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 ForCodeSize = DAG.getMachineFunction().getFunction()->optForSize();
434 /// Runs the dag combiner on all nodes in the work list
435 void Run(CombineLevel AtLevel);
437 SelectionDAG &getDAG() const { return DAG; }
439 /// Returns a type large enough to hold any valid shift amount - before type
440 /// legalization these can be huge.
441 EVT getShiftAmountTy(EVT LHSTy) {
442 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
443 if (LHSTy.isVector())
445 auto &DL = DAG.getDataLayout();
446 return LegalTypes ? TLI.getScalarShiftAmountTy(DL, LHSTy)
447 : TLI.getPointerTy(DL);
450 /// This method returns true if we are running before type legalization or
451 /// if the specified VT is legal.
452 bool isTypeLegal(const EVT &VT) {
453 if (!LegalTypes) return true;
454 return TLI.isTypeLegal(VT);
457 /// Convenience wrapper around TargetLowering::getSetCCResultType
458 EVT getSetCCResultType(EVT VT) const {
459 return TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
466 /// This class is a DAGUpdateListener that removes any deleted
467 /// nodes from the worklist.
468 class WorklistRemover : public SelectionDAG::DAGUpdateListener {
471 explicit WorklistRemover(DAGCombiner &dc)
472 : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {}
474 void NodeDeleted(SDNode *N, SDNode *E) override {
475 DC.removeFromWorklist(N);
480 //===----------------------------------------------------------------------===//
481 // TargetLowering::DAGCombinerInfo implementation
482 //===----------------------------------------------------------------------===//
484 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
485 ((DAGCombiner*)DC)->AddToWorklist(N);
488 void TargetLowering::DAGCombinerInfo::RemoveFromWorklist(SDNode *N) {
489 ((DAGCombiner*)DC)->removeFromWorklist(N);
492 SDValue TargetLowering::DAGCombinerInfo::
493 CombineTo(SDNode *N, ArrayRef<SDValue> To, bool AddTo) {
494 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
497 SDValue TargetLowering::DAGCombinerInfo::
498 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
499 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
503 SDValue TargetLowering::DAGCombinerInfo::
504 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
505 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
508 void TargetLowering::DAGCombinerInfo::
509 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
510 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
513 //===----------------------------------------------------------------------===//
515 //===----------------------------------------------------------------------===//
517 void DAGCombiner::deleteAndRecombine(SDNode *N) {
518 removeFromWorklist(N);
520 // If the operands of this node are only used by the node, they will now be
521 // dead. Make sure to re-visit them and recursively delete dead nodes.
522 for (const SDValue &Op : N->ops())
523 // For an operand generating multiple values, one of the values may
524 // become dead allowing further simplification (e.g. split index
525 // arithmetic from an indexed load).
526 if (Op->hasOneUse() || Op->getNumValues() > 1)
527 AddToWorklist(Op.getNode());
532 /// Return 1 if we can compute the negated form of the specified expression for
533 /// the same cost as the expression itself, or 2 if we can compute the negated
534 /// form more cheaply than the expression itself.
535 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
536 const TargetLowering &TLI,
537 const TargetOptions *Options,
538 unsigned Depth = 0) {
539 // fneg is removable even if it has multiple uses.
540 if (Op.getOpcode() == ISD::FNEG) return 2;
542 // Don't allow anything with multiple uses.
543 if (!Op.hasOneUse()) return 0;
545 // Don't recurse exponentially.
546 if (Depth > 6) return 0;
548 switch (Op.getOpcode()) {
549 default: return false;
550 case ISD::ConstantFP:
551 // Don't invert constant FP values after legalize. The negated constant
552 // isn't necessarily legal.
553 return LegalOperations ? 0 : 1;
555 // FIXME: determine better conditions for this xform.
556 if (!Options->UnsafeFPMath) return 0;
558 // After operation legalization, it might not be legal to create new FSUBs.
559 if (LegalOperations &&
560 !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType()))
563 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
564 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
567 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
568 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
571 // We can't turn -(A-B) into B-A when we honor signed zeros.
572 if (!Options->UnsafeFPMath) return 0;
574 // fold (fneg (fsub A, B)) -> (fsub B, A)
579 if (Options->HonorSignDependentRoundingFPMath()) return 0;
581 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
582 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
586 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
592 return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options,
597 /// If isNegatibleForFree returns true, return the newly negated expression.
598 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
599 bool LegalOperations, unsigned Depth = 0) {
600 const TargetOptions &Options = DAG.getTarget().Options;
601 // fneg is removable even if it has multiple uses.
602 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
604 // Don't allow anything with multiple uses.
605 assert(Op.hasOneUse() && "Unknown reuse!");
607 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
608 switch (Op.getOpcode()) {
609 default: llvm_unreachable("Unknown code");
610 case ISD::ConstantFP: {
611 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
613 return DAG.getConstantFP(V, SDLoc(Op), Op.getValueType());
616 // FIXME: determine better conditions for this xform.
617 assert(Options.UnsafeFPMath);
619 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
620 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
621 DAG.getTargetLoweringInfo(), &Options, Depth+1))
622 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
623 GetNegatedExpression(Op.getOperand(0), DAG,
624 LegalOperations, Depth+1),
626 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
627 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
628 GetNegatedExpression(Op.getOperand(1), DAG,
629 LegalOperations, Depth+1),
632 // We can't turn -(A-B) into B-A when we honor signed zeros.
633 assert(Options.UnsafeFPMath);
635 // fold (fneg (fsub 0, B)) -> B
636 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
638 return Op.getOperand(1);
640 // fold (fneg (fsub A, B)) -> (fsub B, A)
641 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
642 Op.getOperand(1), Op.getOperand(0));
646 assert(!Options.HonorSignDependentRoundingFPMath());
648 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
649 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
650 DAG.getTargetLoweringInfo(), &Options, Depth+1))
651 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
652 GetNegatedExpression(Op.getOperand(0), DAG,
653 LegalOperations, Depth+1),
656 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
657 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
659 GetNegatedExpression(Op.getOperand(1), DAG,
660 LegalOperations, Depth+1));
664 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
665 GetNegatedExpression(Op.getOperand(0), DAG,
666 LegalOperations, Depth+1));
668 return DAG.getNode(ISD::FP_ROUND, SDLoc(Op), Op.getValueType(),
669 GetNegatedExpression(Op.getOperand(0), DAG,
670 LegalOperations, Depth+1),
675 // Return true if this node is a setcc, or is a select_cc
676 // that selects between the target values used for true and false, making it
677 // equivalent to a setcc. Also, set the incoming LHS, RHS, and CC references to
678 // the appropriate nodes based on the type of node we are checking. This
679 // simplifies life a bit for the callers.
680 bool DAGCombiner::isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
682 if (N.getOpcode() == ISD::SETCC) {
683 LHS = N.getOperand(0);
684 RHS = N.getOperand(1);
685 CC = N.getOperand(2);
689 if (N.getOpcode() != ISD::SELECT_CC ||
690 !TLI.isConstTrueVal(N.getOperand(2).getNode()) ||
691 !TLI.isConstFalseVal(N.getOperand(3).getNode()))
694 if (TLI.getBooleanContents(N.getValueType()) ==
695 TargetLowering::UndefinedBooleanContent)
698 LHS = N.getOperand(0);
699 RHS = N.getOperand(1);
700 CC = N.getOperand(4);
704 /// Return true if this is a SetCC-equivalent operation with only one use.
705 /// If this is true, it allows the users to invert the operation for free when
706 /// it is profitable to do so.
707 bool DAGCombiner::isOneUseSetCC(SDValue N) const {
709 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
714 /// Returns true if N is a BUILD_VECTOR node whose
715 /// elements are all the same constant or undefined.
716 static bool isConstantSplatVector(SDNode *N, APInt& SplatValue) {
717 BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(N);
722 unsigned SplatBitSize;
724 EVT EltVT = N->getValueType(0).getVectorElementType();
725 return (C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
727 EltVT.getSizeInBits() >= SplatBitSize);
730 // \brief Returns the SDNode if it is a constant integer BuildVector
731 // or constant integer.
732 static SDNode *isConstantIntBuildVectorOrConstantInt(SDValue N) {
733 if (isa<ConstantSDNode>(N))
735 if (ISD::isBuildVectorOfConstantSDNodes(N.getNode()))
740 // \brief Returns the SDNode if it is a constant float BuildVector
741 // or constant float.
742 static SDNode *isConstantFPBuildVectorOrConstantFP(SDValue N) {
743 if (isa<ConstantFPSDNode>(N))
745 if (ISD::isBuildVectorOfConstantFPSDNodes(N.getNode()))
750 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
752 static ConstantSDNode *isConstOrConstSplat(SDValue N) {
753 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N))
756 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
757 BitVector UndefElements;
758 ConstantSDNode *CN = BV->getConstantSplatNode(&UndefElements);
760 // BuildVectors can truncate their operands. Ignore that case here.
761 // FIXME: We blindly ignore splats which include undef which is overly
763 if (CN && UndefElements.none() &&
764 CN->getValueType(0) == N.getValueType().getScalarType())
771 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
773 static ConstantFPSDNode *isConstOrConstSplatFP(SDValue N) {
774 if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N))
777 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
778 BitVector UndefElements;
779 ConstantFPSDNode *CN = BV->getConstantFPSplatNode(&UndefElements);
781 if (CN && UndefElements.none())
788 SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL,
789 SDValue N0, SDValue N1) {
790 EVT VT = N0.getValueType();
791 if (N0.getOpcode() == Opc) {
792 if (SDNode *L = isConstantIntBuildVectorOrConstantInt(N0.getOperand(1))) {
793 if (SDNode *R = isConstantIntBuildVectorOrConstantInt(N1)) {
794 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
795 if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, DL, VT, L, R))
796 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
799 if (N0.hasOneUse()) {
800 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one
802 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1);
803 if (!OpNode.getNode())
805 AddToWorklist(OpNode.getNode());
806 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
811 if (N1.getOpcode() == Opc) {
812 if (SDNode *R = isConstantIntBuildVectorOrConstantInt(N1.getOperand(1))) {
813 if (SDNode *L = isConstantIntBuildVectorOrConstantInt(N0)) {
814 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
815 if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, DL, VT, R, L))
816 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
819 if (N1.hasOneUse()) {
820 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one
822 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N1.getOperand(0), N0);
823 if (!OpNode.getNode())
825 AddToWorklist(OpNode.getNode());
826 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
834 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
836 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
838 DEBUG(dbgs() << "\nReplacing.1 ";
840 dbgs() << "\nWith: ";
841 To[0].getNode()->dump(&DAG);
842 dbgs() << " and " << NumTo-1 << " other values\n");
843 for (unsigned i = 0, e = NumTo; i != e; ++i)
844 assert((!To[i].getNode() ||
845 N->getValueType(i) == To[i].getValueType()) &&
846 "Cannot combine value to value of different type!");
848 WorklistRemover DeadNodes(*this);
849 DAG.ReplaceAllUsesWith(N, To);
851 // Push the new nodes and any users onto the worklist
852 for (unsigned i = 0, e = NumTo; i != e; ++i) {
853 if (To[i].getNode()) {
854 AddToWorklist(To[i].getNode());
855 AddUsersToWorklist(To[i].getNode());
860 // Finally, if the node is now dead, remove it from the graph. The node
861 // may not be dead if the replacement process recursively simplified to
862 // something else needing this node.
864 deleteAndRecombine(N);
865 return SDValue(N, 0);
869 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
870 // Replace all uses. If any nodes become isomorphic to other nodes and
871 // are deleted, make sure to remove them from our worklist.
872 WorklistRemover DeadNodes(*this);
873 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New);
875 // Push the new node and any (possibly new) users onto the worklist.
876 AddToWorklist(TLO.New.getNode());
877 AddUsersToWorklist(TLO.New.getNode());
879 // Finally, if the node is now dead, remove it from the graph. The node
880 // may not be dead if the replacement process recursively simplified to
881 // something else needing this node.
882 if (TLO.Old.getNode()->use_empty())
883 deleteAndRecombine(TLO.Old.getNode());
886 /// Check the specified integer node value to see if it can be simplified or if
887 /// things it uses can be simplified by bit propagation. If so, return true.
888 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
889 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
890 APInt KnownZero, KnownOne;
891 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
895 AddToWorklist(Op.getNode());
897 // Replace the old value with the new one.
899 DEBUG(dbgs() << "\nReplacing.2 ";
900 TLO.Old.getNode()->dump(&DAG);
901 dbgs() << "\nWith: ";
902 TLO.New.getNode()->dump(&DAG);
905 CommitTargetLoweringOpt(TLO);
909 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
911 EVT VT = Load->getValueType(0);
912 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
914 DEBUG(dbgs() << "\nReplacing.9 ";
916 dbgs() << "\nWith: ";
917 Trunc.getNode()->dump(&DAG);
919 WorklistRemover DeadNodes(*this);
920 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc);
921 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1));
922 deleteAndRecombine(Load);
923 AddToWorklist(Trunc.getNode());
926 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
929 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
930 EVT MemVT = LD->getMemoryVT();
931 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
932 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, PVT, MemVT) ? ISD::ZEXTLOAD
934 : LD->getExtensionType();
936 return DAG.getExtLoad(ExtType, dl, PVT,
937 LD->getChain(), LD->getBasePtr(),
938 MemVT, LD->getMemOperand());
941 unsigned Opc = Op.getOpcode();
944 case ISD::AssertSext:
945 return DAG.getNode(ISD::AssertSext, dl, PVT,
946 SExtPromoteOperand(Op.getOperand(0), PVT),
948 case ISD::AssertZext:
949 return DAG.getNode(ISD::AssertZext, dl, PVT,
950 ZExtPromoteOperand(Op.getOperand(0), PVT),
952 case ISD::Constant: {
954 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
955 return DAG.getNode(ExtOpc, dl, PVT, Op);
959 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
961 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
964 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
965 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
967 EVT OldVT = Op.getValueType();
969 bool Replace = false;
970 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
971 if (!NewOp.getNode())
973 AddToWorklist(NewOp.getNode());
976 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
977 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
978 DAG.getValueType(OldVT));
981 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
982 EVT OldVT = Op.getValueType();
984 bool Replace = false;
985 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
986 if (!NewOp.getNode())
988 AddToWorklist(NewOp.getNode());
991 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
992 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
995 /// Promote the specified integer binary operation if the target indicates it is
996 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
997 /// i32 since i16 instructions are longer.
998 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
999 if (!LegalOperations)
1002 EVT VT = Op.getValueType();
1003 if (VT.isVector() || !VT.isInteger())
1006 // If operation type is 'undesirable', e.g. i16 on x86, consider
1008 unsigned Opc = Op.getOpcode();
1009 if (TLI.isTypeDesirableForOp(Opc, VT))
1013 // Consult target whether it is a good idea to promote this operation and
1014 // what's the right type to promote it to.
1015 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1016 assert(PVT != VT && "Don't know what type to promote to!");
1018 bool Replace0 = false;
1019 SDValue N0 = Op.getOperand(0);
1020 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
1024 bool Replace1 = false;
1025 SDValue N1 = Op.getOperand(1);
1030 NN1 = PromoteOperand(N1, PVT, Replace1);
1035 AddToWorklist(NN0.getNode());
1037 AddToWorklist(NN1.getNode());
1040 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
1042 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
1044 DEBUG(dbgs() << "\nPromoting ";
1045 Op.getNode()->dump(&DAG));
1047 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1048 DAG.getNode(Opc, dl, PVT, NN0, NN1));
1053 /// Promote the specified integer shift operation if the target indicates it is
1054 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
1055 /// i32 since i16 instructions are longer.
1056 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
1057 if (!LegalOperations)
1060 EVT VT = Op.getValueType();
1061 if (VT.isVector() || !VT.isInteger())
1064 // If operation type is 'undesirable', e.g. i16 on x86, consider
1066 unsigned Opc = Op.getOpcode();
1067 if (TLI.isTypeDesirableForOp(Opc, VT))
1071 // Consult target whether it is a good idea to promote this operation and
1072 // what's the right type to promote it to.
1073 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1074 assert(PVT != VT && "Don't know what type to promote to!");
1076 bool Replace = false;
1077 SDValue N0 = Op.getOperand(0);
1078 if (Opc == ISD::SRA)
1079 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
1080 else if (Opc == ISD::SRL)
1081 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
1083 N0 = PromoteOperand(N0, PVT, Replace);
1087 AddToWorklist(N0.getNode());
1089 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
1091 DEBUG(dbgs() << "\nPromoting ";
1092 Op.getNode()->dump(&DAG));
1094 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1095 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
1100 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
1101 if (!LegalOperations)
1104 EVT VT = Op.getValueType();
1105 if (VT.isVector() || !VT.isInteger())
1108 // If operation type is 'undesirable', e.g. i16 on x86, consider
1110 unsigned Opc = Op.getOpcode();
1111 if (TLI.isTypeDesirableForOp(Opc, VT))
1115 // Consult target whether it is a good idea to promote this operation and
1116 // what's the right type to promote it to.
1117 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1118 assert(PVT != VT && "Don't know what type to promote to!");
1119 // fold (aext (aext x)) -> (aext x)
1120 // fold (aext (zext x)) -> (zext x)
1121 // fold (aext (sext x)) -> (sext x)
1122 DEBUG(dbgs() << "\nPromoting ";
1123 Op.getNode()->dump(&DAG));
1124 return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0));
1129 bool DAGCombiner::PromoteLoad(SDValue Op) {
1130 if (!LegalOperations)
1133 EVT VT = Op.getValueType();
1134 if (VT.isVector() || !VT.isInteger())
1137 // If operation type is 'undesirable', e.g. i16 on x86, consider
1139 unsigned Opc = Op.getOpcode();
1140 if (TLI.isTypeDesirableForOp(Opc, VT))
1144 // Consult target whether it is a good idea to promote this operation and
1145 // what's the right type to promote it to.
1146 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1147 assert(PVT != VT && "Don't know what type to promote to!");
1150 SDNode *N = Op.getNode();
1151 LoadSDNode *LD = cast<LoadSDNode>(N);
1152 EVT MemVT = LD->getMemoryVT();
1153 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
1154 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, PVT, MemVT) ? ISD::ZEXTLOAD
1156 : LD->getExtensionType();
1157 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
1158 LD->getChain(), LD->getBasePtr(),
1159 MemVT, LD->getMemOperand());
1160 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
1162 DEBUG(dbgs() << "\nPromoting ";
1165 Result.getNode()->dump(&DAG);
1167 WorklistRemover DeadNodes(*this);
1168 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
1169 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1));
1170 deleteAndRecombine(N);
1171 AddToWorklist(Result.getNode());
1177 /// \brief Recursively delete a node which has no uses and any operands for
1178 /// which it is the only use.
1180 /// Note that this both deletes the nodes and removes them from the worklist.
1181 /// It also adds any nodes who have had a user deleted to the worklist as they
1182 /// may now have only one use and subject to other combines.
1183 bool DAGCombiner::recursivelyDeleteUnusedNodes(SDNode *N) {
1184 if (!N->use_empty())
1187 SmallSetVector<SDNode *, 16> Nodes;
1190 N = Nodes.pop_back_val();
1194 if (N->use_empty()) {
1195 for (const SDValue &ChildN : N->op_values())
1196 Nodes.insert(ChildN.getNode());
1198 removeFromWorklist(N);
1203 } while (!Nodes.empty());
1207 //===----------------------------------------------------------------------===//
1208 // Main DAG Combiner implementation
1209 //===----------------------------------------------------------------------===//
1211 void DAGCombiner::Run(CombineLevel AtLevel) {
1212 // set the instance variables, so that the various visit routines may use it.
1214 LegalOperations = Level >= AfterLegalizeVectorOps;
1215 LegalTypes = Level >= AfterLegalizeTypes;
1217 // Add all the dag nodes to the worklist.
1218 for (SDNode &Node : DAG.allnodes())
1219 AddToWorklist(&Node);
1221 // Create a dummy node (which is not added to allnodes), that adds a reference
1222 // to the root node, preventing it from being deleted, and tracking any
1223 // changes of the root.
1224 HandleSDNode Dummy(DAG.getRoot());
1226 // while the worklist isn't empty, find a node and
1227 // try and combine it.
1228 while (!WorklistMap.empty()) {
1230 // The Worklist holds the SDNodes in order, but it may contain null entries.
1232 N = Worklist.pop_back_val();
1235 bool GoodWorklistEntry = WorklistMap.erase(N);
1236 (void)GoodWorklistEntry;
1237 assert(GoodWorklistEntry &&
1238 "Found a worklist entry without a corresponding map entry!");
1240 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1241 // N is deleted from the DAG, since they too may now be dead or may have a
1242 // reduced number of uses, allowing other xforms.
1243 if (recursivelyDeleteUnusedNodes(N))
1246 WorklistRemover DeadNodes(*this);
1248 // If this combine is running after legalizing the DAG, re-legalize any
1249 // nodes pulled off the worklist.
1250 if (Level == AfterLegalizeDAG) {
1251 SmallSetVector<SDNode *, 16> UpdatedNodes;
1252 bool NIsValid = DAG.LegalizeOp(N, UpdatedNodes);
1254 for (SDNode *LN : UpdatedNodes) {
1256 AddUsersToWorklist(LN);
1262 DEBUG(dbgs() << "\nCombining: "; N->dump(&DAG));
1264 // Add any operands of the new node which have not yet been combined to the
1265 // worklist as well. Because the worklist uniques things already, this
1266 // won't repeatedly process the same operand.
1267 CombinedNodes.insert(N);
1268 for (const SDValue &ChildN : N->op_values())
1269 if (!CombinedNodes.count(ChildN.getNode()))
1270 AddToWorklist(ChildN.getNode());
1272 SDValue RV = combine(N);
1279 // If we get back the same node we passed in, rather than a new node or
1280 // zero, we know that the node must have defined multiple values and
1281 // CombineTo was used. Since CombineTo takes care of the worklist
1282 // mechanics for us, we have no work to do in this case.
1283 if (RV.getNode() == N)
1286 assert(N->getOpcode() != ISD::DELETED_NODE &&
1287 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1288 "Node was deleted but visit returned new node!");
1290 DEBUG(dbgs() << " ... into: ";
1291 RV.getNode()->dump(&DAG));
1293 // Transfer debug value.
1294 DAG.TransferDbgValues(SDValue(N, 0), RV);
1295 if (N->getNumValues() == RV.getNode()->getNumValues())
1296 DAG.ReplaceAllUsesWith(N, RV.getNode());
1298 assert(N->getValueType(0) == RV.getValueType() &&
1299 N->getNumValues() == 1 && "Type mismatch");
1301 DAG.ReplaceAllUsesWith(N, &OpV);
1304 // Push the new node and any users onto the worklist
1305 AddToWorklist(RV.getNode());
1306 AddUsersToWorklist(RV.getNode());
1308 // Finally, if the node is now dead, remove it from the graph. The node
1309 // may not be dead if the replacement process recursively simplified to
1310 // something else needing this node. This will also take care of adding any
1311 // operands which have lost a user to the worklist.
1312 recursivelyDeleteUnusedNodes(N);
1315 // If the root changed (e.g. it was a dead load, update the root).
1316 DAG.setRoot(Dummy.getValue());
1317 DAG.RemoveDeadNodes();
1320 SDValue DAGCombiner::visit(SDNode *N) {
1321 switch (N->getOpcode()) {
1323 case ISD::TokenFactor: return visitTokenFactor(N);
1324 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1325 case ISD::ADD: return visitADD(N);
1326 case ISD::SUB: return visitSUB(N);
1327 case ISD::ADDC: return visitADDC(N);
1328 case ISD::SUBC: return visitSUBC(N);
1329 case ISD::ADDE: return visitADDE(N);
1330 case ISD::SUBE: return visitSUBE(N);
1331 case ISD::MUL: return visitMUL(N);
1332 case ISD::SDIV: return visitSDIV(N);
1333 case ISD::UDIV: return visitUDIV(N);
1334 case ISD::SREM: return visitSREM(N);
1335 case ISD::UREM: return visitUREM(N);
1336 case ISD::MULHU: return visitMULHU(N);
1337 case ISD::MULHS: return visitMULHS(N);
1338 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1339 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1340 case ISD::SMULO: return visitSMULO(N);
1341 case ISD::UMULO: return visitUMULO(N);
1342 case ISD::SDIVREM: return visitSDIVREM(N);
1343 case ISD::UDIVREM: return visitUDIVREM(N);
1344 case ISD::AND: return visitAND(N);
1345 case ISD::OR: return visitOR(N);
1346 case ISD::XOR: return visitXOR(N);
1347 case ISD::SHL: return visitSHL(N);
1348 case ISD::SRA: return visitSRA(N);
1349 case ISD::SRL: return visitSRL(N);
1351 case ISD::ROTL: return visitRotate(N);
1352 case ISD::BSWAP: return visitBSWAP(N);
1353 case ISD::CTLZ: return visitCTLZ(N);
1354 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1355 case ISD::CTTZ: return visitCTTZ(N);
1356 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1357 case ISD::CTPOP: return visitCTPOP(N);
1358 case ISD::SELECT: return visitSELECT(N);
1359 case ISD::VSELECT: return visitVSELECT(N);
1360 case ISD::SELECT_CC: return visitSELECT_CC(N);
1361 case ISD::SETCC: return visitSETCC(N);
1362 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1363 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1364 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1365 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1366 case ISD::SIGN_EXTEND_VECTOR_INREG: return visitSIGN_EXTEND_VECTOR_INREG(N);
1367 case ISD::TRUNCATE: return visitTRUNCATE(N);
1368 case ISD::BITCAST: return visitBITCAST(N);
1369 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1370 case ISD::FADD: return visitFADD(N);
1371 case ISD::FSUB: return visitFSUB(N);
1372 case ISD::FMUL: return visitFMUL(N);
1373 case ISD::FMA: return visitFMA(N);
1374 case ISD::FDIV: return visitFDIV(N);
1375 case ISD::FREM: return visitFREM(N);
1376 case ISD::FSQRT: return visitFSQRT(N);
1377 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1378 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1379 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1380 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1381 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1382 case ISD::FP_ROUND: return visitFP_ROUND(N);
1383 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1384 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1385 case ISD::FNEG: return visitFNEG(N);
1386 case ISD::FABS: return visitFABS(N);
1387 case ISD::FFLOOR: return visitFFLOOR(N);
1388 case ISD::FMINNUM: return visitFMINNUM(N);
1389 case ISD::FMAXNUM: return visitFMAXNUM(N);
1390 case ISD::FCEIL: return visitFCEIL(N);
1391 case ISD::FTRUNC: return visitFTRUNC(N);
1392 case ISD::BRCOND: return visitBRCOND(N);
1393 case ISD::BR_CC: return visitBR_CC(N);
1394 case ISD::LOAD: return visitLOAD(N);
1395 case ISD::STORE: return visitSTORE(N);
1396 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1397 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1398 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1399 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1400 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1401 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1402 case ISD::SCALAR_TO_VECTOR: return visitSCALAR_TO_VECTOR(N);
1403 case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N);
1404 case ISD::MGATHER: return visitMGATHER(N);
1405 case ISD::MLOAD: return visitMLOAD(N);
1406 case ISD::MSCATTER: return visitMSCATTER(N);
1407 case ISD::MSTORE: return visitMSTORE(N);
1408 case ISD::FP_TO_FP16: return visitFP_TO_FP16(N);
1413 SDValue DAGCombiner::combine(SDNode *N) {
1414 SDValue RV = visit(N);
1416 // If nothing happened, try a target-specific DAG combine.
1417 if (!RV.getNode()) {
1418 assert(N->getOpcode() != ISD::DELETED_NODE &&
1419 "Node was deleted but visit returned NULL!");
1421 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1422 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1424 // Expose the DAG combiner to the target combiner impls.
1425 TargetLowering::DAGCombinerInfo
1426 DagCombineInfo(DAG, Level, false, this);
1428 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1432 // If nothing happened still, try promoting the operation.
1433 if (!RV.getNode()) {
1434 switch (N->getOpcode()) {
1442 RV = PromoteIntBinOp(SDValue(N, 0));
1447 RV = PromoteIntShiftOp(SDValue(N, 0));
1449 case ISD::SIGN_EXTEND:
1450 case ISD::ZERO_EXTEND:
1451 case ISD::ANY_EXTEND:
1452 RV = PromoteExtend(SDValue(N, 0));
1455 if (PromoteLoad(SDValue(N, 0)))
1461 // If N is a commutative binary node, try commuting it to enable more
1463 if (!RV.getNode() && SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1464 N->getNumValues() == 1) {
1465 SDValue N0 = N->getOperand(0);
1466 SDValue N1 = N->getOperand(1);
1468 // Constant operands are canonicalized to RHS.
1469 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1470 SDValue Ops[] = {N1, N0};
1472 if (const auto *BinNode = dyn_cast<BinaryWithFlagsSDNode>(N)) {
1473 CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops,
1476 CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops);
1479 return SDValue(CSENode, 0);
1486 /// Given a node, return its input chain if it has one, otherwise return a null
1488 static SDValue getInputChainForNode(SDNode *N) {
1489 if (unsigned NumOps = N->getNumOperands()) {
1490 if (N->getOperand(0).getValueType() == MVT::Other)
1491 return N->getOperand(0);
1492 if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1493 return N->getOperand(NumOps-1);
1494 for (unsigned i = 1; i < NumOps-1; ++i)
1495 if (N->getOperand(i).getValueType() == MVT::Other)
1496 return N->getOperand(i);
1501 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1502 // If N has two operands, where one has an input chain equal to the other,
1503 // the 'other' chain is redundant.
1504 if (N->getNumOperands() == 2) {
1505 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1506 return N->getOperand(0);
1507 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1508 return N->getOperand(1);
1511 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1512 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1513 SmallPtrSet<SDNode*, 16> SeenOps;
1514 bool Changed = false; // If we should replace this token factor.
1516 // Start out with this token factor.
1519 // Iterate through token factors. The TFs grows when new token factors are
1521 for (unsigned i = 0; i < TFs.size(); ++i) {
1522 SDNode *TF = TFs[i];
1524 // Check each of the operands.
1525 for (const SDValue &Op : TF->op_values()) {
1527 switch (Op.getOpcode()) {
1528 case ISD::EntryToken:
1529 // Entry tokens don't need to be added to the list. They are
1534 case ISD::TokenFactor:
1535 if (Op.hasOneUse() &&
1536 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1537 // Queue up for processing.
1538 TFs.push_back(Op.getNode());
1539 // Clean up in case the token factor is removed.
1540 AddToWorklist(Op.getNode());
1547 // Only add if it isn't already in the list.
1548 if (SeenOps.insert(Op.getNode()).second)
1559 // If we've changed things around then replace token factor.
1562 // The entry token is the only possible outcome.
1563 Result = DAG.getEntryNode();
1565 // New and improved token factor.
1566 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Ops);
1569 // Add users to worklist if AA is enabled, since it may introduce
1570 // a lot of new chained token factors while removing memory deps.
1571 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
1572 : DAG.getSubtarget().useAA();
1573 return CombineTo(N, Result, UseAA /*add to worklist*/);
1579 /// MERGE_VALUES can always be eliminated.
1580 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1581 WorklistRemover DeadNodes(*this);
1582 // Replacing results may cause a different MERGE_VALUES to suddenly
1583 // be CSE'd with N, and carry its uses with it. Iterate until no
1584 // uses remain, to ensure that the node can be safely deleted.
1585 // First add the users of this node to the work list so that they
1586 // can be tried again once they have new operands.
1587 AddUsersToWorklist(N);
1589 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1590 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1591 } while (!N->use_empty());
1592 deleteAndRecombine(N);
1593 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1596 static bool isNullConstant(SDValue V) {
1597 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(V);
1598 return Const != nullptr && Const->isNullValue();
1601 static bool isNullFPConstant(SDValue V) {
1602 ConstantFPSDNode *Const = dyn_cast<ConstantFPSDNode>(V);
1603 return Const != nullptr && Const->isZero() && !Const->isNegative();
1606 static bool isAllOnesConstant(SDValue V) {
1607 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(V);
1608 return Const != nullptr && Const->isAllOnesValue();
1611 static bool isOneConstant(SDValue V) {
1612 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(V);
1613 return Const != nullptr && Const->isOne();
1616 /// If \p N is a ContantSDNode with isOpaque() == false return it casted to a
1617 /// ContantSDNode pointer else nullptr.
1618 static ConstantSDNode *getAsNonOpaqueConstant(SDValue N) {
1619 ConstantSDNode *Const = dyn_cast<ConstantSDNode>(N);
1620 return Const != nullptr && !Const->isOpaque() ? Const : nullptr;
1623 SDValue DAGCombiner::visitADD(SDNode *N) {
1624 SDValue N0 = N->getOperand(0);
1625 SDValue N1 = N->getOperand(1);
1626 EVT VT = N0.getValueType();
1629 if (VT.isVector()) {
1630 if (SDValue FoldedVOp = SimplifyVBinOp(N))
1633 // fold (add x, 0) -> x, vector edition
1634 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1636 if (ISD::isBuildVectorAllZeros(N0.getNode()))
1640 // fold (add x, undef) -> undef
1641 if (N0.getOpcode() == ISD::UNDEF)
1643 if (N1.getOpcode() == ISD::UNDEF)
1645 // fold (add c1, c2) -> c1+c2
1646 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
1647 ConstantSDNode *N1C = getAsNonOpaqueConstant(N1);
1649 return DAG.FoldConstantArithmetic(ISD::ADD, SDLoc(N), VT, N0C, N1C);
1650 // canonicalize constant to RHS
1651 if (isConstantIntBuildVectorOrConstantInt(N0) &&
1652 !isConstantIntBuildVectorOrConstantInt(N1))
1653 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, N0);
1654 // fold (add x, 0) -> x
1655 if (isNullConstant(N1))
1657 // fold (add Sym, c) -> Sym+c
1658 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1659 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1660 GA->getOpcode() == ISD::GlobalAddress)
1661 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1663 (uint64_t)N1C->getSExtValue());
1664 // fold ((c1-A)+c2) -> (c1+c2)-A
1665 if (N1C && N0.getOpcode() == ISD::SUB)
1666 if (ConstantSDNode *N0C = getAsNonOpaqueConstant(N0.getOperand(0))) {
1668 return DAG.getNode(ISD::SUB, DL, VT,
1669 DAG.getConstant(N1C->getAPIntValue()+
1670 N0C->getAPIntValue(), DL, VT),
1674 if (SDValue RADD = ReassociateOps(ISD::ADD, SDLoc(N), N0, N1))
1676 // fold ((0-A) + B) -> B-A
1677 if (N0.getOpcode() == ISD::SUB && isNullConstant(N0.getOperand(0)))
1678 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, N0.getOperand(1));
1679 // fold (A + (0-B)) -> A-B
1680 if (N1.getOpcode() == ISD::SUB && isNullConstant(N1.getOperand(0)))
1681 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1.getOperand(1));
1682 // fold (A+(B-A)) -> B
1683 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1684 return N1.getOperand(0);
1685 // fold ((B-A)+A) -> B
1686 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1687 return N0.getOperand(0);
1688 // fold (A+(B-(A+C))) to (B-C)
1689 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1690 N0 == N1.getOperand(1).getOperand(0))
1691 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1692 N1.getOperand(1).getOperand(1));
1693 // fold (A+(B-(C+A))) to (B-C)
1694 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1695 N0 == N1.getOperand(1).getOperand(1))
1696 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1697 N1.getOperand(1).getOperand(0));
1698 // fold (A+((B-A)+or-C)) to (B+or-C)
1699 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1700 N1.getOperand(0).getOpcode() == ISD::SUB &&
1701 N0 == N1.getOperand(0).getOperand(1))
1702 return DAG.getNode(N1.getOpcode(), SDLoc(N), VT,
1703 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1705 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1706 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1707 SDValue N00 = N0.getOperand(0);
1708 SDValue N01 = N0.getOperand(1);
1709 SDValue N10 = N1.getOperand(0);
1710 SDValue N11 = N1.getOperand(1);
1712 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1713 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1714 DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10),
1715 DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11));
1718 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1719 return SDValue(N, 0);
1721 // fold (a+b) -> (a|b) iff a and b share no bits.
1722 if (VT.isInteger() && !VT.isVector()) {
1723 APInt LHSZero, LHSOne;
1724 APInt RHSZero, RHSOne;
1725 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1727 if (LHSZero.getBoolValue()) {
1728 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1730 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1731 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1732 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero){
1733 if (!LegalOperations || TLI.isOperationLegal(ISD::OR, VT))
1734 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1);
1739 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1740 if (N1.getOpcode() == ISD::SHL && N1.getOperand(0).getOpcode() == ISD::SUB &&
1741 isNullConstant(N1.getOperand(0).getOperand(0)))
1742 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0,
1743 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1744 N1.getOperand(0).getOperand(1),
1746 if (N0.getOpcode() == ISD::SHL && N0.getOperand(0).getOpcode() == ISD::SUB &&
1747 isNullConstant(N0.getOperand(0).getOperand(0)))
1748 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1,
1749 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1750 N0.getOperand(0).getOperand(1),
1753 if (N1.getOpcode() == ISD::AND) {
1754 SDValue AndOp0 = N1.getOperand(0);
1755 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1756 unsigned DestBits = VT.getScalarType().getSizeInBits();
1758 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1759 // and similar xforms where the inner op is either ~0 or 0.
1760 if (NumSignBits == DestBits && isOneConstant(N1->getOperand(1))) {
1762 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1766 // add (sext i1), X -> sub X, (zext i1)
1767 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1768 N0.getOperand(0).getValueType() == MVT::i1 &&
1769 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1771 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1772 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1775 // add X, (sextinreg Y i1) -> sub X, (and Y 1)
1776 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
1777 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
1778 if (TN->getVT() == MVT::i1) {
1780 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
1781 DAG.getConstant(1, DL, VT));
1782 return DAG.getNode(ISD::SUB, DL, VT, N0, ZExt);
1789 SDValue DAGCombiner::visitADDC(SDNode *N) {
1790 SDValue N0 = N->getOperand(0);
1791 SDValue N1 = N->getOperand(1);
1792 EVT VT = N0.getValueType();
1794 // If the flag result is dead, turn this into an ADD.
1795 if (!N->hasAnyUseOfValue(1))
1796 return CombineTo(N, DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N1),
1797 DAG.getNode(ISD::CARRY_FALSE,
1798 SDLoc(N), MVT::Glue));
1800 // canonicalize constant to RHS.
1801 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1802 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1804 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N1, N0);
1806 // fold (addc x, 0) -> x + no carry out
1807 if (isNullConstant(N1))
1808 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1809 SDLoc(N), MVT::Glue));
1811 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1812 APInt LHSZero, LHSOne;
1813 APInt RHSZero, RHSOne;
1814 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1816 if (LHSZero.getBoolValue()) {
1817 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1819 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1820 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1821 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1822 return CombineTo(N, DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1),
1823 DAG.getNode(ISD::CARRY_FALSE,
1824 SDLoc(N), MVT::Glue));
1830 SDValue DAGCombiner::visitADDE(SDNode *N) {
1831 SDValue N0 = N->getOperand(0);
1832 SDValue N1 = N->getOperand(1);
1833 SDValue CarryIn = N->getOperand(2);
1835 // canonicalize constant to RHS
1836 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1837 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1839 return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(),
1842 // fold (adde x, y, false) -> (addc x, y)
1843 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1844 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1);
1849 // Since it may not be valid to emit a fold to zero for vector initializers
1850 // check if we can before folding.
1851 static SDValue tryFoldToZero(SDLoc DL, const TargetLowering &TLI, EVT VT,
1853 bool LegalOperations, bool LegalTypes) {
1855 return DAG.getConstant(0, DL, VT);
1856 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
1857 return DAG.getConstant(0, DL, VT);
1861 SDValue DAGCombiner::visitSUB(SDNode *N) {
1862 SDValue N0 = N->getOperand(0);
1863 SDValue N1 = N->getOperand(1);
1864 EVT VT = N0.getValueType();
1867 if (VT.isVector()) {
1868 if (SDValue FoldedVOp = SimplifyVBinOp(N))
1871 // fold (sub x, 0) -> x, vector edition
1872 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1876 // fold (sub x, x) -> 0
1877 // FIXME: Refactor this and xor and other similar operations together.
1879 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
1880 // fold (sub c1, c2) -> c1-c2
1881 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
1882 ConstantSDNode *N1C = getAsNonOpaqueConstant(N1);
1884 return DAG.FoldConstantArithmetic(ISD::SUB, SDLoc(N), VT, N0C, N1C);
1885 // fold (sub x, c) -> (add x, -c)
1888 return DAG.getNode(ISD::ADD, DL, VT, N0,
1889 DAG.getConstant(-N1C->getAPIntValue(), DL, VT));
1891 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1892 if (isAllOnesConstant(N0))
1893 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
1894 // fold A-(A-B) -> B
1895 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1896 return N1.getOperand(1);
1897 // fold (A+B)-A -> B
1898 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1899 return N0.getOperand(1);
1900 // fold (A+B)-B -> A
1901 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1902 return N0.getOperand(0);
1903 // fold C2-(A+C1) -> (C2-C1)-A
1904 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? nullptr :
1905 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1906 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1908 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(),
1910 return DAG.getNode(ISD::SUB, DL, VT, NewC,
1913 // fold ((A+(B+or-C))-B) -> A+or-C
1914 if (N0.getOpcode() == ISD::ADD &&
1915 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1916 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1917 N0.getOperand(1).getOperand(0) == N1)
1918 return DAG.getNode(N0.getOperand(1).getOpcode(), SDLoc(N), VT,
1919 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1920 // fold ((A+(C+B))-B) -> A+C
1921 if (N0.getOpcode() == ISD::ADD &&
1922 N0.getOperand(1).getOpcode() == ISD::ADD &&
1923 N0.getOperand(1).getOperand(1) == N1)
1924 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1925 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1926 // fold ((A-(B-C))-C) -> A-B
1927 if (N0.getOpcode() == ISD::SUB &&
1928 N0.getOperand(1).getOpcode() == ISD::SUB &&
1929 N0.getOperand(1).getOperand(1) == N1)
1930 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1931 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1933 // If either operand of a sub is undef, the result is undef
1934 if (N0.getOpcode() == ISD::UNDEF)
1936 if (N1.getOpcode() == ISD::UNDEF)
1939 // If the relocation model supports it, consider symbol offsets.
1940 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1941 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1942 // fold (sub Sym, c) -> Sym-c
1943 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1944 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1946 (uint64_t)N1C->getSExtValue());
1947 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1948 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1949 if (GA->getGlobal() == GB->getGlobal())
1950 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1954 // sub X, (sextinreg Y i1) -> add X, (and Y 1)
1955 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
1956 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
1957 if (TN->getVT() == MVT::i1) {
1959 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
1960 DAG.getConstant(1, DL, VT));
1961 return DAG.getNode(ISD::ADD, DL, VT, N0, ZExt);
1968 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1969 SDValue N0 = N->getOperand(0);
1970 SDValue N1 = N->getOperand(1);
1971 EVT VT = N0.getValueType();
1973 // If the flag result is dead, turn this into an SUB.
1974 if (!N->hasAnyUseOfValue(1))
1975 return CombineTo(N, DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1),
1976 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1979 // fold (subc x, x) -> 0 + no borrow
1982 return CombineTo(N, DAG.getConstant(0, DL, VT),
1983 DAG.getNode(ISD::CARRY_FALSE, DL,
1987 // fold (subc x, 0) -> x + no borrow
1988 if (isNullConstant(N1))
1989 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1992 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1993 if (isAllOnesConstant(N0))
1994 return CombineTo(N, DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0),
1995 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
2001 SDValue DAGCombiner::visitSUBE(SDNode *N) {
2002 SDValue N0 = N->getOperand(0);
2003 SDValue N1 = N->getOperand(1);
2004 SDValue CarryIn = N->getOperand(2);
2006 // fold (sube x, y, false) -> (subc x, y)
2007 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
2008 return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1);
2013 SDValue DAGCombiner::visitMUL(SDNode *N) {
2014 SDValue N0 = N->getOperand(0);
2015 SDValue N1 = N->getOperand(1);
2016 EVT VT = N0.getValueType();
2018 // fold (mul x, undef) -> 0
2019 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2020 return DAG.getConstant(0, SDLoc(N), VT);
2022 bool N0IsConst = false;
2023 bool N1IsConst = false;
2024 bool N1IsOpaqueConst = false;
2025 bool N0IsOpaqueConst = false;
2026 APInt ConstValue0, ConstValue1;
2028 if (VT.isVector()) {
2029 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2032 N0IsConst = isConstantSplatVector(N0.getNode(), ConstValue0);
2033 N1IsConst = isConstantSplatVector(N1.getNode(), ConstValue1);
2035 N0IsConst = isa<ConstantSDNode>(N0);
2037 ConstValue0 = cast<ConstantSDNode>(N0)->getAPIntValue();
2038 N0IsOpaqueConst = cast<ConstantSDNode>(N0)->isOpaque();
2040 N1IsConst = isa<ConstantSDNode>(N1);
2042 ConstValue1 = cast<ConstantSDNode>(N1)->getAPIntValue();
2043 N1IsOpaqueConst = cast<ConstantSDNode>(N1)->isOpaque();
2047 // fold (mul c1, c2) -> c1*c2
2048 if (N0IsConst && N1IsConst && !N0IsOpaqueConst && !N1IsOpaqueConst)
2049 return DAG.FoldConstantArithmetic(ISD::MUL, SDLoc(N), VT,
2050 N0.getNode(), N1.getNode());
2052 // canonicalize constant to RHS (vector doesn't have to splat)
2053 if (isConstantIntBuildVectorOrConstantInt(N0) &&
2054 !isConstantIntBuildVectorOrConstantInt(N1))
2055 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0);
2056 // fold (mul x, 0) -> 0
2057 if (N1IsConst && ConstValue1 == 0)
2059 // We require a splat of the entire scalar bit width for non-contiguous
2062 ConstValue1.getBitWidth() == VT.getScalarType().getSizeInBits();
2063 // fold (mul x, 1) -> x
2064 if (N1IsConst && ConstValue1 == 1 && IsFullSplat)
2066 // fold (mul x, -1) -> 0-x
2067 if (N1IsConst && ConstValue1.isAllOnesValue()) {
2069 return DAG.getNode(ISD::SUB, DL, VT,
2070 DAG.getConstant(0, DL, VT), N0);
2072 // fold (mul x, (1 << c)) -> x << c
2073 if (N1IsConst && !N1IsOpaqueConst && ConstValue1.isPowerOf2() &&
2076 return DAG.getNode(ISD::SHL, DL, VT, N0,
2077 DAG.getConstant(ConstValue1.logBase2(), DL,
2078 getShiftAmountTy(N0.getValueType())));
2080 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
2081 if (N1IsConst && !N1IsOpaqueConst && (-ConstValue1).isPowerOf2() &&
2083 unsigned Log2Val = (-ConstValue1).logBase2();
2085 // FIXME: If the input is something that is easily negated (e.g. a
2086 // single-use add), we should put the negate there.
2087 return DAG.getNode(ISD::SUB, DL, VT,
2088 DAG.getConstant(0, DL, VT),
2089 DAG.getNode(ISD::SHL, DL, VT, N0,
2090 DAG.getConstant(Log2Val, DL,
2091 getShiftAmountTy(N0.getValueType()))));
2095 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
2096 if (N1IsConst && N0.getOpcode() == ISD::SHL &&
2097 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2098 isa<ConstantSDNode>(N0.getOperand(1)))) {
2099 SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT,
2100 N1, N0.getOperand(1));
2101 AddToWorklist(C3.getNode());
2102 return DAG.getNode(ISD::MUL, SDLoc(N), VT,
2103 N0.getOperand(0), C3);
2106 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
2109 SDValue Sh(nullptr,0), Y(nullptr,0);
2110 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
2111 if (N0.getOpcode() == ISD::SHL &&
2112 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2113 isa<ConstantSDNode>(N0.getOperand(1))) &&
2114 N0.getNode()->hasOneUse()) {
2116 } else if (N1.getOpcode() == ISD::SHL &&
2117 isa<ConstantSDNode>(N1.getOperand(1)) &&
2118 N1.getNode()->hasOneUse()) {
2123 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2124 Sh.getOperand(0), Y);
2125 return DAG.getNode(ISD::SHL, SDLoc(N), VT,
2126 Mul, Sh.getOperand(1));
2130 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
2131 if (N1IsConst && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
2132 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2133 isa<ConstantSDNode>(N0.getOperand(1))))
2134 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
2135 DAG.getNode(ISD::MUL, SDLoc(N0), VT,
2136 N0.getOperand(0), N1),
2137 DAG.getNode(ISD::MUL, SDLoc(N1), VT,
2138 N0.getOperand(1), N1));
2141 if (SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1))
2147 SDValue DAGCombiner::visitSDIV(SDNode *N) {
2148 SDValue N0 = N->getOperand(0);
2149 SDValue N1 = N->getOperand(1);
2150 EVT VT = N->getValueType(0);
2154 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2157 // fold (sdiv c1, c2) -> c1/c2
2158 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2159 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2160 if (N0C && N1C && !N0C->isOpaque() && !N1C->isOpaque())
2161 return DAG.FoldConstantArithmetic(ISD::SDIV, SDLoc(N), VT, N0C, N1C);
2162 // fold (sdiv X, 1) -> X
2163 if (N1C && N1C->isOne())
2165 // fold (sdiv X, -1) -> 0-X
2166 if (N1C && N1C->isAllOnesValue()) {
2168 return DAG.getNode(ISD::SUB, DL, VT,
2169 DAG.getConstant(0, DL, VT), N0);
2171 // If we know the sign bits of both operands are zero, strength reduce to a
2172 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
2173 if (!VT.isVector()) {
2174 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2175 return DAG.getNode(ISD::UDIV, SDLoc(N), N1.getValueType(),
2179 // fold (sdiv X, pow2) -> simple ops after legalize
2180 // FIXME: We check for the exact bit here because the generic lowering gives
2181 // better results in that case. The target-specific lowering should learn how
2182 // to handle exact sdivs efficiently.
2183 if (N1C && !N1C->isNullValue() && !N1C->isOpaque() &&
2184 !cast<BinaryWithFlagsSDNode>(N)->Flags.hasExact() &&
2185 (N1C->getAPIntValue().isPowerOf2() ||
2186 (-N1C->getAPIntValue()).isPowerOf2())) {
2187 // If dividing by powers of two is cheap, then don't perform the following
2189 if (TLI.isPow2SDivCheap())
2192 // Target-specific implementation of sdiv x, pow2.
2193 if (SDValue Res = BuildSDIVPow2(N))
2196 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
2199 // Splat the sign bit into the register
2201 DAG.getNode(ISD::SRA, DL, VT, N0,
2202 DAG.getConstant(VT.getScalarSizeInBits() - 1, DL,
2203 getShiftAmountTy(N0.getValueType())));
2204 AddToWorklist(SGN.getNode());
2206 // Add (N0 < 0) ? abs2 - 1 : 0;
2208 DAG.getNode(ISD::SRL, DL, VT, SGN,
2209 DAG.getConstant(VT.getScalarSizeInBits() - lg2, DL,
2210 getShiftAmountTy(SGN.getValueType())));
2211 SDValue ADD = DAG.getNode(ISD::ADD, DL, VT, N0, SRL);
2212 AddToWorklist(SRL.getNode());
2213 AddToWorklist(ADD.getNode()); // Divide by pow2
2214 SDValue SRA = DAG.getNode(ISD::SRA, DL, VT, ADD,
2215 DAG.getConstant(lg2, DL,
2216 getShiftAmountTy(ADD.getValueType())));
2218 // If we're dividing by a positive value, we're done. Otherwise, we must
2219 // negate the result.
2220 if (N1C->getAPIntValue().isNonNegative())
2223 AddToWorklist(SRA.getNode());
2224 return DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), SRA);
2227 // If integer divide is expensive and we satisfy the requirements, emit an
2228 // alternate sequence.
2229 if (N1C && !TLI.isIntDivCheap())
2230 if (SDValue Op = BuildSDIV(N))
2234 if (N0.getOpcode() == ISD::UNDEF)
2235 return DAG.getConstant(0, SDLoc(N), VT);
2236 // X / undef -> undef
2237 if (N1.getOpcode() == ISD::UNDEF)
2243 SDValue DAGCombiner::visitUDIV(SDNode *N) {
2244 SDValue N0 = N->getOperand(0);
2245 SDValue N1 = N->getOperand(1);
2246 EVT VT = N->getValueType(0);
2250 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2253 // fold (udiv c1, c2) -> c1/c2
2254 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2255 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2257 if (SDValue Folded = DAG.FoldConstantArithmetic(ISD::UDIV, SDLoc(N), VT,
2260 // fold (udiv x, (1 << c)) -> x >>u c
2261 if (N1C && !N1C->isOpaque() && N1C->getAPIntValue().isPowerOf2()) {
2263 return DAG.getNode(ISD::SRL, DL, VT, N0,
2264 DAG.getConstant(N1C->getAPIntValue().logBase2(), DL,
2265 getShiftAmountTy(N0.getValueType())));
2267 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
2268 if (N1.getOpcode() == ISD::SHL) {
2269 if (ConstantSDNode *SHC = getAsNonOpaqueConstant(N1.getOperand(0))) {
2270 if (SHC->getAPIntValue().isPowerOf2()) {
2271 EVT ADDVT = N1.getOperand(1).getValueType();
2273 SDValue Add = DAG.getNode(ISD::ADD, DL, ADDVT,
2275 DAG.getConstant(SHC->getAPIntValue()
2278 AddToWorklist(Add.getNode());
2279 return DAG.getNode(ISD::SRL, DL, VT, N0, Add);
2283 // fold (udiv x, c) -> alternate
2284 if (N1C && !TLI.isIntDivCheap())
2285 if (SDValue Op = BuildUDIV(N))
2289 if (N0.getOpcode() == ISD::UNDEF)
2290 return DAG.getConstant(0, SDLoc(N), VT);
2291 // X / undef -> undef
2292 if (N1.getOpcode() == ISD::UNDEF)
2298 SDValue DAGCombiner::visitSREM(SDNode *N) {
2299 SDValue N0 = N->getOperand(0);
2300 SDValue N1 = N->getOperand(1);
2301 EVT VT = N->getValueType(0);
2303 // fold (srem c1, c2) -> c1%c2
2304 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2305 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2307 if (SDValue Folded = DAG.FoldConstantArithmetic(ISD::SREM, SDLoc(N), VT,
2310 // If we know the sign bits of both operands are zero, strength reduce to a
2311 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
2312 if (!VT.isVector()) {
2313 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2314 return DAG.getNode(ISD::UREM, SDLoc(N), VT, N0, N1);
2317 // If X/C can be simplified by the division-by-constant logic, lower
2318 // X%C to the equivalent of X-X/C*C.
2319 if (N1C && !N1C->isNullValue()) {
2320 SDValue Div = DAG.getNode(ISD::SDIV, SDLoc(N), VT, N0, N1);
2321 AddToWorklist(Div.getNode());
2322 SDValue OptimizedDiv = combine(Div.getNode());
2323 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2324 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2326 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2327 AddToWorklist(Mul.getNode());
2333 if (N0.getOpcode() == ISD::UNDEF)
2334 return DAG.getConstant(0, SDLoc(N), VT);
2335 // X % undef -> undef
2336 if (N1.getOpcode() == ISD::UNDEF)
2342 SDValue DAGCombiner::visitUREM(SDNode *N) {
2343 SDValue N0 = N->getOperand(0);
2344 SDValue N1 = N->getOperand(1);
2345 EVT VT = N->getValueType(0);
2347 // fold (urem c1, c2) -> c1%c2
2348 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2349 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2351 if (SDValue Folded = DAG.FoldConstantArithmetic(ISD::UREM, SDLoc(N), VT,
2354 // fold (urem x, pow2) -> (and x, pow2-1)
2355 if (N1C && !N1C->isNullValue() && !N1C->isOpaque() &&
2356 N1C->getAPIntValue().isPowerOf2()) {
2358 return DAG.getNode(ISD::AND, DL, VT, N0,
2359 DAG.getConstant(N1C->getAPIntValue() - 1, DL, VT));
2361 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2362 if (N1.getOpcode() == ISD::SHL) {
2363 if (ConstantSDNode *SHC = getAsNonOpaqueConstant(N1.getOperand(0))) {
2364 if (SHC->getAPIntValue().isPowerOf2()) {
2367 DAG.getNode(ISD::ADD, DL, VT, N1,
2368 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL,
2370 AddToWorklist(Add.getNode());
2371 return DAG.getNode(ISD::AND, DL, VT, N0, Add);
2376 // If X/C can be simplified by the division-by-constant logic, lower
2377 // X%C to the equivalent of X-X/C*C.
2378 if (N1C && !N1C->isNullValue()) {
2379 SDValue Div = DAG.getNode(ISD::UDIV, SDLoc(N), VT, N0, N1);
2380 AddToWorklist(Div.getNode());
2381 SDValue OptimizedDiv = combine(Div.getNode());
2382 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2383 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2385 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2386 AddToWorklist(Mul.getNode());
2392 if (N0.getOpcode() == ISD::UNDEF)
2393 return DAG.getConstant(0, SDLoc(N), VT);
2394 // X % undef -> undef
2395 if (N1.getOpcode() == ISD::UNDEF)
2401 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2402 SDValue N0 = N->getOperand(0);
2403 SDValue N1 = N->getOperand(1);
2404 EVT VT = N->getValueType(0);
2407 // fold (mulhs x, 0) -> 0
2408 if (isNullConstant(N1))
2410 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2411 if (isOneConstant(N1)) {
2413 return DAG.getNode(ISD::SRA, DL, N0.getValueType(), N0,
2414 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2416 getShiftAmountTy(N0.getValueType())));
2418 // fold (mulhs x, undef) -> 0
2419 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2420 return DAG.getConstant(0, SDLoc(N), VT);
2422 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2424 if (VT.isSimple() && !VT.isVector()) {
2425 MVT Simple = VT.getSimpleVT();
2426 unsigned SimpleSize = Simple.getSizeInBits();
2427 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2428 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2429 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2430 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2431 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2432 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2433 DAG.getConstant(SimpleSize, DL,
2434 getShiftAmountTy(N1.getValueType())));
2435 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2442 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2443 SDValue N0 = N->getOperand(0);
2444 SDValue N1 = N->getOperand(1);
2445 EVT VT = N->getValueType(0);
2448 // fold (mulhu x, 0) -> 0
2449 if (isNullConstant(N1))
2451 // fold (mulhu x, 1) -> 0
2452 if (isOneConstant(N1))
2453 return DAG.getConstant(0, DL, N0.getValueType());
2454 // fold (mulhu x, undef) -> 0
2455 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2456 return DAG.getConstant(0, DL, VT);
2458 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2460 if (VT.isSimple() && !VT.isVector()) {
2461 MVT Simple = VT.getSimpleVT();
2462 unsigned SimpleSize = Simple.getSizeInBits();
2463 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2464 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2465 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2466 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2467 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2468 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2469 DAG.getConstant(SimpleSize, DL,
2470 getShiftAmountTy(N1.getValueType())));
2471 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2478 /// Perform optimizations common to nodes that compute two values. LoOp and HiOp
2479 /// give the opcodes for the two computations that are being performed. Return
2480 /// true if a simplification was made.
2481 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2483 // If the high half is not needed, just compute the low half.
2484 bool HiExists = N->hasAnyUseOfValue(1);
2486 (!LegalOperations ||
2487 TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) {
2488 SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2489 return CombineTo(N, Res, Res);
2492 // If the low half is not needed, just compute the high half.
2493 bool LoExists = N->hasAnyUseOfValue(0);
2495 (!LegalOperations ||
2496 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2497 SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2498 return CombineTo(N, Res, Res);
2501 // If both halves are used, return as it is.
2502 if (LoExists && HiExists)
2505 // If the two computed results can be simplified separately, separate them.
2507 SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2508 AddToWorklist(Lo.getNode());
2509 SDValue LoOpt = combine(Lo.getNode());
2510 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2511 (!LegalOperations ||
2512 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2513 return CombineTo(N, LoOpt, LoOpt);
2517 SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2518 AddToWorklist(Hi.getNode());
2519 SDValue HiOpt = combine(Hi.getNode());
2520 if (HiOpt.getNode() && HiOpt != Hi &&
2521 (!LegalOperations ||
2522 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2523 return CombineTo(N, HiOpt, HiOpt);
2529 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2530 if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS))
2533 EVT VT = N->getValueType(0);
2536 // If the type is twice as wide is legal, transform the mulhu to a wider
2537 // multiply plus a shift.
2538 if (VT.isSimple() && !VT.isVector()) {
2539 MVT Simple = VT.getSimpleVT();
2540 unsigned SimpleSize = Simple.getSizeInBits();
2541 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2542 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2543 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2544 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2545 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2546 // Compute the high part as N1.
2547 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2548 DAG.getConstant(SimpleSize, DL,
2549 getShiftAmountTy(Lo.getValueType())));
2550 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2551 // Compute the low part as N0.
2552 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2553 return CombineTo(N, Lo, Hi);
2560 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2561 if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU))
2564 EVT VT = N->getValueType(0);
2567 // If the type is twice as wide is legal, transform the mulhu to a wider
2568 // multiply plus a shift.
2569 if (VT.isSimple() && !VT.isVector()) {
2570 MVT Simple = VT.getSimpleVT();
2571 unsigned SimpleSize = Simple.getSizeInBits();
2572 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2573 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2574 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2575 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2576 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2577 // Compute the high part as N1.
2578 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2579 DAG.getConstant(SimpleSize, DL,
2580 getShiftAmountTy(Lo.getValueType())));
2581 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2582 // Compute the low part as N0.
2583 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2584 return CombineTo(N, Lo, Hi);
2591 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2592 // (smulo x, 2) -> (saddo x, x)
2593 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2594 if (C2->getAPIntValue() == 2)
2595 return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(),
2596 N->getOperand(0), N->getOperand(0));
2601 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2602 // (umulo x, 2) -> (uaddo x, x)
2603 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2604 if (C2->getAPIntValue() == 2)
2605 return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(),
2606 N->getOperand(0), N->getOperand(0));
2611 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2612 if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM))
2618 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2619 if (SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM))
2625 /// If this is a binary operator with two operands of the same opcode, try to
2627 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2628 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2629 EVT VT = N0.getValueType();
2630 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2632 // Bail early if none of these transforms apply.
2633 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2635 // For each of OP in AND/OR/XOR:
2636 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2637 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2638 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2639 // fold (OP (bswap x), (bswap y)) -> (bswap (OP x, y))
2640 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2642 // do not sink logical op inside of a vector extend, since it may combine
2644 EVT Op0VT = N0.getOperand(0).getValueType();
2645 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2646 N0.getOpcode() == ISD::SIGN_EXTEND ||
2647 N0.getOpcode() == ISD::BSWAP ||
2648 // Avoid infinite looping with PromoteIntBinOp.
2649 (N0.getOpcode() == ISD::ANY_EXTEND &&
2650 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2651 (N0.getOpcode() == ISD::TRUNCATE &&
2652 (!TLI.isZExtFree(VT, Op0VT) ||
2653 !TLI.isTruncateFree(Op0VT, VT)) &&
2654 TLI.isTypeLegal(Op0VT))) &&
2656 Op0VT == N1.getOperand(0).getValueType() &&
2657 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2658 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2659 N0.getOperand(0).getValueType(),
2660 N0.getOperand(0), N1.getOperand(0));
2661 AddToWorklist(ORNode.getNode());
2662 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode);
2665 // For each of OP in SHL/SRL/SRA/AND...
2666 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2667 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2668 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2669 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2670 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2671 N0.getOperand(1) == N1.getOperand(1)) {
2672 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2673 N0.getOperand(0).getValueType(),
2674 N0.getOperand(0), N1.getOperand(0));
2675 AddToWorklist(ORNode.getNode());
2676 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
2677 ORNode, N0.getOperand(1));
2680 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2681 // Only perform this optimization after type legalization and before
2682 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2683 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2684 // we don't want to undo this promotion.
2685 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2687 if ((N0.getOpcode() == ISD::BITCAST ||
2688 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) &&
2689 Level == AfterLegalizeTypes) {
2690 SDValue In0 = N0.getOperand(0);
2691 SDValue In1 = N1.getOperand(0);
2692 EVT In0Ty = In0.getValueType();
2693 EVT In1Ty = In1.getValueType();
2695 // If both incoming values are integers, and the original types are the
2697 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2698 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1);
2699 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op);
2700 AddToWorklist(Op.getNode());
2705 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2706 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2707 // If both shuffles use the same mask, and both shuffle within a single
2708 // vector, then it is worthwhile to move the swizzle after the operation.
2709 // The type-legalizer generates this pattern when loading illegal
2710 // vector types from memory. In many cases this allows additional shuffle
2712 // There are other cases where moving the shuffle after the xor/and/or
2713 // is profitable even if shuffles don't perform a swizzle.
2714 // If both shuffles use the same mask, and both shuffles have the same first
2715 // or second operand, then it might still be profitable to move the shuffle
2716 // after the xor/and/or operation.
2717 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) {
2718 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2719 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2721 assert(N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() &&
2722 "Inputs to shuffles are not the same type");
2724 // Check that both shuffles use the same mask. The masks are known to be of
2725 // the same length because the result vector type is the same.
2726 // Check also that shuffles have only one use to avoid introducing extra
2728 if (SVN0->hasOneUse() && SVN1->hasOneUse() &&
2729 SVN0->getMask().equals(SVN1->getMask())) {
2730 SDValue ShOp = N0->getOperand(1);
2732 // Don't try to fold this node if it requires introducing a
2733 // build vector of all zeros that might be illegal at this stage.
2734 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2736 ShOp = DAG.getConstant(0, SDLoc(N), VT);
2741 // (AND (shuf (A, C), shuf (B, C)) -> shuf (AND (A, B), C)
2742 // (OR (shuf (A, C), shuf (B, C)) -> shuf (OR (A, B), C)
2743 // (XOR (shuf (A, C), shuf (B, C)) -> shuf (XOR (A, B), V_0)
2744 if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) {
2745 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2746 N0->getOperand(0), N1->getOperand(0));
2747 AddToWorklist(NewNode.getNode());
2748 return DAG.getVectorShuffle(VT, SDLoc(N), NewNode, ShOp,
2749 &SVN0->getMask()[0]);
2752 // Don't try to fold this node if it requires introducing a
2753 // build vector of all zeros that might be illegal at this stage.
2754 ShOp = N0->getOperand(0);
2755 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2757 ShOp = DAG.getConstant(0, SDLoc(N), VT);
2762 // (AND (shuf (C, A), shuf (C, B)) -> shuf (C, AND (A, B))
2763 // (OR (shuf (C, A), shuf (C, B)) -> shuf (C, OR (A, B))
2764 // (XOR (shuf (C, A), shuf (C, B)) -> shuf (V_0, XOR (A, B))
2765 if (N0->getOperand(0) == N1->getOperand(0) && ShOp.getNode()) {
2766 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2767 N0->getOperand(1), N1->getOperand(1));
2768 AddToWorklist(NewNode.getNode());
2769 return DAG.getVectorShuffle(VT, SDLoc(N), ShOp, NewNode,
2770 &SVN0->getMask()[0]);
2778 /// This contains all DAGCombine rules which reduce two values combined by
2779 /// an And operation to a single value. This makes them reusable in the context
2780 /// of visitSELECT(). Rules involving constants are not included as
2781 /// visitSELECT() already handles those cases.
2782 SDValue DAGCombiner::visitANDLike(SDValue N0, SDValue N1,
2783 SDNode *LocReference) {
2784 EVT VT = N1.getValueType();
2786 // fold (and x, undef) -> 0
2787 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2788 return DAG.getConstant(0, SDLoc(LocReference), VT);
2789 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2790 SDValue LL, LR, RL, RR, CC0, CC1;
2791 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2792 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2793 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2795 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2796 LL.getValueType().isInteger()) {
2797 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2798 if (isNullConstant(LR) && Op1 == ISD::SETEQ) {
2799 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2800 LR.getValueType(), LL, RL);
2801 AddToWorklist(ORNode.getNode());
2802 return DAG.getSetCC(SDLoc(LocReference), VT, ORNode, LR, Op1);
2804 if (isAllOnesConstant(LR)) {
2805 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2806 if (Op1 == ISD::SETEQ) {
2807 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(N0),
2808 LR.getValueType(), LL, RL);
2809 AddToWorklist(ANDNode.getNode());
2810 return DAG.getSetCC(SDLoc(LocReference), VT, ANDNode, LR, Op1);
2812 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2813 if (Op1 == ISD::SETGT) {
2814 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2815 LR.getValueType(), LL, RL);
2816 AddToWorklist(ORNode.getNode());
2817 return DAG.getSetCC(SDLoc(LocReference), VT, ORNode, LR, Op1);
2821 // Simplify (and (setne X, 0), (setne X, -1)) -> (setuge (add X, 1), 2)
2822 if (LL == RL && isa<ConstantSDNode>(LR) && isa<ConstantSDNode>(RR) &&
2823 Op0 == Op1 && LL.getValueType().isInteger() &&
2824 Op0 == ISD::SETNE && ((isNullConstant(LR) && isAllOnesConstant(RR)) ||
2825 (isAllOnesConstant(LR) && isNullConstant(RR)))) {
2827 SDValue ADDNode = DAG.getNode(ISD::ADD, DL, LL.getValueType(),
2828 LL, DAG.getConstant(1, DL,
2829 LL.getValueType()));
2830 AddToWorklist(ADDNode.getNode());
2831 return DAG.getSetCC(SDLoc(LocReference), VT, ADDNode,
2832 DAG.getConstant(2, DL, LL.getValueType()),
2835 // canonicalize equivalent to ll == rl
2836 if (LL == RR && LR == RL) {
2837 Op1 = ISD::getSetCCSwappedOperands(Op1);
2840 if (LL == RL && LR == RR) {
2841 bool isInteger = LL.getValueType().isInteger();
2842 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2843 if (Result != ISD::SETCC_INVALID &&
2844 (!LegalOperations ||
2845 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
2846 TLI.isOperationLegal(ISD::SETCC,
2847 getSetCCResultType(N0.getSimpleValueType())))))
2848 return DAG.getSetCC(SDLoc(LocReference), N0.getValueType(),
2853 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
2854 VT.getSizeInBits() <= 64) {
2855 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2856 APInt ADDC = ADDI->getAPIntValue();
2857 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2858 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
2859 // immediate for an add, but it is legal if its top c2 bits are set,
2860 // transform the ADD so the immediate doesn't need to be materialized
2862 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
2863 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
2864 SRLI->getZExtValue());
2865 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
2867 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2870 DAG.getNode(ISD::ADD, DL, VT,
2871 N0.getOperand(0), DAG.getConstant(ADDC, DL, VT));
2872 CombineTo(N0.getNode(), NewAdd);
2873 // Return N so it doesn't get rechecked!
2874 return SDValue(LocReference, 0);
2885 SDValue DAGCombiner::visitAND(SDNode *N) {
2886 SDValue N0 = N->getOperand(0);
2887 SDValue N1 = N->getOperand(1);
2888 EVT VT = N1.getValueType();
2891 if (VT.isVector()) {
2892 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2895 // fold (and x, 0) -> 0, vector edition
2896 if (ISD::isBuildVectorAllZeros(N0.getNode()))
2897 // do not return N0, because undef node may exist in N0
2898 return DAG.getConstant(
2899 APInt::getNullValue(
2900 N0.getValueType().getScalarType().getSizeInBits()),
2901 SDLoc(N), N0.getValueType());
2902 if (ISD::isBuildVectorAllZeros(N1.getNode()))
2903 // do not return N1, because undef node may exist in N1
2904 return DAG.getConstant(
2905 APInt::getNullValue(
2906 N1.getValueType().getScalarType().getSizeInBits()),
2907 SDLoc(N), N1.getValueType());
2909 // fold (and x, -1) -> x, vector edition
2910 if (ISD::isBuildVectorAllOnes(N0.getNode()))
2912 if (ISD::isBuildVectorAllOnes(N1.getNode()))
2916 // fold (and c1, c2) -> c1&c2
2917 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
2918 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2919 if (N0C && N1C && !N1C->isOpaque())
2920 return DAG.FoldConstantArithmetic(ISD::AND, SDLoc(N), VT, N0C, N1C);
2921 // canonicalize constant to RHS
2922 if (isConstantIntBuildVectorOrConstantInt(N0) &&
2923 !isConstantIntBuildVectorOrConstantInt(N1))
2924 return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0);
2925 // fold (and x, -1) -> x
2926 if (isAllOnesConstant(N1))
2928 // if (and x, c) is known to be zero, return 0
2929 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2930 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2931 APInt::getAllOnesValue(BitWidth)))
2932 return DAG.getConstant(0, SDLoc(N), VT);
2934 if (SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1))
2936 // fold (and (or x, C), D) -> D if (C & D) == D
2937 if (N1C && N0.getOpcode() == ISD::OR)
2938 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2939 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2941 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2942 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2943 SDValue N0Op0 = N0.getOperand(0);
2944 APInt Mask = ~N1C->getAPIntValue();
2945 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2946 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2947 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
2948 N0.getValueType(), N0Op0);
2950 // Replace uses of the AND with uses of the Zero extend node.
2953 // We actually want to replace all uses of the any_extend with the
2954 // zero_extend, to avoid duplicating things. This will later cause this
2955 // AND to be folded.
2956 CombineTo(N0.getNode(), Zext);
2957 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2960 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2961 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2962 // already be zero by virtue of the width of the base type of the load.
2964 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2966 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2967 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2968 N0.getOpcode() == ISD::LOAD) {
2969 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2970 N0 : N0.getOperand(0) );
2972 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2973 // This can be a pure constant or a vector splat, in which case we treat the
2974 // vector as a scalar and use the splat value.
2975 APInt Constant = APInt::getNullValue(1);
2976 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2977 Constant = C->getAPIntValue();
2978 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2979 APInt SplatValue, SplatUndef;
2980 unsigned SplatBitSize;
2982 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2983 SplatBitSize, HasAnyUndefs);
2985 // Undef bits can contribute to a possible optimisation if set, so
2987 SplatValue |= SplatUndef;
2989 // The splat value may be something like "0x00FFFFFF", which means 0 for
2990 // the first vector value and FF for the rest, repeating. We need a mask
2991 // that will apply equally to all members of the vector, so AND all the
2992 // lanes of the constant together.
2993 EVT VT = Vector->getValueType(0);
2994 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2996 // If the splat value has been compressed to a bitlength lower
2997 // than the size of the vector lane, we need to re-expand it to
2999 if (BitWidth > SplatBitSize)
3000 for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
3001 SplatBitSize < BitWidth;
3002 SplatBitSize = SplatBitSize * 2)
3003 SplatValue |= SplatValue.shl(SplatBitSize);
3005 // Make sure that variable 'Constant' is only set if 'SplatBitSize' is a
3006 // multiple of 'BitWidth'. Otherwise, we could propagate a wrong value.
3007 if (SplatBitSize % BitWidth == 0) {
3008 Constant = APInt::getAllOnesValue(BitWidth);
3009 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
3010 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
3015 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
3016 // actually legal and isn't going to get expanded, else this is a false
3018 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
3019 Load->getValueType(0),
3020 Load->getMemoryVT());
3022 // Resize the constant to the same size as the original memory access before
3023 // extension. If it is still the AllOnesValue then this AND is completely
3026 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
3029 switch (Load->getExtensionType()) {
3030 default: B = false; break;
3031 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
3033 case ISD::NON_EXTLOAD: B = true; break;
3036 if (B && Constant.isAllOnesValue()) {
3037 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
3038 // preserve semantics once we get rid of the AND.
3039 SDValue NewLoad(Load, 0);
3040 if (Load->getExtensionType() == ISD::EXTLOAD) {
3041 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
3042 Load->getValueType(0), SDLoc(Load),
3043 Load->getChain(), Load->getBasePtr(),
3044 Load->getOffset(), Load->getMemoryVT(),
3045 Load->getMemOperand());
3046 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
3047 if (Load->getNumValues() == 3) {
3048 // PRE/POST_INC loads have 3 values.
3049 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
3050 NewLoad.getValue(2) };
3051 CombineTo(Load, To, 3, true);
3053 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
3057 // Fold the AND away, taking care not to fold to the old load node if we
3059 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
3061 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3065 // fold (and (load x), 255) -> (zextload x, i8)
3066 // fold (and (extload x, i16), 255) -> (zextload x, i8)
3067 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
3068 if (N1C && (N0.getOpcode() == ISD::LOAD ||
3069 (N0.getOpcode() == ISD::ANY_EXTEND &&
3070 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
3071 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
3072 LoadSDNode *LN0 = HasAnyExt
3073 ? cast<LoadSDNode>(N0.getOperand(0))
3074 : cast<LoadSDNode>(N0);
3075 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
3076 LN0->isUnindexed() && N0.hasOneUse() && SDValue(LN0, 0).hasOneUse()) {
3077 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
3078 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
3079 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
3080 EVT LoadedVT = LN0->getMemoryVT();
3081 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
3083 if (ExtVT == LoadedVT &&
3084 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy,
3088 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
3089 LN0->getChain(), LN0->getBasePtr(), ExtVT,
3090 LN0->getMemOperand());
3092 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
3093 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3096 // Do not change the width of a volatile load.
3097 // Do not generate loads of non-round integer types since these can
3098 // be expensive (and would be wrong if the type is not byte sized).
3099 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
3100 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy,
3102 EVT PtrType = LN0->getOperand(1).getValueType();
3104 unsigned Alignment = LN0->getAlignment();
3105 SDValue NewPtr = LN0->getBasePtr();
3107 // For big endian targets, we need to add an offset to the pointer
3108 // to load the correct bytes. For little endian systems, we merely
3109 // need to read fewer bytes from the same pointer.
3110 if (DAG.getDataLayout().isBigEndian()) {
3111 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
3112 unsigned EVTStoreBytes = ExtVT.getStoreSize();
3113 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
3115 NewPtr = DAG.getNode(ISD::ADD, DL, PtrType,
3116 NewPtr, DAG.getConstant(PtrOff, DL, PtrType));
3117 Alignment = MinAlign(Alignment, PtrOff);
3120 AddToWorklist(NewPtr.getNode());
3123 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
3124 LN0->getChain(), NewPtr,
3125 LN0->getPointerInfo(),
3126 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
3127 LN0->isInvariant(), Alignment, LN0->getAAInfo());
3129 CombineTo(LN0, Load, Load.getValue(1));
3130 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3136 if (SDValue Combined = visitANDLike(N0, N1, N))
3139 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
3140 if (N0.getOpcode() == N1.getOpcode())
3141 if (SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N))
3144 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
3145 // fold (and (sra)) -> (and (srl)) when possible.
3146 if (!VT.isVector() &&
3147 SimplifyDemandedBits(SDValue(N, 0)))
3148 return SDValue(N, 0);
3150 // fold (zext_inreg (extload x)) -> (zextload x)
3151 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
3152 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3153 EVT MemVT = LN0->getMemoryVT();
3154 // If we zero all the possible extended bits, then we can turn this into
3155 // a zextload if we are running before legalize or the operation is legal.
3156 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
3157 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
3158 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
3159 ((!LegalOperations && !LN0->isVolatile()) ||
3160 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) {
3161 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
3162 LN0->getChain(), LN0->getBasePtr(),
3163 MemVT, LN0->getMemOperand());
3165 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
3166 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3169 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
3170 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
3172 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3173 EVT MemVT = LN0->getMemoryVT();
3174 // If we zero all the possible extended bits, then we can turn this into
3175 // a zextload if we are running before legalize or the operation is legal.
3176 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
3177 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
3178 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
3179 ((!LegalOperations && !LN0->isVolatile()) ||
3180 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) {
3181 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
3182 LN0->getChain(), LN0->getBasePtr(),
3183 MemVT, LN0->getMemOperand());
3185 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
3186 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3189 // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const)
3190 if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) {
3191 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
3192 N0.getOperand(1), false);
3193 if (BSwap.getNode())
3200 /// Match (a >> 8) | (a << 8) as (bswap a) >> 16.
3201 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
3202 bool DemandHighBits) {
3203 if (!LegalOperations)
3206 EVT VT = N->getValueType(0);
3207 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
3209 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3212 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
3213 bool LookPassAnd0 = false;
3214 bool LookPassAnd1 = false;
3215 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
3217 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
3219 if (N0.getOpcode() == ISD::AND) {
3220 if (!N0.getNode()->hasOneUse())
3222 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3223 if (!N01C || N01C->getZExtValue() != 0xFF00)
3225 N0 = N0.getOperand(0);
3226 LookPassAnd0 = true;
3229 if (N1.getOpcode() == ISD::AND) {
3230 if (!N1.getNode()->hasOneUse())
3232 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3233 if (!N11C || N11C->getZExtValue() != 0xFF)
3235 N1 = N1.getOperand(0);
3236 LookPassAnd1 = true;
3239 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
3241 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
3243 if (!N0.getNode()->hasOneUse() ||
3244 !N1.getNode()->hasOneUse())
3247 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3248 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3251 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
3254 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
3255 SDValue N00 = N0->getOperand(0);
3256 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
3257 if (!N00.getNode()->hasOneUse())
3259 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
3260 if (!N001C || N001C->getZExtValue() != 0xFF)
3262 N00 = N00.getOperand(0);
3263 LookPassAnd0 = true;
3266 SDValue N10 = N1->getOperand(0);
3267 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
3268 if (!N10.getNode()->hasOneUse())
3270 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
3271 if (!N101C || N101C->getZExtValue() != 0xFF00)
3273 N10 = N10.getOperand(0);
3274 LookPassAnd1 = true;
3280 // Make sure everything beyond the low halfword gets set to zero since the SRL
3281 // 16 will clear the top bits.
3282 unsigned OpSizeInBits = VT.getSizeInBits();
3283 if (DemandHighBits && OpSizeInBits > 16) {
3284 // If the left-shift isn't masked out then the only way this is a bswap is
3285 // if all bits beyond the low 8 are 0. In that case the entire pattern
3286 // reduces to a left shift anyway: leave it for other parts of the combiner.
3290 // However, if the right shift isn't masked out then it might be because
3291 // it's not needed. See if we can spot that too.
3292 if (!LookPassAnd1 &&
3293 !DAG.MaskedValueIsZero(
3294 N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16)))
3298 SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00);
3299 if (OpSizeInBits > 16) {
3301 Res = DAG.getNode(ISD::SRL, DL, VT, Res,
3302 DAG.getConstant(OpSizeInBits - 16, DL,
3303 getShiftAmountTy(VT)));
3308 /// Return true if the specified node is an element that makes up a 32-bit
3309 /// packed halfword byteswap.
3310 /// ((x & 0x000000ff) << 8) |
3311 /// ((x & 0x0000ff00) >> 8) |
3312 /// ((x & 0x00ff0000) << 8) |
3313 /// ((x & 0xff000000) >> 8)
3314 static bool isBSwapHWordElement(SDValue N, MutableArrayRef<SDNode *> Parts) {
3315 if (!N.getNode()->hasOneUse())
3318 unsigned Opc = N.getOpcode();
3319 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
3322 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3327 switch (N1C->getZExtValue()) {
3330 case 0xFF: Num = 0; break;
3331 case 0xFF00: Num = 1; break;
3332 case 0xFF0000: Num = 2; break;
3333 case 0xFF000000: Num = 3; break;
3336 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
3337 SDValue N0 = N.getOperand(0);
3338 if (Opc == ISD::AND) {
3339 if (Num == 0 || Num == 2) {
3341 // (x >> 8) & 0xff0000
3342 if (N0.getOpcode() != ISD::SRL)
3344 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3345 if (!C || C->getZExtValue() != 8)
3348 // (x << 8) & 0xff00
3349 // (x << 8) & 0xff000000
3350 if (N0.getOpcode() != ISD::SHL)
3352 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3353 if (!C || C->getZExtValue() != 8)
3356 } else if (Opc == ISD::SHL) {
3358 // (x & 0xff0000) << 8
3359 if (Num != 0 && Num != 2)
3361 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3362 if (!C || C->getZExtValue() != 8)
3364 } else { // Opc == ISD::SRL
3365 // (x & 0xff00) >> 8
3366 // (x & 0xff000000) >> 8
3367 if (Num != 1 && Num != 3)
3369 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3370 if (!C || C->getZExtValue() != 8)
3377 Parts[Num] = N0.getOperand(0).getNode();
3381 /// Match a 32-bit packed halfword bswap. That is
3382 /// ((x & 0x000000ff) << 8) |
3383 /// ((x & 0x0000ff00) >> 8) |
3384 /// ((x & 0x00ff0000) << 8) |
3385 /// ((x & 0xff000000) >> 8)
3386 /// => (rotl (bswap x), 16)
3387 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
3388 if (!LegalOperations)
3391 EVT VT = N->getValueType(0);
3394 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3398 // (or (or (and), (and)), (or (and), (and)))
3399 // (or (or (or (and), (and)), (and)), (and))
3400 if (N0.getOpcode() != ISD::OR)
3402 SDValue N00 = N0.getOperand(0);
3403 SDValue N01 = N0.getOperand(1);
3404 SDNode *Parts[4] = {};
3406 if (N1.getOpcode() == ISD::OR &&
3407 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) {
3408 // (or (or (and), (and)), (or (and), (and)))
3409 SDValue N000 = N00.getOperand(0);
3410 if (!isBSwapHWordElement(N000, Parts))
3413 SDValue N001 = N00.getOperand(1);
3414 if (!isBSwapHWordElement(N001, Parts))
3416 SDValue N010 = N01.getOperand(0);
3417 if (!isBSwapHWordElement(N010, Parts))
3419 SDValue N011 = N01.getOperand(1);
3420 if (!isBSwapHWordElement(N011, Parts))
3423 // (or (or (or (and), (and)), (and)), (and))
3424 if (!isBSwapHWordElement(N1, Parts))
3426 if (!isBSwapHWordElement(N01, Parts))
3428 if (N00.getOpcode() != ISD::OR)
3430 SDValue N000 = N00.getOperand(0);
3431 if (!isBSwapHWordElement(N000, Parts))
3433 SDValue N001 = N00.getOperand(1);
3434 if (!isBSwapHWordElement(N001, Parts))
3438 // Make sure the parts are all coming from the same node.
3439 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
3443 SDValue BSwap = DAG.getNode(ISD::BSWAP, DL, VT,
3444 SDValue(Parts[0], 0));
3446 // Result of the bswap should be rotated by 16. If it's not legal, then
3447 // do (x << 16) | (x >> 16).
3448 SDValue ShAmt = DAG.getConstant(16, DL, getShiftAmountTy(VT));
3449 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3450 return DAG.getNode(ISD::ROTL, DL, VT, BSwap, ShAmt);
3451 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3452 return DAG.getNode(ISD::ROTR, DL, VT, BSwap, ShAmt);
3453 return DAG.getNode(ISD::OR, DL, VT,
3454 DAG.getNode(ISD::SHL, DL, VT, BSwap, ShAmt),
3455 DAG.getNode(ISD::SRL, DL, VT, BSwap, ShAmt));
3458 /// This contains all DAGCombine rules which reduce two values combined by
3459 /// an Or operation to a single value \see visitANDLike().
3460 SDValue DAGCombiner::visitORLike(SDValue N0, SDValue N1, SDNode *LocReference) {
3461 EVT VT = N1.getValueType();
3462 // fold (or x, undef) -> -1
3463 if (!LegalOperations &&
3464 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3465 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3466 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()),
3467 SDLoc(LocReference), VT);
3469 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3470 SDValue LL, LR, RL, RR, CC0, CC1;
3471 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3472 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3473 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3475 if (LR == RR && Op0 == Op1 && LL.getValueType().isInteger()) {
3476 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3477 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3478 if (isNullConstant(LR) && (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3479 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(LR),
3480 LR.getValueType(), LL, RL);
3481 AddToWorklist(ORNode.getNode());
3482 return DAG.getSetCC(SDLoc(LocReference), VT, ORNode, LR, Op1);
3484 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3485 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3486 if (isAllOnesConstant(LR) && (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3487 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(LR),
3488 LR.getValueType(), LL, RL);
3489 AddToWorklist(ANDNode.getNode());
3490 return DAG.getSetCC(SDLoc(LocReference), VT, ANDNode, LR, Op1);
3493 // canonicalize equivalent to ll == rl
3494 if (LL == RR && LR == RL) {
3495 Op1 = ISD::getSetCCSwappedOperands(Op1);
3498 if (LL == RL && LR == RR) {
3499 bool isInteger = LL.getValueType().isInteger();
3500 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3501 if (Result != ISD::SETCC_INVALID &&
3502 (!LegalOperations ||
3503 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
3504 TLI.isOperationLegal(ISD::SETCC,
3505 getSetCCResultType(N0.getValueType())))))
3506 return DAG.getSetCC(SDLoc(LocReference), N0.getValueType(),
3511 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3512 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == ISD::AND &&
3513 // Don't increase # computations.
3514 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3515 // We can only do this xform if we know that bits from X that are set in C2
3516 // but not in C1 are already zero. Likewise for Y.
3517 if (const ConstantSDNode *N0O1C =
3518 getAsNonOpaqueConstant(N0.getOperand(1))) {
3519 if (const ConstantSDNode *N1O1C =
3520 getAsNonOpaqueConstant(N1.getOperand(1))) {
3521 // We can only do this xform if we know that bits from X that are set in
3522 // C2 but not in C1 are already zero. Likewise for Y.
3523 const APInt &LHSMask = N0O1C->getAPIntValue();
3524 const APInt &RHSMask = N1O1C->getAPIntValue();
3526 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3527 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3528 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3529 N0.getOperand(0), N1.getOperand(0));
3530 SDLoc DL(LocReference);
3531 return DAG.getNode(ISD::AND, DL, VT, X,
3532 DAG.getConstant(LHSMask | RHSMask, DL, VT));
3538 // (or (and X, M), (and X, N)) -> (and X, (or M, N))
3539 if (N0.getOpcode() == ISD::AND &&
3540 N1.getOpcode() == ISD::AND &&
3541 N0.getOperand(0) == N1.getOperand(0) &&
3542 // Don't increase # computations.
3543 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3544 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3545 N0.getOperand(1), N1.getOperand(1));
3546 return DAG.getNode(ISD::AND, SDLoc(LocReference), VT, N0.getOperand(0), X);
3552 SDValue DAGCombiner::visitOR(SDNode *N) {
3553 SDValue N0 = N->getOperand(0);
3554 SDValue N1 = N->getOperand(1);
3555 EVT VT = N1.getValueType();
3558 if (VT.isVector()) {
3559 if (SDValue FoldedVOp = SimplifyVBinOp(N))
3562 // fold (or x, 0) -> x, vector edition
3563 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3565 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3568 // fold (or x, -1) -> -1, vector edition
3569 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3570 // do not return N0, because undef node may exist in N0
3571 return DAG.getConstant(
3572 APInt::getAllOnesValue(
3573 N0.getValueType().getScalarType().getSizeInBits()),
3574 SDLoc(N), N0.getValueType());
3575 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3576 // do not return N1, because undef node may exist in N1
3577 return DAG.getConstant(
3578 APInt::getAllOnesValue(
3579 N1.getValueType().getScalarType().getSizeInBits()),
3580 SDLoc(N), N1.getValueType());
3582 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask1)
3583 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf B, A, Mask2)
3584 // Do this only if the resulting shuffle is legal.
3585 if (isa<ShuffleVectorSDNode>(N0) &&
3586 isa<ShuffleVectorSDNode>(N1) &&
3587 // Avoid folding a node with illegal type.
3588 TLI.isTypeLegal(VT) &&
3589 N0->getOperand(1) == N1->getOperand(1) &&
3590 ISD::isBuildVectorAllZeros(N0.getOperand(1).getNode())) {
3591 bool CanFold = true;
3592 unsigned NumElts = VT.getVectorNumElements();
3593 const ShuffleVectorSDNode *SV0 = cast<ShuffleVectorSDNode>(N0);
3594 const ShuffleVectorSDNode *SV1 = cast<ShuffleVectorSDNode>(N1);
3595 // We construct two shuffle masks:
3596 // - Mask1 is a shuffle mask for a shuffle with N0 as the first operand
3597 // and N1 as the second operand.
3598 // - Mask2 is a shuffle mask for a shuffle with N1 as the first operand
3599 // and N0 as the second operand.
3600 // We do this because OR is commutable and therefore there might be
3601 // two ways to fold this node into a shuffle.
3602 SmallVector<int,4> Mask1;
3603 SmallVector<int,4> Mask2;
3605 for (unsigned i = 0; i != NumElts && CanFold; ++i) {
3606 int M0 = SV0->getMaskElt(i);
3607 int M1 = SV1->getMaskElt(i);
3609 // Both shuffle indexes are undef. Propagate Undef.
3610 if (M0 < 0 && M1 < 0) {
3611 Mask1.push_back(M0);
3612 Mask2.push_back(M0);
3616 if (M0 < 0 || M1 < 0 ||
3617 (M0 < (int)NumElts && M1 < (int)NumElts) ||
3618 (M0 >= (int)NumElts && M1 >= (int)NumElts)) {
3623 Mask1.push_back(M0 < (int)NumElts ? M0 : M1 + NumElts);
3624 Mask2.push_back(M1 < (int)NumElts ? M1 : M0 + NumElts);
3628 // Fold this sequence only if the resulting shuffle is 'legal'.
3629 if (TLI.isShuffleMaskLegal(Mask1, VT))
3630 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0),
3631 N1->getOperand(0), &Mask1[0]);
3632 if (TLI.isShuffleMaskLegal(Mask2, VT))
3633 return DAG.getVectorShuffle(VT, SDLoc(N), N1->getOperand(0),
3634 N0->getOperand(0), &Mask2[0]);
3639 // fold (or c1, c2) -> c1|c2
3640 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
3641 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3642 if (N0C && N1C && !N1C->isOpaque())
3643 return DAG.FoldConstantArithmetic(ISD::OR, SDLoc(N), VT, N0C, N1C);
3644 // canonicalize constant to RHS
3645 if (isConstantIntBuildVectorOrConstantInt(N0) &&
3646 !isConstantIntBuildVectorOrConstantInt(N1))
3647 return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0);
3648 // fold (or x, 0) -> x
3649 if (isNullConstant(N1))
3651 // fold (or x, -1) -> -1
3652 if (isAllOnesConstant(N1))
3654 // fold (or x, c) -> c iff (x & ~c) == 0
3655 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3658 if (SDValue Combined = visitORLike(N0, N1, N))
3661 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3662 if (SDValue BSwap = MatchBSwapHWord(N, N0, N1))
3664 if (SDValue BSwap = MatchBSwapHWordLow(N, N0, N1))
3668 if (SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1))
3670 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3671 // iff (c1 & c2) == 0.
3672 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3673 isa<ConstantSDNode>(N0.getOperand(1))) {
3674 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3675 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) {
3676 if (SDValue COR = DAG.FoldConstantArithmetic(ISD::OR, SDLoc(N1), VT,
3679 ISD::AND, SDLoc(N), VT,
3680 DAG.getNode(ISD::OR, SDLoc(N0), VT, N0.getOperand(0), N1), COR);
3684 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3685 if (N0.getOpcode() == N1.getOpcode())
3686 if (SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N))
3689 // See if this is some rotate idiom.
3690 if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N)))
3691 return SDValue(Rot, 0);
3693 // Simplify the operands using demanded-bits information.
3694 if (!VT.isVector() &&
3695 SimplifyDemandedBits(SDValue(N, 0)))
3696 return SDValue(N, 0);
3701 /// Match "(X shl/srl V1) & V2" where V2 may not be present.
3702 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3703 if (Op.getOpcode() == ISD::AND) {
3704 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3705 Mask = Op.getOperand(1);
3706 Op = Op.getOperand(0);
3712 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3720 // Return true if we can prove that, whenever Neg and Pos are both in the
3721 // range [0, OpSize), Neg == (Pos == 0 ? 0 : OpSize - Pos). This means that
3722 // for two opposing shifts shift1 and shift2 and a value X with OpBits bits:
3724 // (or (shift1 X, Neg), (shift2 X, Pos))
3726 // reduces to a rotate in direction shift2 by Pos or (equivalently) a rotate
3727 // in direction shift1 by Neg. The range [0, OpSize) means that we only need
3728 // to consider shift amounts with defined behavior.
3729 static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned OpSize) {
3730 // If OpSize is a power of 2 then:
3732 // (a) (Pos == 0 ? 0 : OpSize - Pos) == (OpSize - Pos) & (OpSize - 1)
3733 // (b) Neg == Neg & (OpSize - 1) whenever Neg is in [0, OpSize).
3735 // So if OpSize is a power of 2 and Neg is (and Neg', OpSize-1), we check
3736 // for the stronger condition:
3738 // Neg & (OpSize - 1) == (OpSize - Pos) & (OpSize - 1) [A]
3740 // for all Neg and Pos. Since Neg & (OpSize - 1) == Neg' & (OpSize - 1)
3741 // we can just replace Neg with Neg' for the rest of the function.
3743 // In other cases we check for the even stronger condition:
3745 // Neg == OpSize - Pos [B]
3747 // for all Neg and Pos. Note that the (or ...) then invokes undefined
3748 // behavior if Pos == 0 (and consequently Neg == OpSize).
3750 // We could actually use [A] whenever OpSize is a power of 2, but the
3751 // only extra cases that it would match are those uninteresting ones
3752 // where Neg and Pos are never in range at the same time. E.g. for
3753 // OpSize == 32, using [A] would allow a Neg of the form (sub 64, Pos)
3754 // as well as (sub 32, Pos), but:
3756 // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos))
3758 // always invokes undefined behavior for 32-bit X.
3760 // Below, Mask == OpSize - 1 when using [A] and is all-ones otherwise.
3761 unsigned MaskLoBits = 0;
3762 if (Neg.getOpcode() == ISD::AND &&
3763 isPowerOf2_64(OpSize) &&
3764 Neg.getOperand(1).getOpcode() == ISD::Constant &&
3765 cast<ConstantSDNode>(Neg.getOperand(1))->getAPIntValue() == OpSize - 1) {
3766 Neg = Neg.getOperand(0);
3767 MaskLoBits = Log2_64(OpSize);
3770 // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1.
3771 if (Neg.getOpcode() != ISD::SUB)
3773 ConstantSDNode *NegC = dyn_cast<ConstantSDNode>(Neg.getOperand(0));
3776 SDValue NegOp1 = Neg.getOperand(1);
3778 // On the RHS of [A], if Pos is Pos' & (OpSize - 1), just replace Pos with
3779 // Pos'. The truncation is redundant for the purpose of the equality.
3781 Pos.getOpcode() == ISD::AND &&
3782 Pos.getOperand(1).getOpcode() == ISD::Constant &&
3783 cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() == OpSize - 1)
3784 Pos = Pos.getOperand(0);
3786 // The condition we need is now:
3788 // (NegC - NegOp1) & Mask == (OpSize - Pos) & Mask
3790 // If NegOp1 == Pos then we need:
3792 // OpSize & Mask == NegC & Mask
3794 // (because "x & Mask" is a truncation and distributes through subtraction).
3797 Width = NegC->getAPIntValue();
3798 // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC.
3799 // Then the condition we want to prove becomes:
3801 // (NegC - NegOp1) & Mask == (OpSize - (NegOp1 + PosC)) & Mask
3803 // which, again because "x & Mask" is a truncation, becomes:
3805 // NegC & Mask == (OpSize - PosC) & Mask
3806 // OpSize & Mask == (NegC + PosC) & Mask
3807 else if (Pos.getOpcode() == ISD::ADD &&
3808 Pos.getOperand(0) == NegOp1 &&
3809 Pos.getOperand(1).getOpcode() == ISD::Constant)
3810 Width = (cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() +
3811 NegC->getAPIntValue());
3815 // Now we just need to check that OpSize & Mask == Width & Mask.
3817 // Opsize & Mask is 0 since Mask is Opsize - 1.
3818 return Width.getLoBits(MaskLoBits) == 0;
3819 return Width == OpSize;
3822 // A subroutine of MatchRotate used once we have found an OR of two opposite
3823 // shifts of Shifted. If Neg == <operand size> - Pos then the OR reduces
3824 // to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the
3825 // former being preferred if supported. InnerPos and InnerNeg are Pos and
3826 // Neg with outer conversions stripped away.
3827 SDNode *DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos,
3828 SDValue Neg, SDValue InnerPos,
3829 SDValue InnerNeg, unsigned PosOpcode,
3830 unsigned NegOpcode, SDLoc DL) {
3831 // fold (or (shl x, (*ext y)),
3832 // (srl x, (*ext (sub 32, y)))) ->
3833 // (rotl x, y) or (rotr x, (sub 32, y))
3835 // fold (or (shl x, (*ext (sub 32, y))),
3836 // (srl x, (*ext y))) ->
3837 // (rotr x, y) or (rotl x, (sub 32, y))
3838 EVT VT = Shifted.getValueType();
3839 if (matchRotateSub(InnerPos, InnerNeg, VT.getSizeInBits())) {
3840 bool HasPos = TLI.isOperationLegalOrCustom(PosOpcode, VT);
3841 return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted,
3842 HasPos ? Pos : Neg).getNode();
3848 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3849 // idioms for rotate, and if the target supports rotation instructions, generate
3851 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL) {
3852 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3853 EVT VT = LHS.getValueType();
3854 if (!TLI.isTypeLegal(VT)) return nullptr;
3856 // The target must have at least one rotate flavor.
3857 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3858 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3859 if (!HasROTL && !HasROTR) return nullptr;
3861 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3862 SDValue LHSShift; // The shift.
3863 SDValue LHSMask; // AND value if any.
3864 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3865 return nullptr; // Not part of a rotate.
3867 SDValue RHSShift; // The shift.
3868 SDValue RHSMask; // AND value if any.
3869 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3870 return nullptr; // Not part of a rotate.
3872 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3873 return nullptr; // Not shifting the same value.
3875 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3876 return nullptr; // Shifts must disagree.
3878 // Canonicalize shl to left side in a shl/srl pair.
3879 if (RHSShift.getOpcode() == ISD::SHL) {
3880 std::swap(LHS, RHS);
3881 std::swap(LHSShift, RHSShift);
3882 std::swap(LHSMask , RHSMask );
3885 unsigned OpSizeInBits = VT.getSizeInBits();
3886 SDValue LHSShiftArg = LHSShift.getOperand(0);
3887 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3888 SDValue RHSShiftArg = RHSShift.getOperand(0);
3889 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3891 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3892 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3893 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3894 RHSShiftAmt.getOpcode() == ISD::Constant) {
3895 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3896 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3897 if ((LShVal + RShVal) != OpSizeInBits)
3900 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3901 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
3903 // If there is an AND of either shifted operand, apply it to the result.
3904 if (LHSMask.getNode() || RHSMask.getNode()) {
3905 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3907 if (LHSMask.getNode()) {
3908 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3909 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3911 if (RHSMask.getNode()) {
3912 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3913 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3916 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, DL, VT));
3919 return Rot.getNode();
3922 // If there is a mask here, and we have a variable shift, we can't be sure
3923 // that we're masking out the right stuff.
3924 if (LHSMask.getNode() || RHSMask.getNode())
3927 // If the shift amount is sign/zext/any-extended just peel it off.
3928 SDValue LExtOp0 = LHSShiftAmt;
3929 SDValue RExtOp0 = RHSShiftAmt;
3930 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3931 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3932 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3933 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3934 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3935 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3936 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3937 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3938 LExtOp0 = LHSShiftAmt.getOperand(0);
3939 RExtOp0 = RHSShiftAmt.getOperand(0);
3942 SDNode *TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt,
3943 LExtOp0, RExtOp0, ISD::ROTL, ISD::ROTR, DL);
3947 SDNode *TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt,
3948 RExtOp0, LExtOp0, ISD::ROTR, ISD::ROTL, DL);
3955 SDValue DAGCombiner::visitXOR(SDNode *N) {
3956 SDValue N0 = N->getOperand(0);
3957 SDValue N1 = N->getOperand(1);
3958 EVT VT = N0.getValueType();
3961 if (VT.isVector()) {
3962 if (SDValue FoldedVOp = SimplifyVBinOp(N))
3965 // fold (xor x, 0) -> x, vector edition
3966 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3968 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3972 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3973 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3974 return DAG.getConstant(0, SDLoc(N), VT);
3975 // fold (xor x, undef) -> undef
3976 if (N0.getOpcode() == ISD::UNDEF)
3978 if (N1.getOpcode() == ISD::UNDEF)
3980 // fold (xor c1, c2) -> c1^c2
3981 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
3982 ConstantSDNode *N1C = getAsNonOpaqueConstant(N1);
3984 return DAG.FoldConstantArithmetic(ISD::XOR, SDLoc(N), VT, N0C, N1C);
3985 // canonicalize constant to RHS
3986 if (isConstantIntBuildVectorOrConstantInt(N0) &&
3987 !isConstantIntBuildVectorOrConstantInt(N1))
3988 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
3989 // fold (xor x, 0) -> x
3990 if (isNullConstant(N1))
3993 if (SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1))
3996 // fold !(x cc y) -> (x !cc y)
3997 SDValue LHS, RHS, CC;
3998 if (TLI.isConstTrueVal(N1.getNode()) && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3999 bool isInt = LHS.getValueType().isInteger();
4000 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
4003 if (!LegalOperations ||
4004 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) {
4005 switch (N0.getOpcode()) {
4007 llvm_unreachable("Unhandled SetCC Equivalent!");
4009 return DAG.getSetCC(SDLoc(N), VT, LHS, RHS, NotCC);
4010 case ISD::SELECT_CC:
4011 return DAG.getSelectCC(SDLoc(N), LHS, RHS, N0.getOperand(2),
4012 N0.getOperand(3), NotCC);
4017 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
4018 if (isOneConstant(N1) && N0.getOpcode() == ISD::ZERO_EXTEND &&
4019 N0.getNode()->hasOneUse() &&
4020 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
4021 SDValue V = N0.getOperand(0);
4023 V = DAG.getNode(ISD::XOR, DL, V.getValueType(), V,
4024 DAG.getConstant(1, DL, V.getValueType()));
4025 AddToWorklist(V.getNode());
4026 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V);
4029 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
4030 if (isOneConstant(N1) && VT == MVT::i1 &&
4031 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
4032 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
4033 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
4034 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
4035 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
4036 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
4037 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
4038 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
4041 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
4042 if (isAllOnesConstant(N1) &&
4043 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
4044 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
4045 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
4046 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
4047 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
4048 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
4049 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
4050 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
4053 // fold (xor (and x, y), y) -> (and (not x), y)
4054 if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
4055 N0->getOperand(1) == N1) {
4056 SDValue X = N0->getOperand(0);
4057 SDValue NotX = DAG.getNOT(SDLoc(X), X, VT);
4058 AddToWorklist(NotX.getNode());
4059 return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1);
4061 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
4062 if (N1C && N0.getOpcode() == ISD::XOR) {
4063 if (const ConstantSDNode *N00C = getAsNonOpaqueConstant(N0.getOperand(0))) {
4065 return DAG.getNode(ISD::XOR, DL, VT, N0.getOperand(1),
4066 DAG.getConstant(N1C->getAPIntValue() ^
4067 N00C->getAPIntValue(), DL, VT));
4069 if (const ConstantSDNode *N01C = getAsNonOpaqueConstant(N0.getOperand(1))) {
4071 return DAG.getNode(ISD::XOR, DL, VT, N0.getOperand(0),
4072 DAG.getConstant(N1C->getAPIntValue() ^
4073 N01C->getAPIntValue(), DL, VT));
4076 // fold (xor x, x) -> 0
4078 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
4080 // fold (xor (shl 1, x), -1) -> (rotl ~1, x)
4081 // Here is a concrete example of this equivalence:
4083 // i16 shl == 1 << 14 == 16384 == 0b0100000000000000
4084 // i16 xor == ~(1 << 14) == 49151 == 0b1011111111111111
4088 // i16 ~1 == 0b1111111111111110
4089 // i16 rol(~1, 14) == 0b1011111111111111
4091 // Some additional tips to help conceptualize this transform:
4092 // - Try to see the operation as placing a single zero in a value of all ones.
4093 // - There exists no value for x which would allow the result to contain zero.
4094 // - Values of x larger than the bitwidth are undefined and do not require a
4095 // consistent result.
4096 // - Pushing the zero left requires shifting one bits in from the right.
4097 // A rotate left of ~1 is a nice way of achieving the desired result.
4098 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT) && N0.getOpcode() == ISD::SHL
4099 && isAllOnesConstant(N1) && isOneConstant(N0.getOperand(0))) {
4101 return DAG.getNode(ISD::ROTL, DL, VT, DAG.getConstant(~1, DL, VT),
4105 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
4106 if (N0.getOpcode() == N1.getOpcode())
4107 if (SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N))
4110 // Simplify the expression using non-local knowledge.
4111 if (!VT.isVector() &&
4112 SimplifyDemandedBits(SDValue(N, 0)))
4113 return SDValue(N, 0);
4118 /// Handle transforms common to the three shifts, when the shift amount is a
4120 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, ConstantSDNode *Amt) {
4121 SDNode *LHS = N->getOperand(0).getNode();
4122 if (!LHS->hasOneUse()) return SDValue();
4124 // We want to pull some binops through shifts, so that we have (and (shift))
4125 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
4126 // thing happens with address calculations, so it's important to canonicalize
4128 bool HighBitSet = false; // Can we transform this if the high bit is set?
4130 switch (LHS->getOpcode()) {
4131 default: return SDValue();
4134 HighBitSet = false; // We can only transform sra if the high bit is clear.
4137 HighBitSet = true; // We can only transform sra if the high bit is set.
4140 if (N->getOpcode() != ISD::SHL)
4141 return SDValue(); // only shl(add) not sr[al](add).
4142 HighBitSet = false; // We can only transform sra if the high bit is clear.
4146 // We require the RHS of the binop to be a constant and not opaque as well.
4147 ConstantSDNode *BinOpCst = getAsNonOpaqueConstant(LHS->getOperand(1));
4148 if (!BinOpCst) return SDValue();
4150 // FIXME: disable this unless the input to the binop is a shift by a constant.
4151 // If it is not a shift, it pessimizes some common cases like:
4153 // void foo(int *X, int i) { X[i & 1235] = 1; }
4154 // int bar(int *X, int i) { return X[i & 255]; }
4155 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
4156 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
4157 BinOpLHSVal->getOpcode() != ISD::SRA &&
4158 BinOpLHSVal->getOpcode() != ISD::SRL) ||
4159 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
4162 EVT VT = N->getValueType(0);
4164 // If this is a signed shift right, and the high bit is modified by the
4165 // logical operation, do not perform the transformation. The highBitSet
4166 // boolean indicates the value of the high bit of the constant which would
4167 // cause it to be modified for this operation.
4168 if (N->getOpcode() == ISD::SRA) {
4169 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
4170 if (BinOpRHSSignSet != HighBitSet)
4174 if (!TLI.isDesirableToCommuteWithShift(LHS))
4177 // Fold the constants, shifting the binop RHS by the shift amount.
4178 SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)),
4180 LHS->getOperand(1), N->getOperand(1));
4181 assert(isa<ConstantSDNode>(NewRHS) && "Folding was not successful!");
4183 // Create the new shift.
4184 SDValue NewShift = DAG.getNode(N->getOpcode(),
4185 SDLoc(LHS->getOperand(0)),
4186 VT, LHS->getOperand(0), N->getOperand(1));
4188 // Create the new binop.
4189 return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS);
4192 SDValue DAGCombiner::distributeTruncateThroughAnd(SDNode *N) {
4193 assert(N->getOpcode() == ISD::TRUNCATE);
4194 assert(N->getOperand(0).getOpcode() == ISD::AND);
4196 // (truncate:TruncVT (and N00, N01C)) -> (and (truncate:TruncVT N00), TruncC)
4197 if (N->hasOneUse() && N->getOperand(0).hasOneUse()) {
4198 SDValue N01 = N->getOperand(0).getOperand(1);
4200 if (ConstantSDNode *N01C = isConstOrConstSplat(N01)) {
4201 if (!N01C->isOpaque()) {
4202 EVT TruncVT = N->getValueType(0);
4203 SDValue N00 = N->getOperand(0).getOperand(0);
4204 APInt TruncC = N01C->getAPIntValue();
4205 TruncC = TruncC.trunc(TruncVT.getScalarSizeInBits());
4208 return DAG.getNode(ISD::AND, DL, TruncVT,
4209 DAG.getNode(ISD::TRUNCATE, DL, TruncVT, N00),
4210 DAG.getConstant(TruncC, DL, TruncVT));
4218 SDValue DAGCombiner::visitRotate(SDNode *N) {
4219 // fold (rot* x, (trunc (and y, c))) -> (rot* x, (and (trunc y), (trunc c))).
4220 if (N->getOperand(1).getOpcode() == ISD::TRUNCATE &&
4221 N->getOperand(1).getOperand(0).getOpcode() == ISD::AND) {
4222 SDValue NewOp1 = distributeTruncateThroughAnd(N->getOperand(1).getNode());
4223 if (NewOp1.getNode())
4224 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
4225 N->getOperand(0), NewOp1);
4230 SDValue DAGCombiner::visitSHL(SDNode *N) {
4231 SDValue N0 = N->getOperand(0);
4232 SDValue N1 = N->getOperand(1);
4233 EVT VT = N0.getValueType();
4234 unsigned OpSizeInBits = VT.getScalarSizeInBits();
4237 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4238 if (VT.isVector()) {
4239 if (SDValue FoldedVOp = SimplifyVBinOp(N))
4242 BuildVectorSDNode *N1CV = dyn_cast<BuildVectorSDNode>(N1);
4243 // If setcc produces all-one true value then:
4244 // (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<<N1CV)
4245 if (N1CV && N1CV->isConstant()) {
4246 if (N0.getOpcode() == ISD::AND) {
4247 SDValue N00 = N0->getOperand(0);
4248 SDValue N01 = N0->getOperand(1);
4249 BuildVectorSDNode *N01CV = dyn_cast<BuildVectorSDNode>(N01);
4251 if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC &&
4252 TLI.getBooleanContents(N00.getOperand(0).getValueType()) ==
4253 TargetLowering::ZeroOrNegativeOneBooleanContent) {
4254 if (SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, SDLoc(N), VT,
4256 return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C);
4259 N1C = isConstOrConstSplat(N1);
4264 // fold (shl c1, c2) -> c1<<c2
4265 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
4266 if (N0C && N1C && !N1C->isOpaque())
4267 return DAG.FoldConstantArithmetic(ISD::SHL, SDLoc(N), VT, N0C, N1C);
4268 // fold (shl 0, x) -> 0
4269 if (isNullConstant(N0))
4271 // fold (shl x, c >= size(x)) -> undef
4272 if (N1C && N1C->getAPIntValue().uge(OpSizeInBits))
4273 return DAG.getUNDEF(VT);
4274 // fold (shl x, 0) -> x
4275 if (N1C && N1C->isNullValue())
4277 // fold (shl undef, x) -> 0
4278 if (N0.getOpcode() == ISD::UNDEF)
4279 return DAG.getConstant(0, SDLoc(N), VT);
4280 // if (shl x, c) is known to be zero, return 0
4281 if (DAG.MaskedValueIsZero(SDValue(N, 0),
4282 APInt::getAllOnesValue(OpSizeInBits)))
4283 return DAG.getConstant(0, SDLoc(N), VT);
4284 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
4285 if (N1.getOpcode() == ISD::TRUNCATE &&
4286 N1.getOperand(0).getOpcode() == ISD::AND) {
4287 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4288 if (NewOp1.getNode())
4289 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, NewOp1);
4292 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4293 return SDValue(N, 0);
4295 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
4296 if (N1C && N0.getOpcode() == ISD::SHL) {
4297 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4298 uint64_t c1 = N0C1->getZExtValue();
4299 uint64_t c2 = N1C->getZExtValue();
4301 if (c1 + c2 >= OpSizeInBits)
4302 return DAG.getConstant(0, DL, VT);
4303 return DAG.getNode(ISD::SHL, DL, VT, N0.getOperand(0),
4304 DAG.getConstant(c1 + c2, DL, N1.getValueType()));
4308 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
4309 // For this to be valid, the second form must not preserve any of the bits
4310 // that are shifted out by the inner shift in the first form. This means
4311 // the outer shift size must be >= the number of bits added by the ext.
4312 // As a corollary, we don't care what kind of ext it is.
4313 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
4314 N0.getOpcode() == ISD::ANY_EXTEND ||
4315 N0.getOpcode() == ISD::SIGN_EXTEND) &&
4316 N0.getOperand(0).getOpcode() == ISD::SHL) {
4317 SDValue N0Op0 = N0.getOperand(0);
4318 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4319 uint64_t c1 = N0Op0C1->getZExtValue();
4320 uint64_t c2 = N1C->getZExtValue();
4321 EVT InnerShiftVT = N0Op0.getValueType();
4322 uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits();
4323 if (c2 >= OpSizeInBits - InnerShiftSize) {
4325 if (c1 + c2 >= OpSizeInBits)
4326 return DAG.getConstant(0, DL, VT);
4327 return DAG.getNode(ISD::SHL, DL, VT,
4328 DAG.getNode(N0.getOpcode(), DL, VT,
4329 N0Op0->getOperand(0)),
4330 DAG.getConstant(c1 + c2, DL, N1.getValueType()));
4335 // fold (shl (zext (srl x, C)), C) -> (zext (shl (srl x, C), C))
4336 // Only fold this if the inner zext has no other uses to avoid increasing
4337 // the total number of instructions.
4338 if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() &&
4339 N0.getOperand(0).getOpcode() == ISD::SRL) {
4340 SDValue N0Op0 = N0.getOperand(0);
4341 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4342 uint64_t c1 = N0Op0C1->getZExtValue();
4343 if (c1 < VT.getScalarSizeInBits()) {
4344 uint64_t c2 = N1C->getZExtValue();
4346 SDValue NewOp0 = N0.getOperand(0);
4347 EVT CountVT = NewOp0.getOperand(1).getValueType();
4349 SDValue NewSHL = DAG.getNode(ISD::SHL, DL, NewOp0.getValueType(),
4351 DAG.getConstant(c2, DL, CountVT));
4352 AddToWorklist(NewSHL.getNode());
4353 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
4359 // fold (shl (sr[la] exact X, C1), C2) -> (shl X, (C2-C1)) if C1 <= C2
4360 // fold (shl (sr[la] exact X, C1), C2) -> (sr[la] X, (C2-C1)) if C1 > C2
4361 if (N1C && (N0.getOpcode() == ISD::SRL || N0.getOpcode() == ISD::SRA) &&
4362 cast<BinaryWithFlagsSDNode>(N0)->Flags.hasExact()) {
4363 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4364 uint64_t C1 = N0C1->getZExtValue();
4365 uint64_t C2 = N1C->getZExtValue();
4368 return DAG.getNode(ISD::SHL, DL, VT, N0.getOperand(0),
4369 DAG.getConstant(C2 - C1, DL, N1.getValueType()));
4370 return DAG.getNode(N0.getOpcode(), DL, VT, N0.getOperand(0),
4371 DAG.getConstant(C1 - C2, DL, N1.getValueType()));
4375 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
4376 // (and (srl x, (sub c1, c2), MASK)
4377 // Only fold this if the inner shift has no other uses -- if it does, folding
4378 // this will increase the total number of instructions.
4379 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
4380 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4381 uint64_t c1 = N0C1->getZExtValue();
4382 if (c1 < OpSizeInBits) {
4383 uint64_t c2 = N1C->getZExtValue();
4384 APInt Mask = APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - c1);
4387 Mask = Mask.shl(c2 - c1);
4389 Shift = DAG.getNode(ISD::SHL, DL, VT, N0.getOperand(0),
4390 DAG.getConstant(c2 - c1, DL, N1.getValueType()));
4392 Mask = Mask.lshr(c1 - c2);
4394 Shift = DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4395 DAG.getConstant(c1 - c2, DL, N1.getValueType()));
4398 return DAG.getNode(ISD::AND, DL, VT, Shift,
4399 DAG.getConstant(Mask, DL, VT));
4403 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
4404 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
4405 unsigned BitSize = VT.getScalarSizeInBits();
4407 SDValue HiBitsMask =
4408 DAG.getConstant(APInt::getHighBitsSet(BitSize,
4409 BitSize - N1C->getZExtValue()),
4411 return DAG.getNode(ISD::AND, DL, VT, N0.getOperand(0),
4415 // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
4416 // Variant of version done on multiply, except mul by a power of 2 is turned
4419 if (N1C && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
4420 (isa<ConstantSDNode>(N0.getOperand(1)) ||
4421 isConstantSplatVector(N0.getOperand(1).getNode(), Val))) {
4422 SDValue Shl0 = DAG.getNode(ISD::SHL, SDLoc(N0), VT, N0.getOperand(0), N1);
4423 SDValue Shl1 = DAG.getNode(ISD::SHL, SDLoc(N1), VT, N0.getOperand(1), N1);
4424 return DAG.getNode(ISD::ADD, SDLoc(N), VT, Shl0, Shl1);
4427 if (N1C && !N1C->isOpaque())
4428 if (SDValue NewSHL = visitShiftByConstant(N, N1C))
4434 SDValue DAGCombiner::visitSRA(SDNode *N) {
4435 SDValue N0 = N->getOperand(0);
4436 SDValue N1 = N->getOperand(1);
4437 EVT VT = N0.getValueType();
4438 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4441 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4442 if (VT.isVector()) {
4443 if (SDValue FoldedVOp = SimplifyVBinOp(N))
4446 N1C = isConstOrConstSplat(N1);
4449 // fold (sra c1, c2) -> (sra c1, c2)
4450 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
4451 if (N0C && N1C && !N1C->isOpaque())
4452 return DAG.FoldConstantArithmetic(ISD::SRA, SDLoc(N), VT, N0C, N1C);
4453 // fold (sra 0, x) -> 0
4454 if (isNullConstant(N0))
4456 // fold (sra -1, x) -> -1
4457 if (isAllOnesConstant(N0))
4459 // fold (sra x, (setge c, size(x))) -> undef
4460 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4461 return DAG.getUNDEF(VT);
4462 // fold (sra x, 0) -> x
4463 if (N1C && N1C->isNullValue())
4465 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
4467 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
4468 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
4469 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
4471 ExtVT = EVT::getVectorVT(*DAG.getContext(),
4472 ExtVT, VT.getVectorNumElements());
4473 if ((!LegalOperations ||
4474 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
4475 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
4476 N0.getOperand(0), DAG.getValueType(ExtVT));
4479 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
4480 if (N1C && N0.getOpcode() == ISD::SRA) {
4481 if (ConstantSDNode *C1 = isConstOrConstSplat(N0.getOperand(1))) {
4482 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
4483 if (Sum >= OpSizeInBits)
4484 Sum = OpSizeInBits - 1;
4486 return DAG.getNode(ISD::SRA, DL, VT, N0.getOperand(0),
4487 DAG.getConstant(Sum, DL, N1.getValueType()));
4491 // fold (sra (shl X, m), (sub result_size, n))
4492 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
4493 // result_size - n != m.
4494 // If truncate is free for the target sext(shl) is likely to result in better
4496 if (N0.getOpcode() == ISD::SHL && N1C) {
4497 // Get the two constanst of the shifts, CN0 = m, CN = n.
4498 const ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1));
4500 LLVMContext &Ctx = *DAG.getContext();
4501 // Determine what the truncate's result bitsize and type would be.
4502 EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - N1C->getZExtValue());
4505 TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorNumElements());
4507 // Determine the residual right-shift amount.
4508 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
4510 // If the shift is not a no-op (in which case this should be just a sign
4511 // extend already), the truncated to type is legal, sign_extend is legal
4512 // on that type, and the truncate to that type is both legal and free,
4513 // perform the transform.
4514 if ((ShiftAmt > 0) &&
4515 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
4516 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
4517 TLI.isTruncateFree(VT, TruncVT)) {
4520 SDValue Amt = DAG.getConstant(ShiftAmt, DL,
4521 getShiftAmountTy(N0.getOperand(0).getValueType()));
4522 SDValue Shift = DAG.getNode(ISD::SRL, DL, VT,
4523 N0.getOperand(0), Amt);
4524 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, TruncVT,
4526 return DAG.getNode(ISD::SIGN_EXTEND, DL,
4527 N->getValueType(0), Trunc);
4532 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
4533 if (N1.getOpcode() == ISD::TRUNCATE &&
4534 N1.getOperand(0).getOpcode() == ISD::AND) {
4535 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4536 if (NewOp1.getNode())
4537 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, NewOp1);
4540 // fold (sra (trunc (srl x, c1)), c2) -> (trunc (sra x, c1 + c2))
4541 // if c1 is equal to the number of bits the trunc removes
4542 if (N0.getOpcode() == ISD::TRUNCATE &&
4543 (N0.getOperand(0).getOpcode() == ISD::SRL ||
4544 N0.getOperand(0).getOpcode() == ISD::SRA) &&
4545 N0.getOperand(0).hasOneUse() &&
4546 N0.getOperand(0).getOperand(1).hasOneUse() &&
4548 SDValue N0Op0 = N0.getOperand(0);
4549 if (ConstantSDNode *LargeShift = isConstOrConstSplat(N0Op0.getOperand(1))) {
4550 unsigned LargeShiftVal = LargeShift->getZExtValue();
4551 EVT LargeVT = N0Op0.getValueType();
4553 if (LargeVT.getScalarSizeInBits() - OpSizeInBits == LargeShiftVal) {
4556 DAG.getConstant(LargeShiftVal + N1C->getZExtValue(), DL,
4557 getShiftAmountTy(N0Op0.getOperand(0).getValueType()));
4558 SDValue SRA = DAG.getNode(ISD::SRA, DL, LargeVT,
4559 N0Op0.getOperand(0), Amt);
4560 return DAG.getNode(ISD::TRUNCATE, DL, VT, SRA);
4565 // Simplify, based on bits shifted out of the LHS.
4566 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4567 return SDValue(N, 0);
4570 // If the sign bit is known to be zero, switch this to a SRL.
4571 if (DAG.SignBitIsZero(N0))
4572 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1);
4574 if (N1C && !N1C->isOpaque())
4575 if (SDValue NewSRA = visitShiftByConstant(N, N1C))
4581 SDValue DAGCombiner::visitSRL(SDNode *N) {
4582 SDValue N0 = N->getOperand(0);
4583 SDValue N1 = N->getOperand(1);
4584 EVT VT = N0.getValueType();
4585 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4588 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4589 if (VT.isVector()) {
4590 if (SDValue FoldedVOp = SimplifyVBinOp(N))
4593 N1C = isConstOrConstSplat(N1);
4596 // fold (srl c1, c2) -> c1 >>u c2
4597 ConstantSDNode *N0C = getAsNonOpaqueConstant(N0);
4598 if (N0C && N1C && !N1C->isOpaque())
4599 return DAG.FoldConstantArithmetic(ISD::SRL, SDLoc(N), VT, N0C, N1C);
4600 // fold (srl 0, x) -> 0
4601 if (isNullConstant(N0))
4603 // fold (srl x, c >= size(x)) -> undef
4604 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4605 return DAG.getUNDEF(VT);
4606 // fold (srl x, 0) -> x
4607 if (N1C && N1C->isNullValue())
4609 // if (srl x, c) is known to be zero, return 0
4610 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
4611 APInt::getAllOnesValue(OpSizeInBits)))
4612 return DAG.getConstant(0, SDLoc(N), VT);
4614 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
4615 if (N1C && N0.getOpcode() == ISD::SRL) {
4616 if (ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1))) {
4617 uint64_t c1 = N01C->getZExtValue();
4618 uint64_t c2 = N1C->getZExtValue();
4620 if (c1 + c2 >= OpSizeInBits)
4621 return DAG.getConstant(0, DL, VT);
4622 return DAG.getNode(ISD::SRL, DL, VT, N0.getOperand(0),
4623 DAG.getConstant(c1 + c2, DL, N1.getValueType()));
4627 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
4628 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
4629 N0.getOperand(0).getOpcode() == ISD::SRL &&
4630 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
4632 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
4633 uint64_t c2 = N1C->getZExtValue();
4634 EVT InnerShiftVT = N0.getOperand(0).getValueType();
4635 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
4636 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
4637 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
4638 if (c1 + OpSizeInBits == InnerShiftSize) {
4640 if (c1 + c2 >= InnerShiftSize)
4641 return DAG.getConstant(0, DL, VT);
4642 return DAG.getNode(ISD::TRUNCATE, DL, VT,
4643 DAG.getNode(ISD::SRL, DL, InnerShiftVT,
4644 N0.getOperand(0)->getOperand(0),
4645 DAG.getConstant(c1 + c2, DL,
4650 // fold (srl (shl x, c), c) -> (and x, cst2)
4651 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1) {
4652 unsigned BitSize = N0.getScalarValueSizeInBits();
4653 if (BitSize <= 64) {
4654 uint64_t ShAmt = N1C->getZExtValue() + 64 - BitSize;
4656 return DAG.getNode(ISD::AND, DL, VT, N0.getOperand(0),
4657 DAG.getConstant(~0ULL >> ShAmt, DL, VT));
4661 // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask)
4662 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
4663 // Shifting in all undef bits?
4664 EVT SmallVT = N0.getOperand(0).getValueType();
4665 unsigned BitSize = SmallVT.getScalarSizeInBits();
4666 if (N1C->getZExtValue() >= BitSize)
4667 return DAG.getUNDEF(VT);
4669 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
4670 uint64_t ShiftAmt = N1C->getZExtValue();
4672 SDValue SmallShift = DAG.getNode(ISD::SRL, DL0, SmallVT,
4674 DAG.getConstant(ShiftAmt, DL0,
4675 getShiftAmountTy(SmallVT)));
4676 AddToWorklist(SmallShift.getNode());
4677 APInt Mask = APInt::getAllOnesValue(OpSizeInBits).lshr(ShiftAmt);
4679 return DAG.getNode(ISD::AND, DL, VT,
4680 DAG.getNode(ISD::ANY_EXTEND, DL, VT, SmallShift),
4681 DAG.getConstant(Mask, DL, VT));
4685 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
4686 // bit, which is unmodified by sra.
4687 if (N1C && N1C->getZExtValue() + 1 == OpSizeInBits) {
4688 if (N0.getOpcode() == ISD::SRA)
4689 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1);
4692 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
4693 if (N1C && N0.getOpcode() == ISD::CTLZ &&
4694 N1C->getAPIntValue() == Log2_32(OpSizeInBits)) {
4695 APInt KnownZero, KnownOne;
4696 DAG.computeKnownBits(N0.getOperand(0), KnownZero, KnownOne);
4698 // If any of the input bits are KnownOne, then the input couldn't be all
4699 // zeros, thus the result of the srl will always be zero.
4700 if (KnownOne.getBoolValue()) return DAG.getConstant(0, SDLoc(N0), VT);
4702 // If all of the bits input the to ctlz node are known to be zero, then
4703 // the result of the ctlz is "32" and the result of the shift is one.
4704 APInt UnknownBits = ~KnownZero;
4705 if (UnknownBits == 0) return DAG.getConstant(1, SDLoc(N0), VT);
4707 // Otherwise, check to see if there is exactly one bit input to the ctlz.
4708 if ((UnknownBits & (UnknownBits - 1)) == 0) {
4709 // Okay, we know that only that the single bit specified by UnknownBits
4710 // could be set on input to the CTLZ node. If this bit is set, the SRL
4711 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
4712 // to an SRL/XOR pair, which is likely to simplify more.
4713 unsigned ShAmt = UnknownBits.countTrailingZeros();
4714 SDValue Op = N0.getOperand(0);
4718 Op = DAG.getNode(ISD::SRL, DL, VT, Op,
4719 DAG.getConstant(ShAmt, DL,
4720 getShiftAmountTy(Op.getValueType())));
4721 AddToWorklist(Op.getNode());
4725 return DAG.getNode(ISD::XOR, DL, VT,
4726 Op, DAG.getConstant(1, DL, VT));
4730 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
4731 if (N1.getOpcode() == ISD::TRUNCATE &&
4732 N1.getOperand(0).getOpcode() == ISD::AND) {
4733 if (SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode()))
4734 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, NewOp1);
4737 // fold operands of srl based on knowledge that the low bits are not
4739 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4740 return SDValue(N, 0);
4742 if (N1C && !N1C->isOpaque())
4743 if (SDValue NewSRL = visitShiftByConstant(N, N1C))
4746 // Attempt to convert a srl of a load into a narrower zero-extending load.
4747 if (SDValue NarrowLoad = ReduceLoadWidth(N))
4750 // Here is a common situation. We want to optimize:
4753 // %b = and i32 %a, 2
4754 // %c = srl i32 %b, 1
4755 // brcond i32 %c ...
4761 // %c = setcc eq %b, 0
4764 // However when after the source operand of SRL is optimized into AND, the SRL
4765 // itself may not be optimized further. Look for it and add the BRCOND into
4767 if (N->hasOneUse()) {
4768 SDNode *Use = *N->use_begin();
4769 if (Use->getOpcode() == ISD::BRCOND)
4771 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
4772 // Also look pass the truncate.
4773 Use = *Use->use_begin();
4774 if (Use->getOpcode() == ISD::BRCOND)
4782 SDValue DAGCombiner::visitBSWAP(SDNode *N) {
4783 SDValue N0 = N->getOperand(0);
4784 EVT VT = N->getValueType(0);
4786 // fold (bswap c1) -> c2
4787 if (isConstantIntBuildVectorOrConstantInt(N0))
4788 return DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N0);
4789 // fold (bswap (bswap x)) -> x
4790 if (N0.getOpcode() == ISD::BSWAP)
4791 return N0->getOperand(0);
4795 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
4796 SDValue N0 = N->getOperand(0);
4797 EVT VT = N->getValueType(0);
4799 // fold (ctlz c1) -> c2
4800 if (isConstantIntBuildVectorOrConstantInt(N0))
4801 return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0);
4805 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4806 SDValue N0 = N->getOperand(0);
4807 EVT VT = N->getValueType(0);
4809 // fold (ctlz_zero_undef c1) -> c2
4810 if (isConstantIntBuildVectorOrConstantInt(N0))
4811 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4815 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4816 SDValue N0 = N->getOperand(0);
4817 EVT VT = N->getValueType(0);
4819 // fold (cttz c1) -> c2
4820 if (isConstantIntBuildVectorOrConstantInt(N0))
4821 return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0);
4825 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4826 SDValue N0 = N->getOperand(0);
4827 EVT VT = N->getValueType(0);
4829 // fold (cttz_zero_undef c1) -> c2
4830 if (isConstantIntBuildVectorOrConstantInt(N0))
4831 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4835 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4836 SDValue N0 = N->getOperand(0);
4837 EVT VT = N->getValueType(0);
4839 // fold (ctpop c1) -> c2
4840 if (isConstantIntBuildVectorOrConstantInt(N0))
4841 return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0);
4846 /// \brief Generate Min/Max node
4847 static SDValue combineMinNumMaxNum(SDLoc DL, EVT VT, SDValue LHS, SDValue RHS,
4848 SDValue True, SDValue False,
4849 ISD::CondCode CC, const TargetLowering &TLI,
4850 SelectionDAG &DAG) {
4851 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
4861 unsigned Opcode = (LHS == True) ? ISD::FMINNUM : ISD::FMAXNUM;
4862 if (TLI.isOperationLegal(Opcode, VT))
4863 return DAG.getNode(Opcode, DL, VT, LHS, RHS);
4872 unsigned Opcode = (LHS == True) ? ISD::FMAXNUM : ISD::FMINNUM;
4873 if (TLI.isOperationLegal(Opcode, VT))
4874 return DAG.getNode(Opcode, DL, VT, LHS, RHS);
4882 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4883 SDValue N0 = N->getOperand(0);
4884 SDValue N1 = N->getOperand(1);
4885 SDValue N2 = N->getOperand(2);
4886 EVT VT = N->getValueType(0);
4887 EVT VT0 = N0.getValueType();
4889 // fold (select C, X, X) -> X
4892 if (const ConstantSDNode *N0C = dyn_cast<const ConstantSDNode>(N0)) {
4893 // fold (select true, X, Y) -> X
4894 // fold (select false, X, Y) -> Y
4895 return !N0C->isNullValue() ? N1 : N2;
4897 // fold (select C, 1, X) -> (or C, X)
4898 if (VT == MVT::i1 && isOneConstant(N1))
4899 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4900 // fold (select C, 0, 1) -> (xor C, 1)
4901 // We can't do this reliably if integer based booleans have different contents
4902 // to floating point based booleans. This is because we can't tell whether we
4903 // have an integer-based boolean or a floating-point-based boolean unless we
4904 // can find the SETCC that produced it and inspect its operands. This is
4905 // fairly easy if C is the SETCC node, but it can potentially be
4906 // undiscoverable (or not reasonably discoverable). For example, it could be
4907 // in another basic block or it could require searching a complicated
4909 if (VT.isInteger() &&
4910 (VT0 == MVT::i1 || (VT0.isInteger() &&
4911 TLI.getBooleanContents(false, false) ==
4912 TLI.getBooleanContents(false, true) &&
4913 TLI.getBooleanContents(false, false) ==
4914 TargetLowering::ZeroOrOneBooleanContent)) &&
4915 isNullConstant(N1) && isOneConstant(N2)) {
4919 return DAG.getNode(ISD::XOR, DL, VT0,
4920 N0, DAG.getConstant(1, DL, VT0));
4923 XORNode = DAG.getNode(ISD::XOR, DL0, VT0,
4924 N0, DAG.getConstant(1, DL0, VT0));
4925 AddToWorklist(XORNode.getNode());
4927 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, XORNode);
4928 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, XORNode);
4930 // fold (select C, 0, X) -> (and (not C), X)
4931 if (VT == VT0 && VT == MVT::i1 && isNullConstant(N1)) {
4932 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4933 AddToWorklist(NOTNode.getNode());
4934 return DAG.getNode(ISD::AND, SDLoc(N), VT, NOTNode, N2);
4936 // fold (select C, X, 1) -> (or (not C), X)
4937 if (VT == VT0 && VT == MVT::i1 && isOneConstant(N2)) {
4938 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4939 AddToWorklist(NOTNode.getNode());
4940 return DAG.getNode(ISD::OR, SDLoc(N), VT, NOTNode, N1);
4942 // fold (select C, X, 0) -> (and C, X)
4943 if (VT == MVT::i1 && isNullConstant(N2))
4944 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4945 // fold (select X, X, Y) -> (or X, Y)
4946 // fold (select X, 1, Y) -> (or X, Y)
4947 if (VT == MVT::i1 && (N0 == N1 || isOneConstant(N1)))
4948 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4949 // fold (select X, Y, X) -> (and X, Y)
4950 // fold (select X, Y, 0) -> (and X, Y)
4951 if (VT == MVT::i1 && (N0 == N2 || isNullConstant(N2)))
4952 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4954 // If we can fold this based on the true/false value, do so.
4955 if (SimplifySelectOps(N, N1, N2))
4956 return SDValue(N, 0); // Don't revisit N.
4958 if (VT0 == MVT::i1) {
4959 // The code in this block deals with the following 2 equivalences:
4960 // select(C0|C1, x, y) <=> select(C0, x, select(C1, x, y))
4961 // select(C0&C1, x, y) <=> select(C0, select(C1, x, y), y)
4962 // The target can specify its prefered form with the
4963 // shouldNormalizeToSelectSequence() callback. However we always transform
4964 // to the right anyway if we find the inner select exists in the DAG anyway
4965 // and we always transform to the left side if we know that we can further
4966 // optimize the combination of the conditions.
4967 bool normalizeToSequence
4968 = TLI.shouldNormalizeToSelectSequence(*DAG.getContext(), VT);
4969 // select (and Cond0, Cond1), X, Y
4970 // -> select Cond0, (select Cond1, X, Y), Y
4971 if (N0->getOpcode() == ISD::AND && N0->hasOneUse()) {
4972 SDValue Cond0 = N0->getOperand(0);
4973 SDValue Cond1 = N0->getOperand(1);
4974 SDValue InnerSelect = DAG.getNode(ISD::SELECT, SDLoc(N),
4975 N1.getValueType(), Cond1, N1, N2);
4976 if (normalizeToSequence || !InnerSelect.use_empty())
4977 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Cond0,
4980 // select (or Cond0, Cond1), X, Y -> select Cond0, X, (select Cond1, X, Y)
4981 if (N0->getOpcode() == ISD::OR && N0->hasOneUse()) {
4982 SDValue Cond0 = N0->getOperand(0);
4983 SDValue Cond1 = N0->getOperand(1);
4984 SDValue InnerSelect = DAG.getNode(ISD::SELECT, SDLoc(N),
4985 N1.getValueType(), Cond1, N1, N2);
4986 if (normalizeToSequence || !InnerSelect.use_empty())
4987 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Cond0, N1,
4991 // select Cond0, (select Cond1, X, Y), Y -> select (and Cond0, Cond1), X, Y
4992 if (N1->getOpcode() == ISD::SELECT && N1->hasOneUse()) {
4993 SDValue N1_0 = N1->getOperand(0);
4994 SDValue N1_1 = N1->getOperand(1);
4995 SDValue N1_2 = N1->getOperand(2);
4996 if (N1_2 == N2 && N0.getValueType() == N1_0.getValueType()) {
4997 // Create the actual and node if we can generate good code for it.
4998 if (!normalizeToSequence) {
4999 SDValue And = DAG.getNode(ISD::AND, SDLoc(N), N0.getValueType(),
5001 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), And,
5004 // Otherwise see if we can optimize the "and" to a better pattern.
5005 if (SDValue Combined = visitANDLike(N0, N1_0, N))
5006 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Combined,
5010 // select Cond0, X, (select Cond1, X, Y) -> select (or Cond0, Cond1), X, Y
5011 if (N2->getOpcode() == ISD::SELECT && N2->hasOneUse()) {
5012 SDValue N2_0 = N2->getOperand(0);
5013 SDValue N2_1 = N2->getOperand(1);
5014 SDValue N2_2 = N2->getOperand(2);
5015 if (N2_1 == N1 && N0.getValueType() == N2_0.getValueType()) {
5016 // Create the actual or node if we can generate good code for it.
5017 if (!normalizeToSequence) {
5018 SDValue Or = DAG.getNode(ISD::OR, SDLoc(N), N0.getValueType(),
5020 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Or,
5023 // Otherwise see if we can optimize to a better pattern.
5024 if (SDValue Combined = visitORLike(N0, N2_0, N))
5025 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Combined,
5031 // fold selects based on a setcc into other things, such as min/max/abs
5032 if (N0.getOpcode() == ISD::SETCC) {
5033 // select x, y (fcmp lt x, y) -> fminnum x, y
5034 // select x, y (fcmp gt x, y) -> fmaxnum x, y
5036 // This is OK if we don't care about what happens if either operand is a
5040 // FIXME: Instead of testing for UnsafeFPMath, this should be checking for
5041 // no signed zeros as well as no nans.
5042 const TargetOptions &Options = DAG.getTarget().Options;
5043 if (Options.UnsafeFPMath &&
5044 VT.isFloatingPoint() && N0.hasOneUse() &&
5045 DAG.isKnownNeverNaN(N1) && DAG.isKnownNeverNaN(N2)) {
5046 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5048 if (SDValue FMinMax = combineMinNumMaxNum(SDLoc(N), VT, N0.getOperand(0),
5049 N0.getOperand(1), N1, N2, CC,
5054 if ((!LegalOperations &&
5055 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT)) ||
5056 TLI.isOperationLegal(ISD::SELECT_CC, VT))
5057 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT,
5058 N0.getOperand(0), N0.getOperand(1),
5059 N1, N2, N0.getOperand(2));
5060 return SimplifySelect(SDLoc(N), N0, N1, N2);
5067 std::pair<SDValue, SDValue> SplitVSETCC(const SDNode *N, SelectionDAG &DAG) {
5070 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
5072 // Split the inputs.
5073 SDValue Lo, Hi, LL, LH, RL, RH;
5074 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
5075 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
5077 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));
5078 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));
5080 return std::make_pair(Lo, Hi);
5083 // This function assumes all the vselect's arguments are CONCAT_VECTOR
5084 // nodes and that the condition is a BV of ConstantSDNodes (or undefs).
5085 static SDValue ConvertSelectToConcatVector(SDNode *N, SelectionDAG &DAG) {
5087 SDValue Cond = N->getOperand(0);
5088 SDValue LHS = N->getOperand(1);
5089 SDValue RHS = N->getOperand(2);
5090 EVT VT = N->getValueType(0);
5091 int NumElems = VT.getVectorNumElements();
5092 assert(LHS.getOpcode() == ISD::CONCAT_VECTORS &&
5093 RHS.getOpcode() == ISD::CONCAT_VECTORS &&
5094 Cond.getOpcode() == ISD::BUILD_VECTOR);
5096 // CONCAT_VECTOR can take an arbitrary number of arguments. We only care about
5097 // binary ones here.
5098 if (LHS->getNumOperands() != 2 || RHS->getNumOperands() != 2)
5101 // We're sure we have an even number of elements due to the
5102 // concat_vectors we have as arguments to vselect.
5103 // Skip BV elements until we find one that's not an UNDEF
5104 // After we find an UNDEF element, keep looping until we get to half the
5105 // length of the BV and see if all the non-undef nodes are the same.
5106 ConstantSDNode *BottomHalf = nullptr;
5107 for (int i = 0; i < NumElems / 2; ++i) {
5108 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
5111 if (BottomHalf == nullptr)
5112 BottomHalf = cast<ConstantSDNode>(Cond.getOperand(i));
5113 else if (Cond->getOperand(i).getNode() != BottomHalf)
5117 // Do the same for the second half of the BuildVector
5118 ConstantSDNode *TopHalf = nullptr;
5119 for (int i = NumElems / 2; i < NumElems; ++i) {
5120 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
5123 if (TopHalf == nullptr)
5124 TopHalf = cast<ConstantSDNode>(Cond.getOperand(i));
5125 else if (Cond->getOperand(i).getNode() != TopHalf)
5129 assert(TopHalf && BottomHalf &&
5130 "One half of the selector was all UNDEFs and the other was all the "
5131 "same value. This should have been addressed before this function.");
5133 ISD::CONCAT_VECTORS, dl, VT,
5134 BottomHalf->isNullValue() ? RHS->getOperand(0) : LHS->getOperand(0),
5135 TopHalf->isNullValue() ? RHS->getOperand(1) : LHS->getOperand(1));
5138 SDValue DAGCombiner::visitMSCATTER(SDNode *N) {
5140 if (Level >= AfterLegalizeTypes)
5143 MaskedScatterSDNode *MSC = cast<MaskedScatterSDNode>(N);
5144 SDValue Mask = MSC->getMask();
5145 SDValue Data = MSC->getValue();
5148 // If the MSCATTER data type requires splitting and the mask is provided by a
5149 // SETCC, then split both nodes and its operands before legalization. This
5150 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5151 // and enables future optimizations (e.g. min/max pattern matching on X86).
5152 if (Mask.getOpcode() != ISD::SETCC)
5155 // Check if any splitting is required.
5156 if (TLI.getTypeAction(*DAG.getContext(), Data.getValueType()) !=
5157 TargetLowering::TypeSplitVector)
5159 SDValue MaskLo, MaskHi, Lo, Hi;
5160 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
5163 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MSC->getValueType(0));
5165 SDValue Chain = MSC->getChain();
5167 EVT MemoryVT = MSC->getMemoryVT();
5168 unsigned Alignment = MSC->getOriginalAlignment();
5170 EVT LoMemVT, HiMemVT;
5171 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
5173 SDValue DataLo, DataHi;
5174 std::tie(DataLo, DataHi) = DAG.SplitVector(Data, DL);
5176 SDValue BasePtr = MSC->getBasePtr();
5177 SDValue IndexLo, IndexHi;
5178 std::tie(IndexLo, IndexHi) = DAG.SplitVector(MSC->getIndex(), DL);
5180 MachineMemOperand *MMO = DAG.getMachineFunction().
5181 getMachineMemOperand(MSC->getPointerInfo(),
5182 MachineMemOperand::MOStore, LoMemVT.getStoreSize(),
5183 Alignment, MSC->getAAInfo(), MSC->getRanges());
5185 SDValue OpsLo[] = { Chain, DataLo, MaskLo, BasePtr, IndexLo };
5186 Lo = DAG.getMaskedScatter(DAG.getVTList(MVT::Other), DataLo.getValueType(),
5189 SDValue OpsHi[] = {Chain, DataHi, MaskHi, BasePtr, IndexHi};
5190 Hi = DAG.getMaskedScatter(DAG.getVTList(MVT::Other), DataHi.getValueType(),
5193 AddToWorklist(Lo.getNode());
5194 AddToWorklist(Hi.getNode());
5196 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);
5199 SDValue DAGCombiner::visitMSTORE(SDNode *N) {
5201 if (Level >= AfterLegalizeTypes)
5204 MaskedStoreSDNode *MST = dyn_cast<MaskedStoreSDNode>(N);
5205 SDValue Mask = MST->getMask();
5206 SDValue Data = MST->getValue();
5209 // If the MSTORE data type requires splitting and the mask is provided by a
5210 // SETCC, then split both nodes and its operands before legalization. This
5211 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5212 // and enables future optimizations (e.g. min/max pattern matching on X86).
5213 if (Mask.getOpcode() == ISD::SETCC) {
5215 // Check if any splitting is required.
5216 if (TLI.getTypeAction(*DAG.getContext(), Data.getValueType()) !=
5217 TargetLowering::TypeSplitVector)
5220 SDValue MaskLo, MaskHi, Lo, Hi;
5221 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
5224 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MST->getValueType(0));
5226 SDValue Chain = MST->getChain();
5227 SDValue Ptr = MST->getBasePtr();
5229 EVT MemoryVT = MST->getMemoryVT();
5230 unsigned Alignment = MST->getOriginalAlignment();
5232 // if Alignment is equal to the vector size,
5233 // take the half of it for the second part
5234 unsigned SecondHalfAlignment =
5235 (Alignment == Data->getValueType(0).getSizeInBits()/8) ?
5236 Alignment/2 : Alignment;
5238 EVT LoMemVT, HiMemVT;
5239 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
5241 SDValue DataLo, DataHi;
5242 std::tie(DataLo, DataHi) = DAG.SplitVector(Data, DL);
5244 MachineMemOperand *MMO = DAG.getMachineFunction().
5245 getMachineMemOperand(MST->getPointerInfo(),
5246 MachineMemOperand::MOStore, LoMemVT.getStoreSize(),
5247 Alignment, MST->getAAInfo(), MST->getRanges());
5249 Lo = DAG.getMaskedStore(Chain, DL, DataLo, Ptr, MaskLo, LoMemVT, MMO,
5250 MST->isTruncatingStore());
5252 unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
5253 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5254 DAG.getConstant(IncrementSize, DL, Ptr.getValueType()));
5256 MMO = DAG.getMachineFunction().
5257 getMachineMemOperand(MST->getPointerInfo(),
5258 MachineMemOperand::MOStore, HiMemVT.getStoreSize(),
5259 SecondHalfAlignment, MST->getAAInfo(),
5262 Hi = DAG.getMaskedStore(Chain, DL, DataHi, Ptr, MaskHi, HiMemVT, MMO,
5263 MST->isTruncatingStore());
5265 AddToWorklist(Lo.getNode());
5266 AddToWorklist(Hi.getNode());
5268 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);
5273 SDValue DAGCombiner::visitMGATHER(SDNode *N) {
5275 if (Level >= AfterLegalizeTypes)
5278 MaskedGatherSDNode *MGT = dyn_cast<MaskedGatherSDNode>(N);
5279 SDValue Mask = MGT->getMask();
5282 // If the MGATHER result requires splitting and the mask is provided by a
5283 // SETCC, then split both nodes and its operands before legalization. This
5284 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5285 // and enables future optimizations (e.g. min/max pattern matching on X86).
5287 if (Mask.getOpcode() != ISD::SETCC)
5290 EVT VT = N->getValueType(0);
5292 // Check if any splitting is required.
5293 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
5294 TargetLowering::TypeSplitVector)
5297 SDValue MaskLo, MaskHi, Lo, Hi;
5298 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
5300 SDValue Src0 = MGT->getValue();
5301 SDValue Src0Lo, Src0Hi;
5302 std::tie(Src0Lo, Src0Hi) = DAG.SplitVector(Src0, DL);
5305 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
5307 SDValue Chain = MGT->getChain();
5308 EVT MemoryVT = MGT->getMemoryVT();
5309 unsigned Alignment = MGT->getOriginalAlignment();
5311 EVT LoMemVT, HiMemVT;
5312 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
5314 SDValue BasePtr = MGT->getBasePtr();
5315 SDValue Index = MGT->getIndex();
5316 SDValue IndexLo, IndexHi;
5317 std::tie(IndexLo, IndexHi) = DAG.SplitVector(Index, DL);
5319 MachineMemOperand *MMO = DAG.getMachineFunction().
5320 getMachineMemOperand(MGT->getPointerInfo(),
5321 MachineMemOperand::MOLoad, LoMemVT.getStoreSize(),
5322 Alignment, MGT->getAAInfo(), MGT->getRanges());
5324 SDValue OpsLo[] = { Chain, Src0Lo, MaskLo, BasePtr, IndexLo };
5325 Lo = DAG.getMaskedGather(DAG.getVTList(LoVT, MVT::Other), LoVT, DL, OpsLo,
5328 SDValue OpsHi[] = {Chain, Src0Hi, MaskHi, BasePtr, IndexHi};
5329 Hi = DAG.getMaskedGather(DAG.getVTList(HiVT, MVT::Other), HiVT, DL, OpsHi,
5332 AddToWorklist(Lo.getNode());
5333 AddToWorklist(Hi.getNode());
5335 // Build a factor node to remember that this load is independent of the
5337 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo.getValue(1),
5340 // Legalized the chain result - switch anything that used the old chain to
5342 DAG.ReplaceAllUsesOfValueWith(SDValue(MGT, 1), Chain);
5344 SDValue GatherRes = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
5346 SDValue RetOps[] = { GatherRes, Chain };
5347 return DAG.getMergeValues(RetOps, DL);
5350 SDValue DAGCombiner::visitMLOAD(SDNode *N) {
5352 if (Level >= AfterLegalizeTypes)
5355 MaskedLoadSDNode *MLD = dyn_cast<MaskedLoadSDNode>(N);
5356 SDValue Mask = MLD->getMask();
5359 // If the MLOAD result requires splitting and the mask is provided by a
5360 // SETCC, then split both nodes and its operands before legalization. This
5361 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5362 // and enables future optimizations (e.g. min/max pattern matching on X86).
5364 if (Mask.getOpcode() == ISD::SETCC) {
5365 EVT VT = N->getValueType(0);
5367 // Check if any splitting is required.
5368 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
5369 TargetLowering::TypeSplitVector)
5372 SDValue MaskLo, MaskHi, Lo, Hi;
5373 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
5375 SDValue Src0 = MLD->getSrc0();
5376 SDValue Src0Lo, Src0Hi;
5377 std::tie(Src0Lo, Src0Hi) = DAG.SplitVector(Src0, DL);
5380 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MLD->getValueType(0));
5382 SDValue Chain = MLD->getChain();
5383 SDValue Ptr = MLD->getBasePtr();
5384 EVT MemoryVT = MLD->getMemoryVT();
5385 unsigned Alignment = MLD->getOriginalAlignment();
5387 // if Alignment is equal to the vector size,
5388 // take the half of it for the second part
5389 unsigned SecondHalfAlignment =
5390 (Alignment == MLD->getValueType(0).getSizeInBits()/8) ?
5391 Alignment/2 : Alignment;
5393 EVT LoMemVT, HiMemVT;
5394 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
5396 MachineMemOperand *MMO = DAG.getMachineFunction().
5397 getMachineMemOperand(MLD->getPointerInfo(),
5398 MachineMemOperand::MOLoad, LoMemVT.getStoreSize(),
5399 Alignment, MLD->getAAInfo(), MLD->getRanges());
5401 Lo = DAG.getMaskedLoad(LoVT, DL, Chain, Ptr, MaskLo, Src0Lo, LoMemVT, MMO,
5404 unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
5405 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5406 DAG.getConstant(IncrementSize, DL, Ptr.getValueType()));
5408 MMO = DAG.getMachineFunction().
5409 getMachineMemOperand(MLD->getPointerInfo(),
5410 MachineMemOperand::MOLoad, HiMemVT.getStoreSize(),
5411 SecondHalfAlignment, MLD->getAAInfo(), MLD->getRanges());
5413 Hi = DAG.getMaskedLoad(HiVT, DL, Chain, Ptr, MaskHi, Src0Hi, HiMemVT, MMO,
5416 AddToWorklist(Lo.getNode());
5417 AddToWorklist(Hi.getNode());
5419 // Build a factor node to remember that this load is independent of the
5421 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo.getValue(1),
5424 // Legalized the chain result - switch anything that used the old chain to
5426 DAG.ReplaceAllUsesOfValueWith(SDValue(MLD, 1), Chain);
5428 SDValue LoadRes = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
5430 SDValue RetOps[] = { LoadRes, Chain };
5431 return DAG.getMergeValues(RetOps, DL);
5436 SDValue DAGCombiner::visitVSELECT(SDNode *N) {
5437 SDValue N0 = N->getOperand(0);
5438 SDValue N1 = N->getOperand(1);
5439 SDValue N2 = N->getOperand(2);
5442 // Canonicalize integer abs.
5443 // vselect (setg[te] X, 0), X, -X ->
5444 // vselect (setgt X, -1), X, -X ->
5445 // vselect (setl[te] X, 0), -X, X ->
5446 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
5447 if (N0.getOpcode() == ISD::SETCC) {
5448 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
5449 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5451 bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
5453 if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
5454 (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) &&
5455 N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1))
5456 isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode());
5457 else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) &&
5458 N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1))
5459 isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode());
5462 EVT VT = LHS.getValueType();
5463 SDValue Shift = DAG.getNode(
5464 ISD::SRA, DL, VT, LHS,
5465 DAG.getConstant(VT.getScalarType().getSizeInBits() - 1, DL, VT));
5466 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift);
5467 AddToWorklist(Shift.getNode());
5468 AddToWorklist(Add.getNode());
5469 return DAG.getNode(ISD::XOR, DL, VT, Add, Shift);
5473 if (SimplifySelectOps(N, N1, N2))
5474 return SDValue(N, 0); // Don't revisit N.
5476 // If the VSELECT result requires splitting and the mask is provided by a
5477 // SETCC, then split both nodes and its operands before legalization. This
5478 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5479 // and enables future optimizations (e.g. min/max pattern matching on X86).
5480 if (N0.getOpcode() == ISD::SETCC) {
5481 EVT VT = N->getValueType(0);
5483 // Check if any splitting is required.
5484 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
5485 TargetLowering::TypeSplitVector)
5488 SDValue Lo, Hi, CCLo, CCHi, LL, LH, RL, RH;
5489 std::tie(CCLo, CCHi) = SplitVSETCC(N0.getNode(), DAG);
5490 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 1);
5491 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 2);
5493 Lo = DAG.getNode(N->getOpcode(), DL, LL.getValueType(), CCLo, LL, RL);
5494 Hi = DAG.getNode(N->getOpcode(), DL, LH.getValueType(), CCHi, LH, RH);
5496 // Add the new VSELECT nodes to the work list in case they need to be split
5498 AddToWorklist(Lo.getNode());
5499 AddToWorklist(Hi.getNode());
5501 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
5504 // Fold (vselect (build_vector all_ones), N1, N2) -> N1
5505 if (ISD::isBuildVectorAllOnes(N0.getNode()))
5507 // Fold (vselect (build_vector all_zeros), N1, N2) -> N2
5508 if (ISD::isBuildVectorAllZeros(N0.getNode()))
5511 // The ConvertSelectToConcatVector function is assuming both the above
5512 // checks for (vselect (build_vector all{ones,zeros) ...) have been made
5514 if (N1.getOpcode() == ISD::CONCAT_VECTORS &&
5515 N2.getOpcode() == ISD::CONCAT_VECTORS &&
5516 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
5517 if (SDValue CV = ConvertSelectToConcatVector(N, DAG))
5524 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
5525 SDValue N0 = N->getOperand(0);
5526 SDValue N1 = N->getOperand(1);
5527 SDValue N2 = N->getOperand(2);
5528 SDValue N3 = N->getOperand(3);
5529 SDValue N4 = N->getOperand(4);
5530 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
5532 // fold select_cc lhs, rhs, x, x, cc -> x
5536 // Determine if the condition we're dealing with is constant
5537 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
5538 N0, N1, CC, SDLoc(N), false);
5539 if (SCC.getNode()) {
5540 AddToWorklist(SCC.getNode());
5542 if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) {
5543 if (!SCCC->isNullValue())
5544 return N2; // cond always true -> true val
5546 return N3; // cond always false -> false val
5547 } else if (SCC->getOpcode() == ISD::UNDEF) {
5548 // When the condition is UNDEF, just return the first operand. This is
5549 // coherent the DAG creation, no setcc node is created in this case
5551 } else if (SCC.getOpcode() == ISD::SETCC) {
5552 // Fold to a simpler select_cc
5553 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(),
5554 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
5559 // If we can fold this based on the true/false value, do so.
5560 if (SimplifySelectOps(N, N2, N3))
5561 return SDValue(N, 0); // Don't revisit N.
5563 // fold select_cc into other things, such as min/max/abs
5564 return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC);
5567 SDValue DAGCombiner::visitSETCC(SDNode *N) {
5568 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
5569 cast<CondCodeSDNode>(N->getOperand(2))->get(),
5573 /// Try to fold a sext/zext/aext dag node into a ConstantSDNode or
5574 /// a build_vector of constants.
5575 /// This function is called by the DAGCombiner when visiting sext/zext/aext
5576 /// dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND).
5577 /// Vector extends are not folded if operations are legal; this is to
5578 /// avoid introducing illegal build_vector dag nodes.
5579 static SDNode *tryToFoldExtendOfConstant(SDNode *N, const TargetLowering &TLI,
5580 SelectionDAG &DAG, bool LegalTypes,
5581 bool LegalOperations) {
5582 unsigned Opcode = N->getOpcode();
5583 SDValue N0 = N->getOperand(0);
5584 EVT VT = N->getValueType(0);
5586 assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND ||
5587 Opcode == ISD::ANY_EXTEND || Opcode == ISD::SIGN_EXTEND_VECTOR_INREG)
5588 && "Expected EXTEND dag node in input!");
5590 // fold (sext c1) -> c1
5591 // fold (zext c1) -> c1
5592 // fold (aext c1) -> c1
5593 if (isa<ConstantSDNode>(N0))
5594 return DAG.getNode(Opcode, SDLoc(N), VT, N0).getNode();
5596 // fold (sext (build_vector AllConstants) -> (build_vector AllConstants)
5597 // fold (zext (build_vector AllConstants) -> (build_vector AllConstants)
5598 // fold (aext (build_vector AllConstants) -> (build_vector AllConstants)
5599 EVT SVT = VT.getScalarType();
5600 if (!(VT.isVector() &&
5601 (!LegalTypes || (!LegalOperations && TLI.isTypeLegal(SVT))) &&
5602 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())))
5605 // We can fold this node into a build_vector.
5606 unsigned VTBits = SVT.getSizeInBits();
5607 unsigned EVTBits = N0->getValueType(0).getScalarType().getSizeInBits();
5608 SmallVector<SDValue, 8> Elts;
5609 unsigned NumElts = VT.getVectorNumElements();
5612 for (unsigned i=0; i != NumElts; ++i) {
5613 SDValue Op = N0->getOperand(i);
5614 if (Op->getOpcode() == ISD::UNDEF) {
5615 Elts.push_back(DAG.getUNDEF(SVT));
5620 // Get the constant value and if needed trunc it to the size of the type.
5621 // Nodes like build_vector might have constants wider than the scalar type.
5622 APInt C = cast<ConstantSDNode>(Op)->getAPIntValue().zextOrTrunc(EVTBits);
5623 if (Opcode == ISD::SIGN_EXTEND || Opcode == ISD::SIGN_EXTEND_VECTOR_INREG)
5624 Elts.push_back(DAG.getConstant(C.sext(VTBits), DL, SVT));
5626 Elts.push_back(DAG.getConstant(C.zext(VTBits), DL, SVT));
5629 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Elts).getNode();
5632 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
5633 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
5634 // transformation. Returns true if extension are possible and the above
5635 // mentioned transformation is profitable.
5636 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
5638 SmallVectorImpl<SDNode *> &ExtendNodes,
5639 const TargetLowering &TLI) {
5640 bool HasCopyToRegUses = false;
5641 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
5642 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
5643 UE = N0.getNode()->use_end();
5648 if (UI.getUse().getResNo() != N0.getResNo())
5650 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
5651 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
5652 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
5653 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
5654 // Sign bits will be lost after a zext.
5657 for (unsigned i = 0; i != 2; ++i) {
5658 SDValue UseOp = User->getOperand(i);
5661 if (!isa<ConstantSDNode>(UseOp))
5666 ExtendNodes.push_back(User);
5669 // If truncates aren't free and there are users we can't
5670 // extend, it isn't worthwhile.
5673 // Remember if this value is live-out.
5674 if (User->getOpcode() == ISD::CopyToReg)
5675 HasCopyToRegUses = true;
5678 if (HasCopyToRegUses) {
5679 bool BothLiveOut = false;
5680 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
5682 SDUse &Use = UI.getUse();
5683 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
5689 // Both unextended and extended values are live out. There had better be
5690 // a good reason for the transformation.
5691 return ExtendNodes.size();
5696 void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
5697 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
5698 ISD::NodeType ExtType) {
5699 // Extend SetCC uses if necessary.
5700 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
5701 SDNode *SetCC = SetCCs[i];
5702 SmallVector<SDValue, 4> Ops;
5704 for (unsigned j = 0; j != 2; ++j) {
5705 SDValue SOp = SetCC->getOperand(j);
5707 Ops.push_back(ExtLoad);
5709 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
5712 Ops.push_back(SetCC->getOperand(2));
5713 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), Ops));
5717 // FIXME: Bring more similar combines here, common to sext/zext (maybe aext?).
5718 SDValue DAGCombiner::CombineExtLoad(SDNode *N) {
5719 SDValue N0 = N->getOperand(0);
5720 EVT DstVT = N->getValueType(0);
5721 EVT SrcVT = N0.getValueType();
5723 assert((N->getOpcode() == ISD::SIGN_EXTEND ||
5724 N->getOpcode() == ISD::ZERO_EXTEND) &&
5725 "Unexpected node type (not an extend)!");
5727 // fold (sext (load x)) to multiple smaller sextloads; same for zext.
5728 // For example, on a target with legal v4i32, but illegal v8i32, turn:
5729 // (v8i32 (sext (v8i16 (load x))))
5731 // (v8i32 (concat_vectors (v4i32 (sextload x)),
5732 // (v4i32 (sextload (x + 16)))))
5733 // Where uses of the original load, i.e.:
5735 // are replaced with:
5737 // (v8i32 (concat_vectors (v4i32 (sextload x)),
5738 // (v4i32 (sextload (x + 16)))))))
5740 // This combine is only applicable to illegal, but splittable, vectors.
5741 // All legal types, and illegal non-vector types, are handled elsewhere.
5742 // This combine is controlled by TargetLowering::isVectorLoadExtDesirable.
5744 if (N0->getOpcode() != ISD::LOAD)
5747 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5749 if (!ISD::isNON_EXTLoad(LN0) || !ISD::isUNINDEXEDLoad(LN0) ||
5750 !N0.hasOneUse() || LN0->isVolatile() || !DstVT.isVector() ||
5751 !DstVT.isPow2VectorType() || !TLI.isVectorLoadExtDesirable(SDValue(N, 0)))
5754 SmallVector<SDNode *, 4> SetCCs;
5755 if (!ExtendUsesToFormExtLoad(N, N0, N->getOpcode(), SetCCs, TLI))
5758 ISD::LoadExtType ExtType =
5759 N->getOpcode() == ISD::SIGN_EXTEND ? ISD::SEXTLOAD : ISD::ZEXTLOAD;
5761 // Try to split the vector types to get down to legal types.
5762 EVT SplitSrcVT = SrcVT;
5763 EVT SplitDstVT = DstVT;
5764 while (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT) &&
5765 SplitSrcVT.getVectorNumElements() > 1) {
5766 SplitDstVT = DAG.GetSplitDestVTs(SplitDstVT).first;
5767 SplitSrcVT = DAG.GetSplitDestVTs(SplitSrcVT).first;
5770 if (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT))
5774 const unsigned NumSplits =
5775 DstVT.getVectorNumElements() / SplitDstVT.getVectorNumElements();
5776 const unsigned Stride = SplitSrcVT.getStoreSize();
5777 SmallVector<SDValue, 4> Loads;
5778 SmallVector<SDValue, 4> Chains;
5780 SDValue BasePtr = LN0->getBasePtr();
5781 for (unsigned Idx = 0; Idx < NumSplits; Idx++) {
5782 const unsigned Offset = Idx * Stride;
5783 const unsigned Align = MinAlign(LN0->getAlignment(), Offset);
5785 SDValue SplitLoad = DAG.getExtLoad(
5786 ExtType, DL, SplitDstVT, LN0->getChain(), BasePtr,
5787 LN0->getPointerInfo().getWithOffset(Offset), SplitSrcVT,
5788 LN0->isVolatile(), LN0->isNonTemporal(), LN0->isInvariant(),
5789 Align, LN0->getAAInfo());
5791 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
5792 DAG.getConstant(Stride, DL, BasePtr.getValueType()));
5794 Loads.push_back(SplitLoad.getValue(0));
5795 Chains.push_back(SplitLoad.getValue(1));
5798 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
5799 SDValue NewValue = DAG.getNode(ISD::CONCAT_VECTORS, DL, DstVT, Loads);
5801 CombineTo(N, NewValue);
5803 // Replace uses of the original load (before extension)
5804 // with a truncate of the concatenated sextloaded vectors.
5806 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(), NewValue);
5807 CombineTo(N0.getNode(), Trunc, NewChain);
5808 ExtendSetCCUses(SetCCs, Trunc, NewValue, DL,
5809 (ISD::NodeType)N->getOpcode());
5810 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5813 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
5814 SDValue N0 = N->getOperand(0);
5815 EVT VT = N->getValueType(0);
5817 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5819 return SDValue(Res, 0);
5821 // fold (sext (sext x)) -> (sext x)
5822 // fold (sext (aext x)) -> (sext x)
5823 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5824 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT,
5827 if (N0.getOpcode() == ISD::TRUNCATE) {
5828 // fold (sext (truncate (load x))) -> (sext (smaller load x))
5829 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
5830 if (SDValue NarrowLoad = ReduceLoadWidth(N0.getNode())) {
5831 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5832 if (NarrowLoad.getNode() != N0.getNode()) {
5833 CombineTo(N0.getNode(), NarrowLoad);
5834 // CombineTo deleted the truncate, if needed, but not what's under it.
5837 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5840 // See if the value being truncated is already sign extended. If so, just
5841 // eliminate the trunc/sext pair.
5842 SDValue Op = N0.getOperand(0);
5843 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
5844 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
5845 unsigned DestBits = VT.getScalarType().getSizeInBits();
5846 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
5848 if (OpBits == DestBits) {
5849 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
5850 // bits, it is already ready.
5851 if (NumSignBits > DestBits-MidBits)
5853 } else if (OpBits < DestBits) {
5854 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
5855 // bits, just sext from i32.
5856 if (NumSignBits > OpBits-MidBits)
5857 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, Op);
5859 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
5860 // bits, just truncate to i32.
5861 if (NumSignBits > OpBits-MidBits)
5862 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5865 // fold (sext (truncate x)) -> (sextinreg x).
5866 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
5867 N0.getValueType())) {
5868 if (OpBits < DestBits)
5869 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op);
5870 else if (OpBits > DestBits)
5871 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op);
5872 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, Op,
5873 DAG.getValueType(N0.getValueType()));
5877 // fold (sext (load x)) -> (sext (truncate (sextload x)))
5878 // Only generate vector extloads when 1) they're legal, and 2) they are
5879 // deemed desirable by the target.
5880 if (ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5881 ((!LegalOperations && !VT.isVector() &&
5882 !cast<LoadSDNode>(N0)->isVolatile()) ||
5883 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, N0.getValueType()))) {
5884 bool DoXform = true;
5885 SmallVector<SDNode*, 4> SetCCs;
5886 if (!N0.hasOneUse())
5887 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
5889 DoXform &= TLI.isVectorLoadExtDesirable(SDValue(N, 0));
5891 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5892 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5894 LN0->getBasePtr(), N0.getValueType(),
5895 LN0->getMemOperand());
5896 CombineTo(N, ExtLoad);
5897 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5898 N0.getValueType(), ExtLoad);
5899 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5900 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5902 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5906 // fold (sext (load x)) to multiple smaller sextloads.
5907 // Only on illegal but splittable vectors.
5908 if (SDValue ExtLoad = CombineExtLoad(N))
5911 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
5912 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
5913 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5914 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5915 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5916 EVT MemVT = LN0->getMemoryVT();
5917 if ((!LegalOperations && !LN0->isVolatile()) ||
5918 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, MemVT)) {
5919 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5921 LN0->getBasePtr(), MemVT,
5922 LN0->getMemOperand());
5923 CombineTo(N, ExtLoad);
5924 CombineTo(N0.getNode(),
5925 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5926 N0.getValueType(), ExtLoad),
5927 ExtLoad.getValue(1));
5928 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5932 // fold (sext (and/or/xor (load x), cst)) ->
5933 // (and/or/xor (sextload x), (sext cst))
5934 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5935 N0.getOpcode() == ISD::XOR) &&
5936 isa<LoadSDNode>(N0.getOperand(0)) &&
5937 N0.getOperand(1).getOpcode() == ISD::Constant &&
5938 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, N0.getValueType()) &&
5939 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5940 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5941 if (LN0->getExtensionType() != ISD::ZEXTLOAD && LN0->isUnindexed()) {
5942 bool DoXform = true;
5943 SmallVector<SDNode*, 4> SetCCs;
5944 if (!N0.hasOneUse())
5945 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
5948 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN0), VT,
5949 LN0->getChain(), LN0->getBasePtr(),
5951 LN0->getMemOperand());
5952 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5953 Mask = Mask.sext(VT.getSizeInBits());
5955 SDValue And = DAG.getNode(N0.getOpcode(), DL, VT,
5956 ExtLoad, DAG.getConstant(Mask, DL, VT));
5957 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5958 SDLoc(N0.getOperand(0)),
5959 N0.getOperand(0).getValueType(), ExtLoad);
5961 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5962 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, DL,
5964 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5969 if (N0.getOpcode() == ISD::SETCC) {
5970 EVT N0VT = N0.getOperand(0).getValueType();
5971 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
5972 // Only do this before legalize for now.
5973 if (VT.isVector() && !LegalOperations &&
5974 TLI.getBooleanContents(N0VT) ==
5975 TargetLowering::ZeroOrNegativeOneBooleanContent) {
5976 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
5977 // of the same size as the compared operands. Only optimize sext(setcc())
5978 // if this is the case.
5979 EVT SVT = getSetCCResultType(N0VT);
5981 // We know that the # elements of the results is the same as the
5982 // # elements of the compare (and the # elements of the compare result
5983 // for that matter). Check to see that they are the same size. If so,
5984 // we know that the element size of the sext'd result matches the
5985 // element size of the compare operands.
5986 if (VT.getSizeInBits() == SVT.getSizeInBits())
5987 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5989 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5991 // If the desired elements are smaller or larger than the source
5992 // elements we can use a matching integer vector type and then
5993 // truncate/sign extend
5994 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5995 if (SVT == MatchingVectorType) {
5996 SDValue VsetCC = DAG.getSetCC(SDLoc(N), MatchingVectorType,
5997 N0.getOperand(0), N0.getOperand(1),
5998 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5999 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
6003 // sext(setcc x, y, cc) -> (select (setcc x, y, cc), -1, 0)
6004 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
6007 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), DL, VT);
6009 SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1),
6010 NegOne, DAG.getConstant(0, DL, VT),
6011 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
6012 if (SCC.getNode()) return SCC;
6014 if (!VT.isVector()) {
6015 EVT SetCCVT = getSetCCResultType(N0.getOperand(0).getValueType());
6016 if (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, SetCCVT)) {
6018 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
6019 SDValue SetCC = DAG.getSetCC(DL, SetCCVT,
6020 N0.getOperand(0), N0.getOperand(1), CC);
6021 return DAG.getSelect(DL, VT, SetCC,
6022 NegOne, DAG.getConstant(0, DL, VT));
6027 // fold (sext x) -> (zext x) if the sign bit is known zero.
6028 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
6029 DAG.SignBitIsZero(N0))
6030 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0);
6035 // isTruncateOf - If N is a truncate of some other value, return true, record
6036 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
6037 // This function computes KnownZero to avoid a duplicated call to
6038 // computeKnownBits in the caller.
6039 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
6042 if (N->getOpcode() == ISD::TRUNCATE) {
6043 Op = N->getOperand(0);
6044 DAG.computeKnownBits(Op, KnownZero, KnownOne);
6048 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
6049 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
6052 SDValue Op0 = N->getOperand(0);
6053 SDValue Op1 = N->getOperand(1);
6054 assert(Op0.getValueType() == Op1.getValueType());
6056 if (isNullConstant(Op0))
6058 else if (isNullConstant(Op1))
6063 DAG.computeKnownBits(Op, KnownZero, KnownOne);
6065 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
6071 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
6072 SDValue N0 = N->getOperand(0);
6073 EVT VT = N->getValueType(0);
6075 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
6077 return SDValue(Res, 0);
6079 // fold (zext (zext x)) -> (zext x)
6080 // fold (zext (aext x)) -> (zext x)
6081 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
6082 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT,
6085 // fold (zext (truncate x)) -> (zext x) or
6086 // (zext (truncate x)) -> (truncate x)
6087 // This is valid when the truncated bits of x are already zero.
6088 // FIXME: We should extend this to work for vectors too.
6091 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
6092 APInt TruncatedBits =
6093 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
6094 APInt(Op.getValueSizeInBits(), 0) :
6095 APInt::getBitsSet(Op.getValueSizeInBits(),
6096 N0.getValueSizeInBits(),
6097 std::min(Op.getValueSizeInBits(),
6098 VT.getSizeInBits()));
6099 if (TruncatedBits == (KnownZero & TruncatedBits)) {
6100 if (VT.bitsGT(Op.getValueType()))
6101 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, Op);
6102 if (VT.bitsLT(Op.getValueType()))
6103 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
6109 // fold (zext (truncate (load x))) -> (zext (smaller load x))
6110 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
6111 if (N0.getOpcode() == ISD::TRUNCATE) {
6112 if (SDValue NarrowLoad = ReduceLoadWidth(N0.getNode())) {
6113 SDNode* oye = N0.getNode()->getOperand(0).getNode();
6114 if (NarrowLoad.getNode() != N0.getNode()) {
6115 CombineTo(N0.getNode(), NarrowLoad);
6116 // CombineTo deleted the truncate, if needed, but not what's under it.
6119 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6123 // fold (zext (truncate x)) -> (and x, mask)
6124 if (N0.getOpcode() == ISD::TRUNCATE) {
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 EVT SrcVT = N0.getOperand(0).getValueType();
6138 EVT MinVT = N0.getValueType();
6140 // Try to mask before the extension to avoid having to generate a larger mask,
6141 // possibly over several sub-vectors.
6142 if (SrcVT.bitsLT(VT)) {
6143 if (!LegalOperations || (TLI.isOperationLegal(ISD::AND, SrcVT) &&
6144 TLI.isOperationLegal(ISD::ZERO_EXTEND, VT))) {
6145 SDValue Op = N0.getOperand(0);
6146 Op = DAG.getZeroExtendInReg(Op, SDLoc(N), MinVT.getScalarType());
6147 AddToWorklist(Op.getNode());
6148 return DAG.getZExtOrTrunc(Op, SDLoc(N), VT);
6152 if (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT)) {
6153 SDValue Op = N0.getOperand(0);
6154 if (SrcVT.bitsLT(VT)) {
6155 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, Op);
6156 AddToWorklist(Op.getNode());
6157 } else if (SrcVT.bitsGT(VT)) {
6158 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
6159 AddToWorklist(Op.getNode());
6161 return DAG.getZeroExtendInReg(Op, SDLoc(N), MinVT.getScalarType());
6165 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
6166 // if either of the casts is not free.
6167 if (N0.getOpcode() == ISD::AND &&
6168 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
6169 N0.getOperand(1).getOpcode() == ISD::Constant &&
6170 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
6171 N0.getValueType()) ||
6172 !TLI.isZExtFree(N0.getValueType(), VT))) {
6173 SDValue X = N0.getOperand(0).getOperand(0);
6174 if (X.getValueType().bitsLT(VT)) {
6175 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(X), VT, X);
6176 } else if (X.getValueType().bitsGT(VT)) {
6177 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
6179 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
6180 Mask = Mask.zext(VT.getSizeInBits());
6182 return DAG.getNode(ISD::AND, DL, VT,
6183 X, DAG.getConstant(Mask, DL, VT));
6186 // fold (zext (load x)) -> (zext (truncate (zextload x)))
6187 // Only generate vector extloads when 1) they're legal, and 2) they are
6188 // deemed desirable by the target.
6189 if (ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
6190 ((!LegalOperations && !VT.isVector() &&
6191 !cast<LoadSDNode>(N0)->isVolatile()) ||
6192 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, N0.getValueType()))) {
6193 bool DoXform = true;
6194 SmallVector<SDNode*, 4> SetCCs;
6195 if (!N0.hasOneUse())
6196 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
6198 DoXform &= TLI.isVectorLoadExtDesirable(SDValue(N, 0));
6200 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6201 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
6203 LN0->getBasePtr(), N0.getValueType(),
6204 LN0->getMemOperand());
6205 CombineTo(N, ExtLoad);
6206 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
6207 N0.getValueType(), ExtLoad);
6208 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
6210 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
6212 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6216 // fold (zext (load x)) to multiple smaller zextloads.
6217 // Only on illegal but splittable vectors.
6218 if (SDValue ExtLoad = CombineExtLoad(N))
6221 // fold (zext (and/or/xor (load x), cst)) ->
6222 // (and/or/xor (zextload x), (zext cst))
6223 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
6224 N0.getOpcode() == ISD::XOR) &&
6225 isa<LoadSDNode>(N0.getOperand(0)) &&
6226 N0.getOperand(1).getOpcode() == ISD::Constant &&
6227 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, N0.getValueType()) &&
6228 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
6229 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
6230 if (LN0->getExtensionType() != ISD::SEXTLOAD && LN0->isUnindexed()) {
6231 bool DoXform = true;
6232 SmallVector<SDNode*, 4> SetCCs;
6233 if (!N0.hasOneUse())
6234 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
6237 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), VT,
6238 LN0->getChain(), LN0->getBasePtr(),
6240 LN0->getMemOperand());
6241 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
6242 Mask = Mask.zext(VT.getSizeInBits());
6244 SDValue And = DAG.getNode(N0.getOpcode(), DL, VT,
6245 ExtLoad, DAG.getConstant(Mask, DL, VT));
6246 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
6247 SDLoc(N0.getOperand(0)),
6248 N0.getOperand(0).getValueType(), ExtLoad);
6250 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
6251 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, DL,
6253 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6258 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
6259 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
6260 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
6261 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
6262 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6263 EVT MemVT = LN0->getMemoryVT();
6264 if ((!LegalOperations && !LN0->isVolatile()) ||
6265 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT)) {
6266 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
6268 LN0->getBasePtr(), MemVT,
6269 LN0->getMemOperand());
6270 CombineTo(N, ExtLoad);
6271 CombineTo(N0.getNode(),
6272 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(),
6274 ExtLoad.getValue(1));
6275 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6279 if (N0.getOpcode() == ISD::SETCC) {
6280 if (!LegalOperations && VT.isVector() &&
6281 N0.getValueType().getVectorElementType() == MVT::i1) {
6282 EVT N0VT = N0.getOperand(0).getValueType();
6283 if (getSetCCResultType(N0VT) == N0.getValueType())
6286 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
6287 // Only do this before legalize for now.
6288 EVT EltVT = VT.getVectorElementType();
6290 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
6291 DAG.getConstant(1, DL, EltVT));
6292 if (VT.getSizeInBits() == N0VT.getSizeInBits())
6293 // We know that the # elements of the results is the same as the
6294 // # elements of the compare (and the # elements of the compare result
6295 // for that matter). Check to see that they are the same size. If so,
6296 // we know that the element size of the sext'd result matches the
6297 // element size of the compare operands.
6298 return DAG.getNode(ISD::AND, DL, VT,
6299 DAG.getSetCC(DL, VT, N0.getOperand(0),
6301 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
6302 DAG.getNode(ISD::BUILD_VECTOR, DL, VT,
6305 // If the desired elements are smaller or larger than the source
6306 // elements we can use a matching integer vector type and then
6307 // truncate/sign extend
6308 EVT MatchingElementType =
6309 EVT::getIntegerVT(*DAG.getContext(),
6310 N0VT.getScalarType().getSizeInBits());
6311 EVT MatchingVectorType =
6312 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
6313 N0VT.getVectorNumElements());
6315 DAG.getSetCC(DL, MatchingVectorType, N0.getOperand(0),
6317 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6318 return DAG.getNode(ISD::AND, DL, VT,
6319 DAG.getSExtOrTrunc(VsetCC, DL, VT),
6320 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, OneOps));
6323 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
6326 SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1),
6327 DAG.getConstant(1, DL, VT), DAG.getConstant(0, DL, VT),
6328 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
6329 if (SCC.getNode()) return SCC;
6332 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
6333 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
6334 isa<ConstantSDNode>(N0.getOperand(1)) &&
6335 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
6337 SDValue ShAmt = N0.getOperand(1);
6338 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
6339 if (N0.getOpcode() == ISD::SHL) {
6340 SDValue InnerZExt = N0.getOperand(0);
6341 // If the original shl may be shifting out bits, do not perform this
6343 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
6344 InnerZExt.getOperand(0).getValueType().getSizeInBits();
6345 if (ShAmtVal > KnownZeroBits)
6351 // Ensure that the shift amount is wide enough for the shifted value.
6352 if (VT.getSizeInBits() >= 256)
6353 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
6355 return DAG.getNode(N0.getOpcode(), DL, VT,
6356 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
6363 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
6364 SDValue N0 = N->getOperand(0);
6365 EVT VT = N->getValueType(0);
6367 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
6369 return SDValue(Res, 0);
6371 // fold (aext (aext x)) -> (aext x)
6372 // fold (aext (zext x)) -> (zext x)
6373 // fold (aext (sext x)) -> (sext x)
6374 if (N0.getOpcode() == ISD::ANY_EXTEND ||
6375 N0.getOpcode() == ISD::ZERO_EXTEND ||
6376 N0.getOpcode() == ISD::SIGN_EXTEND)
6377 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0));
6379 // fold (aext (truncate (load x))) -> (aext (smaller load x))
6380 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
6381 if (N0.getOpcode() == ISD::TRUNCATE) {
6382 if (SDValue NarrowLoad = ReduceLoadWidth(N0.getNode())) {
6383 SDNode* oye = N0.getNode()->getOperand(0).getNode();
6384 if (NarrowLoad.getNode() != N0.getNode()) {
6385 CombineTo(N0.getNode(), NarrowLoad);
6386 // CombineTo deleted the truncate, if needed, but not what's under it.
6389 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6393 // fold (aext (truncate x))
6394 if (N0.getOpcode() == ISD::TRUNCATE) {
6395 SDValue TruncOp = N0.getOperand(0);
6396 if (TruncOp.getValueType() == VT)
6397 return TruncOp; // x iff x size == zext size.
6398 if (TruncOp.getValueType().bitsGT(VT))
6399 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, TruncOp);
6400 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, TruncOp);
6403 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
6404 // if the trunc is not free.
6405 if (N0.getOpcode() == ISD::AND &&
6406 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
6407 N0.getOperand(1).getOpcode() == ISD::Constant &&
6408 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
6409 N0.getValueType())) {
6410 SDValue X = N0.getOperand(0).getOperand(0);
6411 if (X.getValueType().bitsLT(VT)) {
6412 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, X);
6413 } else if (X.getValueType().bitsGT(VT)) {
6414 X = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, X);
6416 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
6417 Mask = Mask.zext(VT.getSizeInBits());
6419 return DAG.getNode(ISD::AND, DL, VT,
6420 X, DAG.getConstant(Mask, DL, VT));
6423 // fold (aext (load x)) -> (aext (truncate (extload x)))
6424 // None of the supported targets knows how to perform load and any_ext
6425 // on vectors in one instruction. We only perform this transformation on
6427 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
6428 ISD::isUNINDEXEDLoad(N0.getNode()) &&
6429 TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) {
6430 bool DoXform = true;
6431 SmallVector<SDNode*, 4> SetCCs;
6432 if (!N0.hasOneUse())
6433 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
6435 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6436 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
6438 LN0->getBasePtr(), N0.getValueType(),
6439 LN0->getMemOperand());
6440 CombineTo(N, ExtLoad);
6441 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
6442 N0.getValueType(), ExtLoad);
6443 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
6444 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
6446 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6450 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
6451 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
6452 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
6453 if (N0.getOpcode() == ISD::LOAD &&
6454 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
6456 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6457 ISD::LoadExtType ExtType = LN0->getExtensionType();
6458 EVT MemVT = LN0->getMemoryVT();
6459 if (!LegalOperations || TLI.isLoadExtLegal(ExtType, VT, MemVT)) {
6460 SDValue ExtLoad = DAG.getExtLoad(ExtType, SDLoc(N),
6461 VT, LN0->getChain(), LN0->getBasePtr(),
6462 MemVT, LN0->getMemOperand());
6463 CombineTo(N, ExtLoad);
6464 CombineTo(N0.getNode(),
6465 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
6466 N0.getValueType(), ExtLoad),
6467 ExtLoad.getValue(1));
6468 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6472 if (N0.getOpcode() == ISD::SETCC) {
6474 // aext(setcc) -> vsetcc
6475 // aext(setcc) -> truncate(vsetcc)
6476 // aext(setcc) -> aext(vsetcc)
6477 // Only do this before legalize for now.
6478 if (VT.isVector() && !LegalOperations) {
6479 EVT N0VT = N0.getOperand(0).getValueType();
6480 // We know that the # elements of the results is the same as the
6481 // # elements of the compare (and the # elements of the compare result
6482 // for that matter). Check to see that they are the same size. If so,
6483 // we know that the element size of the sext'd result matches the
6484 // element size of the compare operands.
6485 if (VT.getSizeInBits() == N0VT.getSizeInBits())
6486 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
6488 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6489 // If the desired elements are smaller or larger than the source
6490 // elements we can use a matching integer vector type and then
6491 // truncate/any extend
6493 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
6495 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
6497 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6498 return DAG.getAnyExtOrTrunc(VsetCC, SDLoc(N), VT);
6502 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
6505 SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1),
6506 DAG.getConstant(1, DL, VT), DAG.getConstant(0, DL, VT),
6507 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
6515 /// See if the specified operand can be simplified with the knowledge that only
6516 /// the bits specified by Mask are used. If so, return the simpler operand,
6517 /// otherwise return a null SDValue.
6518 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
6519 switch (V.getOpcode()) {
6521 case ISD::Constant: {
6522 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
6523 assert(CV && "Const value should be ConstSDNode.");
6524 const APInt &CVal = CV->getAPIntValue();
6525 APInt NewVal = CVal & Mask;
6527 return DAG.getConstant(NewVal, SDLoc(V), V.getValueType());
6532 // If the LHS or RHS don't contribute bits to the or, drop them.
6533 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
6534 return V.getOperand(1);
6535 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
6536 return V.getOperand(0);
6539 // Only look at single-use SRLs.
6540 if (!V.getNode()->hasOneUse())
6542 if (ConstantSDNode *RHSC = getAsNonOpaqueConstant(V.getOperand(1))) {
6543 // See if we can recursively simplify the LHS.
6544 unsigned Amt = RHSC->getZExtValue();
6546 // Watch out for shift count overflow though.
6547 if (Amt >= Mask.getBitWidth()) break;
6548 APInt NewMask = Mask << Amt;
6549 if (SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask))
6550 return DAG.getNode(ISD::SRL, SDLoc(V), V.getValueType(),
6551 SimplifyLHS, V.getOperand(1));
6557 /// If the result of a wider load is shifted to right of N bits and then
6558 /// truncated to a narrower type and where N is a multiple of number of bits of
6559 /// the narrower type, transform it to a narrower load from address + N / num of
6560 /// bits of new type. If the result is to be extended, also fold the extension
6561 /// to form a extending load.
6562 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
6563 unsigned Opc = N->getOpcode();
6565 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
6566 SDValue N0 = N->getOperand(0);
6567 EVT VT = N->getValueType(0);
6570 // This transformation isn't valid for vector loads.
6574 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
6576 if (Opc == ISD::SIGN_EXTEND_INREG) {
6577 ExtType = ISD::SEXTLOAD;
6578 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
6579 } else if (Opc == ISD::SRL) {
6580 // Another special-case: SRL is basically zero-extending a narrower value.
6581 ExtType = ISD::ZEXTLOAD;
6583 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
6584 if (!N01) return SDValue();
6585 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
6586 VT.getSizeInBits() - N01->getZExtValue());
6588 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, VT, ExtVT))
6591 unsigned EVTBits = ExtVT.getSizeInBits();
6593 // Do not generate loads of non-round integer types since these can
6594 // be expensive (and would be wrong if the type is not byte sized).
6595 if (!ExtVT.isRound())
6599 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
6600 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
6601 ShAmt = N01->getZExtValue();
6602 // Is the shift amount a multiple of size of VT?
6603 if ((ShAmt & (EVTBits-1)) == 0) {
6604 N0 = N0.getOperand(0);
6605 // Is the load width a multiple of size of VT?
6606 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
6610 // At this point, we must have a load or else we can't do the transform.
6611 if (!isa<LoadSDNode>(N0)) return SDValue();
6613 // Because a SRL must be assumed to *need* to zero-extend the high bits
6614 // (as opposed to anyext the high bits), we can't combine the zextload
6615 // lowering of SRL and an sextload.
6616 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
6619 // If the shift amount is larger than the input type then we're not
6620 // accessing any of the loaded bytes. If the load was a zextload/extload
6621 // then the result of the shift+trunc is zero/undef (handled elsewhere).
6622 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
6627 // If the load is shifted left (and the result isn't shifted back right),
6628 // we can fold the truncate through the shift.
6629 unsigned ShLeftAmt = 0;
6630 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
6631 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
6632 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
6633 ShLeftAmt = N01->getZExtValue();
6634 N0 = N0.getOperand(0);
6638 // If we haven't found a load, we can't narrow it. Don't transform one with
6639 // multiple uses, this would require adding a new load.
6640 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse())
6643 // Don't change the width of a volatile load.
6644 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6645 if (LN0->isVolatile())
6648 // Verify that we are actually reducing a load width here.
6649 if (LN0->getMemoryVT().getSizeInBits() < EVTBits)
6652 // For the transform to be legal, the load must produce only two values
6653 // (the value loaded and the chain). Don't transform a pre-increment
6654 // load, for example, which produces an extra value. Otherwise the
6655 // transformation is not equivalent, and the downstream logic to replace
6656 // uses gets things wrong.
6657 if (LN0->getNumValues() > 2)
6660 // If the load that we're shrinking is an extload and we're not just
6661 // discarding the extension we can't simply shrink the load. Bail.
6662 // TODO: It would be possible to merge the extensions in some cases.
6663 if (LN0->getExtensionType() != ISD::NON_EXTLOAD &&
6664 LN0->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits() + ShAmt)
6667 if (!TLI.shouldReduceLoadWidth(LN0, ExtType, ExtVT))
6670 EVT PtrType = N0.getOperand(1).getValueType();
6672 if (PtrType == MVT::Untyped || PtrType.isExtended())
6673 // It's not possible to generate a constant of extended or untyped type.
6676 // For big endian targets, we need to adjust the offset to the pointer to
6677 // load the correct bytes.
6678 if (DAG.getDataLayout().isBigEndian()) {
6679 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
6680 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
6681 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
6684 uint64_t PtrOff = ShAmt / 8;
6685 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
6687 SDValue NewPtr = DAG.getNode(ISD::ADD, DL,
6688 PtrType, LN0->getBasePtr(),
6689 DAG.getConstant(PtrOff, DL, PtrType));
6690 AddToWorklist(NewPtr.getNode());
6693 if (ExtType == ISD::NON_EXTLOAD)
6694 Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr,
6695 LN0->getPointerInfo().getWithOffset(PtrOff),
6696 LN0->isVolatile(), LN0->isNonTemporal(),
6697 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
6699 Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(),NewPtr,
6700 LN0->getPointerInfo().getWithOffset(PtrOff),
6701 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
6702 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
6704 // Replace the old load's chain with the new load's chain.
6705 WorklistRemover DeadNodes(*this);
6706 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
6708 // Shift the result left, if we've swallowed a left shift.
6709 SDValue Result = Load;
6710 if (ShLeftAmt != 0) {
6711 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
6712 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
6714 // If the shift amount is as large as the result size (but, presumably,
6715 // no larger than the source) then the useful bits of the result are
6716 // zero; we can't simply return the shortened shift, because the result
6717 // of that operation is undefined.
6719 if (ShLeftAmt >= VT.getSizeInBits())
6720 Result = DAG.getConstant(0, DL, VT);
6722 Result = DAG.getNode(ISD::SHL, DL, VT,
6723 Result, DAG.getConstant(ShLeftAmt, DL, ShImmTy));
6726 // Return the new loaded value.
6730 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
6731 SDValue N0 = N->getOperand(0);
6732 SDValue N1 = N->getOperand(1);
6733 EVT VT = N->getValueType(0);
6734 EVT EVT = cast<VTSDNode>(N1)->getVT();
6735 unsigned VTBits = VT.getScalarType().getSizeInBits();
6736 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
6738 // fold (sext_in_reg c1) -> c1
6739 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
6740 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1);
6742 // If the input is already sign extended, just drop the extension.
6743 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
6746 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
6747 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
6748 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT()))
6749 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
6750 N0.getOperand(0), N1);
6752 // fold (sext_in_reg (sext x)) -> (sext x)
6753 // fold (sext_in_reg (aext x)) -> (sext x)
6754 // if x is small enough.
6755 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
6756 SDValue N00 = N0.getOperand(0);
6757 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
6758 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
6759 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
6762 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
6763 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
6764 return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT);
6766 // fold operands of sext_in_reg based on knowledge that the top bits are not
6768 if (SimplifyDemandedBits(SDValue(N, 0)))
6769 return SDValue(N, 0);
6771 // fold (sext_in_reg (load x)) -> (smaller sextload x)
6772 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
6773 if (SDValue NarrowLoad = ReduceLoadWidth(N))
6776 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
6777 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
6778 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
6779 if (N0.getOpcode() == ISD::SRL) {
6780 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
6781 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
6782 // We can turn this into an SRA iff the input to the SRL is already sign
6784 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
6785 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
6786 return DAG.getNode(ISD::SRA, SDLoc(N), VT,
6787 N0.getOperand(0), N0.getOperand(1));
6791 // fold (sext_inreg (extload x)) -> (sextload x)
6792 if (ISD::isEXTLoad(N0.getNode()) &&
6793 ISD::isUNINDEXEDLoad(N0.getNode()) &&
6794 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
6795 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6796 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) {
6797 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6798 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
6800 LN0->getBasePtr(), EVT,
6801 LN0->getMemOperand());
6802 CombineTo(N, ExtLoad);
6803 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
6804 AddToWorklist(ExtLoad.getNode());
6805 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6807 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
6808 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
6810 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
6811 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6812 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) {
6813 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6814 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
6816 LN0->getBasePtr(), EVT,
6817 LN0->getMemOperand());
6818 CombineTo(N, ExtLoad);
6819 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
6820 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6823 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
6824 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
6825 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
6826 N0.getOperand(1), false);
6827 if (BSwap.getNode())
6828 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
6832 // Fold a sext_inreg of a build_vector of ConstantSDNodes or undefs
6833 // into a build_vector.
6834 if (ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
6835 SmallVector<SDValue, 8> Elts;
6836 unsigned NumElts = N0->getNumOperands();
6837 unsigned ShAmt = VTBits - EVTBits;
6839 for (unsigned i = 0; i != NumElts; ++i) {
6840 SDValue Op = N0->getOperand(i);
6841 if (Op->getOpcode() == ISD::UNDEF) {
6846 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
6847 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
6848 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
6849 SDLoc(Op), Op.getValueType()));
6852 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Elts);
6858 SDValue DAGCombiner::visitSIGN_EXTEND_VECTOR_INREG(SDNode *N) {
6859 SDValue N0 = N->getOperand(0);
6860 EVT VT = N->getValueType(0);
6862 if (N0.getOpcode() == ISD::UNDEF)
6863 return DAG.getUNDEF(VT);
6865 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
6867 return SDValue(Res, 0);
6872 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
6873 SDValue N0 = N->getOperand(0);
6874 EVT VT = N->getValueType(0);
6875 bool isLE = DAG.getDataLayout().isLittleEndian();
6878 if (N0.getValueType() == N->getValueType(0))
6880 // fold (truncate c1) -> c1
6881 if (isConstantIntBuildVectorOrConstantInt(N0))
6882 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0);
6883 // fold (truncate (truncate x)) -> (truncate x)
6884 if (N0.getOpcode() == ISD::TRUNCATE)
6885 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6886 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
6887 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
6888 N0.getOpcode() == ISD::SIGN_EXTEND ||
6889 N0.getOpcode() == ISD::ANY_EXTEND) {
6890 if (N0.getOperand(0).getValueType().bitsLT(VT))
6891 // if the source is smaller than the dest, we still need an extend
6892 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
6894 if (N0.getOperand(0).getValueType().bitsGT(VT))
6895 // if the source is larger than the dest, than we just need the truncate
6896 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6897 // if the source and dest are the same type, we can drop both the extend
6898 // and the truncate.
6899 return N0.getOperand(0);
6902 // Fold extract-and-trunc into a narrow extract. For example:
6903 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
6904 // i32 y = TRUNCATE(i64 x)
6906 // v16i8 b = BITCAST (v2i64 val)
6907 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
6909 // Note: We only run this optimization after type legalization (which often
6910 // creates this pattern) and before operation legalization after which
6911 // we need to be more careful about the vector instructions that we generate.
6912 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6913 LegalTypes && !LegalOperations && N0->hasOneUse() && VT != MVT::i1) {
6915 EVT VecTy = N0.getOperand(0).getValueType();
6916 EVT ExTy = N0.getValueType();
6917 EVT TrTy = N->getValueType(0);
6919 unsigned NumElem = VecTy.getVectorNumElements();
6920 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
6922 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
6923 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
6925 SDValue EltNo = N0->getOperand(1);
6926 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
6927 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
6928 EVT IndexTy = TLI.getVectorIdxTy(DAG.getDataLayout());
6929 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
6931 SDValue V = DAG.getNode(ISD::BITCAST, SDLoc(N),
6932 NVT, N0.getOperand(0));
6935 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
6937 DAG.getConstant(Index, DL, IndexTy));
6941 // trunc (select c, a, b) -> select c, (trunc a), (trunc b)
6942 if (N0.getOpcode() == ISD::SELECT) {
6943 EVT SrcVT = N0.getValueType();
6944 if ((!LegalOperations || TLI.isOperationLegal(ISD::SELECT, SrcVT)) &&
6945 TLI.isTruncateFree(SrcVT, VT)) {
6947 SDValue Cond = N0.getOperand(0);
6948 SDValue TruncOp0 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(1));
6949 SDValue TruncOp1 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(2));
6950 return DAG.getNode(ISD::SELECT, SDLoc(N), VT, Cond, TruncOp0, TruncOp1);
6954 // Fold a series of buildvector, bitcast, and truncate if possible.
6956 // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to
6957 // (2xi32 (buildvector x, y)).
6958 if (Level == AfterLegalizeVectorOps && VT.isVector() &&
6959 N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
6960 N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
6961 N0.getOperand(0).hasOneUse()) {
6963 SDValue BuildVect = N0.getOperand(0);
6964 EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType();
6965 EVT TruncVecEltTy = VT.getVectorElementType();
6967 // Check that the element types match.
6968 if (BuildVectEltTy == TruncVecEltTy) {
6969 // Now we only need to compute the offset of the truncated elements.
6970 unsigned BuildVecNumElts = BuildVect.getNumOperands();
6971 unsigned TruncVecNumElts = VT.getVectorNumElements();
6972 unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts;
6974 assert((BuildVecNumElts % TruncVecNumElts) == 0 &&
6975 "Invalid number of elements");
6977 SmallVector<SDValue, 8> Opnds;
6978 for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset)
6979 Opnds.push_back(BuildVect.getOperand(i));
6981 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
6985 // See if we can simplify the input to this truncate through knowledge that
6986 // only the low bits are being used.
6987 // For example "trunc (or (shl x, 8), y)" // -> trunc y
6988 // Currently we only perform this optimization on scalars because vectors
6989 // may have different active low bits.
6990 if (!VT.isVector()) {
6992 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
6993 VT.getSizeInBits()));
6994 if (Shorter.getNode())
6995 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter);
6997 // fold (truncate (load x)) -> (smaller load x)
6998 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
6999 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
7000 if (SDValue Reduced = ReduceLoadWidth(N))
7003 // Handle the case where the load remains an extending load even
7004 // after truncation.
7005 if (N0.hasOneUse() && ISD::isUNINDEXEDLoad(N0.getNode())) {
7006 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7007 if (!LN0->isVolatile() &&
7008 LN0->getMemoryVT().getStoreSizeInBits() < VT.getSizeInBits()) {
7009 SDValue NewLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(LN0),
7010 VT, LN0->getChain(), LN0->getBasePtr(),
7012 LN0->getMemOperand());
7013 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLoad.getValue(1));
7018 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
7019 // where ... are all 'undef'.
7020 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
7021 SmallVector<EVT, 8> VTs;
7024 unsigned NumDefs = 0;
7026 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
7027 SDValue X = N0.getOperand(i);
7028 if (X.getOpcode() != ISD::UNDEF) {
7033 // Stop if more than one members are non-undef.
7036 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
7037 VT.getVectorElementType(),
7038 X.getValueType().getVectorNumElements()));
7042 return DAG.getUNDEF(VT);
7045 assert(V.getNode() && "The single defined operand is empty!");
7046 SmallVector<SDValue, 8> Opnds;
7047 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
7049 Opnds.push_back(DAG.getUNDEF(VTs[i]));
7052 SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V);
7053 AddToWorklist(NV.getNode());
7054 Opnds.push_back(NV);
7056 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Opnds);
7060 // Simplify the operands using demanded-bits information.
7061 if (!VT.isVector() &&
7062 SimplifyDemandedBits(SDValue(N, 0)))
7063 return SDValue(N, 0);
7068 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
7069 SDValue Elt = N->getOperand(i);
7070 if (Elt.getOpcode() != ISD::MERGE_VALUES)
7071 return Elt.getNode();
7072 return Elt.getOperand(Elt.getResNo()).getNode();
7075 /// build_pair (load, load) -> load
7076 /// if load locations are consecutive.
7077 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
7078 assert(N->getOpcode() == ISD::BUILD_PAIR);
7080 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
7081 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
7082 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
7083 LD1->getAddressSpace() != LD2->getAddressSpace())
7085 EVT LD1VT = LD1->getValueType(0);
7087 if (ISD::isNON_EXTLoad(LD2) &&
7089 // If both are volatile this would reduce the number of volatile loads.
7090 // If one is volatile it might be ok, but play conservative and bail out.
7091 !LD1->isVolatile() &&
7092 !LD2->isVolatile() &&
7093 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
7094 unsigned Align = LD1->getAlignment();
7095 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
7096 VT.getTypeForEVT(*DAG.getContext()));
7098 if (NewAlign <= Align &&
7099 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
7100 return DAG.getLoad(VT, SDLoc(N), LD1->getChain(),
7101 LD1->getBasePtr(), LD1->getPointerInfo(),
7102 false, false, false, Align);
7108 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
7109 SDValue N0 = N->getOperand(0);
7110 EVT VT = N->getValueType(0);
7112 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
7113 // Only do this before legalize, since afterward the target may be depending
7114 // on the bitconvert.
7115 // First check to see if this is all constant.
7117 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
7119 bool isSimple = cast<BuildVectorSDNode>(N0)->isConstant();
7121 EVT DestEltVT = N->getValueType(0).getVectorElementType();
7122 assert(!DestEltVT.isVector() &&
7123 "Element type of vector ValueType must not be vector!");
7125 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
7128 // If the input is a constant, let getNode fold it.
7129 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
7130 // If we can't allow illegal operations, we need to check that this is just
7131 // a fp -> int or int -> conversion and that the resulting operation will
7133 if (!LegalOperations ||
7134 (isa<ConstantSDNode>(N0) && VT.isFloatingPoint() && !VT.isVector() &&
7135 TLI.isOperationLegal(ISD::ConstantFP, VT)) ||
7136 (isa<ConstantFPSDNode>(N0) && VT.isInteger() && !VT.isVector() &&
7137 TLI.isOperationLegal(ISD::Constant, VT)))
7138 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, N0);
7141 // (conv (conv x, t1), t2) -> (conv x, t2)
7142 if (N0.getOpcode() == ISD::BITCAST)
7143 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT,
7146 // fold (conv (load x)) -> (load (conv*)x)
7147 // If the resultant load doesn't need a higher alignment than the original!
7148 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7149 // Do not change the width of a volatile load.
7150 !cast<LoadSDNode>(N0)->isVolatile() &&
7151 // Do not remove the cast if the types differ in endian layout.
7152 TLI.hasBigEndianPartOrdering(N0.getValueType(), DAG.getDataLayout()) ==
7153 TLI.hasBigEndianPartOrdering(VT, DAG.getDataLayout()) &&
7154 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)) &&
7155 TLI.isLoadBitCastBeneficial(N0.getValueType(), VT)) {
7156 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7157 unsigned Align = DAG.getDataLayout().getABITypeAlignment(
7158 VT.getTypeForEVT(*DAG.getContext()));
7159 unsigned OrigAlign = LN0->getAlignment();
7161 if (Align <= OrigAlign) {
7162 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(),
7163 LN0->getBasePtr(), LN0->getPointerInfo(),
7164 LN0->isVolatile(), LN0->isNonTemporal(),
7165 LN0->isInvariant(), OrigAlign,
7167 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
7172 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
7173 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
7174 // This often reduces constant pool loads.
7175 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) ||
7176 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) &&
7177 N0.getNode()->hasOneUse() && VT.isInteger() &&
7178 !VT.isVector() && !N0.getValueType().isVector()) {
7179 SDValue NewConv = DAG.getNode(ISD::BITCAST, SDLoc(N0), VT,
7181 AddToWorklist(NewConv.getNode());
7184 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
7185 if (N0.getOpcode() == ISD::FNEG)
7186 return DAG.getNode(ISD::XOR, DL, VT,
7187 NewConv, DAG.getConstant(SignBit, DL, VT));
7188 assert(N0.getOpcode() == ISD::FABS);
7189 return DAG.getNode(ISD::AND, DL, VT,
7190 NewConv, DAG.getConstant(~SignBit, DL, VT));
7193 // fold (bitconvert (fcopysign cst, x)) ->
7194 // (or (and (bitconvert x), sign), (and cst, (not sign)))
7195 // Note that we don't handle (copysign x, cst) because this can always be
7196 // folded to an fneg or fabs.
7197 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
7198 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
7199 VT.isInteger() && !VT.isVector()) {
7200 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
7201 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
7202 if (isTypeLegal(IntXVT)) {
7203 SDValue X = DAG.getNode(ISD::BITCAST, SDLoc(N0),
7204 IntXVT, N0.getOperand(1));
7205 AddToWorklist(X.getNode());
7207 // If X has a different width than the result/lhs, sext it or truncate it.
7208 unsigned VTWidth = VT.getSizeInBits();
7209 if (OrigXWidth < VTWidth) {
7210 X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X);
7211 AddToWorklist(X.getNode());
7212 } else if (OrigXWidth > VTWidth) {
7213 // To get the sign bit in the right place, we have to shift it right
7214 // before truncating.
7216 X = DAG.getNode(ISD::SRL, DL,
7217 X.getValueType(), X,
7218 DAG.getConstant(OrigXWidth-VTWidth, DL,
7220 AddToWorklist(X.getNode());
7221 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
7222 AddToWorklist(X.getNode());
7225 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
7226 X = DAG.getNode(ISD::AND, SDLoc(X), VT,
7227 X, DAG.getConstant(SignBit, SDLoc(X), VT));
7228 AddToWorklist(X.getNode());
7230 SDValue Cst = DAG.getNode(ISD::BITCAST, SDLoc(N0),
7231 VT, N0.getOperand(0));
7232 Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT,
7233 Cst, DAG.getConstant(~SignBit, SDLoc(Cst), VT));
7234 AddToWorklist(Cst.getNode());
7236 return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst);
7240 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
7241 if (N0.getOpcode() == ISD::BUILD_PAIR)
7242 if (SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT))
7245 // Remove double bitcasts from shuffles - this is often a legacy of
7246 // XformToShuffleWithZero being used to combine bitmaskings (of
7247 // float vectors bitcast to integer vectors) into shuffles.
7248 // bitcast(shuffle(bitcast(s0),bitcast(s1))) -> shuffle(s0,s1)
7249 if (Level < AfterLegalizeDAG && TLI.isTypeLegal(VT) && VT.isVector() &&
7250 N0->getOpcode() == ISD::VECTOR_SHUFFLE &&
7251 VT.getVectorNumElements() >= N0.getValueType().getVectorNumElements() &&
7252 !(VT.getVectorNumElements() % N0.getValueType().getVectorNumElements())) {
7253 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N0);
7255 // If operands are a bitcast, peek through if it casts the original VT.
7256 // If operands are a constant, just bitcast back to original VT.
7257 auto PeekThroughBitcast = [&](SDValue Op) {
7258 if (Op.getOpcode() == ISD::BITCAST &&
7259 Op.getOperand(0).getValueType() == VT)
7260 return SDValue(Op.getOperand(0));
7261 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) ||
7262 ISD::isBuildVectorOfConstantFPSDNodes(Op.getNode()))
7263 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
7267 SDValue SV0 = PeekThroughBitcast(N0->getOperand(0));
7268 SDValue SV1 = PeekThroughBitcast(N0->getOperand(1));
7273 VT.getVectorNumElements() / N0.getValueType().getVectorNumElements();
7274 SmallVector<int, 8> NewMask;
7275 for (int M : SVN->getMask())
7276 for (int i = 0; i != MaskScale; ++i)
7277 NewMask.push_back(M < 0 ? -1 : M * MaskScale + i);
7279 bool LegalMask = TLI.isShuffleMaskLegal(NewMask, VT);
7281 std::swap(SV0, SV1);
7282 ShuffleVectorSDNode::commuteMask(NewMask);
7283 LegalMask = TLI.isShuffleMaskLegal(NewMask, VT);
7287 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, NewMask);
7293 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
7294 EVT VT = N->getValueType(0);
7295 return CombineConsecutiveLoads(N, VT);
7298 /// We know that BV is a build_vector node with Constant, ConstantFP or Undef
7299 /// operands. DstEltVT indicates the destination element value type.
7300 SDValue DAGCombiner::
7301 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
7302 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
7304 // If this is already the right type, we're done.
7305 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
7307 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
7308 unsigned DstBitSize = DstEltVT.getSizeInBits();
7310 // If this is a conversion of N elements of one type to N elements of another
7311 // type, convert each element. This handles FP<->INT cases.
7312 if (SrcBitSize == DstBitSize) {
7313 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
7314 BV->getValueType(0).getVectorNumElements());
7316 // Due to the FP element handling below calling this routine recursively,
7317 // we can end up with a scalar-to-vector node here.
7318 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
7319 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
7320 DAG.getNode(ISD::BITCAST, SDLoc(BV),
7321 DstEltVT, BV->getOperand(0)));
7323 SmallVector<SDValue, 8> Ops;
7324 for (SDValue Op : BV->op_values()) {
7325 // If the vector element type is not legal, the BUILD_VECTOR operands
7326 // are promoted and implicitly truncated. Make that explicit here.
7327 if (Op.getValueType() != SrcEltVT)
7328 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op);
7329 Ops.push_back(DAG.getNode(ISD::BITCAST, SDLoc(BV),
7331 AddToWorklist(Ops.back().getNode());
7333 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
7336 // Otherwise, we're growing or shrinking the elements. To avoid having to
7337 // handle annoying details of growing/shrinking FP values, we convert them to
7339 if (SrcEltVT.isFloatingPoint()) {
7340 // Convert the input float vector to a int vector where the elements are the
7342 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
7343 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
7347 // Now we know the input is an integer vector. If the output is a FP type,
7348 // convert to integer first, then to FP of the right size.
7349 if (DstEltVT.isFloatingPoint()) {
7350 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
7351 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
7353 // Next, convert to FP elements of the same size.
7354 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
7359 // Okay, we know the src/dst types are both integers of differing types.
7360 // Handling growing first.
7361 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
7362 if (SrcBitSize < DstBitSize) {
7363 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
7365 SmallVector<SDValue, 8> Ops;
7366 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
7367 i += NumInputsPerOutput) {
7368 bool isLE = DAG.getDataLayout().isLittleEndian();
7369 APInt NewBits = APInt(DstBitSize, 0);
7370 bool EltIsUndef = true;
7371 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
7372 // Shift the previously computed bits over.
7373 NewBits <<= SrcBitSize;
7374 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
7375 if (Op.getOpcode() == ISD::UNDEF) continue;
7378 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
7379 zextOrTrunc(SrcBitSize).zext(DstBitSize);
7383 Ops.push_back(DAG.getUNDEF(DstEltVT));
7385 Ops.push_back(DAG.getConstant(NewBits, DL, DstEltVT));
7388 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
7389 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Ops);
7392 // Finally, this must be the case where we are shrinking elements: each input
7393 // turns into multiple outputs.
7394 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
7395 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
7396 NumOutputsPerInput*BV->getNumOperands());
7397 SmallVector<SDValue, 8> Ops;
7399 for (const SDValue &Op : BV->op_values()) {
7400 if (Op.getOpcode() == ISD::UNDEF) {
7401 Ops.append(NumOutputsPerInput, DAG.getUNDEF(DstEltVT));
7405 APInt OpVal = cast<ConstantSDNode>(Op)->
7406 getAPIntValue().zextOrTrunc(SrcBitSize);
7408 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
7409 APInt ThisVal = OpVal.trunc(DstBitSize);
7410 Ops.push_back(DAG.getConstant(ThisVal, DL, DstEltVT));
7411 OpVal = OpVal.lshr(DstBitSize);
7414 // For big endian targets, swap the order of the pieces of each element.
7415 if (DAG.getDataLayout().isBigEndian())
7416 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
7419 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Ops);
7422 /// Try to perform FMA combining on a given FADD node.
7423 SDValue DAGCombiner::visitFADDForFMACombine(SDNode *N) {
7424 SDValue N0 = N->getOperand(0);
7425 SDValue N1 = N->getOperand(1);
7426 EVT VT = N->getValueType(0);
7429 const TargetOptions &Options = DAG.getTarget().Options;
7430 bool UnsafeFPMath = (Options.AllowFPOpFusion == FPOpFusion::Fast ||
7431 Options.UnsafeFPMath);
7433 // Floating-point multiply-add with intermediate rounding.
7434 bool HasFMAD = (LegalOperations &&
7435 TLI.isOperationLegal(ISD::FMAD, VT));
7437 // Floating-point multiply-add without intermediate rounding.
7438 bool HasFMA = ((!LegalOperations ||
7439 TLI.isOperationLegalOrCustom(ISD::FMA, VT)) &&
7440 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
7443 // No valid opcode, do not combine.
7444 if (!HasFMAD && !HasFMA)
7447 // Always prefer FMAD to FMA for precision.
7448 unsigned int PreferredFusedOpcode = HasFMAD ? ISD::FMAD : ISD::FMA;
7449 bool Aggressive = TLI.enableAggressiveFMAFusion(VT);
7450 bool LookThroughFPExt = TLI.isFPExtFree(VT);
7452 // If we have two choices trying to fold (fadd (fmul u, v), (fmul x, y)),
7453 // prefer to fold the multiply with fewer uses.
7454 if (Aggressive && N0.getOpcode() == ISD::FMUL &&
7455 N1.getOpcode() == ISD::FMUL) {
7456 if (N0.getNode()->use_size() > N1.getNode()->use_size())
7460 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
7461 if (N0.getOpcode() == ISD::FMUL &&
7462 (Aggressive || N0->hasOneUse())) {
7463 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7464 N0.getOperand(0), N0.getOperand(1), N1);
7467 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
7468 // Note: Commutes FADD operands.
7469 if (N1.getOpcode() == ISD::FMUL &&
7470 (Aggressive || N1->hasOneUse())) {
7471 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7472 N1.getOperand(0), N1.getOperand(1), N0);
7475 // Look through FP_EXTEND nodes to do more combining.
7476 if (UnsafeFPMath && LookThroughFPExt) {
7477 // fold (fadd (fpext (fmul x, y)), z) -> (fma (fpext x), (fpext y), z)
7478 if (N0.getOpcode() == ISD::FP_EXTEND) {
7479 SDValue N00 = N0.getOperand(0);
7480 if (N00.getOpcode() == ISD::FMUL)
7481 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7482 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7484 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7485 N00.getOperand(1)), N1);
7488 // fold (fadd x, (fpext (fmul y, z))) -> (fma (fpext y), (fpext z), x)
7489 // Note: Commutes FADD operands.
7490 if (N1.getOpcode() == ISD::FP_EXTEND) {
7491 SDValue N10 = N1.getOperand(0);
7492 if (N10.getOpcode() == ISD::FMUL)
7493 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7494 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7496 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7497 N10.getOperand(1)), N0);
7501 // More folding opportunities when target permits.
7502 if ((UnsafeFPMath || HasFMAD) && Aggressive) {
7503 // fold (fadd (fma x, y, (fmul u, v)), z) -> (fma x, y (fma u, v, z))
7504 if (N0.getOpcode() == PreferredFusedOpcode &&
7505 N0.getOperand(2).getOpcode() == ISD::FMUL) {
7506 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7507 N0.getOperand(0), N0.getOperand(1),
7508 DAG.getNode(PreferredFusedOpcode, SL, VT,
7509 N0.getOperand(2).getOperand(0),
7510 N0.getOperand(2).getOperand(1),
7514 // fold (fadd x, (fma y, z, (fmul u, v)) -> (fma y, z (fma u, v, x))
7515 if (N1->getOpcode() == PreferredFusedOpcode &&
7516 N1.getOperand(2).getOpcode() == ISD::FMUL) {
7517 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7518 N1.getOperand(0), N1.getOperand(1),
7519 DAG.getNode(PreferredFusedOpcode, SL, VT,
7520 N1.getOperand(2).getOperand(0),
7521 N1.getOperand(2).getOperand(1),
7525 if (UnsafeFPMath && LookThroughFPExt) {
7526 // fold (fadd (fma x, y, (fpext (fmul u, v))), z)
7527 // -> (fma x, y, (fma (fpext u), (fpext v), z))
7528 auto FoldFAddFMAFPExtFMul = [&] (
7529 SDValue X, SDValue Y, SDValue U, SDValue V, SDValue Z) {
7530 return DAG.getNode(PreferredFusedOpcode, SL, VT, X, Y,
7531 DAG.getNode(PreferredFusedOpcode, SL, VT,
7532 DAG.getNode(ISD::FP_EXTEND, SL, VT, U),
7533 DAG.getNode(ISD::FP_EXTEND, SL, VT, V),
7536 if (N0.getOpcode() == PreferredFusedOpcode) {
7537 SDValue N02 = N0.getOperand(2);
7538 if (N02.getOpcode() == ISD::FP_EXTEND) {
7539 SDValue N020 = N02.getOperand(0);
7540 if (N020.getOpcode() == ISD::FMUL)
7541 return FoldFAddFMAFPExtFMul(N0.getOperand(0), N0.getOperand(1),
7542 N020.getOperand(0), N020.getOperand(1),
7547 // fold (fadd (fpext (fma x, y, (fmul u, v))), z)
7548 // -> (fma (fpext x), (fpext y), (fma (fpext u), (fpext v), z))
7549 // FIXME: This turns two single-precision and one double-precision
7550 // operation into two double-precision operations, which might not be
7551 // interesting for all targets, especially GPUs.
7552 auto FoldFAddFPExtFMAFMul = [&] (
7553 SDValue X, SDValue Y, SDValue U, SDValue V, SDValue Z) {
7554 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7555 DAG.getNode(ISD::FP_EXTEND, SL, VT, X),
7556 DAG.getNode(ISD::FP_EXTEND, SL, VT, Y),
7557 DAG.getNode(PreferredFusedOpcode, SL, VT,
7558 DAG.getNode(ISD::FP_EXTEND, SL, VT, U),
7559 DAG.getNode(ISD::FP_EXTEND, SL, VT, V),
7562 if (N0.getOpcode() == ISD::FP_EXTEND) {
7563 SDValue N00 = N0.getOperand(0);
7564 if (N00.getOpcode() == PreferredFusedOpcode) {
7565 SDValue N002 = N00.getOperand(2);
7566 if (N002.getOpcode() == ISD::FMUL)
7567 return FoldFAddFPExtFMAFMul(N00.getOperand(0), N00.getOperand(1),
7568 N002.getOperand(0), N002.getOperand(1),
7573 // fold (fadd x, (fma y, z, (fpext (fmul u, v)))
7574 // -> (fma y, z, (fma (fpext u), (fpext v), x))
7575 if (N1.getOpcode() == PreferredFusedOpcode) {
7576 SDValue N12 = N1.getOperand(2);
7577 if (N12.getOpcode() == ISD::FP_EXTEND) {
7578 SDValue N120 = N12.getOperand(0);
7579 if (N120.getOpcode() == ISD::FMUL)
7580 return FoldFAddFMAFPExtFMul(N1.getOperand(0), N1.getOperand(1),
7581 N120.getOperand(0), N120.getOperand(1),
7586 // fold (fadd x, (fpext (fma y, z, (fmul u, v)))
7587 // -> (fma (fpext y), (fpext z), (fma (fpext u), (fpext v), x))
7588 // FIXME: This turns two single-precision and one double-precision
7589 // operation into two double-precision operations, which might not be
7590 // interesting for all targets, especially GPUs.
7591 if (N1.getOpcode() == ISD::FP_EXTEND) {
7592 SDValue N10 = N1.getOperand(0);
7593 if (N10.getOpcode() == PreferredFusedOpcode) {
7594 SDValue N102 = N10.getOperand(2);
7595 if (N102.getOpcode() == ISD::FMUL)
7596 return FoldFAddFPExtFMAFMul(N10.getOperand(0), N10.getOperand(1),
7597 N102.getOperand(0), N102.getOperand(1),
7607 /// Try to perform FMA combining on a given FSUB node.
7608 SDValue DAGCombiner::visitFSUBForFMACombine(SDNode *N) {
7609 SDValue N0 = N->getOperand(0);
7610 SDValue N1 = N->getOperand(1);
7611 EVT VT = N->getValueType(0);
7614 const TargetOptions &Options = DAG.getTarget().Options;
7615 bool UnsafeFPMath = (Options.AllowFPOpFusion == FPOpFusion::Fast ||
7616 Options.UnsafeFPMath);
7618 // Floating-point multiply-add with intermediate rounding.
7619 bool HasFMAD = (LegalOperations &&
7620 TLI.isOperationLegal(ISD::FMAD, VT));
7622 // Floating-point multiply-add without intermediate rounding.
7623 bool HasFMA = ((!LegalOperations ||
7624 TLI.isOperationLegalOrCustom(ISD::FMA, VT)) &&
7625 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
7628 // No valid opcode, do not combine.
7629 if (!HasFMAD && !HasFMA)
7632 // Always prefer FMAD to FMA for precision.
7633 unsigned int PreferredFusedOpcode = HasFMAD ? ISD::FMAD : ISD::FMA;
7634 bool Aggressive = TLI.enableAggressiveFMAFusion(VT);
7635 bool LookThroughFPExt = TLI.isFPExtFree(VT);
7637 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
7638 if (N0.getOpcode() == ISD::FMUL &&
7639 (Aggressive || N0->hasOneUse())) {
7640 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7641 N0.getOperand(0), N0.getOperand(1),
7642 DAG.getNode(ISD::FNEG, SL, VT, N1));
7645 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
7646 // Note: Commutes FSUB operands.
7647 if (N1.getOpcode() == ISD::FMUL &&
7648 (Aggressive || N1->hasOneUse()))
7649 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7650 DAG.getNode(ISD::FNEG, SL, VT,
7652 N1.getOperand(1), N0);
7654 // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
7655 if (N0.getOpcode() == ISD::FNEG &&
7656 N0.getOperand(0).getOpcode() == ISD::FMUL &&
7657 (Aggressive || (N0->hasOneUse() && N0.getOperand(0).hasOneUse()))) {
7658 SDValue N00 = N0.getOperand(0).getOperand(0);
7659 SDValue N01 = N0.getOperand(0).getOperand(1);
7660 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7661 DAG.getNode(ISD::FNEG, SL, VT, N00), N01,
7662 DAG.getNode(ISD::FNEG, SL, VT, N1));
7665 // Look through FP_EXTEND nodes to do more combining.
7666 if (UnsafeFPMath && LookThroughFPExt) {
7667 // fold (fsub (fpext (fmul x, y)), z)
7668 // -> (fma (fpext x), (fpext y), (fneg z))
7669 if (N0.getOpcode() == ISD::FP_EXTEND) {
7670 SDValue N00 = N0.getOperand(0);
7671 if (N00.getOpcode() == ISD::FMUL)
7672 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7673 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7675 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7677 DAG.getNode(ISD::FNEG, SL, VT, N1));
7680 // fold (fsub x, (fpext (fmul y, z)))
7681 // -> (fma (fneg (fpext y)), (fpext z), x)
7682 // Note: Commutes FSUB operands.
7683 if (N1.getOpcode() == ISD::FP_EXTEND) {
7684 SDValue N10 = N1.getOperand(0);
7685 if (N10.getOpcode() == ISD::FMUL)
7686 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7687 DAG.getNode(ISD::FNEG, SL, VT,
7688 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7689 N10.getOperand(0))),
7690 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7695 // fold (fsub (fpext (fneg (fmul, x, y))), z)
7696 // -> (fneg (fma (fpext x), (fpext y), z))
7697 // Note: This could be removed with appropriate canonicalization of the
7698 // input expression into (fneg (fadd (fpext (fmul, x, y)), z). However, the
7699 // orthogonal flags -fp-contract=fast and -enable-unsafe-fp-math prevent
7700 // from implementing the canonicalization in visitFSUB.
7701 if (N0.getOpcode() == ISD::FP_EXTEND) {
7702 SDValue N00 = N0.getOperand(0);
7703 if (N00.getOpcode() == ISD::FNEG) {
7704 SDValue N000 = N00.getOperand(0);
7705 if (N000.getOpcode() == ISD::FMUL) {
7706 return DAG.getNode(ISD::FNEG, SL, VT,
7707 DAG.getNode(PreferredFusedOpcode, SL, VT,
7708 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7709 N000.getOperand(0)),
7710 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7711 N000.getOperand(1)),
7717 // fold (fsub (fneg (fpext (fmul, x, y))), z)
7718 // -> (fneg (fma (fpext x)), (fpext y), z)
7719 // Note: This could be removed with appropriate canonicalization of the
7720 // input expression into (fneg (fadd (fpext (fmul, x, y)), z). However, the
7721 // orthogonal flags -fp-contract=fast and -enable-unsafe-fp-math prevent
7722 // from implementing the canonicalization in visitFSUB.
7723 if (N0.getOpcode() == ISD::FNEG) {
7724 SDValue N00 = N0.getOperand(0);
7725 if (N00.getOpcode() == ISD::FP_EXTEND) {
7726 SDValue N000 = N00.getOperand(0);
7727 if (N000.getOpcode() == ISD::FMUL) {
7728 return DAG.getNode(ISD::FNEG, SL, VT,
7729 DAG.getNode(PreferredFusedOpcode, SL, VT,
7730 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7731 N000.getOperand(0)),
7732 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7733 N000.getOperand(1)),
7741 // More folding opportunities when target permits.
7742 if ((UnsafeFPMath || HasFMAD) && Aggressive) {
7743 // fold (fsub (fma x, y, (fmul u, v)), z)
7744 // -> (fma x, y (fma u, v, (fneg z)))
7745 if (N0.getOpcode() == PreferredFusedOpcode &&
7746 N0.getOperand(2).getOpcode() == ISD::FMUL) {
7747 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7748 N0.getOperand(0), N0.getOperand(1),
7749 DAG.getNode(PreferredFusedOpcode, SL, VT,
7750 N0.getOperand(2).getOperand(0),
7751 N0.getOperand(2).getOperand(1),
7752 DAG.getNode(ISD::FNEG, SL, VT,
7756 // fold (fsub x, (fma y, z, (fmul u, v)))
7757 // -> (fma (fneg y), z, (fma (fneg u), v, x))
7758 if (N1.getOpcode() == PreferredFusedOpcode &&
7759 N1.getOperand(2).getOpcode() == ISD::FMUL) {
7760 SDValue N20 = N1.getOperand(2).getOperand(0);
7761 SDValue N21 = N1.getOperand(2).getOperand(1);
7762 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7763 DAG.getNode(ISD::FNEG, SL, VT,
7766 DAG.getNode(PreferredFusedOpcode, SL, VT,
7767 DAG.getNode(ISD::FNEG, SL, VT, N20),
7772 if (UnsafeFPMath && LookThroughFPExt) {
7773 // fold (fsub (fma x, y, (fpext (fmul u, v))), z)
7774 // -> (fma x, y (fma (fpext u), (fpext v), (fneg z)))
7775 if (N0.getOpcode() == PreferredFusedOpcode) {
7776 SDValue N02 = N0.getOperand(2);
7777 if (N02.getOpcode() == ISD::FP_EXTEND) {
7778 SDValue N020 = N02.getOperand(0);
7779 if (N020.getOpcode() == ISD::FMUL)
7780 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7781 N0.getOperand(0), N0.getOperand(1),
7782 DAG.getNode(PreferredFusedOpcode, SL, VT,
7783 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7784 N020.getOperand(0)),
7785 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7786 N020.getOperand(1)),
7787 DAG.getNode(ISD::FNEG, SL, VT,
7792 // fold (fsub (fpext (fma x, y, (fmul u, v))), z)
7793 // -> (fma (fpext x), (fpext y),
7794 // (fma (fpext u), (fpext v), (fneg z)))
7795 // FIXME: This turns two single-precision and one double-precision
7796 // operation into two double-precision operations, which might not be
7797 // interesting for all targets, especially GPUs.
7798 if (N0.getOpcode() == ISD::FP_EXTEND) {
7799 SDValue N00 = N0.getOperand(0);
7800 if (N00.getOpcode() == PreferredFusedOpcode) {
7801 SDValue N002 = N00.getOperand(2);
7802 if (N002.getOpcode() == ISD::FMUL)
7803 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7804 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7806 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7808 DAG.getNode(PreferredFusedOpcode, SL, VT,
7809 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7810 N002.getOperand(0)),
7811 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7812 N002.getOperand(1)),
7813 DAG.getNode(ISD::FNEG, SL, VT,
7818 // fold (fsub x, (fma y, z, (fpext (fmul u, v))))
7819 // -> (fma (fneg y), z, (fma (fneg (fpext u)), (fpext v), x))
7820 if (N1.getOpcode() == PreferredFusedOpcode &&
7821 N1.getOperand(2).getOpcode() == ISD::FP_EXTEND) {
7822 SDValue N120 = N1.getOperand(2).getOperand(0);
7823 if (N120.getOpcode() == ISD::FMUL) {
7824 SDValue N1200 = N120.getOperand(0);
7825 SDValue N1201 = N120.getOperand(1);
7826 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7827 DAG.getNode(ISD::FNEG, SL, VT, N1.getOperand(0)),
7829 DAG.getNode(PreferredFusedOpcode, SL, VT,
7830 DAG.getNode(ISD::FNEG, SL, VT,
7831 DAG.getNode(ISD::FP_EXTEND, SL,
7833 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7839 // fold (fsub x, (fpext (fma y, z, (fmul u, v))))
7840 // -> (fma (fneg (fpext y)), (fpext z),
7841 // (fma (fneg (fpext u)), (fpext v), x))
7842 // FIXME: This turns two single-precision and one double-precision
7843 // operation into two double-precision operations, which might not be
7844 // interesting for all targets, especially GPUs.
7845 if (N1.getOpcode() == ISD::FP_EXTEND &&
7846 N1.getOperand(0).getOpcode() == PreferredFusedOpcode) {
7847 SDValue N100 = N1.getOperand(0).getOperand(0);
7848 SDValue N101 = N1.getOperand(0).getOperand(1);
7849 SDValue N102 = N1.getOperand(0).getOperand(2);
7850 if (N102.getOpcode() == ISD::FMUL) {
7851 SDValue N1020 = N102.getOperand(0);
7852 SDValue N1021 = N102.getOperand(1);
7853 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7854 DAG.getNode(ISD::FNEG, SL, VT,
7855 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7857 DAG.getNode(ISD::FP_EXTEND, SL, VT, N101),
7858 DAG.getNode(PreferredFusedOpcode, SL, VT,
7859 DAG.getNode(ISD::FNEG, SL, VT,
7860 DAG.getNode(ISD::FP_EXTEND, SL,
7862 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7873 SDValue DAGCombiner::visitFADD(SDNode *N) {
7874 SDValue N0 = N->getOperand(0);
7875 SDValue N1 = N->getOperand(1);
7876 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7877 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7878 EVT VT = N->getValueType(0);
7880 const TargetOptions &Options = DAG.getTarget().Options;
7884 if (SDValue FoldedVOp = SimplifyVBinOp(N))
7887 // fold (fadd c1, c2) -> c1 + c2
7889 return DAG.getNode(ISD::FADD, DL, VT, N0, N1);
7891 // canonicalize constant to RHS
7892 if (N0CFP && !N1CFP)
7893 return DAG.getNode(ISD::FADD, DL, VT, N1, N0);
7895 // fold (fadd A, (fneg B)) -> (fsub A, B)
7896 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
7897 isNegatibleForFree(N1, LegalOperations, TLI, &Options) == 2)
7898 return DAG.getNode(ISD::FSUB, DL, VT, N0,
7899 GetNegatedExpression(N1, DAG, LegalOperations));
7901 // fold (fadd (fneg A), B) -> (fsub B, A)
7902 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
7903 isNegatibleForFree(N0, LegalOperations, TLI, &Options) == 2)
7904 return DAG.getNode(ISD::FSUB, DL, VT, N1,
7905 GetNegatedExpression(N0, DAG, LegalOperations));
7907 // If 'unsafe math' is enabled, fold lots of things.
7908 if (Options.UnsafeFPMath) {
7909 // No FP constant should be created after legalization as Instruction
7910 // Selection pass has a hard time dealing with FP constants.
7911 bool AllowNewConst = (Level < AfterLegalizeDAG);
7913 // fold (fadd A, 0) -> A
7914 if (N1CFP && N1CFP->isZero())
7917 // fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
7918 if (N1CFP && N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
7919 isa<ConstantFPSDNode>(N0.getOperand(1)))
7920 return DAG.getNode(ISD::FADD, DL, VT, N0.getOperand(0),
7921 DAG.getNode(ISD::FADD, DL, VT, N0.getOperand(1), N1));
7923 // If allowed, fold (fadd (fneg x), x) -> 0.0
7924 if (AllowNewConst && N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1)
7925 return DAG.getConstantFP(0.0, DL, VT);
7927 // If allowed, fold (fadd x, (fneg x)) -> 0.0
7928 if (AllowNewConst && N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0)
7929 return DAG.getConstantFP(0.0, DL, VT);
7931 // We can fold chains of FADD's of the same value into multiplications.
7932 // This transform is not safe in general because we are reducing the number
7933 // of rounding steps.
7934 if (TLI.isOperationLegalOrCustom(ISD::FMUL, VT) && !N0CFP && !N1CFP) {
7935 if (N0.getOpcode() == ISD::FMUL) {
7936 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
7937 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
7939 // (fadd (fmul x, c), x) -> (fmul x, c+1)
7940 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
7941 SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, SDValue(CFP01, 0),
7942 DAG.getConstantFP(1.0, DL, VT));
7943 return DAG.getNode(ISD::FMUL, DL, VT, N1, NewCFP);
7946 // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2)
7947 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
7948 N1.getOperand(0) == N1.getOperand(1) &&
7949 N0.getOperand(0) == N1.getOperand(0)) {
7950 SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, SDValue(CFP01, 0),
7951 DAG.getConstantFP(2.0, DL, VT));
7952 return DAG.getNode(ISD::FMUL, DL, VT, N0.getOperand(0), NewCFP);
7956 if (N1.getOpcode() == ISD::FMUL) {
7957 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
7958 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
7960 // (fadd x, (fmul x, c)) -> (fmul x, c+1)
7961 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
7962 SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, SDValue(CFP11, 0),
7963 DAG.getConstantFP(1.0, DL, VT));
7964 return DAG.getNode(ISD::FMUL, DL, VT, N0, NewCFP);
7967 // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2)
7968 if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD &&
7969 N0.getOperand(0) == N0.getOperand(1) &&
7970 N1.getOperand(0) == N0.getOperand(0)) {
7971 SDValue NewCFP = DAG.getNode(ISD::FADD, DL, VT, SDValue(CFP11, 0),
7972 DAG.getConstantFP(2.0, DL, VT));
7973 return DAG.getNode(ISD::FMUL, DL, VT, N1.getOperand(0), NewCFP);
7977 if (N0.getOpcode() == ISD::FADD && AllowNewConst) {
7978 ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
7979 // (fadd (fadd x, x), x) -> (fmul x, 3.0)
7980 if (!CFP && N0.getOperand(0) == N0.getOperand(1) &&
7981 (N0.getOperand(0) == N1)) {
7982 return DAG.getNode(ISD::FMUL, DL, VT,
7983 N1, DAG.getConstantFP(3.0, DL, VT));
7987 if (N1.getOpcode() == ISD::FADD && AllowNewConst) {
7988 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
7989 // (fadd x, (fadd x, x)) -> (fmul x, 3.0)
7990 if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) &&
7991 N1.getOperand(0) == N0) {
7992 return DAG.getNode(ISD::FMUL, DL, VT,
7993 N0, DAG.getConstantFP(3.0, DL, VT));
7997 // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0)
7998 if (AllowNewConst &&
7999 N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
8000 N0.getOperand(0) == N0.getOperand(1) &&
8001 N1.getOperand(0) == N1.getOperand(1) &&
8002 N0.getOperand(0) == N1.getOperand(0)) {
8003 return DAG.getNode(ISD::FMUL, DL, VT,
8004 N0.getOperand(0), DAG.getConstantFP(4.0, DL, VT));
8007 } // enable-unsafe-fp-math
8009 // FADD -> FMA combines:
8010 if (SDValue Fused = visitFADDForFMACombine(N)) {
8011 AddToWorklist(Fused.getNode());
8018 SDValue DAGCombiner::visitFSUB(SDNode *N) {
8019 SDValue N0 = N->getOperand(0);
8020 SDValue N1 = N->getOperand(1);
8021 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
8022 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
8023 EVT VT = N->getValueType(0);
8025 const TargetOptions &Options = DAG.getTarget().Options;
8029 if (SDValue FoldedVOp = SimplifyVBinOp(N))
8032 // fold (fsub c1, c2) -> c1-c2
8034 return DAG.getNode(ISD::FSUB, dl, VT, N0, N1);
8036 // fold (fsub A, (fneg B)) -> (fadd A, B)
8037 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
8038 return DAG.getNode(ISD::FADD, dl, VT, N0,
8039 GetNegatedExpression(N1, DAG, LegalOperations));
8041 // If 'unsafe math' is enabled, fold lots of things.
8042 if (Options.UnsafeFPMath) {
8044 if (N1CFP && N1CFP->isZero())
8047 // (fsub 0, B) -> -B
8048 if (N0CFP && N0CFP->isZero()) {
8049 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
8050 return GetNegatedExpression(N1, DAG, LegalOperations);
8051 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
8052 return DAG.getNode(ISD::FNEG, dl, VT, N1);
8055 // (fsub x, x) -> 0.0
8057 return DAG.getConstantFP(0.0f, dl, VT);
8059 // (fsub x, (fadd x, y)) -> (fneg y)
8060 // (fsub x, (fadd y, x)) -> (fneg y)
8061 if (N1.getOpcode() == ISD::FADD) {
8062 SDValue N10 = N1->getOperand(0);
8063 SDValue N11 = N1->getOperand(1);
8065 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI, &Options))
8066 return GetNegatedExpression(N11, DAG, LegalOperations);
8068 if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI, &Options))
8069 return GetNegatedExpression(N10, DAG, LegalOperations);
8073 // FSUB -> FMA combines:
8074 if (SDValue Fused = visitFSUBForFMACombine(N)) {
8075 AddToWorklist(Fused.getNode());
8082 SDValue DAGCombiner::visitFMUL(SDNode *N) {
8083 SDValue N0 = N->getOperand(0);
8084 SDValue N1 = N->getOperand(1);
8085 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
8086 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
8087 EVT VT = N->getValueType(0);
8089 const TargetOptions &Options = DAG.getTarget().Options;
8092 if (VT.isVector()) {
8093 // This just handles C1 * C2 for vectors. Other vector folds are below.
8094 if (SDValue FoldedVOp = SimplifyVBinOp(N))
8098 // fold (fmul c1, c2) -> c1*c2
8100 return DAG.getNode(ISD::FMUL, DL, VT, N0, N1);
8102 // canonicalize constant to RHS
8103 if (isConstantFPBuildVectorOrConstantFP(N0) &&
8104 !isConstantFPBuildVectorOrConstantFP(N1))
8105 return DAG.getNode(ISD::FMUL, DL, VT, N1, N0);
8107 // fold (fmul A, 1.0) -> A
8108 if (N1CFP && N1CFP->isExactlyValue(1.0))
8111 if (Options.UnsafeFPMath) {
8112 // fold (fmul A, 0) -> 0
8113 if (N1CFP && N1CFP->isZero())
8116 // fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
8117 if (N0.getOpcode() == ISD::FMUL) {
8118 // Fold scalars or any vector constants (not just splats).
8119 // This fold is done in general by InstCombine, but extra fmul insts
8120 // may have been generated during lowering.
8121 SDValue N00 = N0.getOperand(0);
8122 SDValue N01 = N0.getOperand(1);
8123 auto *BV1 = dyn_cast<BuildVectorSDNode>(N1);
8124 auto *BV00 = dyn_cast<BuildVectorSDNode>(N00);
8125 auto *BV01 = dyn_cast<BuildVectorSDNode>(N01);
8127 // Check 1: Make sure that the first operand of the inner multiply is NOT
8128 // a constant. Otherwise, we may induce infinite looping.
8129 if (!(isConstOrConstSplatFP(N00) || (BV00 && BV00->isConstant()))) {
8130 // Check 2: Make sure that the second operand of the inner multiply and
8131 // the second operand of the outer multiply are constants.
8132 if ((N1CFP && isConstOrConstSplatFP(N01)) ||
8133 (BV1 && BV01 && BV1->isConstant() && BV01->isConstant())) {
8134 SDValue MulConsts = DAG.getNode(ISD::FMUL, DL, VT, N01, N1);
8135 return DAG.getNode(ISD::FMUL, DL, VT, N00, MulConsts);
8140 // fold (fmul (fadd x, x), c) -> (fmul x, (fmul 2.0, c))
8141 // Undo the fmul 2.0, x -> fadd x, x transformation, since if it occurs
8142 // during an early run of DAGCombiner can prevent folding with fmuls
8143 // inserted during lowering.
8144 if (N0.getOpcode() == ISD::FADD &&
8145 (N0.getOperand(0) == N0.getOperand(1)) &&
8147 const SDValue Two = DAG.getConstantFP(2.0, DL, VT);
8148 SDValue MulConsts = DAG.getNode(ISD::FMUL, DL, VT, Two, N1);
8149 return DAG.getNode(ISD::FMUL, DL, VT, N0.getOperand(0), MulConsts);
8153 // fold (fmul X, 2.0) -> (fadd X, X)
8154 if (N1CFP && N1CFP->isExactlyValue(+2.0))
8155 return DAG.getNode(ISD::FADD, DL, VT, N0, N0);
8157 // fold (fmul X, -1.0) -> (fneg X)
8158 if (N1CFP && N1CFP->isExactlyValue(-1.0))
8159 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
8160 return DAG.getNode(ISD::FNEG, DL, VT, N0);
8162 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
8163 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
8164 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
8165 // Both can be negated for free, check to see if at least one is cheaper
8167 if (LHSNeg == 2 || RHSNeg == 2)
8168 return DAG.getNode(ISD::FMUL, DL, VT,
8169 GetNegatedExpression(N0, DAG, LegalOperations),
8170 GetNegatedExpression(N1, DAG, LegalOperations));
8177 SDValue DAGCombiner::visitFMA(SDNode *N) {
8178 SDValue N0 = N->getOperand(0);
8179 SDValue N1 = N->getOperand(1);
8180 SDValue N2 = N->getOperand(2);
8181 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8182 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8183 EVT VT = N->getValueType(0);
8185 const TargetOptions &Options = DAG.getTarget().Options;
8187 // Constant fold FMA.
8188 if (isa<ConstantFPSDNode>(N0) &&
8189 isa<ConstantFPSDNode>(N1) &&
8190 isa<ConstantFPSDNode>(N2)) {
8191 return DAG.getNode(ISD::FMA, dl, VT, N0, N1, N2);
8194 if (Options.UnsafeFPMath) {
8195 if (N0CFP && N0CFP->isZero())
8197 if (N1CFP && N1CFP->isZero())
8200 if (N0CFP && N0CFP->isExactlyValue(1.0))
8201 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2);
8202 if (N1CFP && N1CFP->isExactlyValue(1.0))
8203 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2);
8205 // Canonicalize (fma c, x, y) -> (fma x, c, y)
8206 if (N0CFP && !N1CFP)
8207 return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2);
8209 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
8210 if (Options.UnsafeFPMath && N1CFP &&
8211 N2.getOpcode() == ISD::FMUL &&
8212 N0 == N2.getOperand(0) &&
8213 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
8214 return DAG.getNode(ISD::FMUL, dl, VT, N0,
8215 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
8219 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
8220 if (Options.UnsafeFPMath &&
8221 N0.getOpcode() == ISD::FMUL && N1CFP &&
8222 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
8223 return DAG.getNode(ISD::FMA, dl, VT,
8225 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
8229 // (fma x, 1, y) -> (fadd x, y)
8230 // (fma x, -1, y) -> (fadd (fneg x), y)
8232 if (N1CFP->isExactlyValue(1.0))
8233 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
8235 if (N1CFP->isExactlyValue(-1.0) &&
8236 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
8237 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
8238 AddToWorklist(RHSNeg.getNode());
8239 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
8243 // (fma x, c, x) -> (fmul x, (c+1))
8244 if (Options.UnsafeFPMath && N1CFP && N0 == N2)
8245 return DAG.getNode(ISD::FMUL, dl, VT, N0,
8246 DAG.getNode(ISD::FADD, dl, VT,
8247 N1, DAG.getConstantFP(1.0, dl, VT)));
8249 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
8250 if (Options.UnsafeFPMath && N1CFP &&
8251 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0)
8252 return DAG.getNode(ISD::FMUL, dl, VT, N0,
8253 DAG.getNode(ISD::FADD, dl, VT,
8254 N1, DAG.getConstantFP(-1.0, dl, VT)));
8260 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
8262 // E.g., (a / D; b / D;) -> (recip = 1.0 / D; a * recip; b * recip)
8263 // Notice that this is not always beneficial. One reason is different target
8264 // may have different costs for FDIV and FMUL, so sometimes the cost of two
8265 // FDIVs may be lower than the cost of one FDIV and two FMULs. Another reason
8266 // is the critical path is increased from "one FDIV" to "one FDIV + one FMUL".
8267 SDValue DAGCombiner::combineRepeatedFPDivisors(SDNode *N) {
8268 if (!DAG.getTarget().Options.UnsafeFPMath)
8271 // Skip if current node is a reciprocal.
8272 SDValue N0 = N->getOperand(0);
8273 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8274 if (N0CFP && N0CFP->isExactlyValue(1.0))
8277 // Exit early if the target does not want this transform or if there can't
8278 // possibly be enough uses of the divisor to make the transform worthwhile.
8279 SDValue N1 = N->getOperand(1);
8280 unsigned MinUses = TLI.combineRepeatedFPDivisors();
8281 if (!MinUses || N1->use_size() < MinUses)
8284 // Find all FDIV users of the same divisor.
8285 // Use a set because duplicates may be present in the user list.
8286 SetVector<SDNode *> Users;
8287 for (auto *U : N1->uses())
8288 if (U->getOpcode() == ISD::FDIV && U->getOperand(1) == N1)
8291 // Now that we have the actual number of divisor uses, make sure it meets
8292 // the minimum threshold specified by the target.
8293 if (Users.size() < MinUses)
8296 EVT VT = N->getValueType(0);
8298 SDValue FPOne = DAG.getConstantFP(1.0, DL, VT);
8299 // FIXME: This optimization requires some level of fast-math, so the
8300 // created reciprocal node should at least have the 'allowReciprocal'
8301 // fast-math-flag set.
8302 SDValue Reciprocal = DAG.getNode(ISD::FDIV, DL, VT, FPOne, N1);
8304 // Dividend / Divisor -> Dividend * Reciprocal
8305 for (auto *U : Users) {
8306 SDValue Dividend = U->getOperand(0);
8307 if (Dividend != FPOne) {
8308 SDValue NewNode = DAG.getNode(ISD::FMUL, SDLoc(U), VT, Dividend,
8310 CombineTo(U, NewNode);
8311 } else if (U != Reciprocal.getNode()) {
8312 // In the absence of fast-math-flags, this user node is always the
8313 // same node as Reciprocal, but with FMF they may be different nodes.
8314 CombineTo(U, Reciprocal);
8317 return SDValue(N, 0); // N was replaced.
8320 SDValue DAGCombiner::visitFDIV(SDNode *N) {
8321 SDValue N0 = N->getOperand(0);
8322 SDValue N1 = N->getOperand(1);
8323 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8324 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8325 EVT VT = N->getValueType(0);
8327 const TargetOptions &Options = DAG.getTarget().Options;
8331 if (SDValue FoldedVOp = SimplifyVBinOp(N))
8334 // fold (fdiv c1, c2) -> c1/c2
8336 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1);
8338 if (Options.UnsafeFPMath) {
8339 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
8341 // Compute the reciprocal 1.0 / c2.
8342 APFloat N1APF = N1CFP->getValueAPF();
8343 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
8344 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
8345 // Only do the transform if the reciprocal is a legal fp immediate that
8346 // isn't too nasty (eg NaN, denormal, ...).
8347 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
8348 (!LegalOperations ||
8349 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
8350 // backend)... we should handle this gracefully after Legalize.
8351 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
8352 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
8353 TLI.isFPImmLegal(Recip, VT)))
8354 return DAG.getNode(ISD::FMUL, DL, VT, N0,
8355 DAG.getConstantFP(Recip, DL, VT));
8358 // If this FDIV is part of a reciprocal square root, it may be folded
8359 // into a target-specific square root estimate instruction.
8360 if (N1.getOpcode() == ISD::FSQRT) {
8361 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0))) {
8362 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
8364 } else if (N1.getOpcode() == ISD::FP_EXTEND &&
8365 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
8366 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0).getOperand(0))) {
8367 RV = DAG.getNode(ISD::FP_EXTEND, SDLoc(N1), VT, RV);
8368 AddToWorklist(RV.getNode());
8369 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
8371 } else if (N1.getOpcode() == ISD::FP_ROUND &&
8372 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
8373 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0).getOperand(0))) {
8374 RV = DAG.getNode(ISD::FP_ROUND, SDLoc(N1), VT, RV, N1.getOperand(1));
8375 AddToWorklist(RV.getNode());
8376 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
8378 } else if (N1.getOpcode() == ISD::FMUL) {
8379 // Look through an FMUL. Even though this won't remove the FDIV directly,
8380 // it's still worthwhile to get rid of the FSQRT if possible.
8383 if (N1.getOperand(0).getOpcode() == ISD::FSQRT) {
8384 SqrtOp = N1.getOperand(0);
8385 OtherOp = N1.getOperand(1);
8386 } else if (N1.getOperand(1).getOpcode() == ISD::FSQRT) {
8387 SqrtOp = N1.getOperand(1);
8388 OtherOp = N1.getOperand(0);
8390 if (SqrtOp.getNode()) {
8391 // We found a FSQRT, so try to make this fold:
8392 // x / (y * sqrt(z)) -> x * (rsqrt(z) / y)
8393 if (SDValue RV = BuildRsqrtEstimate(SqrtOp.getOperand(0))) {
8394 RV = DAG.getNode(ISD::FDIV, SDLoc(N1), VT, RV, OtherOp);
8395 AddToWorklist(RV.getNode());
8396 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
8401 // Fold into a reciprocal estimate and multiply instead of a real divide.
8402 if (SDValue RV = BuildReciprocalEstimate(N1)) {
8403 AddToWorklist(RV.getNode());
8404 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
8408 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
8409 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
8410 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
8411 // Both can be negated for free, check to see if at least one is cheaper
8413 if (LHSNeg == 2 || RHSNeg == 2)
8414 return DAG.getNode(ISD::FDIV, SDLoc(N), VT,
8415 GetNegatedExpression(N0, DAG, LegalOperations),
8416 GetNegatedExpression(N1, DAG, LegalOperations));
8420 if (SDValue CombineRepeatedDivisors = combineRepeatedFPDivisors(N))
8421 return CombineRepeatedDivisors;
8426 SDValue DAGCombiner::visitFREM(SDNode *N) {
8427 SDValue N0 = N->getOperand(0);
8428 SDValue N1 = N->getOperand(1);
8429 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8430 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8431 EVT VT = N->getValueType(0);
8433 // fold (frem c1, c2) -> fmod(c1,c2)
8435 return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1);
8440 SDValue DAGCombiner::visitFSQRT(SDNode *N) {
8441 if (!DAG.getTarget().Options.UnsafeFPMath || TLI.isFsqrtCheap())
8444 // Compute this as X * (1/sqrt(X)) = X * (X ** -0.5)
8445 SDValue RV = BuildRsqrtEstimate(N->getOperand(0));
8449 EVT VT = RV.getValueType();
8451 RV = DAG.getNode(ISD::FMUL, DL, VT, N->getOperand(0), RV);
8452 AddToWorklist(RV.getNode());
8454 // Unfortunately, RV is now NaN if the input was exactly 0.
8455 // Select out this case and force the answer to 0.
8456 SDValue Zero = DAG.getConstantFP(0.0, DL, VT);
8457 EVT CCVT = TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
8458 SDValue ZeroCmp = DAG.getSetCC(DL, CCVT, N->getOperand(0), Zero, ISD::SETEQ);
8459 AddToWorklist(ZeroCmp.getNode());
8460 AddToWorklist(RV.getNode());
8462 return DAG.getNode(VT.isVector() ? ISD::VSELECT : ISD::SELECT, DL, VT,
8466 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
8467 SDValue N0 = N->getOperand(0);
8468 SDValue N1 = N->getOperand(1);
8469 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8470 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8471 EVT VT = N->getValueType(0);
8473 if (N0CFP && N1CFP) // Constant fold
8474 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1);
8477 const APFloat& V = N1CFP->getValueAPF();
8478 // copysign(x, c1) -> fabs(x) iff ispos(c1)
8479 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
8480 if (!V.isNegative()) {
8481 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
8482 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
8484 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
8485 return DAG.getNode(ISD::FNEG, SDLoc(N), VT,
8486 DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0));
8490 // copysign(fabs(x), y) -> copysign(x, y)
8491 // copysign(fneg(x), y) -> copysign(x, y)
8492 // copysign(copysign(x,z), y) -> copysign(x, y)
8493 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
8494 N0.getOpcode() == ISD::FCOPYSIGN)
8495 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
8496 N0.getOperand(0), N1);
8498 // copysign(x, abs(y)) -> abs(x)
8499 if (N1.getOpcode() == ISD::FABS)
8500 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
8502 // copysign(x, copysign(y,z)) -> copysign(x, z)
8503 if (N1.getOpcode() == ISD::FCOPYSIGN)
8504 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
8505 N0, N1.getOperand(1));
8507 // copysign(x, fp_extend(y)) -> copysign(x, y)
8508 // copysign(x, fp_round(y)) -> copysign(x, y)
8509 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
8510 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
8511 N0, N1.getOperand(0));
8516 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
8517 SDValue N0 = N->getOperand(0);
8518 EVT VT = N->getValueType(0);
8519 EVT OpVT = N0.getValueType();
8521 // fold (sint_to_fp c1) -> c1fp
8522 if (isConstantIntBuildVectorOrConstantInt(N0) &&
8523 // ...but only if the target supports immediate floating-point values
8524 (!LegalOperations ||
8525 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
8526 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
8528 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
8529 // but UINT_TO_FP is legal on this target, try to convert.
8530 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
8531 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
8532 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
8533 if (DAG.SignBitIsZero(N0))
8534 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
8537 // The next optimizations are desirable only if SELECT_CC can be lowered.
8538 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
8539 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
8540 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
8542 (!LegalOperations ||
8543 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
8546 { N0.getOperand(0), N0.getOperand(1),
8547 DAG.getConstantFP(-1.0, DL, VT), DAG.getConstantFP(0.0, DL, VT),
8549 return DAG.getNode(ISD::SELECT_CC, DL, VT, Ops);
8552 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
8553 // (select_cc x, y, 1.0, 0.0,, cc)
8554 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
8555 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
8556 (!LegalOperations ||
8557 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
8560 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
8561 DAG.getConstantFP(1.0, DL, VT), DAG.getConstantFP(0.0, DL, VT),
8562 N0.getOperand(0).getOperand(2) };
8563 return DAG.getNode(ISD::SELECT_CC, DL, VT, Ops);
8570 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
8571 SDValue N0 = N->getOperand(0);
8572 EVT VT = N->getValueType(0);
8573 EVT OpVT = N0.getValueType();
8575 // fold (uint_to_fp c1) -> c1fp
8576 if (isConstantIntBuildVectorOrConstantInt(N0) &&
8577 // ...but only if the target supports immediate floating-point values
8578 (!LegalOperations ||
8579 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
8580 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
8582 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
8583 // but SINT_TO_FP is legal on this target, try to convert.
8584 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
8585 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
8586 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
8587 if (DAG.SignBitIsZero(N0))
8588 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
8591 // The next optimizations are desirable only if SELECT_CC can be lowered.
8592 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
8593 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
8595 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
8596 (!LegalOperations ||
8597 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
8600 { N0.getOperand(0), N0.getOperand(1),
8601 DAG.getConstantFP(1.0, DL, VT), DAG.getConstantFP(0.0, DL, VT),
8603 return DAG.getNode(ISD::SELECT_CC, DL, VT, Ops);
8610 // Fold (fp_to_{s/u}int ({s/u}int_to_fpx)) -> zext x, sext x, trunc x, or x
8611 static SDValue FoldIntToFPToInt(SDNode *N, SelectionDAG &DAG) {
8612 SDValue N0 = N->getOperand(0);
8613 EVT VT = N->getValueType(0);
8615 if (N0.getOpcode() != ISD::UINT_TO_FP && N0.getOpcode() != ISD::SINT_TO_FP)
8618 SDValue Src = N0.getOperand(0);
8619 EVT SrcVT = Src.getValueType();
8620 bool IsInputSigned = N0.getOpcode() == ISD::SINT_TO_FP;
8621 bool IsOutputSigned = N->getOpcode() == ISD::FP_TO_SINT;
8623 // We can safely assume the conversion won't overflow the output range,
8624 // because (for example) (uint8_t)18293.f is undefined behavior.
8626 // Since we can assume the conversion won't overflow, our decision as to
8627 // whether the input will fit in the float should depend on the minimum
8628 // of the input range and output range.
8630 // This means this is also safe for a signed input and unsigned output, since
8631 // a negative input would lead to undefined behavior.
8632 unsigned InputSize = (int)SrcVT.getScalarSizeInBits() - IsInputSigned;
8633 unsigned OutputSize = (int)VT.getScalarSizeInBits() - IsOutputSigned;
8634 unsigned ActualSize = std::min(InputSize, OutputSize);
8635 const fltSemantics &sem = DAG.EVTToAPFloatSemantics(N0.getValueType());
8637 // We can only fold away the float conversion if the input range can be
8638 // represented exactly in the float range.
8639 if (APFloat::semanticsPrecision(sem) >= ActualSize) {
8640 if (VT.getScalarSizeInBits() > SrcVT.getScalarSizeInBits()) {
8641 unsigned ExtOp = IsInputSigned && IsOutputSigned ? ISD::SIGN_EXTEND
8643 return DAG.getNode(ExtOp, SDLoc(N), VT, Src);
8645 if (VT.getScalarSizeInBits() < SrcVT.getScalarSizeInBits())
8646 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Src);
8649 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Src);
8654 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
8655 SDValue N0 = N->getOperand(0);
8656 EVT VT = N->getValueType(0);
8658 // fold (fp_to_sint c1fp) -> c1
8659 if (isConstantFPBuildVectorOrConstantFP(N0))
8660 return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0);
8662 return FoldIntToFPToInt(N, DAG);
8665 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
8666 SDValue N0 = N->getOperand(0);
8667 EVT VT = N->getValueType(0);
8669 // fold (fp_to_uint c1fp) -> c1
8670 if (isConstantFPBuildVectorOrConstantFP(N0))
8671 return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0);
8673 return FoldIntToFPToInt(N, DAG);
8676 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
8677 SDValue N0 = N->getOperand(0);
8678 SDValue N1 = N->getOperand(1);
8679 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8680 EVT VT = N->getValueType(0);
8682 // fold (fp_round c1fp) -> c1fp
8684 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1);
8686 // fold (fp_round (fp_extend x)) -> x
8687 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
8688 return N0.getOperand(0);
8690 // fold (fp_round (fp_round x)) -> (fp_round x)
8691 if (N0.getOpcode() == ISD::FP_ROUND) {
8692 const bool NIsTrunc = N->getConstantOperandVal(1) == 1;
8693 const bool N0IsTrunc = N0.getNode()->getConstantOperandVal(1) == 1;
8694 // If the first fp_round isn't a value preserving truncation, it might
8695 // introduce a tie in the second fp_round, that wouldn't occur in the
8696 // single-step fp_round we want to fold to.
8697 // In other words, double rounding isn't the same as rounding.
8698 // Also, this is a value preserving truncation iff both fp_round's are.
8699 if (DAG.getTarget().Options.UnsafeFPMath || N0IsTrunc) {
8701 return DAG.getNode(ISD::FP_ROUND, DL, VT, N0.getOperand(0),
8702 DAG.getIntPtrConstant(NIsTrunc && N0IsTrunc, DL));
8706 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
8707 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
8708 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT,
8709 N0.getOperand(0), N1);
8710 AddToWorklist(Tmp.getNode());
8711 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
8712 Tmp, N0.getOperand(1));
8718 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
8719 SDValue N0 = N->getOperand(0);
8720 EVT VT = N->getValueType(0);
8721 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
8722 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8724 // fold (fp_round_inreg c1fp) -> c1fp
8725 if (N0CFP && isTypeLegal(EVT)) {
8727 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), DL, EVT);
8728 return DAG.getNode(ISD::FP_EXTEND, DL, VT, Round);
8734 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
8735 SDValue N0 = N->getOperand(0);
8736 EVT VT = N->getValueType(0);
8738 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
8739 if (N->hasOneUse() &&
8740 N->use_begin()->getOpcode() == ISD::FP_ROUND)
8743 // fold (fp_extend c1fp) -> c1fp
8744 if (isConstantFPBuildVectorOrConstantFP(N0))
8745 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0);
8747 // fold (fp_extend (fp16_to_fp op)) -> (fp16_to_fp op)
8748 if (N0.getOpcode() == ISD::FP16_TO_FP &&
8749 TLI.getOperationAction(ISD::FP16_TO_FP, VT) == TargetLowering::Legal)
8750 return DAG.getNode(ISD::FP16_TO_FP, SDLoc(N), VT, N0.getOperand(0));
8752 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
8754 if (N0.getOpcode() == ISD::FP_ROUND
8755 && N0.getNode()->getConstantOperandVal(1) == 1) {
8756 SDValue In = N0.getOperand(0);
8757 if (In.getValueType() == VT) return In;
8758 if (VT.bitsLT(In.getValueType()))
8759 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT,
8760 In, N0.getOperand(1));
8761 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In);
8764 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
8765 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
8766 TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) {
8767 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
8768 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
8770 LN0->getBasePtr(), N0.getValueType(),
8771 LN0->getMemOperand());
8772 CombineTo(N, ExtLoad);
8773 CombineTo(N0.getNode(),
8774 DAG.getNode(ISD::FP_ROUND, SDLoc(N0),
8775 N0.getValueType(), ExtLoad,
8776 DAG.getIntPtrConstant(1, SDLoc(N0))),
8777 ExtLoad.getValue(1));
8778 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8784 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
8785 SDValue N0 = N->getOperand(0);
8786 EVT VT = N->getValueType(0);
8788 // fold (fceil c1) -> fceil(c1)
8789 if (isConstantFPBuildVectorOrConstantFP(N0))
8790 return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0);
8795 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
8796 SDValue N0 = N->getOperand(0);
8797 EVT VT = N->getValueType(0);
8799 // fold (ftrunc c1) -> ftrunc(c1)
8800 if (isConstantFPBuildVectorOrConstantFP(N0))
8801 return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0);
8806 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
8807 SDValue N0 = N->getOperand(0);
8808 EVT VT = N->getValueType(0);
8810 // fold (ffloor c1) -> ffloor(c1)
8811 if (isConstantFPBuildVectorOrConstantFP(N0))
8812 return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0);
8817 // FIXME: FNEG and FABS have a lot in common; refactor.
8818 SDValue DAGCombiner::visitFNEG(SDNode *N) {
8819 SDValue N0 = N->getOperand(0);
8820 EVT VT = N->getValueType(0);
8822 // Constant fold FNEG.
8823 if (isConstantFPBuildVectorOrConstantFP(N0))
8824 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
8826 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
8827 &DAG.getTarget().Options))
8828 return GetNegatedExpression(N0, DAG, LegalOperations);
8830 // Transform fneg(bitconvert(x)) -> bitconvert(x ^ sign) to avoid loading
8831 // constant pool values.
8832 if (!TLI.isFNegFree(VT) &&
8833 N0.getOpcode() == ISD::BITCAST &&
8834 N0.getNode()->hasOneUse()) {
8835 SDValue Int = N0.getOperand(0);
8836 EVT IntVT = Int.getValueType();
8837 if (IntVT.isInteger() && !IntVT.isVector()) {
8839 if (N0.getValueType().isVector()) {
8840 // For a vector, get a mask such as 0x80... per scalar element
8842 SignMask = APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
8843 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
8845 // For a scalar, just generate 0x80...
8846 SignMask = APInt::getSignBit(IntVT.getSizeInBits());
8849 Int = DAG.getNode(ISD::XOR, DL0, IntVT, Int,
8850 DAG.getConstant(SignMask, DL0, IntVT));
8851 AddToWorklist(Int.getNode());
8852 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Int);
8856 // (fneg (fmul c, x)) -> (fmul -c, x)
8857 if (N0.getOpcode() == ISD::FMUL &&
8858 (N0.getNode()->hasOneUse() || !TLI.isFNegFree(VT))) {
8859 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
8861 APFloat CVal = CFP1->getValueAPF();
8863 if (Level >= AfterLegalizeDAG &&
8864 (TLI.isFPImmLegal(CVal, N->getValueType(0)) ||
8865 TLI.isOperationLegal(ISD::ConstantFP, N->getValueType(0))))
8867 ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
8868 DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0.getOperand(1)));
8875 SDValue DAGCombiner::visitFMINNUM(SDNode *N) {
8876 SDValue N0 = N->getOperand(0);
8877 SDValue N1 = N->getOperand(1);
8878 const ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8879 const ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8881 if (N0CFP && N1CFP) {
8882 const APFloat &C0 = N0CFP->getValueAPF();
8883 const APFloat &C1 = N1CFP->getValueAPF();
8884 return DAG.getConstantFP(minnum(C0, C1), SDLoc(N), N->getValueType(0));
8888 EVT VT = N->getValueType(0);
8889 // Canonicalize to constant on RHS.
8890 return DAG.getNode(ISD::FMINNUM, SDLoc(N), VT, N1, N0);
8896 SDValue DAGCombiner::visitFMAXNUM(SDNode *N) {
8897 SDValue N0 = N->getOperand(0);
8898 SDValue N1 = N->getOperand(1);
8899 const ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8900 const ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8902 if (N0CFP && N1CFP) {
8903 const APFloat &C0 = N0CFP->getValueAPF();
8904 const APFloat &C1 = N1CFP->getValueAPF();
8905 return DAG.getConstantFP(maxnum(C0, C1), SDLoc(N), N->getValueType(0));
8909 EVT VT = N->getValueType(0);
8910 // Canonicalize to constant on RHS.
8911 return DAG.getNode(ISD::FMAXNUM, SDLoc(N), VT, N1, N0);
8917 SDValue DAGCombiner::visitFABS(SDNode *N) {
8918 SDValue N0 = N->getOperand(0);
8919 EVT VT = N->getValueType(0);
8921 // fold (fabs c1) -> fabs(c1)
8922 if (isConstantFPBuildVectorOrConstantFP(N0))
8923 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
8925 // fold (fabs (fabs x)) -> (fabs x)
8926 if (N0.getOpcode() == ISD::FABS)
8927 return N->getOperand(0);
8929 // fold (fabs (fneg x)) -> (fabs x)
8930 // fold (fabs (fcopysign x, y)) -> (fabs x)
8931 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
8932 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0));
8934 // Transform fabs(bitconvert(x)) -> bitconvert(x & ~sign) to avoid loading
8935 // constant pool values.
8936 if (!TLI.isFAbsFree(VT) &&
8937 N0.getOpcode() == ISD::BITCAST &&
8938 N0.getNode()->hasOneUse()) {
8939 SDValue Int = N0.getOperand(0);
8940 EVT IntVT = Int.getValueType();
8941 if (IntVT.isInteger() && !IntVT.isVector()) {
8943 if (N0.getValueType().isVector()) {
8944 // For a vector, get a mask such as 0x7f... per scalar element
8946 SignMask = ~APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
8947 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
8949 // For a scalar, just generate 0x7f...
8950 SignMask = ~APInt::getSignBit(IntVT.getSizeInBits());
8953 Int = DAG.getNode(ISD::AND, DL, IntVT, Int,
8954 DAG.getConstant(SignMask, DL, IntVT));
8955 AddToWorklist(Int.getNode());
8956 return DAG.getNode(ISD::BITCAST, SDLoc(N), N->getValueType(0), Int);
8963 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
8964 SDValue Chain = N->getOperand(0);
8965 SDValue N1 = N->getOperand(1);
8966 SDValue N2 = N->getOperand(2);
8968 // If N is a constant we could fold this into a fallthrough or unconditional
8969 // branch. However that doesn't happen very often in normal code, because
8970 // Instcombine/SimplifyCFG should have handled the available opportunities.
8971 // If we did this folding here, it would be necessary to update the
8972 // MachineBasicBlock CFG, which is awkward.
8974 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
8976 if (N1.getOpcode() == ISD::SETCC &&
8977 TLI.isOperationLegalOrCustom(ISD::BR_CC,
8978 N1.getOperand(0).getValueType())) {
8979 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
8980 Chain, N1.getOperand(2),
8981 N1.getOperand(0), N1.getOperand(1), N2);
8984 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
8985 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
8986 (N1.getOperand(0).hasOneUse() &&
8987 N1.getOperand(0).getOpcode() == ISD::SRL))) {
8988 SDNode *Trunc = nullptr;
8989 if (N1.getOpcode() == ISD::TRUNCATE) {
8990 // Look pass the truncate.
8991 Trunc = N1.getNode();
8992 N1 = N1.getOperand(0);
8995 // Match this pattern so that we can generate simpler code:
8998 // %b = and i32 %a, 2
8999 // %c = srl i32 %b, 1
9000 // brcond i32 %c ...
9005 // %b = and i32 %a, 2
9006 // %c = setcc eq %b, 0
9009 // This applies only when the AND constant value has one bit set and the
9010 // SRL constant is equal to the log2 of the AND constant. The back-end is
9011 // smart enough to convert the result into a TEST/JMP sequence.
9012 SDValue Op0 = N1.getOperand(0);
9013 SDValue Op1 = N1.getOperand(1);
9015 if (Op0.getOpcode() == ISD::AND &&
9016 Op1.getOpcode() == ISD::Constant) {
9017 SDValue AndOp1 = Op0.getOperand(1);
9019 if (AndOp1.getOpcode() == ISD::Constant) {
9020 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
9022 if (AndConst.isPowerOf2() &&
9023 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
9027 getSetCCResultType(Op0.getValueType()),
9028 Op0, DAG.getConstant(0, DL, Op0.getValueType()),
9031 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, DL,
9032 MVT::Other, Chain, SetCC, N2);
9033 // Don't add the new BRCond into the worklist or else SimplifySelectCC
9034 // will convert it back to (X & C1) >> C2.
9035 CombineTo(N, NewBRCond, false);
9036 // Truncate is dead.
9038 deleteAndRecombine(Trunc);
9039 // Replace the uses of SRL with SETCC
9040 WorklistRemover DeadNodes(*this);
9041 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
9042 deleteAndRecombine(N1.getNode());
9043 return SDValue(N, 0); // Return N so it doesn't get rechecked!
9049 // Restore N1 if the above transformation doesn't match.
9050 N1 = N->getOperand(1);
9053 // Transform br(xor(x, y)) -> br(x != y)
9054 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
9055 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
9056 SDNode *TheXor = N1.getNode();
9057 SDValue Op0 = TheXor->getOperand(0);
9058 SDValue Op1 = TheXor->getOperand(1);
9059 if (Op0.getOpcode() == Op1.getOpcode()) {
9060 // Avoid missing important xor optimizations.
9061 if (SDValue Tmp = visitXOR(TheXor)) {
9062 if (Tmp.getNode() != TheXor) {
9063 DEBUG(dbgs() << "\nReplacing.8 ";
9065 dbgs() << "\nWith: ";
9066 Tmp.getNode()->dump(&DAG);
9068 WorklistRemover DeadNodes(*this);
9069 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
9070 deleteAndRecombine(TheXor);
9071 return DAG.getNode(ISD::BRCOND, SDLoc(N),
9072 MVT::Other, Chain, Tmp, N2);
9075 // visitXOR has changed XOR's operands or replaced the XOR completely,
9077 return SDValue(N, 0);
9081 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
9083 if (isOneConstant(Op0) && Op0.hasOneUse() &&
9084 Op0.getOpcode() == ISD::XOR) {
9085 TheXor = Op0.getNode();
9089 EVT SetCCVT = N1.getValueType();
9091 SetCCVT = getSetCCResultType(SetCCVT);
9092 SDValue SetCC = DAG.getSetCC(SDLoc(TheXor),
9095 Equal ? ISD::SETEQ : ISD::SETNE);
9096 // Replace the uses of XOR with SETCC
9097 WorklistRemover DeadNodes(*this);
9098 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
9099 deleteAndRecombine(N1.getNode());
9100 return DAG.getNode(ISD::BRCOND, SDLoc(N),
9101 MVT::Other, Chain, SetCC, N2);
9108 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
9110 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
9111 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
9112 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
9114 // If N is a constant we could fold this into a fallthrough or unconditional
9115 // branch. However that doesn't happen very often in normal code, because
9116 // Instcombine/SimplifyCFG should have handled the available opportunities.
9117 // If we did this folding here, it would be necessary to update the
9118 // MachineBasicBlock CFG, which is awkward.
9120 // Use SimplifySetCC to simplify SETCC's.
9121 SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()),
9122 CondLHS, CondRHS, CC->get(), SDLoc(N),
9124 if (Simp.getNode()) AddToWorklist(Simp.getNode());
9126 // fold to a simpler setcc
9127 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
9128 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
9129 N->getOperand(0), Simp.getOperand(2),
9130 Simp.getOperand(0), Simp.getOperand(1),
9136 /// Return true if 'Use' is a load or a store that uses N as its base pointer
9137 /// and that N may be folded in the load / store addressing mode.
9138 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
9140 const TargetLowering &TLI) {
9144 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
9145 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
9147 VT = LD->getMemoryVT();
9148 AS = LD->getAddressSpace();
9149 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
9150 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
9152 VT = ST->getMemoryVT();
9153 AS = ST->getAddressSpace();
9157 TargetLowering::AddrMode AM;
9158 if (N->getOpcode() == ISD::ADD) {
9159 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
9162 AM.BaseOffs = Offset->getSExtValue();
9166 } else if (N->getOpcode() == ISD::SUB) {
9167 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
9170 AM.BaseOffs = -Offset->getSExtValue();
9177 return TLI.isLegalAddressingMode(DAG.getDataLayout(), AM,
9178 VT.getTypeForEVT(*DAG.getContext()), AS);
9181 /// Try turning a load/store into a pre-indexed load/store when the base
9182 /// pointer is an add or subtract and it has other uses besides the load/store.
9183 /// After the transformation, the new indexed load/store has effectively folded
9184 /// the add/subtract in and all of its other uses are redirected to the
9186 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
9187 if (Level < AfterLegalizeDAG)
9193 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9194 if (LD->isIndexed())
9196 VT = LD->getMemoryVT();
9197 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
9198 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
9200 Ptr = LD->getBasePtr();
9201 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9202 if (ST->isIndexed())
9204 VT = ST->getMemoryVT();
9205 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
9206 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
9208 Ptr = ST->getBasePtr();
9214 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
9215 // out. There is no reason to make this a preinc/predec.
9216 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
9217 Ptr.getNode()->hasOneUse())
9220 // Ask the target to do addressing mode selection.
9223 ISD::MemIndexedMode AM = ISD::UNINDEXED;
9224 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
9227 // Backends without true r+i pre-indexed forms may need to pass a
9228 // constant base with a variable offset so that constant coercion
9229 // will work with the patterns in canonical form.
9230 bool Swapped = false;
9231 if (isa<ConstantSDNode>(BasePtr)) {
9232 std::swap(BasePtr, Offset);
9236 // Don't create a indexed load / store with zero offset.
9237 if (isNullConstant(Offset))
9240 // Try turning it into a pre-indexed load / store except when:
9241 // 1) The new base ptr is a frame index.
9242 // 2) If N is a store and the new base ptr is either the same as or is a
9243 // predecessor of the value being stored.
9244 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
9245 // that would create a cycle.
9246 // 4) All uses are load / store ops that use it as old base ptr.
9248 // Check #1. Preinc'ing a frame index would require copying the stack pointer
9249 // (plus the implicit offset) to a register to preinc anyway.
9250 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
9255 SDValue Val = cast<StoreSDNode>(N)->getValue();
9256 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
9260 // If the offset is a constant, there may be other adds of constants that
9261 // can be folded with this one. We should do this to avoid having to keep
9262 // a copy of the original base pointer.
9263 SmallVector<SDNode *, 16> OtherUses;
9264 if (isa<ConstantSDNode>(Offset))
9265 for (SDNode::use_iterator UI = BasePtr.getNode()->use_begin(),
9266 UE = BasePtr.getNode()->use_end();
9268 SDUse &Use = UI.getUse();
9269 // Skip the use that is Ptr and uses of other results from BasePtr's
9270 // node (important for nodes that return multiple results).
9271 if (Use.getUser() == Ptr.getNode() || Use != BasePtr)
9274 if (Use.getUser()->isPredecessorOf(N))
9277 if (Use.getUser()->getOpcode() != ISD::ADD &&
9278 Use.getUser()->getOpcode() != ISD::SUB) {
9283 SDValue Op1 = Use.getUser()->getOperand((UI.getOperandNo() + 1) & 1);
9284 if (!isa<ConstantSDNode>(Op1)) {
9289 // FIXME: In some cases, we can be smarter about this.
9290 if (Op1.getValueType() != Offset.getValueType()) {
9295 OtherUses.push_back(Use.getUser());
9299 std::swap(BasePtr, Offset);
9301 // Now check for #3 and #4.
9302 bool RealUse = false;
9304 // Caches for hasPredecessorHelper
9305 SmallPtrSet<const SDNode *, 32> Visited;
9306 SmallVector<const SDNode *, 16> Worklist;
9308 for (SDNode *Use : Ptr.getNode()->uses()) {
9311 if (N->hasPredecessorHelper(Use, Visited, Worklist))
9314 // If Ptr may be folded in addressing mode of other use, then it's
9315 // not profitable to do this transformation.
9316 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
9325 Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
9326 BasePtr, Offset, AM);
9328 Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
9329 BasePtr, Offset, AM);
9332 DEBUG(dbgs() << "\nReplacing.4 ";
9334 dbgs() << "\nWith: ";
9335 Result.getNode()->dump(&DAG);
9337 WorklistRemover DeadNodes(*this);
9339 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
9340 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
9342 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
9345 // Finally, since the node is now dead, remove it from the graph.
9346 deleteAndRecombine(N);
9349 std::swap(BasePtr, Offset);
9351 // Replace other uses of BasePtr that can be updated to use Ptr
9352 for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) {
9353 unsigned OffsetIdx = 1;
9354 if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode())
9356 assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() ==
9357 BasePtr.getNode() && "Expected BasePtr operand");
9359 // We need to replace ptr0 in the following expression:
9360 // x0 * offset0 + y0 * ptr0 = t0
9362 // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store)
9364 // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the
9365 // indexed load/store and the expresion that needs to be re-written.
9367 // Therefore, we have:
9368 // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1
9370 ConstantSDNode *CN =
9371 cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx));
9373 APInt Offset0 = CN->getAPIntValue();
9374 APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue();
9376 X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1;
9377 Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1;
9378 X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1;
9379 Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1;
9381 unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD;
9383 APInt CNV = Offset0;
9384 if (X0 < 0) CNV = -CNV;
9385 if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1;
9386 else CNV = CNV - Offset1;
9388 SDLoc DL(OtherUses[i]);
9390 // We can now generate the new expression.
9391 SDValue NewOp1 = DAG.getConstant(CNV, DL, CN->getValueType(0));
9392 SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0);
9394 SDValue NewUse = DAG.getNode(Opcode,
9396 OtherUses[i]->getValueType(0), NewOp1, NewOp2);
9397 DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse);
9398 deleteAndRecombine(OtherUses[i]);
9401 // Replace the uses of Ptr with uses of the updated base value.
9402 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
9403 deleteAndRecombine(Ptr.getNode());
9408 /// Try to combine a load/store with a add/sub of the base pointer node into a
9409 /// post-indexed load/store. The transformation folded the add/subtract into the
9410 /// new indexed load/store effectively and all of its uses are redirected to the
9412 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
9413 if (Level < AfterLegalizeDAG)
9419 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9420 if (LD->isIndexed())
9422 VT = LD->getMemoryVT();
9423 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
9424 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
9426 Ptr = LD->getBasePtr();
9427 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9428 if (ST->isIndexed())
9430 VT = ST->getMemoryVT();
9431 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
9432 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
9434 Ptr = ST->getBasePtr();
9440 if (Ptr.getNode()->hasOneUse())
9443 for (SDNode *Op : Ptr.getNode()->uses()) {
9445 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
9450 ISD::MemIndexedMode AM = ISD::UNINDEXED;
9451 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
9452 // Don't create a indexed load / store with zero offset.
9453 if (isNullConstant(Offset))
9456 // Try turning it into a post-indexed load / store except when
9457 // 1) All uses are load / store ops that use it as base ptr (and
9458 // it may be folded as addressing mmode).
9459 // 2) Op must be independent of N, i.e. Op is neither a predecessor
9460 // nor a successor of N. Otherwise, if Op is folded that would
9463 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
9467 bool TryNext = false;
9468 for (SDNode *Use : BasePtr.getNode()->uses()) {
9469 if (Use == Ptr.getNode())
9472 // If all the uses are load / store addresses, then don't do the
9474 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
9475 bool RealUse = false;
9476 for (SDNode *UseUse : Use->uses()) {
9477 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
9492 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
9493 SDValue Result = isLoad
9494 ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
9495 BasePtr, Offset, AM)
9496 : DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
9497 BasePtr, Offset, AM);
9500 DEBUG(dbgs() << "\nReplacing.5 ";
9502 dbgs() << "\nWith: ";
9503 Result.getNode()->dump(&DAG);
9505 WorklistRemover DeadNodes(*this);
9507 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
9508 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
9510 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
9513 // Finally, since the node is now dead, remove it from the graph.
9514 deleteAndRecombine(N);
9516 // Replace the uses of Use with uses of the updated base value.
9517 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
9518 Result.getValue(isLoad ? 1 : 0));
9519 deleteAndRecombine(Op);
9528 /// \brief Return the base-pointer arithmetic from an indexed \p LD.
9529 SDValue DAGCombiner::SplitIndexingFromLoad(LoadSDNode *LD) {
9530 ISD::MemIndexedMode AM = LD->getAddressingMode();
9531 assert(AM != ISD::UNINDEXED);
9532 SDValue BP = LD->getOperand(1);
9533 SDValue Inc = LD->getOperand(2);
9535 // Some backends use TargetConstants for load offsets, but don't expect
9536 // TargetConstants in general ADD nodes. We can convert these constants into
9537 // regular Constants (if the constant is not opaque).
9538 assert((Inc.getOpcode() != ISD::TargetConstant ||
9539 !cast<ConstantSDNode>(Inc)->isOpaque()) &&
9540 "Cannot split out indexing using opaque target constants");
9541 if (Inc.getOpcode() == ISD::TargetConstant) {
9542 ConstantSDNode *ConstInc = cast<ConstantSDNode>(Inc);
9543 Inc = DAG.getConstant(*ConstInc->getConstantIntValue(), SDLoc(Inc),
9544 ConstInc->getValueType(0));
9548 (AM == ISD::PRE_INC || AM == ISD::POST_INC ? ISD::ADD : ISD::SUB);
9549 return DAG.getNode(Opc, SDLoc(LD), BP.getSimpleValueType(), BP, Inc);
9552 SDValue DAGCombiner::visitLOAD(SDNode *N) {
9553 LoadSDNode *LD = cast<LoadSDNode>(N);
9554 SDValue Chain = LD->getChain();
9555 SDValue Ptr = LD->getBasePtr();
9557 // If load is not volatile and there are no uses of the loaded value (and
9558 // the updated indexed value in case of indexed loads), change uses of the
9559 // chain value into uses of the chain input (i.e. delete the dead load).
9560 if (!LD->isVolatile()) {
9561 if (N->getValueType(1) == MVT::Other) {
9563 if (!N->hasAnyUseOfValue(0)) {
9564 // It's not safe to use the two value CombineTo variant here. e.g.
9565 // v1, chain2 = load chain1, loc
9566 // v2, chain3 = load chain2, loc
9568 // Now we replace use of chain2 with chain1. This makes the second load
9569 // isomorphic to the one we are deleting, and thus makes this load live.
9570 DEBUG(dbgs() << "\nReplacing.6 ";
9572 dbgs() << "\nWith chain: ";
9573 Chain.getNode()->dump(&DAG);
9575 WorklistRemover DeadNodes(*this);
9576 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
9579 deleteAndRecombine(N);
9581 return SDValue(N, 0); // Return N so it doesn't get rechecked!
9585 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
9587 // If this load has an opaque TargetConstant offset, then we cannot split
9588 // the indexing into an add/sub directly (that TargetConstant may not be
9589 // valid for a different type of node, and we cannot convert an opaque
9590 // target constant into a regular constant).
9591 bool HasOTCInc = LD->getOperand(2).getOpcode() == ISD::TargetConstant &&
9592 cast<ConstantSDNode>(LD->getOperand(2))->isOpaque();
9594 if (!N->hasAnyUseOfValue(0) &&
9595 ((MaySplitLoadIndex && !HasOTCInc) || !N->hasAnyUseOfValue(1))) {
9596 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
9598 if (N->hasAnyUseOfValue(1) && MaySplitLoadIndex && !HasOTCInc) {
9599 Index = SplitIndexingFromLoad(LD);
9600 // Try to fold the base pointer arithmetic into subsequent loads and
9602 AddUsersToWorklist(N);
9604 Index = DAG.getUNDEF(N->getValueType(1));
9605 DEBUG(dbgs() << "\nReplacing.7 ";
9607 dbgs() << "\nWith: ";
9608 Undef.getNode()->dump(&DAG);
9609 dbgs() << " and 2 other values\n");
9610 WorklistRemover DeadNodes(*this);
9611 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
9612 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Index);
9613 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
9614 deleteAndRecombine(N);
9615 return SDValue(N, 0); // Return N so it doesn't get rechecked!
9620 // If this load is directly stored, replace the load value with the stored
9622 // TODO: Handle store large -> read small portion.
9623 // TODO: Handle TRUNCSTORE/LOADEXT
9624 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
9625 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
9626 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
9627 if (PrevST->getBasePtr() == Ptr &&
9628 PrevST->getValue().getValueType() == N->getValueType(0))
9629 return CombineTo(N, Chain.getOperand(1), Chain);
9633 // Try to infer better alignment information than the load already has.
9634 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
9635 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
9636 if (Align > LD->getMemOperand()->getBaseAlignment()) {
9638 DAG.getExtLoad(LD->getExtensionType(), SDLoc(N),
9639 LD->getValueType(0),
9640 Chain, Ptr, LD->getPointerInfo(),
9642 LD->isVolatile(), LD->isNonTemporal(),
9643 LD->isInvariant(), Align, LD->getAAInfo());
9644 if (NewLoad.getNode() != N)
9645 return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true);
9650 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
9651 : DAG.getSubtarget().useAA();
9653 if (CombinerAAOnlyFunc.getNumOccurrences() &&
9654 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
9657 if (UseAA && LD->isUnindexed()) {
9658 // Walk up chain skipping non-aliasing memory nodes.
9659 SDValue BetterChain = FindBetterChain(N, Chain);
9661 // If there is a better chain.
9662 if (Chain != BetterChain) {
9665 // Replace the chain to void dependency.
9666 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
9667 ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD),
9668 BetterChain, Ptr, LD->getMemOperand());
9670 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD),
9671 LD->getValueType(0),
9672 BetterChain, Ptr, LD->getMemoryVT(),
9673 LD->getMemOperand());
9676 // Create token factor to keep old chain connected.
9677 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
9678 MVT::Other, Chain, ReplLoad.getValue(1));
9680 // Make sure the new and old chains are cleaned up.
9681 AddToWorklist(Token.getNode());
9683 // Replace uses with load result and token factor. Don't add users
9685 return CombineTo(N, ReplLoad.getValue(0), Token, false);
9689 // Try transforming N to an indexed load.
9690 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
9691 return SDValue(N, 0);
9693 // Try to slice up N to more direct loads if the slices are mapped to
9694 // different register banks or pairing can take place.
9696 return SDValue(N, 0);
9702 /// \brief Helper structure used to slice a load in smaller loads.
9703 /// Basically a slice is obtained from the following sequence:
9704 /// Origin = load Ty1, Base
9705 /// Shift = srl Ty1 Origin, CstTy Amount
9706 /// Inst = trunc Shift to Ty2
9708 /// Then, it will be rewriten into:
9709 /// Slice = load SliceTy, Base + SliceOffset
9710 /// [Inst = zext Slice to Ty2], only if SliceTy <> Ty2
9712 /// SliceTy is deduced from the number of bits that are actually used to
9714 struct LoadedSlice {
9715 /// \brief Helper structure used to compute the cost of a slice.
9717 /// Are we optimizing for code size.
9722 unsigned CrossRegisterBanksCopies;
9726 Cost(bool ForCodeSize = false)
9727 : ForCodeSize(ForCodeSize), Loads(0), Truncates(0),
9728 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {}
9730 /// \brief Get the cost of one isolated slice.
9731 Cost(const LoadedSlice &LS, bool ForCodeSize = false)
9732 : ForCodeSize(ForCodeSize), Loads(1), Truncates(0),
9733 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {
9734 EVT TruncType = LS.Inst->getValueType(0);
9735 EVT LoadedType = LS.getLoadedType();
9736 if (TruncType != LoadedType &&
9737 !LS.DAG->getTargetLoweringInfo().isZExtFree(LoadedType, TruncType))
9741 /// \brief Account for slicing gain in the current cost.
9742 /// Slicing provide a few gains like removing a shift or a
9743 /// truncate. This method allows to grow the cost of the original
9744 /// load with the gain from this slice.
9745 void addSliceGain(const LoadedSlice &LS) {
9746 // Each slice saves a truncate.
9747 const TargetLowering &TLI = LS.DAG->getTargetLoweringInfo();
9748 if (!TLI.isTruncateFree(LS.Inst->getValueType(0),
9749 LS.Inst->getOperand(0).getValueType()))
9751 // If there is a shift amount, this slice gets rid of it.
9754 // If this slice can merge a cross register bank copy, account for it.
9755 if (LS.canMergeExpensiveCrossRegisterBankCopy())
9756 ++CrossRegisterBanksCopies;
9759 Cost &operator+=(const Cost &RHS) {
9761 Truncates += RHS.Truncates;
9762 CrossRegisterBanksCopies += RHS.CrossRegisterBanksCopies;
9768 bool operator==(const Cost &RHS) const {
9769 return Loads == RHS.Loads && Truncates == RHS.Truncates &&
9770 CrossRegisterBanksCopies == RHS.CrossRegisterBanksCopies &&
9771 ZExts == RHS.ZExts && Shift == RHS.Shift;
9774 bool operator!=(const Cost &RHS) const { return !(*this == RHS); }
9776 bool operator<(const Cost &RHS) const {
9777 // Assume cross register banks copies are as expensive as loads.
9778 // FIXME: Do we want some more target hooks?
9779 unsigned ExpensiveOpsLHS = Loads + CrossRegisterBanksCopies;
9780 unsigned ExpensiveOpsRHS = RHS.Loads + RHS.CrossRegisterBanksCopies;
9781 // Unless we are optimizing for code size, consider the
9782 // expensive operation first.
9783 if (!ForCodeSize && ExpensiveOpsLHS != ExpensiveOpsRHS)
9784 return ExpensiveOpsLHS < ExpensiveOpsRHS;
9785 return (Truncates + ZExts + Shift + ExpensiveOpsLHS) <
9786 (RHS.Truncates + RHS.ZExts + RHS.Shift + ExpensiveOpsRHS);
9789 bool operator>(const Cost &RHS) const { return RHS < *this; }
9791 bool operator<=(const Cost &RHS) const { return !(RHS < *this); }
9793 bool operator>=(const Cost &RHS) const { return !(*this < RHS); }
9795 // The last instruction that represent the slice. This should be a
9796 // truncate instruction.
9798 // The original load instruction.
9800 // The right shift amount in bits from the original load.
9802 // The DAG from which Origin came from.
9803 // This is used to get some contextual information about legal types, etc.
9806 LoadedSlice(SDNode *Inst = nullptr, LoadSDNode *Origin = nullptr,
9807 unsigned Shift = 0, SelectionDAG *DAG = nullptr)
9808 : Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {}
9810 /// \brief Get the bits used in a chunk of bits \p BitWidth large.
9811 /// \return Result is \p BitWidth and has used bits set to 1 and
9812 /// not used bits set to 0.
9813 APInt getUsedBits() const {
9814 // Reproduce the trunc(lshr) sequence:
9815 // - Start from the truncated value.
9816 // - Zero extend to the desired bit width.
9818 assert(Origin && "No original load to compare against.");
9819 unsigned BitWidth = Origin->getValueSizeInBits(0);
9820 assert(Inst && "This slice is not bound to an instruction");
9821 assert(Inst->getValueSizeInBits(0) <= BitWidth &&
9822 "Extracted slice is bigger than the whole type!");
9823 APInt UsedBits(Inst->getValueSizeInBits(0), 0);
9824 UsedBits.setAllBits();
9825 UsedBits = UsedBits.zext(BitWidth);
9830 /// \brief Get the size of the slice to be loaded in bytes.
9831 unsigned getLoadedSize() const {
9832 unsigned SliceSize = getUsedBits().countPopulation();
9833 assert(!(SliceSize & 0x7) && "Size is not a multiple of a byte.");
9834 return SliceSize / 8;
9837 /// \brief Get the type that will be loaded for this slice.
9838 /// Note: This may not be the final type for the slice.
9839 EVT getLoadedType() const {
9840 assert(DAG && "Missing context");
9841 LLVMContext &Ctxt = *DAG->getContext();
9842 return EVT::getIntegerVT(Ctxt, getLoadedSize() * 8);
9845 /// \brief Get the alignment of the load used for this slice.
9846 unsigned getAlignment() const {
9847 unsigned Alignment = Origin->getAlignment();
9848 unsigned Offset = getOffsetFromBase();
9850 Alignment = MinAlign(Alignment, Alignment + Offset);
9854 /// \brief Check if this slice can be rewritten with legal operations.
9855 bool isLegal() const {
9856 // An invalid slice is not legal.
9857 if (!Origin || !Inst || !DAG)
9860 // Offsets are for indexed load only, we do not handle that.
9861 if (Origin->getOffset().getOpcode() != ISD::UNDEF)
9864 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
9866 // Check that the type is legal.
9867 EVT SliceType = getLoadedType();
9868 if (!TLI.isTypeLegal(SliceType))
9871 // Check that the load is legal for this type.
9872 if (!TLI.isOperationLegal(ISD::LOAD, SliceType))
9875 // Check that the offset can be computed.
9876 // 1. Check its type.
9877 EVT PtrType = Origin->getBasePtr().getValueType();
9878 if (PtrType == MVT::Untyped || PtrType.isExtended())
9881 // 2. Check that it fits in the immediate.
9882 if (!TLI.isLegalAddImmediate(getOffsetFromBase()))
9885 // 3. Check that the computation is legal.
9886 if (!TLI.isOperationLegal(ISD::ADD, PtrType))
9889 // Check that the zext is legal if it needs one.
9890 EVT TruncateType = Inst->getValueType(0);
9891 if (TruncateType != SliceType &&
9892 !TLI.isOperationLegal(ISD::ZERO_EXTEND, TruncateType))
9898 /// \brief Get the offset in bytes of this slice in the original chunk of
9900 /// \pre DAG != nullptr.
9901 uint64_t getOffsetFromBase() const {
9902 assert(DAG && "Missing context.");
9903 bool IsBigEndian = DAG->getDataLayout().isBigEndian();
9904 assert(!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported.");
9905 uint64_t Offset = Shift / 8;
9906 unsigned TySizeInBytes = Origin->getValueSizeInBits(0) / 8;
9907 assert(!(Origin->getValueSizeInBits(0) & 0x7) &&
9908 "The size of the original loaded type is not a multiple of a"
9910 // If Offset is bigger than TySizeInBytes, it means we are loading all
9911 // zeros. This should have been optimized before in the process.
9912 assert(TySizeInBytes > Offset &&
9913 "Invalid shift amount for given loaded size");
9915 Offset = TySizeInBytes - Offset - getLoadedSize();
9919 /// \brief Generate the sequence of instructions to load the slice
9920 /// represented by this object and redirect the uses of this slice to
9921 /// this new sequence of instructions.
9922 /// \pre this->Inst && this->Origin are valid Instructions and this
9923 /// object passed the legal check: LoadedSlice::isLegal returned true.
9924 /// \return The last instruction of the sequence used to load the slice.
9925 SDValue loadSlice() const {
9926 assert(Inst && Origin && "Unable to replace a non-existing slice.");
9927 const SDValue &OldBaseAddr = Origin->getBasePtr();
9928 SDValue BaseAddr = OldBaseAddr;
9929 // Get the offset in that chunk of bytes w.r.t. the endianess.
9930 int64_t Offset = static_cast<int64_t>(getOffsetFromBase());
9931 assert(Offset >= 0 && "Offset too big to fit in int64_t!");
9933 // BaseAddr = BaseAddr + Offset.
9934 EVT ArithType = BaseAddr.getValueType();
9936 BaseAddr = DAG->getNode(ISD::ADD, DL, ArithType, BaseAddr,
9937 DAG->getConstant(Offset, DL, ArithType));
9940 // Create the type of the loaded slice according to its size.
9941 EVT SliceType = getLoadedType();
9943 // Create the load for the slice.
9944 SDValue LastInst = DAG->getLoad(
9945 SliceType, SDLoc(Origin), Origin->getChain(), BaseAddr,
9946 Origin->getPointerInfo().getWithOffset(Offset), Origin->isVolatile(),
9947 Origin->isNonTemporal(), Origin->isInvariant(), getAlignment());
9948 // If the final type is not the same as the loaded type, this means that
9949 // we have to pad with zero. Create a zero extend for that.
9950 EVT FinalType = Inst->getValueType(0);
9951 if (SliceType != FinalType)
9953 DAG->getNode(ISD::ZERO_EXTEND, SDLoc(LastInst), FinalType, LastInst);
9957 /// \brief Check if this slice can be merged with an expensive cross register
9958 /// bank copy. E.g.,
9960 /// f = bitcast i32 i to float
9961 bool canMergeExpensiveCrossRegisterBankCopy() const {
9962 if (!Inst || !Inst->hasOneUse())
9964 SDNode *Use = *Inst->use_begin();
9965 if (Use->getOpcode() != ISD::BITCAST)
9967 assert(DAG && "Missing context");
9968 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
9969 EVT ResVT = Use->getValueType(0);
9970 const TargetRegisterClass *ResRC = TLI.getRegClassFor(ResVT.getSimpleVT());
9971 const TargetRegisterClass *ArgRC =
9972 TLI.getRegClassFor(Use->getOperand(0).getValueType().getSimpleVT());
9973 if (ArgRC == ResRC || !TLI.isOperationLegal(ISD::LOAD, ResVT))
9976 // At this point, we know that we perform a cross-register-bank copy.
9977 // Check if it is expensive.
9978 const TargetRegisterInfo *TRI = DAG->getSubtarget().getRegisterInfo();
9979 // Assume bitcasts are cheap, unless both register classes do not
9980 // explicitly share a common sub class.
9981 if (!TRI || TRI->getCommonSubClass(ArgRC, ResRC))
9984 // Check if it will be merged with the load.
9985 // 1. Check the alignment constraint.
9986 unsigned RequiredAlignment = DAG->getDataLayout().getABITypeAlignment(
9987 ResVT.getTypeForEVT(*DAG->getContext()));
9989 if (RequiredAlignment > getAlignment())
9992 // 2. Check that the load is a legal operation for that type.
9993 if (!TLI.isOperationLegal(ISD::LOAD, ResVT))
9996 // 3. Check that we do not have a zext in the way.
9997 if (Inst->getValueType(0) != getLoadedType())
10005 /// \brief Check that all bits set in \p UsedBits form a dense region, i.e.,
10006 /// \p UsedBits looks like 0..0 1..1 0..0.
10007 static bool areUsedBitsDense(const APInt &UsedBits) {
10008 // If all the bits are one, this is dense!
10009 if (UsedBits.isAllOnesValue())
10012 // Get rid of the unused bits on the right.
10013 APInt NarrowedUsedBits = UsedBits.lshr(UsedBits.countTrailingZeros());
10014 // Get rid of the unused bits on the left.
10015 if (NarrowedUsedBits.countLeadingZeros())
10016 NarrowedUsedBits = NarrowedUsedBits.trunc(NarrowedUsedBits.getActiveBits());
10017 // Check that the chunk of bits is completely used.
10018 return NarrowedUsedBits.isAllOnesValue();
10021 /// \brief Check whether or not \p First and \p Second are next to each other
10022 /// in memory. This means that there is no hole between the bits loaded
10023 /// by \p First and the bits loaded by \p Second.
10024 static bool areSlicesNextToEachOther(const LoadedSlice &First,
10025 const LoadedSlice &Second) {
10026 assert(First.Origin == Second.Origin && First.Origin &&
10027 "Unable to match different memory origins.");
10028 APInt UsedBits = First.getUsedBits();
10029 assert((UsedBits & Second.getUsedBits()) == 0 &&
10030 "Slices are not supposed to overlap.");
10031 UsedBits |= Second.getUsedBits();
10032 return areUsedBitsDense(UsedBits);
10035 /// \brief Adjust the \p GlobalLSCost according to the target
10036 /// paring capabilities and the layout of the slices.
10037 /// \pre \p GlobalLSCost should account for at least as many loads as
10038 /// there is in the slices in \p LoadedSlices.
10039 static void adjustCostForPairing(SmallVectorImpl<LoadedSlice> &LoadedSlices,
10040 LoadedSlice::Cost &GlobalLSCost) {
10041 unsigned NumberOfSlices = LoadedSlices.size();
10042 // If there is less than 2 elements, no pairing is possible.
10043 if (NumberOfSlices < 2)
10046 // Sort the slices so that elements that are likely to be next to each
10047 // other in memory are next to each other in the list.
10048 std::sort(LoadedSlices.begin(), LoadedSlices.end(),
10049 [](const LoadedSlice &LHS, const LoadedSlice &RHS) {
10050 assert(LHS.Origin == RHS.Origin && "Different bases not implemented.");
10051 return LHS.getOffsetFromBase() < RHS.getOffsetFromBase();
10053 const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo();
10054 // First (resp. Second) is the first (resp. Second) potentially candidate
10055 // to be placed in a paired load.
10056 const LoadedSlice *First = nullptr;
10057 const LoadedSlice *Second = nullptr;
10058 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice,
10059 // Set the beginning of the pair.
10062 Second = &LoadedSlices[CurrSlice];
10064 // If First is NULL, it means we start a new pair.
10065 // Get to the next slice.
10069 EVT LoadedType = First->getLoadedType();
10071 // If the types of the slices are different, we cannot pair them.
10072 if (LoadedType != Second->getLoadedType())
10075 // Check if the target supplies paired loads for this type.
10076 unsigned RequiredAlignment = 0;
10077 if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) {
10078 // move to the next pair, this type is hopeless.
10082 // Check if we meet the alignment requirement.
10083 if (RequiredAlignment > First->getAlignment())
10086 // Check that both loads are next to each other in memory.
10087 if (!areSlicesNextToEachOther(*First, *Second))
10090 assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!");
10091 --GlobalLSCost.Loads;
10092 // Move to the next pair.
10097 /// \brief Check the profitability of all involved LoadedSlice.
10098 /// Currently, it is considered profitable if there is exactly two
10099 /// involved slices (1) which are (2) next to each other in memory, and
10100 /// whose cost (\see LoadedSlice::Cost) is smaller than the original load (3).
10102 /// Note: The order of the elements in \p LoadedSlices may be modified, but not
10103 /// the elements themselves.
10105 /// FIXME: When the cost model will be mature enough, we can relax
10106 /// constraints (1) and (2).
10107 static bool isSlicingProfitable(SmallVectorImpl<LoadedSlice> &LoadedSlices,
10108 const APInt &UsedBits, bool ForCodeSize) {
10109 unsigned NumberOfSlices = LoadedSlices.size();
10110 if (StressLoadSlicing)
10111 return NumberOfSlices > 1;
10114 if (NumberOfSlices != 2)
10118 if (!areUsedBitsDense(UsedBits))
10122 LoadedSlice::Cost OrigCost(ForCodeSize), GlobalSlicingCost(ForCodeSize);
10123 // The original code has one big load.
10124 OrigCost.Loads = 1;
10125 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice) {
10126 const LoadedSlice &LS = LoadedSlices[CurrSlice];
10127 // Accumulate the cost of all the slices.
10128 LoadedSlice::Cost SliceCost(LS, ForCodeSize);
10129 GlobalSlicingCost += SliceCost;
10131 // Account as cost in the original configuration the gain obtained
10132 // with the current slices.
10133 OrigCost.addSliceGain(LS);
10136 // If the target supports paired load, adjust the cost accordingly.
10137 adjustCostForPairing(LoadedSlices, GlobalSlicingCost);
10138 return OrigCost > GlobalSlicingCost;
10141 /// \brief If the given load, \p LI, is used only by trunc or trunc(lshr)
10142 /// operations, split it in the various pieces being extracted.
10144 /// This sort of thing is introduced by SROA.
10145 /// This slicing takes care not to insert overlapping loads.
10146 /// \pre LI is a simple load (i.e., not an atomic or volatile load).
10147 bool DAGCombiner::SliceUpLoad(SDNode *N) {
10148 if (Level < AfterLegalizeDAG)
10151 LoadSDNode *LD = cast<LoadSDNode>(N);
10152 if (LD->isVolatile() || !ISD::isNormalLoad(LD) ||
10153 !LD->getValueType(0).isInteger())
10156 // Keep track of already used bits to detect overlapping values.
10157 // In that case, we will just abort the transformation.
10158 APInt UsedBits(LD->getValueSizeInBits(0), 0);
10160 SmallVector<LoadedSlice, 4> LoadedSlices;
10162 // Check if this load is used as several smaller chunks of bits.
10163 // Basically, look for uses in trunc or trunc(lshr) and record a new chain
10164 // of computation for each trunc.
10165 for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end();
10166 UI != UIEnd; ++UI) {
10167 // Skip the uses of the chain.
10168 if (UI.getUse().getResNo() != 0)
10171 SDNode *User = *UI;
10172 unsigned Shift = 0;
10174 // Check if this is a trunc(lshr).
10175 if (User->getOpcode() == ISD::SRL && User->hasOneUse() &&
10176 isa<ConstantSDNode>(User->getOperand(1))) {
10177 Shift = cast<ConstantSDNode>(User->getOperand(1))->getZExtValue();
10178 User = *User->use_begin();
10181 // At this point, User is a Truncate, iff we encountered, trunc or
10183 if (User->getOpcode() != ISD::TRUNCATE)
10186 // The width of the type must be a power of 2 and greater than 8-bits.
10187 // Otherwise the load cannot be represented in LLVM IR.
10188 // Moreover, if we shifted with a non-8-bits multiple, the slice
10189 // will be across several bytes. We do not support that.
10190 unsigned Width = User->getValueSizeInBits(0);
10191 if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7))
10194 // Build the slice for this chain of computations.
10195 LoadedSlice LS(User, LD, Shift, &DAG);
10196 APInt CurrentUsedBits = LS.getUsedBits();
10198 // Check if this slice overlaps with another.
10199 if ((CurrentUsedBits & UsedBits) != 0)
10201 // Update the bits used globally.
10202 UsedBits |= CurrentUsedBits;
10204 // Check if the new slice would be legal.
10208 // Record the slice.
10209 LoadedSlices.push_back(LS);
10212 // Abort slicing if it does not seem to be profitable.
10213 if (!isSlicingProfitable(LoadedSlices, UsedBits, ForCodeSize))
10218 // Rewrite each chain to use an independent load.
10219 // By construction, each chain can be represented by a unique load.
10221 // Prepare the argument for the new token factor for all the slices.
10222 SmallVector<SDValue, 8> ArgChains;
10223 for (SmallVectorImpl<LoadedSlice>::const_iterator
10224 LSIt = LoadedSlices.begin(),
10225 LSItEnd = LoadedSlices.end();
10226 LSIt != LSItEnd; ++LSIt) {
10227 SDValue SliceInst = LSIt->loadSlice();
10228 CombineTo(LSIt->Inst, SliceInst, true);
10229 if (SliceInst.getNode()->getOpcode() != ISD::LOAD)
10230 SliceInst = SliceInst.getOperand(0);
10231 assert(SliceInst->getOpcode() == ISD::LOAD &&
10232 "It takes more than a zext to get to the loaded slice!!");
10233 ArgChains.push_back(SliceInst.getValue(1));
10236 SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other,
10238 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
10242 /// Check to see if V is (and load (ptr), imm), where the load is having
10243 /// specific bytes cleared out. If so, return the byte size being masked out
10244 /// and the shift amount.
10245 static std::pair<unsigned, unsigned>
10246 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
10247 std::pair<unsigned, unsigned> Result(0, 0);
10249 // Check for the structure we're looking for.
10250 if (V->getOpcode() != ISD::AND ||
10251 !isa<ConstantSDNode>(V->getOperand(1)) ||
10252 !ISD::isNormalLoad(V->getOperand(0).getNode()))
10255 // Check the chain and pointer.
10256 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
10257 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
10259 // The store should be chained directly to the load or be an operand of a
10261 if (LD == Chain.getNode())
10263 else if (Chain->getOpcode() != ISD::TokenFactor)
10264 return Result; // Fail.
10267 for (const SDValue &ChainOp : Chain->op_values())
10268 if (ChainOp.getNode() == LD) {
10272 if (!isOk) return Result;
10275 // This only handles simple types.
10276 if (V.getValueType() != MVT::i16 &&
10277 V.getValueType() != MVT::i32 &&
10278 V.getValueType() != MVT::i64)
10281 // Check the constant mask. Invert it so that the bits being masked out are
10282 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
10283 // follow the sign bit for uniformity.
10284 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
10285 unsigned NotMaskLZ = countLeadingZeros(NotMask);
10286 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
10287 unsigned NotMaskTZ = countTrailingZeros(NotMask);
10288 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
10289 if (NotMaskLZ == 64) return Result; // All zero mask.
10291 // See if we have a continuous run of bits. If so, we have 0*1+0*
10292 if (countTrailingOnes(NotMask >> NotMaskTZ) + NotMaskTZ + NotMaskLZ != 64)
10295 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
10296 if (V.getValueType() != MVT::i64 && NotMaskLZ)
10297 NotMaskLZ -= 64-V.getValueSizeInBits();
10299 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
10300 switch (MaskedBytes) {
10304 default: return Result; // All one mask, or 5-byte mask.
10307 // Verify that the first bit starts at a multiple of mask so that the access
10308 // is aligned the same as the access width.
10309 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
10311 Result.first = MaskedBytes;
10312 Result.second = NotMaskTZ/8;
10317 /// Check to see if IVal is something that provides a value as specified by
10318 /// MaskInfo. If so, replace the specified store with a narrower store of
10319 /// truncated IVal.
10321 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
10322 SDValue IVal, StoreSDNode *St,
10324 unsigned NumBytes = MaskInfo.first;
10325 unsigned ByteShift = MaskInfo.second;
10326 SelectionDAG &DAG = DC->getDAG();
10328 // Check to see if IVal is all zeros in the part being masked in by the 'or'
10329 // that uses this. If not, this is not a replacement.
10330 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
10331 ByteShift*8, (ByteShift+NumBytes)*8);
10332 if (!DAG.MaskedValueIsZero(IVal, Mask)) return nullptr;
10334 // Check that it is legal on the target to do this. It is legal if the new
10335 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
10337 MVT VT = MVT::getIntegerVT(NumBytes*8);
10338 if (!DC->isTypeLegal(VT))
10341 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
10342 // shifted by ByteShift and truncated down to NumBytes.
10345 IVal = DAG.getNode(ISD::SRL, DL, IVal.getValueType(), IVal,
10346 DAG.getConstant(ByteShift*8, DL,
10347 DC->getShiftAmountTy(IVal.getValueType())));
10350 // Figure out the offset for the store and the alignment of the access.
10352 unsigned NewAlign = St->getAlignment();
10354 if (DAG.getDataLayout().isLittleEndian())
10355 StOffset = ByteShift;
10357 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
10359 SDValue Ptr = St->getBasePtr();
10362 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(),
10363 Ptr, DAG.getConstant(StOffset, DL, Ptr.getValueType()));
10364 NewAlign = MinAlign(NewAlign, StOffset);
10367 // Truncate down to the new size.
10368 IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal);
10371 return DAG.getStore(St->getChain(), SDLoc(St), IVal, Ptr,
10372 St->getPointerInfo().getWithOffset(StOffset),
10373 false, false, NewAlign).getNode();
10377 /// Look for sequence of load / op / store where op is one of 'or', 'xor', and
10378 /// 'and' of immediates. If 'op' is only touching some of the loaded bits, try
10379 /// narrowing the load and store if it would end up being a win for performance
10381 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
10382 StoreSDNode *ST = cast<StoreSDNode>(N);
10383 if (ST->isVolatile())
10386 SDValue Chain = ST->getChain();
10387 SDValue Value = ST->getValue();
10388 SDValue Ptr = ST->getBasePtr();
10389 EVT VT = Value.getValueType();
10391 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
10394 unsigned Opc = Value.getOpcode();
10396 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
10397 // is a byte mask indicating a consecutive number of bytes, check to see if
10398 // Y is known to provide just those bytes. If so, we try to replace the
10399 // load + replace + store sequence with a single (narrower) store, which makes
10401 if (Opc == ISD::OR) {
10402 std::pair<unsigned, unsigned> MaskedLoad;
10403 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
10404 if (MaskedLoad.first)
10405 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
10406 Value.getOperand(1), ST,this))
10407 return SDValue(NewST, 0);
10409 // Or is commutative, so try swapping X and Y.
10410 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
10411 if (MaskedLoad.first)
10412 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
10413 Value.getOperand(0), ST,this))
10414 return SDValue(NewST, 0);
10417 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
10418 Value.getOperand(1).getOpcode() != ISD::Constant)
10421 SDValue N0 = Value.getOperand(0);
10422 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
10423 Chain == SDValue(N0.getNode(), 1)) {
10424 LoadSDNode *LD = cast<LoadSDNode>(N0);
10425 if (LD->getBasePtr() != Ptr ||
10426 LD->getPointerInfo().getAddrSpace() !=
10427 ST->getPointerInfo().getAddrSpace())
10430 // Find the type to narrow it the load / op / store to.
10431 SDValue N1 = Value.getOperand(1);
10432 unsigned BitWidth = N1.getValueSizeInBits();
10433 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
10434 if (Opc == ISD::AND)
10435 Imm ^= APInt::getAllOnesValue(BitWidth);
10436 if (Imm == 0 || Imm.isAllOnesValue())
10438 unsigned ShAmt = Imm.countTrailingZeros();
10439 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
10440 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
10441 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
10442 // The narrowing should be profitable, the load/store operation should be
10443 // legal (or custom) and the store size should be equal to the NewVT width.
10444 while (NewBW < BitWidth &&
10445 (NewVT.getStoreSizeInBits() != NewBW ||
10446 !TLI.isOperationLegalOrCustom(Opc, NewVT) ||
10447 !TLI.isNarrowingProfitable(VT, NewVT))) {
10448 NewBW = NextPowerOf2(NewBW);
10449 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
10451 if (NewBW >= BitWidth)
10454 // If the lsb changed does not start at the type bitwidth boundary,
10455 // start at the previous one.
10457 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
10458 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt,
10459 std::min(BitWidth, ShAmt + NewBW));
10460 if ((Imm & Mask) == Imm) {
10461 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
10462 if (Opc == ISD::AND)
10463 NewImm ^= APInt::getAllOnesValue(NewBW);
10464 uint64_t PtrOff = ShAmt / 8;
10465 // For big endian targets, we need to adjust the offset to the pointer to
10466 // load the correct bytes.
10467 if (DAG.getDataLayout().isBigEndian())
10468 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
10470 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
10471 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
10472 if (NewAlign < DAG.getDataLayout().getABITypeAlignment(NewVTTy))
10475 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD),
10476 Ptr.getValueType(), Ptr,
10477 DAG.getConstant(PtrOff, SDLoc(LD),
10478 Ptr.getValueType()));
10479 SDValue NewLD = DAG.getLoad(NewVT, SDLoc(N0),
10480 LD->getChain(), NewPtr,
10481 LD->getPointerInfo().getWithOffset(PtrOff),
10482 LD->isVolatile(), LD->isNonTemporal(),
10483 LD->isInvariant(), NewAlign,
10485 SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD,
10486 DAG.getConstant(NewImm, SDLoc(Value),
10488 SDValue NewST = DAG.getStore(Chain, SDLoc(N),
10490 ST->getPointerInfo().getWithOffset(PtrOff),
10491 false, false, NewAlign);
10493 AddToWorklist(NewPtr.getNode());
10494 AddToWorklist(NewLD.getNode());
10495 AddToWorklist(NewVal.getNode());
10496 WorklistRemover DeadNodes(*this);
10497 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
10506 /// For a given floating point load / store pair, if the load value isn't used
10507 /// by any other operations, then consider transforming the pair to integer
10508 /// load / store operations if the target deems the transformation profitable.
10509 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
10510 StoreSDNode *ST = cast<StoreSDNode>(N);
10511 SDValue Chain = ST->getChain();
10512 SDValue Value = ST->getValue();
10513 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
10514 Value.hasOneUse() &&
10515 Chain == SDValue(Value.getNode(), 1)) {
10516 LoadSDNode *LD = cast<LoadSDNode>(Value);
10517 EVT VT = LD->getMemoryVT();
10518 if (!VT.isFloatingPoint() ||
10519 VT != ST->getMemoryVT() ||
10520 LD->isNonTemporal() ||
10521 ST->isNonTemporal() ||
10522 LD->getPointerInfo().getAddrSpace() != 0 ||
10523 ST->getPointerInfo().getAddrSpace() != 0)
10526 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
10527 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
10528 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
10529 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
10530 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
10533 unsigned LDAlign = LD->getAlignment();
10534 unsigned STAlign = ST->getAlignment();
10535 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
10536 unsigned ABIAlign = DAG.getDataLayout().getABITypeAlignment(IntVTTy);
10537 if (LDAlign < ABIAlign || STAlign < ABIAlign)
10540 SDValue NewLD = DAG.getLoad(IntVT, SDLoc(Value),
10541 LD->getChain(), LD->getBasePtr(),
10542 LD->getPointerInfo(),
10543 false, false, false, LDAlign);
10545 SDValue NewST = DAG.getStore(NewLD.getValue(1), SDLoc(N),
10546 NewLD, ST->getBasePtr(),
10547 ST->getPointerInfo(),
10548 false, false, STAlign);
10550 AddToWorklist(NewLD.getNode());
10551 AddToWorklist(NewST.getNode());
10552 WorklistRemover DeadNodes(*this);
10553 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
10562 /// Helper struct to parse and store a memory address as base + index + offset.
10563 /// We ignore sign extensions when it is safe to do so.
10564 /// The following two expressions are not equivalent. To differentiate we need
10565 /// to store whether there was a sign extension involved in the index
10567 /// (load (i64 add (i64 copyfromreg %c)
10568 /// (i64 signextend (add (i8 load %index)
10572 /// (load (i64 add (i64 copyfromreg %c)
10573 /// (i64 signextend (i32 add (i32 signextend (i8 load %index))
10575 struct BaseIndexOffset {
10579 bool IsIndexSignExt;
10581 BaseIndexOffset() : Offset(0), IsIndexSignExt(false) {}
10583 BaseIndexOffset(SDValue Base, SDValue Index, int64_t Offset,
10584 bool IsIndexSignExt) :
10585 Base(Base), Index(Index), Offset(Offset), IsIndexSignExt(IsIndexSignExt) {}
10587 bool equalBaseIndex(const BaseIndexOffset &Other) {
10588 return Other.Base == Base && Other.Index == Index &&
10589 Other.IsIndexSignExt == IsIndexSignExt;
10592 /// Parses tree in Ptr for base, index, offset addresses.
10593 static BaseIndexOffset match(SDValue Ptr) {
10594 bool IsIndexSignExt = false;
10596 // We only can pattern match BASE + INDEX + OFFSET. If Ptr is not an ADD
10597 // instruction, then it could be just the BASE or everything else we don't
10598 // know how to handle. Just use Ptr as BASE and give up.
10599 if (Ptr->getOpcode() != ISD::ADD)
10600 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
10602 // We know that we have at least an ADD instruction. Try to pattern match
10603 // the simple case of BASE + OFFSET.
10604 if (isa<ConstantSDNode>(Ptr->getOperand(1))) {
10605 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
10606 return BaseIndexOffset(Ptr->getOperand(0), SDValue(), Offset,
10610 // Inside a loop the current BASE pointer is calculated using an ADD and a
10611 // MUL instruction. In this case Ptr is the actual BASE pointer.
10612 // (i64 add (i64 %array_ptr)
10613 // (i64 mul (i64 %induction_var)
10614 // (i64 %element_size)))
10615 if (Ptr->getOperand(1)->getOpcode() == ISD::MUL)
10616 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
10618 // Look at Base + Index + Offset cases.
10619 SDValue Base = Ptr->getOperand(0);
10620 SDValue IndexOffset = Ptr->getOperand(1);
10622 // Skip signextends.
10623 if (IndexOffset->getOpcode() == ISD::SIGN_EXTEND) {
10624 IndexOffset = IndexOffset->getOperand(0);
10625 IsIndexSignExt = true;
10628 // Either the case of Base + Index (no offset) or something else.
10629 if (IndexOffset->getOpcode() != ISD::ADD)
10630 return BaseIndexOffset(Base, IndexOffset, 0, IsIndexSignExt);
10632 // Now we have the case of Base + Index + offset.
10633 SDValue Index = IndexOffset->getOperand(0);
10634 SDValue Offset = IndexOffset->getOperand(1);
10636 if (!isa<ConstantSDNode>(Offset))
10637 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
10639 // Ignore signextends.
10640 if (Index->getOpcode() == ISD::SIGN_EXTEND) {
10641 Index = Index->getOperand(0);
10642 IsIndexSignExt = true;
10643 } else IsIndexSignExt = false;
10645 int64_t Off = cast<ConstantSDNode>(Offset)->getSExtValue();
10646 return BaseIndexOffset(Base, Index, Off, IsIndexSignExt);
10651 SDValue DAGCombiner::getMergedConstantVectorStore(SelectionDAG &DAG,
10653 ArrayRef<MemOpLink> Stores,
10655 SmallVector<SDValue, 8> BuildVector;
10657 for (unsigned I = 0, E = Ty.getVectorNumElements(); I != E; ++I)
10658 BuildVector.push_back(cast<StoreSDNode>(Stores[I].MemNode)->getValue());
10660 return DAG.getNode(ISD::BUILD_VECTOR, SL, Ty, BuildVector);
10663 bool DAGCombiner::MergeStoresOfConstantsOrVecElts(
10664 SmallVectorImpl<MemOpLink> &StoreNodes, EVT MemVT,
10665 unsigned NumElem, bool IsConstantSrc, bool UseVector) {
10666 // Make sure we have something to merge.
10670 int64_t ElementSizeBytes = MemVT.getSizeInBits() / 8;
10671 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
10672 unsigned LatestNodeUsed = 0;
10674 for (unsigned i=0; i < NumElem; ++i) {
10675 // Find a chain for the new wide-store operand. Notice that some
10676 // of the store nodes that we found may not be selected for inclusion
10677 // in the wide store. The chain we use needs to be the chain of the
10678 // latest store node which is *used* and replaced by the wide store.
10679 if (StoreNodes[i].SequenceNum < StoreNodes[LatestNodeUsed].SequenceNum)
10680 LatestNodeUsed = i;
10683 // The latest Node in the DAG.
10684 LSBaseSDNode *LatestOp = StoreNodes[LatestNodeUsed].MemNode;
10685 SDLoc DL(StoreNodes[0].MemNode);
10689 // Find a legal type for the vector store.
10690 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
10691 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
10692 if (IsConstantSrc) {
10693 StoredVal = getMergedConstantVectorStore(DAG, DL, StoreNodes, Ty);
10695 SmallVector<SDValue, 8> Ops;
10696 for (unsigned i = 0; i < NumElem ; ++i) {
10697 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10698 SDValue Val = St->getValue();
10699 // All of the operands of a BUILD_VECTOR must have the same type.
10700 if (Val.getValueType() != MemVT)
10702 Ops.push_back(Val);
10705 // Build the extracted vector elements back into a vector.
10706 StoredVal = DAG.getNode(ISD::BUILD_VECTOR, DL, Ty, Ops);
10709 // We should always use a vector store when merging extracted vector
10710 // elements, so this path implies a store of constants.
10711 assert(IsConstantSrc && "Merged vector elements should use vector store");
10713 unsigned SizeInBits = NumElem * ElementSizeBytes * 8;
10714 APInt StoreInt(SizeInBits, 0);
10716 // Construct a single integer constant which is made of the smaller
10717 // constant inputs.
10718 bool IsLE = DAG.getDataLayout().isLittleEndian();
10719 for (unsigned i = 0; i < NumElem ; ++i) {
10720 unsigned Idx = IsLE ? (NumElem - 1 - i) : i;
10721 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
10722 SDValue Val = St->getValue();
10723 StoreInt <<= ElementSizeBytes * 8;
10724 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
10725 StoreInt |= C->getAPIntValue().zext(SizeInBits);
10726 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
10727 StoreInt |= C->getValueAPF().bitcastToAPInt().zext(SizeInBits);
10729 llvm_unreachable("Invalid constant element type");
10733 // Create the new Load and Store operations.
10734 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), SizeInBits);
10735 StoredVal = DAG.getConstant(StoreInt, DL, StoreTy);
10738 SDValue NewStore = DAG.getStore(LatestOp->getChain(), DL, StoredVal,
10739 FirstInChain->getBasePtr(),
10740 FirstInChain->getPointerInfo(),
10742 FirstInChain->getAlignment());
10744 // Replace the last store with the new store
10745 CombineTo(LatestOp, NewStore);
10746 // Erase all other stores.
10747 for (unsigned i = 0; i < NumElem ; ++i) {
10748 if (StoreNodes[i].MemNode == LatestOp)
10750 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10751 // ReplaceAllUsesWith will replace all uses that existed when it was
10752 // called, but graph optimizations may cause new ones to appear. For
10753 // example, the case in pr14333 looks like
10755 // St's chain -> St -> another store -> X
10757 // And the only difference from St to the other store is the chain.
10758 // When we change it's chain to be St's chain they become identical,
10759 // get CSEed and the net result is that X is now a use of St.
10760 // Since we know that St is redundant, just iterate.
10761 while (!St->use_empty())
10762 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain());
10763 deleteAndRecombine(St);
10769 void DAGCombiner::getStoreMergeAndAliasCandidates(
10770 StoreSDNode* St, SmallVectorImpl<MemOpLink> &StoreNodes,
10771 SmallVectorImpl<LSBaseSDNode*> &AliasLoadNodes) {
10772 // This holds the base pointer, index, and the offset in bytes from the base
10774 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr());
10776 // We must have a base and an offset.
10777 if (!BasePtr.Base.getNode())
10780 // Do not handle stores to undef base pointers.
10781 if (BasePtr.Base.getOpcode() == ISD::UNDEF)
10784 // Walk up the chain and look for nodes with offsets from the same
10785 // base pointer. Stop when reaching an instruction with a different kind
10786 // or instruction which has a different base pointer.
10787 EVT MemVT = St->getMemoryVT();
10789 StoreSDNode *Index = St;
10791 // If the chain has more than one use, then we can't reorder the mem ops.
10792 if (Index != St && !SDValue(Index, 0)->hasOneUse())
10795 // Find the base pointer and offset for this memory node.
10796 BaseIndexOffset Ptr = BaseIndexOffset::match(Index->getBasePtr());
10798 // Check that the base pointer is the same as the original one.
10799 if (!Ptr.equalBaseIndex(BasePtr))
10802 // The memory operands must not be volatile.
10803 if (Index->isVolatile() || Index->isIndexed())
10807 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
10808 if (St->isTruncatingStore())
10811 // The stored memory type must be the same.
10812 if (Index->getMemoryVT() != MemVT)
10815 // We found a potential memory operand to merge.
10816 StoreNodes.push_back(MemOpLink(Index, Ptr.Offset, Seq++));
10818 // Find the next memory operand in the chain. If the next operand in the
10819 // chain is a store then move up and continue the scan with the next
10820 // memory operand. If the next operand is a load save it and use alias
10821 // information to check if it interferes with anything.
10822 SDNode *NextInChain = Index->getChain().getNode();
10824 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
10825 // We found a store node. Use it for the next iteration.
10828 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
10829 if (Ldn->isVolatile()) {
10834 // Save the load node for later. Continue the scan.
10835 AliasLoadNodes.push_back(Ldn);
10836 NextInChain = Ldn->getChain().getNode();
10846 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
10847 if (OptLevel == CodeGenOpt::None)
10850 EVT MemVT = St->getMemoryVT();
10851 int64_t ElementSizeBytes = MemVT.getSizeInBits() / 8;
10852 bool NoVectors = DAG.getMachineFunction().getFunction()->hasFnAttribute(
10853 Attribute::NoImplicitFloat);
10855 // This function cannot currently deal with non-byte-sized memory sizes.
10856 if (ElementSizeBytes * 8 != MemVT.getSizeInBits())
10859 // Don't merge vectors into wider inputs.
10860 if (MemVT.isVector() || !MemVT.isSimple())
10863 // Perform an early exit check. Do not bother looking at stored values that
10864 // are not constants, loads, or extracted vector elements.
10865 SDValue StoredVal = St->getValue();
10866 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
10867 bool IsConstantSrc = isa<ConstantSDNode>(StoredVal) ||
10868 isa<ConstantFPSDNode>(StoredVal);
10869 bool IsExtractVecEltSrc = (StoredVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT);
10871 if (!IsConstantSrc && !IsLoadSrc && !IsExtractVecEltSrc)
10874 // Only look at ends of store sequences.
10875 SDValue Chain = SDValue(St, 0);
10876 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
10879 // Save the LoadSDNodes that we find in the chain.
10880 // We need to make sure that these nodes do not interfere with
10881 // any of the store nodes.
10882 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
10884 // Save the StoreSDNodes that we find in the chain.
10885 SmallVector<MemOpLink, 8> StoreNodes;
10887 getStoreMergeAndAliasCandidates(St, StoreNodes, AliasLoadNodes);
10889 // Check if there is anything to merge.
10890 if (StoreNodes.size() < 2)
10893 // Sort the memory operands according to their distance from the base pointer.
10894 std::sort(StoreNodes.begin(), StoreNodes.end(),
10895 [](MemOpLink LHS, MemOpLink RHS) {
10896 return LHS.OffsetFromBase < RHS.OffsetFromBase ||
10897 (LHS.OffsetFromBase == RHS.OffsetFromBase &&
10898 LHS.SequenceNum > RHS.SequenceNum);
10901 // Scan the memory operations on the chain and find the first non-consecutive
10902 // store memory address.
10903 unsigned LastConsecutiveStore = 0;
10904 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
10905 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) {
10907 // Check that the addresses are consecutive starting from the second
10908 // element in the list of stores.
10910 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
10911 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
10915 bool Alias = false;
10916 // Check if this store interferes with any of the loads that we found.
10917 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld)
10918 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) {
10922 // We found a load that alias with this store. Stop the sequence.
10926 // Mark this node as useful.
10927 LastConsecutiveStore = i;
10930 // The node with the lowest store address.
10931 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
10932 unsigned FirstStoreAS = FirstInChain->getAddressSpace();
10933 unsigned FirstStoreAlign = FirstInChain->getAlignment();
10934 LLVMContext &Context = *DAG.getContext();
10935 const DataLayout &DL = DAG.getDataLayout();
10937 // Store the constants into memory as one consecutive store.
10938 if (IsConstantSrc) {
10939 unsigned LastLegalType = 0;
10940 unsigned LastLegalVectorType = 0;
10941 bool NonZero = false;
10942 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
10943 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10944 SDValue StoredVal = St->getValue();
10946 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
10947 NonZero |= !C->isNullValue();
10948 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) {
10949 NonZero |= !C->getConstantFPValue()->isNullValue();
10955 // Find a legal type for the constant store.
10956 unsigned SizeInBits = (i+1) * ElementSizeBytes * 8;
10957 EVT StoreTy = EVT::getIntegerVT(Context, SizeInBits);
10958 if (TLI.isTypeLegal(StoreTy) &&
10959 TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstStoreAS,
10960 FirstStoreAlign)) {
10961 LastLegalType = i+1;
10962 // Or check whether a truncstore is legal.
10963 } else if (TLI.getTypeAction(Context, StoreTy) ==
10964 TargetLowering::TypePromoteInteger) {
10965 EVT LegalizedStoredValueTy =
10966 TLI.getTypeToTransformTo(Context, StoredVal.getValueType());
10967 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
10968 TLI.allowsMemoryAccess(Context, DL, LegalizedStoredValueTy,
10969 FirstStoreAS, FirstStoreAlign)) {
10970 LastLegalType = i + 1;
10974 // Find a legal type for the vector store.
10975 EVT Ty = EVT::getVectorVT(Context, MemVT, i+1);
10976 if (TLI.isTypeLegal(Ty) &&
10977 TLI.allowsMemoryAccess(Context, DL, Ty, FirstStoreAS,
10978 FirstStoreAlign)) {
10979 LastLegalVectorType = i + 1;
10984 // We only use vectors if the constant is known to be zero or the target
10985 // allows it and the function is not marked with the noimplicitfloat
10988 LastLegalVectorType = 0;
10989 } else if (NonZero && !TLI.storeOfVectorConstantIsCheap(MemVT,
10990 LastLegalVectorType,
10992 LastLegalVectorType = 0;
10995 // Check if we found a legal integer type to store.
10996 if (LastLegalType == 0 && LastLegalVectorType == 0)
10999 bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors;
11000 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType;
11002 return MergeStoresOfConstantsOrVecElts(StoreNodes, MemVT, NumElem,
11006 // When extracting multiple vector elements, try to store them
11007 // in one vector store rather than a sequence of scalar stores.
11008 if (IsExtractVecEltSrc) {
11009 unsigned NumElem = 0;
11010 for (unsigned i = 0; i < LastConsecutiveStore + 1; ++i) {
11011 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
11012 SDValue StoredVal = St->getValue();
11013 // This restriction could be loosened.
11014 // Bail out if any stored values are not elements extracted from a vector.
11015 // It should be possible to handle mixed sources, but load sources need
11016 // more careful handling (see the block of code below that handles
11017 // consecutive loads).
11018 if (StoredVal.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
11021 // Find a legal type for the vector store.
11022 EVT Ty = EVT::getVectorVT(Context, MemVT, i+1);
11023 if (TLI.isTypeLegal(Ty) &&
11024 TLI.allowsMemoryAccess(Context, DL, Ty, FirstStoreAS,
11029 return MergeStoresOfConstantsOrVecElts(StoreNodes, MemVT, NumElem,
11033 // Below we handle the case of multiple consecutive stores that
11034 // come from multiple consecutive loads. We merge them into a single
11035 // wide load and a single wide store.
11037 // Look for load nodes which are used by the stored values.
11038 SmallVector<MemOpLink, 8> LoadNodes;
11040 // Find acceptable loads. Loads need to have the same chain (token factor),
11041 // must not be zext, volatile, indexed, and they must be consecutive.
11042 BaseIndexOffset LdBasePtr;
11043 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
11044 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
11045 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
11048 // Loads must only have one use.
11049 if (!Ld->hasNUsesOfValue(1, 0))
11052 // The memory operands must not be volatile.
11053 if (Ld->isVolatile() || Ld->isIndexed())
11056 // We do not accept ext loads.
11057 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
11060 // The stored memory type must be the same.
11061 if (Ld->getMemoryVT() != MemVT)
11064 BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld->getBasePtr());
11065 // If this is not the first ptr that we check.
11066 if (LdBasePtr.Base.getNode()) {
11067 // The base ptr must be the same.
11068 if (!LdPtr.equalBaseIndex(LdBasePtr))
11071 // Check that all other base pointers are the same as this one.
11075 // We found a potential memory operand to merge.
11076 LoadNodes.push_back(MemOpLink(Ld, LdPtr.Offset, 0));
11079 if (LoadNodes.size() < 2)
11082 // If we have load/store pair instructions and we only have two values,
11084 unsigned RequiredAlignment;
11085 if (LoadNodes.size() == 2 && TLI.hasPairedLoad(MemVT, RequiredAlignment) &&
11086 St->getAlignment() >= RequiredAlignment)
11089 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
11090 unsigned FirstLoadAS = FirstLoad->getAddressSpace();
11091 unsigned FirstLoadAlign = FirstLoad->getAlignment();
11093 // Scan the memory operations on the chain and find the first non-consecutive
11094 // load memory address. These variables hold the index in the store node
11096 unsigned LastConsecutiveLoad = 0;
11097 // This variable refers to the size and not index in the array.
11098 unsigned LastLegalVectorType = 0;
11099 unsigned LastLegalIntegerType = 0;
11100 StartAddress = LoadNodes[0].OffsetFromBase;
11101 SDValue FirstChain = FirstLoad->getChain();
11102 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
11103 // All loads much share the same chain.
11104 if (LoadNodes[i].MemNode->getChain() != FirstChain)
11107 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
11108 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
11110 LastConsecutiveLoad = i;
11112 // Find a legal type for the vector store.
11113 EVT StoreTy = EVT::getVectorVT(Context, MemVT, i+1);
11114 if (TLI.isTypeLegal(StoreTy) &&
11115 TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstStoreAS,
11116 FirstStoreAlign) &&
11117 TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstLoadAS,
11119 LastLegalVectorType = i + 1;
11122 // Find a legal type for the integer store.
11123 unsigned SizeInBits = (i+1) * ElementSizeBytes * 8;
11124 StoreTy = EVT::getIntegerVT(Context, SizeInBits);
11125 if (TLI.isTypeLegal(StoreTy) &&
11126 TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstStoreAS,
11127 FirstStoreAlign) &&
11128 TLI.allowsMemoryAccess(Context, DL, StoreTy, FirstLoadAS,
11130 LastLegalIntegerType = i + 1;
11131 // Or check whether a truncstore and extload is legal.
11132 else if (TLI.getTypeAction(Context, StoreTy) ==
11133 TargetLowering::TypePromoteInteger) {
11134 EVT LegalizedStoredValueTy =
11135 TLI.getTypeToTransformTo(Context, StoreTy);
11136 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
11137 TLI.isLoadExtLegal(ISD::ZEXTLOAD, LegalizedStoredValueTy, StoreTy) &&
11138 TLI.isLoadExtLegal(ISD::SEXTLOAD, LegalizedStoredValueTy, StoreTy) &&
11139 TLI.isLoadExtLegal(ISD::EXTLOAD, LegalizedStoredValueTy, StoreTy) &&
11140 TLI.allowsMemoryAccess(Context, DL, LegalizedStoredValueTy,
11141 FirstStoreAS, FirstStoreAlign) &&
11142 TLI.allowsMemoryAccess(Context, DL, LegalizedStoredValueTy,
11143 FirstLoadAS, FirstLoadAlign))
11144 LastLegalIntegerType = i+1;
11148 // Only use vector types if the vector type is larger than the integer type.
11149 // If they are the same, use integers.
11150 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors;
11151 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType);
11153 // We add +1 here because the LastXXX variables refer to location while
11154 // the NumElem refers to array/index size.
11155 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1;
11156 NumElem = std::min(LastLegalType, NumElem);
11161 // The latest Node in the DAG.
11162 unsigned LatestNodeUsed = 0;
11163 for (unsigned i=1; i<NumElem; ++i) {
11164 // Find a chain for the new wide-store operand. Notice that some
11165 // of the store nodes that we found may not be selected for inclusion
11166 // in the wide store. The chain we use needs to be the chain of the
11167 // latest store node which is *used* and replaced by the wide store.
11168 if (StoreNodes[i].SequenceNum < StoreNodes[LatestNodeUsed].SequenceNum)
11169 LatestNodeUsed = i;
11172 LSBaseSDNode *LatestOp = StoreNodes[LatestNodeUsed].MemNode;
11174 // Find if it is better to use vectors or integers to load and store
11178 JointMemOpVT = EVT::getVectorVT(Context, MemVT, NumElem);
11180 unsigned SizeInBits = NumElem * ElementSizeBytes * 8;
11181 JointMemOpVT = EVT::getIntegerVT(Context, SizeInBits);
11184 SDLoc LoadDL(LoadNodes[0].MemNode);
11185 SDLoc StoreDL(StoreNodes[0].MemNode);
11187 SDValue NewLoad = DAG.getLoad(
11188 JointMemOpVT, LoadDL, FirstLoad->getChain(), FirstLoad->getBasePtr(),
11189 FirstLoad->getPointerInfo(), false, false, false, FirstLoadAlign);
11191 SDValue NewStore = DAG.getStore(
11192 LatestOp->getChain(), StoreDL, NewLoad, FirstInChain->getBasePtr(),
11193 FirstInChain->getPointerInfo(), false, false, FirstStoreAlign);
11195 // Replace one of the loads with the new load.
11196 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode);
11197 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
11198 SDValue(NewLoad.getNode(), 1));
11200 // Remove the rest of the load chains.
11201 for (unsigned i = 1; i < NumElem ; ++i) {
11202 // Replace all chain users of the old load nodes with the chain of the new
11204 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
11205 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain());
11208 // Replace the last store with the new store.
11209 CombineTo(LatestOp, NewStore);
11210 // Erase all other stores.
11211 for (unsigned i = 0; i < NumElem ; ++i) {
11212 // Remove all Store nodes.
11213 if (StoreNodes[i].MemNode == LatestOp)
11215 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
11216 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain());
11217 deleteAndRecombine(St);
11223 SDValue DAGCombiner::visitSTORE(SDNode *N) {
11224 StoreSDNode *ST = cast<StoreSDNode>(N);
11225 SDValue Chain = ST->getChain();
11226 SDValue Value = ST->getValue();
11227 SDValue Ptr = ST->getBasePtr();
11229 // If this is a store of a bit convert, store the input value if the
11230 // resultant store does not need a higher alignment than the original.
11231 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
11232 ST->isUnindexed()) {
11233 unsigned OrigAlign = ST->getAlignment();
11234 EVT SVT = Value.getOperand(0).getValueType();
11235 unsigned Align = DAG.getDataLayout().getABITypeAlignment(
11236 SVT.getTypeForEVT(*DAG.getContext()));
11237 if (Align <= OrigAlign &&
11238 ((!LegalOperations && !ST->isVolatile()) ||
11239 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
11240 return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0),
11241 Ptr, ST->getPointerInfo(), ST->isVolatile(),
11242 ST->isNonTemporal(), OrigAlign,
11246 // Turn 'store undef, Ptr' -> nothing.
11247 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
11250 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
11251 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
11252 // NOTE: If the original store is volatile, this transform must not increase
11253 // the number of stores. For example, on x86-32 an f64 can be stored in one
11254 // processor operation but an i64 (which is not legal) requires two. So the
11255 // transform should not be done in this case.
11256 if (Value.getOpcode() != ISD::TargetConstantFP) {
11258 switch (CFP->getSimpleValueType(0).SimpleTy) {
11259 default: llvm_unreachable("Unknown FP type");
11260 case MVT::f16: // We don't do this for these yet.
11266 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
11267 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
11269 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
11270 bitcastToAPInt().getZExtValue(), SDLoc(CFP),
11272 return DAG.getStore(Chain, SDLoc(N), Tmp,
11273 Ptr, ST->getMemOperand());
11277 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
11278 !ST->isVolatile()) ||
11279 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
11281 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
11282 getZExtValue(), SDLoc(CFP), MVT::i64);
11283 return DAG.getStore(Chain, SDLoc(N), Tmp,
11284 Ptr, ST->getMemOperand());
11287 if (!ST->isVolatile() &&
11288 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
11289 // Many FP stores are not made apparent until after legalize, e.g. for
11290 // argument passing. Since this is so common, custom legalize the
11291 // 64-bit integer store into two 32-bit stores.
11292 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
11293 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, SDLoc(CFP), MVT::i32);
11294 SDValue Hi = DAG.getConstant(Val >> 32, SDLoc(CFP), MVT::i32);
11295 if (DAG.getDataLayout().isBigEndian())
11298 unsigned Alignment = ST->getAlignment();
11299 bool isVolatile = ST->isVolatile();
11300 bool isNonTemporal = ST->isNonTemporal();
11301 AAMDNodes AAInfo = ST->getAAInfo();
11305 SDValue St0 = DAG.getStore(Chain, SDLoc(ST), Lo,
11306 Ptr, ST->getPointerInfo(),
11307 isVolatile, isNonTemporal,
11308 ST->getAlignment(), AAInfo);
11309 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
11310 DAG.getConstant(4, DL, Ptr.getValueType()));
11311 Alignment = MinAlign(Alignment, 4U);
11312 SDValue St1 = DAG.getStore(Chain, SDLoc(ST), Hi,
11313 Ptr, ST->getPointerInfo().getWithOffset(4),
11314 isVolatile, isNonTemporal,
11315 Alignment, AAInfo);
11316 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
11325 // Try to infer better alignment information than the store already has.
11326 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
11327 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
11328 if (Align > ST->getAlignment()) {
11330 DAG.getTruncStore(Chain, SDLoc(N), Value,
11331 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
11332 ST->isVolatile(), ST->isNonTemporal(), Align,
11334 if (NewStore.getNode() != N)
11335 return CombineTo(ST, NewStore, true);
11340 // Try transforming a pair floating point load / store ops to integer
11341 // load / store ops.
11342 if (SDValue NewST = TransformFPLoadStorePair(N))
11345 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
11346 : DAG.getSubtarget().useAA();
11348 if (CombinerAAOnlyFunc.getNumOccurrences() &&
11349 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
11352 if (UseAA && ST->isUnindexed()) {
11353 // Walk up chain skipping non-aliasing memory nodes.
11354 SDValue BetterChain = FindBetterChain(N, Chain);
11356 // If there is a better chain.
11357 if (Chain != BetterChain) {
11360 // Replace the chain to avoid dependency.
11361 if (ST->isTruncatingStore()) {
11362 ReplStore = DAG.getTruncStore(BetterChain, SDLoc(N), Value, Ptr,
11363 ST->getMemoryVT(), ST->getMemOperand());
11365 ReplStore = DAG.getStore(BetterChain, SDLoc(N), Value, Ptr,
11366 ST->getMemOperand());
11369 // Create token to keep both nodes around.
11370 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
11371 MVT::Other, Chain, ReplStore);
11373 // Make sure the new and old chains are cleaned up.
11374 AddToWorklist(Token.getNode());
11376 // Don't add users to work list.
11377 return CombineTo(N, Token, false);
11381 // Try transforming N to an indexed store.
11382 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
11383 return SDValue(N, 0);
11385 // FIXME: is there such a thing as a truncating indexed store?
11386 if (ST->isTruncatingStore() && ST->isUnindexed() &&
11387 Value.getValueType().isInteger()) {
11388 // See if we can simplify the input to this truncstore with knowledge that
11389 // only the low bits are being used. For example:
11390 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
11392 GetDemandedBits(Value,
11393 APInt::getLowBitsSet(
11394 Value.getValueType().getScalarType().getSizeInBits(),
11395 ST->getMemoryVT().getScalarType().getSizeInBits()));
11396 AddToWorklist(Value.getNode());
11397 if (Shorter.getNode())
11398 return DAG.getTruncStore(Chain, SDLoc(N), Shorter,
11399 Ptr, ST->getMemoryVT(), ST->getMemOperand());
11401 // Otherwise, see if we can simplify the operation with
11402 // SimplifyDemandedBits, which only works if the value has a single use.
11403 if (SimplifyDemandedBits(Value,
11404 APInt::getLowBitsSet(
11405 Value.getValueType().getScalarType().getSizeInBits(),
11406 ST->getMemoryVT().getScalarType().getSizeInBits())))
11407 return SDValue(N, 0);
11410 // If this is a load followed by a store to the same location, then the store
11412 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
11413 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
11414 ST->isUnindexed() && !ST->isVolatile() &&
11415 // There can't be any side effects between the load and store, such as
11416 // a call or store.
11417 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
11418 // The store is dead, remove it.
11423 // If this is a store followed by a store with the same value to the same
11424 // location, then the store is dead/noop.
11425 if (StoreSDNode *ST1 = dyn_cast<StoreSDNode>(Chain)) {
11426 if (ST1->getBasePtr() == Ptr && ST->getMemoryVT() == ST1->getMemoryVT() &&
11427 ST1->getValue() == Value && ST->isUnindexed() && !ST->isVolatile() &&
11428 ST1->isUnindexed() && !ST1->isVolatile()) {
11429 // The store is dead, remove it.
11434 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
11435 // truncating store. We can do this even if this is already a truncstore.
11436 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
11437 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
11438 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
11439 ST->getMemoryVT())) {
11440 return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0),
11441 Ptr, ST->getMemoryVT(), ST->getMemOperand());
11444 // Only perform this optimization before the types are legal, because we
11445 // don't want to perform this optimization on every DAGCombine invocation.
11447 bool EverChanged = false;
11450 // There can be multiple store sequences on the same chain.
11451 // Keep trying to merge store sequences until we are unable to do so
11452 // or until we merge the last store on the chain.
11453 bool Changed = MergeConsecutiveStores(ST);
11454 EverChanged |= Changed;
11455 if (!Changed) break;
11456 } while (ST->getOpcode() != ISD::DELETED_NODE);
11459 return SDValue(N, 0);
11462 return ReduceLoadOpStoreWidth(N);
11465 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
11466 SDValue InVec = N->getOperand(0);
11467 SDValue InVal = N->getOperand(1);
11468 SDValue EltNo = N->getOperand(2);
11471 // If the inserted element is an UNDEF, just use the input vector.
11472 if (InVal.getOpcode() == ISD::UNDEF)
11475 EVT VT = InVec.getValueType();
11477 // If we can't generate a legal BUILD_VECTOR, exit
11478 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
11481 // Check that we know which element is being inserted
11482 if (!isa<ConstantSDNode>(EltNo))
11484 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
11486 // Canonicalize insert_vector_elt dag nodes.
11488 // (insert_vector_elt (insert_vector_elt A, Idx0), Idx1)
11489 // -> (insert_vector_elt (insert_vector_elt A, Idx1), Idx0)
11491 // Do this only if the child insert_vector node has one use; also
11492 // do this only if indices are both constants and Idx1 < Idx0.
11493 if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT && InVec.hasOneUse()
11494 && isa<ConstantSDNode>(InVec.getOperand(2))) {
11495 unsigned OtherElt =
11496 cast<ConstantSDNode>(InVec.getOperand(2))->getZExtValue();
11497 if (Elt < OtherElt) {
11499 SDValue NewOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(N), VT,
11500 InVec.getOperand(0), InVal, EltNo);
11501 AddToWorklist(NewOp.getNode());
11502 return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(InVec.getNode()),
11503 VT, NewOp, InVec.getOperand(1), InVec.getOperand(2));
11507 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
11508 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
11509 // vector elements.
11510 SmallVector<SDValue, 8> Ops;
11511 // Do not combine these two vectors if the output vector will not replace
11512 // the input vector.
11513 if (InVec.getOpcode() == ISD::BUILD_VECTOR && InVec.hasOneUse()) {
11514 Ops.append(InVec.getNode()->op_begin(),
11515 InVec.getNode()->op_end());
11516 } else if (InVec.getOpcode() == ISD::UNDEF) {
11517 unsigned NElts = VT.getVectorNumElements();
11518 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
11523 // Insert the element
11524 if (Elt < Ops.size()) {
11525 // All the operands of BUILD_VECTOR must have the same type;
11526 // we enforce that here.
11527 EVT OpVT = Ops[0].getValueType();
11528 if (InVal.getValueType() != OpVT)
11529 InVal = OpVT.bitsGT(InVal.getValueType()) ?
11530 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
11531 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
11535 // Return the new vector
11536 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
11539 SDValue DAGCombiner::ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
11540 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad) {
11541 EVT ResultVT = EVE->getValueType(0);
11542 EVT VecEltVT = InVecVT.getVectorElementType();
11543 unsigned Align = OriginalLoad->getAlignment();
11544 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
11545 VecEltVT.getTypeForEVT(*DAG.getContext()));
11547 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VecEltVT))
11552 SDValue NewPtr = OriginalLoad->getBasePtr();
11554 EVT PtrType = NewPtr.getValueType();
11555 MachinePointerInfo MPI;
11557 if (auto *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo)) {
11558 int Elt = ConstEltNo->getZExtValue();
11559 unsigned PtrOff = VecEltVT.getSizeInBits() * Elt / 8;
11560 Offset = DAG.getConstant(PtrOff, DL, PtrType);
11561 MPI = OriginalLoad->getPointerInfo().getWithOffset(PtrOff);
11563 Offset = DAG.getZExtOrTrunc(EltNo, DL, PtrType);
11564 Offset = DAG.getNode(
11565 ISD::MUL, DL, PtrType, Offset,
11566 DAG.getConstant(VecEltVT.getStoreSize(), DL, PtrType));
11567 MPI = OriginalLoad->getPointerInfo();
11569 NewPtr = DAG.getNode(ISD::ADD, DL, PtrType, NewPtr, Offset);
11571 // The replacement we need to do here is a little tricky: we need to
11572 // replace an extractelement of a load with a load.
11573 // Use ReplaceAllUsesOfValuesWith to do the replacement.
11574 // Note that this replacement assumes that the extractvalue is the only
11575 // use of the load; that's okay because we don't want to perform this
11576 // transformation in other cases anyway.
11579 if (ResultVT.bitsGT(VecEltVT)) {
11580 // If the result type of vextract is wider than the load, then issue an
11581 // extending load instead.
11582 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, ResultVT,
11586 Load = DAG.getExtLoad(
11587 ExtType, SDLoc(EVE), ResultVT, OriginalLoad->getChain(), NewPtr, MPI,
11588 VecEltVT, OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
11589 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
11590 Chain = Load.getValue(1);
11592 Load = DAG.getLoad(
11593 VecEltVT, SDLoc(EVE), OriginalLoad->getChain(), NewPtr, MPI,
11594 OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
11595 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
11596 Chain = Load.getValue(1);
11597 if (ResultVT.bitsLT(VecEltVT))
11598 Load = DAG.getNode(ISD::TRUNCATE, SDLoc(EVE), ResultVT, Load);
11600 Load = DAG.getNode(ISD::BITCAST, SDLoc(EVE), ResultVT, Load);
11602 WorklistRemover DeadNodes(*this);
11603 SDValue From[] = { SDValue(EVE, 0), SDValue(OriginalLoad, 1) };
11604 SDValue To[] = { Load, Chain };
11605 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
11606 // Since we're explicitly calling ReplaceAllUses, add the new node to the
11607 // worklist explicitly as well.
11608 AddToWorklist(Load.getNode());
11609 AddUsersToWorklist(Load.getNode()); // Add users too
11610 // Make sure to revisit this node to clean it up; it will usually be dead.
11611 AddToWorklist(EVE);
11613 return SDValue(EVE, 0);
11616 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
11617 // (vextract (scalar_to_vector val, 0) -> val
11618 SDValue InVec = N->getOperand(0);
11619 EVT VT = InVec.getValueType();
11620 EVT NVT = N->getValueType(0);
11622 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
11623 // Check if the result type doesn't match the inserted element type. A
11624 // SCALAR_TO_VECTOR may truncate the inserted element and the
11625 // EXTRACT_VECTOR_ELT may widen the extracted vector.
11626 SDValue InOp = InVec.getOperand(0);
11627 if (InOp.getValueType() != NVT) {
11628 assert(InOp.getValueType().isInteger() && NVT.isInteger());
11629 return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT);
11634 SDValue EltNo = N->getOperand(1);
11635 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
11637 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
11638 // We only perform this optimization before the op legalization phase because
11639 // we may introduce new vector instructions which are not backed by TD
11640 // patterns. For example on AVX, extracting elements from a wide vector
11641 // without using extract_subvector. However, if we can find an underlying
11642 // scalar value, then we can always use that.
11643 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
11645 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
11646 int NumElem = VT.getVectorNumElements();
11647 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
11648 // Find the new index to extract from.
11649 int OrigElt = SVOp->getMaskElt(Elt);
11651 // Extracting an undef index is undef.
11653 return DAG.getUNDEF(NVT);
11655 // Select the right vector half to extract from.
11657 if (OrigElt < NumElem) {
11658 SVInVec = InVec->getOperand(0);
11660 SVInVec = InVec->getOperand(1);
11661 OrigElt -= NumElem;
11664 if (SVInVec.getOpcode() == ISD::BUILD_VECTOR) {
11665 SDValue InOp = SVInVec.getOperand(OrigElt);
11666 if (InOp.getValueType() != NVT) {
11667 assert(InOp.getValueType().isInteger() && NVT.isInteger());
11668 InOp = DAG.getSExtOrTrunc(InOp, SDLoc(SVInVec), NVT);
11674 // FIXME: We should handle recursing on other vector shuffles and
11675 // scalar_to_vector here as well.
11677 if (!LegalOperations) {
11678 EVT IndexTy = TLI.getVectorIdxTy(DAG.getDataLayout());
11679 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT, SVInVec,
11680 DAG.getConstant(OrigElt, SDLoc(SVOp), IndexTy));
11684 bool BCNumEltsChanged = false;
11685 EVT ExtVT = VT.getVectorElementType();
11688 // If the result of load has to be truncated, then it's not necessarily
11690 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
11693 if (InVec.getOpcode() == ISD::BITCAST) {
11694 // Don't duplicate a load with other uses.
11695 if (!InVec.hasOneUse())
11698 EVT BCVT = InVec.getOperand(0).getValueType();
11699 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
11701 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
11702 BCNumEltsChanged = true;
11703 InVec = InVec.getOperand(0);
11704 ExtVT = BCVT.getVectorElementType();
11707 // (vextract (vN[if]M load $addr), i) -> ([if]M load $addr + i * size)
11708 if (!LegalOperations && !ConstEltNo && InVec.hasOneUse() &&
11709 ISD::isNormalLoad(InVec.getNode()) &&
11710 !N->getOperand(1)->hasPredecessor(InVec.getNode())) {
11711 SDValue Index = N->getOperand(1);
11712 if (LoadSDNode *OrigLoad = dyn_cast<LoadSDNode>(InVec))
11713 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, Index,
11717 // Perform only after legalization to ensure build_vector / vector_shuffle
11718 // optimizations have already been done.
11719 if (!LegalOperations) return SDValue();
11721 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
11722 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
11723 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
11726 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
11728 LoadSDNode *LN0 = nullptr;
11729 const ShuffleVectorSDNode *SVN = nullptr;
11730 if (ISD::isNormalLoad(InVec.getNode())) {
11731 LN0 = cast<LoadSDNode>(InVec);
11732 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
11733 InVec.getOperand(0).getValueType() == ExtVT &&
11734 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
11735 // Don't duplicate a load with other uses.
11736 if (!InVec.hasOneUse())
11739 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
11740 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
11741 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
11743 // (load $addr+1*size)
11745 // Don't duplicate a load with other uses.
11746 if (!InVec.hasOneUse())
11749 // If the bit convert changed the number of elements, it is unsafe
11750 // to examine the mask.
11751 if (BCNumEltsChanged)
11754 // Select the input vector, guarding against out of range extract vector.
11755 unsigned NumElems = VT.getVectorNumElements();
11756 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
11757 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
11759 if (InVec.getOpcode() == ISD::BITCAST) {
11760 // Don't duplicate a load with other uses.
11761 if (!InVec.hasOneUse())
11764 InVec = InVec.getOperand(0);
11766 if (ISD::isNormalLoad(InVec.getNode())) {
11767 LN0 = cast<LoadSDNode>(InVec);
11768 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
11769 EltNo = DAG.getConstant(Elt, SDLoc(EltNo), EltNo.getValueType());
11773 // Make sure we found a non-volatile load and the extractelement is
11775 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
11778 // If Idx was -1 above, Elt is going to be -1, so just return undef.
11780 return DAG.getUNDEF(LVT);
11782 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, EltNo, LN0);
11788 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
11789 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
11790 // We perform this optimization post type-legalization because
11791 // the type-legalizer often scalarizes integer-promoted vectors.
11792 // Performing this optimization before may create bit-casts which
11793 // will be type-legalized to complex code sequences.
11794 // We perform this optimization only before the operation legalizer because we
11795 // may introduce illegal operations.
11796 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
11799 unsigned NumInScalars = N->getNumOperands();
11801 EVT VT = N->getValueType(0);
11803 // Check to see if this is a BUILD_VECTOR of a bunch of values
11804 // which come from any_extend or zero_extend nodes. If so, we can create
11805 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
11806 // optimizations. We do not handle sign-extend because we can't fill the sign
11808 EVT SourceType = MVT::Other;
11809 bool AllAnyExt = true;
11811 for (unsigned i = 0; i != NumInScalars; ++i) {
11812 SDValue In = N->getOperand(i);
11813 // Ignore undef inputs.
11814 if (In.getOpcode() == ISD::UNDEF) continue;
11816 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
11817 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
11819 // Abort if the element is not an extension.
11820 if (!ZeroExt && !AnyExt) {
11821 SourceType = MVT::Other;
11825 // The input is a ZeroExt or AnyExt. Check the original type.
11826 EVT InTy = In.getOperand(0).getValueType();
11828 // Check that all of the widened source types are the same.
11829 if (SourceType == MVT::Other)
11832 else if (InTy != SourceType) {
11833 // Multiple income types. Abort.
11834 SourceType = MVT::Other;
11838 // Check if all of the extends are ANY_EXTENDs.
11839 AllAnyExt &= AnyExt;
11842 // In order to have valid types, all of the inputs must be extended from the
11843 // same source type and all of the inputs must be any or zero extend.
11844 // Scalar sizes must be a power of two.
11845 EVT OutScalarTy = VT.getScalarType();
11846 bool ValidTypes = SourceType != MVT::Other &&
11847 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
11848 isPowerOf2_32(SourceType.getSizeInBits());
11850 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
11851 // turn into a single shuffle instruction.
11855 bool isLE = DAG.getDataLayout().isLittleEndian();
11856 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
11857 assert(ElemRatio > 1 && "Invalid element size ratio");
11858 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
11859 DAG.getConstant(0, SDLoc(N), SourceType);
11861 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
11862 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
11864 // Populate the new build_vector
11865 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
11866 SDValue Cast = N->getOperand(i);
11867 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
11868 Cast.getOpcode() == ISD::ZERO_EXTEND ||
11869 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
11871 if (Cast.getOpcode() == ISD::UNDEF)
11872 In = DAG.getUNDEF(SourceType);
11874 In = Cast->getOperand(0);
11875 unsigned Index = isLE ? (i * ElemRatio) :
11876 (i * ElemRatio + (ElemRatio - 1));
11878 assert(Index < Ops.size() && "Invalid index");
11882 // The type of the new BUILD_VECTOR node.
11883 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
11884 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
11885 "Invalid vector size");
11886 // Check if the new vector type is legal.
11887 if (!isTypeLegal(VecVT)) return SDValue();
11889 // Make the new BUILD_VECTOR.
11890 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
11892 // The new BUILD_VECTOR node has the potential to be further optimized.
11893 AddToWorklist(BV.getNode());
11894 // Bitcast to the desired type.
11895 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
11898 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
11899 EVT VT = N->getValueType(0);
11901 unsigned NumInScalars = N->getNumOperands();
11904 EVT SrcVT = MVT::Other;
11905 unsigned Opcode = ISD::DELETED_NODE;
11906 unsigned NumDefs = 0;
11908 for (unsigned i = 0; i != NumInScalars; ++i) {
11909 SDValue In = N->getOperand(i);
11910 unsigned Opc = In.getOpcode();
11912 if (Opc == ISD::UNDEF)
11915 // If all scalar values are floats and converted from integers.
11916 if (Opcode == ISD::DELETED_NODE &&
11917 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
11924 EVT InVT = In.getOperand(0).getValueType();
11926 // If all scalar values are typed differently, bail out. It's chosen to
11927 // simplify BUILD_VECTOR of integer types.
11928 if (SrcVT == MVT::Other)
11935 // If the vector has just one element defined, it's not worth to fold it into
11936 // a vectorized one.
11940 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
11941 && "Should only handle conversion from integer to float.");
11942 assert(SrcVT != MVT::Other && "Cannot determine source type!");
11944 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
11946 if (!TLI.isOperationLegalOrCustom(Opcode, NVT))
11949 // Just because the floating-point vector type is legal does not necessarily
11950 // mean that the corresponding integer vector type is.
11951 if (!isTypeLegal(NVT))
11954 SmallVector<SDValue, 8> Opnds;
11955 for (unsigned i = 0; i != NumInScalars; ++i) {
11956 SDValue In = N->getOperand(i);
11958 if (In.getOpcode() == ISD::UNDEF)
11959 Opnds.push_back(DAG.getUNDEF(SrcVT));
11961 Opnds.push_back(In.getOperand(0));
11963 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Opnds);
11964 AddToWorklist(BV.getNode());
11966 return DAG.getNode(Opcode, dl, VT, BV);
11969 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
11970 unsigned NumInScalars = N->getNumOperands();
11972 EVT VT = N->getValueType(0);
11974 // A vector built entirely of undefs is undef.
11975 if (ISD::allOperandsUndef(N))
11976 return DAG.getUNDEF(VT);
11978 if (SDValue V = reduceBuildVecExtToExtBuildVec(N))
11981 if (SDValue V = reduceBuildVecConvertToConvertBuildVec(N))
11984 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
11985 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
11986 // at most two distinct vectors, turn this into a shuffle node.
11988 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
11989 if (!isTypeLegal(VT))
11992 // May only combine to shuffle after legalize if shuffle is legal.
11993 if (LegalOperations && !TLI.isOperationLegal(ISD::VECTOR_SHUFFLE, VT))
11996 SDValue VecIn1, VecIn2;
11997 bool UsesZeroVector = false;
11998 for (unsigned i = 0; i != NumInScalars; ++i) {
11999 SDValue Op = N->getOperand(i);
12000 // Ignore undef inputs.
12001 if (Op.getOpcode() == ISD::UNDEF) continue;
12003 // See if we can combine this build_vector into a blend with a zero vector.
12004 if (!VecIn2.getNode() && (isNullConstant(Op) || isNullFPConstant(Op))) {
12005 UsesZeroVector = true;
12009 // If this input is something other than a EXTRACT_VECTOR_ELT with a
12010 // constant index, bail out.
12011 if (Op.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
12012 !isa<ConstantSDNode>(Op.getOperand(1))) {
12013 VecIn1 = VecIn2 = SDValue(nullptr, 0);
12017 // We allow up to two distinct input vectors.
12018 SDValue ExtractedFromVec = Op.getOperand(0);
12019 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
12022 if (!VecIn1.getNode()) {
12023 VecIn1 = ExtractedFromVec;
12024 } else if (!VecIn2.getNode() && !UsesZeroVector) {
12025 VecIn2 = ExtractedFromVec;
12027 // Too many inputs.
12028 VecIn1 = VecIn2 = SDValue(nullptr, 0);
12033 // If everything is good, we can make a shuffle operation.
12034 if (VecIn1.getNode()) {
12035 unsigned InNumElements = VecIn1.getValueType().getVectorNumElements();
12036 SmallVector<int, 8> Mask;
12037 for (unsigned i = 0; i != NumInScalars; ++i) {
12038 unsigned Opcode = N->getOperand(i).getOpcode();
12039 if (Opcode == ISD::UNDEF) {
12040 Mask.push_back(-1);
12044 // Operands can also be zero.
12045 if (Opcode != ISD::EXTRACT_VECTOR_ELT) {
12046 assert(UsesZeroVector &&
12047 (Opcode == ISD::Constant || Opcode == ISD::ConstantFP) &&
12048 "Unexpected node found!");
12049 Mask.push_back(NumInScalars+i);
12053 // If extracting from the first vector, just use the index directly.
12054 SDValue Extract = N->getOperand(i);
12055 SDValue ExtVal = Extract.getOperand(1);
12056 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
12057 if (Extract.getOperand(0) == VecIn1) {
12058 Mask.push_back(ExtIndex);
12062 // Otherwise, use InIdx + InputVecSize
12063 Mask.push_back(InNumElements + ExtIndex);
12066 // Avoid introducing illegal shuffles with zero.
12067 if (UsesZeroVector && !TLI.isVectorClearMaskLegal(Mask, VT))
12070 // We can't generate a shuffle node with mismatched input and output types.
12071 // Attempt to transform a single input vector to the correct type.
12072 if ((VT != VecIn1.getValueType())) {
12073 // If the input vector type has a different base type to the output
12074 // vector type, bail out.
12075 EVT VTElemType = VT.getVectorElementType();
12076 if ((VecIn1.getValueType().getVectorElementType() != VTElemType) ||
12077 (VecIn2.getNode() &&
12078 (VecIn2.getValueType().getVectorElementType() != VTElemType)))
12081 // If the input vector is too small, widen it.
12082 // We only support widening of vectors which are half the size of the
12083 // output registers. For example XMM->YMM widening on X86 with AVX.
12084 EVT VecInT = VecIn1.getValueType();
12085 if (VecInT.getSizeInBits() * 2 == VT.getSizeInBits()) {
12086 // If we only have one small input, widen it by adding undef values.
12087 if (!VecIn2.getNode())
12088 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, VecIn1,
12089 DAG.getUNDEF(VecIn1.getValueType()));
12090 else if (VecIn1.getValueType() == VecIn2.getValueType()) {
12091 // If we have two small inputs of the same type, try to concat them.
12092 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, VecIn1, VecIn2);
12093 VecIn2 = SDValue(nullptr, 0);
12096 } else if (VecInT.getSizeInBits() == VT.getSizeInBits() * 2) {
12097 // If the input vector is too large, try to split it.
12098 // We don't support having two input vectors that are too large.
12099 // If the zero vector was used, we can not split the vector,
12100 // since we'd need 3 inputs.
12101 if (UsesZeroVector || VecIn2.getNode())
12104 if (!TLI.isExtractSubvectorCheap(VT, VT.getVectorNumElements()))
12107 // Try to replace VecIn1 with two extract_subvectors
12108 // No need to update the masks, they should still be correct.
12109 VecIn2 = DAG.getNode(
12110 ISD::EXTRACT_SUBVECTOR, dl, VT, VecIn1,
12111 DAG.getConstant(VT.getVectorNumElements(), dl,
12112 TLI.getVectorIdxTy(DAG.getDataLayout())));
12113 VecIn1 = DAG.getNode(
12114 ISD::EXTRACT_SUBVECTOR, dl, VT, VecIn1,
12115 DAG.getConstant(0, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
12120 if (UsesZeroVector)
12121 VecIn2 = VT.isInteger() ? DAG.getConstant(0, dl, VT) :
12122 DAG.getConstantFP(0.0, dl, VT);
12124 // If VecIn2 is unused then change it to undef.
12125 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
12127 // Check that we were able to transform all incoming values to the same
12129 if (VecIn2.getValueType() != VecIn1.getValueType() ||
12130 VecIn1.getValueType() != VT)
12133 // Return the new VECTOR_SHUFFLE node.
12137 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]);
12143 static SDValue combineConcatVectorOfScalars(SDNode *N, SelectionDAG &DAG) {
12144 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12145 EVT OpVT = N->getOperand(0).getValueType();
12147 // If the operands are legal vectors, leave them alone.
12148 if (TLI.isTypeLegal(OpVT))
12152 EVT VT = N->getValueType(0);
12153 SmallVector<SDValue, 8> Ops;
12155 EVT SVT = EVT::getIntegerVT(*DAG.getContext(), OpVT.getSizeInBits());
12156 SDValue ScalarUndef = DAG.getNode(ISD::UNDEF, DL, SVT);
12158 // Keep track of what we encounter.
12159 bool AnyInteger = false;
12160 bool AnyFP = false;
12161 for (const SDValue &Op : N->ops()) {
12162 if (ISD::BITCAST == Op.getOpcode() &&
12163 !Op.getOperand(0).getValueType().isVector())
12164 Ops.push_back(Op.getOperand(0));
12165 else if (ISD::UNDEF == Op.getOpcode())
12166 Ops.push_back(ScalarUndef);
12170 // Note whether we encounter an integer or floating point scalar.
12171 // If it's neither, bail out, it could be something weird like x86mmx.
12172 EVT LastOpVT = Ops.back().getValueType();
12173 if (LastOpVT.isFloatingPoint())
12175 else if (LastOpVT.isInteger())
12181 // If any of the operands is a floating point scalar bitcast to a vector,
12182 // use floating point types throughout, and bitcast everything.
12183 // Replace UNDEFs by another scalar UNDEF node, of the final desired type.
12185 SVT = EVT::getFloatingPointVT(OpVT.getSizeInBits());
12186 ScalarUndef = DAG.getNode(ISD::UNDEF, DL, SVT);
12188 for (SDValue &Op : Ops) {
12189 if (Op.getValueType() == SVT)
12191 if (Op.getOpcode() == ISD::UNDEF)
12194 Op = DAG.getNode(ISD::BITCAST, DL, SVT, Op);
12199 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SVT,
12200 VT.getSizeInBits() / SVT.getSizeInBits());
12201 return DAG.getNode(ISD::BITCAST, DL, VT,
12202 DAG.getNode(ISD::BUILD_VECTOR, DL, VecVT, Ops));
12205 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
12206 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
12207 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
12208 // inputs come from at most two distinct vectors, turn this into a shuffle
12211 // If we only have one input vector, we don't need to do any concatenation.
12212 if (N->getNumOperands() == 1)
12213 return N->getOperand(0);
12215 // Check if all of the operands are undefs.
12216 EVT VT = N->getValueType(0);
12217 if (ISD::allOperandsUndef(N))
12218 return DAG.getUNDEF(VT);
12220 // Optimize concat_vectors where all but the first of the vectors are undef.
12221 if (std::all_of(std::next(N->op_begin()), N->op_end(), [](const SDValue &Op) {
12222 return Op.getOpcode() == ISD::UNDEF;
12224 SDValue In = N->getOperand(0);
12225 assert(In.getValueType().isVector() && "Must concat vectors");
12227 // Transform: concat_vectors(scalar, undef) -> scalar_to_vector(sclr).
12228 if (In->getOpcode() == ISD::BITCAST &&
12229 !In->getOperand(0)->getValueType(0).isVector()) {
12230 SDValue Scalar = In->getOperand(0);
12232 // If the bitcast type isn't legal, it might be a trunc of a legal type;
12233 // look through the trunc so we can still do the transform:
12234 // concat_vectors(trunc(scalar), undef) -> scalar_to_vector(scalar)
12235 if (Scalar->getOpcode() == ISD::TRUNCATE &&
12236 !TLI.isTypeLegal(Scalar.getValueType()) &&
12237 TLI.isTypeLegal(Scalar->getOperand(0).getValueType()))
12238 Scalar = Scalar->getOperand(0);
12240 EVT SclTy = Scalar->getValueType(0);
12242 if (!SclTy.isFloatingPoint() && !SclTy.isInteger())
12245 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SclTy,
12246 VT.getSizeInBits() / SclTy.getSizeInBits());
12247 if (!TLI.isTypeLegal(NVT) || !TLI.isTypeLegal(Scalar.getValueType()))
12250 SDLoc dl = SDLoc(N);
12251 SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NVT, Scalar);
12252 return DAG.getNode(ISD::BITCAST, dl, VT, Res);
12256 // Fold any combination of BUILD_VECTOR or UNDEF nodes into one BUILD_VECTOR.
12257 // We have already tested above for an UNDEF only concatenation.
12258 // fold (concat_vectors (BUILD_VECTOR A, B, ...), (BUILD_VECTOR C, D, ...))
12259 // -> (BUILD_VECTOR A, B, ..., C, D, ...)
12260 auto IsBuildVectorOrUndef = [](const SDValue &Op) {
12261 return ISD::UNDEF == Op.getOpcode() || ISD::BUILD_VECTOR == Op.getOpcode();
12263 bool AllBuildVectorsOrUndefs =
12264 std::all_of(N->op_begin(), N->op_end(), IsBuildVectorOrUndef);
12265 if (AllBuildVectorsOrUndefs) {
12266 SmallVector<SDValue, 8> Opnds;
12267 EVT SVT = VT.getScalarType();
12270 if (!SVT.isFloatingPoint()) {
12271 // If BUILD_VECTOR are from built from integer, they may have different
12272 // operand types. Get the smallest type and truncate all operands to it.
12273 bool FoundMinVT = false;
12274 for (const SDValue &Op : N->ops())
12275 if (ISD::BUILD_VECTOR == Op.getOpcode()) {
12276 EVT OpSVT = Op.getOperand(0)->getValueType(0);
12277 MinVT = (!FoundMinVT || OpSVT.bitsLE(MinVT)) ? OpSVT : MinVT;
12280 assert(FoundMinVT && "Concat vector type mismatch");
12283 for (const SDValue &Op : N->ops()) {
12284 EVT OpVT = Op.getValueType();
12285 unsigned NumElts = OpVT.getVectorNumElements();
12287 if (ISD::UNDEF == Op.getOpcode())
12288 Opnds.append(NumElts, DAG.getUNDEF(MinVT));
12290 if (ISD::BUILD_VECTOR == Op.getOpcode()) {
12291 if (SVT.isFloatingPoint()) {
12292 assert(SVT == OpVT.getScalarType() && "Concat vector type mismatch");
12293 Opnds.append(Op->op_begin(), Op->op_begin() + NumElts);
12295 for (unsigned i = 0; i != NumElts; ++i)
12297 DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinVT, Op.getOperand(i)));
12302 assert(VT.getVectorNumElements() == Opnds.size() &&
12303 "Concat vector type mismatch");
12304 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
12307 // Fold CONCAT_VECTORS of only bitcast scalars (or undef) to BUILD_VECTOR.
12308 if (SDValue V = combineConcatVectorOfScalars(N, DAG))
12311 // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
12312 // nodes often generate nop CONCAT_VECTOR nodes.
12313 // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that
12314 // place the incoming vectors at the exact same location.
12315 SDValue SingleSource = SDValue();
12316 unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements();
12318 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
12319 SDValue Op = N->getOperand(i);
12321 if (Op.getOpcode() == ISD::UNDEF)
12324 // Check if this is the identity extract:
12325 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
12328 // Find the single incoming vector for the extract_subvector.
12329 if (SingleSource.getNode()) {
12330 if (Op.getOperand(0) != SingleSource)
12333 SingleSource = Op.getOperand(0);
12335 // Check the source type is the same as the type of the result.
12336 // If not, this concat may extend the vector, so we can not
12337 // optimize it away.
12338 if (SingleSource.getValueType() != N->getValueType(0))
12342 unsigned IdentityIndex = i * PartNumElem;
12343 ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1));
12344 // The extract index must be constant.
12348 // Check that we are reading from the identity index.
12349 if (CS->getZExtValue() != IdentityIndex)
12353 if (SingleSource.getNode())
12354 return SingleSource;
12359 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
12360 EVT NVT = N->getValueType(0);
12361 SDValue V = N->getOperand(0);
12363 if (V->getOpcode() == ISD::CONCAT_VECTORS) {
12365 // (extract_subvec (concat V1, V2, ...), i)
12368 // Only operand 0 is checked as 'concat' assumes all inputs of the same
12370 if (V->getOperand(0).getValueType() != NVT)
12372 unsigned Idx = N->getConstantOperandVal(1);
12373 unsigned NumElems = NVT.getVectorNumElements();
12374 assert((Idx % NumElems) == 0 &&
12375 "IDX in concat is not a multiple of the result vector length.");
12376 return V->getOperand(Idx / NumElems);
12380 if (V->getOpcode() == ISD::BITCAST)
12381 V = V.getOperand(0);
12383 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
12385 // Handle only simple case where vector being inserted and vector
12386 // being extracted are of same type, and are half size of larger vectors.
12387 EVT BigVT = V->getOperand(0).getValueType();
12388 EVT SmallVT = V->getOperand(1).getValueType();
12389 if (!NVT.bitsEq(SmallVT) || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
12392 // Only handle cases where both indexes are constants with the same type.
12393 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
12394 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
12396 if (InsIdx && ExtIdx &&
12397 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
12398 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
12400 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
12402 // indices are equal or bit offsets are equal => V1
12403 // otherwise => (extract_subvec V1, ExtIdx)
12404 if (InsIdx->getZExtValue() * SmallVT.getScalarType().getSizeInBits() ==
12405 ExtIdx->getZExtValue() * NVT.getScalarType().getSizeInBits())
12406 return DAG.getNode(ISD::BITCAST, dl, NVT, V->getOperand(1));
12407 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT,
12408 DAG.getNode(ISD::BITCAST, dl,
12409 N->getOperand(0).getValueType(),
12410 V->getOperand(0)), N->getOperand(1));
12417 static SDValue simplifyShuffleOperandRecursively(SmallBitVector &UsedElements,
12418 SDValue V, SelectionDAG &DAG) {
12420 EVT VT = V.getValueType();
12422 switch (V.getOpcode()) {
12426 case ISD::CONCAT_VECTORS: {
12427 EVT OpVT = V->getOperand(0).getValueType();
12428 int OpSize = OpVT.getVectorNumElements();
12429 SmallBitVector OpUsedElements(OpSize, false);
12430 bool FoundSimplification = false;
12431 SmallVector<SDValue, 4> NewOps;
12432 NewOps.reserve(V->getNumOperands());
12433 for (int i = 0, NumOps = V->getNumOperands(); i < NumOps; ++i) {
12434 SDValue Op = V->getOperand(i);
12435 bool OpUsed = false;
12436 for (int j = 0; j < OpSize; ++j)
12437 if (UsedElements[i * OpSize + j]) {
12438 OpUsedElements[j] = true;
12442 OpUsed ? simplifyShuffleOperandRecursively(OpUsedElements, Op, DAG)
12443 : DAG.getUNDEF(OpVT));
12444 FoundSimplification |= Op == NewOps.back();
12445 OpUsedElements.reset();
12447 if (FoundSimplification)
12448 V = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, NewOps);
12452 case ISD::INSERT_SUBVECTOR: {
12453 SDValue BaseV = V->getOperand(0);
12454 SDValue SubV = V->getOperand(1);
12455 auto *IdxN = dyn_cast<ConstantSDNode>(V->getOperand(2));
12459 int SubSize = SubV.getValueType().getVectorNumElements();
12460 int Idx = IdxN->getZExtValue();
12461 bool SubVectorUsed = false;
12462 SmallBitVector SubUsedElements(SubSize, false);
12463 for (int i = 0; i < SubSize; ++i)
12464 if (UsedElements[i + Idx]) {
12465 SubVectorUsed = true;
12466 SubUsedElements[i] = true;
12467 UsedElements[i + Idx] = false;
12470 // Now recurse on both the base and sub vectors.
12471 SDValue SimplifiedSubV =
12473 ? simplifyShuffleOperandRecursively(SubUsedElements, SubV, DAG)
12474 : DAG.getUNDEF(SubV.getValueType());
12475 SDValue SimplifiedBaseV = simplifyShuffleOperandRecursively(UsedElements, BaseV, DAG);
12476 if (SimplifiedSubV != SubV || SimplifiedBaseV != BaseV)
12477 V = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT,
12478 SimplifiedBaseV, SimplifiedSubV, V->getOperand(2));
12484 static SDValue simplifyShuffleOperands(ShuffleVectorSDNode *SVN, SDValue N0,
12485 SDValue N1, SelectionDAG &DAG) {
12486 EVT VT = SVN->getValueType(0);
12487 int NumElts = VT.getVectorNumElements();
12488 SmallBitVector N0UsedElements(NumElts, false), N1UsedElements(NumElts, false);
12489 for (int M : SVN->getMask())
12490 if (M >= 0 && M < NumElts)
12491 N0UsedElements[M] = true;
12492 else if (M >= NumElts)
12493 N1UsedElements[M - NumElts] = true;
12495 SDValue S0 = simplifyShuffleOperandRecursively(N0UsedElements, N0, DAG);
12496 SDValue S1 = simplifyShuffleOperandRecursively(N1UsedElements, N1, DAG);
12497 if (S0 == N0 && S1 == N1)
12500 return DAG.getVectorShuffle(VT, SDLoc(SVN), S0, S1, SVN->getMask());
12503 // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat,
12504 // or turn a shuffle of a single concat into simpler shuffle then concat.
12505 static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
12506 EVT VT = N->getValueType(0);
12507 unsigned NumElts = VT.getVectorNumElements();
12509 SDValue N0 = N->getOperand(0);
12510 SDValue N1 = N->getOperand(1);
12511 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
12513 SmallVector<SDValue, 4> Ops;
12514 EVT ConcatVT = N0.getOperand(0).getValueType();
12515 unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements();
12516 unsigned NumConcats = NumElts / NumElemsPerConcat;
12518 // Special case: shuffle(concat(A,B)) can be more efficiently represented
12519 // as concat(shuffle(A,B),UNDEF) if the shuffle doesn't set any of the high
12520 // half vector elements.
12521 if (NumElemsPerConcat * 2 == NumElts && N1.getOpcode() == ISD::UNDEF &&
12522 std::all_of(SVN->getMask().begin() + NumElemsPerConcat,
12523 SVN->getMask().end(), [](int i) { return i == -1; })) {
12524 N0 = DAG.getVectorShuffle(ConcatVT, SDLoc(N), N0.getOperand(0), N0.getOperand(1),
12525 ArrayRef<int>(SVN->getMask().begin(), NumElemsPerConcat));
12526 N1 = DAG.getUNDEF(ConcatVT);
12527 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, N0, N1);
12530 // Look at every vector that's inserted. We're looking for exact
12531 // subvector-sized copies from a concatenated vector
12532 for (unsigned I = 0; I != NumConcats; ++I) {
12533 // Make sure we're dealing with a copy.
12534 unsigned Begin = I * NumElemsPerConcat;
12535 bool AllUndef = true, NoUndef = true;
12536 for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) {
12537 if (SVN->getMaskElt(J) >= 0)
12544 if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0)
12547 for (unsigned J = 1; J != NumElemsPerConcat; ++J)
12548 if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J))
12551 unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat;
12552 if (FirstElt < N0.getNumOperands())
12553 Ops.push_back(N0.getOperand(FirstElt));
12555 Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands()));
12557 } else if (AllUndef) {
12558 Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType()));
12559 } else { // Mixed with general masks and undefs, can't do optimization.
12564 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops);
12567 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
12568 EVT VT = N->getValueType(0);
12569 unsigned NumElts = VT.getVectorNumElements();
12571 SDValue N0 = N->getOperand(0);
12572 SDValue N1 = N->getOperand(1);
12574 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
12576 // Canonicalize shuffle undef, undef -> undef
12577 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
12578 return DAG.getUNDEF(VT);
12580 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
12582 // Canonicalize shuffle v, v -> v, undef
12584 SmallVector<int, 8> NewMask;
12585 for (unsigned i = 0; i != NumElts; ++i) {
12586 int Idx = SVN->getMaskElt(i);
12587 if (Idx >= (int)NumElts) Idx -= NumElts;
12588 NewMask.push_back(Idx);
12590 return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT),
12594 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
12595 if (N0.getOpcode() == ISD::UNDEF) {
12596 SmallVector<int, 8> NewMask;
12597 for (unsigned i = 0; i != NumElts; ++i) {
12598 int Idx = SVN->getMaskElt(i);
12600 if (Idx >= (int)NumElts)
12603 Idx = -1; // remove reference to lhs
12605 NewMask.push_back(Idx);
12607 return DAG.getVectorShuffle(VT, SDLoc(N), N1, DAG.getUNDEF(VT),
12611 // Remove references to rhs if it is undef
12612 if (N1.getOpcode() == ISD::UNDEF) {
12613 bool Changed = false;
12614 SmallVector<int, 8> NewMask;
12615 for (unsigned i = 0; i != NumElts; ++i) {
12616 int Idx = SVN->getMaskElt(i);
12617 if (Idx >= (int)NumElts) {
12621 NewMask.push_back(Idx);
12624 return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, &NewMask[0]);
12627 // If it is a splat, check if the argument vector is another splat or a
12629 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
12630 SDNode *V = N0.getNode();
12632 // If this is a bit convert that changes the element type of the vector but
12633 // not the number of vector elements, look through it. Be careful not to
12634 // look though conversions that change things like v4f32 to v2f64.
12635 if (V->getOpcode() == ISD::BITCAST) {
12636 SDValue ConvInput = V->getOperand(0);
12637 if (ConvInput.getValueType().isVector() &&
12638 ConvInput.getValueType().getVectorNumElements() == NumElts)
12639 V = ConvInput.getNode();
12642 if (V->getOpcode() == ISD::BUILD_VECTOR) {
12643 assert(V->getNumOperands() == NumElts &&
12644 "BUILD_VECTOR has wrong number of operands");
12646 bool AllSame = true;
12647 for (unsigned i = 0; i != NumElts; ++i) {
12648 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
12649 Base = V->getOperand(i);
12653 // Splat of <u, u, u, u>, return <u, u, u, u>
12654 if (!Base.getNode())
12656 for (unsigned i = 0; i != NumElts; ++i) {
12657 if (V->getOperand(i) != Base) {
12662 // Splat of <x, x, x, x>, return <x, x, x, x>
12666 // Canonicalize any other splat as a build_vector.
12667 const SDValue &Splatted = V->getOperand(SVN->getSplatIndex());
12668 SmallVector<SDValue, 8> Ops(NumElts, Splatted);
12669 SDValue NewBV = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
12670 V->getValueType(0), Ops);
12672 // We may have jumped through bitcasts, so the type of the
12673 // BUILD_VECTOR may not match the type of the shuffle.
12674 if (V->getValueType(0) != VT)
12675 NewBV = DAG.getNode(ISD::BITCAST, SDLoc(N), VT, NewBV);
12680 // There are various patterns used to build up a vector from smaller vectors,
12681 // subvectors, or elements. Scan chains of these and replace unused insertions
12682 // or components with undef.
12683 if (SDValue S = simplifyShuffleOperands(SVN, N0, N1, DAG))
12686 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
12687 Level < AfterLegalizeVectorOps &&
12688 (N1.getOpcode() == ISD::UNDEF ||
12689 (N1.getOpcode() == ISD::CONCAT_VECTORS &&
12690 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) {
12691 SDValue V = partitionShuffleOfConcats(N, DAG);
12697 // Attempt to combine a shuffle of 2 inputs of 'scalar sources' -
12698 // BUILD_VECTOR or SCALAR_TO_VECTOR into a single BUILD_VECTOR.
12699 if (Level < AfterLegalizeVectorOps && TLI.isTypeLegal(VT)) {
12700 SmallVector<SDValue, 8> Ops;
12701 for (int M : SVN->getMask()) {
12702 SDValue Op = DAG.getUNDEF(VT.getScalarType());
12704 int Idx = M % NumElts;
12705 SDValue &S = (M < (int)NumElts ? N0 : N1);
12706 if (S.getOpcode() == ISD::BUILD_VECTOR && S.hasOneUse()) {
12707 Op = S.getOperand(Idx);
12708 } else if (S.getOpcode() == ISD::SCALAR_TO_VECTOR && S.hasOneUse()) {
12710 Op = S.getOperand(0);
12712 // Operand can't be combined - bail out.
12718 if (Ops.size() == VT.getVectorNumElements()) {
12719 // BUILD_VECTOR requires all inputs to be of the same type, find the
12720 // maximum type and extend them all.
12721 EVT SVT = VT.getScalarType();
12722 if (SVT.isInteger())
12723 for (SDValue &Op : Ops)
12724 SVT = (SVT.bitsLT(Op.getValueType()) ? Op.getValueType() : SVT);
12725 if (SVT != VT.getScalarType())
12726 for (SDValue &Op : Ops)
12727 Op = TLI.isZExtFree(Op.getValueType(), SVT)
12728 ? DAG.getZExtOrTrunc(Op, SDLoc(N), SVT)
12729 : DAG.getSExtOrTrunc(Op, SDLoc(N), SVT);
12730 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Ops);
12734 // If this shuffle only has a single input that is a bitcasted shuffle,
12735 // attempt to merge the 2 shuffles and suitably bitcast the inputs/output
12736 // back to their original types.
12737 if (N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
12738 N1.getOpcode() == ISD::UNDEF && Level < AfterLegalizeVectorOps &&
12739 TLI.isTypeLegal(VT)) {
12741 // Peek through the bitcast only if there is one user.
12743 while (BC0.getOpcode() == ISD::BITCAST) {
12744 if (!BC0.hasOneUse())
12746 BC0 = BC0.getOperand(0);
12749 auto ScaleShuffleMask = [](ArrayRef<int> Mask, int Scale) {
12751 return SmallVector<int, 8>(Mask.begin(), Mask.end());
12753 SmallVector<int, 8> NewMask;
12755 for (int s = 0; s != Scale; ++s)
12756 NewMask.push_back(M < 0 ? -1 : Scale * M + s);
12760 if (BC0.getOpcode() == ISD::VECTOR_SHUFFLE && BC0.hasOneUse()) {
12761 EVT SVT = VT.getScalarType();
12762 EVT InnerVT = BC0->getValueType(0);
12763 EVT InnerSVT = InnerVT.getScalarType();
12765 // Determine which shuffle works with the smaller scalar type.
12766 EVT ScaleVT = SVT.bitsLT(InnerSVT) ? VT : InnerVT;
12767 EVT ScaleSVT = ScaleVT.getScalarType();
12769 if (TLI.isTypeLegal(ScaleVT) &&
12770 0 == (InnerSVT.getSizeInBits() % ScaleSVT.getSizeInBits()) &&
12771 0 == (SVT.getSizeInBits() % ScaleSVT.getSizeInBits())) {
12773 int InnerScale = InnerSVT.getSizeInBits() / ScaleSVT.getSizeInBits();
12774 int OuterScale = SVT.getSizeInBits() / ScaleSVT.getSizeInBits();
12776 // Scale the shuffle masks to the smaller scalar type.
12777 ShuffleVectorSDNode *InnerSVN = cast<ShuffleVectorSDNode>(BC0);
12778 SmallVector<int, 8> InnerMask =
12779 ScaleShuffleMask(InnerSVN->getMask(), InnerScale);
12780 SmallVector<int, 8> OuterMask =
12781 ScaleShuffleMask(SVN->getMask(), OuterScale);
12783 // Merge the shuffle masks.
12784 SmallVector<int, 8> NewMask;
12785 for (int M : OuterMask)
12786 NewMask.push_back(M < 0 ? -1 : InnerMask[M]);
12788 // Test for shuffle mask legality over both commutations.
12789 SDValue SV0 = BC0->getOperand(0);
12790 SDValue SV1 = BC0->getOperand(1);
12791 bool LegalMask = TLI.isShuffleMaskLegal(NewMask, ScaleVT);
12793 std::swap(SV0, SV1);
12794 ShuffleVectorSDNode::commuteMask(NewMask);
12795 LegalMask = TLI.isShuffleMaskLegal(NewMask, ScaleVT);
12799 SV0 = DAG.getNode(ISD::BITCAST, SDLoc(N), ScaleVT, SV0);
12800 SV1 = DAG.getNode(ISD::BITCAST, SDLoc(N), ScaleVT, SV1);
12801 return DAG.getNode(
12802 ISD::BITCAST, SDLoc(N), VT,
12803 DAG.getVectorShuffle(ScaleVT, SDLoc(N), SV0, SV1, NewMask));
12809 // Canonicalize shuffles according to rules:
12810 // shuffle(A, shuffle(A, B)) -> shuffle(shuffle(A,B), A)
12811 // shuffle(B, shuffle(A, B)) -> shuffle(shuffle(A,B), B)
12812 // shuffle(B, shuffle(A, Undef)) -> shuffle(shuffle(A, Undef), B)
12813 if (N1.getOpcode() == ISD::VECTOR_SHUFFLE &&
12814 N0.getOpcode() != ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
12815 TLI.isTypeLegal(VT)) {
12816 // The incoming shuffle must be of the same type as the result of the
12817 // current shuffle.
12818 assert(N1->getOperand(0).getValueType() == VT &&
12819 "Shuffle types don't match");
12821 SDValue SV0 = N1->getOperand(0);
12822 SDValue SV1 = N1->getOperand(1);
12823 bool HasSameOp0 = N0 == SV0;
12824 bool IsSV1Undef = SV1.getOpcode() == ISD::UNDEF;
12825 if (HasSameOp0 || IsSV1Undef || N0 == SV1)
12826 // Commute the operands of this shuffle so that next rule
12828 return DAG.getCommutedVectorShuffle(*SVN);
12831 // Try to fold according to rules:
12832 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2)
12833 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2)
12834 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2)
12835 // Don't try to fold shuffles with illegal type.
12836 // Only fold if this shuffle is the only user of the other shuffle.
12837 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && N->isOnlyUserOf(N0.getNode()) &&
12838 Level < AfterLegalizeDAG && TLI.isTypeLegal(VT)) {
12839 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
12841 // The incoming shuffle must be of the same type as the result of the
12842 // current shuffle.
12843 assert(OtherSV->getOperand(0).getValueType() == VT &&
12844 "Shuffle types don't match");
12847 SmallVector<int, 4> Mask;
12848 // Compute the combined shuffle mask for a shuffle with SV0 as the first
12849 // operand, and SV1 as the second operand.
12850 for (unsigned i = 0; i != NumElts; ++i) {
12851 int Idx = SVN->getMaskElt(i);
12853 // Propagate Undef.
12854 Mask.push_back(Idx);
12858 SDValue CurrentVec;
12859 if (Idx < (int)NumElts) {
12860 // This shuffle index refers to the inner shuffle N0. Lookup the inner
12861 // shuffle mask to identify which vector is actually referenced.
12862 Idx = OtherSV->getMaskElt(Idx);
12864 // Propagate Undef.
12865 Mask.push_back(Idx);
12869 CurrentVec = (Idx < (int) NumElts) ? OtherSV->getOperand(0)
12870 : OtherSV->getOperand(1);
12872 // This shuffle index references an element within N1.
12876 // Simple case where 'CurrentVec' is UNDEF.
12877 if (CurrentVec.getOpcode() == ISD::UNDEF) {
12878 Mask.push_back(-1);
12882 // Canonicalize the shuffle index. We don't know yet if CurrentVec
12883 // will be the first or second operand of the combined shuffle.
12884 Idx = Idx % NumElts;
12885 if (!SV0.getNode() || SV0 == CurrentVec) {
12886 // Ok. CurrentVec is the left hand side.
12887 // Update the mask accordingly.
12889 Mask.push_back(Idx);
12893 // Bail out if we cannot convert the shuffle pair into a single shuffle.
12894 if (SV1.getNode() && SV1 != CurrentVec)
12897 // Ok. CurrentVec is the right hand side.
12898 // Update the mask accordingly.
12900 Mask.push_back(Idx + NumElts);
12903 // Check if all indices in Mask are Undef. In case, propagate Undef.
12904 bool isUndefMask = true;
12905 for (unsigned i = 0; i != NumElts && isUndefMask; ++i)
12906 isUndefMask &= Mask[i] < 0;
12909 return DAG.getUNDEF(VT);
12911 if (!SV0.getNode())
12912 SV0 = DAG.getUNDEF(VT);
12913 if (!SV1.getNode())
12914 SV1 = DAG.getUNDEF(VT);
12916 // Avoid introducing shuffles with illegal mask.
12917 if (!TLI.isShuffleMaskLegal(Mask, VT)) {
12918 ShuffleVectorSDNode::commuteMask(Mask);
12920 if (!TLI.isShuffleMaskLegal(Mask, VT))
12923 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, A, M2)
12924 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, A, M2)
12925 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, B, M2)
12926 std::swap(SV0, SV1);
12929 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2)
12930 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2)
12931 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2)
12932 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, &Mask[0]);
12938 SDValue DAGCombiner::visitSCALAR_TO_VECTOR(SDNode *N) {
12939 SDValue InVal = N->getOperand(0);
12940 EVT VT = N->getValueType(0);
12942 // Replace a SCALAR_TO_VECTOR(EXTRACT_VECTOR_ELT(V,C0)) pattern
12943 // with a VECTOR_SHUFFLE.
12944 if (InVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
12945 SDValue InVec = InVal->getOperand(0);
12946 SDValue EltNo = InVal->getOperand(1);
12948 // FIXME: We could support implicit truncation if the shuffle can be
12949 // scaled to a smaller vector scalar type.
12950 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(EltNo);
12951 if (C0 && VT == InVec.getValueType() &&
12952 VT.getScalarType() == InVal.getValueType()) {
12953 SmallVector<int, 8> NewMask(VT.getVectorNumElements(), -1);
12954 int Elt = C0->getZExtValue();
12957 if (TLI.isShuffleMaskLegal(NewMask, VT))
12958 return DAG.getVectorShuffle(VT, SDLoc(N), InVec, DAG.getUNDEF(VT),
12966 SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
12967 SDValue N0 = N->getOperand(0);
12968 SDValue N2 = N->getOperand(2);
12970 // If the input vector is a concatenation, and the insert replaces
12971 // one of the halves, we can optimize into a single concat_vectors.
12972 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
12973 N0->getNumOperands() == 2 && N2.getOpcode() == ISD::Constant) {
12974 APInt InsIdx = cast<ConstantSDNode>(N2)->getAPIntValue();
12975 EVT VT = N->getValueType(0);
12977 // Lower half: fold (insert_subvector (concat_vectors X, Y), Z) ->
12978 // (concat_vectors Z, Y)
12980 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
12981 N->getOperand(1), N0.getOperand(1));
12983 // Upper half: fold (insert_subvector (concat_vectors X, Y), Z) ->
12984 // (concat_vectors X, Z)
12985 if (InsIdx == VT.getVectorNumElements()/2)
12986 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
12987 N0.getOperand(0), N->getOperand(1));
12993 SDValue DAGCombiner::visitFP_TO_FP16(SDNode *N) {
12994 SDValue N0 = N->getOperand(0);
12996 // fold (fp_to_fp16 (fp16_to_fp op)) -> op
12997 if (N0->getOpcode() == ISD::FP16_TO_FP)
12998 return N0->getOperand(0);
13003 /// Returns a vector_shuffle if it able to transform an AND to a vector_shuffle
13004 /// with the destination vector and a zero vector.
13005 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
13006 /// vector_shuffle V, Zero, <0, 4, 2, 4>
13007 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
13008 EVT VT = N->getValueType(0);
13009 SDValue LHS = N->getOperand(0);
13010 SDValue RHS = N->getOperand(1);
13013 // Make sure we're not running after operation legalization where it
13014 // may have custom lowered the vector shuffles.
13015 if (LegalOperations)
13018 if (N->getOpcode() != ISD::AND)
13021 if (RHS.getOpcode() == ISD::BITCAST)
13022 RHS = RHS.getOperand(0);
13024 if (RHS.getOpcode() != ISD::BUILD_VECTOR)
13027 EVT RVT = RHS.getValueType();
13028 unsigned NumElts = RHS.getNumOperands();
13030 // Attempt to create a valid clear mask, splitting the mask into
13031 // sub elements and checking to see if each is
13032 // all zeros or all ones - suitable for shuffle masking.
13033 auto BuildClearMask = [&](int Split) {
13034 int NumSubElts = NumElts * Split;
13035 int NumSubBits = RVT.getScalarSizeInBits() / Split;
13037 SmallVector<int, 8> Indices;
13038 for (int i = 0; i != NumSubElts; ++i) {
13039 int EltIdx = i / Split;
13040 int SubIdx = i % Split;
13041 SDValue Elt = RHS.getOperand(EltIdx);
13042 if (Elt.getOpcode() == ISD::UNDEF) {
13043 Indices.push_back(-1);
13048 if (isa<ConstantSDNode>(Elt))
13049 Bits = cast<ConstantSDNode>(Elt)->getAPIntValue();
13050 else if (isa<ConstantFPSDNode>(Elt))
13051 Bits = cast<ConstantFPSDNode>(Elt)->getValueAPF().bitcastToAPInt();
13055 // Extract the sub element from the constant bit mask.
13056 if (DAG.getDataLayout().isBigEndian()) {
13057 Bits = Bits.lshr((Split - SubIdx - 1) * NumSubBits);
13059 Bits = Bits.lshr(SubIdx * NumSubBits);
13063 Bits = Bits.trunc(NumSubBits);
13065 if (Bits.isAllOnesValue())
13066 Indices.push_back(i);
13067 else if (Bits == 0)
13068 Indices.push_back(i + NumSubElts);
13073 // Let's see if the target supports this vector_shuffle.
13074 EVT ClearSVT = EVT::getIntegerVT(*DAG.getContext(), NumSubBits);
13075 EVT ClearVT = EVT::getVectorVT(*DAG.getContext(), ClearSVT, NumSubElts);
13076 if (!TLI.isVectorClearMaskLegal(Indices, ClearVT))
13079 SDValue Zero = DAG.getConstant(0, dl, ClearVT);
13080 return DAG.getBitcast(VT, DAG.getVectorShuffle(ClearVT, dl,
13081 DAG.getBitcast(ClearVT, LHS),
13082 Zero, &Indices[0]));
13085 // Determine maximum split level (byte level masking).
13087 if (RVT.getScalarSizeInBits() % 8 == 0)
13088 MaxSplit = RVT.getScalarSizeInBits() / 8;
13090 for (int Split = 1; Split <= MaxSplit; ++Split)
13091 if (RVT.getScalarSizeInBits() % Split == 0)
13092 if (SDValue S = BuildClearMask(Split))
13098 /// Visit a binary vector operation, like ADD.
13099 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
13100 assert(N->getValueType(0).isVector() &&
13101 "SimplifyVBinOp only works on vectors!");
13103 SDValue LHS = N->getOperand(0);
13104 SDValue RHS = N->getOperand(1);
13106 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
13108 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
13109 RHS.getOpcode() == ISD::BUILD_VECTOR) {
13110 // Check if both vectors are constants. If not bail out.
13111 if (!(cast<BuildVectorSDNode>(LHS)->isConstant() &&
13112 cast<BuildVectorSDNode>(RHS)->isConstant()))
13115 SmallVector<SDValue, 8> Ops;
13116 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
13117 SDValue LHSOp = LHS.getOperand(i);
13118 SDValue RHSOp = RHS.getOperand(i);
13120 // Can't fold divide by zero.
13121 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
13122 N->getOpcode() == ISD::FDIV) {
13123 if (isNullConstant(RHSOp) || (RHSOp.getOpcode() == ISD::ConstantFP &&
13124 cast<ConstantFPSDNode>(RHSOp.getNode())->isZero()))
13128 EVT VT = LHSOp.getValueType();
13129 EVT RVT = RHSOp.getValueType();
13131 // Integer BUILD_VECTOR operands may have types larger than the element
13132 // size (e.g., when the element type is not legal). Prior to type
13133 // legalization, the types may not match between the two BUILD_VECTORS.
13134 // Truncate one of the operands to make them match.
13135 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
13136 RHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, RHSOp);
13138 LHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), RVT, LHSOp);
13142 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(LHS), VT,
13144 if (FoldOp.getOpcode() != ISD::UNDEF &&
13145 FoldOp.getOpcode() != ISD::Constant &&
13146 FoldOp.getOpcode() != ISD::ConstantFP)
13148 Ops.push_back(FoldOp);
13149 AddToWorklist(FoldOp.getNode());
13152 if (Ops.size() == LHS.getNumOperands())
13153 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), LHS.getValueType(), Ops);
13156 // Try to convert a constant mask AND into a shuffle clear mask.
13157 if (SDValue Shuffle = XformToShuffleWithZero(N))
13160 // Type legalization might introduce new shuffles in the DAG.
13161 // Fold (VBinOp (shuffle (A, Undef, Mask)), (shuffle (B, Undef, Mask)))
13162 // -> (shuffle (VBinOp (A, B)), Undef, Mask).
13163 if (LegalTypes && isa<ShuffleVectorSDNode>(LHS) &&
13164 isa<ShuffleVectorSDNode>(RHS) && LHS.hasOneUse() && RHS.hasOneUse() &&
13165 LHS.getOperand(1).getOpcode() == ISD::UNDEF &&
13166 RHS.getOperand(1).getOpcode() == ISD::UNDEF) {
13167 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(LHS);
13168 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(RHS);
13170 if (SVN0->getMask().equals(SVN1->getMask())) {
13171 EVT VT = N->getValueType(0);
13172 SDValue UndefVector = LHS.getOperand(1);
13173 SDValue NewBinOp = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
13174 LHS.getOperand(0), RHS.getOperand(0));
13175 AddUsersToWorklist(N);
13176 return DAG.getVectorShuffle(VT, SDLoc(N), NewBinOp, UndefVector,
13177 &SVN0->getMask()[0]);
13184 SDValue DAGCombiner::SimplifySelect(SDLoc DL, SDValue N0,
13185 SDValue N1, SDValue N2){
13186 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
13188 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
13189 cast<CondCodeSDNode>(N0.getOperand(2))->get());
13191 // If we got a simplified select_cc node back from SimplifySelectCC, then
13192 // break it down into a new SETCC node, and a new SELECT node, and then return
13193 // the SELECT node, since we were called with a SELECT node.
13194 if (SCC.getNode()) {
13195 // Check to see if we got a select_cc back (to turn into setcc/select).
13196 // Otherwise, just return whatever node we got back, like fabs.
13197 if (SCC.getOpcode() == ISD::SELECT_CC) {
13198 SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0),
13200 SCC.getOperand(0), SCC.getOperand(1),
13201 SCC.getOperand(4));
13202 AddToWorklist(SETCC.getNode());
13203 return DAG.getSelect(SDLoc(SCC), SCC.getValueType(), SETCC,
13204 SCC.getOperand(2), SCC.getOperand(3));
13212 /// Given a SELECT or a SELECT_CC node, where LHS and RHS are the two values
13213 /// being selected between, see if we can simplify the select. Callers of this
13214 /// should assume that TheSelect is deleted if this returns true. As such, they
13215 /// should return the appropriate thing (e.g. the node) back to the top-level of
13216 /// the DAG combiner loop to avoid it being looked at.
13217 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
13220 // fold (select (setcc x, -0.0, *lt), NaN, (fsqrt x))
13221 // The select + setcc is redundant, because fsqrt returns NaN for X < -0.
13222 if (const ConstantFPSDNode *NaN = isConstOrConstSplatFP(LHS)) {
13223 if (NaN->isNaN() && RHS.getOpcode() == ISD::FSQRT) {
13224 // We have: (select (setcc ?, ?, ?), NaN, (fsqrt ?))
13225 SDValue Sqrt = RHS;
13228 const ConstantFPSDNode *NegZero = nullptr;
13230 if (TheSelect->getOpcode() == ISD::SELECT_CC) {
13231 CC = dyn_cast<CondCodeSDNode>(TheSelect->getOperand(4))->get();
13232 CmpLHS = TheSelect->getOperand(0);
13233 NegZero = isConstOrConstSplatFP(TheSelect->getOperand(1));
13235 // SELECT or VSELECT
13236 SDValue Cmp = TheSelect->getOperand(0);
13237 if (Cmp.getOpcode() == ISD::SETCC) {
13238 CC = dyn_cast<CondCodeSDNode>(Cmp.getOperand(2))->get();
13239 CmpLHS = Cmp.getOperand(0);
13240 NegZero = isConstOrConstSplatFP(Cmp.getOperand(1));
13243 if (NegZero && NegZero->isNegative() && NegZero->isZero() &&
13244 Sqrt.getOperand(0) == CmpLHS && (CC == ISD::SETOLT ||
13245 CC == ISD::SETULT || CC == ISD::SETLT)) {
13246 // We have: (select (setcc x, -0.0, *lt), NaN, (fsqrt x))
13247 CombineTo(TheSelect, Sqrt);
13252 // Cannot simplify select with vector condition
13253 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
13255 // If this is a select from two identical things, try to pull the operation
13256 // through the select.
13257 if (LHS.getOpcode() != RHS.getOpcode() ||
13258 !LHS.hasOneUse() || !RHS.hasOneUse())
13261 // If this is a load and the token chain is identical, replace the select
13262 // of two loads with a load through a select of the address to load from.
13263 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
13264 // constants have been dropped into the constant pool.
13265 if (LHS.getOpcode() == ISD::LOAD) {
13266 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
13267 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
13269 // Token chains must be identical.
13270 if (LHS.getOperand(0) != RHS.getOperand(0) ||
13271 // Do not let this transformation reduce the number of volatile loads.
13272 LLD->isVolatile() || RLD->isVolatile() ||
13273 // FIXME: If either is a pre/post inc/dec load,
13274 // we'd need to split out the address adjustment.
13275 LLD->isIndexed() || RLD->isIndexed() ||
13276 // If this is an EXTLOAD, the VT's must match.
13277 LLD->getMemoryVT() != RLD->getMemoryVT() ||
13278 // If this is an EXTLOAD, the kind of extension must match.
13279 (LLD->getExtensionType() != RLD->getExtensionType() &&
13280 // The only exception is if one of the extensions is anyext.
13281 LLD->getExtensionType() != ISD::EXTLOAD &&
13282 RLD->getExtensionType() != ISD::EXTLOAD) ||
13283 // FIXME: this discards src value information. This is
13284 // over-conservative. It would be beneficial to be able to remember
13285 // both potential memory locations. Since we are discarding
13286 // src value info, don't do the transformation if the memory
13287 // locations are not in the default address space.
13288 LLD->getPointerInfo().getAddrSpace() != 0 ||
13289 RLD->getPointerInfo().getAddrSpace() != 0 ||
13290 !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(),
13291 LLD->getBasePtr().getValueType()))
13294 // Check that the select condition doesn't reach either load. If so,
13295 // folding this will induce a cycle into the DAG. If not, this is safe to
13296 // xform, so create a select of the addresses.
13298 if (TheSelect->getOpcode() == ISD::SELECT) {
13299 SDNode *CondNode = TheSelect->getOperand(0).getNode();
13300 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
13301 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
13303 // The loads must not depend on one another.
13304 if (LLD->isPredecessorOf(RLD) ||
13305 RLD->isPredecessorOf(LLD))
13307 Addr = DAG.getSelect(SDLoc(TheSelect),
13308 LLD->getBasePtr().getValueType(),
13309 TheSelect->getOperand(0), LLD->getBasePtr(),
13310 RLD->getBasePtr());
13311 } else { // Otherwise SELECT_CC
13312 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
13313 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
13315 if ((LLD->hasAnyUseOfValue(1) &&
13316 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
13317 (RLD->hasAnyUseOfValue(1) &&
13318 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
13321 Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect),
13322 LLD->getBasePtr().getValueType(),
13323 TheSelect->getOperand(0),
13324 TheSelect->getOperand(1),
13325 LLD->getBasePtr(), RLD->getBasePtr(),
13326 TheSelect->getOperand(4));
13330 // It is safe to replace the two loads if they have different alignments,
13331 // but the new load must be the minimum (most restrictive) alignment of the
13333 bool isInvariant = LLD->isInvariant() & RLD->isInvariant();
13334 unsigned Alignment = std::min(LLD->getAlignment(), RLD->getAlignment());
13335 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
13336 Load = DAG.getLoad(TheSelect->getValueType(0),
13338 // FIXME: Discards pointer and AA info.
13339 LLD->getChain(), Addr, MachinePointerInfo(),
13340 LLD->isVolatile(), LLD->isNonTemporal(),
13341 isInvariant, Alignment);
13343 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
13344 RLD->getExtensionType() : LLD->getExtensionType(),
13346 TheSelect->getValueType(0),
13347 // FIXME: Discards pointer and AA info.
13348 LLD->getChain(), Addr, MachinePointerInfo(),
13349 LLD->getMemoryVT(), LLD->isVolatile(),
13350 LLD->isNonTemporal(), isInvariant, Alignment);
13353 // Users of the select now use the result of the load.
13354 CombineTo(TheSelect, Load);
13356 // Users of the old loads now use the new load's chain. We know the
13357 // old-load value is dead now.
13358 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
13359 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
13366 /// Simplify an expression of the form (N0 cond N1) ? N2 : N3
13367 /// where 'cond' is the comparison specified by CC.
13368 SDValue DAGCombiner::SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1,
13369 SDValue N2, SDValue N3,
13370 ISD::CondCode CC, bool NotExtCompare) {
13371 // (x ? y : y) -> y.
13372 if (N2 == N3) return N2;
13374 EVT VT = N2.getValueType();
13375 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
13376 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
13378 // Determine if the condition we're dealing with is constant
13379 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
13380 N0, N1, CC, DL, false);
13381 if (SCC.getNode()) AddToWorklist(SCC.getNode());
13383 if (ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode())) {
13384 // fold select_cc true, x, y -> x
13385 // fold select_cc false, x, y -> y
13386 return !SCCC->isNullValue() ? N2 : N3;
13389 // Check to see if we can simplify the select into an fabs node
13390 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
13391 // Allow either -0.0 or 0.0
13392 if (CFP->isZero()) {
13393 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
13394 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
13395 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
13396 N2 == N3.getOperand(0))
13397 return DAG.getNode(ISD::FABS, DL, VT, N0);
13399 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
13400 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
13401 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
13402 N2.getOperand(0) == N3)
13403 return DAG.getNode(ISD::FABS, DL, VT, N3);
13407 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
13408 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
13409 // in it. This is a win when the constant is not otherwise available because
13410 // it replaces two constant pool loads with one. We only do this if the FP
13411 // type is known to be legal, because if it isn't, then we are before legalize
13412 // types an we want the other legalization to happen first (e.g. to avoid
13413 // messing with soft float) and if the ConstantFP is not legal, because if
13414 // it is legal, we may not need to store the FP constant in a constant pool.
13415 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
13416 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
13417 if (TLI.isTypeLegal(N2.getValueType()) &&
13418 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
13419 TargetLowering::Legal &&
13420 !TLI.isFPImmLegal(TV->getValueAPF(), TV->getValueType(0)) &&
13421 !TLI.isFPImmLegal(FV->getValueAPF(), FV->getValueType(0))) &&
13422 // If both constants have multiple uses, then we won't need to do an
13423 // extra load, they are likely around in registers for other users.
13424 (TV->hasOneUse() || FV->hasOneUse())) {
13425 Constant *Elts[] = {
13426 const_cast<ConstantFP*>(FV->getConstantFPValue()),
13427 const_cast<ConstantFP*>(TV->getConstantFPValue())
13429 Type *FPTy = Elts[0]->getType();
13430 const DataLayout &TD = DAG.getDataLayout();
13432 // Create a ConstantArray of the two constants.
13433 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
13435 DAG.getConstantPool(CA, TLI.getPointerTy(DAG.getDataLayout()),
13436 TD.getPrefTypeAlignment(FPTy));
13437 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
13439 // Get the offsets to the 0 and 1 element of the array so that we can
13440 // select between them.
13441 SDValue Zero = DAG.getIntPtrConstant(0, DL);
13442 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
13443 SDValue One = DAG.getIntPtrConstant(EltSize, SDLoc(FV));
13445 SDValue Cond = DAG.getSetCC(DL,
13446 getSetCCResultType(N0.getValueType()),
13448 AddToWorklist(Cond.getNode());
13449 SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(),
13451 AddToWorklist(CstOffset.getNode());
13452 CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx,
13454 AddToWorklist(CPIdx.getNode());
13455 return DAG.getLoad(
13456 TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
13457 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13458 false, false, false, Alignment);
13462 // Check to see if we can perform the "gzip trick", transforming
13463 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
13464 if (isNullConstant(N3) && CC == ISD::SETLT &&
13465 (isNullConstant(N1) || // (a < 0) ? b : 0
13466 (isOneConstant(N1) && N0 == N2))) { // (a < 1) ? a : 0
13467 EVT XType = N0.getValueType();
13468 EVT AType = N2.getValueType();
13469 if (XType.bitsGE(AType)) {
13470 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
13471 // single-bit constant.
13472 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue() - 1)) == 0)) {
13473 unsigned ShCtV = N2C->getAPIntValue().logBase2();
13474 ShCtV = XType.getSizeInBits() - ShCtV - 1;
13475 SDValue ShCt = DAG.getConstant(ShCtV, SDLoc(N0),
13476 getShiftAmountTy(N0.getValueType()));
13477 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0),
13479 AddToWorklist(Shift.getNode());
13481 if (XType.bitsGT(AType)) {
13482 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
13483 AddToWorklist(Shift.getNode());
13486 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
13489 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0),
13491 DAG.getConstant(XType.getSizeInBits() - 1,
13493 getShiftAmountTy(N0.getValueType())));
13494 AddToWorklist(Shift.getNode());
13496 if (XType.bitsGT(AType)) {
13497 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
13498 AddToWorklist(Shift.getNode());
13501 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
13505 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
13506 // where y is has a single bit set.
13507 // A plaintext description would be, we can turn the SELECT_CC into an AND
13508 // when the condition can be materialized as an all-ones register. Any
13509 // single bit-test can be materialized as an all-ones register with
13510 // shift-left and shift-right-arith.
13511 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
13512 N0->getValueType(0) == VT && isNullConstant(N1) && isNullConstant(N2)) {
13513 SDValue AndLHS = N0->getOperand(0);
13514 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
13515 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
13516 // Shift the tested bit over the sign bit.
13517 APInt AndMask = ConstAndRHS->getAPIntValue();
13519 DAG.getConstant(AndMask.countLeadingZeros(), SDLoc(AndLHS),
13520 getShiftAmountTy(AndLHS.getValueType()));
13521 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt);
13523 // Now arithmetic right shift it all the way over, so the result is either
13524 // all-ones, or zero.
13526 DAG.getConstant(AndMask.getBitWidth() - 1, SDLoc(Shl),
13527 getShiftAmountTy(Shl.getValueType()));
13528 SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt);
13530 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
13534 // fold select C, 16, 0 -> shl C, 4
13535 if (N2C && isNullConstant(N3) && N2C->getAPIntValue().isPowerOf2() &&
13536 TLI.getBooleanContents(N0.getValueType()) ==
13537 TargetLowering::ZeroOrOneBooleanContent) {
13539 // If the caller doesn't want us to simplify this into a zext of a compare,
13541 if (NotExtCompare && N2C->isOne())
13544 // Get a SetCC of the condition
13545 // NOTE: Don't create a SETCC if it's not legal on this target.
13546 if (!LegalOperations ||
13547 TLI.isOperationLegal(ISD::SETCC,
13548 LegalTypes ? getSetCCResultType(N0.getValueType()) : MVT::i1)) {
13550 // cast from setcc result type to select result type
13552 SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()),
13554 if (N2.getValueType().bitsLT(SCC.getValueType()))
13555 Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2),
13556 N2.getValueType());
13558 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
13559 N2.getValueType(), SCC);
13561 SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC);
13562 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
13563 N2.getValueType(), SCC);
13566 AddToWorklist(SCC.getNode());
13567 AddToWorklist(Temp.getNode());
13572 // shl setcc result by log2 n2c
13573 return DAG.getNode(
13574 ISD::SHL, DL, N2.getValueType(), Temp,
13575 DAG.getConstant(N2C->getAPIntValue().logBase2(), SDLoc(Temp),
13576 getShiftAmountTy(Temp.getValueType())));
13580 // Check to see if this is the equivalent of setcc
13581 // FIXME: Turn all of these into setcc if setcc if setcc is legal
13582 // otherwise, go ahead with the folds.
13583 if (0 && isNullConstant(N3) && isOneConstant(N2)) {
13584 EVT XType = N0.getValueType();
13585 if (!LegalOperations ||
13586 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(XType))) {
13587 SDValue Res = DAG.getSetCC(DL, getSetCCResultType(XType), N0, N1, CC);
13588 if (Res.getValueType() != VT)
13589 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
13593 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
13594 if (isNullConstant(N1) && CC == ISD::SETEQ &&
13595 (!LegalOperations ||
13596 TLI.isOperationLegal(ISD::CTLZ, XType))) {
13597 SDValue Ctlz = DAG.getNode(ISD::CTLZ, SDLoc(N0), XType, N0);
13598 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
13599 DAG.getConstant(Log2_32(XType.getSizeInBits()),
13601 getShiftAmountTy(Ctlz.getValueType())));
13603 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
13604 if (isNullConstant(N1) && CC == ISD::SETGT) {
13606 SDValue NegN0 = DAG.getNode(ISD::SUB, DL,
13607 XType, DAG.getConstant(0, DL, XType), N0);
13608 SDValue NotN0 = DAG.getNOT(DL, N0, XType);
13609 return DAG.getNode(ISD::SRL, DL, XType,
13610 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
13611 DAG.getConstant(XType.getSizeInBits() - 1, DL,
13612 getShiftAmountTy(XType)));
13614 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
13615 if (isAllOnesConstant(N1) && CC == ISD::SETGT) {
13617 SDValue Sign = DAG.getNode(ISD::SRL, DL, XType, N0,
13618 DAG.getConstant(XType.getSizeInBits() - 1, DL,
13619 getShiftAmountTy(N0.getValueType())));
13620 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, DL,
13625 // Check to see if this is an integer abs.
13626 // select_cc setg[te] X, 0, X, -X ->
13627 // select_cc setgt X, -1, X, -X ->
13628 // select_cc setl[te] X, 0, -X, X ->
13629 // select_cc setlt X, 1, -X, X ->
13630 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
13632 ConstantSDNode *SubC = nullptr;
13633 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
13634 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
13635 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
13636 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
13637 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
13638 (N1C->isOne() && CC == ISD::SETLT)) &&
13639 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
13640 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
13642 EVT XType = N0.getValueType();
13643 if (SubC && SubC->isNullValue() && XType.isInteger()) {
13645 SDValue Shift = DAG.getNode(ISD::SRA, DL, XType,
13647 DAG.getConstant(XType.getSizeInBits() - 1, DL,
13648 getShiftAmountTy(N0.getValueType())));
13649 SDValue Add = DAG.getNode(ISD::ADD, DL,
13651 AddToWorklist(Shift.getNode());
13652 AddToWorklist(Add.getNode());
13653 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
13660 /// This is a stub for TargetLowering::SimplifySetCC.
13661 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
13662 SDValue N1, ISD::CondCode Cond,
13663 SDLoc DL, bool foldBooleans) {
13664 TargetLowering::DAGCombinerInfo
13665 DagCombineInfo(DAG, Level, false, this);
13666 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
13669 /// Given an ISD::SDIV node expressing a divide by constant, return
13670 /// a DAG expression to select that will generate the same value by multiplying
13671 /// by a magic number.
13672 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
13673 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
13674 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
13678 // Avoid division by zero.
13679 if (C->isNullValue())
13682 std::vector<SDNode*> Built;
13684 TLI.BuildSDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
13686 for (SDNode *N : Built)
13691 /// Given an ISD::SDIV node expressing a divide by constant power of 2, return a
13692 /// DAG expression that will generate the same value by right shifting.
13693 SDValue DAGCombiner::BuildSDIVPow2(SDNode *N) {
13694 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
13698 // Avoid division by zero.
13699 if (C->isNullValue())
13702 std::vector<SDNode *> Built;
13703 SDValue S = TLI.BuildSDIVPow2(N, C->getAPIntValue(), DAG, &Built);
13705 for (SDNode *N : Built)
13710 /// Given an ISD::UDIV node expressing a divide by constant, return a DAG
13711 /// expression that will generate the same value by multiplying by a magic
13713 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
13714 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
13715 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
13719 // Avoid division by zero.
13720 if (C->isNullValue())
13723 std::vector<SDNode*> Built;
13725 TLI.BuildUDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
13727 for (SDNode *N : Built)
13732 SDValue DAGCombiner::BuildReciprocalEstimate(SDValue Op) {
13733 if (Level >= AfterLegalizeDAG)
13736 // Expose the DAG combiner to the target combiner implementations.
13737 TargetLowering::DAGCombinerInfo DCI(DAG, Level, false, this);
13739 unsigned Iterations = 0;
13740 if (SDValue Est = TLI.getRecipEstimate(Op, DCI, Iterations)) {
13742 // Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
13743 // For the reciprocal, we need to find the zero of the function:
13744 // F(X) = A X - 1 [which has a zero at X = 1/A]
13746 // X_{i+1} = X_i (2 - A X_i) = X_i + X_i (1 - A X_i) [this second form
13747 // does not require additional intermediate precision]
13748 EVT VT = Op.getValueType();
13750 SDValue FPOne = DAG.getConstantFP(1.0, DL, VT);
13752 AddToWorklist(Est.getNode());
13754 // Newton iterations: Est = Est + Est (1 - Arg * Est)
13755 for (unsigned i = 0; i < Iterations; ++i) {
13756 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Op, Est);
13757 AddToWorklist(NewEst.getNode());
13759 NewEst = DAG.getNode(ISD::FSUB, DL, VT, FPOne, NewEst);
13760 AddToWorklist(NewEst.getNode());
13762 NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst);
13763 AddToWorklist(NewEst.getNode());
13765 Est = DAG.getNode(ISD::FADD, DL, VT, Est, NewEst);
13766 AddToWorklist(Est.getNode());
13775 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
13776 /// For the reciprocal sqrt, we need to find the zero of the function:
13777 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
13779 /// X_{i+1} = X_i (1.5 - A X_i^2 / 2)
13780 /// As a result, we precompute A/2 prior to the iteration loop.
13781 SDValue DAGCombiner::BuildRsqrtNROneConst(SDValue Arg, SDValue Est,
13782 unsigned Iterations) {
13783 EVT VT = Arg.getValueType();
13785 SDValue ThreeHalves = DAG.getConstantFP(1.5, DL, VT);
13787 // We now need 0.5 * Arg which we can write as (1.5 * Arg - Arg) so that
13788 // this entire sequence requires only one FP constant.
13789 SDValue HalfArg = DAG.getNode(ISD::FMUL, DL, VT, ThreeHalves, Arg);
13790 AddToWorklist(HalfArg.getNode());
13792 HalfArg = DAG.getNode(ISD::FSUB, DL, VT, HalfArg, Arg);
13793 AddToWorklist(HalfArg.getNode());
13795 // Newton iterations: Est = Est * (1.5 - HalfArg * Est * Est)
13796 for (unsigned i = 0; i < Iterations; ++i) {
13797 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, Est);
13798 AddToWorklist(NewEst.getNode());
13800 NewEst = DAG.getNode(ISD::FMUL, DL, VT, HalfArg, NewEst);
13801 AddToWorklist(NewEst.getNode());
13803 NewEst = DAG.getNode(ISD::FSUB, DL, VT, ThreeHalves, NewEst);
13804 AddToWorklist(NewEst.getNode());
13806 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst);
13807 AddToWorklist(Est.getNode());
13812 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
13813 /// For the reciprocal sqrt, we need to find the zero of the function:
13814 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
13816 /// X_{i+1} = (-0.5 * X_i) * (A * X_i * X_i + (-3.0))
13817 SDValue DAGCombiner::BuildRsqrtNRTwoConst(SDValue Arg, SDValue Est,
13818 unsigned Iterations) {
13819 EVT VT = Arg.getValueType();
13821 SDValue MinusThree = DAG.getConstantFP(-3.0, DL, VT);
13822 SDValue MinusHalf = DAG.getConstantFP(-0.5, DL, VT);
13824 // Newton iterations: Est = -0.5 * Est * (-3.0 + Arg * Est * Est)
13825 for (unsigned i = 0; i < Iterations; ++i) {
13826 SDValue HalfEst = DAG.getNode(ISD::FMUL, DL, VT, Est, MinusHalf);
13827 AddToWorklist(HalfEst.getNode());
13829 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Est);
13830 AddToWorklist(Est.getNode());
13832 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Arg);
13833 AddToWorklist(Est.getNode());
13835 Est = DAG.getNode(ISD::FADD, DL, VT, Est, MinusThree);
13836 AddToWorklist(Est.getNode());
13838 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, HalfEst);
13839 AddToWorklist(Est.getNode());
13844 SDValue DAGCombiner::BuildRsqrtEstimate(SDValue Op) {
13845 if (Level >= AfterLegalizeDAG)
13848 // Expose the DAG combiner to the target combiner implementations.
13849 TargetLowering::DAGCombinerInfo DCI(DAG, Level, false, this);
13850 unsigned Iterations = 0;
13851 bool UseOneConstNR = false;
13852 if (SDValue Est = TLI.getRsqrtEstimate(Op, DCI, Iterations, UseOneConstNR)) {
13853 AddToWorklist(Est.getNode());
13855 Est = UseOneConstNR ?
13856 BuildRsqrtNROneConst(Op, Est, Iterations) :
13857 BuildRsqrtNRTwoConst(Op, Est, Iterations);
13865 /// Return true if base is a frame index, which is known not to alias with
13866 /// anything but itself. Provides base object and offset as results.
13867 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
13868 const GlobalValue *&GV, const void *&CV) {
13869 // Assume it is a primitive operation.
13870 Base = Ptr; Offset = 0; GV = nullptr; CV = nullptr;
13872 // If it's an adding a simple constant then integrate the offset.
13873 if (Base.getOpcode() == ISD::ADD) {
13874 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
13875 Base = Base.getOperand(0);
13876 Offset += C->getZExtValue();
13880 // Return the underlying GlobalValue, and update the Offset. Return false
13881 // for GlobalAddressSDNode since the same GlobalAddress may be represented
13882 // by multiple nodes with different offsets.
13883 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
13884 GV = G->getGlobal();
13885 Offset += G->getOffset();
13889 // Return the underlying Constant value, and update the Offset. Return false
13890 // for ConstantSDNodes since the same constant pool entry may be represented
13891 // by multiple nodes with different offsets.
13892 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
13893 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
13894 : (const void *)C->getConstVal();
13895 Offset += C->getOffset();
13898 // If it's any of the following then it can't alias with anything but itself.
13899 return isa<FrameIndexSDNode>(Base);
13902 /// Return true if there is any possibility that the two addresses overlap.
13903 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const {
13904 // If they are the same then they must be aliases.
13905 if (Op0->getBasePtr() == Op1->getBasePtr()) return true;
13907 // If they are both volatile then they cannot be reordered.
13908 if (Op0->isVolatile() && Op1->isVolatile()) return true;
13910 // If one operation reads from invariant memory, and the other may store, they
13911 // cannot alias. These should really be checking the equivalent of mayWrite,
13912 // but it only matters for memory nodes other than load /store.
13913 if (Op0->isInvariant() && Op1->writeMem())
13916 if (Op1->isInvariant() && Op0->writeMem())
13919 // Gather base node and offset information.
13920 SDValue Base1, Base2;
13921 int64_t Offset1, Offset2;
13922 const GlobalValue *GV1, *GV2;
13923 const void *CV1, *CV2;
13924 bool isFrameIndex1 = FindBaseOffset(Op0->getBasePtr(),
13925 Base1, Offset1, GV1, CV1);
13926 bool isFrameIndex2 = FindBaseOffset(Op1->getBasePtr(),
13927 Base2, Offset2, GV2, CV2);
13929 // If they have a same base address then check to see if they overlap.
13930 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
13931 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
13932 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
13934 // It is possible for different frame indices to alias each other, mostly
13935 // when tail call optimization reuses return address slots for arguments.
13936 // To catch this case, look up the actual index of frame indices to compute
13937 // the real alias relationship.
13938 if (isFrameIndex1 && isFrameIndex2) {
13939 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
13940 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
13941 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
13942 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
13943 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
13946 // Otherwise, if we know what the bases are, and they aren't identical, then
13947 // we know they cannot alias.
13948 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
13951 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
13952 // compared to the size and offset of the access, we may be able to prove they
13953 // do not alias. This check is conservative for now to catch cases created by
13954 // splitting vector types.
13955 if ((Op0->getOriginalAlignment() == Op1->getOriginalAlignment()) &&
13956 (Op0->getSrcValueOffset() != Op1->getSrcValueOffset()) &&
13957 (Op0->getMemoryVT().getSizeInBits() >> 3 ==
13958 Op1->getMemoryVT().getSizeInBits() >> 3) &&
13959 (Op0->getOriginalAlignment() > Op0->getMemoryVT().getSizeInBits()) >> 3) {
13960 int64_t OffAlign1 = Op0->getSrcValueOffset() % Op0->getOriginalAlignment();
13961 int64_t OffAlign2 = Op1->getSrcValueOffset() % Op1->getOriginalAlignment();
13963 // There is no overlap between these relatively aligned accesses of similar
13964 // size, return no alias.
13965 if ((OffAlign1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign2 ||
13966 (OffAlign2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign1)
13970 bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0
13972 : DAG.getSubtarget().useAA();
13974 if (CombinerAAOnlyFunc.getNumOccurrences() &&
13975 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
13979 Op0->getMemOperand()->getValue() && Op1->getMemOperand()->getValue()) {
13980 // Use alias analysis information.
13981 int64_t MinOffset = std::min(Op0->getSrcValueOffset(),
13982 Op1->getSrcValueOffset());
13983 int64_t Overlap1 = (Op0->getMemoryVT().getSizeInBits() >> 3) +
13984 Op0->getSrcValueOffset() - MinOffset;
13985 int64_t Overlap2 = (Op1->getMemoryVT().getSizeInBits() >> 3) +
13986 Op1->getSrcValueOffset() - MinOffset;
13987 AliasResult AAResult =
13988 AA.alias(MemoryLocation(Op0->getMemOperand()->getValue(), Overlap1,
13989 UseTBAA ? Op0->getAAInfo() : AAMDNodes()),
13990 MemoryLocation(Op1->getMemOperand()->getValue(), Overlap2,
13991 UseTBAA ? Op1->getAAInfo() : AAMDNodes()));
13992 if (AAResult == NoAlias)
13996 // Otherwise we have to assume they alias.
14000 /// Walk up chain skipping non-aliasing memory nodes,
14001 /// looking for aliasing nodes and adding them to the Aliases vector.
14002 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
14003 SmallVectorImpl<SDValue> &Aliases) {
14004 SmallVector<SDValue, 8> Chains; // List of chains to visit.
14005 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
14007 // Get alias information for node.
14008 bool IsLoad = isa<LoadSDNode>(N) && !cast<LSBaseSDNode>(N)->isVolatile();
14011 Chains.push_back(OriginalChain);
14012 unsigned Depth = 0;
14014 // Look at each chain and determine if it is an alias. If so, add it to the
14015 // aliases list. If not, then continue up the chain looking for the next
14017 while (!Chains.empty()) {
14018 SDValue Chain = Chains.pop_back_val();
14020 // For TokenFactor nodes, look at each operand and only continue up the
14021 // chain until we find two aliases. If we've seen two aliases, assume we'll
14022 // find more and revert to original chain since the xform is unlikely to be
14025 // FIXME: The depth check could be made to return the last non-aliasing
14026 // chain we found before we hit a tokenfactor rather than the original
14028 if (Depth > 6 || Aliases.size() == 2) {
14030 Aliases.push_back(OriginalChain);
14034 // Don't bother if we've been before.
14035 if (!Visited.insert(Chain.getNode()).second)
14038 switch (Chain.getOpcode()) {
14039 case ISD::EntryToken:
14040 // Entry token is ideal chain operand, but handled in FindBetterChain.
14045 // Get alias information for Chain.
14046 bool IsOpLoad = isa<LoadSDNode>(Chain.getNode()) &&
14047 !cast<LSBaseSDNode>(Chain.getNode())->isVolatile();
14049 // If chain is alias then stop here.
14050 if (!(IsLoad && IsOpLoad) &&
14051 isAlias(cast<LSBaseSDNode>(N), cast<LSBaseSDNode>(Chain.getNode()))) {
14052 Aliases.push_back(Chain);
14054 // Look further up the chain.
14055 Chains.push_back(Chain.getOperand(0));
14061 case ISD::TokenFactor:
14062 // We have to check each of the operands of the token factor for "small"
14063 // token factors, so we queue them up. Adding the operands to the queue
14064 // (stack) in reverse order maintains the original order and increases the
14065 // likelihood that getNode will find a matching token factor (CSE.)
14066 if (Chain.getNumOperands() > 16) {
14067 Aliases.push_back(Chain);
14070 for (unsigned n = Chain.getNumOperands(); n;)
14071 Chains.push_back(Chain.getOperand(--n));
14076 // For all other instructions we will just have to take what we can get.
14077 Aliases.push_back(Chain);
14082 // We need to be careful here to also search for aliases through the
14083 // value operand of a store, etc. Consider the following situation:
14085 // L1 = load Token1, %52
14086 // S1 = store Token1, L1, %51
14087 // L2 = load Token1, %52+8
14088 // S2 = store Token1, L2, %51+8
14089 // Token2 = Token(S1, S2)
14090 // L3 = load Token2, %53
14091 // S3 = store Token2, L3, %52
14092 // L4 = load Token2, %53+8
14093 // S4 = store Token2, L4, %52+8
14094 // If we search for aliases of S3 (which loads address %52), and we look
14095 // only through the chain, then we'll miss the trivial dependence on L1
14096 // (which also loads from %52). We then might change all loads and
14097 // stores to use Token1 as their chain operand, which could result in
14098 // copying %53 into %52 before copying %52 into %51 (which should
14101 // The problem is, however, that searching for such data dependencies
14102 // can become expensive, and the cost is not directly related to the
14103 // chain depth. Instead, we'll rule out such configurations here by
14104 // insisting that we've visited all chain users (except for users
14105 // of the original chain, which is not necessary). When doing this,
14106 // we need to look through nodes we don't care about (otherwise, things
14107 // like register copies will interfere with trivial cases).
14109 SmallVector<const SDNode *, 16> Worklist;
14110 for (const SDNode *N : Visited)
14111 if (N != OriginalChain.getNode())
14112 Worklist.push_back(N);
14114 while (!Worklist.empty()) {
14115 const SDNode *M = Worklist.pop_back_val();
14117 // We have already visited M, and want to make sure we've visited any uses
14118 // of M that we care about. For uses that we've not visisted, and don't
14119 // care about, queue them to the worklist.
14121 for (SDNode::use_iterator UI = M->use_begin(),
14122 UIE = M->use_end(); UI != UIE; ++UI)
14123 if (UI.getUse().getValueType() == MVT::Other &&
14124 Visited.insert(*UI).second) {
14125 if (isa<MemSDNode>(*UI)) {
14126 // We've not visited this use, and we care about it (it could have an
14127 // ordering dependency with the original node).
14129 Aliases.push_back(OriginalChain);
14133 // We've not visited this use, but we don't care about it. Mark it as
14134 // visited and enqueue it to the worklist.
14135 Worklist.push_back(*UI);
14140 /// Walk up chain skipping non-aliasing memory nodes, looking for a better chain
14141 /// (aliasing node.)
14142 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
14143 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
14145 // Accumulate all the aliases to this node.
14146 GatherAllAliases(N, OldChain, Aliases);
14148 // If no operands then chain to entry token.
14149 if (Aliases.size() == 0)
14150 return DAG.getEntryNode();
14152 // If a single operand then chain to it. We don't need to revisit it.
14153 if (Aliases.size() == 1)
14156 // Construct a custom tailored token factor.
14157 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Aliases);
14160 /// This is the entry point for the file.
14161 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
14162 CodeGenOpt::Level OptLevel) {
14163 /// This is the main entry point to this class.
14164 DAGCombiner(*this, AA, OptLevel).Run(Level);