo(N, ExtLoad);
6804 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
6805 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6808 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
6809 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
6810 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
6811 N0.getOperand(1), false);
6812 if (BSwap.getNode())
6813 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
6817 // Fold a sext_inreg of a build_vector of ConstantSDNodes or undefs
6818 // into a build_vector.
6819 if (ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
6820 SmallVector<SDValue, 8> Elts;
6821 unsigned NumElts = N0->getNumOperands();
6822 unsigned ShAmt = VTBits - EVTBits;
6824 for (unsigned i = 0; i != NumElts; ++i) {
6825 SDValue Op = N0->getOperand(i);
6826 if (Op->getOpcode() == ISD::UNDEF) {
6831 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
6832 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
6833 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
6834 SDLoc(Op), Op.getValueType()));
6837 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Elts);
6843 SDValue DAGCombiner::visitSIGN_EXTEND_VECTOR_INREG(SDNode *N) {
6844 SDValue N0 = N->getOperand(0);
6845 EVT VT = N->getValueType(0);
6847 if (N0.getOpcode() == ISD::UNDEF)
6848 return DAG.getUNDEF(VT);
6850 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
6852 return SDValue(Res, 0);
6857 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
6858 SDValue N0 = N->getOperand(0);
6859 EVT VT = N->getValueType(0);
6860 bool isLE = DAG.getDataLayout().isLittleEndian();
6863 if (N0.getValueType() == N->getValueType(0))
6865 // fold (truncate c1) -> c1
6866 if (isConstantIntBuildVectorOrConstantInt(N0))
6867 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0);
6868 // fold (truncate (truncate x)) -> (truncate x)
6869 if (N0.getOpcode() == ISD::TRUNCATE)
6870 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6871 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
6872 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
6873 N0.getOpcode() == ISD::SIGN_EXTEND ||
6874 N0.getOpcode() == ISD::ANY_EXTEND) {
6875 if (N0.getOperand(0).getValueType().bitsLT(VT))
6876 // if the source is smaller than the dest, we still need an extend
6877 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
6879 if (N0.getOperand(0).getValueType().bitsGT(VT))
6880 // if the source is larger than the dest, than we just need the truncate
6881 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6882 // if the source and dest are the same type, we can drop both the extend
6883 // and the truncate.
6884 return N0.getOperand(0);
6887 // Fold extract-and-trunc into a narrow extract. For example:
6888 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
6889 // i32 y = TRUNCATE(i64 x)
6891 // v16i8 b = BITCAST (v2i64 val)
6892 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
6894 // Note: We only run this optimization after type legalization (which often
6895 // creates this pattern) and before operation legalization after which
6896 // we need to be more careful about the vector instructions that we generate.
6897 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6898 LegalTypes && !LegalOperations && N0->hasOneUse() && VT != MVT::i1) {
6900 EVT VecTy = N0.getOperand(0).getValueType();
6901 EVT ExTy = N0.getValueType();
6902 EVT TrTy = N->getValueType(0);
6904 unsigned NumElem = VecTy.getVectorNumElements();
6905 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
6907 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
6908 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
6910 SDValue EltNo = N0->getOperand(1);
6911 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
6912 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
6913 EVT IndexTy = TLI.getVectorIdxTy(DAG.getDataLayout());
6914 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
6916 SDValue V = DAG.getNode(ISD::BITCAST, SDLoc(N),
6917 NVT, N0.getOperand(0));
6920 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
6922 DAG.getConstant(Index, DL, IndexTy));
6926 // trunc (select c, a, b) -> select c, (trunc a), (trunc b)
6927 if (N0.getOpcode() == ISD::SELECT) {
6928 EVT SrcVT = N0.getValueType();
6929 if ((!LegalOperations || TLI.isOperationLegal(ISD::SELECT, SrcVT)) &&
6930 TLI.isTruncateFree(SrcVT, VT)) {
6932 SDValue Cond = N0.getOperand(0);
6933 SDValue TruncOp0 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(1));
6934 SDValue TruncOp1 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(2));
6935 return DAG.getNode(ISD::SELECT, SDLoc(N), VT, Cond, TruncOp0, TruncOp1);
6939 // Fold a series of buildvector, bitcast, and truncate if possible.
6941 // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to
6942 // (2xi32 (buildvector x, y)).
6943 if (Level == AfterLegalizeVectorOps && VT.isVector() &&
6944 N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
6945 N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
6946 N0.getOperand(0).hasOneUse()) {
6948 SDValue BuildVect = N0.getOperand(0);
6949 EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType();
6950 EVT TruncVecEltTy = VT.getVectorElementType();
6952 // Check that the element types match.
6953 if (BuildVectEltTy == TruncVecEltTy) {
6954 // Now we only need to compute the offset of the truncated elements.
6955 unsigned BuildVecNumElts = BuildVect.getNumOperands();
6956 unsigned TruncVecNumElts = VT.getVectorNumElements();
6957 unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts;
6959 assert((BuildVecNumElts % TruncVecNumElts) == 0 &&
6960 "Invalid number of elements");
6962 SmallVector<SDValue, 8> Opnds;
6963 for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset)
6964 Opnds.push_back(BuildVect.getOperand(i));
6966 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
6970 // See if we can simplify the input to this truncate through knowledge that
6971 // only the low bits are being used.
6972 // For example "trunc (or (shl x, 8), y)" // -> trunc y
6973 // Currently we only perform this optimization on scalars because vectors
6974 // may have different active low bits.
6975 if (!VT.isVector()) {
6977 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
6978 VT.getSizeInBits()));
6979 if (Shorter.getNode())
6980 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter);
6982 // fold (truncate (load x)) -> (smaller load x)
6983 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
6984 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
6985 if (SDValue Reduced = ReduceLoadWidth(N))
6988 // Handle the case where the load remains an extending load even
6989 // after truncation.
6990 if (N0.hasOneUse() && ISD::isUNINDEXEDLoad(N0.getNode())) {
6991 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6992 if (!LN0->isVolatile() &&
6993 LN0->getMemoryVT().getStoreSizeInBits() < VT.getSizeInBits()) {
6994 SDValue NewLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(LN0),
6995 VT, LN0->getChain(), LN0->getBasePtr(),
6997 LN0->getMemOperand());
6998 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLoad.getValue(1));
7003 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
7004 // where ... are all 'undef'.
7005 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
7006 SmallVector<EVT, 8> VTs;
7009 unsigned NumDefs = 0;
7011 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
7012 SDValue X = N0.getOperand(i);
7013 if (X.getOpcode() != ISD::UNDEF) {
7018 // Stop if more than one members are non-undef.
7021 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
7022 VT.getVectorElementType(),
7023 X.getValueType().getVectorNumElements()));
7027 return DAG.getUNDEF(VT);
7030 assert(V.getNode() && "The single defined operand is empty!");
7031 SmallVector<SDValue, 8> Opnds;
7032 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
7034 Opnds.push_back(DAG.getUNDEF(VTs[i]));
7037 SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V);
7038 AddToWorklist(NV.getNode());
7039 Opnds.push_back(NV);
7041 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Opnds);
7045 // Simplify the operands using demanded-bits information.
7046 if (!VT.isVector() &&
7047 SimplifyDemandedBits(SDValue(N, 0)))
7048 return SDValue(N, 0);
7053 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
7054 SDValue Elt = N->getOperand(i);
7055 if (Elt.getOpcode() != ISD::MERGE_VALUES)
7056 return Elt.getNode();
7057 return Elt.getOperand(Elt.getResNo()).getNode();
7060 /// build_pair (load, load) -> load
7061 /// if load locations are consecutive.
7062 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
7063 assert(N->getOpcode() == ISD::BUILD_PAIR);
7065 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
7066 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
7067 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
7068 LD1->getAddressSpace() != LD2->getAddressSpace())
7070 EVT LD1VT = LD1->getValueType(0);
7072 if (ISD::isNON_EXTLoad(LD2) &&
7074 // If both are volatile this would reduce the number of volatile loads.
7075 // If one is volatile it might be ok, but play conservative and bail out.
7076 !LD1->isVolatile() &&
7077 !LD2->isVolatile() &&
7078 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
7079 unsigned Align = LD1->getAlignment();
7080 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
7081 VT.getTypeForEVT(*DAG.getContext()));
7083 if (NewAlign <= Align &&
7084 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
7085 return DAG.getLoad(VT, SDLoc(N), LD1->getChain(),
7086 LD1->getBasePtr(), LD1->getPointerInfo(),
7087 false, false, false, Align);
7093 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
7094 SDValue N0 = N->getOperand(0);
7095 EVT VT = N->getValueType(0);
7097 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
7098 // Only do this before legalize, since afterward the target may be depending
7099 // on the bitconvert.
7100 // First check to see if this is all constant.
7102 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
7104 bool isSimple = cast<BuildVectorSDNode>(N0)->isConstant();
7106 EVT DestEltVT = N->getValueType(0).getVectorElementType();
7107 assert(!DestEltVT.isVector() &&
7108 "Element type of vector ValueType must not be vector!");
7110 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
7113 // If the input is a constant, let getNode fold it.
7114 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
7115 // If we can't allow illegal operations, we need to check that this is just
7116 // a fp -> int or int -> conversion and that the resulting operation will
7118 if (!LegalOperations ||
7119 (isa<ConstantSDNode>(N0) && VT.isFloatingPoint() && !VT.isVector() &&
7120 TLI.isOperationLegal(ISD::ConstantFP, VT)) ||
7121 (isa<ConstantFPSDNode>(N0) && VT.isInteger() && !VT.isVector() &&
7122 TLI.isOperationLegal(ISD::Constant, VT)))
7123 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, N0);
7126 // (conv (conv x, t1), t2) -> (conv x, t2)
7127 if (N0.getOpcode() == ISD::BITCAST)
7128 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT,
7131 // fold (conv (load x)) -> (load (conv*)x)
7132 // If the resultant load doesn't need a higher alignment than the original!
7133 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7134 // Do not change the width of a volatile load.
7135 !cast<LoadSDNode>(N0)->isVolatile() &&
7136 // Do not remove the cast if the types differ in endian layout.
7137 TLI.hasBigEndianPartOrdering(N0.getValueType(), DAG.getDataLayout()) ==
7138 TLI.hasBigEndianPartOrdering(VT, DAG.getDataLayout()) &&
7139 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)) &&
7140 TLI.isLoadBitCastBeneficial(N0.getValueType(), VT)) {
7141 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7142 unsigned Align = DAG.getDataLayout().getABITypeAlignment(
7143 VT.getTypeForEVT(*DAG.getContext()));
7144 unsigned OrigAlign = LN0->getAlignment();
7146 if (Align <= OrigAlign) {
7147 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(),
7148 LN0->getBasePtr(), LN0->getPointerInfo(),
7149 LN0->isVolatile(), LN0->isNonTemporal(),
7150 LN0->isInvariant(), OrigAlign,
7152 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
7157 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
7158 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
7159 // This often reduces constant pool loads.
7160 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) ||
7161 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) &&
7162 N0.getNode()->hasOneUse() && VT.isInteger() &&
7163 !VT.isVector() && !N0.getValueType().isVector()) {
7164 SDValue NewConv = DAG.getNode(ISD::BITCAST, SDLoc(N0), VT,
7166 AddToWorklist(NewConv.getNode());
7169 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
7170 if (N0.getOpcode() == ISD::FNEG)
7171 return DAG.getNode(ISD::XOR, DL, VT,
7172 NewConv, DAG.getConstant(SignBit, DL, VT));
7173 assert(N0.getOpcode() == ISD::FABS);
7174 return DAG.getNode(ISD::AND, DL, VT,
7175 NewConv, DAG.getConstant(~SignBit, DL, VT));
7178 // fold (bitconvert (fcopysign cst, x)) ->
7179 // (or (and (bitconvert x), sign), (and cst, (not sign)))
7180 // Note that we don't handle (copysign x, cst) because this can always be
7181 // folded to an fneg or fabs.
7182 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
7183 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
7184 VT.isInteger() && !VT.isVector()) {
7185 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
7186 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
7187 if (isTypeLegal(IntXVT)) {
7188 SDValue X = DAG.getNode(ISD::BITCAST, SDLoc(N0),
7189 IntXVT, N0.getOperand(1));
7190 AddToWorklist(X.getNode());
7192 // If X has a different width than the result/lhs, sext it or truncate it.
7193 unsigned VTWidth = VT.getSizeInBits();
7194 if (OrigXWidth < VTWidth) {
7195 X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X);
7196 AddToWorklist(X.getNode());
7197 } else if (OrigXWidth > VTWidth) {
7198 // To get the sign bit in the right place, we have to shift it right
7199 // before truncating.
7201 X = DAG.getNode(ISD::SRL, DL,
7202 X.getValueType(), X,
7203 DAG.getConstant(OrigXWidth-VTWidth, DL,
7205 AddToWorklist(X.getNode());
7206 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
7207 AddToWorklist(X.getNode());
7210 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
7211 X = DAG.getNode(ISD::AND, SDLoc(X), VT,
7212 X, DAG.getConstant(SignBit, SDLoc(X), VT));
7213 AddToWorklist(X.getNode());
7215 SDValue Cst = DAG.getNode(ISD::BITCAST, SDLoc(N0),
7216 VT, N0.getOperand(0));
7217 Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT,
7218 Cst, DAG.getConstant(~SignBit, SDLoc(Cst), VT));
7219 AddToWorklist(Cst.getNode());
7221 return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst);
7225 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
7226 if (N0.getOpcode() == ISD::BUILD_PAIR)
7227 if (SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT))
7230 // Remove double bitcasts from shuffles - this is often a legacy of
7231 // XformToShuffleWithZero being used to combine bitmaskings (of
7232 // float vectors bitcast to integer vectors) into shuffles.
7233 // bitcast(shuffle(bitcast(s0),bitcast(s1))) -> shuffle(s0,s1)
7234 if (Level < AfterLegalizeDAG && TLI.isTypeLegal(VT) && VT.isVector() &&
7235 N0->getOpcode() == ISD::VECTOR_SHUFFLE &&
7236 VT.getVectorNumElements() >= N0.getValueType().getVectorNumElements() &&
7237 !(VT.getVectorNumElements() % N0.getValueType().getVectorNumElements())) {
7238 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N0);
7240 // If operands are a bitcast, peek through if it casts the original VT.
7241 // If operands are a constant, just bitcast back to original VT.
7242 auto PeekThroughBitcast = [&](SDValue Op) {
7243 if (Op.getOpcode() == ISD::BITCAST &&
7244 Op.getOperand(0).getValueType() == VT)
7245 return SDValue(Op.getOperand(0));
7246 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) ||
7247 ISD::isBuildVectorOfConstantFPSDNodes(Op.getNode()))
7248 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
7252 SDValue SV0 = PeekThroughBitcast(N0->getOperand(0));
7253 SDValue SV1 = PeekThroughBitcast(N0->getOperand(1));
7258 VT.getVectorNumElements() / N0.getValueType().getVectorNumElements();
7259 SmallVector<int, 8> NewMask;
7260 for (int M : SVN->getMask())
7261 for (int i = 0; i != MaskScale; ++i)
7262 NewMask.push_back(M < 0 ? -1 : M * MaskScale + i);
7264 bool LegalMask = TLI.isShuffleMaskLegal(NewMask, VT);
7266 std::swap(SV0, SV1);
7267 ShuffleVectorSDNode::commuteMask(NewMask);
7268 LegalMask = TLI.isShuffleMaskLegal(NewMask, VT);
7272 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, NewMask);
7278 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
7279 EVT VT = N->getValueType(0);
7280 return CombineConsecutiveLoads(N, VT);
7283 /// We know that BV is a build_vector node with Constant, ConstantFP or Undef
7284 /// operands. DstEltVT indicates the destination element value type.
7285 SDValue DAGCombiner::
7286 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
7287 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
7289 // If this is already the right type, we're done.
7290 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
7292 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
7293 unsigned DstBitSize = DstEltVT.getSizeInBits();
7295 // If this is a conversion of N elements of one type to N elements of another
7296 // type, convert each element. This handles FP<->INT cases.
7297 if (SrcBitSize == DstBitSize) {
7298 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
7299 BV->getValueType(0).getVectorNumElements());
7301 // Due to the FP element handling below calling this routine recursively,
7302 // we can end up with a scalar-to-vector node here.
7303 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
7304 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
7305 DAG.getNode(ISD::BITCAST, SDLoc(BV),
7306 DstEltVT, BV->getOperand(0)));
7308 SmallVector<SDValue, 8> Ops;
7309 for (SDValue Op : BV->op_values()) {
7310 // If the vector element type is not legal, the BUILD_VECTOR operands
7311 // are promoted and implicitly truncated. Make that explicit here.
7312 if (Op.getValueType() != SrcEltVT)
7313 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op);
7314 Ops.push_back(DAG.getNode(ISD::BITCAST, SDLoc(BV),
7316 AddToWorklist(Ops.back().getNode());
7318 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
7321 // Otherwise, we're growing or shrinking the elements. To avoid having to
7322 // handle annoying details of growing/shrinking FP values, we convert them to
7324 if (SrcEltVT.isFloatingPoint()) {
7325 // Convert the input float vector to a int vector where the elements are the
7327 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
7328 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
7332 // Now we know the input is an integer vector. If the output is a FP type,
7333 // convert to integer first, then to FP of the right size.
7334 if (DstEltVT.isFloatingPoint()) {
7335 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
7336 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
7338 // Next, convert to FP elements of the same size.
7339 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
7344 // Okay, we know the src/dst types are both integers of differing types.
7345 // Handling growing first.
7346 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
7347 if (SrcBitSize < DstBitSize) {
7348 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
7350 SmallVector<SDValue, 8> Ops;
7351 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
7352 i += NumInputsPerOutput) {
7353 bool isLE = DAG.getDataLayout().isLittleEndian();
7354 APInt NewBits = APInt(DstBitSize, 0);
7355 bool EltIsUndef = true;
7356 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
7357 // Shift the previously computed bits over.
7358 NewBits <<= SrcBitSize;
7359 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
7360 if (Op.getOpcode() == ISD::UNDEF) continue;
7363 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
7364 zextOrTrunc(SrcBitSize).zext(DstBitSize);
7368 Ops.push_back(DAG.getUNDEF(DstEltVT));
7370 Ops.push_back(DAG.getConstant(NewBits, DL, DstEltVT));
7373 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
7374 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Ops);
7377 // Finally, this must be the case where we are shrinking elements: each input
7378 // turns into multiple outputs.
7379 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
7380 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
7381 NumOutputsPerInput*BV->getNumOperands());
7382 SmallVector<SDValue, 8> Ops;
7384 for (const SDValue &Op : BV->op_values()) {
7385 if (Op.getOpcode() == ISD::UNDEF) {
7386 Ops.append(NumOutputsPerInput, DAG.getUNDEF(DstEltVT));
7390 APInt OpVal = cast<ConstantSDNode>(Op)->
7391 getAPIntValue().zextOrTrunc(SrcBitSize);
7393 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
7394 APInt ThisVal = OpVal.trunc(DstBitSize);
7395 Ops.push_back(DAG.getConstant(ThisVal, DL, DstEltVT));
7396 OpVal = OpVal.lshr(DstBitSize);
7399 // For big endian targets, swap the order of the pieces of each element.
7400 if (DAG.getDataLayout().isBigEndian())
7401 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
7404 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Ops);
7407 /// Try to perform FMA combining on a given FADD node.
7408 SDValue DAGCombiner::visitFADDForFMACombine(SDNode *N) {
7409 SDValue N0 = N->getOperand(0);
7410 SDValue N1 = N->getOperand(1);
7411 EVT VT = N->getValueType(0);
7414 const TargetOptions &Options = DAG.getTarget().Options;
7415 bool UnsafeFPMath = (Options.AllowFPOpFusion == FPOpFusion::Fast ||
7416 Options.UnsafeFPMath);
7418 // Floating-point multiply-add with intermediate rounding.
7419 bool HasFMAD = (LegalOperations &&
7420 TLI.isOperationLegal(ISD::FMAD, VT));
7422 // Floating-point multiply-add without intermediate rounding.
7423 bool HasFMA = ((!LegalOperations ||
7424 TLI.isOperationLegalOrCustom(ISD::FMA, VT)) &&
7425 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
7428 // No valid opcode, do not combine.
7429 if (!HasFMAD && !HasFMA)
7432 // Always prefer FMAD to FMA for precision.
7433 unsigned int PreferredFusedOpcode = HasFMAD ? ISD::FMAD : ISD::FMA;
7434 bool Aggressive = TLI.enableAggressiveFMAFusion(VT);
7435 bool LookThroughFPExt = TLI.isFPExtFree(VT);
7437 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
7438 if (N0.getOpcode() == ISD::FMUL &&
7439 (Aggressive || N0->hasOneUse())) {
7440 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7441 N0.getOperand(0), N0.getOperand(1), N1);
7444 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
7445 // Note: Commutes FADD operands.
7446 if (N1.getOpcode() == ISD::FMUL &&
7447 (Aggressive || N1->hasOneUse())) {
7448 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7449 N1.getOperand(0), N1.getOperand(1), N0);
7452 // Look through FP_EXTEND nodes to do more combining.
7453 if (UnsafeFPMath && LookThroughFPExt) {
7454 // fold (fadd (fpext (fmul x, y)), z) -> (fma (fpext x), (fpext y), z)
7455 if (N0.getOpcode() == ISD::FP_EXTEND) {
7456 SDValue N00 = N0.getOperand(0);
7457 if (N00.getOpcode() == ISD::FMUL)
7458 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7459 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7461 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7462 N00.getOperand(1)), N1);
7465 // fold (fadd x, (fpext (fmul y, z))) -> (fma (fpext y), (fpext z), x)
7466 // Note: Commutes FADD operands.
7467 if (N1.getOpcode() == ISD::FP_EXTEND) {
7468 SDValue N10 = N1.getOperand(0);
7469 if (N10.getOpcode() == ISD::FMUL)
7470 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7471 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7473 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7474 N10.getOperand(1)), N0);
7478 // More folding opportunities when target permits.
7479 if ((UnsafeFPMath || HasFMAD) && Aggressive) {
7480 // fold (fadd (fma x, y, (fmul u, v)), z) -> (fma x, y (fma u, v, z))
7481 if (N0.getOpcode() == PreferredFusedOpcode &&
7482 N0.getOperand(2).getOpcode() == ISD::FMUL) {
7483 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7484 N0.getOperand(0), N0.getOperand(1),
7485 DAG.getNode(PreferredFusedOpcode, SL, VT,
7486 N0.getOperand(2).getOperand(0),
7487 N0.getOperand(2).getOperand(1),
7491 // fold (fadd x, (fma y, z, (fmul u, v)) -> (fma y, z (fma u, v, x))
7492 if (N1->getOpcode() == PreferredFusedOpcode &&
7493 N1.getOperand(2).getOpcode() == ISD::FMUL) {
7494 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7495 N1.getOperand(0), N1.getOperand(1),
7496 DAG.getNode(PreferredFusedOpcode, SL, VT,
7497 N1.getOperand(2).getOperand(0),
7498 N1.getOperand(2).getOperand(1),
7502 if (UnsafeFPMath && LookThroughFPExt) {
7503 // fold (fadd (fma x, y, (fpext (fmul u, v))), z)
7504 // -> (fma x, y, (fma (fpext u), (fpext v), z))
7505 auto FoldFAddFMAFPExtFMul = [&] (
7506 SDValue X, SDValue Y, SDValue U, SDValue V, SDValue Z) {
7507 return DAG.getNode(PreferredFusedOpcode, SL, VT, X, Y,
7508 DAG.getNode(PreferredFusedOpcode, SL, VT,
7509 DAG.getNode(ISD::FP_EXTEND, SL, VT, U),
7510 DAG.getNode(ISD::FP_EXTEND, SL, VT, V),
7513 if (N0.getOpcode() == PreferredFusedOpcode) {
7514 SDValue N02 = N0.getOperand(2);
7515 if (N02.getOpcode() == ISD::FP_EXTEND) {
7516 SDValue N020 = N02.getOperand(0);
7517 if (N020.getOpcode() == ISD::FMUL)
7518 return FoldFAddFMAFPExtFMul(N0.getOperand(0), N0.getOperand(1),
7519 N020.getOperand(0), N020.getOperand(1),
7524 // fold (fadd (fpext (fma x, y, (fmul u, v))), z)
7525 // -> (fma (fpext x), (fpext y), (fma (fpext u), (fpext v), z))
7526 // FIXME: This turns two single-precision and one double-precision
7527 // operation into two double-precision operations, which might not be
7528 // interesting for all targets, especially GPUs.
7529 auto FoldFAddFPExtFMAFMul = [&] (
7530 SDValue X, SDValue Y, SDValue U, SDValue V, SDValue Z) {
7531 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7532 DAG.getNode(ISD::FP_EXTEND, SL, VT, X),
7533 DAG.getNode(ISD::FP_EXTEND, SL, VT, Y),
7534 DAG.getNode(PreferredFusedOpcode, SL, VT,
7535 DAG.getNode(ISD::FP_EXTEND, SL, VT, U),
7536 DAG.getNode(ISD::FP_EXTEND, SL, VT, V),
7539 if (N0.getOpcode() == ISD::FP_EXTEND) {
7540 SDValue N00 = N0.getOperand(0);
7541 if (N00.getOpcode() == PreferredFusedOpcode) {
7542 SDValue N002 = N00.getOperand(2);
7543 if (N002.getOpcode() == ISD::FMUL)
7544 return FoldFAddFPExtFMAFMul(N00.getOperand(0), N00.getOperand(1),
7545 N002.getOperand(0), N002.getOperand(1),
7550 // fold (fadd x, (fma y, z, (fpext (fmul u, v)))
7551 // -> (fma y, z, (fma (fpext u), (fpext v), x))
7552 if (N1.getOpcode() == PreferredFusedOpcode) {
7553 SDValue N12 = N1.getOperand(2);
7554 if (N12.getOpcode() == ISD::FP_EXTEND) {
7555 SDValue N120 = N12.getOperand(0);
7556 if (N120.getOpcode() == ISD::FMUL)
7557 return FoldFAddFMAFPExtFMul(N1.getOperand(0), N1.getOperand(1),
7558 N120.getOperand(0), N120.getOperand(1),
7563 // fold (fadd x, (fpext (fma y, z, (fmul u, v)))
7564 // -> (fma (fpext y), (fpext z), (fma (fpext u), (fpext v), x))
7565 // FIXME: This turns two single-precision and one double-precision
7566 // operation into two double-precision operations, which might not be
7567 // interesting for all targets, especially GPUs.
7568 if (N1.getOpcode() == ISD::FP_EXTEND) {
7569 SDValue N10 = N1.getOperand(0);
7570 if (N10.getOpcode() == PreferredFusedOpcode) {
7571 SDValue N102 = N10.getOperand(2);
7572 if (N102.getOpcode() == ISD::FMUL)
7573 return FoldFAddFPExtFMAFMul(N10.getOperand(0), N10.getOperand(1),
7574 N102.getOperand(0), N102.getOperand(1),
7584 /// Try to perform FMA combining on a given FSUB node.
7585 SDValue DAGCombiner::visitFSUBForFMACombine(SDNode *N) {
7586 SDValue N0 = N->getOperand(0);
7587 SDValue N1 = N->getOperand(1);
7588 EVT VT = N->getValueType(0);
7591 const TargetOptions &Options = DAG.getTarget().Options;
7592 bool UnsafeFPMath = (Options.AllowFPOpFusion == FPOpFusion::Fast ||
7593 Options.UnsafeFPMath);
7595 // Floating-point multiply-add with intermediate rounding.
7596 bool HasFMAD = (LegalOperations &&
7597 TLI.isOperationLegal(ISD::FMAD, VT));
7599 // Floating-point multiply-add without intermediate rounding.
7600 bool HasFMA = ((!LegalOperations ||
7601 TLI.isOperationLegalOrCustom(ISD::FMA, VT)) &&
7602 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
7605 // No valid opcode, do not combine.
7606 if (!HasFMAD && !HasFMA)
7609 // Always prefer FMAD to FMA for precision.
7610 unsigned int PreferredFusedOpcode = HasFMAD ? ISD::FMAD : ISD::FMA;
7611 bool Aggressive = TLI.enableAggressiveFMAFusion(VT);
7612 bool LookThroughFPExt = TLI.isFPExtFree(VT);
7614 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
7615 if (N0.getOpcode() == ISD::FMUL &&
7616 (Aggressive || N0->hasOneUse())) {
7617 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7618 N0.getOperand(0), N0.getOperand(1),
7619 DAG.getNode(ISD::FNEG, SL, VT, N1));
7622 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
7623 // Note: Commutes FSUB operands.
7624 if (N1.getOpcode() == ISD::FMUL &&
7625 (Aggressive || N1->hasOneUse()))
7626 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7627 DAG.getNode(ISD::FNEG, SL, VT,
7629 N1.getOperand(1), N0);
7631 // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
7632 if (N0.getOpcode() == ISD::FNEG &&
7633 N0.getOperand(0).getOpcode() == ISD::FMUL &&
7634 (Aggressive || (N0->hasOneUse() && N0.getOperand(0).hasOneUse()))) {
7635 SDValue N00 = N0.getOperand(0).getOperand(0);
7636 SDValue N01 = N0.getOperand(0).getOperand(1);
7637 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7638 DAG.getNode(ISD::FNEG, SL, VT, N00), N01,
7639 DAG.getNode(ISD::FNEG, SL, VT, N1));
7642 // Look through FP_EXTEND nodes to do more combining.
7643 if (UnsafeFPMath && LookThroughFPExt) {
7644 // fold (fsub (fpext (fmul x, y)), z)
7645 // -> (fma (fpext x), (fpext y), (fneg z))
7646 if (N0.getOpcode() == ISD::FP_EXTEND) {
7647 SDValue N00 = N0.getOperand(0);
7648 if (N00.getOpcode() == ISD::FMUL)
7649 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7650 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7652 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7654 DAG.getNode(ISD::FNEG, SL, VT, N1));
7657 // fold (fsub x, (fpext (fmul y, z)))
7658 // -> (fma (fneg (fpext y)), (fpext z), x)
7659 // Note: Commutes FSUB operands.
7660 if (N1.getOpcode() == ISD::FP_EXTEND) {
7661 SDValue N10 = N1.getOperand(0);
7662 if (N10.getOpcode() == ISD::FMUL)
7663 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7664 DAG.getNode(ISD::FNEG, SL, VT,
7665 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7666 N10.getOperand(0))),
7667 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7672 // fold (fsub (fpext (fneg (fmul, x, y))), z)
7673 // -> (fneg (fma (fpext x), (fpext y), z))
7674 // Note: This could be removed with appropriate canonicalization of the
7675 // input expression into (fneg (fadd (fpext (fmul, x, y)), z). However, the
7676 // orthogonal flags -fp-contract=fast and -enable-unsafe-fp-math prevent
7677 // from implementing the canonicalization in visitFSUB.
7678 if (N0.getOpcode() == ISD::FP_EXTEND) {
7679 SDValue N00 = N0.getOperand(0);
7680 if (N00.getOpcode() == ISD::FNEG) {
7681 SDValue N000 = N00.getOperand(0);
7682 if (N000.getOpcode() == ISD::FMUL) {
7683 return DAG.getNode(ISD::FNEG, SL, VT,
7684 DAG.getNode(PreferredFusedOpcode, SL, VT,
7685 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7686 N000.getOperand(0)),
7687 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7688 N000.getOperand(1)),
7694 // fold (fsub (fneg (fpext (fmul, x, y))), z)
7695 // -> (fneg (fma (fpext x)), (fpext y), z)
7696 // Note: This could be removed with appropriate canonicalization of the
7697 // input expression into (fneg (fadd (fpext (fmul, x, y)), z). However, the
7698 // orthogonal flags -fp-contract=fast and -enable-unsafe-fp-math prevent
7699 // from implementing the canonicalization in visitFSUB.
7700 if (N0.getOpcode() == ISD::FNEG) {
7701 SDValue N00 = N0.getOperand(0);
7702 if (N00.getOpcode() == ISD::FP_EXTEND) {
7703 SDValue N000 = N00.getOperand(0);
7704 if (N000.getOpcode() == ISD::FMUL) {
7705 return DAG.getNode(ISD::FNEG, SL, VT,
7706 DAG.getNode(PreferredFusedOpcode, SL, VT,
7707 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7708 N000.getOperand(0)),
7709 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7710 N000.getOperand(1)),
7718 // More folding opportunities when target permits.
7719 if ((UnsafeFPMath || HasFMAD) && Aggressive) {
7720 // fold (fsub (fma x, y, (fmul u, v)), z)
7721 // -> (fma x, y (fma u, v, (fneg z)))
7722 if (N0.getOpcode() == PreferredFusedOpcode &&
7723 N0.getOperand(2).getOpcode() == ISD::FMUL) {
7724 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7725 N0.getOperand(0), N0.getOperand(1),
7726 DAG.getNode(PreferredFusedOpcode, SL, VT,
7727 N0.getOperand(2).getOperand(0),
7728 N0.getOperand(2).getOperand(1),
7729 DAG.getNode(ISD::FNEG, SL, VT,
7733 // fold (fsub x, (fma y, z, (fmul u, v)))
7734 // -> (fma (fneg y), z, (fma (fneg u), v, x))
7735 if (N1.getOpcode() == PreferredFusedOpcode &&
7736 N1.getOperand(2).getOpcode() == ISD::FMUL) {
7737 SDValue N20 = N1.getOperand(2).getOperand(0);
7738 SDValue N21 = N1.getOperand(2).getOperand(1);
7739 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7740 DAG.getNode(ISD::FNEG, SL, VT,
7743 DAG.getNode(PreferredFusedOpcode, SL, VT,
7744 DAG.getNode(ISD::FNEG, SL, VT, N20),
7749 if (UnsafeFPMath && LookThroughFPExt) {
7750 // fold (fsub (fma x, y, (fpext (fmul u, v))), z)
7751 // -> (fma x, y (fma (fpext u), (fpext v), (fneg z)))
7752 if (N0.getOpcode() == PreferredFusedOpcode) {
7753 SDValue N02 = N0.getOperand(2);
7754 if (N02.getOpcode() == ISD::FP_EXTEND) {
7755 SDValue N020 = N02.getOperand(0);
7756 if (N020.getOpcode() == ISD::FMUL)
7757 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7758 N0.getOperand(0), N0.getOperand(1),
7759 DAG.getNode(PreferredFusedOpcode, SL, VT,
7760 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7761 N020.getOperand(0)),
7762 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7763 N020.getOperand(1)),
7764 DAG.getNode(ISD::FNEG, SL, VT,
7769 // fold (fsub (fpext (fma x, y, (fmul u, v))), z)
7770 // -> (fma (fpext x), (fpext y),
7771 // (fma (fpext u), (fpext v), (fneg z)))
7772 // FIXME: This turns two single-precision and one double-precision
7773 // operation into two double-precision operations, which might not be
7774 // interesting for all targets, especially GPUs.
7775 if (N0.getOpcode() == ISD::FP_EXTEND) {
7776 SDValue N00 = N0.getOperand(0);
7777 if (N00.getOpcode() == PreferredFusedOpcode) {
7778 SDValue N002 = N00.getOperand(2);
7779 if (N002.getOpcode() == ISD::FMUL)
7780 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7781 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7783 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7785 DAG.getNode(PreferredFusedOpcode, SL, VT,
7786 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7787 N002.getOperand(0)),
7788 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7789 N002.getOperand(1)),
7790 DAG.getNode(ISD::FNEG, SL, VT,
7795 // fold (fsub x, (fma y, z, (fpext (fmul u, v))))
7796 // -> (fma (fneg y), z, (fma (fneg (fpext u)), (fpext v), x))
7797 if (N1.getOpcode() == PreferredFusedOpcode &&
7798 N1.getOperand(2).getOpcode() == ISD::FP_EXTEND) {
7799 SDValue N120 = N1.getOperand(2).getOperand(0);
7800 if (N120.getOpcode() == ISD::FMUL) {
7801 SDValue N1200 = N120.getOperand(0);
7802 SDValue N1201 = N120.getOperand(1);
7803 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7804 DAG.getNode(ISD::FNEG, SL, VT, N1.getOperand(0)),
7806 DAG.getNode(PreferredFusedOpcode, SL, VT,
7807 DAG.getNode(ISD::FNEG, SL, VT,
7808 DAG.getNode(ISD::FP_EXTEND, SL,
7810 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7816 // fold (fsub x, (fpext (fma y, z, (fmul u, v))))
7817 // -> (fma (fneg (fpext y)), (fpext z),
7818 // (fma (fneg (fpext u)), (fpext v), x))
7819 // FIXME: This turns two single-precision and one double-precision
7820 // operation into two double-precision operations, which might not be
7821 // interesting for all targets, especially GPUs.
7822 if (N1.getOpcode() == ISD::FP_EXTEND &&
7823 N1.getOperand(0).getOpcode() == PreferredFusedOpcode) {
7824 SDValue N100 = N1.getOperand(0).getOperand(0);
7825 SDValue N101 = N1.getOperand(0).getOperand(1);
7826 SDValue N102 = N1.getOperand(0).getOperand(2);
7827 if (N102.getOpcode() == ISD::FMUL) {
7828 SDValue N1020 = N102.getOperand(0);
7829 SDValue N1021 = N102.getOperand(1);
7830 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7831 DAG.getNode(ISD::FNEG, SL, VT,
7832 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7834 DAG.getNode(ISD::FP_EXTEND, SL, VT, N101),
7835 DAG.getNode(PreferredFusedOpcode, SL, VT,
7836 DAG.getNode(ISD::FNEG, SL, VT,
7837 DAG.getNode(ISD::FP_EXTEND, SL,
7839 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7850 SDValue DAGCombiner::visitFADD(SDNode *N) {
7851 SDValue N0 = N->getOperand(0);
7852 SDValue N1 = N->getOperand(1);
7853 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7854 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7855 EVT VT = N->getValueType(0);
7857 const TargetOptions &Options = DAG.getTarget().Options;
7861 if (SDValue FoldedVOp = SimplifyVBinOp(N))
7864 // fold (fadd c1, c2) -> c1 + c2
7866 return DAG.getNode(ISD::FADD, DL, VT, N0, N1);
7868 // canonicalize constant to RHS
7869 if (N0CFP && !N1CFP)
7870 return DAG.getNode(ISD::FADD, DL, VT, N1, N0);
7872 // fold (fadd A, (fneg B)) -> (fsub A, B)
7873 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
7874 isNegatibleForFree(N1, LegalOperations, TLI, &Options) == 2)
7875 return DAG.getNode(ISD::FSUB, DL, VT, N0,
7876 GetNegatedExpression(N1, DAG, LegalOperations));
7878 // fold (fadd (fneg A), B) -> (fsub B, A)
7879 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
7880 isNegatibleForFree(N0, LegalOperations, TLI, &Options) == 2)
7881 return DAG.getNode(ISD::FSUB, DL, VT, N1,
7882 GetNegatedExpression(N0, DAG, LegalOperations));
7884 // If 'unsafe math' is enabled, fold lots of things.
7885 if (Options.UnsafeFPMath) {
7886 // No FP constant should be created after legalization as Instruction
7887 // Selection pass has a hard time dealing with FP constants.
7888 bool AllowNewConst = (Level < AfterLegalizeDAG);
7890 // fold (fadd A, 0) -> A
7891 if (N1CFP && N1CFP->isZero())
7894 // fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
7895 if (N1CFP && N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
7896 isa<ConstantFPSDNode>(N0.getOperand(1)))
7897 return DAG.getNode(ISD::FADD, DL, VT, N0.getOperand(0),
7898 DAG.getNode(ISD::FADD, DL, VT, N0.getOperand(1), N1));
7900 // If allowed, fold (fadd (fneg x), x) -> 0.0
7901 if (AllowNewConst && N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1)
7902 return DAG.getConstantFP(0.0, DL, VT);
7904 // If allowed, fold (fadd x, (fneg x)) -> 0.0
7905 if (AllowNewConst && N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0)
7906 return DAG.getConstantFP(0.0, DL, VT);
7908 // We can fold chains of FADD's of the same value into multiplications.
7909 // This transform is not safe in general because we are reducing the number
7910 // of rounding steps.
7911 if (TLI.isOperationLegalOrCustom(ISD::FMUL, VT) && !N0CFP && !N1CFP) {
7912 if (N0.getOpcode() == ISD::FMUL) {
7913 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
7914 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
7916 // (fadd (fmul x, c), x) -> (fmul x, c+1)
7917 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
7918 SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, SDValue(CFP01, 0),
7919 DAG.getConstantFP(1.0, DL, VT));
7920 return DAG.getNode(ISD::FMUL, DL, VT, N1, NewCFP);
7923 // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2)
7924 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
7925 N1.getOperand(0) == N1.getOperand(1) &&
7926 N0.getOperand(0) == N1.getOperand(0)) {
7927 SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, SDValue(CFP01, 0),
7928 DAG.getConstantFP(2.0, DL, VT));
7929 return DAG.getNode(ISD::FMUL, DL, VT, N0.getOperand(0), NewCFP);
7933 if (N1.getOpcode() == ISD::FMUL) {
7934 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
7935 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
7937 // (fadd x, (fmul x, c)) -> (fmul x, c+1)
7938 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
7939 SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, SDValue(CFP11, 0),
7940 DAG.getConstantFP(1.0, DL, VT));
7941 return DAG.getNode(ISD::FMUL, DL, VT, N0, NewCFP);
7944 // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2)
7945 if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD &&
7946 N0.getOperand(0) == N0.getOperand(1) &&
7947 N1.getOperand(0) == N0.getOperand(0)) {
7948 SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, SDValue(CFP11, 0),
7949 DAG.getConstantFP(2.0, DL, VT));
7950 return DAG.getNode(ISD::FMUL, DL, VT, N1.getOperand(0), NewCFP);
7954 if (N0.getOpcode() == ISD::FADD && AllowNewConst) {
7955 ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
7956 // (fadd (fadd x, x), x) -> (fmul x, 3.0)
7957 if (!CFP && N0.getOperand(0) == N0.getOperand(1) &&
7958 (N0.getOperand(0) == N1)) {
7959 return DAG.getNode(ISD::FMUL, DL, VT,
7960 N1, DAG.getConstantFP(3.0, DL, VT));
7964 if (N1.getOpcode() == ISD::FADD && AllowNewConst) {
7965 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
7966 // (fadd x, (fadd x, x)) -> (fmul x, 3.0)
7967 if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) &&
7968 N1.getOperand(0) == N0) {
7969 return DAG.getNode(ISD::FMUL, DL, VT,
7970 N0, DAG.getConstantFP(3.0, DL, VT));
7974 // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0)
7975 if (AllowNewConst &&
7976 N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
7977 N0.getOperand(0) == N0.getOperand(1) &&
7978 N1.getOperand(0) == N1.getOperand(1) &&
7979 N0.getOperand(0) == N1.getOperand(0)) {
7980 return DAG.getNode(ISD::FMUL, DL, VT,
7981 N0.getOperand(0), DAG.getConstantFP(4.0, DL, VT));
7984 } // enable-unsafe-fp-math
7986 // FADD -> FMA combines:
7987 SDValue Fused = visitFADDForFMACombine(N);
7989 AddToWorklist(Fused.getNode());
7996 SDValue DAGCombiner::visitFSUB(SDNode *N) {
7997 SDValue N0 = N->getOperand(0);
7998 SDValue N1 = N->getOperand(1);
7999 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
8000 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
8001 EVT VT = N->getValueType(0);
8003 const TargetOptions &Options = DAG.getTarget().Options;
8007 if (SDValue FoldedVOp = SimplifyVBinOp(N))
8010 // fold (fsub c1, c2) -> c1-c2
8012 return DAG.getNode(ISD::FSUB, dl, VT, N0, N1);
8014 // fold (fsub A, (fneg B)) -> (fadd A, B)
8015 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
8016 return DAG.getNode(ISD::FADD, dl, VT, N0,
8017 GetNegatedExpression(N1, DAG, LegalOperations));
8019 // If 'unsafe math' is enabled, fold lots of things.
8020 if (Options.UnsafeFPMath) {
8022 if (N1CFP && N1CFP->isZero())
8025 // (fsub 0, B) -> -B
8026 if (N0CFP && N0CFP->isZero()) {
8027 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
8028 return GetNegatedExpression(N1, DAG, LegalOperations);
8029 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
8030 return DAG.getNode(ISD::FNEG, dl, VT, N1);
8033 // (fsub x, x) -> 0.0
8035 return DAG.getConstantFP(0.0f, dl, VT);
8037 // (fsub x, (fadd x, y)) -> (fneg y)
8038 // (fsub x, (fadd y, x)) -> (fneg y)
8039 if (N1.getOpcode() == ISD::FADD) {
8040 SDValue N10 = N1->getOperand(0);
8041 SDValue N11 = N1->getOperand(1);
8043 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI, &Options))
8044 return GetNegatedExpression(N11, DAG, LegalOperations);
8046 if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI, &Options))
8047 return GetNegatedExpression(N10, DAG, LegalOperations);
8051 // FSUB -> FMA combines:
8052 SDValue Fused = visitFSUBForFMACombine(N);
8054 AddToWorklist(Fused.getNode());
8061 SDValue DAGCombiner::visitFMUL(SDNode *N) {
8062 SDValue N0 = N->getOperand(0);
8063 SDValue N1 = N->getOperand(1);
8064 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
8065 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
8066 EVT VT = N->getValueType(0);
8068 const TargetOptions &Options = DAG.getTarget().Options;
8071 if (VT.isVector()) {
8072 // This just handles C1 * C2 for vectors. Other vector folds are below.
8073 if (SDValue FoldedVOp = SimplifyVBinOp(N))
8077 // fold (fmul c1, c2) -> c1*c2
8079 return DAG.getNode(ISD::FMUL, DL, VT, N0, N1);
8081 // canonicalize constant to RHS
8082 if (isConstantFPBuildVectorOrConstantFP(N0) &&
8083 !isConstantFPBuildVectorOrConstantFP(N1))
8084 return DAG.getNode(ISD::FMUL, DL, VT, N1, N0);
8086 // fold (fmul A, 1.0) -> A
8087 if (N1CFP && N1CFP->isExactlyValue(1.0))
8090 if (Options.UnsafeFPMath) {
8091 // fold (fmul A, 0) -> 0
8092 if (N1CFP && N1CFP->isZero())
8095 // fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
8096 if (N0.getOpcode() == ISD::FMUL) {
8097 // Fold scalars or any vector constants (not just splats).
8098 // This fold is done in general by InstCombine, but extra fmul insts
8099 // may have been generated during lowering.
8100 SDValue N00 = N0.getOperand(0);
8101 SDValue N01 = N0.getOperand(1);
8102 auto *BV1 = dyn_cast<BuildVectorSDNode>(N1);
8103 auto *BV00 = dyn_cast<BuildVectorSDNode>(N00);
8104 auto *BV01 = dyn_cast<BuildVectorSDNode>(N01);
8106 // Check 1: Make sure that the first operand of the inner multiply is NOT
8107 // a constant. Otherwise, we may induce infinite looping.
8108 if (!(isConstOrConstSplatFP(N00) || (BV00 && BV00->isConstant()))) {
8109 // Check 2: Make sure that the second operand of the inner multiply and
8110 // the second operand of the outer multiply are constants.
8111 if ((N1CFP && isConstOrConstSplatFP(N01)) ||
8112 (BV1 && BV01 && BV1->isConstant() && BV01->isConstant())) {
8113 SDValue MulConsts = DAG.getNode(ISD::FMUL, DL, VT, N01, N1);
8114 return DAG.getNode(ISD::FMUL, DL, VT, N00, MulConsts);
8119 // fold (fmul (fadd x, x), c) -> (fmul x, (fmul 2.0, c))
8120 // Undo the fmul 2.0, x -> fadd x, x transformation, since if it occurs
8121 // during an early run of DAGCombiner can prevent folding with fmuls
8122 // inserted during lowering.
8123 if (N0.getOpcode() == ISD::FADD &&
8124 (N0.getOperand(0) == N0.getOperand(1)) &&
8126 const SDValue Two = DAG.getConstantFP(2.0, DL, VT);
8127 SDValue MulConsts = DAG.getNode(ISD::FMUL, DL, VT, Two, N1);
8128 return DAG.getNode(ISD::FMUL, DL, VT, N0.getOperand(0), MulConsts);
8132 // fold (fmul X, 2.0) -> (fadd X, X)
8133 if (N1CFP && N1CFP->isExactlyValue(+2.0))
8134 return DAG.getNode(ISD::FADD, DL, VT, N0, N0);
8136 // fold (fmul X, -1.0) -> (fneg X)
8137 if (N1CFP && N1CFP->isExactlyValue(-1.0))
8138 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
8139 return DAG.getNode(ISD::FNEG, DL, VT, N0);
8141 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
8142 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
8143 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
8144 // Both can be negated for free, check to see if at least one is cheaper
8146 if (LHSNeg == 2 || RHSNeg == 2)
8147 return DAG.getNode(ISD::FMUL, DL, VT,
8148 GetNegatedExpression(N0, DAG, LegalOperations),
8149 GetNegatedExpression(N1, DAG, LegalOperations));
8156 SDValue DAGCombiner::visitFMA(SDNode *N) {
8157 SDValue N0 = N->getOperand(0);
8158 SDValue N1 = N->getOperand(1);
8159 SDValue N2 = N->getOperand(2);
8160 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8161 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8162 EVT VT = N->getValueType(0);
8164 const TargetOptions &Options = DAG.getTarget().Options;
8166 // Constant fold FMA.
8167 if (isa<ConstantFPSDNode>(N0) &&
8168 isa<ConstantFPSDNode>(N1) &&
8169 isa<ConstantFPSDNode>(N2)) {
8170 return DAG.getNode(ISD::FMA, dl, VT, N0, N1, N2);
8173 if (Options.UnsafeFPMath) {
8174 if (N0CFP && N0CFP->isZero())
8176 if (N1CFP && N1CFP->isZero())
8179 if (N0CFP && N0CFP->isExactlyValue(1.0))
8180 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2);
8181 if (N1CFP && N1CFP->isExactlyValue(1.0))
8182 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2);
8184 // Canonicalize (fma c, x, y) -> (fma x, c, y)
8185 if (N0CFP && !N1CFP)
8186 return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2);
8188 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
8189 if (Options.UnsafeFPMath && N1CFP &&
8190 N2.getOpcode() == ISD::FMUL &&
8191 N0 == N2.getOperand(0) &&
8192 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
8193 return DAG.getNode(ISD::FMUL, dl, VT, N0,
8194 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
8198 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
8199 if (Options.UnsafeFPMath &&
8200 N0.getOpcode() == ISD::FMUL && N1CFP &&
8201 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
8202 return DAG.getNode(ISD::FMA, dl, VT,
8204 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
8208 // (fma x, 1, y) -> (fadd x, y)
8209 // (fma x, -1, y) -> (fadd (fneg x), y)
8211 if (N1CFP->isExactlyValue(1.0))
8212 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
8214 if (N1CFP->isExactlyValue(-1.0) &&
8215 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
8216 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
8217 AddToWorklist(RHSNeg.getNode());
8218 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
8222 // (fma x, c, x) -> (fmul x, (c+1))
8223 if (Options.UnsafeFPMath && N1CFP && N0 == N2)
8224 return DAG.getNode(ISD::FMUL, dl, VT, N0,
8225 DAG.getNode(ISD::FADD, dl, VT,
8226 N1, DAG.getConstantFP(1.0, dl, VT)));
8228 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
8229 if (Options.UnsafeFPMath && N1CFP &&
8230 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0)
8231 return DAG.getNode(ISD::FMUL, dl, VT, N0,
8232 DAG.getNode(ISD::FADD, dl, VT,
8233 N1, DAG.getConstantFP(-1.0, dl, VT)));
8239 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
8241 // E.g., (a / D; b / D;) -> (recip = 1.0 / D; a * recip; b * recip)
8242 // Notice that this is not always beneficial. One reason is different target
8243 // may have different costs for FDIV and FMUL, so sometimes the cost of two
8244 // FDIVs may be lower than the cost of one FDIV and two FMULs. Another reason
8245 // is the critical path is increased from "one FDIV" to "one FDIV + one FMUL".
8246 SDValue DAGCombiner::combineRepeatedFPDivisors(SDNode *N) {
8247 if (!DAG.getTarget().Options.UnsafeFPMath)
8250 SDValue N0 = N->getOperand(0);
8251 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8253 // Skip if current node is a reciprocal.
8254 if (N0CFP && N0CFP->isExactlyValue(1.0))
8257 SDValue N1 = N->getOperand(1);
8258 SmallVector<SDNode *, 4> Users;
8260 // Find all FDIV users of the same divisor.
8261 for (auto *U : N1->uses()) {
8262 if (U->getOpcode() == ISD::FDIV && U->getOperand(1) == N1)
8266 if (!TLI.combineRepeatedFPDivisors(Users.size()))
8269 EVT VT = N->getValueType(0);
8271 SDValue FPOne = DAG.getConstantFP(1.0, DL, VT);
8272 // FIXME: This optimization requires some level of fast-math, so the
8273 // created reciprocal node should at least have the 'allowReciprocal'
8274 // fast-math-flag set.
8275 SDValue Reciprocal = DAG.getNode(ISD::FDIV, DL, VT, FPOne, N1);
8277 // Dividend / Divisor -> Dividend * Reciprocal
8278 for (auto *U : Users) {
8279 SDValue Dividend = U->getOperand(0);
8280 if (Dividend != FPOne) {
8281 SDValue NewNode = DAG.getNode(ISD::FMUL, SDLoc(U), VT, Dividend,
8283 CombineTo(U, NewNode);
8284 } else if (U != Reciprocal.getNode()) {
8285 // In the absence of fast-math-flags, this user node is always the
8286 // same node as Reciprocal, but with FMF they may be different nodes.
8287 CombineTo(U, Reciprocal);
8290 return SDValue(N, 0); // N was replaced.
8293 SDValue DAGCombiner::visitFDIV(SDNode *N) {
8294 SDValue N0 = N->getOperand(0);
8295 SDValue N1 = N->getOperand(1);
8296 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8297 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8298 EVT VT = N->getValueType(0);
8300 const TargetOptions &Options = DAG.getTarget().Options;
8304 if (SDValue FoldedVOp = SimplifyVBinOp(N))
8307 // fold (fdiv c1, c2) -> c1/c2
8309 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1);
8311 if (Options.UnsafeFPMath) {
8312 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
8314 // Compute the reciprocal 1.0 / c2.
8315 APFloat N1APF = N1CFP->getValueAPF();
8316 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
8317 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
8318 // Only do the transform if the reciprocal is a legal fp immediate that
8319 // isn't too nasty (eg NaN, denormal, ...).
8320 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
8321 (!LegalOperations ||
8322 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
8323 // backend)... we should handle this gracefully after Legalize.
8324 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
8325 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
8326 TLI.isFPImmLegal(Recip, VT)))
8327 return DAG.getNode(ISD::FMUL, DL, VT, N0,
8328 DAG.getConstantFP(Recip, DL, VT));
8331 // If this FDIV is part of a reciprocal square root, it may be folded
8332 // into a target-specific square root estimate instruction.
8333 if (N1.getOpcode() == ISD::FSQRT) {
8334 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0))) {
8335 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
8337 } else if (N1.getOpcode() == ISD::FP_EXTEND &&
8338 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
8339 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0).getOperand(0))) {
8340 RV = DAG.getNode(ISD::FP_EXTEND, SDLoc(N1), VT, RV);
8341 AddToWorklist(RV.getNode());
8342 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
8344 } else if (N1.getOpcode() == ISD::FP_ROUND &&
8345 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
8346 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0).getOperand(0))) {
8347 RV = DAG.getNode(ISD::FP_ROUND, SDLoc(N1), VT, RV, N1.getOperand(1));
8348 AddToWorklist(RV.getNode());
8349 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
8351 } else if (N1.getOpcode() == ISD::FMUL) {
8352 // Look through an FMUL. Even though this won't remove the FDIV directly,
8353 // it's still worthwhile to get rid of the FSQRT if possible.
8356 if (N1.getOperand(0).getOpcode() == ISD::FSQRT) {
8357 SqrtOp = N1.getOperand(0);
8358 OtherOp = N1.getOperand(1);
8359 } else if (N1.getOperand(1).getOpcode() == ISD::FSQRT) {
8360 SqrtOp = N1.getOperand(1);
8361 OtherOp = N1.getOperand(0);
8363 if (SqrtOp.getNode()) {
8364 // We found a FSQRT, so try to make this fold:
8365 // x / (y * sqrt(z)) -> x * (rsqrt(z) / y)
8366 if (SDValue RV = BuildRsqrtEstimate(SqrtOp.getOperand(0))) {
8367 RV = DAG.getNode(ISD::FDIV, SDLoc(N1), VT, RV, OtherOp);
8368 AddToWorklist(RV.getNode());
8369 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
8374 // Fold into a reciprocal estimate and multiply instead of a real divide.
8375 if (SDValue RV = BuildReciprocalEstimate(N1)) {
8376 AddToWorklist(RV.getNode());
8377 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
8381 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
8382 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
8383 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
8384 // Both can be negated for free, check to see if at least one is cheaper
8386 if (LHSNeg == 2 || RHSNeg == 2)
8387 return DAG.getNode(ISD::FDIV, SDLoc(N), VT,
8388 GetNegatedExpression(N0, DAG, LegalOperations),
8389 GetNegatedExpression(N1, DAG, LegalOperations));
8393 if (SDValue CombineRepeatedDivisors = combineRepeatedFPDivisors(N))
8394 return CombineRepeatedDivisors;
8399 SDValue DAGCombiner::visitFREM(SDNode *N) {
8400 SDValue N0 = N->getOperand(0);
8401 SDValue N1 = N->getOperand(1);
8402 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8403 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8404 EVT VT = N->getValueType(0);
8406 // fold (frem c1, c2) -> fmod(c1,c2)
8408 return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1);
8413 SDValue DAGCombiner::visitFSQRT(SDNode *N) {
8414 if (!DAG.getTarget().Options.UnsafeFPMath || TLI.isFsqrtCheap())
8417 // Compute this as X * (1/sqrt(X)) = X * (X ** -0.5)
8418 SDValue RV = BuildRsqrtEstimate(N->getOperand(0));
8422 EVT VT = RV.getValueType();
8424 RV = DAG.getNode(ISD::FMUL, DL, VT, N->getOperand(0), RV);
8425 AddToWorklist(RV.getNode());
8427 // Unfortunately, RV is now NaN if the input was exactly 0.
8428 // Select out this case and force the answer to 0.
8429 SDValue Zero = DAG.getConstantFP(0.0, DL, VT);
8430 EVT CCVT = TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
8431 SDValue ZeroCmp = DAG.getSetCC(DL, CCVT, N->getOperand(0), Zero, ISD::SETEQ);
8432 AddToWorklist(ZeroCmp.getNode());
8433 AddToWorklist(RV.getNode());
8435 return DAG.getNode(VT.isVector() ? ISD::VSELECT : ISD::SELECT, DL, VT,
8439 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
8440 SDValue N0 = N->getOperand(0);
8441 SDValue N1 = N->getOperand(1);
8442 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8443 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8444 EVT VT = N->getValueType(0);
8446 if (N0CFP && N1CFP) // Constant fold
8447 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1);
8450 const APFloat& V = N1CFP->getValueAPF();
8451 // copysign(x, c1) -> fabs(x) iff ispos(c1)
8452 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
8453 if (!V.isNegative()) {
8454 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
8455 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
8457 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
8458 return DAG.getNode(ISD::FNEG, SDLoc(N), VT,
8459 DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0));
8463 // copysign(fabs(x), y) -> copysign(x, y)
8464 // copysign(fneg(x), y) -> copysign(x, y)
8465 // copysign(copysign(x,z), y) -> copysign(x, y)
8466 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
8467 N0.getOpcode() == ISD::FCOPYSIGN)
8468 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
8469 N0.getOperand(0), N1);
8471 // copysign(x, abs(y)) -> abs(x)
8472 if (N1.getOpcode() == ISD::FABS)
8473 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
8475 // copysign(x, copysign(y,z)) -> copysign(x, z)
8476 if (N1.getOpcode() == ISD::FCOPYSIGN)
8477 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
8478 N0, N1.getOperand(1));
8480 // copysign(x, fp_extend(y)) -> copysign(x, y)
8481 // copysign(x, fp_round(y)) -> copysign(x, y)
8482 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
8483 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
8484 N0, N1.getOperand(0));
8489 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
8490 SDValue N0 = N->getOperand(0);
8491 EVT VT = N->getValueType(0);
8492 EVT OpVT = N0.getValueType();
8494 // fold (sint_to_fp c1) -> c1fp
8495 if (isConstantIntBuildVectorOrConstantInt(N0) &&
8496 // ...but only if the target supports immediate floating-point values
8497 (!LegalOperations ||
8498 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
8499 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
8501 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
8502 // but UINT_TO_FP is legal on this target, try to convert.
8503 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
8504 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
8505 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
8506 if (DAG.SignBitIsZero(N0))
8507 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
8510 // The next optimizations are desirable only if SELECT_CC can be lowered.
8511 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
8512 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
8513 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
8515 (!LegalOperations ||
8516 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
8519 { N0.getOperand(0), N0.getOperand(1),
8520 DAG.getConstantFP(-1.0, DL, VT), DAG.getConstantFP(0.0, DL, VT),
8522 return DAG.getNode(ISD::SELECT_CC, DL, VT, Ops);
8525 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
8526 // (select_cc x, y, 1.0, 0.0,, cc)
8527 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
8528 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
8529 (!LegalOperations ||
8530 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
8533 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
8534 DAG.getConstantFP(1.0, DL, VT), DAG.getConstantFP(0.0, DL, VT),
8535 N0.getOperand(0).getOperand(2) };
8536 return DAG.getNode(ISD::SELECT_CC, DL, VT, Ops);
8543 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
8544 SDValue N0 = N->getOperand(0);
8545 EVT VT = N->getValueType(0);
8546 EVT OpVT = N0.getValueType();
8548 // fold (uint_to_fp c1) -> c1fp
8549 if (isConstantIntBuildVectorOrConstantInt(N0) &&
8550 // ...but only if the target supports immediate floating-point values
8551 (!LegalOperations ||
8552 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
8553 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
8555 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
8556 // but SINT_TO_FP is legal on this target, try to convert.
8557 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
8558 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
8559 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
8560 if (DAG.SignBitIsZero(N0))
8561 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
8564 // The next optimizations are desirable only if SELECT_CC can be lowered.
8565 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
8566 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
8568 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
8569 (!LegalOperations ||
8570 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
8573 { N0.getOperand(0), N0.getOperand(1),
8574 DAG.getConstantFP(1.0, DL, VT), DAG.getConstantFP(0.0, DL, VT),
8576 return DAG.getNode(ISD::SELECT_CC, DL, VT, Ops);
8583 // Fold (fp_to_{s/u}int ({s/u}int_to_fpx)) -> zext x, sext x, trunc x, or x
8584 static SDValue FoldIntToFPToInt(SDNode *N, SelectionDAG &DAG) {
8585 SDValue N0 = N->getOperand(0);
8586 EVT VT = N->getValueType(0);
8588 if (N0.getOpcode() != ISD::UINT_TO_FP && N0.getOpcode() != ISD::SINT_TO_FP)
8591 SDValue Src = N0.getOperand(0);
8592 EVT SrcVT = Src.getValueType();
8593 bool IsInputSigned = N0.getOpcode() == ISD::SINT_TO_FP;
8594 bool IsOutputSigned = N->getOpcode() == ISD::FP_TO_SINT;
8596 // We can safely assume the conversion won't overflow the output range,
8597 // because (for example) (uint8_t)18293.f is undefined behavior.
8599 // Since we can assume the conversion won't overflow, our decision as to
8600 // whether the input will fit in the float should depend on the minimum
8601 // of the input range and output range.
8603 // This means this is also safe for a signed input and unsigned output, since
8604 // a negative input would lead to undefined behavior.
8605 unsigned InputSize = (int)SrcVT.getScalarSizeInBits() - IsInputSigned;
8606 unsigned OutputSize = (int)VT.getScalarSizeInBits() - IsOutputSigned;
8607 unsigned ActualSize = std::min(InputSize, OutputSize);
8608 const fltSemantics &sem = DAG.EVTToAPFloatSemantics(N0.getValueType());
8610 // We can only fold away the float conversion if the input range can be
8611 // represented exactly in the float range.
8612 if (APFloat::semanticsPrecision(sem) >= ActualSize) {
8613 if (VT.getScalarSizeInBits() > SrcVT.getScalarSizeInBits()) {
8614 unsigned ExtOp = IsInputSigned && IsOutputSigned ? ISD::SIGN_EXTEND
8616 return DAG.getNode(ExtOp, SDLoc(N), VT, Src);
8618 if (VT.getScalarSizeInBits() < SrcVT.getScalarSizeInBits())
8619 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Src);
8622 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Src);
8627 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
8628 SDValue N0 = N->getOperand(0);
8629 EVT VT = N->getValueType(0);
8631 // fold (fp_to_sint c1fp) -> c1
8632 if (isConstantFPBuildVectorOrConstantFP(N0))
8633 return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0);
8635 return FoldIntToFPToInt(N, DAG);
8638 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
8639 SDValue N0 = N->getOperand(0);
8640 EVT VT = N->getValueType(0);
8642 // fold (fp_to_uint c1fp) -> c1
8643 if (isConstantFPBuildVectorOrConstantFP(N0))
8644 return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0);
8646 return FoldIntToFPToInt(N, DAG);
8649 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
8650 SDValue N0 = N->getOperand(0);
8651 SDValue N1 = N->getOperand(1);
8652 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8653 EVT VT = N->getValueType(0);
8655 // fold (fp_round c1fp) -> c1fp
8657 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1);
8659 // fold (fp_round (fp_extend x)) -> x
8660 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
8661 return N0.getOperand(0);
8663 // fold (fp_round (fp_round x)) -> (fp_round x)
8664 if (N0.getOpcode() == ISD::FP_ROUND) {
8665 const bool NIsTrunc = N->getConstantOperandVal(1) == 1;
8666 const bool N0IsTrunc = N0.getNode()->getConstantOperandVal(1) == 1;
8667 // If the first fp_round isn't a value preserving truncation, it might
8668 // introduce a tie in the second fp_round, that wouldn't occur in the
8669 // single-step fp_round we want to fold to.
8670 // In other words, double rounding isn't the same as rounding.
8671 // Also, this is a value preserving truncation iff both fp_round's are.
8672 if (DAG.getTarget().Options.UnsafeFPMath || N0IsTrunc) {
8674 return DAG.getNode(ISD::FP_ROUND, DL, VT, N0.getOperand(0),
8675 DAG.getIntPtrConstant(NIsTrunc && N0IsTrunc, DL));
8679 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
8680 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
8681 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT,
8682 N0.getOperand(0), N1);
8683 AddToWorklist(Tmp.getNode());
8684 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
8685 Tmp, N0.getOperand(1));
8691 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
8692 SDValue N0 = N->getOperand(0);
8693 EVT VT = N->getValueType(0);
8694 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
8695 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8697 // fold (fp_round_inreg c1fp) -> c1fp
8698 if (N0CFP && isTypeLegal(EVT)) {
8700 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), DL, EVT);
8701 return DAG.getNode(ISD::FP_EXTEND, DL, VT, Round);
8707 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
8708 SDValue N0 = N->getOperand(0);
8709 EVT VT = N->getValueType(0);
8711 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
8712 if (N->hasOneUse() &&
8713 N->use_begin()->getOpcode() == ISD::FP_ROUND)
8716 // fold (fp_extend c1fp) -> c1fp
8717 if (isConstantFPBuildVectorOrConstantFP(N0))
8718 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0);
8720 // fold (fp_extend (fp16_to_fp op)) -> (fp16_to_fp op)
8721 if (N0.getOpcode() == ISD::FP16_TO_FP &&
8722 TLI.getOperationAction(ISD::FP16_TO_FP, VT) == TargetLowering::Legal)
8723 return DAG.getNode(ISD::FP16_TO_FP, SDLoc(N), VT, N0.getOperand(0));
8725 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
8727 if (N0.getOpcode() == ISD::FP_ROUND
8728 && N0.getNode()->getConstantOperandVal(1) == 1) {
8729 SDValue In = N0.getOperand(0);
8730 if (In.getValueType() == VT) return In;
8731 if (VT.bitsLT(In.getValueType()))
8732 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT,
8733 In, N0.getOperand(1));
8734 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In);
8737 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
8738 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
8739 TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) {
8740 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
8741 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
8743 LN0->getBasePtr(), N0.getValueType(),
8744 LN0->getMemOperand());
8745 CombineTo(N, ExtLoad);
8746 CombineTo(N0.getNode(),
8747 DAG.getNode(ISD::FP_ROUND, SDLoc(N0),
8748 N0.getValueType(), ExtLoad,
8749 DAG.getIntPtrConstant(1, SDLoc(N0))),
8750 ExtLoad.getValue(1));
8751 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8757 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
8758 SDValue N0 = N->getOperand(0);
8759 EVT VT = N->getValueType(0);
8761 // fold (fceil c1) -> fceil(c1)
8762 if (isConstantFPBuildVectorOrConstantFP(N0))
8763 return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0);
8768 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
8769 SDValue N0 = N->getOperand(0);
8770 EVT VT = N->getValueType(0);
8772 // fold (ftrunc c1) -> ftrunc(c1)
8773 if (isConstantFPBuildVectorOrConstantFP(N0))
8774 return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0);
8779 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
8780 SDValue N0 = N->getOperand(0);
8781 EVT VT = N->getValueType(0);
8783 // fold (ffloor c1) -> ffloor(c1)
8784 if (isConstantFPBuildVectorOrConstantFP(N0))
8785 return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0);
8790 // FIXME: FNEG and FABS have a lot in common; refactor.
8791 SDValue DAGCombiner::visitFNEG(SDNode *N) {
8792 SDValue N0 = N->getOperand(0);
8793 EVT VT = N->getValueType(0);
8795 // Constant fold FNEG.
8796 if (isConstantFPBuildVectorOrConstantFP(N0))
8797 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
8799 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
8800 &DAG.getTarget().Options))
8801 return GetNegatedExpression(N0, DAG, LegalOperations);
8803 // Transform fneg(bitconvert(x)) -> bitconvert(x ^ sign) to avoid loading
8804 // constant pool values.
8805 if (!TLI.isFNegFree(VT) &&
8806 N0.getOpcode() == ISD::BITCAST &&
8807 N0.getNode()->hasOneUse()) {
8808 SDValue Int = N0.getOperand(0);
8809 EVT IntVT = Int.getValueType();
8810 if (IntVT.isInteger() && !IntVT.isVector()) {
8812 if (N0.getValueType().isVector()) {
8813 // For a vector, get a mask such as 0x80... per scalar element
8815 SignMask = APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
8816 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
8818 // For a scalar, just generate 0x80...
8819 SignMask = APInt::getSignBit(IntVT.getSizeInBits());
8822 Int = DAG.getNode(ISD::XOR, DL0, IntVT, Int,
8823 DAG.getConstant(SignMask, DL0, IntVT));
8824 AddToWorklist(Int.getNode());
8825 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Int);
8829 // (fneg (fmul c, x)) -> (fmul -c, x)
8830 if (N0.getOpcode() == ISD::FMUL &&
8831 (N0.getNode()->hasOneUse() || !TLI.isFNegFree(VT))) {
8832 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
8834 APFloat CVal = CFP1->getValueAPF();
8836 if (Level >= AfterLegalizeDAG &&
8837 (TLI.isFPImmLegal(CVal, N->getValueType(0)) ||
8838 TLI.isOperationLegal(ISD::ConstantFP, N->getValueType(0))))
8840 ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
8841 DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0.getOperand(1)));
8848 SDValue DAGCombiner::visitFMINNUM(SDNode *N) {
8849 SDValue N0 = N->getOperand(0);
8850 SDValue N1 = N->getOperand(1);
8851 const ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8852 const ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8854 if (N0CFP && N1CFP) {
8855 const APFloat &C0 = N0CFP->getValueAPF();
8856 const APFloat &C1 = N1CFP->getValueAPF();
8857 return DAG.getConstantFP(minnum(C0, C1), SDLoc(N), N->getValueType(0));
8861 EVT VT = N->getValueType(0);
8862 // Canonicalize to constant on RHS.
8863 return DAG.getNode(ISD::FMINNUM, SDLoc(N), VT, N1, N0);
8869 SDValue DAGCombiner::visitFMAXNUM(SDNode *N) {
8870 SDValue N0 = N->getOperand(0);
8871 SDValue N1 = N->getOperand(1);
8872 const ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8873 const ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8875 if (N0CFP && N1CFP) {
8876 const APFloat &C0 = N0CFP->getValueAPF();
8877 const APFloat &C1 = N1CFP->getValueAPF();
8878 return DAG.getConstantFP(maxnum(C0, C1), SDLoc(N), N->getValueType(0));
8882 EVT VT = N->getValueType(0);
8883 // Canonicalize to constant on RHS.
8884 return DAG.getNode(ISD::FMAXNUM, SDLoc(N), VT, N1, N0);
8890 SDValue DAGCombiner::visitFABS(SDNode *N) {
8891 SDValue N0 = N->getOperand(0);
8892 EVT VT = N->getValueType(0);
8894 // fold (fabs c1) -> fabs(c1)
8895 if (isConstantFPBuildVectorOrConstantFP(N0))
8896 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
8898 // fold (fabs (fabs x)) -> (fabs x)
8899 if (N0.getOpcode() == ISD::FABS)
8900 return N->getOperand(0);
8902 // fold (fabs (fneg x)) -> (fabs x)
8903 // fold (fabs (fcopysign x, y)) -> (fabs x)
8904 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
8905 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0));
8907 // Transform fabs(bitconvert(x)) -> bitconvert(x & ~sign) to avoid loading
8908 // constant pool values.
8909 if (!TLI.isFAbsFree(VT) &&
8910 N0.getOpcode() == ISD::BITCAST &&
8911 N0.getNode()->hasOneUse()) {
8912 SDValue Int = N0.getOperand(0);
8913 EVT IntVT = Int.getValueType();
8914 if (IntVT.isInteger() && !IntVT.isVector()) {
8916 if (N0.getValueType().isVector()) {
8917 // For a vector, get a mask such as 0x7f... per scalar element
8919 SignMask = ~APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
8920 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
8922 // For a scalar, just generate 0x7f...
8923 SignMask = ~APInt::getSignBit(IntVT.getSizeInBits());
8926 Int = DAG.getNode(ISD::AND, DL, IntVT, Int,
8927 DAG.getConstant(SignMask, DL, IntVT));
8928 AddToWorklist(Int.getNode());
8929 return DAG.getNode(ISD::BITCAST, SDLoc(N), N->getValueType(0), Int);
8936 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
8937 SDValue Chain = N->getOperand(0);
8938 SDValue N1 = N->getOperand(1);
8939 SDValue N2 = N->getOperand(2);
8941 // If N is a constant we could fold this into a fallthrough or unconditional
8942 // branch. However that doesn't happen very often in normal code, because
8943 // Instcombine/SimplifyCFG should have handled the available opportunities.
8944 // If we did this folding here, it would be necessary to update the
8945 // MachineBasicBlock CFG, which is awkward.
8947 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
8949 if (N1.getOpcode() == ISD::SETCC &&
8950 TLI.isOperationLegalOrCustom(ISD::BR_CC,
8951 N1.getOperand(0).getValueType())) {
8952 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
8953 Chain, N1.getOperand(2),
8954 N1.getOperand(0), N1.getOperand(1), N2);
8957 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
8958 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
8959 (N1.getOperand(0).hasOneUse() &&
8960 N1.getOperand(0).getOpcode() == ISD::SRL))) {
8961 SDNode *Trunc = nullptr;
8962 if (N1.getOpcode() == ISD::TRUNCATE) {
8963 // Look pass the truncate.
8964 Trunc = N1.getNode();
8965 N1 = N1.getOperand(0);
8968 // Match this pattern so that we can generate simpler code:
8971 // %b = and i32 %a, 2
8972 // %c = srl i32 %b, 1
8973 // brcond i32 %c ...
8978 // %b = and i32 %a, 2
8979 // %c = setcc eq %b, 0
8982 // This applies only when the AND constant value has one bit set and the
8983 // SRL constant is equal to the log2 of the AND constant. The back-end is
8984 // smart enough to convert the result into a TEST/JMP sequence.
8985 SDValue Op0 = N1.getOperand(0);
8986 SDValue Op1 = N1.getOperand(1);
8988 if (Op0.getOpcode() == ISD::AND &&
8989 Op1.getOpcode() == ISD::Constant) {
8990 SDValue AndOp1 = Op0.getOperand(1);
8992 if (AndOp1.getOpcode() == ISD::Constant) {
8993 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
8995 if (AndConst.isPowerOf2() &&
8996 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
9000 getSetCCResultType(Op0.getValueType()),
9001 Op0, DAG.getConstant(0, DL, Op0.getValueType()),
9004 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, DL,
9005 MVT::Other, Chain, SetCC, N2);
9006 // Don't add the new BRCond into the worklist or else SimplifySelectCC
9007 // will convert it back to (X & C1) >> C2.
9008 CombineTo(N, NewBRCond, false);
9009 // Truncate is dead.
9011 deleteAndRecombine(Trunc);
9012 // Replace the uses of SRL with SETCC
9013 WorklistRemover DeadNodes(*this);
9014 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
9015 deleteAndRecombine(N1.getNode());
9016 return SDValue(N, 0); // Return N so it doesn't get rechecked!
9022 // Restore N1 if the above transformation doesn't match.
9023 N1 = N->getOperand(1);
9026 // Transform br(xor(x, y)) -> br(x != y)
9027 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
9028 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
9029 SDNode *TheXor = N1.getNode();
9030 SDValue Op0 = TheXor->getOperand(0);
9031 SDValue Op1 = TheXor->getOperand(1);
9032 if (Op0.getOpcode() == Op1.getOpcode()) {
9033 // Avoid missing important xor optimizations.
9034 SDValue Tmp = visitXOR(TheXor);
9035 if (Tmp.getNode()) {
9036 if (Tmp.getNode() != TheXor) {
9037 DEBUG(dbgs() << "\nReplacing.8 ";
9039 dbgs() << "\nWith: ";
9040 Tmp.getNode()->dump(&DAG);
9042 WorklistRemover DeadNodes(*this);
9043 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
9044 deleteAndRecombine(TheXor);
9045 return DAG.getNode(ISD::BRCOND, SDLoc(N),
9046 MVT::Other, Chain, Tmp, N2);
9049 // visitXOR has changed XOR's operands or replaced the XOR completely,
9051 return SDValue(N, 0);
9055 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
9057 if (isOneConstant(Op0) && Op0.hasOneUse() &&
9058 Op0.getOpcode() == ISD::XOR) {
9059 TheXor = Op0.getNode();
9063 EVT SetCCVT = N1.getValueType();
9065 SetCCVT = getSetCCResultType(SetCCVT);
9066 SDValue SetCC = DAG.getSetCC(SDLoc(TheXor),
9069 Equal ? ISD::SETEQ : ISD::SETNE);
9070 // Replace the uses of XOR with SETCC
9071 WorklistRemover DeadNodes(*this);
9072 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
9073 deleteAndRecombine(N1.getNode());
9074 return DAG.getNode(ISD::BRCOND, SDLoc(N),
9075 MVT::Other, Chain, SetCC, N2);
9082 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
9084 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
9085 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
9086 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
9088 // If N is a constant we could fold this into a fallthrough or unconditional
9089 // branch. However that doesn't happen very often in normal code, because
9090 // Instcombine/SimplifyCFG should have handled the available opportunities.
9091 // If we did this folding here, it would be necessary to update the
9092 // MachineBasicBlock CFG, which is awkward.
9094 // Use SimplifySetCC to simplify SETCC's.
9095 SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()),
9096 CondLHS, CondRHS, CC->get(), SDLoc(N),
9098 if (Simp.getNode()) AddToWorklist(Simp.getNode());
9100 // fold to a simpler setcc
9101 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
9102 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
9103 N->getOperand(0), Simp.getOperand(2),
9104 Simp.getOperand(0), Simp.getOperand(1),
9110 /// Return true if 'Use' is a load or a store that uses N as its base pointer
9111 /// and that N may be folded in the load / store addressing mode.
9112 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
9114 const TargetLowering &TLI) {
9118 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
9119 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
9121 VT = LD->getMemoryVT();
9122 AS = LD->getAddressSpace();
9123 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
9124 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
9126 VT = ST->getMemoryVT();
9127 AS = ST->getAddressSpace();
9131 TargetLowering::AddrMode AM;
9132 if (N->getOpcode() == ISD::ADD) {
9133 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
9136 AM.BaseOffs = Offset->getSExtValue();
9140 } else if (N->getOpcode() == ISD::SUB) {
9141 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
9144 AM.BaseOffs = -Offset->getSExtValue();
9151 return TLI.isLegalAddressingMode(DAG.getDataLayout(), AM,
9152 VT.getTypeForEVT(*DAG.getContext()), AS);
9155 /// Try turning a load/store into a pre-indexed load/store when the base
9156 /// pointer is an add or subtract and it has other uses besides the load/store.
9157 /// After the transformation, the new indexed load/store has effectively folded
9158 /// the add/subtract in and all of its other uses are redirected to the
9160 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
9161 if (Level < AfterLegalizeDAG)
9167 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9168 if (LD->isIndexed())
9170 VT = LD->getMemoryVT();
9171 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
9172 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
9174 Ptr = LD->getBasePtr();
9175 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9176 if (ST->isIndexed())
9178 VT = ST->getMemoryVT();
9179 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
9180 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
9182 Ptr = ST->getBasePtr();
9188 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
9189 // out. There is no reason to make this a preinc/predec.
9190 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
9191 Ptr.getNode()->hasOneUse())
9194 // Ask the target to do addressing mode selection.
9197 ISD::MemIndexedMode AM = ISD::UNINDEXED;
9198 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
9201 // Backends without true r+i pre-indexed forms may need to pass a
9202 // constant base with a variable offset so that constant coercion
9203 // will work with the patterns in canonical form.
9204 bool Swapped = false;
9205 if (isa<ConstantSDNode>(BasePtr)) {
9206 std::swap(BasePtr, Offset);
9210 // Don't create a indexed load / store with zero offset.
9211 if (isNullConstant(Offset))
9214 // Try turning it into a pre-indexed load / store except when:
9215 // 1) The new base ptr is a frame index.
9216 // 2) If N is a store and the new base ptr is either the same as or is a
9217 // predecessor of the value being stored.
9218 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
9219 // that would create a cycle.
9220 // 4) All uses are load / store ops that use it as old base ptr.
9222 // Check #1. Preinc'ing a frame index would require copying the stack pointer
9223 // (plus the implicit offset) to a register to preinc anyway.
9224 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
9229 SDValue Val = cast<StoreSDNode>(N)->getValue();
9230 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
9234 // If the offset is a constant, there may be other adds of constants that
9235 // can be folded with this one. We should do this to avoid having to keep
9236 // a copy of the original base pointer.
9237 SmallVector<SDNode *, 16> OtherUses;
9238 if (isa<ConstantSDNode>(Offset))
9239 for (SDNode::use_iterator UI = BasePtr.getNode()->use_begin(),
9240 UE = BasePtr.getNode()->use_end();
9242 SDUse &Use = UI.getUse();
9243 // Skip the use that is Ptr and uses of other results from BasePtr's
9244 // node (important for nodes that return multiple results).
9245 if (Use.getUser() == Ptr.getNode() || Use != BasePtr)
9248 if (Use.getUser()->isPredecessorOf(N))
9251 if (Use.getUser()->getOpcode() != ISD::ADD &&
9252 Use.getUser()->getOpcode() != ISD::SUB) {
9257 SDValue Op1 = Use.getUser()->getOperand((UI.getOperandNo() + 1) & 1);
9258 if (!isa<ConstantSDNode>(Op1)) {
9263 // FIXME: In some cases, we can be smarter about this.
9264 if (Op1.getValueType() != Offset.getValueType()) {
9269 OtherUses.push_back(Use.getUser());
9273 std::swap(BasePtr, Offset);
9275 // Now check for #3 and #4.
9276 bool RealUse = false;
9278 // Caches for hasPredecessorHelper
9279 SmallPtrSet<const SDNode *, 32> Visited;
9280 SmallVector<const SDNode *, 16> Worklist;
9282 for (SDNode *Use : Ptr.getNode()->uses()) {
9285 if (N->hasPredecessorHelper(Use, Visited, Worklist))
9288 // If Ptr may be folded in addressing mode of other use, then it's
9289 // not profitable to do this transformation.
9290 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
9299 Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
9300 BasePtr, Offset, AM);
9302 Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
9303 BasePtr, Offset, AM);
9306 DEBUG(dbgs() << "\nReplacing.4 ";
9308 dbgs() << "\nWith: ";
9309 Result.getNode()->dump(&DAG);
9311 WorklistRemover DeadNodes(*this);
9313 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
9314 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
9316 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
9319 // Finally, since the node is now dead, remove it from the graph.
9320 deleteAndRecombine(N);
9323 std::swap(BasePtr, Offset);
9325 // Replace other uses of BasePtr that can be updated to use Ptr
9326 for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) {
9327 unsigned OffsetIdx = 1;
9328 if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode())
9330 assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() ==
9331 BasePtr.getNode() && "Expected BasePtr operand");
9333 // We need to replace ptr0 in the following expression:
9334 // x0 * offset0 + y0 * ptr0 = t0
9336 // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store)
9338 // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the
9339 // indexed load/store and the expresion that needs to be re-written.
9341 // Therefore, we have:
9342 // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1
9344 ConstantSDNode *CN =
9345 cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx));
9347 APInt Offset0 = CN->getAPIntValue();
9348 APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue();
9350 X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1;
9351 Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1;
9352 X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1;
9353 Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1;
9355 unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD;
9357 APInt CNV = Offset0;
9358 if (X0 < 0) CNV = -CNV;
9359 if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1;
9360 else CNV = CNV - Offset1;
9362 SDLoc DL(OtherUses[i]);
9364 // We can now generate the new expression.
9365 SDValue NewOp1 = DAG.getConstant(CNV, DL, CN->getValueType(0));
9366 SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0);
9368 SDValue NewUse = DAG.getNode(Opcode,
9370 OtherUses[i]->getValueType(0), NewOp1, NewOp2);
9371 DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse);
9372 deleteAndRecombine(OtherUses[i]);
9375 // Replace the uses of Ptr with uses of the updated base value.
9376 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
9377 deleteAndRecombine(Ptr.getNode());
9382 /// Try to combine a load/store with a add/sub of the base pointer node into a
9383 /// post-indexed load/store. The transformation folded the add/subtract into the
9384 /// new indexed load/store effectively and all of its uses are redirected to the
9386 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
9387 if (Level < AfterLegalizeDAG)
9393 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9394 if (LD->isIndexed())
9396 VT = LD->getMemoryVT();
9397 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
9398 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
9400 Ptr = LD->getBasePtr();
9401 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9402 if (ST->isIndexed())
9404 VT = ST->getMemoryVT();
9405 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
9406 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
9408 Ptr = ST->getBasePtr();
9414 if (Ptr.getNode()->hasOneUse())
9417 for (SDNode *Op : Ptr.getNode()->uses()) {
9419 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
9424 ISD::MemIndexedMode AM = ISD::UNINDEXED;
9425 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
9426 // Don't create a indexed load / store with zero offset.
9427 if (isNullConstant(Offset))
9430 // Try turning it into a post-indexed load / store except when
9431 // 1) All uses are load / store ops that use it as base ptr (and
9432 // it may be folded as addressing mmode).
9433 // 2) Op must be independent of N, i.e. Op is neither a predecessor
9434 // nor a successor of N. Otherwise, if Op is folded that would
9437 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
9441 bool TryNext = false;
9442 for (SDNode *Use : BasePtr.getNode()->uses()) {
9443 if (Use == Ptr.getNode())
9446 // If all the uses are load / store addresses, then don't do the
9448 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
9449 bool RealUse = false;
9450 for (SDNode *UseUse : Use->uses()) {
9451 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
9466 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
9467 SDValue Result = isLoad
9468 ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
9469 BasePtr, Offset, AM)
9470 : DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
9471 BasePtr, Offset, AM);
9474 DEBUG(dbgs() << "\nReplacing.5 ";
9476 dbgs() << "\nWith: ";
9477 Result.getNode()->dump(&DAG);
9479 WorklistRemover DeadNodes(*this);
9481 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
9482 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
9484 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
9487 // Finally, since the node is now dead, remove it from the graph.
9488 deleteAndRecombine(N);
9490 // Replace the uses of Use with uses of the updated base value.
9491 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
9492 Result.getValue(isLoad ? 1 : 0));
9493 deleteAndRecombine(Op);
9502 /// \brief Return the base-pointer arithmetic from an indexed \p LD.
9503 SDValue DAGCombiner::SplitIndexingFromLoad(LoadSDNode *LD) {
9504 ISD::MemIndexedMode AM = LD->getAddressingMode();
9505 assert(AM != ISD::UNINDEXED);
9506 SDValue BP = LD->getOperand(1);
9507 SDValue Inc = LD->getOperand(2);
9509 // Some backends use TargetConstants for load offsets, but don't expect
9510 // TargetConstants in general ADD nodes. We can convert these constants into
9511 // regular Constants (if the constant is not opaque).
9512 assert((Inc.getOpcode() != ISD::TargetConstant ||
9513 !cast<ConstantSDNode>(Inc)->isOpaque()) &&
9514 "Cannot split out indexing using opaque target constants");
9515 if (Inc.getOpcode() == ISD::TargetConstant) {
9516 ConstantSDNode *ConstInc = cast<ConstantSDNode>(Inc);
9517 Inc = DAG.getConstant(*ConstInc->getConstantIntValue(), SDLoc(Inc),
9518 ConstInc->getValueType(0));
9522 (AM == ISD::PRE_INC || AM == ISD::POST_INC ? ISD::ADD : ISD::SUB);
9523 return DAG.getNode(Opc, SDLoc(LD), BP.getSimpleValueType(), BP, Inc);
9526 SDValue DAGCombiner::visitLOAD(SDNode *N) {
9527 LoadSDNode *LD = cast<LoadSDNode>(N);
9528 SDValue Chain = LD->getChain();
9529 SDValue Ptr = LD->getBasePtr();
9531 // If load is not volatile and there are no uses of the loaded value (and
9532 // the updated indexed value in case of indexed loads), change uses of the
9533 // chain value into uses of the chain input (i.e. delete the dead load).
9534 if (!LD->isVolatile()) {
9535 if (N->getValueType(1) == MVT::Other) {
9537 if (!N->hasAnyUseOfValue(0)) {
9538 // It's not safe to use the two value CombineTo variant here. e.g.
9539 // v1, chain2 = load chain1, loc
9540 // v2, chain3 = load chain2, loc
9542 // Now we replace use of chain2 with chain1. This makes the second load
9543 // isomorphic to the one we are deleting, and thus makes this load live.
9544 DEBUG(dbgs() << "\nReplacing.6 ";
9546 dbgs() << "\nWith chain: ";
9547 Chain.getNode()->dump(&DAG);
9549 WorklistRemover DeadNodes(*this);
9550 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
9553 deleteAndRecombine(N);
9555 return SDValue(N, 0); // Return N so it doesn't get rechecked!
9559 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
9561 // If this load has an opaque TargetConstant offset, then we cannot split
9562 // the indexing into an add/sub directly (that TargetConstant may not be
9563 // valid for a different type of node, and we cannot convert an opaque
9564 // target constant into a regular constant).
9565 bool HasOTCInc = LD->getOperand(2).getOpcode() == ISD::TargetConstant &&
9566 cast<ConstantSDNode>(LD->getOperand(2))->isOpaque();
9568 if (!N->hasAnyUseOfValue(0) &&
9569 ((MaySplitLoadIndex && !HasOTCInc) || !N->hasAnyUseOfValue(1))) {
9570 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
9572 if (N->hasAnyUseOfValue(1) && MaySplitLoadIndex && !HasOTCInc) {
9573 Index = SplitIndexingFromLoad(LD);
9574 // Try to fold the base pointer arithmetic into subsequent loads and
9576 AddUsersToWorklist(N);
9578 Index = DAG.getUNDEF(N->getValueType(1));
9579 DEBUG(dbgs() << "\nReplacing.7 ";
9581 dbgs() << "\nWith: ";
9582 Undef.getNode()->dump(&DAG);
9583 dbgs() << " and 2 other values\n");
9584 WorklistRemover DeadNodes(*this);
9585 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
9586 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Index);
9587 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
9588 deleteAndRecombine(N);
9589 return SDValue(N, 0); // Return N so it doesn't get rechecked!
9594 // If this load is directly stored, replace the load value with the stored
9596 // TODO: Handle store large -> read small portion.
9597 // TODO: Handle TRUNCSTORE/LOADEXT
9598 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
9599 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
9600 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
9601 if (PrevST->getBasePtr() == Ptr &&
9602 PrevST->getValue().getValueType() == N->getValueType(0))
9603 return CombineTo(N, Chain.getOperand(1), Chain);
9607 // Try to infer better alignment information than the load already has.
9608 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
9609 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
9610 if (Align > LD->getMemOperand()->getBaseAlignment()) {
9612 DAG.getExtLoad(LD->getExtensionType(), SDLoc(N),
9613 LD->getValueType(0),
9614 Chain, Ptr, LD->getPointerInfo(),
9616 LD->isVolatile(), LD->isNonTemporal(),
9617 LD->isInvariant(), Align, LD->getAAInfo());
9618 if (NewLoad.getNode() != N)
9619 return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true);
9624 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
9625 : DAG.getSubtarget().useAA();
9627 if (CombinerAAOnlyFunc.getNumOccurrences() &&
9628 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
9631 if (UseAA && LD->isUnindexed()) {
9632 // Walk up chain skipping non-aliasing memory nodes.
9633 SDValue BetterChain = FindBetterChain(N, Chain);
9635 // If there is a better chain.
9636 if (Chain != BetterChain) {
9639 // Replace the chain to void dependency.
9640 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
9641 ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD),
9642 BetterChain, Ptr, LD->getMemOperand());
9644 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD),
9645 LD->getValueType(0),
9646 BetterChain, Ptr, LD->getMemoryVT(),
9647 LD->getMemOperand());
9650 // Create token factor to keep old chain connected.
9651 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
9652 MVT::Other, Chain, ReplLoad.getValue(1));
9654 // Make sure the new and old chains are cleaned up.
9655 AddToWorklist(Token.getNode());
9657 // Replace uses with load result and token factor. Don't add users
9659 return CombineTo(N, ReplLoad.getValue(0), Token, false);
9663 // Try transforming N to an indexed load.
9664 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
9665 return SDValue(N, 0);
9667 // Try to slice up N to more direct loads if the slices are mapped to
9668 // different register banks or pairing can take place.
9670 return SDValue(N, 0);
9676 /// \brief Helper structure used to slice a load in smaller loads.
9677 /// Basically a slice is obtained from the following sequence:
9678 /// Origin = load Ty1, Base
9679 /// Shift = srl Ty1 Origin, CstTy Amount
9680 /// Inst = trunc Shift to Ty2
9682 /// Then, it will be rewriten into:
9683 /// Slice = load SliceTy, Base + SliceOffset
9684 /// [Inst = zext Slice to Ty2], only if SliceTy <> Ty2
9686 /// SliceTy is deduced from the number of bits that are actually used to
9688 struct LoadedSlice {
9689 /// \brief Helper structure used to compute the cost of a slice.
9691 /// Are we optimizing for code size.
9696 unsigned CrossRegisterBanksCopies;
9700 Cost(bool ForCodeSize = false)
9701 : ForCodeSize(ForCodeSize), Loads(0), Truncates(0),
9702 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {}
9704 /// \brief Get the cost of one isolated slice.
9705 Cost(const LoadedSlice &LS, bool ForCodeSize = false)
9706 : ForCodeSize(ForCodeSize), Loads(1), Truncates(0),
9707 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {
9708 EVT TruncType = LS.Inst->getValueType(0);
9709 EVT LoadedType = LS.getLoadedType();
9710 if (TruncType != LoadedType &&
9711 !LS.DAG->getTargetLoweringInfo().isZExtFree(LoadedType, TruncType))
9715 /// \brief Account for slicing gain in the current cost.
9716 /// Slicing provide a few gains like removing a shift or a
9717 /// truncate. This method allows to grow the cost of the original
9718 /// load with the gain from this slice.
9719 void addSliceGain(const LoadedSlice &LS) {
9720 // Each slice saves a truncate.
9721 const TargetLowering &TLI = LS.DAG->getTargetLoweringInfo();
9722 if (!TLI.isTruncateFree(LS.Inst->getValueType(0),
9723 LS.Inst->getOperand(0).getValueType()))
9725 // If there is a shift amount, this slice gets rid of it.
9728 // If this slice can merge a cross register bank copy, account for it.
9729 if (LS.canMergeExpensiveCrossRegisterBankCopy())
9730 ++CrossRegisterBanksCopies;
9733 Cost &operator+=(const Cost &RHS) {
9735 Truncates += RHS.Truncates;
9736 CrossRegisterBanksCopies += RHS.CrossRegisterBanksCopies;
9742 bool operator==(const Cost &RHS) const {
9743 return Loads == RHS.Loads && Truncates == RHS.Truncates &&
9744 CrossRegisterBanksCopies == RHS.CrossRegisterBanksCopies &&
9745 ZExts == RHS.ZExts && Shift == RHS.Shift;
9748 bool operator!=(const Cost &RHS) const { return !(*this == RHS); }
9750 bool operator<(const Cost &RHS) const {
9751 // Assume cross register banks copies are as expensive as loads.
9752 // FIXME: Do we want some more target hooks?
9753 unsigned ExpensiveOpsLHS = Loads + CrossRegisterBanksCopies;
9754 unsigned ExpensiveOpsRHS = RHS.Loads + RHS.CrossRegisterBanksCopies;
9755 // Unless we are optimizing for code size, consider the
9756 // expensive operation first.
9757 if (!ForCodeSize && ExpensiveOpsLHS != ExpensiveOpsRHS)
9758 return ExpensiveOpsLHS < ExpensiveOpsRHS;
9759 return (Truncates + ZExts + Shift + ExpensiveOpsLHS) <
9760 (RHS.Truncates + RHS.ZExts + RHS.Shift + ExpensiveOpsRHS);
9763 bool operator>(const Cost &RHS) const { return RHS < *this; }
9765 bool operator<=(const Cost &RHS) const { return !(RHS < *this); }
9767 bool operator>=(const Cost &RHS) const { return !(*this < RHS); }
9769 // The last instruction that represent the slice. This should be a
9770 // truncate instruction.
9772 // The original load instruction.
9774 // The right shift amount in bits from the original load.
9776 // The DAG from which Origin came from.
9777 // This is used to get some contextual information about legal types, etc.
9780 LoadedSlice(SDNode *Inst = nullptr, LoadSDNode *Origin = nullptr,
9781 unsigned Shift = 0, SelectionDAG *DAG = nullptr)
9782 : Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {}
9784 /// \brief Get the bits used in a chunk of bits \p BitWidth large.
9785 /// \return Result is \p BitWidth and has used bits set to 1 and
9786 /// not used bits set to 0.
9787 APInt getUsedBits() const {
9788 // Reproduce the trunc(lshr) sequence:
9789 // - Start from the truncated value.
9790 // - Zero extend to the desired bit width.
9792 assert(Origin && "No original load to compare against.");
9793 unsigned BitWidth = Origin->getValueSizeInBits(0);
9794 assert(Inst && "This slice is not bound to an instruction");
9795 assert(Inst->getValueSizeInBits(0) <= BitWidth &&
9796 "Extracted slice is bigger than the whole type!");
9797 APInt UsedBits(Inst->getValueSizeInBits(0), 0);
9798 UsedBits.setAllBits();
9799 UsedBits = UsedBits.zext(BitWidth);
9804 /// \brief Get the size of the slice to be loaded in bytes.
9805 unsigned getLoadedSize() const {
9806 unsigned SliceSize = getUsedBits().countPopulation();
9807 assert(!(SliceSize & 0x7) && "Size is not a multiple of a byte.");
9808 return SliceSize / 8;
9811 /// \brief Get the type that will be loaded for this slice.
9812 /// Note: This may not be the final type for the slice.
9813 EVT getLoadedType() const {
9814 assert(DAG && "Missing context");
9815 LLVMContext &Ctxt = *DAG->getContext();
9816 return EVT::getIntegerVT(Ctxt, getLoadedSize() * 8);
9819 /// \brief Get the alignment of the load used for this slice.
9820 unsigned getAlignment() const {
9821 unsigned Alignment = Origin->getAlignment();
9822 unsigned Offset = getOffsetFromBase();
9824 Alignment = MinAlign(Alignment, Alignment + Offset);
9828 /// \brief Check if this slice can be rewritten with legal operations.
9829 bool isLegal() const {
9830 // An invalid slice is not legal.
9831 if (!Origin || !Inst || !DAG)
9834 // Offsets are for indexed load only, we do not handle that.
9835 if (Origin->getOffset().getOpcode() != ISD::UNDEF)
9838 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
9840 // Check that the type is legal.
9841 EVT SliceType = getLoadedType();
9842 if (!TLI.isTypeLegal(SliceType))
9845 // Check that the load is legal for this type.
9846 if (!TLI.isOperationLegal(ISD::LOAD, SliceType))
9849 // Check that the offset can be computed.
9850 // 1. Check its type.
9851 EVT PtrType = Origin->getBasePtr().getValueType();
9852 if (PtrType == MVT::Untyped || PtrType.isExtended())
9855 // 2. Check that it fits in the immediate.
9856 if (!TLI.isLegalAddImmediate(getOffsetFromBase()))
9859 // 3. Check that the computation is legal.
9860 if (!TLI.isOperationLegal(ISD::ADD, PtrType))
9863 // Check that the zext is legal if it needs one.
9864 EVT TruncateType = Inst->getValueType(0);
9865 if (TruncateType != SliceType &&
9866 !TLI.isOperationLegal(ISD::ZERO_EXTEND, TruncateType))
9872 /// \brief Get the offset in bytes of this slice in the original chunk of
9874 /// \pre DAG != nullptr.
9875 uint64_t getOffsetFromBase() const {
9876 assert(DAG && "Missing context.");
9877 bool IsBigEndian = DAG->getDataLayout().isBigEndian();
9878 assert(!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported.");
9879 uint64_t Offset = Shift / 8;
9880 unsigned TySizeInBytes = Origin->getValueSizeInBits(0) / 8;
9881 assert(!(Origin->getValueSizeInBits(0) & 0x7) &&
9882 "The size of the original loaded type is not a multiple of a"
9884 // If Offset is bigger than TySizeInBytes, it means we are loading all
9885 // zeros. This should have been optimized before in the process.
9886 assert(TySizeInBytes > Offset &&
9887 "Invalid shift amount for given loaded size");
9889 Offset = TySizeInBytes - Offset - getLoadedSize();
9893 /// \brief Generate the sequence of instructions to load the slice
9894 /// represented by this object and redirect the uses of this slice to
9895 /// this new sequence of instructions.
9896 /// \pre this->Inst && this->Origin are valid Instructions and this
9897 /// object passed the legal check: LoadedSlice::isLegal returned true.
9898 /// \return The last instruction of the sequence used to load the slice.
9899 SDValue loadSlice() const {
9900 assert(Inst && Origin && "Unable to replace a non-existing slice.");
9901 const SDValue &OldBaseAddr = Origin->getBasePtr();
9902 SDValue BaseAddr = OldBaseAddr;
9903 // Get the offset in that chunk of bytes w.r.t. the endianess.
9904 int64_t Offset = static_cast<int64_t>(getOffsetFromBase());
9905 assert(Offset >= 0 && "Offset too big to fit in int64_t!");
9907 // BaseAddr = BaseAddr + Offset.
9908 EVT ArithType = BaseAddr.getValueType();
9910 BaseAddr = DAG->getNode(ISD::ADD, DL, ArithType, BaseAddr,
9911 DAG->getConstant(Offset, DL, ArithType));
9914 // Create the type of the loaded slice according to its size.
9915 EVT SliceType = getLoadedType();
9917 // Create the load for the slice.
9918 SDValue LastInst = DAG->getLoad(
9919 SliceType, SDLoc(Origin), Origin->getChain(), BaseAddr,
9920 Origin->getPointerInfo().getWithOffset(Offset), Origin->isVolatile(),
9921 Origin->isNonTemporal(), Origin->isInvariant(), getAlignment());
9922 // If the final type is not the same as the loaded type, this means that
9923 // we have to pad with zero. Create a zero extend for that.
9924 EVT FinalType = Inst->getValueType(0);
9925 if (SliceType != FinalType)
9927 DAG->getNode(ISD::ZERO_EXTEND, SDLoc(LastInst), FinalType, LastInst);
9931 /// \brief Check if this slice can be merged with an expensive cross register
9932 /// bank copy. E.g.,
9934 /// f = bitcast i32 i to float
9935 bool canMergeExpensiveCrossRegisterBankCopy() const {
9936 if (!Inst || !Inst->hasOneUse())
9938 SDNode *Use = *Inst->use_begin();
9939 if (Use->getOpcode() != ISD::BITCAST)
9941 assert(DAG && "Missing context");
9942 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
9943 EVT ResVT = Use->getValueType(0);
9944 const TargetRegisterClass *ResRC = TLI.getRegClassFor(ResVT.getSimpleVT());
9945 const TargetRegisterClass *ArgRC =
9946 TLI.getRegClassFor(Use->getOperand(0).getValueType().getSimpleVT());
9947 if (ArgRC == ResRC || !TLI.isOperationLegal(ISD::LOAD, ResVT))
9950 // At this point, we know that we perform a cross-register-bank copy.
9951 // Check if it is expensive.
9952 const TargetRegisterInfo *TRI = DAG->getSubtarget().getRegisterInfo();
9953 // Assume bitcasts are cheap, unless both register classes do not
9954 // explicitly share a common sub class.
9955 if (!TRI || TRI->getCommonSubClass(ArgRC, ResRC))
9958 // Check if it will be merged with the load.
9959 // 1. Check the alignment constraint.
9960 unsigned RequiredAlignment = DAG->getDataLayout().getABITypeAlignment(
9961 ResVT.getTypeForEVT(*DAG->getContext()));
9963 if (RequiredAlignment > getAlignment())
9966 // 2. Check that the load is a legal operation for that type.
9967 if (!TLI.isOperationLegal(ISD::LOAD, ResVT))
9970 // 3. Check that we do not have a zext in the way.
9971 if (Inst->getValueType(0) != getLoadedType())
9979 /// \brief Check that all bits set in \p UsedBits form a dense region, i.e.,
9980 /// \p UsedBits looks like 0..0 1..1 0..0.
9981 static bool areUsedBitsDense(const APInt &UsedBits) {
9982 // If all the bits are one, this is dense!
9983 if (UsedBits.isAllOnesValue())
9986 // Get rid of the unused bits on the right.
9987 APInt NarrowedUsedBits = UsedBits.lshr(UsedBits.countTrailingZeros());
9988 // Get rid of the unused bits on the left.
9989 if (NarrowedUsedBits.countLeadingZeros())
9990 NarrowedUsedBits = NarrowedUsedBits.trunc(NarrowedUsedBits.getActiveBits());
9991 // Check that the chunk of bits is completely used.
9992 return NarrowedUsedBits.isAllOnesValue();
9995 /// \brief Check whether or not \p First and \p Second are next to each other
9996 /// in memory. This means that there is no hole between the bits loaded
9997 /// by \p First and the bits loaded by \p Second.
9998 static bool areSlicesNextToEachOther(const LoadedSlice &First,
9999 const LoadedSlice &Second) {
10000 assert(First.Origin == Second.Origin && First.Origin &&
10001 "Unable to match different memory origins.");
10002 APInt UsedBits = First.getUsedBits();
10003 assert((UsedBits & Second.getUsedBits()) == 0 &&
10004 "Slices are not supposed to overlap.");
10005 UsedBits |= Second.getUsedBits();
10006 return areUsedBitsDense(UsedBits);
10009 /// \brief Adjust the \p GlobalLSCost according to the target
10010 /// paring capabilities and the layout of the slices.
10011 /// \pre \p GlobalLSCost should account for at least as many loads as
10012 /// there is in the slices in \p LoadedSlices.
10013 static void adjustCostForPairing(SmallVectorImpl<LoadedSlice> &LoadedSlices,
10014 LoadedSlice::Cost &GlobalLSCost) {
10015 unsigned NumberOfSlices = LoadedSlices.size();
10016 // If there is less than 2 elements, no pairing is possible.
10017 if (NumberOfSlices < 2)
10020 // Sort the slices so that elements that are likely to be next to each
10021 // other in memory are next to each other in the list.
10022 std::sort(LoadedSlices.begin(), LoadedSlices.end(),
10023 [](const LoadedSlice &LHS, const LoadedSlice &RHS) {
10024 assert(LHS.Origin == RHS.Origin && "Different bases not implemented.");
10025 return LHS.getOffsetFromBase() < RHS.getOffsetFromBase();
10027 const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo();
10028 // First (resp. Second) is the first (resp. Second) potentially candidate
10029 // to be placed in a paired load.
10030 const LoadedSlice *First = nullptr;
10031 const LoadedSlice *Second = nullptr;
10032 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice,
10033 // Set the beginning of the pair.
10036 Second = &LoadedSlices[CurrSlice];
10038 // If First is NULL, it means we start a new pair.
10039 // Get to the next slice.
10043 EVT LoadedType = First->getLoadedType();
10045 // If the types of the slices are different, we cannot pair them.
10046 if (LoadedType != Second->getLoadedType())
10049 // Check if the target supplies paired loads for this type.
10050 unsigned RequiredAlignment = 0;
10051 if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) {
10052 // move to the next pair, this type is hopeless.
10056 // Check if we meet the alignment requirement.
10057 if (RequiredAlignment > First->getAlignment())
10060 // Check that both loads are next to each other in memory.
10061 if (!areSlicesNextToEachOther(*First, *Second))
10064 assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!");
10065 --GlobalLSCost.Loads;
10066 // Move to the next pair.
10071 /// \brief Check the profitability of all involved LoadedSlice.
10072 /// Currently, it is considered profitable if there is exactly two
10073 /// involved slices (1) which are (2) next to each other in memory, and
10074 /// whose cost (\see LoadedSlice::Cost) is smaller than the original load (3).
10076 /// Note: The order of the elements in \p LoadedSlices may be modified, but not
10077 /// the elements themselves.
10079 /// FIXME: When the cost model will be mature enough, we can relax
10080 /// constraints (1) and (2).
10081 static bool isSlicingProfitable(SmallVectorImpl<LoadedSlice> &LoadedSlices,
10082 const APInt &UsedBits, bool ForCodeSize) {
10083 unsigned NumberOfSlices = LoadedSlices.size();
10084 if (StressLoadSlicing)
10085 return NumberOfSlices > 1;
10088 if (NumberOfSlices != 2)
10092 if (!areUsedBitsDense(UsedBits))
10096 LoadedSlice::Cost OrigCost(ForCodeSize), GlobalSlicingCost(ForCodeSize);
10097 // The original code has one big load.
10098 OrigCost.Loads = 1;
10099 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice) {
10100 const LoadedSlice &LS = LoadedSlices[CurrSlice];
10101 // Accumulate the cost of all the slices.
10102 LoadedSlice::Cost SliceCost(LS, ForCodeSize);
10103 GlobalSlicingCost += SliceCost;
10105 // Account as cost in the original configuration the gain obtained
10106 // with the current slices.
10107 OrigCost.addSliceGain(LS);
10110 // If the target supports paired load, adjust the cost accordingly.
10111 adjustCostForPairing(LoadedSlices, GlobalSlicingCost);
10112 return OrigCost > GlobalSlicingCost;
10115 /// \brief If the given load, \p LI, is used only by trunc or trunc(lshr)
10116 /// operations, split it in the various pieces being extracted.
10118 /// This sort of thing is introduced by SROA.
10119 /// This slicing takes care not to insert overlapping loads.
10120 /// \pre LI is a simple load (i.e., not an atomic or volatile load).
10121 bool DAGCombiner::SliceUpLoad(SDNode *N) {
10122 if (Level < AfterLegalizeDAG)
10125 LoadSDNode *LD = cast<LoadSDNode>(N);
10126 if (LD->isVolatile() || !ISD::isNormalLoad(LD) ||
10127 !LD->getValueType(0).isInteger())
10130 // Keep track of already used bits to detect overlapping values.
10131 // In that case, we will just abort the transformation.
10132 APInt UsedBits(LD->getValueSizeInBits(0), 0);
10134 SmallVector<LoadedSlice, 4> LoadedSlices;
10136 // Check if this load is used as several smaller chunks of bits.
10137 // Basically, look for uses in trunc or trunc(lshr) and record a new chain
10138 // of computation for each trunc.
10139 for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end();
10140 UI != UIEnd; ++UI) {
10141 // Skip the uses of the chain.
10142 if (UI.getUse().getResNo() != 0)
10145 SDNode *User = *UI;
10146 unsigned Shift = 0;
10148 // Check if this is a trunc(lshr).
10149 if (User->getOpcode() == ISD::SRL && User->hasOneUse() &&
10150 isa<ConstantSDNode>(User->getOperand(1))) {
10151 Shift = cast<ConstantSDNode>(User->getOperand(1))->getZExtValue();
10152 User = *User->use_begin();
10155 // At this point, User is a Truncate, iff we encountered, trunc or
10157 if (User->getOpcode() != ISD::TRUNCATE)
10160 // The width of the type must be a power of 2 and greater than 8-bits.
10161 // Otherwise the load cannot be represented in LLVM IR.
10162 // Moreover, if we shifted with a non-8-bits multiple, the slice
10163 // will be across several bytes. We do not support that.
10164 unsigned Width = User->getValueSizeInBits(0);
10165 if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7))
10168 // Build the slice for this chain of computations.
10169 LoadedSlice LS(User, LD, Shift, &DAG);
10170 APInt CurrentUsedBits = LS.getUsedBits();
10172 // Check if this slice overlaps with another.
10173 if ((CurrentUsedBits & UsedBits) != 0)
10175 // Update the bits used globally.
10176 UsedBits |= CurrentUsedBits;
10178 // Check if the new slice would be legal.
10182 // Record the slice.
10183 LoadedSlices.push_back(LS);
10186 // Abort slicing if it does not seem to be profitable.
10187 if (!isSlicingProfitable(LoadedSlices, UsedBits, ForCodeSize))
10192 // Rewrite each chain to use an independent load.
10193 // By construction, each chain can be represented by a unique load.
10195 // Prepare the argument for the new token factor for all the slices.
10196 SmallVector<SDValue, 8> ArgChains;
10197 for (SmallVectorImpl<LoadedSlice>::const_iterator
10198 LSIt = LoadedSlices.begin(),
10199 LSItEnd = LoadedSlices.end();
10200 LSIt != LSItEnd; ++LSIt) {
10201 SDValue SliceInst = LSIt->loadSlice();
10202 CombineTo(LSIt->Inst, SliceInst, true);
10203 if (SliceInst.getNode()->getOpcode() != ISD::LOAD)
10204 SliceInst = SliceInst.getOperand(0);
10205 assert(SliceInst->getOpcode() == ISD::LOAD &&
10206 "It takes more than a zext to get to the loaded slice!!");
10207 ArgChains.push_back(SliceInst.getValue(1));
10210 SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other,
10212 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
10216 /// Check to see if V is (and load (ptr), imm), where the load is having
10217 /// specific bytes cleared out. If so, return the byte size being masked out
10218 /// and the shift amount.
10219 static std::pair<unsigned, unsigned>
10220 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
10221 std::pair<unsigned, unsigned> Result(0, 0);
10223 // Check for the structure we're looking for.
10224 if (V->getOpcode() != ISD::AND ||
10225 !isa<ConstantSDNode>(V->getOperand(1)) ||
10226 !ISD::isNormalLoad(V->getOperand(0).getNode()))
10229 // Check the chain and pointer.
10230 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
10231 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
10233 // The store should be chained directly to the load or be an operand of a
10235 if (LD == Chain.getNode())
10237 else if (Chain->getOpcode() != ISD::TokenFactor)
10238 return Result; // Fail.
10241 for (const SDValue &ChainOp : Chain->op_values())
10242 if (ChainOp.getNode() == LD) {
10246 if (!isOk) return Result;
10249 // This only handles simple types.
10250 if (V.getValueType() != MVT::i16 &&
10251 V.getValueType() != MVT::i32 &&
10252 V.getValueType() != MVT::i64)
10255 // Check the constant mask. Invert it so that the bits being masked out are
10256 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
10257 // follow the sign bit for uniformity.
10258 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
10259 unsigned NotMaskLZ = countLeadingZeros(NotMask);
10260 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
10261 unsigned NotMaskTZ = countTrailingZeros(NotMask);
10262 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
10263 if (NotMaskLZ == 64) return Result; // All zero mask.
10265 // See if we have a continuous run of bits. If so, we have 0*1+0*
10266 if (countTrailingOnes(NotMask >> NotMaskTZ) + NotMaskTZ + NotMaskLZ != 64)
10269 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
10270 if (V.getValueType() != MVT::i64 && NotMaskLZ)
10271 NotMaskLZ -= 64-V.getValueSizeInBits();
10273 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
10274 switch (MaskedBytes) {
10278 default: return Result; // All one mask, or 5-byte mask.
10281 // Verify that the first bit starts at a multiple of mask so that the access
10282 // is aligned the same as the access width.
10283 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
10285 Result.first = MaskedBytes;
10286 Result.second = NotMaskTZ/8;
10291 /// Check to see if IVal is something that provides a value as specified by
10292 /// MaskInfo. If so, replace the specified store with a narrower store of
10293 /// truncated IVal.
10295 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
10296 SDValue IVal, StoreSDNode *St,
10298 unsigned NumBytes = MaskInfo.first;
10299 unsigned ByteShift = MaskInfo.second;
10300 SelectionDAG &DAG = DC->getDAG();
10302 // Check to see if IVal is all zeros in the part being masked in by the 'or'
10303 // that uses this. If not, this is not a replacement.
10304 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
10305 ByteShift*8, (ByteShift+NumBytes)*8);
10306 if (!DAG.MaskedValueIsZero(IVal, Mask)) return nullptr;
10308 // Check that it is legal on the target to do this. It is legal if the new
10309 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
10311 MVT VT = MVT::getIntegerVT(NumBytes*8);
10312 if (!DC->isTypeLegal(VT))
10315 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
10316 // shifted by ByteShift and truncated down to NumBytes.
10319 IVal = DAG.getNode(ISD::SRL, DL, IVal.getValueType(), IVal,
10320 DAG.getConstant(ByteShift*8, DL,
10321 DC->getShiftAmountTy(IVal.getValueType())));
10324 // Figure out the offset for the store and the alignment of the access.
10326 unsigned NewAlign = St->getAlignment();
10328 if (DAG.getDataLayout().isLittleEndian())
10329 StOffset = ByteShift;
10331 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
10333 SDValue Ptr = St->getBasePtr();
10336 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(),
10337 Ptr, DAG.getConstant(StOffset, DL, Ptr.getValueType()));
10338 NewAlign = MinAlign(NewAlign, StOffset);
10341 // Truncate down to the new size.
10342 IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal);
10345 return DAG.getStore(St->getChain(), SDLoc(St), IVal, Ptr,
10346 St->getPointerInfo().getWithOffset(StOffset),
10347 false, false, NewAlign).getNode();
10351 /// Look for sequence of load / op / store where op is one of 'or', 'xor', and
10352 /// 'and' of immediates. If 'op' is only touching some of the loaded bits, try
10353 /// narrowing the load and store if it would end up being a win for performance
10355 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
10356 StoreSDNode *ST = cast<StoreSDNode>(N);
10357 if (ST->isVolatile())
10360 SDValue Chain = ST->getChain();
10361 SDValue Value = ST->getValue();
10362 SDValue Ptr = ST->getBasePtr();
10363 EVT VT = Value.getValueType();
10365 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
10368 unsigned Opc = Value.getOpcode();
10370 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
10371 // is a byte mask indicating a consecutive number of bytes, check to see if
10372 // Y is known to provide just those bytes. If so, we try to replace the
10373 // load + replace + store sequence with a single (narrower) store, which makes
10375 if (Opc == ISD::OR) {
10376 std::pair<unsigned, unsigned> MaskedLoad;
10377 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
10378 if (MaskedLoad.first)
10379 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
10380 Value.getOperand(1), ST,this))
10381 return SDValue(NewST, 0);
10383 // Or is commutative, so try swapping X and Y.
10384 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
10385 if (MaskedLoad.first)
10386 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
10387 Value.getOperand(0), ST,this))
10388 return SDValue(NewST, 0);
10391 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
10392 Value.getOperand(1).getOpcode() != ISD::Constant)
10395 SDValue N0 = Value.getOperand(0);
10396 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
10397 Chain == SDValue(N0.getNode(), 1)) {
10398 LoadSDNode *LD = cast<LoadSDNode>(N0);
10399 if (LD->getBasePtr() != Ptr ||
10400 LD->getPointerInfo().getAddrSpace() !=
10401 ST->getPointerInfo().getAddrSpace())
10404 // Find the type to narrow it the load / op / store to.
10405 SDValue N1 = Value.getOperand(1);
10406 unsigned BitWidth = N1.getValueSizeInBits();
10407 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
10408 if (Opc == ISD::AND)
10409 Imm ^= APInt::getAllOnesValue(BitWidth);
10410 if (Imm == 0 || Imm.isAllOnesValue())
10412 unsigned ShAmt = Imm.countTrailingZeros();
10413 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
10414 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
10415 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
10416 // The narrowing should be profitable, the load/store operation should be
10417 // legal (or custom) and the store size should be equal to the NewVT width.
10418 while (NewBW < BitWidth &&
10419 (NewVT.getStoreSizeInBits() != NewBW ||
10420 !TLI.isOperationLegalOrCustom(Opc, NewVT) ||
10421 !TLI.isNarrowingProfitable(VT, NewVT))) {
10422 NewBW = NextPowerOf2(NewBW);
10423 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
10425 if (NewBW >= BitWidth)
10428 // If the lsb changed does not start at the type bitwidth boundary,
10429 // start at the previous one.
10431 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
10432 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt,
10433 std::min(BitWidth, ShAmt + NewBW));
10434 if ((Imm & Mask) == Imm) {
10435 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
10436 if (Opc == ISD::AND)
10437 NewImm ^= APInt::getAllOnesValue(NewBW);
10438 uint64_t PtrOff = ShAmt / 8;
10439 // For big endian targets, we need to adjust the offset to the pointer to
10440 // load the correct bytes.
10441 if (DAG.getDataLayout().isBigEndian())
10442 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
10444 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
10445 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
10446 if (NewAlign < DAG.getDataLayout().getABITypeAlignment(NewVTTy))
10449 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD),
10450 Ptr.getValueType(), Ptr,
10451 DAG.getConstant(PtrOff, SDLoc(LD),
10452 Ptr.getValueType()));
10453 SDValue NewLD = DAG.getLoad(NewVT, SDLoc(N0),
10454 LD->getChain(), NewPtr,
10455 LD->getPointerInfo().getWithOffset(PtrOff),
10456 LD->isVolatile(), LD->isNonTemporal(),
10457 LD->isInvariant(), NewAlign,
10459 SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD,
10460 DAG.getConstant(NewImm, SDLoc(Value),
10462 SDValue NewST = DAG.getStore(Chain, SDLoc(N),
10464 ST->getPointerInfo().getWithOffset(PtrOff),
10465 false, false, NewAlign);
10467 AddToWorklist(NewPtr.getNode());
10468 AddToWorklist(NewLD.getNode());
10469 AddToWorklist(NewVal.getNode());
10470 WorklistRemover DeadNodes(*this);
10471 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
10480 /// For a given floating point load / store pair, if the load value isn't used
10481 /// by any other operations, then consider transforming the pair to integer
10482 /// load / store operations if the target deems the transformation profitable.
10483 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
10484 StoreSDNode *ST = cast<StoreSDNode>(N);
10485 SDValue Chain = ST->getChain();
10486 SDValue Value = ST->getValue();
10487 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
10488 Value.hasOneUse() &&
10489 Chain == SDValue(Value.getNode(), 1)) {
10490 LoadSDNode *LD = cast<LoadSDNode>(Value);
10491 EVT VT = LD->getMemoryVT();
10492 if (!VT.isFloatingPoint() ||
10493 VT != ST->getMemoryVT() ||
10494 LD->isNonTemporal() ||
10495 ST->isNonTemporal() ||
10496 LD->getPointerInfo().getAddrSpace() != 0 ||
10497 ST->getPointerInfo().getAddrSpace() != 0)
10500 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
10501 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
10502 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
10503 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
10504 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
10507 unsigned LDAlign = LD->getAlignment();
10508 unsigned STAlign = ST->getAlignment();
10509 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
10510 unsigned ABIAlign = DAG.getDataLayout().getABITypeAlignment(IntVTTy);
10511 if (LDAlign < ABIAlign || STAlign < ABIAlign)
10514 SDValue NewLD = DAG.getLoad(IntVT, SDLoc(Value),
10515 LD->getChain(), LD->getBasePtr(),
10516 LD->getPointerInfo(),
10517 false, false, false, LDAlign);
10519 SDValue NewST = DAG.getStore(NewLD.getValue(1), SDLoc(N),
10520 NewLD, ST->getBasePtr(),
10521 ST->getPointerInfo(),
10522 false, false, STAlign);
10524 AddToWorklist(NewLD.getNode());
10525 AddToWorklist(NewST.getNode());
10526 WorklistRemover DeadNodes(*this);
10527 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
10536 /// Helper struct to parse and store a memory address as base + index + offset.
10537 /// We ignore sign extensions when it is safe to do so.
10538 /// The following two expressions are not equivalent. To differentiate we need
10539 /// to store whether there was a sign extension involved in the index
10541 /// (load (i64 add (i64 copyfromreg %c)
10542 /// (i64 signextend (add (i8 load %index)
10546 /// (load (i64 add (i64 copyfromreg %c)
10547 /// (i64 signextend (i32 add (i32 signextend (i8 load %index))
10549 struct BaseIndexOffset {
10553 bool IsIndexSignExt;
10555 BaseIndexOffset() : Offset(0), IsIndexSignExt(false) {}
10557 BaseIndexOffset(SDValue Base, SDValue Index, int64_t Offset,
10558 bool IsIndexSignExt) :
10559 Base(Base), Index(Index), Offset(Offset), IsIndexSignExt(IsIndexSignExt) {}
10561 bool equalBaseIndex(const BaseIndexOffset &Other) {
10562 return Other.Base == Base && Other.Index == Index &&
10563 Other.IsIndexSignExt == IsIndexSignExt;
10566 /// Parses tree in Ptr for base, index, offset addresses.
10567 static BaseIndexOffset match(SDValue Ptr) {
10568 bool IsIndexSignExt = false;
10570 // We only can pattern match BASE + INDEX + OFFSET. If Ptr is not an ADD
10571 // instruction, then it could be just the BASE or everything else we don't
10572 // know how to handle. Just use Ptr as BASE and give up.
10573 if (Ptr->getOpcode() != ISD::ADD)
10574 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
10576 // We know that we have at least an ADD instruction. Try to pattern match
10577 // the simple case of BASE + OFFSET.
10578 if (isa<ConstantSDNode>(Ptr->getOperand(1))) {
10579 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
10580 return BaseIndexOffset(Ptr->getOperand(0), SDValue(), Offset,
10584 // Inside a loop the current BASE pointer is calculated using an ADD and a
10585 // MUL instruction. In this case Ptr is the actual BASE pointer.
10586 // (i64 add (i64 %array_ptr)
10587 // (i64 mul (i64 %induction_var)
10588 // (i64 %element_size)))
10589 if (Ptr->getOperand(1)->getOpcode() == ISD::MUL)
10590 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
10592 // Look at Base + Index + Offset cases.
10593 SDValue Base = Ptr->getOperand(0);
10594 SDValue IndexOffset = Ptr->getOperand(1);
10596 // Skip signextends.
10597 if (IndexOffset->getOpcode() == ISD::SIGN_EXTEND) {
10598 IndexOffset = IndexOffset->getOperand(0);
10599 IsIndexSignExt = true;
10602 // Either the case of Base + Index (no offset) or something else.
10603 if (IndexOffset->getOpcode() != ISD::ADD)
10604 return BaseIndexOffset(Base, IndexOffset, 0, IsIndexSignExt);
10606 // Now we have the case of Base + Index + offset.
10607 SDValue Index = IndexOffset->getOperand(0);
10608 SDValue Offset = IndexOffset->getOperand(1);
10610 if (!isa<ConstantSDNode>(Offset))
10611 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
10613 // Ignore signextends.
10614 if (Index->getOpcode() == ISD::SIGN_EXTEND) {
10615 Index = Index->getOperand(0);
10616 IsIndexSignExt = true;
10617 } else IsIndexSignExt = false;
10619 int64_t Off = cast<ConstantSDNode>(Offset)->getSExtValue();
10620 return BaseIndexOffset(Base, Index, Off, IsIndexSignExt);
10625 SDValue DAGCombiner::getMergedConstantVectorStore(SelectionDAG &DAG,
10627 ArrayRef<MemOpLink> Stores,
10629 SmallVector<SDValue, 8> BuildVector;
10631 for (unsigned I = 0, E = Ty.getVectorNumElements(); I != E; ++I)
10632 BuildVector.push_back(cast<StoreSDNode>(Stores[I].MemNode)->getValue());
10634 return DAG.getNode(ISD::BUILD_VECTOR, SL, Ty, BuildVector);
10637 bool DAGCombiner::MergeStoresOfConstantsOrVecElts(
10638 SmallVectorImpl<MemOpLink> &StoreNodes, EVT MemVT,
10639 unsigned NumElem, bool IsConstantSrc, bool UseVector) {
10640 // Make sure we have something to merge.
10644 int64_t ElementSizeBytes = MemVT.getSizeInBits() / 8;
10645 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
10646 unsigned LatestNodeUsed = 0;
10648 for (unsigned i=0; i < NumElem; ++i) {
10649 // Find a chain for the new wide-store operand. Notice that some
10650 // of the store nodes that we found may not be selected for inclusion
10651 // in the wide store. The chain we use needs to be the chain of the
10652 // latest store node which is *used* and replaced by the wide store.
10653 if (StoreNodes[i].SequenceNum < StoreNodes[LatestNodeUsed].SequenceNum)
10654 LatestNodeUsed = i;
10657 // The latest Node in the DAG.
10658 LSBaseSDNode *LatestOp = StoreNodes[LatestNodeUsed].MemNode;
10659 SDLoc DL(StoreNodes[0].MemNode);
10663 // Find a legal type for the vector store.
10664 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
10665 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
10666 if (IsConstantSrc) {
10667 StoredVal = getMergedConstantVectorStore(DAG, DL, StoreNodes, Ty);
10669 SmallVector<SDValue, 8> Ops;
10670 for (unsigned i = 0; i < NumElem ; ++i) {
10671 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10672 SDValue Val = St->getValue();
10673 // All of the operands of a BUILD_VECTOR must have the same type.
10674 if (Val.getValueType() != MemVT)
10676 Ops.push_back(Val);
10679 // Build the extracted vector elements back into a vector.
10680 StoredVal = DAG.getNode(ISD::BUILD_VECTOR, DL, Ty, Ops);
10683 // We should always use a vector store when merging extracted vector
10684 // elements, so this path implies a store of constants.
10685 assert(IsConstantSrc && "Merged vector elements should use vector store");
10687 unsigned SizeInBits = NumElem * ElementSizeBytes * 8;
10688 APInt StoreInt(SizeInBits, 0);
10690 // Construct a single integer constant which is made of the smaller
10691 // constant inputs.
10692 bool IsLE = DAG.getDataLayout().isLittleEndian();
10693 for (unsigned i = 0; i < NumElem ; ++i) {
10694 unsigned Idx = IsLE ? (NumElem - 1 - i) : i;
10695 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
10696 SDValue Val = St->getValue();
10697 StoreInt <<= ElementSizeBytes * 8;
10698 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
10699 StoreInt |= C->getAPIntValue().zext(SizeInBits);
10700 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
10701 StoreInt |= C->getValueAPF().bitcastToAPInt().zext(SizeInBits);
10703 llvm_unreachable("Invalid constant element type");
10707 // Create the new Load and Store operations.
10708 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), SizeInBits);
10709 StoredVal = DAG.getConstant(StoreInt, DL, StoreTy);
10712 SDValue NewStore = DAG.getStore(LatestOp->getChain(), DL, StoredVal,
10713 FirstInChain->getBasePtr(),
10714 FirstInChain->getPointerInfo(),
10716 FirstInChain->getAlignment());
10718 // Replace the last store with the new store
10719 CombineTo(LatestOp, NewStore);
10720 // Erase all other stores.
10721 for (unsigned i = 0; i < NumElem ; ++i) {
10722 if (StoreNodes[i].MemNode == LatestOp)
10724 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10725 // ReplaceAllUsesWith will replace all uses that existed when it was
10726 // called, but graph optimizations may cause new ones to appear. For
10727 // example, the case in pr14333 looks like
10729 // St's chain -> St -> another store -> X
10731 // And the only difference from St to the other store is the chain.
10732 // When we change it's chain to be St's chain they become identical,
10733 // get CSEed and the net result is that X is now a use of St.
10734 // Since we know that St is redundant, just iterate.
10735 while (!St->use_empty())
10736 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain());
10737 deleteAndRecombine(St);
10743 static bool allowableAlignment(const SelectionDAG &DAG,
10744 const TargetLowering &TLI, EVT EVTTy,
10745 unsigned AS, unsigned Align) {
10746 if (TLI.allowsMisalignedMemoryAccesses(EVTTy, AS, Align))
10749 Type *Ty = EVTTy.getTypeForEVT(*DAG.getContext());
10750 unsigned ABIAlignment = DAG.getDataLayout().getPrefTypeAlignment(Ty);
10751 return (Align >= ABIAlignment);
10754 void DAGCombiner::getStoreMergeAndAliasCandidates(
10755 StoreSDNode* St, SmallVectorImpl<MemOpLink> &StoreNodes,
10756 SmallVectorImpl<LSBaseSDNode*> &AliasLoadNodes) {
10757 // This holds the base pointer, index, and the offset in bytes from the base
10759 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr());
10761 // We must have a base and an offset.
10762 if (!BasePtr.Base.getNode())
10765 // Do not handle stores to undef base pointers.
10766 if (BasePtr.Base.getOpcode() == ISD::UNDEF)
10769 // Walk up the chain and look for nodes with offsets from the same
10770 // base pointer. Stop when reaching an instruction with a different kind
10771 // or instruction which has a different base pointer.
10772 EVT MemVT = St->getMemoryVT();
10774 StoreSDNode *Index = St;
10776 // If the chain has more than one use, then we can't reorder the mem ops.
10777 if (Index != St && !SDValue(Index, 0)->hasOneUse())
10780 // Find the base pointer and offset for this memory node.
10781 BaseIndexOffset Ptr = BaseIndexOffset::match(Index->getBasePtr());
10783 // Check that the base pointer is the same as the original one.
10784 if (!Ptr.equalBaseIndex(BasePtr))
10787 // The memory operands must not be volatile.
10788 if (Index->isVolatile() || Index->isIndexed())
10792 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
10793 if (St->isTruncatingStore())
10796 // The stored memory type must be the same.
10797 if (Index->getMemoryVT() != MemVT)
10800 // We found a potential memory operand to merge.
10801 StoreNodes.push_back(MemOpLink(Index, Ptr.Offset, Seq++));
10803 // Find the next memory operand in the chain. If the next operand in the
10804 // chain is a store then move up and continue the scan with the next
10805 // memory operand. If the next operand is a load save it and use alias
10806 // information to check if it interferes with anything.
10807 SDNode *NextInChain = Index->getChain().getNode();
10809 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
10810 // We found a store node. Use it for the next iteration.
10813 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
10814 if (Ldn->isVolatile()) {
10819 // Save the load node for later. Continue the scan.
10820 AliasLoadNodes.push_back(Ldn);
10821 NextInChain = Ldn->getChain().getNode();
10831 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
10832 if (OptLevel == CodeGenOpt::None)
10835 EVT MemVT = St->getMemoryVT();
10836 int64_t ElementSizeBytes = MemVT.getSizeInBits() / 8;
10837 bool NoVectors = DAG.getMachineFunction().getFunction()->hasFnAttribute(
10838 Attribute::NoImplicitFloat);
10840 // This function cannot currently deal with non-byte-sized memory sizes.
10841 if (ElementSizeBytes * 8 != MemVT.getSizeInBits())
10844 // Don't merge vectors into wider inputs.
10845 if (MemVT.isVector() || !MemVT.isSimple())
10848 // Perform an early exit check. Do not bother looking at stored values that
10849 // are not constants, loads, or extracted vector elements.
10850 SDValue StoredVal = St->getValue();
10851 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
10852 bool IsConstantSrc = isa<ConstantSDNode>(StoredVal) ||
10853 isa<ConstantFPSDNode>(StoredVal);
10854 bool IsExtractVecEltSrc = (StoredVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT);
10856 if (!IsConstantSrc && !IsLoadSrc && !IsExtractVecEltSrc)
10859 // Only look at ends of store sequences.
10860 SDValue Chain = SDValue(St, 0);
10861 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
10864 // Save the LoadSDNodes that we find in the chain.
10865 // We need to make sure that these nodes do not interfere with
10866 // any of the store nodes.
10867 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
10869 // Save the StoreSDNodes that we find in the chain.
10870 SmallVector<MemOpLink, 8> StoreNodes;
10872 getStoreMergeAndAliasCandidates(St, StoreNodes, AliasLoadNodes);
10874 // Check if there is anything to merge.
10875 if (StoreNodes.size() < 2)
10878 // Sort the memory operands according to their distance from the base pointer.
10879 std::sort(StoreNodes.begin(), StoreNodes.end(),
10880 [](MemOpLink LHS, MemOpLink RHS) {
10881 return LHS.OffsetFromBase < RHS.OffsetFromBase ||
10882 (LHS.OffsetFromBase == RHS.OffsetFromBase &&
10883 LHS.SequenceNum > RHS.SequenceNum);
10886 // Scan the memory operations on the chain and find the first non-consecutive
10887 // store memory address.
10888 unsigned LastConsecutiveStore = 0;
10889 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
10890 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) {
10892 // Check that the addresses are consecutive starting from the second
10893 // element in the list of stores.
10895 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
10896 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
10900 bool Alias = false;
10901 // Check if this store interferes with any of the loads that we found.
10902 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld)
10903 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) {
10907 // We found a load that alias with this store. Stop the sequence.
10911 // Mark this node as useful.
10912 LastConsecutiveStore = i;
10915 // The node with the lowest store address.
10916 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
10917 unsigned FirstStoreAS = FirstInChain->getAddressSpace();
10918 unsigned FirstStoreAlign = FirstInChain->getAlignment();
10920 // Store the constants into memory as one consecutive store.
10921 if (IsConstantSrc) {
10922 unsigned LastLegalType = 0;
10923 unsigned LastLegalVectorType = 0;
10924 bool NonZero = false;
10925 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
10926 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10927 SDValue StoredVal = St->getValue();
10929 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
10930 NonZero |= !C->isNullValue();
10931 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) {
10932 NonZero |= !C->getConstantFPValue()->isNullValue();
10938 // Find a legal type for the constant store.
10939 unsigned SizeInBits = (i+1) * ElementSizeBytes * 8;
10940 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), SizeInBits);
10941 if (TLI.isTypeLegal(StoreTy) &&
10942 allowableAlignment(DAG, TLI, StoreTy, FirstStoreAS,
10943 FirstStoreAlign)) {
10944 LastLegalType = i+1;
10945 // Or check whether a truncstore is legal.
10946 } else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
10947 TargetLowering::TypePromoteInteger) {
10948 EVT LegalizedStoredValueTy =
10949 TLI.getTypeToTransformTo(*DAG.getContext(), StoredVal.getValueType());
10950 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
10951 allowableAlignment(DAG, TLI, LegalizedStoredValueTy, FirstStoreAS,
10952 FirstStoreAlign)) {
10953 LastLegalType = i + 1;
10957 // Find a legal type for the vector store.
10958 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
10959 if (TLI.isTypeLegal(Ty) &&
10960 allowableAlignment(DAG, TLI, Ty, FirstStoreAS, FirstStoreAlign)) {
10961 LastLegalVectorType = i + 1;
10966 // We only use vectors if the constant is known to be zero or the target
10967 // allows it and the function is not marked with the noimplicitfloat
10970 LastLegalVectorType = 0;
10971 } else if (NonZero && !TLI.storeOfVectorConstantIsCheap(MemVT,
10972 LastLegalVectorType,
10974 LastLegalVectorType = 0;
10977 // Check if we found a legal integer type to store.
10978 if (LastLegalType == 0 && LastLegalVectorType == 0)
10981 bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors;
10982 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType;
10984 return MergeStoresOfConstantsOrVecElts(StoreNodes, MemVT, NumElem,
10988 // When extracting multiple vector elements, try to store them
10989 // in one vector store rather than a sequence of scalar stores.
10990 if (IsExtractVecEltSrc) {
10991 unsigned NumElem = 0;
10992 for (unsigned i = 0; i < LastConsecutiveStore + 1; ++i) {
10993 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10994 SDValue StoredVal = St->getValue();
10995 // This restriction could be loosened.
10996 // Bail out if any stored values are not elements extracted from a vector.
10997 // It should be possible to handle mixed sources, but load sources need
10998 // more careful handling (see the block of code below that handles
10999 // consecutive loads).
11000 if (StoredVal.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
11003 // Find a legal type for the vector store.
11004 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
11005 if (TLI.isTypeLegal(Ty) &&
11006 allowableAlignment(DAG, TLI, Ty, FirstStoreAS, FirstStoreAlign))
11010 return MergeStoresOfConstantsOrVecElts(StoreNodes, MemVT, NumElem,
11014 // Below we handle the case of multiple consecutive stores that
11015 // come from multiple consecutive loads. We merge them into a single
11016 // wide load and a single wide store.
11018 // Look for load nodes which are used by the stored values.
11019 SmallVector<MemOpLink, 8> LoadNodes;
11021 // Find acceptable loads. Loads need to have the same chain (token factor),
11022 // must not be zext, volatile, indexed, and they must be consecutive.
11023 BaseIndexOffset LdBasePtr;
11024 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
11025 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
11026 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
11029 // Loads must only have one use.
11030 if (!Ld->hasNUsesOfValue(1, 0))
11033 // The memory operands must not be volatile.
11034 if (Ld->isVolatile() || Ld->isIndexed())
11037 // We do not accept ext loads.
11038 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
11041 // The stored memory type must be the same.
11042 if (Ld->getMemoryVT() != MemVT)
11045 BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld->getBasePtr());
11046 // If this is not the first ptr that we check.
11047 if (LdBasePtr.Base.getNode()) {
11048 // The base ptr must be the same.
11049 if (!LdPtr.equalBaseIndex(LdBasePtr))
11052 // Check that all other base pointers are the same as this one.
11056 // We found a potential memory operand to merge.
11057 LoadNodes.push_back(MemOpLink(Ld, LdPtr.Offset, 0));
11060 if (LoadNodes.size() < 2)
11063 // If we have load/store pair instructions and we only have two values,
11065 unsigned RequiredAlignment;
11066 if (LoadNodes.size() == 2 && TLI.hasPairedLoad(MemVT, RequiredAlignment) &&
11067 St->getAlignment() >= RequiredAlignment)
11070 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
11071 unsigned FirstLoadAS = FirstLoad->getAddressSpace();
11072 unsigned FirstLoadAlign = FirstLoad->getAlignment();
11074 // Scan the memory operations on the chain and find the first non-consecutive
11075 // load memory address. These variables hold the index in the store node
11077 unsigned LastConsecutiveLoad = 0;
11078 // This variable refers to the size and not index in the array.
11079 unsigned LastLegalVectorType = 0;
11080 unsigned LastLegalIntegerType = 0;
11081 StartAddress = LoadNodes[0].OffsetFromBase;
11082 SDValue FirstChain = FirstLoad->getChain();
11083 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
11084 // All loads much share the same chain.
11085 if (LoadNodes[i].MemNode->getChain() != FirstChain)
11088 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
11089 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
11091 LastConsecutiveLoad = i;
11093 // Find a legal type for the vector store.
11094 EVT StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
11095 if (TLI.isTypeLegal(StoreTy) &&
11096 allowableAlignment(DAG, TLI, StoreTy, FirstStoreAS, FirstStoreAlign) &&
11097 allowableAlignment(DAG, TLI, StoreTy, FirstLoadAS, FirstLoadAlign)) {
11098 LastLegalVectorType = i + 1;
11101 // Find a legal type for the integer store.
11102 unsigned SizeInBits = (i+1) * ElementSizeBytes * 8;
11103 StoreTy = EVT::getIntegerVT(*DAG.getContext(), SizeInBits);
11104 if (TLI.isTypeLegal(StoreTy) &&
11105 allowableAlignment(DAG, TLI, StoreTy, FirstStoreAS, FirstStoreAlign) &&
11106 allowableAlignment(DAG, TLI, StoreTy, FirstLoadAS, FirstLoadAlign))
11107 LastLegalIntegerType = i + 1;
11108 // Or check whether a truncstore and extload is legal.
11109 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
11110 TargetLowering::TypePromoteInteger) {
11111 EVT LegalizedStoredValueTy =
11112 TLI.getTypeToTransformTo(*DAG.getContext(), StoreTy);
11113 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
11114 TLI.isLoadExtLegal(ISD::ZEXTLOAD, LegalizedStoredValueTy, StoreTy) &&
11115 TLI.isLoadExtLegal(ISD::SEXTLOAD, LegalizedStoredValueTy, StoreTy) &&
11116 TLI.isLoadExtLegal(ISD::EXTLOAD, LegalizedStoredValueTy, StoreTy) &&
11117 allowableAlignment(DAG, TLI, LegalizedStoredValueTy, FirstStoreAS,
11118 FirstStoreAlign) &&
11119 allowableAlignment(DAG, TLI, LegalizedStoredValueTy, FirstLoadAS,
11121 LastLegalIntegerType = i+1;
11125 // Only use vector types if the vector type is larger than the integer type.
11126 // If they are the same, use integers.
11127 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors;
11128 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType);
11130 // We add +1 here because the LastXXX variables refer to location while
11131 // the NumElem refers to array/index size.
11132 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1;
11133 NumElem = std::min(LastLegalType, NumElem);
11138 // The latest Node in the DAG.
11139 unsigned LatestNodeUsed = 0;
11140 for (unsigned i=1; i<NumElem; ++i) {
11141 // Find a chain for the new wide-store operand. Notice that some
11142 // of the store nodes that we found may not be selected for inclusion
11143 // in the wide store. The chain we use needs to be the chain of the
11144 // latest store node which is *used* and replaced by the wide store.
11145 if (StoreNodes[i].SequenceNum < StoreNodes[LatestNodeUsed].SequenceNum)
11146 LatestNodeUsed = i;
11149 LSBaseSDNode *LatestOp = StoreNodes[LatestNodeUsed].MemNode;
11151 // Find if it is better to use vectors or integers to load and store
11155 JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
11157 unsigned SizeInBits = NumElem * ElementSizeBytes * 8;
11158 JointMemOpVT = EVT::getIntegerVT(*DAG.getContext(), SizeInBits);
11161 SDLoc LoadDL(LoadNodes[0].MemNode);
11162 SDLoc StoreDL(StoreNodes[0].MemNode);
11164 SDValue NewLoad = DAG.getLoad(
11165 JointMemOpVT, LoadDL, FirstLoad->getChain(), FirstLoad->getBasePtr(),
11166 FirstLoad->getPointerInfo(), false, false, false, FirstLoadAlign);
11168 SDValue NewStore = DAG.getStore(
11169 LatestOp->getChain(), StoreDL, NewLoad, FirstInChain->getBasePtr(),
11170 FirstInChain->getPointerInfo(), false, false, FirstStoreAlign);
11172 // Replace one of the loads with the new load.
11173 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode);
11174 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
11175 SDValue(NewLoad.getNode(), 1));
11177 // Remove the rest of the load chains.
11178 for (unsigned i = 1; i < NumElem ; ++i) {
11179 // Replace all chain users of the old load nodes with the chain of the new
11181 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
11182 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain());
11185 // Replace the last store with the new store.
11186 CombineTo(LatestOp, NewStore);
11187 // Erase all other stores.
11188 for (unsigned i = 0; i < NumElem ; ++i) {
11189 // Remove all Store nodes.
11190 if (StoreNodes[i].MemNode == LatestOp)
11192 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
11193 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain());
11194 deleteAndRecombine(St);
11200 SDValue DAGCombiner::visitSTORE(SDNode *N) {
11201 StoreSDNode *ST = cast<StoreSDNode>(N);
11202 SDValue Chain = ST->getChain();
11203 SDValue Value = ST->getValue();
11204 SDValue Ptr = ST->getBasePtr();
11206 // If this is a store of a bit convert, store the input value if the
11207 // resultant store does not need a higher alignment than the original.
11208 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
11209 ST->isUnindexed()) {
11210 unsigned OrigAlign = ST->getAlignment();
11211 EVT SVT = Value.getOperand(0).getValueType();
11212 unsigned Align = DAG.getDataLayout().getABITypeAlignment(
11213 SVT.getTypeForEVT(*DAG.getContext()));
11214 if (Align <= OrigAlign &&
11215 ((!LegalOperations && !ST->isVolatile()) ||
11216 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
11217 return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0),
11218 Ptr, ST->getPointerInfo(), ST->isVolatile(),
11219 ST->isNonTemporal(), OrigAlign,
11223 // Turn 'store undef, Ptr' -> nothing.
11224 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
11227 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
11228 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
11229 // NOTE: If the original store is volatile, this transform must not increase
11230 // the number of stores. For example, on x86-32 an f64 can be stored in one
11231 // processor operation but an i64 (which is not legal) requires two. So the
11232 // transform should not be done in this case.
11233 if (Value.getOpcode() != ISD::TargetConstantFP) {
11235 switch (CFP->getSimpleValueType(0).SimpleTy) {
11236 default: llvm_unreachable("Unknown FP type");
11237 case MVT::f16: // We don't do this for these yet.
11243 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
11244 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
11246 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
11247 bitcastToAPInt().getZExtValue(), SDLoc(CFP),
11249 return DAG.getStore(Chain, SDLoc(N), Tmp,
11250 Ptr, ST->getMemOperand());
11254 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
11255 !ST->isVolatile()) ||
11256 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
11258 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
11259 getZExtValue(), SDLoc(CFP), MVT::i64);
11260 return DAG.getStore(Chain, SDLoc(N), Tmp,
11261 Ptr, ST->getMemOperand());
11264 if (!ST->isVolatile() &&
11265 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
11266 // Many FP stores are not made apparent until after legalize, e.g. for
11267 // argument passing. Since this is so common, custom legalize the
11268 // 64-bit integer store into two 32-bit stores.
11269 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
11270 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, SDLoc(CFP), MVT::i32);
11271 SDValue Hi = DAG.getConstant(Val >> 32, SDLoc(CFP), MVT::i32);
11272 if (DAG.getDataLayout().isBigEndian())
11275 unsigned Alignment = ST->getAlignment();
11276 bool isVolatile = ST->isVolatile();
11277 bool isNonTemporal = ST->isNonTemporal();
11278 AAMDNodes AAInfo = ST->getAAInfo();
11282 SDValue St0 = DAG.getStore(Chain, SDLoc(ST), Lo,
11283 Ptr, ST->getPointerInfo(),
11284 isVolatile, isNonTemporal,
11285 ST->getAlignment(), AAInfo);
11286 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
11287 DAG.getConstant(4, DL, Ptr.getValueType()));
11288 Alignment = MinAlign(Alignment, 4U);
11289 SDValue St1 = DAG.getStore(Chain, SDLoc(ST), Hi,
11290 Ptr, ST->getPointerInfo().getWithOffset(4),
11291 isVolatile, isNonTemporal,
11292 Alignment, AAInfo);
11293 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
11302 // Try to infer better alignment information than the store already has.
11303 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
11304 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
11305 if (Align > ST->getAlignment()) {
11307 DAG.getTruncStore(Chain, SDLoc(N), Value,
11308 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
11309 ST->isVolatile(), ST->isNonTemporal(), Align,
11311 if (NewStore.getNode() != N)
11312 return CombineTo(ST, NewStore, true);
11317 // Try transforming a pair floating point load / store ops to integer
11318 // load / store ops.
11319 SDValue NewST = TransformFPLoadStorePair(N);
11320 if (NewST.getNode())
11323 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
11324 : DAG.getSubtarget().useAA();
11326 if (CombinerAAOnlyFunc.getNumOccurrences() &&
11327 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
11330 if (UseAA && ST->isUnindexed()) {
11331 // Walk up chain skipping non-aliasing memory nodes.
11332 SDValue BetterChain = FindBetterChain(N, Chain);
11334 // If there is a better chain.
11335 if (Chain != BetterChain) {
11338 // Replace the chain to avoid dependency.
11339 if (ST->isTruncatingStore()) {
11340 ReplStore = DAG.getTruncStore(BetterChain, SDLoc(N), Value, Ptr,
11341 ST->getMemoryVT(), ST->getMemOperand());
11343 ReplStore = DAG.getStore(BetterChain, SDLoc(N), Value, Ptr,
11344 ST->getMemOperand());
11347 // Create token to keep both nodes around.
11348 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
11349 MVT::Other, Chain, ReplStore);
11351 // Make sure the new and old chains are cleaned up.
11352 AddToWorklist(Token.getNode());
11354 // Don't add users to work list.
11355 return CombineTo(N, Token, false);
11359 // Try transforming N to an indexed store.
11360 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
11361 return SDValue(N, 0);
11363 // FIXME: is there such a thing as a truncating indexed store?
11364 if (ST->isTruncatingStore() && ST->isUnindexed() &&
11365 Value.getValueType().isInteger()) {
11366 // See if we can simplify the input to this truncstore with knowledge that
11367 // only the low bits are being used. For example:
11368 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
11370 GetDemandedBits(Value,
11371 APInt::getLowBitsSet(
11372 Value.getValueType().getScalarType().getSizeInBits(),
11373 ST->getMemoryVT().getScalarType().getSizeInBits()));
11374 AddToWorklist(Value.getNode());
11375 if (Shorter.getNode())
11376 return DAG.getTruncStore(Chain, SDLoc(N), Shorter,
11377 Ptr, ST->getMemoryVT(), ST->getMemOperand());
11379 // Otherwise, see if we can simplify the operation with
11380 // SimplifyDemandedBits, which only works if the value has a single use.
11381 if (SimplifyDemandedBits(Value,
11382 APInt::getLowBitsSet(
11383 Value.getValueType().getScalarType().getSizeInBits(),
11384 ST->getMemoryVT().getScalarType().getSizeInBits())))
11385 return SDValue(N, 0);
11388 // If this is a load followed by a store to the same location, then the store
11390 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
11391 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
11392 ST->isUnindexed() && !ST->isVolatile() &&
11393 // There can't be any side effects between the load and store, such as
11394 // a call or store.
11395 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
11396 // The store is dead, remove it.
11401 // If this is a store followed by a store with the same value to the same
11402 // location, then the store is dead/noop.
11403 if (StoreSDNode *ST1 = dyn_cast<StoreSDNode>(Chain)) {
11404 if (ST1->getBasePtr() == Ptr && ST->getMemoryVT() == ST1->getMemoryVT() &&
11405 ST1->getValue() == Value && ST->isUnindexed() && !ST->isVolatile() &&
11406 ST1->isUnindexed() && !ST1->isVolatile()) {
11407 // The store is dead, remove it.
11412 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
11413 // truncating store. We can do this even if this is already a truncstore.
11414 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
11415 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
11416 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
11417 ST->getMemoryVT())) {
11418 return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0),
11419 Ptr, ST->getMemoryVT(), ST->getMemOperand());
11422 // Only perform this optimization before the types are legal, because we
11423 // don't want to perform this optimization on every DAGCombine invocation.
11425 bool EverChanged = false;
11428 // There can be multiple store sequences on the same chain.
11429 // Keep trying to merge store sequences until we are unable to do so
11430 // or until we merge the last store on the chain.
11431 bool Changed = MergeConsecutiveStores(ST);
11432 EverChanged |= Changed;
11433 if (!Changed) break;
11434 } while (ST->getOpcode() != ISD::DELETED_NODE);
11437 return SDValue(N, 0);
11440 return ReduceLoadOpStoreWidth(N);
11443 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
11444 SDValue InVec = N->getOperand(0);
11445 SDValue InVal = N->getOperand(1);
11446 SDValue EltNo = N->getOperand(2);
11449 // If the inserted element is an UNDEF, just use the input vector.
11450 if (InVal.getOpcode() == ISD::UNDEF)
11453 EVT VT = InVec.getValueType();
11455 // If we can't generate a legal BUILD_VECTOR, exit
11456 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
11459 // Check that we know which element is being inserted
11460 if (!isa<ConstantSDNode>(EltNo))
11462 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
11464 // Canonicalize insert_vector_elt dag nodes.
11466 // (insert_vector_elt (insert_vector_elt A, Idx0), Idx1)
11467 // -> (insert_vector_elt (insert_vector_elt A, Idx1), Idx0)
11469 // Do this only if the child insert_vector node has one use; also
11470 // do this only if indices are both constants and Idx1 < Idx0.
11471 if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT && InVec.hasOneUse()
11472 && isa<ConstantSDNode>(InVec.getOperand(2))) {
11473 unsigned OtherElt =
11474 cast<ConstantSDNode>(InVec.getOperand(2))->getZExtValue();
11475 if (Elt < OtherElt) {
11477 SDValue NewOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(N), VT,
11478 InVec.getOperand(0), InVal, EltNo);
11479 AddToWorklist(NewOp.getNode());
11480 return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(InVec.getNode()),
11481 VT, NewOp, InVec.getOperand(1), InVec.getOperand(2));
11485 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
11486 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
11487 // vector elements.
11488 SmallVector<SDValue, 8> Ops;
11489 // Do not combine these two vectors if the output vector will not replace
11490 // the input vector.
11491 if (InVec.getOpcode() == ISD::BUILD_VECTOR && InVec.hasOneUse()) {
11492 Ops.append(InVec.getNode()->op_begin(),
11493 InVec.getNode()->op_end());
11494 } else if (InVec.getOpcode() == ISD::UNDEF) {
11495 unsigned NElts = VT.getVectorNumElements();
11496 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
11501 // Insert the element
11502 if (Elt < Ops.size()) {
11503 // All the operands of BUILD_VECTOR must have the same type;
11504 // we enforce that here.
11505 EVT OpVT = Ops[0].getValueType();
11506 if (InVal.getValueType() != OpVT)
11507 InVal = OpVT.bitsGT(InVal.getValueType()) ?
11508 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
11509 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
11513 // Return the new vector
11514 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
11517 SDValue DAGCombiner::ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
11518 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad) {
11519 EVT ResultVT = EVE->getValueType(0);
11520 EVT VecEltVT = InVecVT.getVectorElementType();
11521 unsigned Align = OriginalLoad->getAlignment();
11522 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
11523 VecEltVT.getTypeForEVT(*DAG.getContext()));
11525 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VecEltVT))
11530 SDValue NewPtr = OriginalLoad->getBasePtr();
11532 EVT PtrType = NewPtr.getValueType();
11533 MachinePointerInfo MPI;
11535 if (auto *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo)) {
11536 int Elt = ConstEltNo->getZExtValue();
11537 unsigned PtrOff = VecEltVT.getSizeInBits() * Elt / 8;
11538 Offset = DAG.getConstant(PtrOff, DL, PtrType);
11539 MPI = OriginalLoad->getPointerInfo().getWithOffset(PtrOff);
11541 Offset = DAG.getZExtOrTrunc(EltNo, DL, PtrType);
11542 Offset = DAG.getNode(
11543 ISD::MUL, DL, PtrType, Offset,
11544 DAG.getConstant(VecEltVT.getStoreSize(), DL, PtrType));
11545 MPI = OriginalLoad->getPointerInfo();
11547 NewPtr = DAG.getNode(ISD::ADD, DL, PtrType, NewPtr, Offset);
11549 // The replacement we need to do here is a little tricky: we need to
11550 // replace an extractelement of a load with a load.
11551 // Use ReplaceAllUsesOfValuesWith to do the replacement.
11552 // Note that this replacement assumes that the extractvalue is the only
11553 // use of the load; that's okay because we don't want to perform this
11554 // transformation in other cases anyway.
11557 if (ResultVT.bitsGT(VecEltVT)) {
11558 // If the result type of vextract is wider than the load, then issue an
11559 // extending load instead.
11560 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, ResultVT,
11564 Load = DAG.getExtLoad(
11565 ExtType, SDLoc(EVE), ResultVT, OriginalLoad->getChain(), NewPtr, MPI,
11566 VecEltVT, OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
11567 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
11568 Chain = Load.getValue(1);
11570 Load = DAG.getLoad(
11571 VecEltVT, SDLoc(EVE), OriginalLoad->getChain(), NewPtr, MPI,
11572 OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
11573 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
11574 Chain = Load.getValue(1);
11575 if (ResultVT.bitsLT(VecEltVT))
11576 Load = DAG.getNode(ISD::TRUNCATE, SDLoc(EVE), ResultVT, Load);
11578 Load = DAG.getNode(ISD::BITCAST, SDLoc(EVE), ResultVT, Load);
11580 WorklistRemover DeadNodes(*this);
11581 SDValue From[] = { SDValue(EVE, 0), SDValue(OriginalLoad, 1) };
11582 SDValue To[] = { Load, Chain };
11583 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
11584 // Since we're explicitly calling ReplaceAllUses, add the new node to the
11585 // worklist explicitly as well.
11586 AddToWorklist(Load.getNode());
11587 AddUsersToWorklist(Load.getNode()); // Add users too
11588 // Make sure to revisit this node to clean it up; it will usually be dead.
11589 AddToWorklist(EVE);
11591 return SDValue(EVE, 0);
11594 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
11595 // (vextract (scalar_to_vector val, 0) -> val
11596 SDValue InVec = N->getOperand(0);
11597 EVT VT = InVec.getValueType();
11598 EVT NVT = N->getValueType(0);
11600 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
11601 // Check if the result type doesn't match the inserted element type. A
11602 // SCALAR_TO_VECTOR may truncate the inserted element and the
11603 // EXTRACT_VECTOR_ELT may widen the extracted vector.
11604 SDValue InOp = InVec.getOperand(0);
11605 if (InOp.getValueType() != NVT) {
11606 assert(InOp.getValueType().isInteger() && NVT.isInteger());
11607 return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT);
11612 SDValue EltNo = N->getOperand(1);
11613 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
11615 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
11616 // We only perform this optimization before the op legalization phase because
11617 // we may introduce new vector instructions which are not backed by TD
11618 // patterns. For example on AVX, extracting elements from a wide vector
11619 // without using extract_subvector. However, if we can find an underlying
11620 // scalar value, then we can always use that.
11621 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
11623 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
11624 int NumElem = VT.getVectorNumElements();
11625 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
11626 // Find the new index to extract from.
11627 int OrigElt = SVOp->getMaskElt(Elt);
11629 // Extracting an undef index is undef.
11631 return DAG.getUNDEF(NVT);
11633 // Select the right vector half to extract from.
11635 if (OrigElt < NumElem) {
11636 SVInVec = InVec->getOperand(0);
11638 SVInVec = InVec->getOperand(1);
11639 OrigElt -= NumElem;
11642 if (SVInVec.getOpcode() == ISD::BUILD_VECTOR) {
11643 SDValue InOp = SVInVec.getOperand(OrigElt);
11644 if (InOp.getValueType() != NVT) {
11645 assert(InOp.getValueType().isInteger() && NVT.isInteger());
11646 InOp = DAG.getSExtOrTrunc(InOp, SDLoc(SVInVec), NVT);
11652 // FIXME: We should handle recursing on other vector shuffles and
11653 // scalar_to_vector here as well.
11655 if (!LegalOperations) {
11656 EVT IndexTy = TLI.getVectorIdxTy(DAG.getDataLayout());
11657 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT, SVInVec,
11658 DAG.getConstant(OrigElt, SDLoc(SVOp), IndexTy));
11662 bool BCNumEltsChanged = false;
11663 EVT ExtVT = VT.getVectorElementType();
11666 // If the result of load has to be truncated, then it's not necessarily
11668 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
11671 if (InVec.getOpcode() == ISD::BITCAST) {
11672 // Don't duplicate a load with other uses.
11673 if (!InVec.hasOneUse())
11676 EVT BCVT = InVec.getOperand(0).getValueType();
11677 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
11679 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
11680 BCNumEltsChanged = true;
11681 InVec = InVec.getOperand(0);
11682 ExtVT = BCVT.getVectorElementType();
11685 // (vextract (vN[if]M load $addr), i) -> ([if]M load $addr + i * size)
11686 if (!LegalOperations && !ConstEltNo && InVec.hasOneUse() &&
11687 ISD::isNormalLoad(InVec.getNode()) &&
11688 !N->getOperand(1)->hasPredecessor(InVec.getNode())) {
11689 SDValue Index = N->getOperand(1);
11690 if (LoadSDNode *OrigLoad = dyn_cast<LoadSDNode>(InVec))
11691 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, Index,
11695 // Perform only after legalization to ensure build_vector / vector_shuffle
11696 // optimizations have already been done.
11697 if (!LegalOperations) return SDValue();
11699 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
11700 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
11701 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
11704 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
11706 LoadSDNode *LN0 = nullptr;
11707 const ShuffleVectorSDNode *SVN = nullptr;
11708 if (ISD::isNormalLoad(InVec.getNode())) {
11709 LN0 = cast<LoadSDNode>(InVec);
11710 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
11711 InVec.getOperand(0).getValueType() == ExtVT &&
11712 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
11713 // Don't duplicate a load with other uses.
11714 if (!InVec.hasOneUse())
11717 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
11718 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
11719 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
11721 // (load $addr+1*size)
11723 // Don't duplicate a load with other uses.
11724 if (!InVec.hasOneUse())
11727 // If the bit convert changed the number of elements, it is unsafe
11728 // to examine the mask.
11729 if (BCNumEltsChanged)
11732 // Select the input vector, guarding against out of range extract vector.
11733 unsigned NumElems = VT.getVectorNumElements();
11734 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
11735 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
11737 if (InVec.getOpcode() == ISD::BITCAST) {
11738 // Don't duplicate a load with other uses.
11739 if (!InVec.hasOneUse())
11742 InVec = InVec.getOperand(0);
11744 if (ISD::isNormalLoad(InVec.getNode())) {
11745 LN0 = cast<LoadSDNode>(InVec);
11746 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
11747 EltNo = DAG.getConstant(Elt, SDLoc(EltNo), EltNo.getValueType());
11751 // Make sure we found a non-volatile load and the extractelement is
11753 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
11756 // If Idx was -1 above, Elt is going to be -1, so just return undef.
11758 return DAG.getUNDEF(LVT);
11760 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, EltNo, LN0);
11766 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
11767 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
11768 // We perform this optimization post type-legalization because
11769 // the type-legalizer often scalarizes integer-promoted vectors.
11770 // Performing this optimization before may create bit-casts which
11771 // will be type-legalized to complex code sequences.
11772 // We perform this optimization only before the operation legalizer because we
11773 // may introduce illegal operations.
11774 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
11777 unsigned NumInScalars = N->getNumOperands();
11779 EVT VT = N->getValueType(0);
11781 // Check to see if this is a BUILD_VECTOR of a bunch of values
11782 // which come from any_extend or zero_extend nodes. If so, we can create
11783 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
11784 // optimizations. We do not handle sign-extend because we can't fill the sign
11786 EVT SourceType = MVT::Other;
11787 bool AllAnyExt = true;
11789 for (unsigned i = 0; i != NumInScalars; ++i) {
11790 SDValue In = N->getOperand(i);
11791 // Ignore undef inputs.
11792 if (In.getOpcode() == ISD::UNDEF) continue;
11794 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
11795 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
11797 // Abort if the element is not an extension.
11798 if (!ZeroExt && !AnyExt) {
11799 SourceType = MVT::Other;
11803 // The input is a ZeroExt or AnyExt. Check the original type.
11804 EVT InTy = In.getOperand(0).getValueType();
11806 // Check that all of the widened source types are the same.
11807 if (SourceType == MVT::Other)
11810 else if (InTy != SourceType) {
11811 // Multiple income types. Abort.
11812 SourceType = MVT::Other;
11816 // Check if all of the extends are ANY_EXTENDs.
11817 AllAnyExt &= AnyExt;
11820 // In order to have valid types, all of the inputs must be extended from the
11821 // same source type and all of the inputs must be any or zero extend.
11822 // Scalar sizes must be a power of two.
11823 EVT OutScalarTy = VT.getScalarType();
11824 bool ValidTypes = SourceType != MVT::Other &&
11825 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
11826 isPowerOf2_32(SourceType.getSizeInBits());
11828 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
11829 // turn into a single shuffle instruction.
11833 bool isLE = DAG.getDataLayout().isLittleEndian();
11834 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
11835 assert(ElemRatio > 1 && "Invalid element size ratio");
11836 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
11837 DAG.getConstant(0, SDLoc(N), SourceType);
11839 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
11840 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
11842 // Populate the new build_vector
11843 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
11844 SDValue Cast = N->getOperand(i);
11845 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
11846 Cast.getOpcode() == ISD::ZERO_EXTEND ||
11847 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
11849 if (Cast.getOpcode() == ISD::UNDEF)
11850 In = DAG.getUNDEF(SourceType);
11852 In = Cast->getOperand(0);
11853 unsigned Index = isLE ? (i * ElemRatio) :
11854 (i * ElemRatio + (ElemRatio - 1));
11856 assert(Index < Ops.size() && "Invalid index");
11860 // The type of the new BUILD_VECTOR node.
11861 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
11862 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
11863 "Invalid vector size");
11864 // Check if the new vector type is legal.
11865 if (!isTypeLegal(VecVT)) return SDValue();
11867 // Make the new BUILD_VECTOR.
11868 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
11870 // The new BUILD_VECTOR node has the potential to be further optimized.
11871 AddToWorklist(BV.getNode());
11872 // Bitcast to the desired type.
11873 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
11876 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
11877 EVT VT = N->getValueType(0);
11879 unsigned NumInScalars = N->getNumOperands();
11882 EVT SrcVT = MVT::Other;
11883 unsigned Opcode = ISD::DELETED_NODE;
11884 unsigned NumDefs = 0;
11886 for (unsigned i = 0; i != NumInScalars; ++i) {
11887 SDValue In = N->getOperand(i);
11888 unsigned Opc = In.getOpcode();
11890 if (Opc == ISD::UNDEF)
11893 // If all scalar values are floats and converted from integers.
11894 if (Opcode == ISD::DELETED_NODE &&
11895 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
11902 EVT InVT = In.getOperand(0).getValueType();
11904 // If all scalar values are typed differently, bail out. It's chosen to
11905 // simplify BUILD_VECTOR of integer types.
11906 if (SrcVT == MVT::Other)
11913 // If the vector has just one element defined, it's not worth to fold it into
11914 // a vectorized one.
11918 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
11919 && "Should only handle conversion from integer to float.");
11920 assert(SrcVT != MVT::Other && "Cannot determine source type!");
11922 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
11924 if (!TLI.isOperationLegalOrCustom(Opcode, NVT))
11927 // Just because the floating-point vector type is legal does not necessarily
11928 // mean that the corresponding integer vector type is.
11929 if (!isTypeLegal(NVT))
11932 SmallVector<SDValue, 8> Opnds;
11933 for (unsigned i = 0; i != NumInScalars; ++i) {
11934 SDValue In = N->getOperand(i);
11936 if (In.getOpcode() == ISD::UNDEF)
11937 Opnds.push_back(DAG.getUNDEF(SrcVT));
11939 Opnds.push_back(In.getOperand(0));
11941 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Opnds);
11942 AddToWorklist(BV.getNode());
11944 return DAG.getNode(Opcode, dl, VT, BV);
11947 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
11948 unsigned NumInScalars = N->getNumOperands();
11950 EVT VT = N->getValueType(0);
11952 // A vector built entirely of undefs is undef.
11953 if (ISD::allOperandsUndef(N))
11954 return DAG.getUNDEF(VT);
11956 if (SDValue V = reduceBuildVecExtToExtBuildVec(N))
11959 if (SDValue V = reduceBuildVecConvertToConvertBuildVec(N))
11962 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
11963 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
11964 // at most two distinct vectors, turn this into a shuffle node.
11966 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
11967 if (!isTypeLegal(VT))
11970 // May only combine to shuffle after legalize if shuffle is legal.
11971 if (LegalOperations && !TLI.isOperationLegal(ISD::VECTOR_SHUFFLE, VT))
11974 SDValue VecIn1, VecIn2;
11975 bool UsesZeroVector = false;
11976 for (unsigned i = 0; i != NumInScalars; ++i) {
11977 SDValue Op = N->getOperand(i);
11978 // Ignore undef inputs.
11979 if (Op.getOpcode() == ISD::UNDEF) continue;
11981 // See if we can combine this build_vector into a blend with a zero vector.
11982 if (!VecIn2.getNode() && (isNullConstant(Op) || isNullFPConstant(Op))) {
11983 UsesZeroVector = true;
11987 // If this input is something other than a EXTRACT_VECTOR_ELT with a
11988 // constant index, bail out.
11989 if (Op.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
11990 !isa<ConstantSDNode>(Op.getOperand(1))) {
11991 VecIn1 = VecIn2 = SDValue(nullptr, 0);
11995 // We allow up to two distinct input vectors.
11996 SDValue ExtractedFromVec = Op.getOperand(0);
11997 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
12000 if (!VecIn1.getNode()) {
12001 VecIn1 = ExtractedFromVec;
12002 } else if (!VecIn2.getNode() && !UsesZeroVector) {
12003 VecIn2 = ExtractedFromVec;
12005 // Too many inputs.
12006 VecIn1 = VecIn2 = SDValue(nullptr, 0);
12011 // If everything is good, we can make a shuffle operation.
12012 if (VecIn1.getNode()) {
12013 unsigned InNumElements = VecIn1.getValueType().getVectorNumElements();
12014 SmallVector<int, 8> Mask;
12015 for (unsigned i = 0; i != NumInScalars; ++i) {
12016 unsigned Opcode = N->getOperand(i).getOpcode();
12017 if (Opcode == ISD::UNDEF) {
12018 Mask.push_back(-1);
12022 // Operands can also be zero.
12023 if (Opcode != ISD::EXTRACT_VECTOR_ELT) {
12024 assert(UsesZeroVector &&
12025 (Opcode == ISD::Constant || Opcode == ISD::ConstantFP) &&
12026 "Unexpected node found!");
12027 Mask.push_back(NumInScalars+i);
12031 // If extracting from the first vector, just use the index directly.
12032 SDValue Extract = N->getOperand(i);
12033 SDValue ExtVal = Extract.getOperand(1);
12034 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
12035 if (Extract.getOperand(0) == VecIn1) {
12036 Mask.push_back(ExtIndex);
12040 // Otherwise, use InIdx + InputVecSize
12041 Mask.push_back(InNumElements + ExtIndex);
12044 // Avoid introducing illegal shuffles with zero.
12045 if (UsesZeroVector && !TLI.isVectorClearMaskLegal(Mask, VT))
12048 // We can't generate a shuffle node with mismatched input and output types.
12049 // Attempt to transform a single input vector to the correct type.
12050 if ((VT != VecIn1.getValueType())) {
12051 // If the input vector type has a different base type to the output
12052 // vector type, bail out.
12053 EVT VTElemType = VT.getVectorElementType();
12054 if ((VecIn1.getValueType().getVectorElementType() != VTElemType) ||
12055 (VecIn2.getNode() &&
12056 (VecIn2.getValueType().getVectorElementType() != VTElemType)))
12059 // If the input vector is too small, widen it.
12060 // We only support widening of vectors which are half the size of the
12061 // output registers. For example XMM->YMM widening on X86 with AVX.
12062 EVT VecInT = VecIn1.getValueType();
12063 if (VecInT.getSizeInBits() * 2 == VT.getSizeInBits()) {
12064 // If we only have one small input, widen it by adding undef values.
12065 if (!VecIn2.getNode())
12066 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, VecIn1,
12067 DAG.getUNDEF(VecIn1.getValueType()));
12068 else if (VecIn1.getValueType() == VecIn2.getValueType()) {
12069 // If we have two small inputs of the same type, try to concat them.
12070 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, VecIn1, VecIn2);
12071 VecIn2 = SDValue(nullptr, 0);
12074 } else if (VecInT.getSizeInBits() == VT.getSizeInBits() * 2) {
12075 // If the input vector is too large, try to split it.
12076 // We don't support having two input vectors that are too large.
12077 // If the zero vector was used, we can not split the vector,
12078 // since we'd need 3 inputs.
12079 if (UsesZeroVector || VecIn2.getNode())
12082 if (!TLI.isExtractSubvectorCheap(VT, VT.getVectorNumElements()))
12085 // Try to replace VecIn1 with two extract_subvectors
12086 // No need to update the masks, they should still be correct.
12087 VecIn2 = DAG.getNode(
12088 ISD::EXTRACT_SUBVECTOR, dl, VT, VecIn1,
12089 DAG.getConstant(VT.getVectorNumElements(), dl,
12090 TLI.getVectorIdxTy(DAG.getDataLayout())));
12091 VecIn1 = DAG.getNode(
12092 ISD::EXTRACT_SUBVECTOR, dl, VT, VecIn1,
12093 DAG.getConstant(0, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
12098 if (UsesZeroVector)
12099 VecIn2 = VT.isInteger() ? DAG.getConstant(0, dl, VT) :
12100 DAG.getConstantFP(0.0, dl, VT);
12102 // If VecIn2 is unused then change it to undef.
12103 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
12105 // Check that we were able to transform all incoming values to the same
12107 if (VecIn2.getValueType() != VecIn1.getValueType() ||
12108 VecIn1.getValueType() != VT)
12111 // Return the new VECTOR_SHUFFLE node.
12115 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]);
12121 static SDValue combineConcatVectorOfScalars(SDNode *N, SelectionDAG &DAG) {
12122 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12123 EVT OpVT = N->getOperand(0).getValueType();
12125 // If the operands are legal vectors, leave them alone.
12126 if (TLI.isTypeLegal(OpVT))
12130 EVT VT = N->getValueType(0);
12131 SmallVector<SDValue, 8> Ops;
12133 EVT SVT = EVT::getIntegerVT(*DAG.getContext(), OpVT.getSizeInBits());
12134 SDValue ScalarUndef = DAG.getNode(ISD::UNDEF, DL, SVT);
12136 // Keep track of what we encounter.
12137 bool AnyInteger = false;
12138 bool AnyFP = false;
12139 for (const SDValue &Op : N->ops()) {
12140 if (ISD::BITCAST == Op.getOpcode() &&
12141 !Op.getOperand(0).getValueType().isVector())
12142 Ops.push_back(Op.getOperand(0));
12143 else if (ISD::UNDEF == Op.getOpcode())
12144 Ops.push_back(ScalarUndef);
12148 // Note whether we encounter an integer or floating point scalar.
12149 // If it's neither, bail out, it could be something weird like x86mmx.
12150 EVT LastOpVT = Ops.back().getValueType();
12151 if (LastOpVT.isFloatingPoint())
12153 else if (LastOpVT.isInteger())
12159 // If any of the operands is a floating point scalar bitcast to a vector,
12160 // use floating point types throughout, and bitcast everything.
12161 // Replace UNDEFs by another scalar UNDEF node, of the final desired type.
12163 SVT = EVT::getFloatingPointVT(OpVT.getSizeInBits());
12164 ScalarUndef = DAG.getNode(ISD::UNDEF, DL, SVT);
12166 for (SDValue &Op : Ops) {
12167 if (Op.getValueType() == SVT)
12169 if (Op.getOpcode() == ISD::UNDEF)
12172 Op = DAG.getNode(ISD::BITCAST, DL, SVT, Op);
12177 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SVT,
12178 VT.getSizeInBits() / SVT.getSizeInBits());
12179 return DAG.getNode(ISD::BITCAST, DL, VT,
12180 DAG.getNode(ISD::BUILD_VECTOR, DL, VecVT, Ops));
12183 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
12184 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
12185 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
12186 // inputs come from at most two distinct vectors, turn this into a shuffle
12189 // If we only have one input vector, we don't need to do any concatenation.
12190 if (N->getNumOperands() == 1)
12191 return N->getOperand(0);
12193 // Check if all of the operands are undefs.
12194 EVT VT = N->getValueType(0);
12195 if (ISD::allOperandsUndef(N))
12196 return DAG.getUNDEF(VT);
12198 // Optimize concat_vectors where all but the first of the vectors are undef.
12199 if (std::all_of(std::next(N->op_begin()), N->op_end(), [](const SDValue &Op) {
12200 return Op.getOpcode() == ISD::UNDEF;
12202 SDValue In = N->getOperand(0);
12203 assert(In.getValueType().isVector() && "Must concat vectors");
12205 // Transform: concat_vectors(scalar, undef) -> scalar_to_vector(sclr).
12206 if (In->getOpcode() == ISD::BITCAST &&
12207 !In->getOperand(0)->getValueType(0).isVector()) {
12208 SDValue Scalar = In->getOperand(0);
12210 // If the bitcast type isn't legal, it might be a trunc of a legal type;
12211 // look through the trunc so we can still do the transform:
12212 // concat_vectors(trunc(scalar), undef) -> scalar_to_vector(scalar)
12213 if (Scalar->getOpcode() == ISD::TRUNCATE &&
12214 !TLI.isTypeLegal(Scalar.getValueType()) &&
12215 TLI.isTypeLegal(Scalar->getOperand(0).getValueType()))
12216 Scalar = Scalar->getOperand(0);
12218 EVT SclTy = Scalar->getValueType(0);
12220 if (!SclTy.isFloatingPoint() && !SclTy.isInteger())
12223 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SclTy,
12224 VT.getSizeInBits() / SclTy.getSizeInBits());
12225 if (!TLI.isTypeLegal(NVT) || !TLI.isTypeLegal(Scalar.getValueType()))
12228 SDLoc dl = SDLoc(N);
12229 SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NVT, Scalar);
12230 return DAG.getNode(ISD::BITCAST, dl, VT, Res);
12234 // Fold any combination of BUILD_VECTOR or UNDEF nodes into one BUILD_VECTOR.
12235 // We have already tested above for an UNDEF only concatenation.
12236 // fold (concat_vectors (BUILD_VECTOR A, B, ...), (BUILD_VECTOR C, D, ...))
12237 // -> (BUILD_VECTOR A, B, ..., C, D, ...)
12238 auto IsBuildVectorOrUndef = [](const SDValue &Op) {
12239 return ISD::UNDEF == Op.getOpcode() || ISD::BUILD_VECTOR == Op.getOpcode();
12241 bool AllBuildVectorsOrUndefs =
12242 std::all_of(N->op_begin(), N->op_end(), IsBuildVectorOrUndef);
12243 if (AllBuildVectorsOrUndefs) {
12244 SmallVector<SDValue, 8> Opnds;
12245 EVT SVT = VT.getScalarType();
12248 if (!SVT.isFloatingPoint()) {
12249 // If BUILD_VECTOR are from built from integer, they may have different
12250 // operand types. Get the smallest type and truncate all operands to it.
12251 bool FoundMinVT = false;
12252 for (const SDValue &Op : N->ops())
12253 if (ISD::BUILD_VECTOR == Op.getOpcode()) {
12254 EVT OpSVT = Op.getOperand(0)->getValueType(0);
12255 MinVT = (!FoundMinVT || OpSVT.bitsLE(MinVT)) ? OpSVT : MinVT;
12258 assert(FoundMinVT && "Concat vector type mismatch");
12261 for (const SDValue &Op : N->ops()) {
12262 EVT OpVT = Op.getValueType();
12263 unsigned NumElts = OpVT.getVectorNumElements();
12265 if (ISD::UNDEF == Op.getOpcode())
12266 Opnds.append(NumElts, DAG.getUNDEF(MinVT));
12268 if (ISD::BUILD_VECTOR == Op.getOpcode()) {
12269 if (SVT.isFloatingPoint()) {
12270 assert(SVT == OpVT.getScalarType() && "Concat vector type mismatch");
12271 Opnds.append(Op->op_begin(), Op->op_begin() + NumElts);
12273 for (unsigned i = 0; i != NumElts; ++i)
12275 DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinVT, Op.getOperand(i)));
12280 assert(VT.getVectorNumElements() == Opnds.size() &&
12281 "Concat vector type mismatch");
12282 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
12285 // Fold CONCAT_VECTORS of only bitcast scalars (or undef) to BUILD_VECTOR.
12286 if (SDValue V = combineConcatVectorOfScalars(N, DAG))
12289 // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
12290 // nodes often generate nop CONCAT_VECTOR nodes.
12291 // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that
12292 // place the incoming vectors at the exact same location.
12293 SDValue SingleSource = SDValue();
12294 unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements();
12296 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
12297 SDValue Op = N->getOperand(i);
12299 if (Op.getOpcode() == ISD::UNDEF)
12302 // Check if this is the identity extract:
12303 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
12306 // Find the single incoming vector for the extract_subvector.
12307 if (SingleSource.getNode()) {
12308 if (Op.getOperand(0) != SingleSource)
12311 SingleSource = Op.getOperand(0);
12313 // Check the source type is the same as the type of the result.
12314 // If not, this concat may extend the vector, so we can not
12315 // optimize it away.
12316 if (SingleSource.getValueType() != N->getValueType(0))
12320 unsigned IdentityIndex = i * PartNumElem;
12321 ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1));
12322 // The extract index must be constant.
12326 // Check that we are reading from the identity index.
12327 if (CS->getZExtValue() != IdentityIndex)
12331 if (SingleSource.getNode())
12332 return SingleSource;
12337 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
12338 EVT NVT = N->getValueType(0);
12339 SDValue V = N->getOperand(0);
12341 if (V->getOpcode() == ISD::CONCAT_VECTORS) {
12343 // (extract_subvec (concat V1, V2, ...), i)
12346 // Only operand 0 is checked as 'concat' assumes all inputs of the same
12348 if (V->getOperand(0).getValueType() != NVT)
12350 unsigned Idx = N->getConstantOperandVal(1);
12351 unsigned NumElems = NVT.getVectorNumElements();
12352 assert((Idx % NumElems) == 0 &&
12353 "IDX in concat is not a multiple of the result vector length.");
12354 return V->getOperand(Idx / NumElems);
12358 if (V->getOpcode() == ISD::BITCAST)
12359 V = V.getOperand(0);
12361 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
12363 // Handle only simple case where vector being inserted and vector
12364 // being extracted are of same type, and are half size of larger vectors.
12365 EVT BigVT = V->getOperand(0).getValueType();
12366 EVT SmallVT = V->getOperand(1).getValueType();
12367 if (!NVT.bitsEq(SmallVT) || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
12370 // Only handle cases where both indexes are constants with the same type.
12371 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
12372 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
12374 if (InsIdx && ExtIdx &&
12375 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
12376 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
12378 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
12380 // indices are equal or bit offsets are equal => V1
12381 // otherwise => (extract_subvec V1, ExtIdx)
12382 if (InsIdx->getZExtValue() * SmallVT.getScalarType().getSizeInBits() ==
12383 ExtIdx->getZExtValue() * NVT.getScalarType().getSizeInBits())
12384 return DAG.getNode(ISD::BITCAST, dl, NVT, V->getOperand(1));
12385 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT,
12386 DAG.getNode(ISD::BITCAST, dl,
12387 N->getOperand(0).getValueType(),
12388 V->getOperand(0)), N->getOperand(1));
12395 static SDValue simplifyShuffleOperandRecursively(SmallBitVector &UsedElements,
12396 SDValue V, SelectionDAG &DAG) {
12398 EVT VT = V.getValueType();
12400 switch (V.getOpcode()) {
12404 case ISD::CONCAT_VECTORS: {
12405 EVT OpVT = V->getOperand(0).getValueType();
12406 int OpSize = OpVT.getVectorNumElements();
12407 SmallBitVector OpUsedElements(OpSize, false);
12408 bool FoundSimplification = false;
12409 SmallVector<SDValue, 4> NewOps;
12410 NewOps.reserve(V->getNumOperands());
12411 for (int i = 0, NumOps = V->getNumOperands(); i < NumOps; ++i) {
12412 SDValue Op = V->getOperand(i);
12413 bool OpUsed = false;
12414 for (int j = 0; j < OpSize; ++j)
12415 if (UsedElements[i * OpSize + j]) {
12416 OpUsedElements[j] = true;
12420 OpUsed ? simplifyShuffleOperandRecursively(OpUsedElements, Op, DAG)
12421 : DAG.getUNDEF(OpVT));
12422 FoundSimplification |= Op == NewOps.back();
12423 OpUsedElements.reset();
12425 if (FoundSimplification)
12426 V = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, NewOps);
12430 case ISD::INSERT_SUBVECTOR: {
12431 SDValue BaseV = V->getOperand(0);
12432 SDValue SubV = V->getOperand(1);
12433 auto *IdxN = dyn_cast<ConstantSDNode>(V->getOperand(2));
12437 int SubSize = SubV.getValueType().getVectorNumElements();
12438 int Idx = IdxN->getZExtValue();
12439 bool SubVectorUsed = false;
12440 SmallBitVector SubUsedElements(SubSize, false);
12441 for (int i = 0; i < SubSize; ++i)
12442 if (UsedElements[i + Idx]) {
12443 SubVectorUsed = true;
12444 SubUsedElements[i] = true;
12445 UsedElements[i + Idx] = false;
12448 // Now recurse on both the base and sub vectors.
12449 SDValue SimplifiedSubV =
12451 ? simplifyShuffleOperandRecursively(SubUsedElements, SubV, DAG)
12452 : DAG.getUNDEF(SubV.getValueType());
12453 SDValue SimplifiedBaseV = simplifyShuffleOperandRecursively(UsedElements, BaseV, DAG);
12454 if (SimplifiedSubV != SubV || SimplifiedBaseV != BaseV)
12455 V = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT,
12456 SimplifiedBaseV, SimplifiedSubV, V->getOperand(2));
12462 static SDValue simplifyShuffleOperands(ShuffleVectorSDNode *SVN, SDValue N0,
12463 SDValue N1, SelectionDAG &DAG) {
12464 EVT VT = SVN->getValueType(0);
12465 int NumElts = VT.getVectorNumElements();
12466 SmallBitVector N0UsedElements(NumElts, false), N1UsedElements(NumElts, false);
12467 for (int M : SVN->getMask())
12468 if (M >= 0 && M < NumElts)
12469 N0UsedElements[M] = true;
12470 else if (M >= NumElts)
12471 N1UsedElements[M - NumElts] = true;
12473 SDValue S0 = simplifyShuffleOperandRecursively(N0UsedElements, N0, DAG);
12474 SDValue S1 = simplifyShuffleOperandRecursively(N1UsedElements, N1, DAG);
12475 if (S0 == N0 && S1 == N1)
12478 return DAG.getVectorShuffle(VT, SDLoc(SVN), S0, S1, SVN->getMask());
12481 // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat,
12482 // or turn a shuffle of a single concat into simpler shuffle then concat.
12483 static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
12484 EVT VT = N->getValueType(0);
12485 unsigned NumElts = VT.getVectorNumElements();
12487 SDValue N0 = N->getOperand(0);
12488 SDValue N1 = N->getOperand(1);
12489 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
12491 SmallVector<SDValue, 4> Ops;
12492 EVT ConcatVT = N0.getOperand(0).getValueType();
12493 unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements();
12494 unsigned NumConcats = NumElts / NumElemsPerConcat;
12496 // Special case: shuffle(concat(A,B)) can be more efficiently represented
12497 // as concat(shuffle(A,B),UNDEF) if the shuffle doesn't set any of the high
12498 // half vector elements.
12499 if (NumElemsPerConcat * 2 == NumElts && N1.getOpcode() == ISD::UNDEF &&
12500 std::all_of(SVN->getMask().begin() + NumElemsPerConcat,
12501 SVN->getMask().end(), [](int i) { return i == -1; })) {
12502 N0 = DAG.getVectorShuffle(ConcatVT, SDLoc(N), N0.getOperand(0), N0.getOperand(1),
12503 ArrayRef<int>(SVN->getMask().begin(), NumElemsPerConcat));
12504 N1 = DAG.getUNDEF(ConcatVT);
12505 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, N0, N1);
12508 // Look at every vector that's inserted. We're looking for exact
12509 // subvector-sized copies from a concatenated vector
12510 for (unsigned I = 0; I != NumConcats; ++I) {
12511 // Make sure we're dealing with a copy.
12512 unsigned Begin = I * NumElemsPerConcat;
12513 bool AllUndef = true, NoUndef = true;
12514 for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) {
12515 if (SVN->getMaskElt(J) >= 0)
12522 if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0)
12525 for (unsigned J = 1; J != NumElemsPerConcat; ++J)
12526 if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J))
12529 unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat;
12530 if (FirstElt < N0.getNumOperands())
12531 Ops.push_back(N0.getOperand(FirstElt));
12533 Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands()));
12535 } else if (AllUndef) {
12536 Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType()));
12537 } else { // Mixed with general masks and undefs, can't do optimization.
12542 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops);
12545 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
12546 EVT VT = N->getValueType(0);
12547 unsigned NumElts = VT.getVectorNumElements();
12549 SDValue N0 = N->getOperand(0);
12550 SDValue N1 = N->getOperand(1);
12552 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
12554 // Canonicalize shuffle undef, undef -> undef
12555 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
12556 return DAG.getUNDEF(VT);
12558 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
12560 // Canonicalize shuffle v, v -> v, undef
12562 SmallVector<int, 8> NewMask;
12563 for (unsigned i = 0; i != NumElts; ++i) {
12564 int Idx = SVN->getMaskElt(i);
12565 if (Idx >= (int)NumElts) Idx -= NumElts;
12566 NewMask.push_back(Idx);
12568 return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT),
12572 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
12573 if (N0.getOpcode() == ISD::UNDEF) {
12574 SmallVector<int, 8> NewMask;
12575 for (unsigned i = 0; i != NumElts; ++i) {
12576 int Idx = SVN->getMaskElt(i);
12578 if (Idx >= (int)NumElts)
12581 Idx = -1; // remove reference to lhs
12583 NewMask.push_back(Idx);
12585 return DAG.getVectorShuffle(VT, SDLoc(N), N1, DAG.getUNDEF(VT),
12589 // Remove references to rhs if it is undef
12590 if (N1.getOpcode() == ISD::UNDEF) {
12591 bool Changed = false;
12592 SmallVector<int, 8> NewMask;
12593 for (unsigned i = 0; i != NumElts; ++i) {
12594 int Idx = SVN->getMaskElt(i);
12595 if (Idx >= (int)NumElts) {
12599 NewMask.push_back(Idx);
12602 return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, &NewMask[0]);
12605 // If it is a splat, check if the argument vector is another splat or a
12607 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
12608 SDNode *V = N0.getNode();
12610 // If this is a bit convert that changes the element type of the vector but
12611 // not the number of vector elements, look through it. Be careful not to
12612 // look though conversions that change things like v4f32 to v2f64.
12613 if (V->getOpcode() == ISD::BITCAST) {
12614 SDValue ConvInput = V->getOperand(0);
12615 if (ConvInput.getValueType().isVector() &&
12616 ConvInput.getValueType().getVectorNumElements() == NumElts)
12617 V = ConvInput.getNode();
12620 if (V->getOpcode() == ISD::BUILD_VECTOR) {
12621 assert(V->getNumOperands() == NumElts &&
12622 "BUILD_VECTOR has wrong number of operands");
12624 bool AllSame = true;
12625 for (unsigned i = 0; i != NumElts; ++i) {
12626 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
12627 Base = V->getOperand(i);
12631 // Splat of <u, u, u, u>, return <u, u, u, u>
12632 if (!Base.getNode())
12634 for (unsigned i = 0; i != NumElts; ++i) {
12635 if (V->getOperand(i) != Base) {
12640 // Splat of <x, x, x, x>, return <x, x, x, x>
12644 // Canonicalize any other splat as a build_vector.
12645 const SDValue &Splatted = V->getOperand(SVN->getSplatIndex());
12646 SmallVector<SDValue, 8> Ops(NumElts, Splatted);
12647 SDValue NewBV = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
12648 V->getValueType(0), Ops);
12650 // We may have jumped through bitcasts, so the type of the
12651 // BUILD_VECTOR may not match the type of the shuffle.
12652 if (V->getValueType(0) != VT)
12653 NewBV = DAG.getNode(ISD::BITCAST, SDLoc(N), VT, NewBV);
12658 // There are various patterns used to build up a vector from smaller vectors,
12659 // subvectors, or elements. Scan chains of these and replace unused insertions
12660 // or components with undef.
12661 if (SDValue S = simplifyShuffleOperands(SVN, N0, N1, DAG))
12664 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
12665 Level < AfterLegalizeVectorOps &&
12666 (N1.getOpcode() == ISD::UNDEF ||
12667 (N1.getOpcode() == ISD::CONCAT_VECTORS &&
12668 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) {
12669 SDValue V = partitionShuffleOfConcats(N, DAG);
12675 // Attempt to combine a shuffle of 2 inputs of 'scalar sources' -
12676 // BUILD_VECTOR or SCALAR_TO_VECTOR into a single BUILD_VECTOR.
12677 if (Level < AfterLegalizeVectorOps && TLI.isTypeLegal(VT)) {
12678 SmallVector<SDValue, 8> Ops;
12679 for (int M : SVN->getMask()) {
12680 SDValue Op = DAG.getUNDEF(VT.getScalarType());
12682 int Idx = M % NumElts;
12683 SDValue &S = (M < (int)NumElts ? N0 : N1);
12684 if (S.getOpcode() == ISD::BUILD_VECTOR && S.hasOneUse()) {
12685 Op = S.getOperand(Idx);
12686 } else if (S.getOpcode() == ISD::SCALAR_TO_VECTOR && S.hasOneUse()) {
12688 Op = S.getOperand(0);
12690 // Operand can't be combined - bail out.
12696 if (Ops.size() == VT.getVectorNumElements()) {
12697 // BUILD_VECTOR requires all inputs to be of the same type, find the
12698 // maximum type and extend them all.
12699 EVT SVT = VT.getScalarType();
12700 if (SVT.isInteger())
12701 for (SDValue &Op : Ops)
12702 SVT = (SVT.bitsLT(Op.getValueType()) ? Op.getValueType() : SVT);
12703 if (SVT != VT.getScalarType())
12704 for (SDValue &Op : Ops)
12705 Op = TLI.isZExtFree(Op.getValueType(), SVT)
12706 ? DAG.getZExtOrTrunc(Op, SDLoc(N), SVT)
12707 : DAG.getSExtOrTrunc(Op, SDLoc(N), SVT);
12708 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Ops);
12712 // If this shuffle only has a single input that is a bitcasted shuffle,
12713 // attempt to merge the 2 shuffles and suitably bitcast the inputs/output
12714 // back to their original types.
12715 if (N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
12716 N1.getOpcode() == ISD::UNDEF && Level < AfterLegalizeVectorOps &&
12717 TLI.isTypeLegal(VT)) {
12719 // Peek through the bitcast only if there is one user.
12721 while (BC0.getOpcode() == ISD::BITCAST) {
12722 if (!BC0.hasOneUse())
12724 BC0 = BC0.getOperand(0);
12727 auto ScaleShuffleMask = [](ArrayRef<int> Mask, int Scale) {
12729 return SmallVector<int, 8>(Mask.begin(), Mask.end());
12731 SmallVector<int, 8> NewMask;
12733 for (int s = 0; s != Scale; ++s)
12734 NewMask.push_back(M < 0 ? -1 : Scale * M + s);
12738 if (BC0.getOpcode() == ISD::VECTOR_SHUFFLE && BC0.hasOneUse()) {
12739 EVT SVT = VT.getScalarType();
12740 EVT InnerVT = BC0->getValueType(0);
12741 EVT InnerSVT = InnerVT.getScalarType();
12743 // Determine which shuffle works with the smaller scalar type.
12744 EVT ScaleVT = SVT.bitsLT(InnerSVT) ? VT : InnerVT;
12745 EVT ScaleSVT = ScaleVT.getScalarType();
12747 if (TLI.isTypeLegal(ScaleVT) &&
12748 0 == (InnerSVT.getSizeInBits() % ScaleSVT.getSizeInBits()) &&
12749 0 == (SVT.getSizeInBits() % ScaleSVT.getSizeInBits())) {
12751 int InnerScale = InnerSVT.getSizeInBits() / ScaleSVT.getSizeInBits();
12752 int OuterScale = SVT.getSizeInBits() / ScaleSVT.getSizeInBits();
12754 // Scale the shuffle masks to the smaller scalar type.
12755 ShuffleVectorSDNode *InnerSVN = cast<ShuffleVectorSDNode>(BC0);
12756 SmallVector<int, 8> InnerMask =
12757 ScaleShuffleMask(InnerSVN->getMask(), InnerScale);
12758 SmallVector<int, 8> OuterMask =
12759 ScaleShuffleMask(SVN->getMask(), OuterScale);
12761 // Merge the shuffle masks.
12762 SmallVector<int, 8> NewMask;
12763 for (int M : OuterMask)
12764 NewMask.push_back(M < 0 ? -1 : InnerMask[M]);
12766 // Test for shuffle mask legality over both commutations.
12767 SDValue SV0 = BC0->getOperand(0);
12768 SDValue SV1 = BC0->getOperand(1);
12769 bool LegalMask = TLI.isShuffleMaskLegal(NewMask, ScaleVT);
12771 std::swap(SV0, SV1);
12772 ShuffleVectorSDNode::commuteMask(NewMask);
12773 LegalMask = TLI.isShuffleMaskLegal(NewMask, ScaleVT);
12777 SV0 = DAG.getNode(ISD::BITCAST, SDLoc(N), ScaleVT, SV0);
12778 SV1 = DAG.getNode(ISD::BITCAST, SDLoc(N), ScaleVT, SV1);
12779 return DAG.getNode(
12780 ISD::BITCAST, SDLoc(N), VT,
12781 DAG.getVectorShuffle(ScaleVT, SDLoc(N), SV0, SV1, NewMask));
12787 // Canonicalize shuffles according to rules:
12788 // shuffle(A, shuffle(A, B)) -> shuffle(shuffle(A,B), A)
12789 // shuffle(B, shuffle(A, B)) -> shuffle(shuffle(A,B), B)
12790 // shuffle(B, shuffle(A, Undef)) -> shuffle(shuffle(A, Undef), B)
12791 if (N1.getOpcode() == ISD::VECTOR_SHUFFLE &&
12792 N0.getOpcode() != ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
12793 TLI.isTypeLegal(VT)) {
12794 // The incoming shuffle must be of the same type as the result of the
12795 // current shuffle.
12796 assert(N1->getOperand(0).getValueType() == VT &&
12797 "Shuffle types don't match");
12799 SDValue SV0 = N1->getOperand(0);
12800 SDValue SV1 = N1->getOperand(1);
12801 bool HasSameOp0 = N0 == SV0;
12802 bool IsSV1Undef = SV1.getOpcode() == ISD::UNDEF;
12803 if (HasSameOp0 || IsSV1Undef || N0 == SV1)
12804 // Commute the operands of this shuffle so that next rule
12806 return DAG.getCommutedVectorShuffle(*SVN);
12809 // Try to fold according to rules:
12810 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2)
12811 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2)
12812 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2)
12813 // Don't try to fold shuffles with illegal type.
12814 // Only fold if this shuffle is the only user of the other shuffle.
12815 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && N->isOnlyUserOf(N0.getNode()) &&
12816 Level < AfterLegalizeDAG && TLI.isTypeLegal(VT)) {
12817 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
12819 // The incoming shuffle must be of the same type as the result of the
12820 // current shuffle.
12821 assert(OtherSV->getOperand(0).getValueType() == VT &&
12822 "Shuffle types don't match");
12825 SmallVector<int, 4> Mask;
12826 // Compute the combined shuffle mask for a shuffle with SV0 as the first
12827 // operand, and SV1 as the second operand.
12828 for (unsigned i = 0; i != NumElts; ++i) {
12829 int Idx = SVN->getMaskElt(i);
12831 // Propagate Undef.
12832 Mask.push_back(Idx);
12836 SDValue CurrentVec;
12837 if (Idx < (int)NumElts) {
12838 // This shuffle index refers to the inner shuffle N0. Lookup the inner
12839 // shuffle mask to identify which vector is actually referenced.
12840 Idx = OtherSV->getMaskElt(Idx);
12842 // Propagate Undef.
12843 Mask.push_back(Idx);
12847 CurrentVec = (Idx < (int) NumElts) ? OtherSV->getOperand(0)
12848 : OtherSV->getOperand(1);
12850 // This shuffle index references an element within N1.
12854 // Simple case where 'CurrentVec' is UNDEF.
12855 if (CurrentVec.getOpcode() == ISD::UNDEF) {
12856 Mask.push_back(-1);
12860 // Canonicalize the shuffle index. We don't know yet if CurrentVec
12861 // will be the first or second operand of the combined shuffle.
12862 Idx = Idx % NumElts;
12863 if (!SV0.getNode() || SV0 == CurrentVec) {
12864 // Ok. CurrentVec is the left hand side.
12865 // Update the mask accordingly.
12867 Mask.push_back(Idx);
12871 // Bail out if we cannot convert the shuffle pair into a single shuffle.
12872 if (SV1.getNode() && SV1 != CurrentVec)
12875 // Ok. CurrentVec is the right hand side.
12876 // Update the mask accordingly.
12878 Mask.push_back(Idx + NumElts);
12881 // Check if all indices in Mask are Undef. In case, propagate Undef.
12882 bool isUndefMask = true;
12883 for (unsigned i = 0; i != NumElts && isUndefMask; ++i)
12884 isUndefMask &= Mask[i] < 0;
12887 return DAG.getUNDEF(VT);
12889 if (!SV0.getNode())
12890 SV0 = DAG.getUNDEF(VT);
12891 if (!SV1.getNode())
12892 SV1 = DAG.getUNDEF(VT);
12894 // Avoid introducing shuffles with illegal mask.
12895 if (!TLI.isShuffleMaskLegal(Mask, VT)) {
12896 ShuffleVectorSDNode::commuteMask(Mask);
12898 if (!TLI.isShuffleMaskLegal(Mask, VT))
12901 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, A, M2)
12902 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, A, M2)
12903 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, B, M2)
12904 std::swap(SV0, SV1);
12907 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2)
12908 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2)
12909 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2)
12910 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, &Mask[0]);
12916 SDValue DAGCombiner::visitSCALAR_TO_VECTOR(SDNode *N) {
12917 SDValue InVal = N->getOperand(0);
12918 EVT VT = N->getValueType(0);
12920 // Replace a SCALAR_TO_VECTOR(EXTRACT_VECTOR_ELT(V,C0)) pattern
12921 // with a VECTOR_SHUFFLE.
12922 if (InVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
12923 SDValue InVec = InVal->getOperand(0);
12924 SDValue EltNo = InVal->getOperand(1);
12926 // FIXME: We could support implicit truncation if the shuffle can be
12927 // scaled to a smaller vector scalar type.
12928 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(EltNo);
12929 if (C0 && VT == InVec.getValueType() &&
12930 VT.getScalarType() == InVal.getValueType()) {
12931 SmallVector<int, 8> NewMask(VT.getVectorNumElements(), -1);
12932 int Elt = C0->getZExtValue();
12935 if (TLI.isShuffleMaskLegal(NewMask, VT))
12936 return DAG.getVectorShuffle(VT, SDLoc(N), InVec, DAG.getUNDEF(VT),
12944 SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
12945 SDValue N0 = N->getOperand(0);
12946 SDValue N2 = N->getOperand(2);
12948 // If the input vector is a concatenation, and the insert replaces
12949 // one of the halves, we can optimize into a single concat_vectors.
12950 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
12951 N0->getNumOperands() == 2 && N2.getOpcode() == ISD::Constant) {
12952 APInt InsIdx = cast<ConstantSDNode>(N2)->getAPIntValue();
12953 EVT VT = N->getValueType(0);
12955 // Lower half: fold (insert_subvector (concat_vectors X, Y), Z) ->
12956 // (concat_vectors Z, Y)
12958 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
12959 N->getOperand(1), N0.getOperand(1));
12961 // Upper half: fold (insert_subvector (concat_vectors X, Y), Z) ->
12962 // (concat_vectors X, Z)
12963 if (InsIdx == VT.getVectorNumElements()/2)
12964 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
12965 N0.getOperand(0), N->getOperand(1));
12971 SDValue DAGCombiner::visitFP_TO_FP16(SDNode *N) {
12972 SDValue N0 = N->getOperand(0);
12974 // fold (fp_to_fp16 (fp16_to_fp op)) -> op
12975 if (N0->getOpcode() == ISD::FP16_TO_FP)
12976 return N0->getOperand(0);
12981 /// Returns a vector_shuffle if it able to transform an AND to a vector_shuffle
12982 /// with the destination vector and a zero vector.
12983 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
12984 /// vector_shuffle V, Zero, <0, 4, 2, 4>
12985 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
12986 EVT VT = N->getValueType(0);
12987 SDValue LHS = N->getOperand(0);
12988 SDValue RHS = N->getOperand(1);
12991 // Make sure we're not running after operation legalization where it
12992 // may have custom lowered the vector shuffles.
12993 if (LegalOperations)
12996 if (N->getOpcode() != ISD::AND)
12999 if (RHS.getOpcode() == ISD::BITCAST)
13000 RHS = RHS.getOperand(0);
13002 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
13003 SmallVector<int, 8> Indices;
13004 unsigned NumElts = RHS.getNumOperands();
13006 for (unsigned i = 0; i != NumElts; ++i) {
13007 SDValue Elt = RHS.getOperand(i);
13008 if (isAllOnesConstant(Elt))
13009 Indices.push_back(i);
13010 else if (isNullConstant(Elt))
13011 Indices.push_back(NumElts+i);
13016 // Let's see if the target supports this vector_shuffle.
13017 EVT RVT = RHS.getValueType();
13018 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
13021 // Return the new VECTOR_SHUFFLE node.
13022 EVT EltVT = RVT.getVectorElementType();
13023 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
13024 DAG.getConstant(0, dl, EltVT));
13025 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, dl, RVT, ZeroOps);
13026 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
13027 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
13028 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
13034 /// Visit a binary vector operation, like ADD.
13035 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
13036 assert(N->getValueType(0).isVector() &&
13037 "SimplifyVBinOp only works on vectors!");
13039 SDValue LHS = N->getOperand(0);
13040 SDValue RHS = N->getOperand(1);
13042 if (SDValue Shuffle = XformToShuffleWithZero(N))
13045 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
13047 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
13048 RHS.getOpcode() == ISD::BUILD_VECTOR) {
13049 // Check if both vectors are constants. If not bail out.
13050 if (!(cast<BuildVectorSDNode>(LHS)->isConstant() &&
13051 cast<BuildVectorSDNode>(RHS)->isConstant()))
13054 SmallVector<SDValue, 8> Ops;
13055 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
13056 SDValue LHSOp = LHS.getOperand(i);
13057 SDValue RHSOp = RHS.getOperand(i);
13059 // Can't fold divide by zero.
13060 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
13061 N->getOpcode() == ISD::FDIV) {
13062 if (isNullConstant(RHSOp) || (RHSOp.getOpcode() == ISD::ConstantFP &&
13063 cast<ConstantFPSDNode>(RHSOp.getNode())->isZero()))
13067 EVT VT = LHSOp.getValueType();
13068 EVT RVT = RHSOp.getValueType();
13070 // Integer BUILD_VECTOR operands may have types larger than the element
13071 // size (e.g., when the element type is not legal). Prior to type
13072 // legalization, the types may not match between the two BUILD_VECTORS.
13073 // Truncate one of the operands to make them match.
13074 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
13075 RHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, RHSOp);
13077 LHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), RVT, LHSOp);
13081 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(LHS), VT,
13083 if (FoldOp.getOpcode() != ISD::UNDEF &&
13084 FoldOp.getOpcode() != ISD::Constant &&
13085 FoldOp.getOpcode() != ISD::ConstantFP)
13087 Ops.push_back(FoldOp);
13088 AddToWorklist(FoldOp.getNode());
13091 if (Ops.size() == LHS.getNumOperands())
13092 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), LHS.getValueType(), Ops);
13095 // Type legalization might introduce new shuffles in the DAG.
13096 // Fold (VBinOp (shuffle (A, Undef, Mask)), (shuffle (B, Undef, Mask)))
13097 // -> (shuffle (VBinOp (A, B)), Undef, Mask).
13098 if (LegalTypes && isa<ShuffleVectorSDNode>(LHS) &&
13099 isa<ShuffleVectorSDNode>(RHS) && LHS.hasOneUse() && RHS.hasOneUse() &&
13100 LHS.getOperand(1).getOpcode() == ISD::UNDEF &&
13101 RHS.getOperand(1).getOpcode() == ISD::UNDEF) {
13102 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(LHS);
13103 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(RHS);
13105 if (SVN0->getMask().equals(SVN1->getMask())) {
13106 EVT VT = N->getValueType(0);
13107 SDValue UndefVector = LHS.getOperand(1);
13108 SDValue NewBinOp = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
13109 LHS.getOperand(0), RHS.getOperand(0));
13110 AddUsersToWorklist(N);
13111 return DAG.getVectorShuffle(VT, SDLoc(N), NewBinOp, UndefVector,
13112 &SVN0->getMask()[0]);
13119 SDValue DAGCombiner::SimplifySelect(SDLoc DL, SDValue N0,
13120 SDValue N1, SDValue N2){
13121 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
13123 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
13124 cast<CondCodeSDNode>(N0.getOperand(2))->get());
13126 // If we got a simplified select_cc node back from SimplifySelectCC, then
13127 // break it down into a new SETCC node, and a new SELECT node, and then return
13128 // the SELECT node, since we were called with a SELECT node.
13129 if (SCC.getNode()) {
13130 // Check to see if we got a select_cc back (to turn into setcc/select).
13131 // Otherwise, just return whatever node we got back, like fabs.
13132 if (SCC.getOpcode() == ISD::SELECT_CC) {
13133 SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0),
13135 SCC.getOperand(0), SCC.getOperand(1),
13136 SCC.getOperand(4));
13137 AddToWorklist(SETCC.getNode());
13138 return DAG.getSelect(SDLoc(SCC), SCC.getValueType(), SETCC,
13139 SCC.getOperand(2), SCC.getOperand(3));
13147 /// Given a SELECT or a SELECT_CC node, where LHS and RHS are the two values
13148 /// being selected between, see if we can simplify the select. Callers of this
13149 /// should assume that TheSelect is deleted if this returns true. As such, they
13150 /// should return the appropriate thing (e.g. the node) back to the top-level of
13151 /// the DAG combiner loop to avoid it being looked at.
13152 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
13155 // fold (select (setcc x, -0.0, *lt), NaN, (fsqrt x))
13156 // The select + setcc is redundant, because fsqrt returns NaN for X < -0.
13157 if (const ConstantFPSDNode *NaN = isConstOrConstSplatFP(LHS)) {
13158 if (NaN->isNaN() && RHS.getOpcode() == ISD::FSQRT) {
13159 // We have: (select (setcc ?, ?, ?), NaN, (fsqrt ?))
13160 SDValue Sqrt = RHS;
13163 const ConstantFPSDNode *NegZero = nullptr;
13165 if (TheSelect->getOpcode() == ISD::SELECT_CC) {
13166 CC = dyn_cast<CondCodeSDNode>(TheSelect->getOperand(4))->get();
13167 CmpLHS = TheSelect->getOperand(0);
13168 NegZero = isConstOrConstSplatFP(TheSelect->getOperand(1));
13170 // SELECT or VSELECT
13171 SDValue Cmp = TheSelect->getOperand(0);
13172 if (Cmp.getOpcode() == ISD::SETCC) {
13173 CC = dyn_cast<CondCodeSDNode>(Cmp.getOperand(2))->get();
13174 CmpLHS = Cmp.getOperand(0);
13175 NegZero = isConstOrConstSplatFP(Cmp.getOperand(1));
13178 if (NegZero && NegZero->isNegative() && NegZero->isZero() &&
13179 Sqrt.getOperand(0) == CmpLHS && (CC == ISD::SETOLT ||
13180 CC == ISD::SETULT || CC == ISD::SETLT)) {
13181 // We have: (select (setcc x, -0.0, *lt), NaN, (fsqrt x))
13182 CombineTo(TheSelect, Sqrt);
13187 // Cannot simplify select with vector condition
13188 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
13190 // If this is a select from two identical things, try to pull the operation
13191 // through the select.
13192 if (LHS.getOpcode() != RHS.getOpcode() ||
13193 !LHS.hasOneUse() || !RHS.hasOneUse())
13196 // If this is a load and the token chain is identical, replace the select
13197 // of two loads with a load through a select of the address to load from.
13198 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
13199 // constants have been dropped into the constant pool.
13200 if (LHS.getOpcode() == ISD::LOAD) {
13201 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
13202 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
13204 // Token chains must be identical.
13205 if (LHS.getOperand(0) != RHS.getOperand(0) ||
13206 // Do not let this transformation reduce the number of volatile loads.
13207 LLD->isVolatile() || RLD->isVolatile() ||
13208 // FIXME: If either is a pre/post inc/dec load,
13209 // we'd need to split out the address adjustment.
13210 LLD->isIndexed() || RLD->isIndexed() ||
13211 // If this is an EXTLOAD, the VT's must match.
13212 LLD->getMemoryVT() != RLD->getMemoryVT() ||
13213 // If this is an EXTLOAD, the kind of extension must match.
13214 (LLD->getExtensionType() != RLD->getExtensionType() &&
13215 // The only exception is if one of the extensions is anyext.
13216 LLD->getExtensionType() != ISD::EXTLOAD &&
13217 RLD->getExtensionType() != ISD::EXTLOAD) ||
13218 // FIXME: this discards src value information. This is
13219 // over-conservative. It would be beneficial to be able to remember
13220 // both potential memory locations. Since we are discarding
13221 // src value info, don't do the transformation if the memory
13222 // locations are not in the default address space.
13223 LLD->getPointerInfo().getAddrSpace() != 0 ||
13224 RLD->getPointerInfo().getAddrSpace() != 0 ||
13225 !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(),
13226 LLD->getBasePtr().getValueType()))
13229 // Check that the select condition doesn't reach either load. If so,
13230 // folding this will induce a cycle into the DAG. If not, this is safe to
13231 // xform, so create a select of the addresses.
13233 if (TheSelect->getOpcode() == ISD::SELECT) {
13234 SDNode *CondNode = TheSelect->getOperand(0).getNode();
13235 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
13236 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
13238 // The loads must not depend on one another.
13239 if (LLD->isPredecessorOf(RLD) ||
13240 RLD->isPredecessorOf(LLD))
13242 Addr = DAG.getSelect(SDLoc(TheSelect),
13243 LLD->getBasePtr().getValueType(),
13244 TheSelect->getOperand(0), LLD->getBasePtr(),
13245 RLD->getBasePtr());
13246 } else { // Otherwise SELECT_CC
13247 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
13248 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
13250 if ((LLD->hasAnyUseOfValue(1) &&
13251 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
13252 (RLD->hasAnyUseOfValue(1) &&
13253 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
13256 Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect),
13257 LLD->getBasePtr().getValueType(),
13258 TheSelect->getOperand(0),
13259 TheSelect->getOperand(1),
13260 LLD->getBasePtr(), RLD->getBasePtr(),
13261 TheSelect->getOperand(4));
13265 // It is safe to replace the two loads if they have different alignments,
13266 // but the new load must be the minimum (most restrictive) alignment of the
13268 bool isInvariant = LLD->isInvariant() & RLD->isInvariant();
13269 unsigned Alignment = std::min(LLD->getAlignment(), RLD->getAlignment());
13270 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
13271 Load = DAG.getLoad(TheSelect->getValueType(0),
13273 // FIXME: Discards pointer and AA info.
13274 LLD->getChain(), Addr, MachinePointerInfo(),
13275 LLD->isVolatile(), LLD->isNonTemporal(),
13276 isInvariant, Alignment);
13278 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
13279 RLD->getExtensionType() : LLD->getExtensionType(),
13281 TheSelect->getValueType(0),
13282 // FIXME: Discards pointer and AA info.
13283 LLD->getChain(), Addr, MachinePointerInfo(),
13284 LLD->getMemoryVT(), LLD->isVolatile(),
13285 LLD->isNonTemporal(), isInvariant, Alignment);
13288 // Users of the select now use the result of the load.
13289 CombineTo(TheSelect, Load);
13291 // Users of the old loads now use the new load's chain. We know the
13292 // old-load value is dead now.
13293 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
13294 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
13301 /// Simplify an expression of the form (N0 cond N1) ? N2 : N3
13302 /// where 'cond' is the comparison specified by CC.
13303 SDValue DAGCombiner::SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1,
13304 SDValue N2, SDValue N3,
13305 ISD::CondCode CC, bool NotExtCompare) {
13306 // (x ? y : y) -> y.
13307 if (N2 == N3) return N2;
13309 EVT VT = N2.getValueType();
13310 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
13311 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
13313 // Determine if the condition we're dealing with is constant
13314 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
13315 N0, N1, CC, DL, false);
13316 if (SCC.getNode()) AddToWorklist(SCC.getNode());
13318 if (ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode())) {
13319 // fold select_cc true, x, y -> x
13320 // fold select_cc false, x, y -> y
13321 return !SCCC->isNullValue() ? N2 : N3;
13324 // Check to see if we can simplify the select into an fabs node
13325 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
13326 // Allow either -0.0 or 0.0
13327 if (CFP->isZero()) {
13328 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
13329 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
13330 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
13331 N2 == N3.getOperand(0))
13332 return DAG.getNode(ISD::FABS, DL, VT, N0);
13334 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
13335 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
13336 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
13337 N2.getOperand(0) == N3)
13338 return DAG.getNode(ISD::FABS, DL, VT, N3);
13342 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
13343 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
13344 // in it. This is a win when the constant is not otherwise available because
13345 // it replaces two constant pool loads with one. We only do this if the FP
13346 // type is known to be legal, because if it isn't, then we are before legalize
13347 // types an we want the other legalization to happen first (e.g. to avoid
13348 // messing with soft float) and if the ConstantFP is not legal, because if
13349 // it is legal, we may not need to store the FP constant in a constant pool.
13350 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
13351 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
13352 if (TLI.isTypeLegal(N2.getValueType()) &&
13353 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
13354 TargetLowering::Legal &&
13355 !TLI.isFPImmLegal(TV->getValueAPF(), TV->getValueType(0)) &&
13356 !TLI.isFPImmLegal(FV->getValueAPF(), FV->getValueType(0))) &&
13357 // If both constants have multiple uses, then we won't need to do an
13358 // extra load, they are likely around in registers for other users.
13359 (TV->hasOneUse() || FV->hasOneUse())) {
13360 Constant *Elts[] = {
13361 const_cast<ConstantFP*>(FV->getConstantFPValue()),
13362 const_cast<ConstantFP*>(TV->getConstantFPValue())
13364 Type *FPTy = Elts[0]->getType();
13365 const DataLayout &TD = DAG.getDataLayout();
13367 // Create a ConstantArray of the two constants.
13368 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
13370 DAG.getConstantPool(CA, TLI.getPointerTy(DAG.getDataLayout()),
13371 TD.getPrefTypeAlignment(FPTy));
13372 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
13374 // Get the offsets to the 0 and 1 element of the array so that we can
13375 // select between them.
13376 SDValue Zero = DAG.getIntPtrConstant(0, DL);
13377 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
13378 SDValue One = DAG.getIntPtrConstant(EltSize, SDLoc(FV));
13380 SDValue Cond = DAG.getSetCC(DL,
13381 getSetCCResultType(N0.getValueType()),
13383 AddToWorklist(Cond.getNode());
13384 SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(),
13386 AddToWorklist(CstOffset.getNode());
13387 CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx,
13389 AddToWorklist(CPIdx.getNode());
13390 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
13391 MachinePointerInfo::getConstantPool(), false,
13392 false, false, Alignment);
13396 // Check to see if we can perform the "gzip trick", transforming
13397 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
13398 if (isNullConstant(N3) && CC == ISD::SETLT &&
13399 (isNullConstant(N1) || // (a < 0) ? b : 0
13400 (isOneConstant(N1) && N0 == N2))) { // (a < 1) ? a : 0
13401 EVT XType = N0.getValueType();
13402 EVT AType = N2.getValueType();
13403 if (XType.bitsGE(AType)) {
13404 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
13405 // single-bit constant.
13406 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue() - 1)) == 0)) {
13407 unsigned ShCtV = N2C->getAPIntValue().logBase2();
13408 ShCtV = XType.getSizeInBits() - ShCtV - 1;
13409 SDValue ShCt = DAG.getConstant(ShCtV, SDLoc(N0),
13410 getShiftAmountTy(N0.getValueType()));
13411 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0),
13413 AddToWorklist(Shift.getNode());
13415 if (XType.bitsGT(AType)) {
13416 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
13417 AddToWorklist(Shift.getNode());
13420 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
13423 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0),
13425 DAG.getConstant(XType.getSizeInBits() - 1,
13427 getShiftAmountTy(N0.getValueType())));
13428 AddToWorklist(Shift.getNode());
13430 if (XType.bitsGT(AType)) {
13431 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
13432 AddToWorklist(Shift.getNode());
13435 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
13439 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
13440 // where y is has a single bit set.
13441 // A plaintext description would be, we can turn the SELECT_CC into an AND
13442 // when the condition can be materialized as an all-ones register. Any
13443 // single bit-test can be materialized as an all-ones register with
13444 // shift-left and shift-right-arith.
13445 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
13446 N0->getValueType(0) == VT && isNullConstant(N1) && isNullConstant(N2)) {
13447 SDValue AndLHS = N0->getOperand(0);
13448 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
13449 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
13450 // Shift the tested bit over the sign bit.
13451 APInt AndMask = ConstAndRHS->getAPIntValue();
13453 DAG.getConstant(AndMask.countLeadingZeros(), SDLoc(AndLHS),
13454 getShiftAmountTy(AndLHS.getValueType()));
13455 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt);
13457 // Now arithmetic right shift it all the way over, so the result is either
13458 // all-ones, or zero.
13460 DAG.getConstant(AndMask.getBitWidth() - 1, SDLoc(Shl),
13461 getShiftAmountTy(Shl.getValueType()));
13462 SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt);
13464 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
13468 // fold select C, 16, 0 -> shl C, 4
13469 if (N2C && isNullConstant(N3) && N2C->getAPIntValue().isPowerOf2() &&
13470 TLI.getBooleanContents(N0.getValueType()) ==
13471 TargetLowering::ZeroOrOneBooleanContent) {
13473 // If the caller doesn't want us to simplify this into a zext of a compare,
13475 if (NotExtCompare && N2C->isOne())
13478 // Get a SetCC of the condition
13479 // NOTE: Don't create a SETCC if it's not legal on this target.
13480 if (!LegalOperations ||
13481 TLI.isOperationLegal(ISD::SETCC,
13482 LegalTypes ? getSetCCResultType(N0.getValueType()) : MVT::i1)) {
13484 // cast from setcc result type to select result type
13486 SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()),
13488 if (N2.getValueType().bitsLT(SCC.getValueType()))
13489 Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2),
13490 N2.getValueType());
13492 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
13493 N2.getValueType(), SCC);
13495 SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC);
13496 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
13497 N2.getValueType(), SCC);
13500 AddToWorklist(SCC.getNode());
13501 AddToWorklist(Temp.getNode());
13506 // shl setcc result by log2 n2c
13507 return DAG.getNode(
13508 ISD::SHL, DL, N2.getValueType(), Temp,
13509 DAG.getConstant(N2C->getAPIntValue().logBase2(), SDLoc(Temp),
13510 getShiftAmountTy(Temp.getValueType())));
13514 // Check to see if this is the equivalent of setcc
13515 // FIXME: Turn all of these into setcc if setcc if setcc is legal
13516 // otherwise, go ahead with the folds.
13517 if (0 && isNullConstant(N3) && isOneConstant(N2)) {
13518 EVT XType = N0.getValueType();
13519 if (!LegalOperations ||
13520 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(XType))) {
13521 SDValue Res = DAG.getSetCC(DL, getSetCCResultType(XType), N0, N1, CC);
13522 if (Res.getValueType() != VT)
13523 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
13527 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
13528 if (isNullConstant(N1) && CC == ISD::SETEQ &&
13529 (!LegalOperations ||
13530 TLI.isOperationLegal(ISD::CTLZ, XType))) {
13531 SDValue Ctlz = DAG.getNode(ISD::CTLZ, SDLoc(N0), XType, N0);
13532 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
13533 DAG.getConstant(Log2_32(XType.getSizeInBits()),
13535 getShiftAmountTy(Ctlz.getValueType())));
13537 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
13538 if (isNullConstant(N1) && CC == ISD::SETGT) {
13540 SDValue NegN0 = DAG.getNode(ISD::SUB, DL,
13541 XType, DAG.getConstant(0, DL, XType), N0);
13542 SDValue NotN0 = DAG.getNOT(DL, N0, XType);
13543 return DAG.getNode(ISD::SRL, DL, XType,
13544 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
13545 DAG.getConstant(XType.getSizeInBits() - 1, DL,
13546 getShiftAmountTy(XType)));
13548 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
13549 if (isAllOnesConstant(N1) && CC == ISD::SETGT) {
13551 SDValue Sign = DAG.getNode(ISD::SRL, DL, XType, N0,
13552 DAG.getConstant(XType.getSizeInBits() - 1, DL,
13553 getShiftAmountTy(N0.getValueType())));
13554 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, DL,
13559 // Check to see if this is an integer abs.
13560 // select_cc setg[te] X, 0, X, -X ->
13561 // select_cc setgt X, -1, X, -X ->
13562 // select_cc setl[te] X, 0, -X, X ->
13563 // select_cc setlt X, 1, -X, X ->
13564 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
13566 ConstantSDNode *SubC = nullptr;
13567 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
13568 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
13569 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
13570 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
13571 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
13572 (N1C->isOne() && CC == ISD::SETLT)) &&
13573 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
13574 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
13576 EVT XType = N0.getValueType();
13577 if (SubC && SubC->isNullValue() && XType.isInteger()) {
13579 SDValue Shift = DAG.getNode(ISD::SRA, DL, XType,
13581 DAG.getConstant(XType.getSizeInBits() - 1, DL,
13582 getShiftAmountTy(N0.getValueType())));
13583 SDValue Add = DAG.getNode(ISD::ADD, DL,
13585 AddToWorklist(Shift.getNode());
13586 AddToWorklist(Add.getNode());
13587 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
13594 /// This is a stub for TargetLowering::SimplifySetCC.
13595 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
13596 SDValue N1, ISD::CondCode Cond,
13597 SDLoc DL, bool foldBooleans) {
13598 TargetLowering::DAGCombinerInfo
13599 DagCombineInfo(DAG, Level, false, this);
13600 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
13603 /// Given an ISD::SDIV node expressing a divide by constant, return
13604 /// a DAG expression to select that will generate the same value by multiplying
13605 /// by a magic number.
13606 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
13607 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
13608 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
13612 // Avoid division by zero.
13613 if (C->isNullValue())
13616 std::vector<SDNode*> Built;
13618 TLI.BuildSDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
13620 for (SDNode *N : Built)
13625 /// Given an ISD::SDIV node expressing a divide by constant power of 2, return a
13626 /// DAG expression that will generate the same value by right shifting.
13627 SDValue DAGCombiner::BuildSDIVPow2(SDNode *N) {
13628 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
13632 // Avoid division by zero.
13633 if (C->isNullValue())
13636 std::vector<SDNode *> Built;
13637 SDValue S = TLI.BuildSDIVPow2(N, C->getAPIntValue(), DAG, &Built);
13639 for (SDNode *N : Built)
13644 /// Given an ISD::UDIV node expressing a divide by constant, return a DAG
13645 /// expression that will generate the same value by multiplying by a magic
13647 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
13648 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
13649 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
13653 // Avoid division by zero.
13654 if (C->isNullValue())
13657 std::vector<SDNode*> Built;
13659 TLI.BuildUDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
13661 for (SDNode *N : Built)
13666 SDValue DAGCombiner::BuildReciprocalEstimate(SDValue Op) {
13667 if (Level >= AfterLegalizeDAG)
13670 // Expose the DAG combiner to the target combiner implementations.
13671 TargetLowering::DAGCombinerInfo DCI(DAG, Level, false, this);
13673 unsigned Iterations = 0;
13674 if (SDValue Est = TLI.getRecipEstimate(Op, DCI, Iterations)) {
13676 // Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
13677 // For the reciprocal, we need to find the zero of the function:
13678 // F(X) = A X - 1 [which has a zero at X = 1/A]
13680 // X_{i+1} = X_i (2 - A X_i) = X_i + X_i (1 - A X_i) [this second form
13681 // does not require additional intermediate precision]
13682 EVT VT = Op.getValueType();
13684 SDValue FPOne = DAG.getConstantFP(1.0, DL, VT);
13686 AddToWorklist(Est.getNode());
13688 // Newton iterations: Est = Est + Est (1 - Arg * Est)
13689 for (unsigned i = 0; i < Iterations; ++i) {
13690 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Op, Est);
13691 AddToWorklist(NewEst.getNode());
13693 NewEst = DAG.getNode(ISD::FSUB, DL, VT, FPOne, NewEst);
13694 AddToWorklist(NewEst.getNode());
13696 NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst);
13697 AddToWorklist(NewEst.getNode());
13699 Est = DAG.getNode(ISD::FADD, DL, VT, Est, NewEst);
13700 AddToWorklist(Est.getNode());
13709 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
13710 /// For the reciprocal sqrt, we need to find the zero of the function:
13711 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
13713 /// X_{i+1} = X_i (1.5 - A X_i^2 / 2)
13714 /// As a result, we precompute A/2 prior to the iteration loop.
13715 SDValue DAGCombiner::BuildRsqrtNROneConst(SDValue Arg, SDValue Est,
13716 unsigned Iterations) {
13717 EVT VT = Arg.getValueType();
13719 SDValue ThreeHalves = DAG.getConstantFP(1.5, DL, VT);
13721 // We now need 0.5 * Arg which we can write as (1.5 * Arg - Arg) so that
13722 // this entire sequence requires only one FP constant.
13723 SDValue HalfArg = DAG.getNode(ISD::FMUL, DL, VT, ThreeHalves, Arg);
13724 AddToWorklist(HalfArg.getNode());
13726 HalfArg = DAG.getNode(ISD::FSUB, DL, VT, HalfArg, Arg);
13727 AddToWorklist(HalfArg.getNode());
13729 // Newton iterations: Est = Est * (1.5 - HalfArg * Est * Est)
13730 for (unsigned i = 0; i < Iterations; ++i) {
13731 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, Est);
13732 AddToWorklist(NewEst.getNode());
13734 NewEst = DAG.getNode(ISD::FMUL, DL, VT, HalfArg, NewEst);
13735 AddToWorklist(NewEst.getNode());
13737 NewEst = DAG.getNode(ISD::FSUB, DL, VT, ThreeHalves, NewEst);
13738 AddToWorklist(NewEst.getNode());
13740 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst);
13741 AddToWorklist(Est.getNode());
13746 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
13747 /// For the reciprocal sqrt, we need to find the zero of the function:
13748 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
13750 /// X_{i+1} = (-0.5 * X_i) * (A * X_i * X_i + (-3.0))
13751 SDValue DAGCombiner::BuildRsqrtNRTwoConst(SDValue Arg, SDValue Est,
13752 unsigned Iterations) {
13753 EVT VT = Arg.getValueType();
13755 SDValue MinusThree = DAG.getConstantFP(-3.0, DL, VT);
13756 SDValue MinusHalf = DAG.getConstantFP(-0.5, DL, VT);
13758 // Newton iterations: Est = -0.5 * Est * (-3.0 + Arg * Est * Est)
13759 for (unsigned i = 0; i < Iterations; ++i) {
13760 SDValue HalfEst = DAG.getNode(ISD::FMUL, DL, VT, Est, MinusHalf);
13761 AddToWorklist(HalfEst.getNode());
13763 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Est);
13764 AddToWorklist(Est.getNode());
13766 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Arg);
13767 AddToWorklist(Est.getNode());
13769 Est = DAG.getNode(ISD::FADD, DL, VT, Est, MinusThree);
13770 AddToWorklist(Est.getNode());
13772 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, HalfEst);
13773 AddToWorklist(Est.getNode());
13778 SDValue DAGCombiner::BuildRsqrtEstimate(SDValue Op) {
13779 if (Level >= AfterLegalizeDAG)
13782 // Expose the DAG combiner to the target combiner implementations.
13783 TargetLowering::DAGCombinerInfo DCI(DAG, Level, false, this);
13784 unsigned Iterations = 0;
13785 bool UseOneConstNR = false;
13786 if (SDValue Est = TLI.getRsqrtEstimate(Op, DCI, Iterations, UseOneConstNR)) {
13787 AddToWorklist(Est.getNode());
13789 Est = UseOneConstNR ?
13790 BuildRsqrtNROneConst(Op, Est, Iterations) :
13791 BuildRsqrtNRTwoConst(Op, Est, Iterations);
13799 /// Return true if base is a frame index, which is known not to alias with
13800 /// anything but itself. Provides base object and offset as results.
13801 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
13802 const GlobalValue *&GV, const void *&CV) {
13803 // Assume it is a primitive operation.
13804 Base = Ptr; Offset = 0; GV = nullptr; CV = nullptr;
13806 // If it's an adding a simple constant then integrate the offset.
13807 if (Base.getOpcode() == ISD::ADD) {
13808 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
13809 Base = Base.getOperand(0);
13810 Offset += C->getZExtValue();
13814 // Return the underlying GlobalValue, and update the Offset. Return false
13815 // for GlobalAddressSDNode since the same GlobalAddress may be represented
13816 // by multiple nodes with different offsets.
13817 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
13818 GV = G->getGlobal();
13819 Offset += G->getOffset();
13823 // Return the underlying Constant value, and update the Offset. Return false
13824 // for ConstantSDNodes since the same constant pool entry may be represented
13825 // by multiple nodes with different offsets.
13826 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
13827 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
13828 : (const void *)C->getConstVal();
13829 Offset += C->getOffset();
13832 // If it's any of the following then it can't alias with anything but itself.
13833 return isa<FrameIndexSDNode>(Base);
13836 /// Return true if there is any possibility that the two addresses overlap.
13837 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const {
13838 // If they are the same then they must be aliases.
13839 if (Op0->getBasePtr() == Op1->getBasePtr()) return true;
13841 // If they are both volatile then they cannot be reordered.
13842 if (Op0->isVolatile() && Op1->isVolatile()) return true;
13844 // If one operation reads from invariant memory, and the other may store, they
13845 // cannot alias. These should really be checking the equivalent of mayWrite,
13846 // but it only matters for memory nodes other than load /store.
13847 if (Op0->isInvariant() && Op1->writeMem())
13850 if (Op1->isInvariant() && Op0->writeMem())
13853 // Gather base node and offset information.
13854 SDValue Base1, Base2;
13855 int64_t Offset1, Offset2;
13856 const GlobalValue *GV1, *GV2;
13857 const void *CV1, *CV2;
13858 bool isFrameIndex1 = FindBaseOffset(Op0->getBasePtr(),
13859 Base1, Offset1, GV1, CV1);
13860 bool isFrameIndex2 = FindBaseOffset(Op1->getBasePtr(),
13861 Base2, Offset2, GV2, CV2);
13863 // If they have a same base address then check to see if they overlap.
13864 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
13865 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
13866 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
13868 // It is possible for different frame indices to alias each other, mostly
13869 // when tail call optimization reuses return address slots for arguments.
13870 // To catch this case, look up the actual index of frame indices to compute
13871 // the real alias relationship.
13872 if (isFrameIndex1 && isFrameIndex2) {
13873 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
13874 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
13875 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
13876 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
13877 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
13880 // Otherwise, if we know what the bases are, and they aren't identical, then
13881 // we know they cannot alias.
13882 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
13885 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
13886 // compared to the size and offset of the access, we may be able to prove they
13887 // do not alias. This check is conservative for now to catch cases created by
13888 // splitting vector types.
13889 if ((Op0->getOriginalAlignment() == Op1->getOriginalAlignment()) &&
13890 (Op0->getSrcValueOffset() != Op1->getSrcValueOffset()) &&
13891 (Op0->getMemoryVT().getSizeInBits() >> 3 ==
13892 Op1->getMemoryVT().getSizeInBits() >> 3) &&
13893 (Op0->getOriginalAlignment() > Op0->getMemoryVT().getSizeInBits()) >> 3) {
13894 int64_t OffAlign1 = Op0->getSrcValueOffset() % Op0->getOriginalAlignment();
13895 int64_t OffAlign2 = Op1->getSrcValueOffset() % Op1->getOriginalAlignment();
13897 // There is no overlap between these relatively aligned accesses of similar
13898 // size, return no alias.
13899 if ((OffAlign1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign2 ||
13900 (OffAlign2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign1)
13904 bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0
13906 : DAG.getSubtarget().useAA();
13908 if (CombinerAAOnlyFunc.getNumOccurrences() &&
13909 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
13913 Op0->getMemOperand()->getValue() && Op1->getMemOperand()->getValue()) {
13914 // Use alias analysis information.
13915 int64_t MinOffset = std::min(Op0->getSrcValueOffset(),
13916 Op1->getSrcValueOffset());
13917 int64_t Overlap1 = (Op0->getMemoryVT().getSizeInBits() >> 3) +
13918 Op0->getSrcValueOffset() - MinOffset;
13919 int64_t Overlap2 = (Op1->getMemoryVT().getSizeInBits() >> 3) +
13920 Op1->getSrcValueOffset() - MinOffset;
13921 AliasResult AAResult =
13922 AA.alias(MemoryLocation(Op0->getMemOperand()->getValue(), Overlap1,
13923 UseTBAA ? Op0->getAAInfo() : AAMDNodes()),
13924 MemoryLocation(Op1->getMemOperand()->getValue(), Overlap2,
13925 UseTBAA ? Op1->getAAInfo() : AAMDNodes()));
13926 if (AAResult == NoAlias)
13930 // Otherwise we have to assume they alias.
13934 /// Walk up chain skipping non-aliasing memory nodes,
13935 /// looking for aliasing nodes and adding them to the Aliases vector.
13936 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
13937 SmallVectorImpl<SDValue> &Aliases) {
13938 SmallVector<SDValue, 8> Chains; // List of chains to visit.
13939 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
13941 // Get alias information for node.
13942 bool IsLoad = isa<LoadSDNode>(N) && !cast<LSBaseSDNode>(N)->isVolatile();
13945 Chains.push_back(OriginalChain);
13946 unsigned Depth = 0;
13948 // Look at each chain and determine if it is an alias. If so, add it to the
13949 // aliases list. If not, then continue up the chain looking for the next
13951 while (!Chains.empty()) {
13952 SDValue Chain = Chains.pop_back_val();
13954 // For TokenFactor nodes, look at each operand and only continue up the
13955 // chain until we find two aliases. If we've seen two aliases, assume we'll
13956 // find more and revert to original chain since the xform is unlikely to be
13959 // FIXME: The depth check could be made to return the last non-aliasing
13960 // chain we found before we hit a tokenfactor rather than the original
13962 if (Depth > 6 || Aliases.size() == 2) {
13964 Aliases.push_back(OriginalChain);
13968 // Don't bother if we've been before.
13969 if (!Visited.insert(Chain.getNode()).second)
13972 switch (Chain.getOpcode()) {
13973 case ISD::EntryToken:
13974 // Entry token is ideal chain operand, but handled in FindBetterChain.
13979 // Get alias information for Chain.
13980 bool IsOpLoad = isa<LoadSDNode>(Chain.getNode()) &&
13981 !cast<LSBaseSDNode>(Chain.getNode())->isVolatile();
13983 // If chain is alias then stop here.
13984 if (!(IsLoad && IsOpLoad) &&
13985 isAlias(cast<LSBaseSDNode>(N), cast<LSBaseSDNode>(Chain.getNode()))) {
13986 Aliases.push_back(Chain);
13988 // Look further up the chain.
13989 Chains.push_back(Chain.getOperand(0));
13995 case ISD::TokenFactor:
13996 // We have to check each of the operands of the token factor for "small"
13997 // token factors, so we queue them up. Adding the operands to the queue
13998 // (stack) in reverse order maintains the original order and increases the
13999 // likelihood that getNode will find a matching token factor (CSE.)
14000 if (Chain.getNumOperands() > 16) {
14001 Aliases.push_back(Chain);
14004 for (unsigned n = Chain.getNumOperands(); n;)
14005 Chains.push_back(Chain.getOperand(--n));
14010 // For all other instructions we will just have to take what we can get.
14011 Aliases.push_back(Chain);
14016 // We need to be careful here to also search for aliases through the
14017 // value operand of a store, etc. Consider the following situation:
14019 // L1 = load Token1, %52
14020 // S1 = store Token1, L1, %51
14021 // L2 = load Token1, %52+8
14022 // S2 = store Token1, L2, %51+8
14023 // Token2 = Token(S1, S2)
14024 // L3 = load Token2, %53
14025 // S3 = store Token2, L3, %52
14026 // L4 = load Token2, %53+8
14027 // S4 = store Token2, L4, %52+8
14028 // If we search for aliases of S3 (which loads address %52), and we look
14029 // only through the chain, then we'll miss the trivial dependence on L1
14030 // (which also loads from %52). We then might change all loads and
14031 // stores to use Token1 as their chain operand, which could result in
14032 // copying %53 into %52 before copying %52 into %51 (which should
14035 // The problem is, however, that searching for such data dependencies
14036 // can become expensive, and the cost is not directly related to the
14037 // chain depth. Instead, we'll rule out such configurations here by
14038 // insisting that we've visited all chain users (except for users
14039 // of the original chain, which is not necessary). When doing this,
14040 // we need to look through nodes we don't care about (otherwise, things
14041 // like register copies will interfere with trivial cases).
14043 SmallVector<const SDNode *, 16> Worklist;
14044 for (const SDNode *N : Visited)
14045 if (N != OriginalChain.getNode())
14046 Worklist.push_back(N);
14048 while (!Worklist.empty()) {
14049 const SDNode *M = Worklist.pop_back_val();
14051 // We have already visited M, and want to make sure we've visited any uses
14052 // of M that we care about. For uses that we've not visisted, and don't
14053 // care about, queue them to the worklist.
14055 for (SDNode::use_iterator UI = M->use_begin(),
14056 UIE = M->use_end(); UI != UIE; ++UI)
14057 if (UI.getUse().getValueType() == MVT::Other &&
14058 Visited.insert(*UI).second) {
14059 if (isa<MemSDNode>(*UI)) {
14060 // We've not visited this use, and we care about it (it could have an
14061 // ordering dependency with the original node).
14063 Aliases.push_back(OriginalChain);
14067 // We've not visited this use, but we don't care about it. Mark it as
14068 // visited and enqueue it to the worklist.
14069 Worklist.push_back(*UI);
14074 /// Walk up chain skipping non-aliasing memory nodes, looking for a better chain
14075 /// (aliasing node.)
14076 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
14077 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
14079 // Accumulate all the aliases to this node.
14080 GatherAllAliases(N, OldChain, Aliases);
14082 // If no operands then chain to entry token.
14083 if (Aliases.size() == 0)
14084 return DAG.getEntryNode();
14086 // If a single operand then chain to it. We don't need to revisit it.
14087 if (Aliases.size() == 1)
14090 // Construct a custom tailored token factor.
14091 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Aliases);
14094 /// This is the entry point for the file.
14095 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
14096 CodeGenOpt::Level OptLevel) {
14097 /// This is the main entry point to this class.
14098 DAGCombiner(*this, AA, OptLevel).Run(Level);