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 visitOR(SDNode *N);
250 SDValue visitXOR(SDNode *N);
251 SDValue SimplifyVBinOp(SDNode *N);
252 SDValue SimplifyVUnaryOp(SDNode *N);
253 SDValue visitSHL(SDNode *N);
254 SDValue visitSRA(SDNode *N);
255 SDValue visitSRL(SDNode *N);
256 SDValue visitRotate(SDNode *N);
257 SDValue visitCTLZ(SDNode *N);
258 SDValue visitCTLZ_ZERO_UNDEF(SDNode *N);
259 SDValue visitCTTZ(SDNode *N);
260 SDValue visitCTTZ_ZERO_UNDEF(SDNode *N);
261 SDValue visitCTPOP(SDNode *N);
262 SDValue visitSELECT(SDNode *N);
263 SDValue visitVSELECT(SDNode *N);
264 SDValue visitSELECT_CC(SDNode *N);
265 SDValue visitSETCC(SDNode *N);
266 SDValue visitSIGN_EXTEND(SDNode *N);
267 SDValue visitZERO_EXTEND(SDNode *N);
268 SDValue visitANY_EXTEND(SDNode *N);
269 SDValue visitSIGN_EXTEND_INREG(SDNode *N);
270 SDValue visitTRUNCATE(SDNode *N);
271 SDValue visitBITCAST(SDNode *N);
272 SDValue visitBUILD_PAIR(SDNode *N);
273 SDValue visitFADD(SDNode *N);
274 SDValue visitFSUB(SDNode *N);
275 SDValue visitFMUL(SDNode *N);
276 SDValue visitFMA(SDNode *N);
277 SDValue visitFDIV(SDNode *N);
278 SDValue visitFREM(SDNode *N);
279 SDValue visitFSQRT(SDNode *N);
280 SDValue visitFCOPYSIGN(SDNode *N);
281 SDValue visitSINT_TO_FP(SDNode *N);
282 SDValue visitUINT_TO_FP(SDNode *N);
283 SDValue visitFP_TO_SINT(SDNode *N);
284 SDValue visitFP_TO_UINT(SDNode *N);
285 SDValue visitFP_ROUND(SDNode *N);
286 SDValue visitFP_ROUND_INREG(SDNode *N);
287 SDValue visitFP_EXTEND(SDNode *N);
288 SDValue visitFNEG(SDNode *N);
289 SDValue visitFABS(SDNode *N);
290 SDValue visitFCEIL(SDNode *N);
291 SDValue visitFTRUNC(SDNode *N);
292 SDValue visitFFLOOR(SDNode *N);
293 SDValue visitFMINNUM(SDNode *N);
294 SDValue visitFMAXNUM(SDNode *N);
295 SDValue visitBRCOND(SDNode *N);
296 SDValue visitBR_CC(SDNode *N);
297 SDValue visitLOAD(SDNode *N);
298 SDValue visitSTORE(SDNode *N);
299 SDValue visitINSERT_VECTOR_ELT(SDNode *N);
300 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
301 SDValue visitBUILD_VECTOR(SDNode *N);
302 SDValue visitCONCAT_VECTORS(SDNode *N);
303 SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
304 SDValue visitVECTOR_SHUFFLE(SDNode *N);
305 SDValue visitINSERT_SUBVECTOR(SDNode *N);
306 SDValue visitMLOAD(SDNode *N);
307 SDValue visitMSTORE(SDNode *N);
309 SDValue XformToShuffleWithZero(SDNode *N);
310 SDValue ReassociateOps(unsigned Opc, SDLoc DL, SDValue LHS, SDValue RHS);
312 SDValue visitShiftByConstant(SDNode *N, ConstantSDNode *Amt);
314 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
315 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
316 SDValue SimplifySelect(SDLoc DL, SDValue N0, SDValue N1, SDValue N2);
317 SDValue SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1, SDValue N2,
318 SDValue N3, ISD::CondCode CC,
319 bool NotExtCompare = false);
320 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
321 SDLoc DL, bool foldBooleans = true);
323 bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
325 bool isOneUseSetCC(SDValue N) const;
327 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
329 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
330 SDValue CombineExtLoad(SDNode *N);
331 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
332 SDValue BuildSDIV(SDNode *N);
333 SDValue BuildSDIVPow2(SDNode *N);
334 SDValue BuildUDIV(SDNode *N);
335 SDValue BuildReciprocalEstimate(SDValue Op);
336 SDValue BuildRsqrtEstimate(SDValue Op);
337 SDValue BuildRsqrtNROneConst(SDValue Op, SDValue Est, unsigned Iterations);
338 SDValue BuildRsqrtNRTwoConst(SDValue Op, SDValue Est, unsigned Iterations);
339 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
340 bool DemandHighBits = true);
341 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
342 SDNode *MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg,
343 SDValue InnerPos, SDValue InnerNeg,
344 unsigned PosOpcode, unsigned NegOpcode,
346 SDNode *MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL);
347 SDValue ReduceLoadWidth(SDNode *N);
348 SDValue ReduceLoadOpStoreWidth(SDNode *N);
349 SDValue TransformFPLoadStorePair(SDNode *N);
350 SDValue reduceBuildVecExtToExtBuildVec(SDNode *N);
351 SDValue reduceBuildVecConvertToConvertBuildVec(SDNode *N);
353 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
355 /// Walk up chain skipping non-aliasing memory nodes,
356 /// looking for aliasing nodes and adding them to the Aliases vector.
357 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
358 SmallVectorImpl<SDValue> &Aliases);
360 /// Return true if there is any possibility that the two addresses overlap.
361 bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const;
363 /// Walk up chain skipping non-aliasing memory nodes, looking for a better
364 /// chain (aliasing node.)
365 SDValue FindBetterChain(SDNode *N, SDValue Chain);
367 /// Holds a pointer to an LSBaseSDNode as well as information on where it
368 /// is located in a sequence of memory operations connected by a chain.
370 MemOpLink (LSBaseSDNode *N, int64_t Offset, unsigned Seq):
371 MemNode(N), OffsetFromBase(Offset), SequenceNum(Seq) { }
372 // Ptr to the mem node.
373 LSBaseSDNode *MemNode;
374 // Offset from the base ptr.
375 int64_t OffsetFromBase;
376 // What is the sequence number of this mem node.
377 // Lowest mem operand in the DAG starts at zero.
378 unsigned SequenceNum;
381 /// This is a helper function for MergeConsecutiveStores. When the source
382 /// elements of the consecutive stores are all constants or all extracted
383 /// vector elements, try to merge them into one larger store.
384 /// \return True if a merged store was created.
385 bool MergeStoresOfConstantsOrVecElts(SmallVectorImpl<MemOpLink> &StoreNodes,
386 EVT MemVT, unsigned NumElem,
387 bool IsConstantSrc, bool UseVector);
389 /// Merge consecutive store operations into a wide store.
390 /// This optimization uses wide integers or vectors when possible.
391 /// \return True if some memory operations were changed.
392 bool MergeConsecutiveStores(StoreSDNode *N);
394 /// \brief Try to transform a truncation where C is a constant:
395 /// (trunc (and X, C)) -> (and (trunc X), (trunc C))
397 /// \p N needs to be a truncation and its first operand an AND. Other
398 /// requirements are checked by the function (e.g. that trunc is
399 /// single-use) and if missed an empty SDValue is returned.
400 SDValue distributeTruncateThroughAnd(SDNode *N);
403 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
404 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
405 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {
406 AttributeSet FnAttrs =
407 DAG.getMachineFunction().getFunction()->getAttributes();
409 FnAttrs.hasAttribute(AttributeSet::FunctionIndex,
410 Attribute::OptimizeForSize) ||
411 FnAttrs.hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize);
414 /// Runs the dag combiner on all nodes in the work list
415 void Run(CombineLevel AtLevel);
417 SelectionDAG &getDAG() const { return DAG; }
419 /// Returns a type large enough to hold any valid shift amount - before type
420 /// legalization these can be huge.
421 EVT getShiftAmountTy(EVT LHSTy) {
422 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
423 if (LHSTy.isVector())
425 return LegalTypes ? TLI.getScalarShiftAmountTy(LHSTy)
426 : TLI.getPointerTy();
429 /// This method returns true if we are running before type legalization or
430 /// if the specified VT is legal.
431 bool isTypeLegal(const EVT &VT) {
432 if (!LegalTypes) return true;
433 return TLI.isTypeLegal(VT);
436 /// Convenience wrapper around TargetLowering::getSetCCResultType
437 EVT getSetCCResultType(EVT VT) const {
438 return TLI.getSetCCResultType(*DAG.getContext(), VT);
445 /// This class is a DAGUpdateListener that removes any deleted
446 /// nodes from the worklist.
447 class WorklistRemover : public SelectionDAG::DAGUpdateListener {
450 explicit WorklistRemover(DAGCombiner &dc)
451 : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {}
453 void NodeDeleted(SDNode *N, SDNode *E) override {
454 DC.removeFromWorklist(N);
459 //===----------------------------------------------------------------------===//
460 // TargetLowering::DAGCombinerInfo implementation
461 //===----------------------------------------------------------------------===//
463 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
464 ((DAGCombiner*)DC)->AddToWorklist(N);
467 void TargetLowering::DAGCombinerInfo::RemoveFromWorklist(SDNode *N) {
468 ((DAGCombiner*)DC)->removeFromWorklist(N);
471 SDValue TargetLowering::DAGCombinerInfo::
472 CombineTo(SDNode *N, const std::vector<SDValue> &To, bool AddTo) {
473 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
476 SDValue TargetLowering::DAGCombinerInfo::
477 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
478 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
482 SDValue TargetLowering::DAGCombinerInfo::
483 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
484 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
487 void TargetLowering::DAGCombinerInfo::
488 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
489 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
492 //===----------------------------------------------------------------------===//
494 //===----------------------------------------------------------------------===//
496 void DAGCombiner::deleteAndRecombine(SDNode *N) {
497 removeFromWorklist(N);
499 // If the operands of this node are only used by the node, they will now be
500 // dead. Make sure to re-visit them and recursively delete dead nodes.
501 for (const SDValue &Op : N->ops())
502 // For an operand generating multiple values, one of the values may
503 // become dead allowing further simplification (e.g. split index
504 // arithmetic from an indexed load).
505 if (Op->hasOneUse() || Op->getNumValues() > 1)
506 AddToWorklist(Op.getNode());
511 /// Return 1 if we can compute the negated form of the specified expression for
512 /// the same cost as the expression itself, or 2 if we can compute the negated
513 /// form more cheaply than the expression itself.
514 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
515 const TargetLowering &TLI,
516 const TargetOptions *Options,
517 unsigned Depth = 0) {
518 // fneg is removable even if it has multiple uses.
519 if (Op.getOpcode() == ISD::FNEG) return 2;
521 // Don't allow anything with multiple uses.
522 if (!Op.hasOneUse()) return 0;
524 // Don't recurse exponentially.
525 if (Depth > 6) return 0;
527 switch (Op.getOpcode()) {
528 default: return false;
529 case ISD::ConstantFP:
530 // Don't invert constant FP values after legalize. The negated constant
531 // isn't necessarily legal.
532 return LegalOperations ? 0 : 1;
534 // FIXME: determine better conditions for this xform.
535 if (!Options->UnsafeFPMath) return 0;
537 // After operation legalization, it might not be legal to create new FSUBs.
538 if (LegalOperations &&
539 !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType()))
542 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
543 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
546 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
547 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
550 // We can't turn -(A-B) into B-A when we honor signed zeros.
551 if (!Options->UnsafeFPMath) return 0;
553 // fold (fneg (fsub A, B)) -> (fsub B, A)
558 if (Options->HonorSignDependentRoundingFPMath()) return 0;
560 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
561 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
565 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
571 return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options,
576 /// If isNegatibleForFree returns true, return the newly negated expression.
577 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
578 bool LegalOperations, unsigned Depth = 0) {
579 const TargetOptions &Options = DAG.getTarget().Options;
580 // fneg is removable even if it has multiple uses.
581 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
583 // Don't allow anything with multiple uses.
584 assert(Op.hasOneUse() && "Unknown reuse!");
586 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
587 switch (Op.getOpcode()) {
588 default: llvm_unreachable("Unknown code");
589 case ISD::ConstantFP: {
590 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
592 return DAG.getConstantFP(V, Op.getValueType());
595 // FIXME: determine better conditions for this xform.
596 assert(Options.UnsafeFPMath);
598 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
599 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
600 DAG.getTargetLoweringInfo(), &Options, Depth+1))
601 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
602 GetNegatedExpression(Op.getOperand(0), DAG,
603 LegalOperations, Depth+1),
605 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
606 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
607 GetNegatedExpression(Op.getOperand(1), DAG,
608 LegalOperations, Depth+1),
611 // We can't turn -(A-B) into B-A when we honor signed zeros.
612 assert(Options.UnsafeFPMath);
614 // fold (fneg (fsub 0, B)) -> B
615 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
616 if (N0CFP->getValueAPF().isZero())
617 return Op.getOperand(1);
619 // fold (fneg (fsub A, B)) -> (fsub B, A)
620 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
621 Op.getOperand(1), Op.getOperand(0));
625 assert(!Options.HonorSignDependentRoundingFPMath());
627 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
628 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
629 DAG.getTargetLoweringInfo(), &Options, Depth+1))
630 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
631 GetNegatedExpression(Op.getOperand(0), DAG,
632 LegalOperations, Depth+1),
635 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
636 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
638 GetNegatedExpression(Op.getOperand(1), DAG,
639 LegalOperations, Depth+1));
643 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
644 GetNegatedExpression(Op.getOperand(0), DAG,
645 LegalOperations, Depth+1));
647 return DAG.getNode(ISD::FP_ROUND, SDLoc(Op), Op.getValueType(),
648 GetNegatedExpression(Op.getOperand(0), DAG,
649 LegalOperations, Depth+1),
654 // Return true if this node is a setcc, or is a select_cc
655 // that selects between the target values used for true and false, making it
656 // equivalent to a setcc. Also, set the incoming LHS, RHS, and CC references to
657 // the appropriate nodes based on the type of node we are checking. This
658 // simplifies life a bit for the callers.
659 bool DAGCombiner::isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
661 if (N.getOpcode() == ISD::SETCC) {
662 LHS = N.getOperand(0);
663 RHS = N.getOperand(1);
664 CC = N.getOperand(2);
668 if (N.getOpcode() != ISD::SELECT_CC ||
669 !TLI.isConstTrueVal(N.getOperand(2).getNode()) ||
670 !TLI.isConstFalseVal(N.getOperand(3).getNode()))
673 if (TLI.getBooleanContents(N.getValueType()) ==
674 TargetLowering::UndefinedBooleanContent)
677 LHS = N.getOperand(0);
678 RHS = N.getOperand(1);
679 CC = N.getOperand(4);
683 /// Return true if this is a SetCC-equivalent operation with only one use.
684 /// If this is true, it allows the users to invert the operation for free when
685 /// it is profitable to do so.
686 bool DAGCombiner::isOneUseSetCC(SDValue N) const {
688 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
693 /// Returns true if N is a BUILD_VECTOR node whose
694 /// elements are all the same constant or undefined.
695 static bool isConstantSplatVector(SDNode *N, APInt& SplatValue) {
696 BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(N);
701 unsigned SplatBitSize;
703 EVT EltVT = N->getValueType(0).getVectorElementType();
704 return (C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
706 EltVT.getSizeInBits() >= SplatBitSize);
709 // \brief Returns the SDNode if it is a constant BuildVector or constant.
710 static SDNode *isConstantBuildVectorOrConstantInt(SDValue N) {
711 if (isa<ConstantSDNode>(N))
713 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N);
714 if (BV && BV->isConstant())
719 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
721 static ConstantSDNode *isConstOrConstSplat(SDValue N) {
722 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N))
725 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
726 BitVector UndefElements;
727 ConstantSDNode *CN = BV->getConstantSplatNode(&UndefElements);
729 // BuildVectors can truncate their operands. Ignore that case here.
730 // FIXME: We blindly ignore splats which include undef which is overly
732 if (CN && UndefElements.none() &&
733 CN->getValueType(0) == N.getValueType().getScalarType())
740 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
742 static ConstantFPSDNode *isConstOrConstSplatFP(SDValue N) {
743 if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N))
746 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
747 BitVector UndefElements;
748 ConstantFPSDNode *CN = BV->getConstantFPSplatNode(&UndefElements);
750 if (CN && UndefElements.none())
757 SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL,
758 SDValue N0, SDValue N1) {
759 EVT VT = N0.getValueType();
760 if (N0.getOpcode() == Opc) {
761 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0.getOperand(1))) {
762 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1)) {
763 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
764 if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, L, R))
765 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
768 if (N0.hasOneUse()) {
769 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one
771 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1);
772 if (!OpNode.getNode())
774 AddToWorklist(OpNode.getNode());
775 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
780 if (N1.getOpcode() == Opc) {
781 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1.getOperand(1))) {
782 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0)) {
783 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
784 if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, R, L))
785 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
788 if (N1.hasOneUse()) {
789 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one
791 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N1.getOperand(0), N0);
792 if (!OpNode.getNode())
794 AddToWorklist(OpNode.getNode());
795 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
803 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
805 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
807 DEBUG(dbgs() << "\nReplacing.1 ";
809 dbgs() << "\nWith: ";
810 To[0].getNode()->dump(&DAG);
811 dbgs() << " and " << NumTo-1 << " other values\n");
812 for (unsigned i = 0, e = NumTo; i != e; ++i)
813 assert((!To[i].getNode() ||
814 N->getValueType(i) == To[i].getValueType()) &&
815 "Cannot combine value to value of different type!");
817 WorklistRemover DeadNodes(*this);
818 DAG.ReplaceAllUsesWith(N, To);
820 // Push the new nodes and any users onto the worklist
821 for (unsigned i = 0, e = NumTo; i != e; ++i) {
822 if (To[i].getNode()) {
823 AddToWorklist(To[i].getNode());
824 AddUsersToWorklist(To[i].getNode());
829 // Finally, if the node is now dead, remove it from the graph. The node
830 // may not be dead if the replacement process recursively simplified to
831 // something else needing this node.
833 deleteAndRecombine(N);
834 return SDValue(N, 0);
838 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
839 // Replace all uses. If any nodes become isomorphic to other nodes and
840 // are deleted, make sure to remove them from our worklist.
841 WorklistRemover DeadNodes(*this);
842 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New);
844 // Push the new node and any (possibly new) users onto the worklist.
845 AddToWorklist(TLO.New.getNode());
846 AddUsersToWorklist(TLO.New.getNode());
848 // Finally, if the node is now dead, remove it from the graph. The node
849 // may not be dead if the replacement process recursively simplified to
850 // something else needing this node.
851 if (TLO.Old.getNode()->use_empty())
852 deleteAndRecombine(TLO.Old.getNode());
855 /// Check the specified integer node value to see if it can be simplified or if
856 /// things it uses can be simplified by bit propagation. If so, return true.
857 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
858 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
859 APInt KnownZero, KnownOne;
860 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
864 AddToWorklist(Op.getNode());
866 // Replace the old value with the new one.
868 DEBUG(dbgs() << "\nReplacing.2 ";
869 TLO.Old.getNode()->dump(&DAG);
870 dbgs() << "\nWith: ";
871 TLO.New.getNode()->dump(&DAG);
874 CommitTargetLoweringOpt(TLO);
878 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
880 EVT VT = Load->getValueType(0);
881 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
883 DEBUG(dbgs() << "\nReplacing.9 ";
885 dbgs() << "\nWith: ";
886 Trunc.getNode()->dump(&DAG);
888 WorklistRemover DeadNodes(*this);
889 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc);
890 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1));
891 deleteAndRecombine(Load);
892 AddToWorklist(Trunc.getNode());
895 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
898 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
899 EVT MemVT = LD->getMemoryVT();
900 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
901 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, PVT, MemVT) ? ISD::ZEXTLOAD
903 : LD->getExtensionType();
905 return DAG.getExtLoad(ExtType, dl, PVT,
906 LD->getChain(), LD->getBasePtr(),
907 MemVT, LD->getMemOperand());
910 unsigned Opc = Op.getOpcode();
913 case ISD::AssertSext:
914 return DAG.getNode(ISD::AssertSext, dl, PVT,
915 SExtPromoteOperand(Op.getOperand(0), PVT),
917 case ISD::AssertZext:
918 return DAG.getNode(ISD::AssertZext, dl, PVT,
919 ZExtPromoteOperand(Op.getOperand(0), PVT),
921 case ISD::Constant: {
923 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
924 return DAG.getNode(ExtOpc, dl, PVT, Op);
928 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
930 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
933 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
934 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
936 EVT OldVT = Op.getValueType();
938 bool Replace = false;
939 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
940 if (!NewOp.getNode())
942 AddToWorklist(NewOp.getNode());
945 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
946 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
947 DAG.getValueType(OldVT));
950 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
951 EVT OldVT = Op.getValueType();
953 bool Replace = false;
954 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
955 if (!NewOp.getNode())
957 AddToWorklist(NewOp.getNode());
960 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
961 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
964 /// Promote the specified integer binary operation if the target indicates it is
965 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
966 /// i32 since i16 instructions are longer.
967 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
968 if (!LegalOperations)
971 EVT VT = Op.getValueType();
972 if (VT.isVector() || !VT.isInteger())
975 // If operation type is 'undesirable', e.g. i16 on x86, consider
977 unsigned Opc = Op.getOpcode();
978 if (TLI.isTypeDesirableForOp(Opc, VT))
982 // Consult target whether it is a good idea to promote this operation and
983 // what's the right type to promote it to.
984 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
985 assert(PVT != VT && "Don't know what type to promote to!");
987 bool Replace0 = false;
988 SDValue N0 = Op.getOperand(0);
989 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
993 bool Replace1 = false;
994 SDValue N1 = Op.getOperand(1);
999 NN1 = PromoteOperand(N1, PVT, Replace1);
1004 AddToWorklist(NN0.getNode());
1006 AddToWorklist(NN1.getNode());
1009 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
1011 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
1013 DEBUG(dbgs() << "\nPromoting ";
1014 Op.getNode()->dump(&DAG));
1016 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1017 DAG.getNode(Opc, dl, PVT, NN0, NN1));
1022 /// Promote the specified integer shift operation if the target indicates it is
1023 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
1024 /// i32 since i16 instructions are longer.
1025 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
1026 if (!LegalOperations)
1029 EVT VT = Op.getValueType();
1030 if (VT.isVector() || !VT.isInteger())
1033 // If operation type is 'undesirable', e.g. i16 on x86, consider
1035 unsigned Opc = Op.getOpcode();
1036 if (TLI.isTypeDesirableForOp(Opc, VT))
1040 // Consult target whether it is a good idea to promote this operation and
1041 // what's the right type to promote it to.
1042 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1043 assert(PVT != VT && "Don't know what type to promote to!");
1045 bool Replace = false;
1046 SDValue N0 = Op.getOperand(0);
1047 if (Opc == ISD::SRA)
1048 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
1049 else if (Opc == ISD::SRL)
1050 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
1052 N0 = PromoteOperand(N0, PVT, Replace);
1056 AddToWorklist(N0.getNode());
1058 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
1060 DEBUG(dbgs() << "\nPromoting ";
1061 Op.getNode()->dump(&DAG));
1063 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1064 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
1069 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
1070 if (!LegalOperations)
1073 EVT VT = Op.getValueType();
1074 if (VT.isVector() || !VT.isInteger())
1077 // If operation type is 'undesirable', e.g. i16 on x86, consider
1079 unsigned Opc = Op.getOpcode();
1080 if (TLI.isTypeDesirableForOp(Opc, VT))
1084 // Consult target whether it is a good idea to promote this operation and
1085 // what's the right type to promote it to.
1086 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1087 assert(PVT != VT && "Don't know what type to promote to!");
1088 // fold (aext (aext x)) -> (aext x)
1089 // fold (aext (zext x)) -> (zext x)
1090 // fold (aext (sext x)) -> (sext x)
1091 DEBUG(dbgs() << "\nPromoting ";
1092 Op.getNode()->dump(&DAG));
1093 return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0));
1098 bool DAGCombiner::PromoteLoad(SDValue Op) {
1099 if (!LegalOperations)
1102 EVT VT = Op.getValueType();
1103 if (VT.isVector() || !VT.isInteger())
1106 // If operation type is 'undesirable', e.g. i16 on x86, consider
1108 unsigned Opc = Op.getOpcode();
1109 if (TLI.isTypeDesirableForOp(Opc, VT))
1113 // Consult target whether it is a good idea to promote this operation and
1114 // what's the right type to promote it to.
1115 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1116 assert(PVT != VT && "Don't know what type to promote to!");
1119 SDNode *N = Op.getNode();
1120 LoadSDNode *LD = cast<LoadSDNode>(N);
1121 EVT MemVT = LD->getMemoryVT();
1122 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
1123 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, PVT, MemVT) ? ISD::ZEXTLOAD
1125 : LD->getExtensionType();
1126 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
1127 LD->getChain(), LD->getBasePtr(),
1128 MemVT, LD->getMemOperand());
1129 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
1131 DEBUG(dbgs() << "\nPromoting ";
1134 Result.getNode()->dump(&DAG);
1136 WorklistRemover DeadNodes(*this);
1137 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
1138 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1));
1139 deleteAndRecombine(N);
1140 AddToWorklist(Result.getNode());
1146 /// \brief Recursively delete a node which has no uses and any operands for
1147 /// which it is the only use.
1149 /// Note that this both deletes the nodes and removes them from the worklist.
1150 /// It also adds any nodes who have had a user deleted to the worklist as they
1151 /// may now have only one use and subject to other combines.
1152 bool DAGCombiner::recursivelyDeleteUnusedNodes(SDNode *N) {
1153 if (!N->use_empty())
1156 SmallSetVector<SDNode *, 16> Nodes;
1159 N = Nodes.pop_back_val();
1163 if (N->use_empty()) {
1164 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1165 Nodes.insert(N->getOperand(i).getNode());
1167 removeFromWorklist(N);
1172 } while (!Nodes.empty());
1176 //===----------------------------------------------------------------------===//
1177 // Main DAG Combiner implementation
1178 //===----------------------------------------------------------------------===//
1180 void DAGCombiner::Run(CombineLevel AtLevel) {
1181 // set the instance variables, so that the various visit routines may use it.
1183 LegalOperations = Level >= AfterLegalizeVectorOps;
1184 LegalTypes = Level >= AfterLegalizeTypes;
1186 // Early exit if this basic block is in an optnone function.
1187 AttributeSet FnAttrs =
1188 DAG.getMachineFunction().getFunction()->getAttributes();
1189 if (FnAttrs.hasAttribute(AttributeSet::FunctionIndex,
1190 Attribute::OptimizeNone))
1193 // Add all the dag nodes to the worklist.
1194 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
1195 E = DAG.allnodes_end(); I != E; ++I)
1198 // Create a dummy node (which is not added to allnodes), that adds a reference
1199 // to the root node, preventing it from being deleted, and tracking any
1200 // changes of the root.
1201 HandleSDNode Dummy(DAG.getRoot());
1203 // while the worklist isn't empty, find a node and
1204 // try and combine it.
1205 while (!WorklistMap.empty()) {
1207 // The Worklist holds the SDNodes in order, but it may contain null entries.
1209 N = Worklist.pop_back_val();
1212 bool GoodWorklistEntry = WorklistMap.erase(N);
1213 (void)GoodWorklistEntry;
1214 assert(GoodWorklistEntry &&
1215 "Found a worklist entry without a corresponding map entry!");
1217 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1218 // N is deleted from the DAG, since they too may now be dead or may have a
1219 // reduced number of uses, allowing other xforms.
1220 if (recursivelyDeleteUnusedNodes(N))
1223 WorklistRemover DeadNodes(*this);
1225 // If this combine is running after legalizing the DAG, re-legalize any
1226 // nodes pulled off the worklist.
1227 if (Level == AfterLegalizeDAG) {
1228 SmallSetVector<SDNode *, 16> UpdatedNodes;
1229 bool NIsValid = DAG.LegalizeOp(N, UpdatedNodes);
1231 for (SDNode *LN : UpdatedNodes) {
1233 AddUsersToWorklist(LN);
1239 DEBUG(dbgs() << "\nCombining: "; N->dump(&DAG));
1241 // Add any operands of the new node which have not yet been combined to the
1242 // worklist as well. Because the worklist uniques things already, this
1243 // won't repeatedly process the same operand.
1244 CombinedNodes.insert(N);
1245 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1246 if (!CombinedNodes.count(N->getOperand(i).getNode()))
1247 AddToWorklist(N->getOperand(i).getNode());
1249 SDValue RV = combine(N);
1256 // If we get back the same node we passed in, rather than a new node or
1257 // zero, we know that the node must have defined multiple values and
1258 // CombineTo was used. Since CombineTo takes care of the worklist
1259 // mechanics for us, we have no work to do in this case.
1260 if (RV.getNode() == N)
1263 assert(N->getOpcode() != ISD::DELETED_NODE &&
1264 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1265 "Node was deleted but visit returned new node!");
1267 DEBUG(dbgs() << " ... into: ";
1268 RV.getNode()->dump(&DAG));
1270 // Transfer debug value.
1271 DAG.TransferDbgValues(SDValue(N, 0), RV);
1272 if (N->getNumValues() == RV.getNode()->getNumValues())
1273 DAG.ReplaceAllUsesWith(N, RV.getNode());
1275 assert(N->getValueType(0) == RV.getValueType() &&
1276 N->getNumValues() == 1 && "Type mismatch");
1278 DAG.ReplaceAllUsesWith(N, &OpV);
1281 // Push the new node and any users onto the worklist
1282 AddToWorklist(RV.getNode());
1283 AddUsersToWorklist(RV.getNode());
1285 // Finally, if the node is now dead, remove it from the graph. The node
1286 // may not be dead if the replacement process recursively simplified to
1287 // something else needing this node. This will also take care of adding any
1288 // operands which have lost a user to the worklist.
1289 recursivelyDeleteUnusedNodes(N);
1292 // If the root changed (e.g. it was a dead load, update the root).
1293 DAG.setRoot(Dummy.getValue());
1294 DAG.RemoveDeadNodes();
1297 SDValue DAGCombiner::visit(SDNode *N) {
1298 switch (N->getOpcode()) {
1300 case ISD::TokenFactor: return visitTokenFactor(N);
1301 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1302 case ISD::ADD: return visitADD(N);
1303 case ISD::SUB: return visitSUB(N);
1304 case ISD::ADDC: return visitADDC(N);
1305 case ISD::SUBC: return visitSUBC(N);
1306 case ISD::ADDE: return visitADDE(N);
1307 case ISD::SUBE: return visitSUBE(N);
1308 case ISD::MUL: return visitMUL(N);
1309 case ISD::SDIV: return visitSDIV(N);
1310 case ISD::UDIV: return visitUDIV(N);
1311 case ISD::SREM: return visitSREM(N);
1312 case ISD::UREM: return visitUREM(N);
1313 case ISD::MULHU: return visitMULHU(N);
1314 case ISD::MULHS: return visitMULHS(N);
1315 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1316 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1317 case ISD::SMULO: return visitSMULO(N);
1318 case ISD::UMULO: return visitUMULO(N);
1319 case ISD::SDIVREM: return visitSDIVREM(N);
1320 case ISD::UDIVREM: return visitUDIVREM(N);
1321 case ISD::AND: return visitAND(N);
1322 case ISD::OR: return visitOR(N);
1323 case ISD::XOR: return visitXOR(N);
1324 case ISD::SHL: return visitSHL(N);
1325 case ISD::SRA: return visitSRA(N);
1326 case ISD::SRL: return visitSRL(N);
1328 case ISD::ROTL: return visitRotate(N);
1329 case ISD::CTLZ: return visitCTLZ(N);
1330 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1331 case ISD::CTTZ: return visitCTTZ(N);
1332 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1333 case ISD::CTPOP: return visitCTPOP(N);
1334 case ISD::SELECT: return visitSELECT(N);
1335 case ISD::VSELECT: return visitVSELECT(N);
1336 case ISD::SELECT_CC: return visitSELECT_CC(N);
1337 case ISD::SETCC: return visitSETCC(N);
1338 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1339 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1340 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1341 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1342 case ISD::TRUNCATE: return visitTRUNCATE(N);
1343 case ISD::BITCAST: return visitBITCAST(N);
1344 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1345 case ISD::FADD: return visitFADD(N);
1346 case ISD::FSUB: return visitFSUB(N);
1347 case ISD::FMUL: return visitFMUL(N);
1348 case ISD::FMA: return visitFMA(N);
1349 case ISD::FDIV: return visitFDIV(N);
1350 case ISD::FREM: return visitFREM(N);
1351 case ISD::FSQRT: return visitFSQRT(N);
1352 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1353 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1354 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1355 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1356 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1357 case ISD::FP_ROUND: return visitFP_ROUND(N);
1358 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1359 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1360 case ISD::FNEG: return visitFNEG(N);
1361 case ISD::FABS: return visitFABS(N);
1362 case ISD::FFLOOR: return visitFFLOOR(N);
1363 case ISD::FMINNUM: return visitFMINNUM(N);
1364 case ISD::FMAXNUM: return visitFMAXNUM(N);
1365 case ISD::FCEIL: return visitFCEIL(N);
1366 case ISD::FTRUNC: return visitFTRUNC(N);
1367 case ISD::BRCOND: return visitBRCOND(N);
1368 case ISD::BR_CC: return visitBR_CC(N);
1369 case ISD::LOAD: return visitLOAD(N);
1370 case ISD::STORE: return visitSTORE(N);
1371 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1372 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1373 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1374 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1375 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1376 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1377 case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N);
1378 case ISD::MLOAD: return visitMLOAD(N);
1379 case ISD::MSTORE: return visitMSTORE(N);
1384 SDValue DAGCombiner::combine(SDNode *N) {
1385 SDValue RV = visit(N);
1387 // If nothing happened, try a target-specific DAG combine.
1388 if (!RV.getNode()) {
1389 assert(N->getOpcode() != ISD::DELETED_NODE &&
1390 "Node was deleted but visit returned NULL!");
1392 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1393 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1395 // Expose the DAG combiner to the target combiner impls.
1396 TargetLowering::DAGCombinerInfo
1397 DagCombineInfo(DAG, Level, false, this);
1399 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1403 // If nothing happened still, try promoting the operation.
1404 if (!RV.getNode()) {
1405 switch (N->getOpcode()) {
1413 RV = PromoteIntBinOp(SDValue(N, 0));
1418 RV = PromoteIntShiftOp(SDValue(N, 0));
1420 case ISD::SIGN_EXTEND:
1421 case ISD::ZERO_EXTEND:
1422 case ISD::ANY_EXTEND:
1423 RV = PromoteExtend(SDValue(N, 0));
1426 if (PromoteLoad(SDValue(N, 0)))
1432 // If N is a commutative binary node, try commuting it to enable more
1434 if (!RV.getNode() && SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1435 N->getNumValues() == 1) {
1436 SDValue N0 = N->getOperand(0);
1437 SDValue N1 = N->getOperand(1);
1439 // Constant operands are canonicalized to RHS.
1440 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1441 SDValue Ops[] = {N1, N0};
1443 if (const BinaryWithFlagsSDNode *BinNode =
1444 dyn_cast<BinaryWithFlagsSDNode>(N)) {
1445 CSENode = DAG.getNodeIfExists(
1446 N->getOpcode(), N->getVTList(), Ops, BinNode->hasNoUnsignedWrap(),
1447 BinNode->hasNoSignedWrap(), BinNode->isExact());
1449 CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops);
1452 return SDValue(CSENode, 0);
1459 /// Given a node, return its input chain if it has one, otherwise return a null
1461 static SDValue getInputChainForNode(SDNode *N) {
1462 if (unsigned NumOps = N->getNumOperands()) {
1463 if (N->getOperand(0).getValueType() == MVT::Other)
1464 return N->getOperand(0);
1465 if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1466 return N->getOperand(NumOps-1);
1467 for (unsigned i = 1; i < NumOps-1; ++i)
1468 if (N->getOperand(i).getValueType() == MVT::Other)
1469 return N->getOperand(i);
1474 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1475 // If N has two operands, where one has an input chain equal to the other,
1476 // the 'other' chain is redundant.
1477 if (N->getNumOperands() == 2) {
1478 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1479 return N->getOperand(0);
1480 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1481 return N->getOperand(1);
1484 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1485 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1486 SmallPtrSet<SDNode*, 16> SeenOps;
1487 bool Changed = false; // If we should replace this token factor.
1489 // Start out with this token factor.
1492 // Iterate through token factors. The TFs grows when new token factors are
1494 for (unsigned i = 0; i < TFs.size(); ++i) {
1495 SDNode *TF = TFs[i];
1497 // Check each of the operands.
1498 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1499 SDValue Op = TF->getOperand(i);
1501 switch (Op.getOpcode()) {
1502 case ISD::EntryToken:
1503 // Entry tokens don't need to be added to the list. They are
1508 case ISD::TokenFactor:
1509 if (Op.hasOneUse() &&
1510 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1511 // Queue up for processing.
1512 TFs.push_back(Op.getNode());
1513 // Clean up in case the token factor is removed.
1514 AddToWorklist(Op.getNode());
1521 // Only add if it isn't already in the list.
1522 if (SeenOps.insert(Op.getNode()).second)
1533 // If we've change things around then replace token factor.
1536 // The entry token is the only possible outcome.
1537 Result = DAG.getEntryNode();
1539 // New and improved token factor.
1540 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Ops);
1543 // Don't add users to work list.
1544 return CombineTo(N, Result, false);
1550 /// MERGE_VALUES can always be eliminated.
1551 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1552 WorklistRemover DeadNodes(*this);
1553 // Replacing results may cause a different MERGE_VALUES to suddenly
1554 // be CSE'd with N, and carry its uses with it. Iterate until no
1555 // uses remain, to ensure that the node can be safely deleted.
1556 // First add the users of this node to the work list so that they
1557 // can be tried again once they have new operands.
1558 AddUsersToWorklist(N);
1560 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1561 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1562 } while (!N->use_empty());
1563 deleteAndRecombine(N);
1564 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1567 SDValue DAGCombiner::visitADD(SDNode *N) {
1568 SDValue N0 = N->getOperand(0);
1569 SDValue N1 = N->getOperand(1);
1570 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1571 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1572 EVT VT = N0.getValueType();
1575 if (VT.isVector()) {
1576 SDValue FoldedVOp = SimplifyVBinOp(N);
1577 if (FoldedVOp.getNode()) return FoldedVOp;
1579 // fold (add x, 0) -> x, vector edition
1580 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1582 if (ISD::isBuildVectorAllZeros(N0.getNode()))
1586 // fold (add x, undef) -> undef
1587 if (N0.getOpcode() == ISD::UNDEF)
1589 if (N1.getOpcode() == ISD::UNDEF)
1591 // fold (add c1, c2) -> c1+c2
1593 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1594 // canonicalize constant to RHS
1596 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, N0);
1597 // fold (add x, 0) -> x
1598 if (N1C && N1C->isNullValue())
1600 // fold (add Sym, c) -> Sym+c
1601 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1602 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1603 GA->getOpcode() == ISD::GlobalAddress)
1604 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1606 (uint64_t)N1C->getSExtValue());
1607 // fold ((c1-A)+c2) -> (c1+c2)-A
1608 if (N1C && N0.getOpcode() == ISD::SUB)
1609 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1610 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1611 DAG.getConstant(N1C->getAPIntValue()+
1612 N0C->getAPIntValue(), VT),
1615 SDValue RADD = ReassociateOps(ISD::ADD, SDLoc(N), N0, N1);
1618 // fold ((0-A) + B) -> B-A
1619 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1620 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1621 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, N0.getOperand(1));
1622 // fold (A + (0-B)) -> A-B
1623 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1624 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1625 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1.getOperand(1));
1626 // fold (A+(B-A)) -> B
1627 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1628 return N1.getOperand(0);
1629 // fold ((B-A)+A) -> B
1630 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1631 return N0.getOperand(0);
1632 // fold (A+(B-(A+C))) to (B-C)
1633 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1634 N0 == N1.getOperand(1).getOperand(0))
1635 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1636 N1.getOperand(1).getOperand(1));
1637 // fold (A+(B-(C+A))) to (B-C)
1638 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1639 N0 == N1.getOperand(1).getOperand(1))
1640 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1641 N1.getOperand(1).getOperand(0));
1642 // fold (A+((B-A)+or-C)) to (B+or-C)
1643 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1644 N1.getOperand(0).getOpcode() == ISD::SUB &&
1645 N0 == N1.getOperand(0).getOperand(1))
1646 return DAG.getNode(N1.getOpcode(), SDLoc(N), VT,
1647 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1649 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1650 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1651 SDValue N00 = N0.getOperand(0);
1652 SDValue N01 = N0.getOperand(1);
1653 SDValue N10 = N1.getOperand(0);
1654 SDValue N11 = N1.getOperand(1);
1656 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1657 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1658 DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10),
1659 DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11));
1662 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1663 return SDValue(N, 0);
1665 // fold (a+b) -> (a|b) iff a and b share no bits.
1666 if (VT.isInteger() && !VT.isVector()) {
1667 APInt LHSZero, LHSOne;
1668 APInt RHSZero, RHSOne;
1669 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1671 if (LHSZero.getBoolValue()) {
1672 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1674 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1675 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1676 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero){
1677 if (!LegalOperations || TLI.isOperationLegal(ISD::OR, VT))
1678 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1);
1683 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1684 if (N1.getOpcode() == ISD::SHL &&
1685 N1.getOperand(0).getOpcode() == ISD::SUB)
1686 if (ConstantSDNode *C =
1687 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1688 if (C->getAPIntValue() == 0)
1689 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0,
1690 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1691 N1.getOperand(0).getOperand(1),
1693 if (N0.getOpcode() == ISD::SHL &&
1694 N0.getOperand(0).getOpcode() == ISD::SUB)
1695 if (ConstantSDNode *C =
1696 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1697 if (C->getAPIntValue() == 0)
1698 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1,
1699 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1700 N0.getOperand(0).getOperand(1),
1703 if (N1.getOpcode() == ISD::AND) {
1704 SDValue AndOp0 = N1.getOperand(0);
1705 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1706 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1707 unsigned DestBits = VT.getScalarType().getSizeInBits();
1709 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1710 // and similar xforms where the inner op is either ~0 or 0.
1711 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1713 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1717 // add (sext i1), X -> sub X, (zext i1)
1718 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1719 N0.getOperand(0).getValueType() == MVT::i1 &&
1720 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1722 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1723 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1726 // add X, (sextinreg Y i1) -> sub X, (and Y 1)
1727 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
1728 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
1729 if (TN->getVT() == MVT::i1) {
1731 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
1732 DAG.getConstant(1, VT));
1733 return DAG.getNode(ISD::SUB, DL, VT, N0, ZExt);
1740 SDValue DAGCombiner::visitADDC(SDNode *N) {
1741 SDValue N0 = N->getOperand(0);
1742 SDValue N1 = N->getOperand(1);
1743 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1744 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1745 EVT VT = N0.getValueType();
1747 // If the flag result is dead, turn this into an ADD.
1748 if (!N->hasAnyUseOfValue(1))
1749 return CombineTo(N, DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N1),
1750 DAG.getNode(ISD::CARRY_FALSE,
1751 SDLoc(N), MVT::Glue));
1753 // canonicalize constant to RHS.
1755 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N1, N0);
1757 // fold (addc x, 0) -> x + no carry out
1758 if (N1C && N1C->isNullValue())
1759 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1760 SDLoc(N), MVT::Glue));
1762 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1763 APInt LHSZero, LHSOne;
1764 APInt RHSZero, RHSOne;
1765 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1767 if (LHSZero.getBoolValue()) {
1768 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1770 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1771 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1772 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1773 return CombineTo(N, DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1),
1774 DAG.getNode(ISD::CARRY_FALSE,
1775 SDLoc(N), MVT::Glue));
1781 SDValue DAGCombiner::visitADDE(SDNode *N) {
1782 SDValue N0 = N->getOperand(0);
1783 SDValue N1 = N->getOperand(1);
1784 SDValue CarryIn = N->getOperand(2);
1785 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1786 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1788 // canonicalize constant to RHS
1790 return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(),
1793 // fold (adde x, y, false) -> (addc x, y)
1794 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1795 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1);
1800 // Since it may not be valid to emit a fold to zero for vector initializers
1801 // check if we can before folding.
1802 static SDValue tryFoldToZero(SDLoc DL, const TargetLowering &TLI, EVT VT,
1804 bool LegalOperations, bool LegalTypes) {
1806 return DAG.getConstant(0, VT);
1807 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
1808 return DAG.getConstant(0, VT);
1812 SDValue DAGCombiner::visitSUB(SDNode *N) {
1813 SDValue N0 = N->getOperand(0);
1814 SDValue N1 = N->getOperand(1);
1815 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1816 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1817 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? nullptr :
1818 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1819 EVT VT = N0.getValueType();
1822 if (VT.isVector()) {
1823 SDValue FoldedVOp = SimplifyVBinOp(N);
1824 if (FoldedVOp.getNode()) return FoldedVOp;
1826 // fold (sub x, 0) -> x, vector edition
1827 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1831 // fold (sub x, x) -> 0
1832 // FIXME: Refactor this and xor and other similar operations together.
1834 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
1835 // fold (sub c1, c2) -> c1-c2
1837 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1838 // fold (sub x, c) -> (add x, -c)
1840 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0,
1841 DAG.getConstant(-N1C->getAPIntValue(), VT));
1842 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1843 if (N0C && N0C->isAllOnesValue())
1844 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
1845 // fold A-(A-B) -> B
1846 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1847 return N1.getOperand(1);
1848 // fold (A+B)-A -> B
1849 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1850 return N0.getOperand(1);
1851 // fold (A+B)-B -> A
1852 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1853 return N0.getOperand(0);
1854 // fold C2-(A+C1) -> (C2-C1)-A
1855 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1856 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(),
1858 return DAG.getNode(ISD::SUB, SDLoc(N), VT, NewC,
1861 // fold ((A+(B+or-C))-B) -> A+or-C
1862 if (N0.getOpcode() == ISD::ADD &&
1863 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1864 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1865 N0.getOperand(1).getOperand(0) == N1)
1866 return DAG.getNode(N0.getOperand(1).getOpcode(), SDLoc(N), VT,
1867 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1868 // fold ((A+(C+B))-B) -> A+C
1869 if (N0.getOpcode() == ISD::ADD &&
1870 N0.getOperand(1).getOpcode() == ISD::ADD &&
1871 N0.getOperand(1).getOperand(1) == N1)
1872 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1873 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1874 // fold ((A-(B-C))-C) -> A-B
1875 if (N0.getOpcode() == ISD::SUB &&
1876 N0.getOperand(1).getOpcode() == ISD::SUB &&
1877 N0.getOperand(1).getOperand(1) == N1)
1878 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1879 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1881 // If either operand of a sub is undef, the result is undef
1882 if (N0.getOpcode() == ISD::UNDEF)
1884 if (N1.getOpcode() == ISD::UNDEF)
1887 // If the relocation model supports it, consider symbol offsets.
1888 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1889 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1890 // fold (sub Sym, c) -> Sym-c
1891 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1892 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1894 (uint64_t)N1C->getSExtValue());
1895 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1896 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1897 if (GA->getGlobal() == GB->getGlobal())
1898 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1902 // sub X, (sextinreg Y i1) -> add X, (and Y 1)
1903 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
1904 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
1905 if (TN->getVT() == MVT::i1) {
1907 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
1908 DAG.getConstant(1, VT));
1909 return DAG.getNode(ISD::ADD, DL, VT, N0, ZExt);
1916 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1917 SDValue N0 = N->getOperand(0);
1918 SDValue N1 = N->getOperand(1);
1919 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1920 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1921 EVT VT = N0.getValueType();
1923 // If the flag result is dead, turn this into an SUB.
1924 if (!N->hasAnyUseOfValue(1))
1925 return CombineTo(N, DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1),
1926 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1929 // fold (subc x, x) -> 0 + no borrow
1931 return CombineTo(N, DAG.getConstant(0, VT),
1932 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1935 // fold (subc x, 0) -> x + no borrow
1936 if (N1C && N1C->isNullValue())
1937 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1940 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1941 if (N0C && N0C->isAllOnesValue())
1942 return CombineTo(N, DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0),
1943 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1949 SDValue DAGCombiner::visitSUBE(SDNode *N) {
1950 SDValue N0 = N->getOperand(0);
1951 SDValue N1 = N->getOperand(1);
1952 SDValue CarryIn = N->getOperand(2);
1954 // fold (sube x, y, false) -> (subc x, y)
1955 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1956 return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1);
1961 SDValue DAGCombiner::visitMUL(SDNode *N) {
1962 SDValue N0 = N->getOperand(0);
1963 SDValue N1 = N->getOperand(1);
1964 EVT VT = N0.getValueType();
1966 // fold (mul x, undef) -> 0
1967 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1968 return DAG.getConstant(0, VT);
1970 bool N0IsConst = false;
1971 bool N1IsConst = false;
1972 APInt ConstValue0, ConstValue1;
1974 if (VT.isVector()) {
1975 SDValue FoldedVOp = SimplifyVBinOp(N);
1976 if (FoldedVOp.getNode()) return FoldedVOp;
1978 N0IsConst = isConstantSplatVector(N0.getNode(), ConstValue0);
1979 N1IsConst = isConstantSplatVector(N1.getNode(), ConstValue1);
1981 N0IsConst = dyn_cast<ConstantSDNode>(N0) != nullptr;
1982 ConstValue0 = N0IsConst ? (dyn_cast<ConstantSDNode>(N0))->getAPIntValue()
1984 N1IsConst = dyn_cast<ConstantSDNode>(N1) != nullptr;
1985 ConstValue1 = N1IsConst ? (dyn_cast<ConstantSDNode>(N1))->getAPIntValue()
1989 // fold (mul c1, c2) -> c1*c2
1990 if (N0IsConst && N1IsConst)
1991 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0.getNode(), N1.getNode());
1993 // canonicalize constant to RHS
1994 if (N0IsConst && !N1IsConst)
1995 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0);
1996 // fold (mul x, 0) -> 0
1997 if (N1IsConst && ConstValue1 == 0)
1999 // We require a splat of the entire scalar bit width for non-contiguous
2002 ConstValue1.getBitWidth() == VT.getScalarType().getSizeInBits();
2003 // fold (mul x, 1) -> x
2004 if (N1IsConst && ConstValue1 == 1 && IsFullSplat)
2006 // fold (mul x, -1) -> 0-x
2007 if (N1IsConst && ConstValue1.isAllOnesValue())
2008 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2009 DAG.getConstant(0, VT), N0);
2010 // fold (mul x, (1 << c)) -> x << c
2011 if (N1IsConst && ConstValue1.isPowerOf2() && IsFullSplat)
2012 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
2013 DAG.getConstant(ConstValue1.logBase2(),
2014 getShiftAmountTy(N0.getValueType())));
2015 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
2016 if (N1IsConst && (-ConstValue1).isPowerOf2() && IsFullSplat) {
2017 unsigned Log2Val = (-ConstValue1).logBase2();
2018 // FIXME: If the input is something that is easily negated (e.g. a
2019 // single-use add), we should put the negate there.
2020 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2021 DAG.getConstant(0, VT),
2022 DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
2023 DAG.getConstant(Log2Val,
2024 getShiftAmountTy(N0.getValueType()))));
2028 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
2029 if (N1IsConst && N0.getOpcode() == ISD::SHL &&
2030 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2031 isa<ConstantSDNode>(N0.getOperand(1)))) {
2032 SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT,
2033 N1, N0.getOperand(1));
2034 AddToWorklist(C3.getNode());
2035 return DAG.getNode(ISD::MUL, SDLoc(N), VT,
2036 N0.getOperand(0), C3);
2039 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
2042 SDValue Sh(nullptr,0), Y(nullptr,0);
2043 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
2044 if (N0.getOpcode() == ISD::SHL &&
2045 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2046 isa<ConstantSDNode>(N0.getOperand(1))) &&
2047 N0.getNode()->hasOneUse()) {
2049 } else if (N1.getOpcode() == ISD::SHL &&
2050 isa<ConstantSDNode>(N1.getOperand(1)) &&
2051 N1.getNode()->hasOneUse()) {
2056 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2057 Sh.getOperand(0), Y);
2058 return DAG.getNode(ISD::SHL, SDLoc(N), VT,
2059 Mul, Sh.getOperand(1));
2063 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
2064 if (N1IsConst && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
2065 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2066 isa<ConstantSDNode>(N0.getOperand(1))))
2067 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
2068 DAG.getNode(ISD::MUL, SDLoc(N0), VT,
2069 N0.getOperand(0), N1),
2070 DAG.getNode(ISD::MUL, SDLoc(N1), VT,
2071 N0.getOperand(1), N1));
2074 SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1);
2081 SDValue DAGCombiner::visitSDIV(SDNode *N) {
2082 SDValue N0 = N->getOperand(0);
2083 SDValue N1 = N->getOperand(1);
2084 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2085 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2086 EVT VT = N->getValueType(0);
2089 if (VT.isVector()) {
2090 SDValue FoldedVOp = SimplifyVBinOp(N);
2091 if (FoldedVOp.getNode()) return FoldedVOp;
2094 // fold (sdiv c1, c2) -> c1/c2
2095 if (N0C && N1C && !N1C->isNullValue())
2096 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
2097 // fold (sdiv X, 1) -> X
2098 if (N1C && N1C->getAPIntValue() == 1LL)
2100 // fold (sdiv X, -1) -> 0-X
2101 if (N1C && N1C->isAllOnesValue())
2102 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2103 DAG.getConstant(0, VT), N0);
2104 // If we know the sign bits of both operands are zero, strength reduce to a
2105 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
2106 if (!VT.isVector()) {
2107 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2108 return DAG.getNode(ISD::UDIV, SDLoc(N), N1.getValueType(),
2112 // fold (sdiv X, pow2) -> simple ops after legalize
2113 if (N1C && !N1C->isNullValue() && (N1C->getAPIntValue().isPowerOf2() ||
2114 (-N1C->getAPIntValue()).isPowerOf2())) {
2115 // If dividing by powers of two is cheap, then don't perform the following
2117 if (TLI.isPow2SDivCheap())
2120 // Target-specific implementation of sdiv x, pow2.
2121 SDValue Res = BuildSDIVPow2(N);
2125 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
2127 // Splat the sign bit into the register
2129 DAG.getNode(ISD::SRA, SDLoc(N), VT, N0,
2130 DAG.getConstant(VT.getScalarSizeInBits() - 1,
2131 getShiftAmountTy(N0.getValueType())));
2132 AddToWorklist(SGN.getNode());
2134 // Add (N0 < 0) ? abs2 - 1 : 0;
2136 DAG.getNode(ISD::SRL, SDLoc(N), VT, SGN,
2137 DAG.getConstant(VT.getScalarSizeInBits() - lg2,
2138 getShiftAmountTy(SGN.getValueType())));
2139 SDValue ADD = DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, SRL);
2140 AddToWorklist(SRL.getNode());
2141 AddToWorklist(ADD.getNode()); // Divide by pow2
2142 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), VT, ADD,
2143 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType())));
2145 // If we're dividing by a positive value, we're done. Otherwise, we must
2146 // negate the result.
2147 if (N1C->getAPIntValue().isNonNegative())
2150 AddToWorklist(SRA.getNode());
2151 return DAG.getNode(ISD::SUB, SDLoc(N), VT, DAG.getConstant(0, VT), SRA);
2154 // if integer divide is expensive and we satisfy the requirements, emit an
2155 // alternate sequence.
2156 if (N1C && !TLI.isIntDivCheap()) {
2157 SDValue Op = BuildSDIV(N);
2158 if (Op.getNode()) return Op;
2162 if (N0.getOpcode() == ISD::UNDEF)
2163 return DAG.getConstant(0, VT);
2164 // X / undef -> undef
2165 if (N1.getOpcode() == ISD::UNDEF)
2171 SDValue DAGCombiner::visitUDIV(SDNode *N) {
2172 SDValue N0 = N->getOperand(0);
2173 SDValue N1 = N->getOperand(1);
2174 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2175 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2176 EVT VT = N->getValueType(0);
2179 if (VT.isVector()) {
2180 SDValue FoldedVOp = SimplifyVBinOp(N);
2181 if (FoldedVOp.getNode()) return FoldedVOp;
2184 // fold (udiv c1, c2) -> c1/c2
2185 if (N0C && N1C && !N1C->isNullValue())
2186 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
2187 // fold (udiv x, (1 << c)) -> x >>u c
2188 if (N1C && N1C->getAPIntValue().isPowerOf2())
2189 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0,
2190 DAG.getConstant(N1C->getAPIntValue().logBase2(),
2191 getShiftAmountTy(N0.getValueType())));
2192 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
2193 if (N1.getOpcode() == ISD::SHL) {
2194 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2195 if (SHC->getAPIntValue().isPowerOf2()) {
2196 EVT ADDVT = N1.getOperand(1).getValueType();
2197 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N), ADDVT,
2199 DAG.getConstant(SHC->getAPIntValue()
2202 AddToWorklist(Add.getNode());
2203 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, Add);
2207 // fold (udiv x, c) -> alternate
2208 if (N1C && !TLI.isIntDivCheap()) {
2209 SDValue Op = BuildUDIV(N);
2210 if (Op.getNode()) return Op;
2214 if (N0.getOpcode() == ISD::UNDEF)
2215 return DAG.getConstant(0, VT);
2216 // X / undef -> undef
2217 if (N1.getOpcode() == ISD::UNDEF)
2223 SDValue DAGCombiner::visitSREM(SDNode *N) {
2224 SDValue N0 = N->getOperand(0);
2225 SDValue N1 = N->getOperand(1);
2226 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2227 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2228 EVT VT = N->getValueType(0);
2230 // fold (srem c1, c2) -> c1%c2
2231 if (N0C && N1C && !N1C->isNullValue())
2232 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
2233 // If we know the sign bits of both operands are zero, strength reduce to a
2234 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
2235 if (!VT.isVector()) {
2236 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2237 return DAG.getNode(ISD::UREM, SDLoc(N), VT, N0, N1);
2240 // If X/C can be simplified by the division-by-constant logic, lower
2241 // X%C to the equivalent of X-X/C*C.
2242 if (N1C && !N1C->isNullValue()) {
2243 SDValue Div = DAG.getNode(ISD::SDIV, SDLoc(N), VT, N0, N1);
2244 AddToWorklist(Div.getNode());
2245 SDValue OptimizedDiv = combine(Div.getNode());
2246 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2247 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2249 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2250 AddToWorklist(Mul.getNode());
2256 if (N0.getOpcode() == ISD::UNDEF)
2257 return DAG.getConstant(0, VT);
2258 // X % undef -> undef
2259 if (N1.getOpcode() == ISD::UNDEF)
2265 SDValue DAGCombiner::visitUREM(SDNode *N) {
2266 SDValue N0 = N->getOperand(0);
2267 SDValue N1 = N->getOperand(1);
2268 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2269 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2270 EVT VT = N->getValueType(0);
2272 // fold (urem c1, c2) -> c1%c2
2273 if (N0C && N1C && !N1C->isNullValue())
2274 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
2275 // fold (urem x, pow2) -> (and x, pow2-1)
2276 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
2277 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0,
2278 DAG.getConstant(N1C->getAPIntValue()-1,VT));
2279 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2280 if (N1.getOpcode() == ISD::SHL) {
2281 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2282 if (SHC->getAPIntValue().isPowerOf2()) {
2284 DAG.getNode(ISD::ADD, SDLoc(N), VT, N1,
2285 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
2287 AddToWorklist(Add.getNode());
2288 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, Add);
2293 // If X/C can be simplified by the division-by-constant logic, lower
2294 // X%C to the equivalent of X-X/C*C.
2295 if (N1C && !N1C->isNullValue()) {
2296 SDValue Div = DAG.getNode(ISD::UDIV, SDLoc(N), VT, N0, N1);
2297 AddToWorklist(Div.getNode());
2298 SDValue OptimizedDiv = combine(Div.getNode());
2299 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2300 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2302 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2303 AddToWorklist(Mul.getNode());
2309 if (N0.getOpcode() == ISD::UNDEF)
2310 return DAG.getConstant(0, VT);
2311 // X % undef -> undef
2312 if (N1.getOpcode() == ISD::UNDEF)
2318 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2319 SDValue N0 = N->getOperand(0);
2320 SDValue N1 = N->getOperand(1);
2321 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2322 EVT VT = N->getValueType(0);
2325 // fold (mulhs x, 0) -> 0
2326 if (N1C && N1C->isNullValue())
2328 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2329 if (N1C && N1C->getAPIntValue() == 1)
2330 return DAG.getNode(ISD::SRA, SDLoc(N), N0.getValueType(), N0,
2331 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2332 getShiftAmountTy(N0.getValueType())));
2333 // fold (mulhs x, undef) -> 0
2334 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2335 return DAG.getConstant(0, VT);
2337 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2339 if (VT.isSimple() && !VT.isVector()) {
2340 MVT Simple = VT.getSimpleVT();
2341 unsigned SimpleSize = Simple.getSizeInBits();
2342 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2343 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2344 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2345 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2346 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2347 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2348 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2349 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2356 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2357 SDValue N0 = N->getOperand(0);
2358 SDValue N1 = N->getOperand(1);
2359 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2360 EVT VT = N->getValueType(0);
2363 // fold (mulhu x, 0) -> 0
2364 if (N1C && N1C->isNullValue())
2366 // fold (mulhu x, 1) -> 0
2367 if (N1C && N1C->getAPIntValue() == 1)
2368 return DAG.getConstant(0, N0.getValueType());
2369 // fold (mulhu x, undef) -> 0
2370 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2371 return DAG.getConstant(0, VT);
2373 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2375 if (VT.isSimple() && !VT.isVector()) {
2376 MVT Simple = VT.getSimpleVT();
2377 unsigned SimpleSize = Simple.getSizeInBits();
2378 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2379 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2380 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2381 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2382 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2383 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2384 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2385 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2392 /// Perform optimizations common to nodes that compute two values. LoOp and HiOp
2393 /// give the opcodes for the two computations that are being performed. Return
2394 /// true if a simplification was made.
2395 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2397 // If the high half is not needed, just compute the low half.
2398 bool HiExists = N->hasAnyUseOfValue(1);
2400 (!LegalOperations ||
2401 TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) {
2402 SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2403 return CombineTo(N, Res, Res);
2406 // If the low half is not needed, just compute the high half.
2407 bool LoExists = N->hasAnyUseOfValue(0);
2409 (!LegalOperations ||
2410 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2411 SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2412 return CombineTo(N, Res, Res);
2415 // If both halves are used, return as it is.
2416 if (LoExists && HiExists)
2419 // If the two computed results can be simplified separately, separate them.
2421 SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2422 AddToWorklist(Lo.getNode());
2423 SDValue LoOpt = combine(Lo.getNode());
2424 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2425 (!LegalOperations ||
2426 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2427 return CombineTo(N, LoOpt, LoOpt);
2431 SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2432 AddToWorklist(Hi.getNode());
2433 SDValue HiOpt = combine(Hi.getNode());
2434 if (HiOpt.getNode() && HiOpt != Hi &&
2435 (!LegalOperations ||
2436 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2437 return CombineTo(N, HiOpt, HiOpt);
2443 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2444 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2445 if (Res.getNode()) return Res;
2447 EVT VT = N->getValueType(0);
2450 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2452 if (VT.isSimple() && !VT.isVector()) {
2453 MVT Simple = VT.getSimpleVT();
2454 unsigned SimpleSize = Simple.getSizeInBits();
2455 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2456 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2457 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2458 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2459 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2460 // Compute the high part as N1.
2461 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2462 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2463 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2464 // Compute the low part as N0.
2465 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2466 return CombineTo(N, Lo, Hi);
2473 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2474 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2475 if (Res.getNode()) return Res;
2477 EVT VT = N->getValueType(0);
2480 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2482 if (VT.isSimple() && !VT.isVector()) {
2483 MVT Simple = VT.getSimpleVT();
2484 unsigned SimpleSize = Simple.getSizeInBits();
2485 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2486 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2487 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2488 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2489 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2490 // Compute the high part as N1.
2491 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2492 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2493 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2494 // Compute the low part as N0.
2495 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2496 return CombineTo(N, Lo, Hi);
2503 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2504 // (smulo x, 2) -> (saddo x, x)
2505 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2506 if (C2->getAPIntValue() == 2)
2507 return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(),
2508 N->getOperand(0), N->getOperand(0));
2513 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2514 // (umulo x, 2) -> (uaddo x, x)
2515 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2516 if (C2->getAPIntValue() == 2)
2517 return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(),
2518 N->getOperand(0), N->getOperand(0));
2523 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2524 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2525 if (Res.getNode()) return Res;
2530 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2531 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2532 if (Res.getNode()) return Res;
2537 /// If this is a binary operator with two operands of the same opcode, try to
2539 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2540 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2541 EVT VT = N0.getValueType();
2542 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2544 // Bail early if none of these transforms apply.
2545 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2547 // For each of OP in AND/OR/XOR:
2548 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2549 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2550 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2551 // fold (OP (bswap x), (bswap y)) -> (bswap (OP x, y))
2552 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2554 // do not sink logical op inside of a vector extend, since it may combine
2556 EVT Op0VT = N0.getOperand(0).getValueType();
2557 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2558 N0.getOpcode() == ISD::SIGN_EXTEND ||
2559 N0.getOpcode() == ISD::BSWAP ||
2560 // Avoid infinite looping with PromoteIntBinOp.
2561 (N0.getOpcode() == ISD::ANY_EXTEND &&
2562 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2563 (N0.getOpcode() == ISD::TRUNCATE &&
2564 (!TLI.isZExtFree(VT, Op0VT) ||
2565 !TLI.isTruncateFree(Op0VT, VT)) &&
2566 TLI.isTypeLegal(Op0VT))) &&
2568 Op0VT == N1.getOperand(0).getValueType() &&
2569 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2570 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2571 N0.getOperand(0).getValueType(),
2572 N0.getOperand(0), N1.getOperand(0));
2573 AddToWorklist(ORNode.getNode());
2574 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode);
2577 // For each of OP in SHL/SRL/SRA/AND...
2578 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2579 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2580 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2581 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2582 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2583 N0.getOperand(1) == N1.getOperand(1)) {
2584 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2585 N0.getOperand(0).getValueType(),
2586 N0.getOperand(0), N1.getOperand(0));
2587 AddToWorklist(ORNode.getNode());
2588 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
2589 ORNode, N0.getOperand(1));
2592 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2593 // Only perform this optimization after type legalization and before
2594 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2595 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2596 // we don't want to undo this promotion.
2597 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2599 if ((N0.getOpcode() == ISD::BITCAST ||
2600 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) &&
2601 Level == AfterLegalizeTypes) {
2602 SDValue In0 = N0.getOperand(0);
2603 SDValue In1 = N1.getOperand(0);
2604 EVT In0Ty = In0.getValueType();
2605 EVT In1Ty = In1.getValueType();
2607 // If both incoming values are integers, and the original types are the
2609 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2610 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1);
2611 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op);
2612 AddToWorklist(Op.getNode());
2617 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2618 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2619 // If both shuffles use the same mask, and both shuffle within a single
2620 // vector, then it is worthwhile to move the swizzle after the operation.
2621 // The type-legalizer generates this pattern when loading illegal
2622 // vector types from memory. In many cases this allows additional shuffle
2624 // There are other cases where moving the shuffle after the xor/and/or
2625 // is profitable even if shuffles don't perform a swizzle.
2626 // If both shuffles use the same mask, and both shuffles have the same first
2627 // or second operand, then it might still be profitable to move the shuffle
2628 // after the xor/and/or operation.
2629 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) {
2630 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2631 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2633 assert(N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() &&
2634 "Inputs to shuffles are not the same type");
2636 // Check that both shuffles use the same mask. The masks are known to be of
2637 // the same length because the result vector type is the same.
2638 // Check also that shuffles have only one use to avoid introducing extra
2640 if (SVN0->hasOneUse() && SVN1->hasOneUse() &&
2641 SVN0->getMask().equals(SVN1->getMask())) {
2642 SDValue ShOp = N0->getOperand(1);
2644 // Don't try to fold this node if it requires introducing a
2645 // build vector of all zeros that might be illegal at this stage.
2646 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2648 ShOp = DAG.getConstant(0, VT);
2653 // (AND (shuf (A, C), shuf (B, C)) -> shuf (AND (A, B), C)
2654 // (OR (shuf (A, C), shuf (B, C)) -> shuf (OR (A, B), C)
2655 // (XOR (shuf (A, C), shuf (B, C)) -> shuf (XOR (A, B), V_0)
2656 if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) {
2657 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2658 N0->getOperand(0), N1->getOperand(0));
2659 AddToWorklist(NewNode.getNode());
2660 return DAG.getVectorShuffle(VT, SDLoc(N), NewNode, ShOp,
2661 &SVN0->getMask()[0]);
2664 // Don't try to fold this node if it requires introducing a
2665 // build vector of all zeros that might be illegal at this stage.
2666 ShOp = N0->getOperand(0);
2667 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2669 ShOp = DAG.getConstant(0, VT);
2674 // (AND (shuf (C, A), shuf (C, B)) -> shuf (C, AND (A, B))
2675 // (OR (shuf (C, A), shuf (C, B)) -> shuf (C, OR (A, B))
2676 // (XOR (shuf (C, A), shuf (C, B)) -> shuf (V_0, XOR (A, B))
2677 if (N0->getOperand(0) == N1->getOperand(0) && ShOp.getNode()) {
2678 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2679 N0->getOperand(1), N1->getOperand(1));
2680 AddToWorklist(NewNode.getNode());
2681 return DAG.getVectorShuffle(VT, SDLoc(N), ShOp, NewNode,
2682 &SVN0->getMask()[0]);
2690 SDValue DAGCombiner::visitAND(SDNode *N) {
2691 SDValue N0 = N->getOperand(0);
2692 SDValue N1 = N->getOperand(1);
2693 SDValue LL, LR, RL, RR, CC0, CC1;
2694 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2695 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2696 EVT VT = N1.getValueType();
2697 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2700 if (VT.isVector()) {
2701 SDValue FoldedVOp = SimplifyVBinOp(N);
2702 if (FoldedVOp.getNode()) return FoldedVOp;
2704 // fold (and x, 0) -> 0, vector edition
2705 if (ISD::isBuildVectorAllZeros(N0.getNode()))
2706 // do not return N0, because undef node may exist in N0
2707 return DAG.getConstant(
2708 APInt::getNullValue(
2709 N0.getValueType().getScalarType().getSizeInBits()),
2711 if (ISD::isBuildVectorAllZeros(N1.getNode()))
2712 // do not return N1, because undef node may exist in N1
2713 return DAG.getConstant(
2714 APInt::getNullValue(
2715 N1.getValueType().getScalarType().getSizeInBits()),
2718 // fold (and x, -1) -> x, vector edition
2719 if (ISD::isBuildVectorAllOnes(N0.getNode()))
2721 if (ISD::isBuildVectorAllOnes(N1.getNode()))
2725 // fold (and x, undef) -> 0
2726 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2727 return DAG.getConstant(0, VT);
2728 // fold (and c1, c2) -> c1&c2
2730 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2731 // canonicalize constant to RHS
2733 return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0);
2734 // fold (and x, -1) -> x
2735 if (N1C && N1C->isAllOnesValue())
2737 // if (and x, c) is known to be zero, return 0
2738 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2739 APInt::getAllOnesValue(BitWidth)))
2740 return DAG.getConstant(0, VT);
2742 SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1);
2745 // fold (and (or x, C), D) -> D if (C & D) == D
2746 if (N1C && N0.getOpcode() == ISD::OR)
2747 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2748 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2750 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2751 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2752 SDValue N0Op0 = N0.getOperand(0);
2753 APInt Mask = ~N1C->getAPIntValue();
2754 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2755 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2756 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
2757 N0.getValueType(), N0Op0);
2759 // Replace uses of the AND with uses of the Zero extend node.
2762 // We actually want to replace all uses of the any_extend with the
2763 // zero_extend, to avoid duplicating things. This will later cause this
2764 // AND to be folded.
2765 CombineTo(N0.getNode(), Zext);
2766 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2769 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2770 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2771 // already be zero by virtue of the width of the base type of the load.
2773 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2775 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2776 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2777 N0.getOpcode() == ISD::LOAD) {
2778 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2779 N0 : N0.getOperand(0) );
2781 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2782 // This can be a pure constant or a vector splat, in which case we treat the
2783 // vector as a scalar and use the splat value.
2784 APInt Constant = APInt::getNullValue(1);
2785 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2786 Constant = C->getAPIntValue();
2787 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2788 APInt SplatValue, SplatUndef;
2789 unsigned SplatBitSize;
2791 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2792 SplatBitSize, HasAnyUndefs);
2794 // Undef bits can contribute to a possible optimisation if set, so
2796 SplatValue |= SplatUndef;
2798 // The splat value may be something like "0x00FFFFFF", which means 0 for
2799 // the first vector value and FF for the rest, repeating. We need a mask
2800 // that will apply equally to all members of the vector, so AND all the
2801 // lanes of the constant together.
2802 EVT VT = Vector->getValueType(0);
2803 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2805 // If the splat value has been compressed to a bitlength lower
2806 // than the size of the vector lane, we need to re-expand it to
2808 if (BitWidth > SplatBitSize)
2809 for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
2810 SplatBitSize < BitWidth;
2811 SplatBitSize = SplatBitSize * 2)
2812 SplatValue |= SplatValue.shl(SplatBitSize);
2814 Constant = APInt::getAllOnesValue(BitWidth);
2815 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
2816 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
2820 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
2821 // actually legal and isn't going to get expanded, else this is a false
2823 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
2824 Load->getValueType(0),
2825 Load->getMemoryVT());
2827 // Resize the constant to the same size as the original memory access before
2828 // extension. If it is still the AllOnesValue then this AND is completely
2831 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
2834 switch (Load->getExtensionType()) {
2835 default: B = false; break;
2836 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
2838 case ISD::NON_EXTLOAD: B = true; break;
2841 if (B && Constant.isAllOnesValue()) {
2842 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
2843 // preserve semantics once we get rid of the AND.
2844 SDValue NewLoad(Load, 0);
2845 if (Load->getExtensionType() == ISD::EXTLOAD) {
2846 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
2847 Load->getValueType(0), SDLoc(Load),
2848 Load->getChain(), Load->getBasePtr(),
2849 Load->getOffset(), Load->getMemoryVT(),
2850 Load->getMemOperand());
2851 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
2852 if (Load->getNumValues() == 3) {
2853 // PRE/POST_INC loads have 3 values.
2854 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
2855 NewLoad.getValue(2) };
2856 CombineTo(Load, To, 3, true);
2858 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
2862 // Fold the AND away, taking care not to fold to the old load node if we
2864 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
2866 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2869 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2870 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2871 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2872 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2874 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2875 LL.getValueType().isInteger()) {
2876 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2877 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
2878 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2879 LR.getValueType(), LL, RL);
2880 AddToWorklist(ORNode.getNode());
2881 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
2883 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2884 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2885 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(N0),
2886 LR.getValueType(), LL, RL);
2887 AddToWorklist(ANDNode.getNode());
2888 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
2890 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2891 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2892 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2893 LR.getValueType(), LL, RL);
2894 AddToWorklist(ORNode.getNode());
2895 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
2898 // Simplify (and (setne X, 0), (setne X, -1)) -> (setuge (add X, 1), 2)
2899 if (LL == RL && isa<ConstantSDNode>(LR) && isa<ConstantSDNode>(RR) &&
2900 Op0 == Op1 && LL.getValueType().isInteger() &&
2901 Op0 == ISD::SETNE && ((cast<ConstantSDNode>(LR)->isNullValue() &&
2902 cast<ConstantSDNode>(RR)->isAllOnesValue()) ||
2903 (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
2904 cast<ConstantSDNode>(RR)->isNullValue()))) {
2905 SDValue ADDNode = DAG.getNode(ISD::ADD, SDLoc(N0), LL.getValueType(),
2906 LL, DAG.getConstant(1, LL.getValueType()));
2907 AddToWorklist(ADDNode.getNode());
2908 return DAG.getSetCC(SDLoc(N), VT, ADDNode,
2909 DAG.getConstant(2, LL.getValueType()), ISD::SETUGE);
2911 // canonicalize equivalent to ll == rl
2912 if (LL == RR && LR == RL) {
2913 Op1 = ISD::getSetCCSwappedOperands(Op1);
2916 if (LL == RL && LR == RR) {
2917 bool isInteger = LL.getValueType().isInteger();
2918 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2919 if (Result != ISD::SETCC_INVALID &&
2920 (!LegalOperations ||
2921 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
2922 TLI.isOperationLegal(ISD::SETCC,
2923 getSetCCResultType(N0.getSimpleValueType())))))
2924 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
2929 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
2930 if (N0.getOpcode() == N1.getOpcode()) {
2931 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2932 if (Tmp.getNode()) return Tmp;
2935 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
2936 // fold (and (sra)) -> (and (srl)) when possible.
2937 if (!VT.isVector() &&
2938 SimplifyDemandedBits(SDValue(N, 0)))
2939 return SDValue(N, 0);
2941 // fold (zext_inreg (extload x)) -> (zextload x)
2942 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
2943 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2944 EVT MemVT = LN0->getMemoryVT();
2945 // If we zero all the possible extended bits, then we can turn this into
2946 // a zextload if we are running before legalize or the operation is legal.
2947 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2948 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2949 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2950 ((!LegalOperations && !LN0->isVolatile()) ||
2951 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) {
2952 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2953 LN0->getChain(), LN0->getBasePtr(),
2954 MemVT, LN0->getMemOperand());
2956 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2957 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2960 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
2961 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
2963 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2964 EVT MemVT = LN0->getMemoryVT();
2965 // If we zero all the possible extended bits, then we can turn this into
2966 // a zextload if we are running before legalize or the operation is legal.
2967 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2968 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2969 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2970 ((!LegalOperations && !LN0->isVolatile()) ||
2971 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) {
2972 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2973 LN0->getChain(), LN0->getBasePtr(),
2974 MemVT, LN0->getMemOperand());
2976 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2977 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2981 // fold (and (load x), 255) -> (zextload x, i8)
2982 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2983 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2984 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2985 (N0.getOpcode() == ISD::ANY_EXTEND &&
2986 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2987 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2988 LoadSDNode *LN0 = HasAnyExt
2989 ? cast<LoadSDNode>(N0.getOperand(0))
2990 : cast<LoadSDNode>(N0);
2991 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2992 LN0->isUnindexed() && N0.hasOneUse() && SDValue(LN0, 0).hasOneUse()) {
2993 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2994 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2995 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2996 EVT LoadedVT = LN0->getMemoryVT();
2997 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2999 if (ExtVT == LoadedVT &&
3000 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy,
3004 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
3005 LN0->getChain(), LN0->getBasePtr(), ExtVT,
3006 LN0->getMemOperand());
3008 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
3009 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3012 // Do not change the width of a volatile load.
3013 // Do not generate loads of non-round integer types since these can
3014 // be expensive (and would be wrong if the type is not byte sized).
3015 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
3016 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy,
3018 EVT PtrType = LN0->getOperand(1).getValueType();
3020 unsigned Alignment = LN0->getAlignment();
3021 SDValue NewPtr = LN0->getBasePtr();
3023 // For big endian targets, we need to add an offset to the pointer
3024 // to load the correct bytes. For little endian systems, we merely
3025 // need to read fewer bytes from the same pointer.
3026 if (TLI.isBigEndian()) {
3027 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
3028 unsigned EVTStoreBytes = ExtVT.getStoreSize();
3029 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
3030 NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0), PtrType,
3031 NewPtr, DAG.getConstant(PtrOff, PtrType));
3032 Alignment = MinAlign(Alignment, PtrOff);
3035 AddToWorklist(NewPtr.getNode());
3038 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
3039 LN0->getChain(), NewPtr,
3040 LN0->getPointerInfo(),
3041 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
3042 LN0->isInvariant(), Alignment, LN0->getAAInfo());
3044 CombineTo(LN0, Load, Load.getValue(1));
3045 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3051 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
3052 VT.getSizeInBits() <= 64) {
3053 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3054 APInt ADDC = ADDI->getAPIntValue();
3055 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
3056 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
3057 // immediate for an add, but it is legal if its top c2 bits are set,
3058 // transform the ADD so the immediate doesn't need to be materialized
3060 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
3061 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
3062 SRLI->getZExtValue());
3063 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
3065 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
3067 DAG.getNode(ISD::ADD, SDLoc(N0), VT,
3068 N0.getOperand(0), DAG.getConstant(ADDC, VT));
3069 CombineTo(N0.getNode(), NewAdd);
3070 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3078 // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const)
3079 if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) {
3080 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
3081 N0.getOperand(1), false);
3082 if (BSwap.getNode())
3089 /// Match (a >> 8) | (a << 8) as (bswap a) >> 16.
3090 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
3091 bool DemandHighBits) {
3092 if (!LegalOperations)
3095 EVT VT = N->getValueType(0);
3096 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
3098 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3101 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
3102 bool LookPassAnd0 = false;
3103 bool LookPassAnd1 = false;
3104 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
3106 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
3108 if (N0.getOpcode() == ISD::AND) {
3109 if (!N0.getNode()->hasOneUse())
3111 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3112 if (!N01C || N01C->getZExtValue() != 0xFF00)
3114 N0 = N0.getOperand(0);
3115 LookPassAnd0 = true;
3118 if (N1.getOpcode() == ISD::AND) {
3119 if (!N1.getNode()->hasOneUse())
3121 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3122 if (!N11C || N11C->getZExtValue() != 0xFF)
3124 N1 = N1.getOperand(0);
3125 LookPassAnd1 = true;
3128 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
3130 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
3132 if (!N0.getNode()->hasOneUse() ||
3133 !N1.getNode()->hasOneUse())
3136 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3137 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3140 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
3143 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
3144 SDValue N00 = N0->getOperand(0);
3145 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
3146 if (!N00.getNode()->hasOneUse())
3148 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
3149 if (!N001C || N001C->getZExtValue() != 0xFF)
3151 N00 = N00.getOperand(0);
3152 LookPassAnd0 = true;
3155 SDValue N10 = N1->getOperand(0);
3156 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
3157 if (!N10.getNode()->hasOneUse())
3159 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
3160 if (!N101C || N101C->getZExtValue() != 0xFF00)
3162 N10 = N10.getOperand(0);
3163 LookPassAnd1 = true;
3169 // Make sure everything beyond the low halfword gets set to zero since the SRL
3170 // 16 will clear the top bits.
3171 unsigned OpSizeInBits = VT.getSizeInBits();
3172 if (DemandHighBits && OpSizeInBits > 16) {
3173 // If the left-shift isn't masked out then the only way this is a bswap is
3174 // if all bits beyond the low 8 are 0. In that case the entire pattern
3175 // reduces to a left shift anyway: leave it for other parts of the combiner.
3179 // However, if the right shift isn't masked out then it might be because
3180 // it's not needed. See if we can spot that too.
3181 if (!LookPassAnd1 &&
3182 !DAG.MaskedValueIsZero(
3183 N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16)))
3187 SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00);
3188 if (OpSizeInBits > 16)
3189 Res = DAG.getNode(ISD::SRL, SDLoc(N), VT, Res,
3190 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
3194 /// Return true if the specified node is an element that makes up a 32-bit
3195 /// packed halfword byteswap.
3196 /// ((x & 0x000000ff) << 8) |
3197 /// ((x & 0x0000ff00) >> 8) |
3198 /// ((x & 0x00ff0000) << 8) |
3199 /// ((x & 0xff000000) >> 8)
3200 static bool isBSwapHWordElement(SDValue N, MutableArrayRef<SDNode *> Parts) {
3201 if (!N.getNode()->hasOneUse())
3204 unsigned Opc = N.getOpcode();
3205 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
3208 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3213 switch (N1C->getZExtValue()) {
3216 case 0xFF: Num = 0; break;
3217 case 0xFF00: Num = 1; break;
3218 case 0xFF0000: Num = 2; break;
3219 case 0xFF000000: Num = 3; break;
3222 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
3223 SDValue N0 = N.getOperand(0);
3224 if (Opc == ISD::AND) {
3225 if (Num == 0 || Num == 2) {
3227 // (x >> 8) & 0xff0000
3228 if (N0.getOpcode() != ISD::SRL)
3230 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3231 if (!C || C->getZExtValue() != 8)
3234 // (x << 8) & 0xff00
3235 // (x << 8) & 0xff000000
3236 if (N0.getOpcode() != ISD::SHL)
3238 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3239 if (!C || C->getZExtValue() != 8)
3242 } else if (Opc == ISD::SHL) {
3244 // (x & 0xff0000) << 8
3245 if (Num != 0 && Num != 2)
3247 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3248 if (!C || C->getZExtValue() != 8)
3250 } else { // Opc == ISD::SRL
3251 // (x & 0xff00) >> 8
3252 // (x & 0xff000000) >> 8
3253 if (Num != 1 && Num != 3)
3255 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3256 if (!C || C->getZExtValue() != 8)
3263 Parts[Num] = N0.getOperand(0).getNode();
3267 /// Match a 32-bit packed halfword bswap. That is
3268 /// ((x & 0x000000ff) << 8) |
3269 /// ((x & 0x0000ff00) >> 8) |
3270 /// ((x & 0x00ff0000) << 8) |
3271 /// ((x & 0xff000000) >> 8)
3272 /// => (rotl (bswap x), 16)
3273 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
3274 if (!LegalOperations)
3277 EVT VT = N->getValueType(0);
3280 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3284 // (or (or (and), (and)), (or (and), (and)))
3285 // (or (or (or (and), (and)), (and)), (and))
3286 if (N0.getOpcode() != ISD::OR)
3288 SDValue N00 = N0.getOperand(0);
3289 SDValue N01 = N0.getOperand(1);
3290 SDNode *Parts[4] = {};
3292 if (N1.getOpcode() == ISD::OR &&
3293 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) {
3294 // (or (or (and), (and)), (or (and), (and)))
3295 SDValue N000 = N00.getOperand(0);
3296 if (!isBSwapHWordElement(N000, Parts))
3299 SDValue N001 = N00.getOperand(1);
3300 if (!isBSwapHWordElement(N001, Parts))
3302 SDValue N010 = N01.getOperand(0);
3303 if (!isBSwapHWordElement(N010, Parts))
3305 SDValue N011 = N01.getOperand(1);
3306 if (!isBSwapHWordElement(N011, Parts))
3309 // (or (or (or (and), (and)), (and)), (and))
3310 if (!isBSwapHWordElement(N1, Parts))
3312 if (!isBSwapHWordElement(N01, Parts))
3314 if (N00.getOpcode() != ISD::OR)
3316 SDValue N000 = N00.getOperand(0);
3317 if (!isBSwapHWordElement(N000, Parts))
3319 SDValue N001 = N00.getOperand(1);
3320 if (!isBSwapHWordElement(N001, Parts))
3324 // Make sure the parts are all coming from the same node.
3325 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
3328 SDValue BSwap = DAG.getNode(ISD::BSWAP, SDLoc(N), VT,
3329 SDValue(Parts[0],0));
3331 // Result of the bswap should be rotated by 16. If it's not legal, then
3332 // do (x << 16) | (x >> 16).
3333 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
3334 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3335 return DAG.getNode(ISD::ROTL, SDLoc(N), VT, BSwap, ShAmt);
3336 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3337 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, BSwap, ShAmt);
3338 return DAG.getNode(ISD::OR, SDLoc(N), VT,
3339 DAG.getNode(ISD::SHL, SDLoc(N), VT, BSwap, ShAmt),
3340 DAG.getNode(ISD::SRL, SDLoc(N), VT, BSwap, ShAmt));
3343 SDValue DAGCombiner::visitOR(SDNode *N) {
3344 SDValue N0 = N->getOperand(0);
3345 SDValue N1 = N->getOperand(1);
3346 SDValue LL, LR, RL, RR, CC0, CC1;
3347 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3348 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3349 EVT VT = N1.getValueType();
3352 if (VT.isVector()) {
3353 SDValue FoldedVOp = SimplifyVBinOp(N);
3354 if (FoldedVOp.getNode()) return FoldedVOp;
3356 // fold (or x, 0) -> x, vector edition
3357 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3359 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3362 // fold (or x, -1) -> -1, vector edition
3363 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3364 // do not return N0, because undef node may exist in N0
3365 return DAG.getConstant(
3366 APInt::getAllOnesValue(
3367 N0.getValueType().getScalarType().getSizeInBits()),
3369 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3370 // do not return N1, because undef node may exist in N1
3371 return DAG.getConstant(
3372 APInt::getAllOnesValue(
3373 N1.getValueType().getScalarType().getSizeInBits()),
3376 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask1)
3377 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf B, A, Mask2)
3378 // Do this only if the resulting shuffle is legal.
3379 if (isa<ShuffleVectorSDNode>(N0) &&
3380 isa<ShuffleVectorSDNode>(N1) &&
3381 // Avoid folding a node with illegal type.
3382 TLI.isTypeLegal(VT) &&
3383 N0->getOperand(1) == N1->getOperand(1) &&
3384 ISD::isBuildVectorAllZeros(N0.getOperand(1).getNode())) {
3385 bool CanFold = true;
3386 unsigned NumElts = VT.getVectorNumElements();
3387 const ShuffleVectorSDNode *SV0 = cast<ShuffleVectorSDNode>(N0);
3388 const ShuffleVectorSDNode *SV1 = cast<ShuffleVectorSDNode>(N1);
3389 // We construct two shuffle masks:
3390 // - Mask1 is a shuffle mask for a shuffle with N0 as the first operand
3391 // and N1 as the second operand.
3392 // - Mask2 is a shuffle mask for a shuffle with N1 as the first operand
3393 // and N0 as the second operand.
3394 // We do this because OR is commutable and therefore there might be
3395 // two ways to fold this node into a shuffle.
3396 SmallVector<int,4> Mask1;
3397 SmallVector<int,4> Mask2;
3399 for (unsigned i = 0; i != NumElts && CanFold; ++i) {
3400 int M0 = SV0->getMaskElt(i);
3401 int M1 = SV1->getMaskElt(i);
3403 // Both shuffle indexes are undef. Propagate Undef.
3404 if (M0 < 0 && M1 < 0) {
3405 Mask1.push_back(M0);
3406 Mask2.push_back(M0);
3410 if (M0 < 0 || M1 < 0 ||
3411 (M0 < (int)NumElts && M1 < (int)NumElts) ||
3412 (M0 >= (int)NumElts && M1 >= (int)NumElts)) {
3417 Mask1.push_back(M0 < (int)NumElts ? M0 : M1 + NumElts);
3418 Mask2.push_back(M1 < (int)NumElts ? M1 : M0 + NumElts);
3422 // Fold this sequence only if the resulting shuffle is 'legal'.
3423 if (TLI.isShuffleMaskLegal(Mask1, VT))
3424 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0),
3425 N1->getOperand(0), &Mask1[0]);
3426 if (TLI.isShuffleMaskLegal(Mask2, VT))
3427 return DAG.getVectorShuffle(VT, SDLoc(N), N1->getOperand(0),
3428 N0->getOperand(0), &Mask2[0]);
3433 // fold (or x, undef) -> -1
3434 if (!LegalOperations &&
3435 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3436 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3437 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
3439 // fold (or c1, c2) -> c1|c2
3441 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
3442 // canonicalize constant to RHS
3444 return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0);
3445 // fold (or x, 0) -> x
3446 if (N1C && N1C->isNullValue())
3448 // fold (or x, -1) -> -1
3449 if (N1C && N1C->isAllOnesValue())
3451 // fold (or x, c) -> c iff (x & ~c) == 0
3452 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3455 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3456 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
3457 if (BSwap.getNode())
3459 BSwap = MatchBSwapHWordLow(N, N0, N1);
3460 if (BSwap.getNode())
3464 SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1);
3467 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3468 // iff (c1 & c2) == 0.
3469 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3470 isa<ConstantSDNode>(N0.getOperand(1))) {
3471 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3472 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) {
3473 if (SDValue COR = DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1))
3475 ISD::AND, SDLoc(N), VT,
3476 DAG.getNode(ISD::OR, SDLoc(N0), VT, N0.getOperand(0), N1), COR);
3480 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3481 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3482 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3483 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3485 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3486 LL.getValueType().isInteger()) {
3487 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3488 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3489 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3490 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3491 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(LR),
3492 LR.getValueType(), LL, RL);
3493 AddToWorklist(ORNode.getNode());
3494 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
3496 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3497 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3498 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3499 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3500 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(LR),
3501 LR.getValueType(), LL, RL);
3502 AddToWorklist(ANDNode.getNode());
3503 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
3506 // canonicalize equivalent to ll == rl
3507 if (LL == RR && LR == RL) {
3508 Op1 = ISD::getSetCCSwappedOperands(Op1);
3511 if (LL == RL && LR == RR) {
3512 bool isInteger = LL.getValueType().isInteger();
3513 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3514 if (Result != ISD::SETCC_INVALID &&
3515 (!LegalOperations ||
3516 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
3517 TLI.isOperationLegal(ISD::SETCC,
3518 getSetCCResultType(N0.getValueType())))))
3519 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
3524 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3525 if (N0.getOpcode() == N1.getOpcode()) {
3526 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3527 if (Tmp.getNode()) return Tmp;
3530 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3531 if (N0.getOpcode() == ISD::AND &&
3532 N1.getOpcode() == ISD::AND &&
3533 N0.getOperand(1).getOpcode() == ISD::Constant &&
3534 N1.getOperand(1).getOpcode() == ISD::Constant &&
3535 // Don't increase # computations.
3536 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3537 // We can only do this xform if we know that bits from X that are set in C2
3538 // but not in C1 are already zero. Likewise for Y.
3539 const APInt &LHSMask =
3540 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3541 const APInt &RHSMask =
3542 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3544 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3545 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3546 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3547 N0.getOperand(0), N1.getOperand(0));
3548 return DAG.getNode(ISD::AND, SDLoc(N), VT, X,
3549 DAG.getConstant(LHSMask | RHSMask, VT));
3553 // (or (and X, M), (and X, N)) -> (and X, (or M, N))
3554 if (N0.getOpcode() == ISD::AND &&
3555 N1.getOpcode() == ISD::AND &&
3556 N0.getOperand(0) == N1.getOperand(0) &&
3557 // Don't increase # computations.
3558 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3559 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3560 N0.getOperand(1), N1.getOperand(1));
3561 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0), X);
3564 // See if this is some rotate idiom.
3565 if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N)))
3566 return SDValue(Rot, 0);
3568 // Simplify the operands using demanded-bits information.
3569 if (!VT.isVector() &&
3570 SimplifyDemandedBits(SDValue(N, 0)))
3571 return SDValue(N, 0);
3576 /// Match "(X shl/srl V1) & V2" where V2 may not be present.
3577 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3578 if (Op.getOpcode() == ISD::AND) {
3579 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3580 Mask = Op.getOperand(1);
3581 Op = Op.getOperand(0);
3587 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3595 // Return true if we can prove that, whenever Neg and Pos are both in the
3596 // range [0, OpSize), Neg == (Pos == 0 ? 0 : OpSize - Pos). This means that
3597 // for two opposing shifts shift1 and shift2 and a value X with OpBits bits:
3599 // (or (shift1 X, Neg), (shift2 X, Pos))
3601 // reduces to a rotate in direction shift2 by Pos or (equivalently) a rotate
3602 // in direction shift1 by Neg. The range [0, OpSize) means that we only need
3603 // to consider shift amounts with defined behavior.
3604 static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned OpSize) {
3605 // If OpSize is a power of 2 then:
3607 // (a) (Pos == 0 ? 0 : OpSize - Pos) == (OpSize - Pos) & (OpSize - 1)
3608 // (b) Neg == Neg & (OpSize - 1) whenever Neg is in [0, OpSize).
3610 // So if OpSize is a power of 2 and Neg is (and Neg', OpSize-1), we check
3611 // for the stronger condition:
3613 // Neg & (OpSize - 1) == (OpSize - Pos) & (OpSize - 1) [A]
3615 // for all Neg and Pos. Since Neg & (OpSize - 1) == Neg' & (OpSize - 1)
3616 // we can just replace Neg with Neg' for the rest of the function.
3618 // In other cases we check for the even stronger condition:
3620 // Neg == OpSize - Pos [B]
3622 // for all Neg and Pos. Note that the (or ...) then invokes undefined
3623 // behavior if Pos == 0 (and consequently Neg == OpSize).
3625 // We could actually use [A] whenever OpSize is a power of 2, but the
3626 // only extra cases that it would match are those uninteresting ones
3627 // where Neg and Pos are never in range at the same time. E.g. for
3628 // OpSize == 32, using [A] would allow a Neg of the form (sub 64, Pos)
3629 // as well as (sub 32, Pos), but:
3631 // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos))
3633 // always invokes undefined behavior for 32-bit X.
3635 // Below, Mask == OpSize - 1 when using [A] and is all-ones otherwise.
3636 unsigned MaskLoBits = 0;
3637 if (Neg.getOpcode() == ISD::AND &&
3638 isPowerOf2_64(OpSize) &&
3639 Neg.getOperand(1).getOpcode() == ISD::Constant &&
3640 cast<ConstantSDNode>(Neg.getOperand(1))->getAPIntValue() == OpSize - 1) {
3641 Neg = Neg.getOperand(0);
3642 MaskLoBits = Log2_64(OpSize);
3645 // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1.
3646 if (Neg.getOpcode() != ISD::SUB)
3648 ConstantSDNode *NegC = dyn_cast<ConstantSDNode>(Neg.getOperand(0));
3651 SDValue NegOp1 = Neg.getOperand(1);
3653 // On the RHS of [A], if Pos is Pos' & (OpSize - 1), just replace Pos with
3654 // Pos'. The truncation is redundant for the purpose of the equality.
3656 Pos.getOpcode() == ISD::AND &&
3657 Pos.getOperand(1).getOpcode() == ISD::Constant &&
3658 cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() == OpSize - 1)
3659 Pos = Pos.getOperand(0);
3661 // The condition we need is now:
3663 // (NegC - NegOp1) & Mask == (OpSize - Pos) & Mask
3665 // If NegOp1 == Pos then we need:
3667 // OpSize & Mask == NegC & Mask
3669 // (because "x & Mask" is a truncation and distributes through subtraction).
3672 Width = NegC->getAPIntValue();
3673 // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC.
3674 // Then the condition we want to prove becomes:
3676 // (NegC - NegOp1) & Mask == (OpSize - (NegOp1 + PosC)) & Mask
3678 // which, again because "x & Mask" is a truncation, becomes:
3680 // NegC & Mask == (OpSize - PosC) & Mask
3681 // OpSize & Mask == (NegC + PosC) & Mask
3682 else if (Pos.getOpcode() == ISD::ADD &&
3683 Pos.getOperand(0) == NegOp1 &&
3684 Pos.getOperand(1).getOpcode() == ISD::Constant)
3685 Width = (cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() +
3686 NegC->getAPIntValue());
3690 // Now we just need to check that OpSize & Mask == Width & Mask.
3692 // Opsize & Mask is 0 since Mask is Opsize - 1.
3693 return Width.getLoBits(MaskLoBits) == 0;
3694 return Width == OpSize;
3697 // A subroutine of MatchRotate used once we have found an OR of two opposite
3698 // shifts of Shifted. If Neg == <operand size> - Pos then the OR reduces
3699 // to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the
3700 // former being preferred if supported. InnerPos and InnerNeg are Pos and
3701 // Neg with outer conversions stripped away.
3702 SDNode *DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos,
3703 SDValue Neg, SDValue InnerPos,
3704 SDValue InnerNeg, unsigned PosOpcode,
3705 unsigned NegOpcode, SDLoc DL) {
3706 // fold (or (shl x, (*ext y)),
3707 // (srl x, (*ext (sub 32, y)))) ->
3708 // (rotl x, y) or (rotr x, (sub 32, y))
3710 // fold (or (shl x, (*ext (sub 32, y))),
3711 // (srl x, (*ext y))) ->
3712 // (rotr x, y) or (rotl x, (sub 32, y))
3713 EVT VT = Shifted.getValueType();
3714 if (matchRotateSub(InnerPos, InnerNeg, VT.getSizeInBits())) {
3715 bool HasPos = TLI.isOperationLegalOrCustom(PosOpcode, VT);
3716 return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted,
3717 HasPos ? Pos : Neg).getNode();
3723 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3724 // idioms for rotate, and if the target supports rotation instructions, generate
3726 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL) {
3727 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3728 EVT VT = LHS.getValueType();
3729 if (!TLI.isTypeLegal(VT)) return nullptr;
3731 // The target must have at least one rotate flavor.
3732 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3733 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3734 if (!HasROTL && !HasROTR) return nullptr;
3736 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3737 SDValue LHSShift; // The shift.
3738 SDValue LHSMask; // AND value if any.
3739 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3740 return nullptr; // Not part of a rotate.
3742 SDValue RHSShift; // The shift.
3743 SDValue RHSMask; // AND value if any.
3744 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3745 return nullptr; // Not part of a rotate.
3747 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3748 return nullptr; // Not shifting the same value.
3750 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3751 return nullptr; // Shifts must disagree.
3753 // Canonicalize shl to left side in a shl/srl pair.
3754 if (RHSShift.getOpcode() == ISD::SHL) {
3755 std::swap(LHS, RHS);
3756 std::swap(LHSShift, RHSShift);
3757 std::swap(LHSMask , RHSMask );
3760 unsigned OpSizeInBits = VT.getSizeInBits();
3761 SDValue LHSShiftArg = LHSShift.getOperand(0);
3762 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3763 SDValue RHSShiftArg = RHSShift.getOperand(0);
3764 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3766 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3767 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3768 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3769 RHSShiftAmt.getOpcode() == ISD::Constant) {
3770 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3771 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3772 if ((LShVal + RShVal) != OpSizeInBits)
3775 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3776 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
3778 // If there is an AND of either shifted operand, apply it to the result.
3779 if (LHSMask.getNode() || RHSMask.getNode()) {
3780 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3782 if (LHSMask.getNode()) {
3783 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3784 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3786 if (RHSMask.getNode()) {
3787 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3788 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3791 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3794 return Rot.getNode();
3797 // If there is a mask here, and we have a variable shift, we can't be sure
3798 // that we're masking out the right stuff.
3799 if (LHSMask.getNode() || RHSMask.getNode())
3802 // If the shift amount is sign/zext/any-extended just peel it off.
3803 SDValue LExtOp0 = LHSShiftAmt;
3804 SDValue RExtOp0 = RHSShiftAmt;
3805 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3806 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3807 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3808 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3809 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3810 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3811 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3812 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3813 LExtOp0 = LHSShiftAmt.getOperand(0);
3814 RExtOp0 = RHSShiftAmt.getOperand(0);
3817 SDNode *TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt,
3818 LExtOp0, RExtOp0, ISD::ROTL, ISD::ROTR, DL);
3822 SDNode *TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt,
3823 RExtOp0, LExtOp0, ISD::ROTR, ISD::ROTL, DL);
3830 SDValue DAGCombiner::visitXOR(SDNode *N) {
3831 SDValue N0 = N->getOperand(0);
3832 SDValue N1 = N->getOperand(1);
3833 SDValue LHS, RHS, CC;
3834 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3835 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3836 EVT VT = N0.getValueType();
3839 if (VT.isVector()) {
3840 SDValue FoldedVOp = SimplifyVBinOp(N);
3841 if (FoldedVOp.getNode()) return FoldedVOp;
3843 // fold (xor x, 0) -> x, vector edition
3844 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3846 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3850 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3851 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3852 return DAG.getConstant(0, VT);
3853 // fold (xor x, undef) -> undef
3854 if (N0.getOpcode() == ISD::UNDEF)
3856 if (N1.getOpcode() == ISD::UNDEF)
3858 // fold (xor c1, c2) -> c1^c2
3860 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3861 // canonicalize constant to RHS
3863 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
3864 // fold (xor x, 0) -> x
3865 if (N1C && N1C->isNullValue())
3868 SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1);
3872 // fold !(x cc y) -> (x !cc y)
3873 if (TLI.isConstTrueVal(N1.getNode()) && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3874 bool isInt = LHS.getValueType().isInteger();
3875 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3878 if (!LegalOperations ||
3879 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) {
3880 switch (N0.getOpcode()) {
3882 llvm_unreachable("Unhandled SetCC Equivalent!");
3884 return DAG.getSetCC(SDLoc(N), VT, LHS, RHS, NotCC);
3885 case ISD::SELECT_CC:
3886 return DAG.getSelectCC(SDLoc(N), LHS, RHS, N0.getOperand(2),
3887 N0.getOperand(3), NotCC);
3892 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3893 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3894 N0.getNode()->hasOneUse() &&
3895 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3896 SDValue V = N0.getOperand(0);
3897 V = DAG.getNode(ISD::XOR, SDLoc(N0), V.getValueType(), V,
3898 DAG.getConstant(1, V.getValueType()));
3899 AddToWorklist(V.getNode());
3900 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V);
3903 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3904 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3905 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3906 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3907 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3908 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3909 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3910 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3911 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
3912 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3915 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3916 if (N1C && N1C->isAllOnesValue() &&
3917 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3918 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3919 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3920 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3921 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3922 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3923 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
3924 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3927 // fold (xor (and x, y), y) -> (and (not x), y)
3928 if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3929 N0->getOperand(1) == N1) {
3930 SDValue X = N0->getOperand(0);
3931 SDValue NotX = DAG.getNOT(SDLoc(X), X, VT);
3932 AddToWorklist(NotX.getNode());
3933 return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1);
3935 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3936 if (N1C && N0.getOpcode() == ISD::XOR) {
3937 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3938 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3940 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(1),
3941 DAG.getConstant(N1C->getAPIntValue() ^
3942 N00C->getAPIntValue(), VT));
3944 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(0),
3945 DAG.getConstant(N1C->getAPIntValue() ^
3946 N01C->getAPIntValue(), VT));
3948 // fold (xor x, x) -> 0
3950 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
3952 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
3953 if (N0.getOpcode() == N1.getOpcode()) {
3954 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3955 if (Tmp.getNode()) return Tmp;
3958 // Simplify the expression using non-local knowledge.
3959 if (!VT.isVector() &&
3960 SimplifyDemandedBits(SDValue(N, 0)))
3961 return SDValue(N, 0);
3966 /// Handle transforms common to the three shifts, when the shift amount is a
3968 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, ConstantSDNode *Amt) {
3969 // We can't and shouldn't fold opaque constants.
3970 if (Amt->isOpaque())
3973 SDNode *LHS = N->getOperand(0).getNode();
3974 if (!LHS->hasOneUse()) return SDValue();
3976 // We want to pull some binops through shifts, so that we have (and (shift))
3977 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
3978 // thing happens with address calculations, so it's important to canonicalize
3980 bool HighBitSet = false; // Can we transform this if the high bit is set?
3982 switch (LHS->getOpcode()) {
3983 default: return SDValue();
3986 HighBitSet = false; // We can only transform sra if the high bit is clear.
3989 HighBitSet = true; // We can only transform sra if the high bit is set.
3992 if (N->getOpcode() != ISD::SHL)
3993 return SDValue(); // only shl(add) not sr[al](add).
3994 HighBitSet = false; // We can only transform sra if the high bit is clear.
3998 // We require the RHS of the binop to be a constant and not opaque as well.
3999 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
4000 if (!BinOpCst || BinOpCst->isOpaque()) return SDValue();
4002 // FIXME: disable this unless the input to the binop is a shift by a constant.
4003 // If it is not a shift, it pessimizes some common cases like:
4005 // void foo(int *X, int i) { X[i & 1235] = 1; }
4006 // int bar(int *X, int i) { return X[i & 255]; }
4007 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
4008 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
4009 BinOpLHSVal->getOpcode() != ISD::SRA &&
4010 BinOpLHSVal->getOpcode() != ISD::SRL) ||
4011 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
4014 EVT VT = N->getValueType(0);
4016 // If this is a signed shift right, and the high bit is modified by the
4017 // logical operation, do not perform the transformation. The highBitSet
4018 // boolean indicates the value of the high bit of the constant which would
4019 // cause it to be modified for this operation.
4020 if (N->getOpcode() == ISD::SRA) {
4021 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
4022 if (BinOpRHSSignSet != HighBitSet)
4026 if (!TLI.isDesirableToCommuteWithShift(LHS))
4029 // Fold the constants, shifting the binop RHS by the shift amount.
4030 SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)),
4032 LHS->getOperand(1), N->getOperand(1));
4033 assert(isa<ConstantSDNode>(NewRHS) && "Folding was not successful!");
4035 // Create the new shift.
4036 SDValue NewShift = DAG.getNode(N->getOpcode(),
4037 SDLoc(LHS->getOperand(0)),
4038 VT, LHS->getOperand(0), N->getOperand(1));
4040 // Create the new binop.
4041 return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS);
4044 SDValue DAGCombiner::distributeTruncateThroughAnd(SDNode *N) {
4045 assert(N->getOpcode() == ISD::TRUNCATE);
4046 assert(N->getOperand(0).getOpcode() == ISD::AND);
4048 // (truncate:TruncVT (and N00, N01C)) -> (and (truncate:TruncVT N00), TruncC)
4049 if (N->hasOneUse() && N->getOperand(0).hasOneUse()) {
4050 SDValue N01 = N->getOperand(0).getOperand(1);
4052 if (ConstantSDNode *N01C = isConstOrConstSplat(N01)) {
4053 EVT TruncVT = N->getValueType(0);
4054 SDValue N00 = N->getOperand(0).getOperand(0);
4055 APInt TruncC = N01C->getAPIntValue();
4056 TruncC = TruncC.trunc(TruncVT.getScalarSizeInBits());
4058 return DAG.getNode(ISD::AND, SDLoc(N), TruncVT,
4059 DAG.getNode(ISD::TRUNCATE, SDLoc(N), TruncVT, N00),
4060 DAG.getConstant(TruncC, TruncVT));
4067 SDValue DAGCombiner::visitRotate(SDNode *N) {
4068 // fold (rot* x, (trunc (and y, c))) -> (rot* x, (and (trunc y), (trunc c))).
4069 if (N->getOperand(1).getOpcode() == ISD::TRUNCATE &&
4070 N->getOperand(1).getOperand(0).getOpcode() == ISD::AND) {
4071 SDValue NewOp1 = distributeTruncateThroughAnd(N->getOperand(1).getNode());
4072 if (NewOp1.getNode())
4073 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
4074 N->getOperand(0), NewOp1);
4079 SDValue DAGCombiner::visitSHL(SDNode *N) {
4080 SDValue N0 = N->getOperand(0);
4081 SDValue N1 = N->getOperand(1);
4082 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4083 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4084 EVT VT = N0.getValueType();
4085 unsigned OpSizeInBits = VT.getScalarSizeInBits();
4088 if (VT.isVector()) {
4089 SDValue FoldedVOp = SimplifyVBinOp(N);
4090 if (FoldedVOp.getNode()) return FoldedVOp;
4092 BuildVectorSDNode *N1CV = dyn_cast<BuildVectorSDNode>(N1);
4093 // If setcc produces all-one true value then:
4094 // (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<<N1CV)
4095 if (N1CV && N1CV->isConstant()) {
4096 if (N0.getOpcode() == ISD::AND) {
4097 SDValue N00 = N0->getOperand(0);
4098 SDValue N01 = N0->getOperand(1);
4099 BuildVectorSDNode *N01CV = dyn_cast<BuildVectorSDNode>(N01);
4101 if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC &&
4102 TLI.getBooleanContents(N00.getOperand(0).getValueType()) ==
4103 TargetLowering::ZeroOrNegativeOneBooleanContent) {
4104 if (SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, VT, N01CV, N1CV))
4105 return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C);
4108 N1C = isConstOrConstSplat(N1);
4113 // fold (shl c1, c2) -> c1<<c2
4115 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
4116 // fold (shl 0, x) -> 0
4117 if (N0C && N0C->isNullValue())
4119 // fold (shl x, c >= size(x)) -> undef
4120 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4121 return DAG.getUNDEF(VT);
4122 // fold (shl x, 0) -> x
4123 if (N1C && N1C->isNullValue())
4125 // fold (shl undef, x) -> 0
4126 if (N0.getOpcode() == ISD::UNDEF)
4127 return DAG.getConstant(0, VT);
4128 // if (shl x, c) is known to be zero, return 0
4129 if (DAG.MaskedValueIsZero(SDValue(N, 0),
4130 APInt::getAllOnesValue(OpSizeInBits)))
4131 return DAG.getConstant(0, VT);
4132 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
4133 if (N1.getOpcode() == ISD::TRUNCATE &&
4134 N1.getOperand(0).getOpcode() == ISD::AND) {
4135 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4136 if (NewOp1.getNode())
4137 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, NewOp1);
4140 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4141 return SDValue(N, 0);
4143 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
4144 if (N1C && N0.getOpcode() == ISD::SHL) {
4145 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4146 uint64_t c1 = N0C1->getZExtValue();
4147 uint64_t c2 = N1C->getZExtValue();
4148 if (c1 + c2 >= OpSizeInBits)
4149 return DAG.getConstant(0, VT);
4150 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4151 DAG.getConstant(c1 + c2, N1.getValueType()));
4155 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
4156 // For this to be valid, the second form must not preserve any of the bits
4157 // that are shifted out by the inner shift in the first form. This means
4158 // the outer shift size must be >= the number of bits added by the ext.
4159 // As a corollary, we don't care what kind of ext it is.
4160 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
4161 N0.getOpcode() == ISD::ANY_EXTEND ||
4162 N0.getOpcode() == ISD::SIGN_EXTEND) &&
4163 N0.getOperand(0).getOpcode() == ISD::SHL) {
4164 SDValue N0Op0 = N0.getOperand(0);
4165 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4166 uint64_t c1 = N0Op0C1->getZExtValue();
4167 uint64_t c2 = N1C->getZExtValue();
4168 EVT InnerShiftVT = N0Op0.getValueType();
4169 uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits();
4170 if (c2 >= OpSizeInBits - InnerShiftSize) {
4171 if (c1 + c2 >= OpSizeInBits)
4172 return DAG.getConstant(0, VT);
4173 return DAG.getNode(ISD::SHL, SDLoc(N0), VT,
4174 DAG.getNode(N0.getOpcode(), SDLoc(N0), VT,
4175 N0Op0->getOperand(0)),
4176 DAG.getConstant(c1 + c2, N1.getValueType()));
4181 // fold (shl (zext (srl x, C)), C) -> (zext (shl (srl x, C), C))
4182 // Only fold this if the inner zext has no other uses to avoid increasing
4183 // the total number of instructions.
4184 if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() &&
4185 N0.getOperand(0).getOpcode() == ISD::SRL) {
4186 SDValue N0Op0 = N0.getOperand(0);
4187 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4188 uint64_t c1 = N0Op0C1->getZExtValue();
4189 if (c1 < VT.getScalarSizeInBits()) {
4190 uint64_t c2 = N1C->getZExtValue();
4192 SDValue NewOp0 = N0.getOperand(0);
4193 EVT CountVT = NewOp0.getOperand(1).getValueType();
4194 SDValue NewSHL = DAG.getNode(ISD::SHL, SDLoc(N), NewOp0.getValueType(),
4195 NewOp0, DAG.getConstant(c2, CountVT));
4196 AddToWorklist(NewSHL.getNode());
4197 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
4203 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
4204 // (and (srl x, (sub c1, c2), MASK)
4205 // Only fold this if the inner shift has no other uses -- if it does, folding
4206 // this will increase the total number of instructions.
4207 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
4208 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4209 uint64_t c1 = N0C1->getZExtValue();
4210 if (c1 < OpSizeInBits) {
4211 uint64_t c2 = N1C->getZExtValue();
4212 APInt Mask = APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - c1);
4215 Mask = Mask.shl(c2 - c1);
4216 Shift = DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4217 DAG.getConstant(c2 - c1, N1.getValueType()));
4219 Mask = Mask.lshr(c1 - c2);
4220 Shift = DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4221 DAG.getConstant(c1 - c2, N1.getValueType()));
4223 return DAG.getNode(ISD::AND, SDLoc(N0), VT, Shift,
4224 DAG.getConstant(Mask, VT));
4228 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
4229 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
4230 unsigned BitSize = VT.getScalarSizeInBits();
4231 SDValue HiBitsMask =
4232 DAG.getConstant(APInt::getHighBitsSet(BitSize,
4233 BitSize - N1C->getZExtValue()), VT);
4234 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4238 // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
4239 // Variant of version done on multiply, except mul by a power of 2 is turned
4242 if (N1C && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
4243 (isa<ConstantSDNode>(N0.getOperand(1)) ||
4244 isConstantSplatVector(N0.getOperand(1).getNode(), Val))) {
4245 SDValue Shl0 = DAG.getNode(ISD::SHL, SDLoc(N0), VT, N0.getOperand(0), N1);
4246 SDValue Shl1 = DAG.getNode(ISD::SHL, SDLoc(N1), VT, N0.getOperand(1), N1);
4247 return DAG.getNode(ISD::ADD, SDLoc(N), VT, Shl0, Shl1);
4251 SDValue NewSHL = visitShiftByConstant(N, N1C);
4252 if (NewSHL.getNode())
4259 SDValue DAGCombiner::visitSRA(SDNode *N) {
4260 SDValue N0 = N->getOperand(0);
4261 SDValue N1 = N->getOperand(1);
4262 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4263 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4264 EVT VT = N0.getValueType();
4265 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4268 if (VT.isVector()) {
4269 SDValue FoldedVOp = SimplifyVBinOp(N);
4270 if (FoldedVOp.getNode()) return FoldedVOp;
4272 N1C = isConstOrConstSplat(N1);
4275 // fold (sra c1, c2) -> (sra c1, c2)
4277 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
4278 // fold (sra 0, x) -> 0
4279 if (N0C && N0C->isNullValue())
4281 // fold (sra -1, x) -> -1
4282 if (N0C && N0C->isAllOnesValue())
4284 // fold (sra x, (setge c, size(x))) -> undef
4285 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4286 return DAG.getUNDEF(VT);
4287 // fold (sra x, 0) -> x
4288 if (N1C && N1C->isNullValue())
4290 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
4292 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
4293 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
4294 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
4296 ExtVT = EVT::getVectorVT(*DAG.getContext(),
4297 ExtVT, VT.getVectorNumElements());
4298 if ((!LegalOperations ||
4299 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
4300 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
4301 N0.getOperand(0), DAG.getValueType(ExtVT));
4304 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
4305 if (N1C && N0.getOpcode() == ISD::SRA) {
4306 if (ConstantSDNode *C1 = isConstOrConstSplat(N0.getOperand(1))) {
4307 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
4308 if (Sum >= OpSizeInBits)
4309 Sum = OpSizeInBits - 1;
4310 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0.getOperand(0),
4311 DAG.getConstant(Sum, N1.getValueType()));
4315 // fold (sra (shl X, m), (sub result_size, n))
4316 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
4317 // result_size - n != m.
4318 // If truncate is free for the target sext(shl) is likely to result in better
4320 if (N0.getOpcode() == ISD::SHL && N1C) {
4321 // Get the two constanst of the shifts, CN0 = m, CN = n.
4322 const ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1));
4324 LLVMContext &Ctx = *DAG.getContext();
4325 // Determine what the truncate's result bitsize and type would be.
4326 EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - N1C->getZExtValue());
4329 TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorNumElements());
4331 // Determine the residual right-shift amount.
4332 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
4334 // If the shift is not a no-op (in which case this should be just a sign
4335 // extend already), the truncated to type is legal, sign_extend is legal
4336 // on that type, and the truncate to that type is both legal and free,
4337 // perform the transform.
4338 if ((ShiftAmt > 0) &&
4339 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
4340 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
4341 TLI.isTruncateFree(VT, TruncVT)) {
4343 SDValue Amt = DAG.getConstant(ShiftAmt,
4344 getShiftAmountTy(N0.getOperand(0).getValueType()));
4345 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0), VT,
4346 N0.getOperand(0), Amt);
4347 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), TruncVT,
4349 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N),
4350 N->getValueType(0), Trunc);
4355 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
4356 if (N1.getOpcode() == ISD::TRUNCATE &&
4357 N1.getOperand(0).getOpcode() == ISD::AND) {
4358 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4359 if (NewOp1.getNode())
4360 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, NewOp1);
4363 // fold (sra (trunc (srl x, c1)), c2) -> (trunc (sra x, c1 + c2))
4364 // if c1 is equal to the number of bits the trunc removes
4365 if (N0.getOpcode() == ISD::TRUNCATE &&
4366 (N0.getOperand(0).getOpcode() == ISD::SRL ||
4367 N0.getOperand(0).getOpcode() == ISD::SRA) &&
4368 N0.getOperand(0).hasOneUse() &&
4369 N0.getOperand(0).getOperand(1).hasOneUse() &&
4371 SDValue N0Op0 = N0.getOperand(0);
4372 if (ConstantSDNode *LargeShift = isConstOrConstSplat(N0Op0.getOperand(1))) {
4373 unsigned LargeShiftVal = LargeShift->getZExtValue();
4374 EVT LargeVT = N0Op0.getValueType();
4376 if (LargeVT.getScalarSizeInBits() - OpSizeInBits == LargeShiftVal) {
4378 DAG.getConstant(LargeShiftVal + N1C->getZExtValue(),
4379 getShiftAmountTy(N0Op0.getOperand(0).getValueType()));
4380 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), LargeVT,
4381 N0Op0.getOperand(0), Amt);
4382 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, SRA);
4387 // Simplify, based on bits shifted out of the LHS.
4388 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4389 return SDValue(N, 0);
4392 // If the sign bit is known to be zero, switch this to a SRL.
4393 if (DAG.SignBitIsZero(N0))
4394 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1);
4397 SDValue NewSRA = visitShiftByConstant(N, N1C);
4398 if (NewSRA.getNode())
4405 SDValue DAGCombiner::visitSRL(SDNode *N) {
4406 SDValue N0 = N->getOperand(0);
4407 SDValue N1 = N->getOperand(1);
4408 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4409 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4410 EVT VT = N0.getValueType();
4411 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4414 if (VT.isVector()) {
4415 SDValue FoldedVOp = SimplifyVBinOp(N);
4416 if (FoldedVOp.getNode()) return FoldedVOp;
4418 N1C = isConstOrConstSplat(N1);
4421 // fold (srl c1, c2) -> c1 >>u c2
4423 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
4424 // fold (srl 0, x) -> 0
4425 if (N0C && N0C->isNullValue())
4427 // fold (srl x, c >= size(x)) -> undef
4428 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4429 return DAG.getUNDEF(VT);
4430 // fold (srl x, 0) -> x
4431 if (N1C && N1C->isNullValue())
4433 // if (srl x, c) is known to be zero, return 0
4434 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
4435 APInt::getAllOnesValue(OpSizeInBits)))
4436 return DAG.getConstant(0, VT);
4438 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
4439 if (N1C && N0.getOpcode() == ISD::SRL) {
4440 if (ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1))) {
4441 uint64_t c1 = N01C->getZExtValue();
4442 uint64_t c2 = N1C->getZExtValue();
4443 if (c1 + c2 >= OpSizeInBits)
4444 return DAG.getConstant(0, VT);
4445 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4446 DAG.getConstant(c1 + c2, N1.getValueType()));
4450 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
4451 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
4452 N0.getOperand(0).getOpcode() == ISD::SRL &&
4453 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
4455 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
4456 uint64_t c2 = N1C->getZExtValue();
4457 EVT InnerShiftVT = N0.getOperand(0).getValueType();
4458 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
4459 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
4460 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
4461 if (c1 + OpSizeInBits == InnerShiftSize) {
4462 if (c1 + c2 >= InnerShiftSize)
4463 return DAG.getConstant(0, VT);
4464 return DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT,
4465 DAG.getNode(ISD::SRL, SDLoc(N0), InnerShiftVT,
4466 N0.getOperand(0)->getOperand(0),
4467 DAG.getConstant(c1 + c2, ShiftCountVT)));
4471 // fold (srl (shl x, c), c) -> (and x, cst2)
4472 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1) {
4473 unsigned BitSize = N0.getScalarValueSizeInBits();
4474 if (BitSize <= 64) {
4475 uint64_t ShAmt = N1C->getZExtValue() + 64 - BitSize;
4476 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4477 DAG.getConstant(~0ULL >> ShAmt, VT));
4481 // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask)
4482 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
4483 // Shifting in all undef bits?
4484 EVT SmallVT = N0.getOperand(0).getValueType();
4485 unsigned BitSize = SmallVT.getScalarSizeInBits();
4486 if (N1C->getZExtValue() >= BitSize)
4487 return DAG.getUNDEF(VT);
4489 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
4490 uint64_t ShiftAmt = N1C->getZExtValue();
4491 SDValue SmallShift = DAG.getNode(ISD::SRL, SDLoc(N0), SmallVT,
4493 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
4494 AddToWorklist(SmallShift.getNode());
4495 APInt Mask = APInt::getAllOnesValue(OpSizeInBits).lshr(ShiftAmt);
4496 return DAG.getNode(ISD::AND, SDLoc(N), VT,
4497 DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, SmallShift),
4498 DAG.getConstant(Mask, VT));
4502 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
4503 // bit, which is unmodified by sra.
4504 if (N1C && N1C->getZExtValue() + 1 == OpSizeInBits) {
4505 if (N0.getOpcode() == ISD::SRA)
4506 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1);
4509 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
4510 if (N1C && N0.getOpcode() == ISD::CTLZ &&
4511 N1C->getAPIntValue() == Log2_32(OpSizeInBits)) {
4512 APInt KnownZero, KnownOne;
4513 DAG.computeKnownBits(N0.getOperand(0), KnownZero, KnownOne);
4515 // If any of the input bits are KnownOne, then the input couldn't be all
4516 // zeros, thus the result of the srl will always be zero.
4517 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
4519 // If all of the bits input the to ctlz node are known to be zero, then
4520 // the result of the ctlz is "32" and the result of the shift is one.
4521 APInt UnknownBits = ~KnownZero;
4522 if (UnknownBits == 0) return DAG.getConstant(1, VT);
4524 // Otherwise, check to see if there is exactly one bit input to the ctlz.
4525 if ((UnknownBits & (UnknownBits - 1)) == 0) {
4526 // Okay, we know that only that the single bit specified by UnknownBits
4527 // could be set on input to the CTLZ node. If this bit is set, the SRL
4528 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
4529 // to an SRL/XOR pair, which is likely to simplify more.
4530 unsigned ShAmt = UnknownBits.countTrailingZeros();
4531 SDValue Op = N0.getOperand(0);
4534 Op = DAG.getNode(ISD::SRL, SDLoc(N0), VT, Op,
4535 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
4536 AddToWorklist(Op.getNode());
4539 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
4540 Op, DAG.getConstant(1, VT));
4544 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
4545 if (N1.getOpcode() == ISD::TRUNCATE &&
4546 N1.getOperand(0).getOpcode() == ISD::AND) {
4547 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4548 if (NewOp1.getNode())
4549 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, NewOp1);
4552 // fold operands of srl based on knowledge that the low bits are not
4554 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4555 return SDValue(N, 0);
4558 SDValue NewSRL = visitShiftByConstant(N, N1C);
4559 if (NewSRL.getNode())
4563 // Attempt to convert a srl of a load into a narrower zero-extending load.
4564 SDValue NarrowLoad = ReduceLoadWidth(N);
4565 if (NarrowLoad.getNode())
4568 // Here is a common situation. We want to optimize:
4571 // %b = and i32 %a, 2
4572 // %c = srl i32 %b, 1
4573 // brcond i32 %c ...
4579 // %c = setcc eq %b, 0
4582 // However when after the source operand of SRL is optimized into AND, the SRL
4583 // itself may not be optimized further. Look for it and add the BRCOND into
4585 if (N->hasOneUse()) {
4586 SDNode *Use = *N->use_begin();
4587 if (Use->getOpcode() == ISD::BRCOND)
4589 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
4590 // Also look pass the truncate.
4591 Use = *Use->use_begin();
4592 if (Use->getOpcode() == ISD::BRCOND)
4600 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
4601 SDValue N0 = N->getOperand(0);
4602 EVT VT = N->getValueType(0);
4604 // fold (ctlz c1) -> c2
4605 if (isa<ConstantSDNode>(N0))
4606 return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0);
4610 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4611 SDValue N0 = N->getOperand(0);
4612 EVT VT = N->getValueType(0);
4614 // fold (ctlz_zero_undef c1) -> c2
4615 if (isa<ConstantSDNode>(N0))
4616 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4620 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4621 SDValue N0 = N->getOperand(0);
4622 EVT VT = N->getValueType(0);
4624 // fold (cttz c1) -> c2
4625 if (isa<ConstantSDNode>(N0))
4626 return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0);
4630 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4631 SDValue N0 = N->getOperand(0);
4632 EVT VT = N->getValueType(0);
4634 // fold (cttz_zero_undef c1) -> c2
4635 if (isa<ConstantSDNode>(N0))
4636 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4640 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4641 SDValue N0 = N->getOperand(0);
4642 EVT VT = N->getValueType(0);
4644 // fold (ctpop c1) -> c2
4645 if (isa<ConstantSDNode>(N0))
4646 return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0);
4651 /// \brief Generate Min/Max node
4652 static SDValue combineMinNumMaxNum(SDLoc DL, EVT VT, SDValue LHS, SDValue RHS,
4653 SDValue True, SDValue False,
4654 ISD::CondCode CC, const TargetLowering &TLI,
4655 SelectionDAG &DAG) {
4656 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
4666 unsigned Opcode = (LHS == True) ? ISD::FMINNUM : ISD::FMAXNUM;
4667 if (TLI.isOperationLegal(Opcode, VT))
4668 return DAG.getNode(Opcode, DL, VT, LHS, RHS);
4677 unsigned Opcode = (LHS == True) ? ISD::FMAXNUM : ISD::FMINNUM;
4678 if (TLI.isOperationLegal(Opcode, VT))
4679 return DAG.getNode(Opcode, DL, VT, LHS, RHS);
4687 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4688 SDValue N0 = N->getOperand(0);
4689 SDValue N1 = N->getOperand(1);
4690 SDValue N2 = N->getOperand(2);
4691 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4692 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4693 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
4694 EVT VT = N->getValueType(0);
4695 EVT VT0 = N0.getValueType();
4697 // fold (select C, X, X) -> X
4700 // fold (select true, X, Y) -> X
4701 if (N0C && !N0C->isNullValue())
4703 // fold (select false, X, Y) -> Y
4704 if (N0C && N0C->isNullValue())
4706 // fold (select C, 1, X) -> (or C, X)
4707 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4708 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4709 // fold (select C, 0, 1) -> (xor C, 1)
4710 // We can't do this reliably if integer based booleans have different contents
4711 // to floating point based booleans. This is because we can't tell whether we
4712 // have an integer-based boolean or a floating-point-based boolean unless we
4713 // can find the SETCC that produced it and inspect its operands. This is
4714 // fairly easy if C is the SETCC node, but it can potentially be
4715 // undiscoverable (or not reasonably discoverable). For example, it could be
4716 // in another basic block or it could require searching a complicated
4718 if (VT.isInteger() &&
4719 (VT0 == MVT::i1 || (VT0.isInteger() &&
4720 TLI.getBooleanContents(false, false) ==
4721 TLI.getBooleanContents(false, true) &&
4722 TLI.getBooleanContents(false, false) ==
4723 TargetLowering::ZeroOrOneBooleanContent)) &&
4724 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4727 return DAG.getNode(ISD::XOR, SDLoc(N), VT0,
4728 N0, DAG.getConstant(1, VT0));
4729 XORNode = DAG.getNode(ISD::XOR, SDLoc(N0), VT0,
4730 N0, DAG.getConstant(1, VT0));
4731 AddToWorklist(XORNode.getNode());
4733 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, XORNode);
4734 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, XORNode);
4736 // fold (select C, 0, X) -> (and (not C), X)
4737 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4738 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4739 AddToWorklist(NOTNode.getNode());
4740 return DAG.getNode(ISD::AND, SDLoc(N), VT, NOTNode, N2);
4742 // fold (select C, X, 1) -> (or (not C), X)
4743 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4744 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4745 AddToWorklist(NOTNode.getNode());
4746 return DAG.getNode(ISD::OR, SDLoc(N), VT, NOTNode, N1);
4748 // fold (select C, X, 0) -> (and C, X)
4749 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4750 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4751 // fold (select X, X, Y) -> (or X, Y)
4752 // fold (select X, 1, Y) -> (or X, Y)
4753 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4754 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4755 // fold (select X, Y, X) -> (and X, Y)
4756 // fold (select X, Y, 0) -> (and X, Y)
4757 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4758 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4760 // If we can fold this based on the true/false value, do so.
4761 if (SimplifySelectOps(N, N1, N2))
4762 return SDValue(N, 0); // Don't revisit N.
4764 // fold selects based on a setcc into other things, such as min/max/abs
4765 if (N0.getOpcode() == ISD::SETCC) {
4766 // select x, y (fcmp lt x, y) -> fminnum x, y
4767 // select x, y (fcmp gt x, y) -> fmaxnum x, y
4769 // This is OK if we don't care about what happens if either operand is a
4773 // FIXME: Instead of testing for UnsafeFPMath, this should be checking for
4774 // no signed zeros as well as no nans.
4775 const TargetOptions &Options = DAG.getTarget().Options;
4776 if (Options.UnsafeFPMath &&
4777 VT.isFloatingPoint() && N0.hasOneUse() &&
4778 DAG.isKnownNeverNaN(N1) && DAG.isKnownNeverNaN(N2)) {
4779 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
4782 combineMinNumMaxNum(SDLoc(N), VT, N0.getOperand(0), N0.getOperand(1),
4783 N1, N2, CC, TLI, DAG);
4788 if ((!LegalOperations &&
4789 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT)) ||
4790 TLI.isOperationLegal(ISD::SELECT_CC, VT))
4791 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT,
4792 N0.getOperand(0), N0.getOperand(1),
4793 N1, N2, N0.getOperand(2));
4794 return SimplifySelect(SDLoc(N), N0, N1, N2);
4801 std::pair<SDValue, SDValue> SplitVSETCC(const SDNode *N, SelectionDAG &DAG) {
4804 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
4806 // Split the inputs.
4807 SDValue Lo, Hi, LL, LH, RL, RH;
4808 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
4809 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
4811 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));
4812 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));
4814 return std::make_pair(Lo, Hi);
4817 // This function assumes all the vselect's arguments are CONCAT_VECTOR
4818 // nodes and that the condition is a BV of ConstantSDNodes (or undefs).
4819 static SDValue ConvertSelectToConcatVector(SDNode *N, SelectionDAG &DAG) {
4821 SDValue Cond = N->getOperand(0);
4822 SDValue LHS = N->getOperand(1);
4823 SDValue RHS = N->getOperand(2);
4824 EVT VT = N->getValueType(0);
4825 int NumElems = VT.getVectorNumElements();
4826 assert(LHS.getOpcode() == ISD::CONCAT_VECTORS &&
4827 RHS.getOpcode() == ISD::CONCAT_VECTORS &&
4828 Cond.getOpcode() == ISD::BUILD_VECTOR);
4830 // CONCAT_VECTOR can take an arbitrary number of arguments. We only care about
4831 // binary ones here.
4832 if (LHS->getNumOperands() != 2 || RHS->getNumOperands() != 2)
4835 // We're sure we have an even number of elements due to the
4836 // concat_vectors we have as arguments to vselect.
4837 // Skip BV elements until we find one that's not an UNDEF
4838 // After we find an UNDEF element, keep looping until we get to half the
4839 // length of the BV and see if all the non-undef nodes are the same.
4840 ConstantSDNode *BottomHalf = nullptr;
4841 for (int i = 0; i < NumElems / 2; ++i) {
4842 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4845 if (BottomHalf == nullptr)
4846 BottomHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4847 else if (Cond->getOperand(i).getNode() != BottomHalf)
4851 // Do the same for the second half of the BuildVector
4852 ConstantSDNode *TopHalf = nullptr;
4853 for (int i = NumElems / 2; i < NumElems; ++i) {
4854 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4857 if (TopHalf == nullptr)
4858 TopHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4859 else if (Cond->getOperand(i).getNode() != TopHalf)
4863 assert(TopHalf && BottomHalf &&
4864 "One half of the selector was all UNDEFs and the other was all the "
4865 "same value. This should have been addressed before this function.");
4867 ISD::CONCAT_VECTORS, dl, VT,
4868 BottomHalf->isNullValue() ? RHS->getOperand(0) : LHS->getOperand(0),
4869 TopHalf->isNullValue() ? RHS->getOperand(1) : LHS->getOperand(1));
4872 SDValue DAGCombiner::visitMSTORE(SDNode *N) {
4874 if (Level >= AfterLegalizeTypes)
4877 MaskedStoreSDNode *MST = dyn_cast<MaskedStoreSDNode>(N);
4878 SDValue Mask = MST->getMask();
4879 SDValue Data = MST->getValue();
4882 // If the MSTORE data type requires splitting and the mask is provided by a
4883 // SETCC, then split both nodes and its operands before legalization. This
4884 // prevents the type legalizer from unrolling SETCC into scalar comparisons
4885 // and enables future optimizations (e.g. min/max pattern matching on X86).
4886 if (Mask.getOpcode() == ISD::SETCC) {
4888 // Check if any splitting is required.
4889 if (TLI.getTypeAction(*DAG.getContext(), Data.getValueType()) !=
4890 TargetLowering::TypeSplitVector)
4893 SDValue MaskLo, MaskHi, Lo, Hi;
4894 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
4897 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MST->getValueType(0));
4899 SDValue Chain = MST->getChain();
4900 SDValue Ptr = MST->getBasePtr();
4902 EVT MemoryVT = MST->getMemoryVT();
4903 unsigned Alignment = MST->getOriginalAlignment();
4905 // if Alignment is equal to the vector size,
4906 // take the half of it for the second part
4907 unsigned SecondHalfAlignment =
4908 (Alignment == Data->getValueType(0).getSizeInBits()/8) ?
4909 Alignment/2 : Alignment;
4911 EVT LoMemVT, HiMemVT;
4912 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
4914 SDValue DataLo, DataHi;
4915 std::tie(DataLo, DataHi) = DAG.SplitVector(Data, DL);
4917 MachineMemOperand *MMO = DAG.getMachineFunction().
4918 getMachineMemOperand(MST->getPointerInfo(),
4919 MachineMemOperand::MOStore, LoMemVT.getStoreSize(),
4920 Alignment, MST->getAAInfo(), MST->getRanges());
4922 Lo = DAG.getMaskedStore(Chain, DL, DataLo, Ptr, MaskLo, LoMemVT, MMO,
4923 MST->isTruncatingStore());
4925 unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
4926 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
4927 DAG.getConstant(IncrementSize, Ptr.getValueType()));
4929 MMO = DAG.getMachineFunction().
4930 getMachineMemOperand(MST->getPointerInfo(),
4931 MachineMemOperand::MOStore, HiMemVT.getStoreSize(),
4932 SecondHalfAlignment, MST->getAAInfo(),
4935 Hi = DAG.getMaskedStore(Chain, DL, DataHi, Ptr, MaskHi, HiMemVT, MMO,
4936 MST->isTruncatingStore());
4938 AddToWorklist(Lo.getNode());
4939 AddToWorklist(Hi.getNode());
4941 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);
4946 SDValue DAGCombiner::visitMLOAD(SDNode *N) {
4948 if (Level >= AfterLegalizeTypes)
4951 MaskedLoadSDNode *MLD = dyn_cast<MaskedLoadSDNode>(N);
4952 SDValue Mask = MLD->getMask();
4955 // If the MLOAD result requires splitting and the mask is provided by a
4956 // SETCC, then split both nodes and its operands before legalization. This
4957 // prevents the type legalizer from unrolling SETCC into scalar comparisons
4958 // and enables future optimizations (e.g. min/max pattern matching on X86).
4960 if (Mask.getOpcode() == ISD::SETCC) {
4961 EVT VT = N->getValueType(0);
4963 // Check if any splitting is required.
4964 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
4965 TargetLowering::TypeSplitVector)
4968 SDValue MaskLo, MaskHi, Lo, Hi;
4969 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
4971 SDValue Src0 = MLD->getSrc0();
4972 SDValue Src0Lo, Src0Hi;
4973 std::tie(Src0Lo, Src0Hi) = DAG.SplitVector(Src0, DL);
4976 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MLD->getValueType(0));
4978 SDValue Chain = MLD->getChain();
4979 SDValue Ptr = MLD->getBasePtr();
4980 EVT MemoryVT = MLD->getMemoryVT();
4981 unsigned Alignment = MLD->getOriginalAlignment();
4983 // if Alignment is equal to the vector size,
4984 // take the half of it for the second part
4985 unsigned SecondHalfAlignment =
4986 (Alignment == MLD->getValueType(0).getSizeInBits()/8) ?
4987 Alignment/2 : Alignment;
4989 EVT LoMemVT, HiMemVT;
4990 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
4992 MachineMemOperand *MMO = DAG.getMachineFunction().
4993 getMachineMemOperand(MLD->getPointerInfo(),
4994 MachineMemOperand::MOLoad, LoMemVT.getStoreSize(),
4995 Alignment, MLD->getAAInfo(), MLD->getRanges());
4997 Lo = DAG.getMaskedLoad(LoVT, DL, Chain, Ptr, MaskLo, Src0Lo, LoMemVT, MMO,
5000 unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
5001 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5002 DAG.getConstant(IncrementSize, Ptr.getValueType()));
5004 MMO = DAG.getMachineFunction().
5005 getMachineMemOperand(MLD->getPointerInfo(),
5006 MachineMemOperand::MOLoad, HiMemVT.getStoreSize(),
5007 SecondHalfAlignment, MLD->getAAInfo(), MLD->getRanges());
5009 Hi = DAG.getMaskedLoad(HiVT, DL, Chain, Ptr, MaskHi, Src0Hi, HiMemVT, MMO,
5012 AddToWorklist(Lo.getNode());
5013 AddToWorklist(Hi.getNode());
5015 // Build a factor node to remember that this load is independent of the
5017 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo.getValue(1),
5020 // Legalized the chain result - switch anything that used the old chain to
5022 DAG.ReplaceAllUsesOfValueWith(SDValue(MLD, 1), Chain);
5024 SDValue LoadRes = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
5026 SDValue RetOps[] = { LoadRes, Chain };
5027 return DAG.getMergeValues(RetOps, DL);
5032 SDValue DAGCombiner::visitVSELECT(SDNode *N) {
5033 SDValue N0 = N->getOperand(0);
5034 SDValue N1 = N->getOperand(1);
5035 SDValue N2 = N->getOperand(2);
5038 // Canonicalize integer abs.
5039 // vselect (setg[te] X, 0), X, -X ->
5040 // vselect (setgt X, -1), X, -X ->
5041 // vselect (setl[te] X, 0), -X, X ->
5042 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
5043 if (N0.getOpcode() == ISD::SETCC) {
5044 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
5045 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5047 bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
5049 if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
5050 (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) &&
5051 N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1))
5052 isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode());
5053 else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) &&
5054 N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1))
5055 isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode());
5058 EVT VT = LHS.getValueType();
5059 SDValue Shift = DAG.getNode(
5060 ISD::SRA, DL, VT, LHS,
5061 DAG.getConstant(VT.getScalarType().getSizeInBits() - 1, VT));
5062 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift);
5063 AddToWorklist(Shift.getNode());
5064 AddToWorklist(Add.getNode());
5065 return DAG.getNode(ISD::XOR, DL, VT, Add, Shift);
5069 // If the VSELECT result requires splitting and the mask is provided by a
5070 // SETCC, then split both nodes and its operands before legalization. This
5071 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5072 // and enables future optimizations (e.g. min/max pattern matching on X86).
5073 if (N0.getOpcode() == ISD::SETCC) {
5074 EVT VT = N->getValueType(0);
5076 // Check if any splitting is required.
5077 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
5078 TargetLowering::TypeSplitVector)
5081 SDValue Lo, Hi, CCLo, CCHi, LL, LH, RL, RH;
5082 std::tie(CCLo, CCHi) = SplitVSETCC(N0.getNode(), DAG);
5083 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 1);
5084 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 2);
5086 Lo = DAG.getNode(N->getOpcode(), DL, LL.getValueType(), CCLo, LL, RL);
5087 Hi = DAG.getNode(N->getOpcode(), DL, LH.getValueType(), CCHi, LH, RH);
5089 // Add the new VSELECT nodes to the work list in case they need to be split
5091 AddToWorklist(Lo.getNode());
5092 AddToWorklist(Hi.getNode());
5094 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
5097 // Fold (vselect (build_vector all_ones), N1, N2) -> N1
5098 if (ISD::isBuildVectorAllOnes(N0.getNode()))
5100 // Fold (vselect (build_vector all_zeros), N1, N2) -> N2
5101 if (ISD::isBuildVectorAllZeros(N0.getNode()))
5104 // The ConvertSelectToConcatVector function is assuming both the above
5105 // checks for (vselect (build_vector all{ones,zeros) ...) have been made
5107 if (N1.getOpcode() == ISD::CONCAT_VECTORS &&
5108 N2.getOpcode() == ISD::CONCAT_VECTORS &&
5109 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
5110 SDValue CV = ConvertSelectToConcatVector(N, DAG);
5118 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
5119 SDValue N0 = N->getOperand(0);
5120 SDValue N1 = N->getOperand(1);
5121 SDValue N2 = N->getOperand(2);
5122 SDValue N3 = N->getOperand(3);
5123 SDValue N4 = N->getOperand(4);
5124 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
5126 // fold select_cc lhs, rhs, x, x, cc -> x
5130 // Determine if the condition we're dealing with is constant
5131 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
5132 N0, N1, CC, SDLoc(N), false);
5133 if (SCC.getNode()) {
5134 AddToWorklist(SCC.getNode());
5136 if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) {
5137 if (!SCCC->isNullValue())
5138 return N2; // cond always true -> true val
5140 return N3; // cond always false -> false val
5141 } else if (SCC->getOpcode() == ISD::UNDEF) {
5142 // When the condition is UNDEF, just return the first operand. This is
5143 // coherent the DAG creation, no setcc node is created in this case
5145 } else if (SCC.getOpcode() == ISD::SETCC) {
5146 // Fold to a simpler select_cc
5147 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(),
5148 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
5153 // If we can fold this based on the true/false value, do so.
5154 if (SimplifySelectOps(N, N2, N3))
5155 return SDValue(N, 0); // Don't revisit N.
5157 // fold select_cc into other things, such as min/max/abs
5158 return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC);
5161 SDValue DAGCombiner::visitSETCC(SDNode *N) {
5162 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
5163 cast<CondCodeSDNode>(N->getOperand(2))->get(),
5167 // tryToFoldExtendOfConstant - Try to fold a sext/zext/aext
5168 // dag node into a ConstantSDNode or a build_vector of constants.
5169 // This function is called by the DAGCombiner when visiting sext/zext/aext
5170 // dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND).
5171 // Vector extends are not folded if operations are legal; this is to
5172 // avoid introducing illegal build_vector dag nodes.
5173 static SDNode *tryToFoldExtendOfConstant(SDNode *N, const TargetLowering &TLI,
5174 SelectionDAG &DAG, bool LegalTypes,
5175 bool LegalOperations) {
5176 unsigned Opcode = N->getOpcode();
5177 SDValue N0 = N->getOperand(0);
5178 EVT VT = N->getValueType(0);
5180 assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND ||
5181 Opcode == ISD::ANY_EXTEND) && "Expected EXTEND dag node in input!");
5183 // fold (sext c1) -> c1
5184 // fold (zext c1) -> c1
5185 // fold (aext c1) -> c1
5186 if (isa<ConstantSDNode>(N0))
5187 return DAG.getNode(Opcode, SDLoc(N), VT, N0).getNode();
5189 // fold (sext (build_vector AllConstants) -> (build_vector AllConstants)
5190 // fold (zext (build_vector AllConstants) -> (build_vector AllConstants)
5191 // fold (aext (build_vector AllConstants) -> (build_vector AllConstants)
5192 EVT SVT = VT.getScalarType();
5193 if (!(VT.isVector() &&
5194 (!LegalTypes || (!LegalOperations && TLI.isTypeLegal(SVT))) &&
5195 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())))
5198 // We can fold this node into a build_vector.
5199 unsigned VTBits = SVT.getSizeInBits();
5200 unsigned EVTBits = N0->getValueType(0).getScalarType().getSizeInBits();
5201 unsigned ShAmt = VTBits - EVTBits;
5202 SmallVector<SDValue, 8> Elts;
5203 unsigned NumElts = N0->getNumOperands();
5206 for (unsigned i=0; i != NumElts; ++i) {
5207 SDValue Op = N0->getOperand(i);
5208 if (Op->getOpcode() == ISD::UNDEF) {
5209 Elts.push_back(DAG.getUNDEF(SVT));
5213 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
5214 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
5215 if (Opcode == ISD::SIGN_EXTEND)
5216 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
5219 Elts.push_back(DAG.getConstant(C.shl(ShAmt).lshr(ShAmt).getZExtValue(),
5223 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Elts).getNode();
5226 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
5227 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
5228 // transformation. Returns true if extension are possible and the above
5229 // mentioned transformation is profitable.
5230 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
5232 SmallVectorImpl<SDNode *> &ExtendNodes,
5233 const TargetLowering &TLI) {
5234 bool HasCopyToRegUses = false;
5235 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
5236 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
5237 UE = N0.getNode()->use_end();
5242 if (UI.getUse().getResNo() != N0.getResNo())
5244 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
5245 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
5246 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
5247 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
5248 // Sign bits will be lost after a zext.
5251 for (unsigned i = 0; i != 2; ++i) {
5252 SDValue UseOp = User->getOperand(i);
5255 if (!isa<ConstantSDNode>(UseOp))
5260 ExtendNodes.push_back(User);
5263 // If truncates aren't free and there are users we can't
5264 // extend, it isn't worthwhile.
5267 // Remember if this value is live-out.
5268 if (User->getOpcode() == ISD::CopyToReg)
5269 HasCopyToRegUses = true;
5272 if (HasCopyToRegUses) {
5273 bool BothLiveOut = false;
5274 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
5276 SDUse &Use = UI.getUse();
5277 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
5283 // Both unextended and extended values are live out. There had better be
5284 // a good reason for the transformation.
5285 return ExtendNodes.size();
5290 void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
5291 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
5292 ISD::NodeType ExtType) {
5293 // Extend SetCC uses if necessary.
5294 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
5295 SDNode *SetCC = SetCCs[i];
5296 SmallVector<SDValue, 4> Ops;
5298 for (unsigned j = 0; j != 2; ++j) {
5299 SDValue SOp = SetCC->getOperand(j);
5301 Ops.push_back(ExtLoad);
5303 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
5306 Ops.push_back(SetCC->getOperand(2));
5307 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), Ops));
5311 // FIXME: Bring more similar combines here, common to sext/zext (maybe aext?).
5312 SDValue DAGCombiner::CombineExtLoad(SDNode *N) {
5313 SDValue N0 = N->getOperand(0);
5314 EVT DstVT = N->getValueType(0);
5315 EVT SrcVT = N0.getValueType();
5317 assert((N->getOpcode() == ISD::SIGN_EXTEND ||
5318 N->getOpcode() == ISD::ZERO_EXTEND) &&
5319 "Unexpected node type (not an extend)!");
5321 // fold (sext (load x)) to multiple smaller sextloads; same for zext.
5322 // For example, on a target with legal v4i32, but illegal v8i32, turn:
5323 // (v8i32 (sext (v8i16 (load x))))
5325 // (v8i32 (concat_vectors (v4i32 (sextload x)),
5326 // (v4i32 (sextload (x + 16)))))
5327 // Where uses of the original load, i.e.:
5329 // are replaced with:
5331 // (v8i32 (concat_vectors (v4i32 (sextload x)),
5332 // (v4i32 (sextload (x + 16)))))))
5334 // This combine is only applicable to illegal, but splittable, vectors.
5335 // All legal types, and illegal non-vector types, are handled elsewhere.
5336 // This combine is controlled by TargetLowering::isVectorLoadExtDesirable.
5338 if (N0->getOpcode() != ISD::LOAD)
5341 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5343 if (!ISD::isNON_EXTLoad(LN0) || !ISD::isUNINDEXEDLoad(LN0) ||
5344 !N0.hasOneUse() || LN0->isVolatile() || !DstVT.isVector() ||
5345 !DstVT.isPow2VectorType() || !TLI.isVectorLoadExtDesirable(SDValue(N, 0)))
5348 SmallVector<SDNode *, 4> SetCCs;
5349 if (!ExtendUsesToFormExtLoad(N, N0, N->getOpcode(), SetCCs, TLI))
5352 ISD::LoadExtType ExtType =
5353 N->getOpcode() == ISD::SIGN_EXTEND ? ISD::SEXTLOAD : ISD::ZEXTLOAD;
5355 // Try to split the vector types to get down to legal types.
5356 EVT SplitSrcVT = SrcVT;
5357 EVT SplitDstVT = DstVT;
5358 while (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT) &&
5359 SplitSrcVT.getVectorNumElements() > 1) {
5360 SplitDstVT = DAG.GetSplitDestVTs(SplitDstVT).first;
5361 SplitSrcVT = DAG.GetSplitDestVTs(SplitSrcVT).first;
5364 if (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT))
5368 const unsigned NumSplits =
5369 DstVT.getVectorNumElements() / SplitDstVT.getVectorNumElements();
5370 const unsigned Stride = SplitSrcVT.getStoreSize();
5371 SmallVector<SDValue, 4> Loads;
5372 SmallVector<SDValue, 4> Chains;
5374 SDValue BasePtr = LN0->getBasePtr();
5375 for (unsigned Idx = 0; Idx < NumSplits; Idx++) {
5376 const unsigned Offset = Idx * Stride;
5377 const unsigned Align = MinAlign(LN0->getAlignment(), Offset);
5379 SDValue SplitLoad = DAG.getExtLoad(
5380 ExtType, DL, SplitDstVT, LN0->getChain(), BasePtr,
5381 LN0->getPointerInfo().getWithOffset(Offset), SplitSrcVT,
5382 LN0->isVolatile(), LN0->isNonTemporal(), LN0->isInvariant(),
5383 Align, LN0->getAAInfo());
5385 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
5386 DAG.getConstant(Stride, BasePtr.getValueType()));
5388 Loads.push_back(SplitLoad.getValue(0));
5389 Chains.push_back(SplitLoad.getValue(1));
5392 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
5393 SDValue NewValue = DAG.getNode(ISD::CONCAT_VECTORS, DL, DstVT, Loads);
5395 CombineTo(N, NewValue);
5397 // Replace uses of the original load (before extension)
5398 // with a truncate of the concatenated sextloaded vectors.
5400 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(), NewValue);
5401 CombineTo(N0.getNode(), Trunc, NewChain);
5402 ExtendSetCCUses(SetCCs, Trunc, NewValue, DL,
5403 (ISD::NodeType)N->getOpcode());
5404 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5407 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
5408 SDValue N0 = N->getOperand(0);
5409 EVT VT = N->getValueType(0);
5411 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5413 return SDValue(Res, 0);
5415 // fold (sext (sext x)) -> (sext x)
5416 // fold (sext (aext x)) -> (sext x)
5417 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5418 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT,
5421 if (N0.getOpcode() == ISD::TRUNCATE) {
5422 // fold (sext (truncate (load x))) -> (sext (smaller load x))
5423 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
5424 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5425 if (NarrowLoad.getNode()) {
5426 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5427 if (NarrowLoad.getNode() != N0.getNode()) {
5428 CombineTo(N0.getNode(), NarrowLoad);
5429 // CombineTo deleted the truncate, if needed, but not what's under it.
5432 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5435 // See if the value being truncated is already sign extended. If so, just
5436 // eliminate the trunc/sext pair.
5437 SDValue Op = N0.getOperand(0);
5438 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
5439 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
5440 unsigned DestBits = VT.getScalarType().getSizeInBits();
5441 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
5443 if (OpBits == DestBits) {
5444 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
5445 // bits, it is already ready.
5446 if (NumSignBits > DestBits-MidBits)
5448 } else if (OpBits < DestBits) {
5449 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
5450 // bits, just sext from i32.
5451 if (NumSignBits > OpBits-MidBits)
5452 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, Op);
5454 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
5455 // bits, just truncate to i32.
5456 if (NumSignBits > OpBits-MidBits)
5457 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5460 // fold (sext (truncate x)) -> (sextinreg x).
5461 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
5462 N0.getValueType())) {
5463 if (OpBits < DestBits)
5464 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op);
5465 else if (OpBits > DestBits)
5466 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op);
5467 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, Op,
5468 DAG.getValueType(N0.getValueType()));
5472 // fold (sext (load x)) -> (sext (truncate (sextload x)))
5473 // Only generate vector extloads when 1) they're legal, and 2) they are
5474 // deemed desirable by the target.
5475 if (ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5476 ((!LegalOperations && !VT.isVector() &&
5477 !cast<LoadSDNode>(N0)->isVolatile()) ||
5478 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, N0.getValueType()))) {
5479 bool DoXform = true;
5480 SmallVector<SDNode*, 4> SetCCs;
5481 if (!N0.hasOneUse())
5482 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
5484 DoXform &= TLI.isVectorLoadExtDesirable(SDValue(N, 0));
5486 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5487 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5489 LN0->getBasePtr(), N0.getValueType(),
5490 LN0->getMemOperand());
5491 CombineTo(N, ExtLoad);
5492 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5493 N0.getValueType(), ExtLoad);
5494 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5495 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5497 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5501 // fold (sext (load x)) to multiple smaller sextloads.
5502 // Only on illegal but splittable vectors.
5503 if (SDValue ExtLoad = CombineExtLoad(N))
5506 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
5507 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
5508 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5509 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5510 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5511 EVT MemVT = LN0->getMemoryVT();
5512 if ((!LegalOperations && !LN0->isVolatile()) ||
5513 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, MemVT)) {
5514 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5516 LN0->getBasePtr(), MemVT,
5517 LN0->getMemOperand());
5518 CombineTo(N, ExtLoad);
5519 CombineTo(N0.getNode(),
5520 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5521 N0.getValueType(), ExtLoad),
5522 ExtLoad.getValue(1));
5523 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5527 // fold (sext (and/or/xor (load x), cst)) ->
5528 // (and/or/xor (sextload x), (sext cst))
5529 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5530 N0.getOpcode() == ISD::XOR) &&
5531 isa<LoadSDNode>(N0.getOperand(0)) &&
5532 N0.getOperand(1).getOpcode() == ISD::Constant &&
5533 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, N0.getValueType()) &&
5534 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5535 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5536 if (LN0->getExtensionType() != ISD::ZEXTLOAD && LN0->isUnindexed()) {
5537 bool DoXform = true;
5538 SmallVector<SDNode*, 4> SetCCs;
5539 if (!N0.hasOneUse())
5540 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
5543 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN0), VT,
5544 LN0->getChain(), LN0->getBasePtr(),
5546 LN0->getMemOperand());
5547 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5548 Mask = Mask.sext(VT.getSizeInBits());
5549 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5550 ExtLoad, DAG.getConstant(Mask, VT));
5551 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5552 SDLoc(N0.getOperand(0)),
5553 N0.getOperand(0).getValueType(), ExtLoad);
5555 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5556 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5558 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5563 if (N0.getOpcode() == ISD::SETCC) {
5564 EVT N0VT = N0.getOperand(0).getValueType();
5565 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
5566 // Only do this before legalize for now.
5567 if (VT.isVector() && !LegalOperations &&
5568 TLI.getBooleanContents(N0VT) ==
5569 TargetLowering::ZeroOrNegativeOneBooleanContent) {
5570 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
5571 // of the same size as the compared operands. Only optimize sext(setcc())
5572 // if this is the case.
5573 EVT SVT = getSetCCResultType(N0VT);
5575 // We know that the # elements of the results is the same as the
5576 // # elements of the compare (and the # elements of the compare result
5577 // for that matter). Check to see that they are the same size. If so,
5578 // we know that the element size of the sext'd result matches the
5579 // element size of the compare operands.
5580 if (VT.getSizeInBits() == SVT.getSizeInBits())
5581 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5583 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5585 // If the desired elements are smaller or larger than the source
5586 // elements we can use a matching integer vector type and then
5587 // truncate/sign extend
5588 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5589 if (SVT == MatchingVectorType) {
5590 SDValue VsetCC = DAG.getSetCC(SDLoc(N), MatchingVectorType,
5591 N0.getOperand(0), N0.getOperand(1),
5592 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5593 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
5597 // sext(setcc x, y, cc) -> (select (setcc x, y, cc), -1, 0)
5598 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
5600 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
5602 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5603 NegOne, DAG.getConstant(0, VT),
5604 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5605 if (SCC.getNode()) return SCC;
5607 if (!VT.isVector()) {
5608 EVT SetCCVT = getSetCCResultType(N0.getOperand(0).getValueType());
5609 if (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, SetCCVT)) {
5611 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5612 SDValue SetCC = DAG.getSetCC(DL, SetCCVT,
5613 N0.getOperand(0), N0.getOperand(1), CC);
5614 return DAG.getSelect(DL, VT, SetCC,
5615 NegOne, DAG.getConstant(0, VT));
5620 // fold (sext x) -> (zext x) if the sign bit is known zero.
5621 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
5622 DAG.SignBitIsZero(N0))
5623 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0);
5628 // isTruncateOf - If N is a truncate of some other value, return true, record
5629 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
5630 // This function computes KnownZero to avoid a duplicated call to
5631 // computeKnownBits in the caller.
5632 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
5635 if (N->getOpcode() == ISD::TRUNCATE) {
5636 Op = N->getOperand(0);
5637 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5641 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
5642 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
5645 SDValue Op0 = N->getOperand(0);
5646 SDValue Op1 = N->getOperand(1);
5647 assert(Op0.getValueType() == Op1.getValueType());
5649 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
5650 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
5651 if (COp0 && COp0->isNullValue())
5653 else if (COp1 && COp1->isNullValue())
5658 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5660 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
5666 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
5667 SDValue N0 = N->getOperand(0);
5668 EVT VT = N->getValueType(0);
5670 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5672 return SDValue(Res, 0);
5674 // fold (zext (zext x)) -> (zext x)
5675 // fold (zext (aext x)) -> (zext x)
5676 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5677 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT,
5680 // fold (zext (truncate x)) -> (zext x) or
5681 // (zext (truncate x)) -> (truncate x)
5682 // This is valid when the truncated bits of x are already zero.
5683 // FIXME: We should extend this to work for vectors too.
5686 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
5687 APInt TruncatedBits =
5688 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
5689 APInt(Op.getValueSizeInBits(), 0) :
5690 APInt::getBitsSet(Op.getValueSizeInBits(),
5691 N0.getValueSizeInBits(),
5692 std::min(Op.getValueSizeInBits(),
5693 VT.getSizeInBits()));
5694 if (TruncatedBits == (KnownZero & TruncatedBits)) {
5695 if (VT.bitsGT(Op.getValueType()))
5696 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, Op);
5697 if (VT.bitsLT(Op.getValueType()))
5698 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5704 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5705 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
5706 if (N0.getOpcode() == ISD::TRUNCATE) {
5707 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5708 if (NarrowLoad.getNode()) {
5709 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5710 if (NarrowLoad.getNode() != N0.getNode()) {
5711 CombineTo(N0.getNode(), NarrowLoad);
5712 // CombineTo deleted the truncate, if needed, but not what's under it.
5715 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5719 // fold (zext (truncate x)) -> (and x, mask)
5720 if (N0.getOpcode() == ISD::TRUNCATE &&
5721 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
5723 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5724 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
5725 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5726 if (NarrowLoad.getNode()) {
5727 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5728 if (NarrowLoad.getNode() != N0.getNode()) {
5729 CombineTo(N0.getNode(), NarrowLoad);
5730 // CombineTo deleted the truncate, if needed, but not what's under it.
5733 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5736 SDValue Op = N0.getOperand(0);
5737 if (Op.getValueType().bitsLT(VT)) {
5738 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, Op);
5739 AddToWorklist(Op.getNode());
5740 } else if (Op.getValueType().bitsGT(VT)) {
5741 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5742 AddToWorklist(Op.getNode());
5744 return DAG.getZeroExtendInReg(Op, SDLoc(N),
5745 N0.getValueType().getScalarType());
5748 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
5749 // if either of the casts is not free.
5750 if (N0.getOpcode() == ISD::AND &&
5751 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5752 N0.getOperand(1).getOpcode() == ISD::Constant &&
5753 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5754 N0.getValueType()) ||
5755 !TLI.isZExtFree(N0.getValueType(), VT))) {
5756 SDValue X = N0.getOperand(0).getOperand(0);
5757 if (X.getValueType().bitsLT(VT)) {
5758 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(X), VT, X);
5759 } else if (X.getValueType().bitsGT(VT)) {
5760 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
5762 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5763 Mask = Mask.zext(VT.getSizeInBits());
5764 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5765 X, DAG.getConstant(Mask, VT));
5768 // fold (zext (load x)) -> (zext (truncate (zextload x)))
5769 // Only generate vector extloads when 1) they're legal, and 2) they are
5770 // deemed desirable by the target.
5771 if (ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5772 ((!LegalOperations && !VT.isVector() &&
5773 !cast<LoadSDNode>(N0)->isVolatile()) ||
5774 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, N0.getValueType()))) {
5775 bool DoXform = true;
5776 SmallVector<SDNode*, 4> SetCCs;
5777 if (!N0.hasOneUse())
5778 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
5780 DoXform &= TLI.isVectorLoadExtDesirable(SDValue(N, 0));
5782 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5783 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5785 LN0->getBasePtr(), N0.getValueType(),
5786 LN0->getMemOperand());
5787 CombineTo(N, ExtLoad);
5788 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5789 N0.getValueType(), ExtLoad);
5790 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5792 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5794 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5798 // fold (zext (load x)) to multiple smaller zextloads.
5799 // Only on illegal but splittable vectors.
5800 if (SDValue ExtLoad = CombineExtLoad(N))
5803 // fold (zext (and/or/xor (load x), cst)) ->
5804 // (and/or/xor (zextload x), (zext cst))
5805 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5806 N0.getOpcode() == ISD::XOR) &&
5807 isa<LoadSDNode>(N0.getOperand(0)) &&
5808 N0.getOperand(1).getOpcode() == ISD::Constant &&
5809 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, N0.getValueType()) &&
5810 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5811 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5812 if (LN0->getExtensionType() != ISD::SEXTLOAD && LN0->isUnindexed()) {
5813 bool DoXform = true;
5814 SmallVector<SDNode*, 4> SetCCs;
5815 if (!N0.hasOneUse())
5816 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
5819 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), VT,
5820 LN0->getChain(), LN0->getBasePtr(),
5822 LN0->getMemOperand());
5823 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5824 Mask = Mask.zext(VT.getSizeInBits());
5825 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5826 ExtLoad, DAG.getConstant(Mask, VT));
5827 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5828 SDLoc(N0.getOperand(0)),
5829 N0.getOperand(0).getValueType(), ExtLoad);
5831 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5832 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5834 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5839 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
5840 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
5841 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5842 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5843 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5844 EVT MemVT = LN0->getMemoryVT();
5845 if ((!LegalOperations && !LN0->isVolatile()) ||
5846 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT)) {
5847 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5849 LN0->getBasePtr(), MemVT,
5850 LN0->getMemOperand());
5851 CombineTo(N, ExtLoad);
5852 CombineTo(N0.getNode(),
5853 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(),
5855 ExtLoad.getValue(1));
5856 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5860 if (N0.getOpcode() == ISD::SETCC) {
5861 if (!LegalOperations && VT.isVector() &&
5862 N0.getValueType().getVectorElementType() == MVT::i1) {
5863 EVT N0VT = N0.getOperand(0).getValueType();
5864 if (getSetCCResultType(N0VT) == N0.getValueType())
5867 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
5868 // Only do this before legalize for now.
5869 EVT EltVT = VT.getVectorElementType();
5870 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
5871 DAG.getConstant(1, EltVT));
5872 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5873 // We know that the # elements of the results is the same as the
5874 // # elements of the compare (and the # elements of the compare result
5875 // for that matter). Check to see that they are the same size. If so,
5876 // we know that the element size of the sext'd result matches the
5877 // element size of the compare operands.
5878 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5879 DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5881 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
5882 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT,
5885 // If the desired elements are smaller or larger than the source
5886 // elements we can use a matching integer vector type and then
5887 // truncate/sign extend
5888 EVT MatchingElementType =
5889 EVT::getIntegerVT(*DAG.getContext(),
5890 N0VT.getScalarType().getSizeInBits());
5891 EVT MatchingVectorType =
5892 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
5893 N0VT.getVectorNumElements());
5895 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5897 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5898 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5899 DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT),
5900 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, OneOps));
5903 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5905 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5906 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5907 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5908 if (SCC.getNode()) return SCC;
5911 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
5912 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
5913 isa<ConstantSDNode>(N0.getOperand(1)) &&
5914 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
5916 SDValue ShAmt = N0.getOperand(1);
5917 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
5918 if (N0.getOpcode() == ISD::SHL) {
5919 SDValue InnerZExt = N0.getOperand(0);
5920 // If the original shl may be shifting out bits, do not perform this
5922 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
5923 InnerZExt.getOperand(0).getValueType().getSizeInBits();
5924 if (ShAmtVal > KnownZeroBits)
5930 // Ensure that the shift amount is wide enough for the shifted value.
5931 if (VT.getSizeInBits() >= 256)
5932 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
5934 return DAG.getNode(N0.getOpcode(), DL, VT,
5935 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
5942 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
5943 SDValue N0 = N->getOperand(0);
5944 EVT VT = N->getValueType(0);
5946 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5948 return SDValue(Res, 0);
5950 // fold (aext (aext x)) -> (aext x)
5951 // fold (aext (zext x)) -> (zext x)
5952 // fold (aext (sext x)) -> (sext x)
5953 if (N0.getOpcode() == ISD::ANY_EXTEND ||
5954 N0.getOpcode() == ISD::ZERO_EXTEND ||
5955 N0.getOpcode() == ISD::SIGN_EXTEND)
5956 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0));
5958 // fold (aext (truncate (load x))) -> (aext (smaller load x))
5959 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
5960 if (N0.getOpcode() == ISD::TRUNCATE) {
5961 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5962 if (NarrowLoad.getNode()) {
5963 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5964 if (NarrowLoad.getNode() != N0.getNode()) {
5965 CombineTo(N0.getNode(), NarrowLoad);
5966 // CombineTo deleted the truncate, if needed, but not what's under it.
5969 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5973 // fold (aext (truncate x))
5974 if (N0.getOpcode() == ISD::TRUNCATE) {
5975 SDValue TruncOp = N0.getOperand(0);
5976 if (TruncOp.getValueType() == VT)
5977 return TruncOp; // x iff x size == zext size.
5978 if (TruncOp.getValueType().bitsGT(VT))
5979 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, TruncOp);
5980 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, TruncOp);
5983 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
5984 // if the trunc is not free.
5985 if (N0.getOpcode() == ISD::AND &&
5986 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5987 N0.getOperand(1).getOpcode() == ISD::Constant &&
5988 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5989 N0.getValueType())) {
5990 SDValue X = N0.getOperand(0).getOperand(0);
5991 if (X.getValueType().bitsLT(VT)) {
5992 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, X);
5993 } else if (X.getValueType().bitsGT(VT)) {
5994 X = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, X);
5996 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5997 Mask = Mask.zext(VT.getSizeInBits());
5998 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5999 X, DAG.getConstant(Mask, VT));
6002 // fold (aext (load x)) -> (aext (truncate (extload x)))
6003 // None of the supported targets knows how to perform load and any_ext
6004 // on vectors in one instruction. We only perform this transformation on
6006 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
6007 ISD::isUNINDEXEDLoad(N0.getNode()) &&
6008 TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) {
6009 bool DoXform = true;
6010 SmallVector<SDNode*, 4> SetCCs;
6011 if (!N0.hasOneUse())
6012 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
6014 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6015 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
6017 LN0->getBasePtr(), N0.getValueType(),
6018 LN0->getMemOperand());
6019 CombineTo(N, ExtLoad);
6020 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
6021 N0.getValueType(), ExtLoad);
6022 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
6023 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
6025 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6029 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
6030 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
6031 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
6032 if (N0.getOpcode() == ISD::LOAD &&
6033 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
6035 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6036 ISD::LoadExtType ExtType = LN0->getExtensionType();
6037 EVT MemVT = LN0->getMemoryVT();
6038 if (!LegalOperations || TLI.isLoadExtLegal(ExtType, VT, MemVT)) {
6039 SDValue ExtLoad = DAG.getExtLoad(ExtType, SDLoc(N),
6040 VT, LN0->getChain(), LN0->getBasePtr(),
6041 MemVT, LN0->getMemOperand());
6042 CombineTo(N, ExtLoad);
6043 CombineTo(N0.getNode(),
6044 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
6045 N0.getValueType(), ExtLoad),
6046 ExtLoad.getValue(1));
6047 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6051 if (N0.getOpcode() == ISD::SETCC) {
6053 // aext(setcc) -> vsetcc
6054 // aext(setcc) -> truncate(vsetcc)
6055 // aext(setcc) -> aext(vsetcc)
6056 // Only do this before legalize for now.
6057 if (VT.isVector() && !LegalOperations) {
6058 EVT N0VT = N0.getOperand(0).getValueType();
6059 // We know that the # elements of the results is the same as the
6060 // # elements of the compare (and the # elements of the compare result
6061 // for that matter). Check to see that they are the same size. If so,
6062 // we know that the element size of the sext'd result matches the
6063 // element size of the compare operands.
6064 if (VT.getSizeInBits() == N0VT.getSizeInBits())
6065 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
6067 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6068 // If the desired elements are smaller or larger than the source
6069 // elements we can use a matching integer vector type and then
6070 // truncate/any extend
6072 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
6074 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
6076 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6077 return DAG.getAnyExtOrTrunc(VsetCC, SDLoc(N), VT);
6081 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
6083 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
6084 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
6085 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
6093 /// See if the specified operand can be simplified with the knowledge that only
6094 /// the bits specified by Mask are used. If so, return the simpler operand,
6095 /// otherwise return a null SDValue.
6096 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
6097 switch (V.getOpcode()) {
6099 case ISD::Constant: {
6100 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
6101 assert(CV && "Const value should be ConstSDNode.");
6102 const APInt &CVal = CV->getAPIntValue();
6103 APInt NewVal = CVal & Mask;
6105 return DAG.getConstant(NewVal, V.getValueType());
6110 // If the LHS or RHS don't contribute bits to the or, drop them.
6111 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
6112 return V.getOperand(1);
6113 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
6114 return V.getOperand(0);
6117 // Only look at single-use SRLs.
6118 if (!V.getNode()->hasOneUse())
6120 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
6121 // See if we can recursively simplify the LHS.
6122 unsigned Amt = RHSC->getZExtValue();
6124 // Watch out for shift count overflow though.
6125 if (Amt >= Mask.getBitWidth()) break;
6126 APInt NewMask = Mask << Amt;
6127 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
6128 if (SimplifyLHS.getNode())
6129 return DAG.getNode(ISD::SRL, SDLoc(V), V.getValueType(),
6130 SimplifyLHS, V.getOperand(1));
6136 /// If the result of a wider load is shifted to right of N bits and then
6137 /// truncated to a narrower type and where N is a multiple of number of bits of
6138 /// the narrower type, transform it to a narrower load from address + N / num of
6139 /// bits of new type. If the result is to be extended, also fold the extension
6140 /// to form a extending load.
6141 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
6142 unsigned Opc = N->getOpcode();
6144 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
6145 SDValue N0 = N->getOperand(0);
6146 EVT VT = N->getValueType(0);
6149 // This transformation isn't valid for vector loads.
6153 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
6155 if (Opc == ISD::SIGN_EXTEND_INREG) {
6156 ExtType = ISD::SEXTLOAD;
6157 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
6158 } else if (Opc == ISD::SRL) {
6159 // Another special-case: SRL is basically zero-extending a narrower value.
6160 ExtType = ISD::ZEXTLOAD;
6162 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
6163 if (!N01) return SDValue();
6164 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
6165 VT.getSizeInBits() - N01->getZExtValue());
6167 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, VT, ExtVT))
6170 unsigned EVTBits = ExtVT.getSizeInBits();
6172 // Do not generate loads of non-round integer types since these can
6173 // be expensive (and would be wrong if the type is not byte sized).
6174 if (!ExtVT.isRound())
6178 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
6179 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
6180 ShAmt = N01->getZExtValue();
6181 // Is the shift amount a multiple of size of VT?
6182 if ((ShAmt & (EVTBits-1)) == 0) {
6183 N0 = N0.getOperand(0);
6184 // Is the load width a multiple of size of VT?
6185 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
6189 // At this point, we must have a load or else we can't do the transform.
6190 if (!isa<LoadSDNode>(N0)) return SDValue();
6192 // Because a SRL must be assumed to *need* to zero-extend the high bits
6193 // (as opposed to anyext the high bits), we can't combine the zextload
6194 // lowering of SRL and an sextload.
6195 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
6198 // If the shift amount is larger than the input type then we're not
6199 // accessing any of the loaded bytes. If the load was a zextload/extload
6200 // then the result of the shift+trunc is zero/undef (handled elsewhere).
6201 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
6206 // If the load is shifted left (and the result isn't shifted back right),
6207 // we can fold the truncate through the shift.
6208 unsigned ShLeftAmt = 0;
6209 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
6210 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
6211 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
6212 ShLeftAmt = N01->getZExtValue();
6213 N0 = N0.getOperand(0);
6217 // If we haven't found a load, we can't narrow it. Don't transform one with
6218 // multiple uses, this would require adding a new load.
6219 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse())
6222 // Don't change the width of a volatile load.
6223 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6224 if (LN0->isVolatile())
6227 // Verify that we are actually reducing a load width here.
6228 if (LN0->getMemoryVT().getSizeInBits() < EVTBits)
6231 // For the transform to be legal, the load must produce only two values
6232 // (the value loaded and the chain). Don't transform a pre-increment
6233 // load, for example, which produces an extra value. Otherwise the
6234 // transformation is not equivalent, and the downstream logic to replace
6235 // uses gets things wrong.
6236 if (LN0->getNumValues() > 2)
6239 // If the load that we're shrinking is an extload and we're not just
6240 // discarding the extension we can't simply shrink the load. Bail.
6241 // TODO: It would be possible to merge the extensions in some cases.
6242 if (LN0->getExtensionType() != ISD::NON_EXTLOAD &&
6243 LN0->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits() + ShAmt)
6246 if (!TLI.shouldReduceLoadWidth(LN0, ExtType, ExtVT))
6249 EVT PtrType = N0.getOperand(1).getValueType();
6251 if (PtrType == MVT::Untyped || PtrType.isExtended())
6252 // It's not possible to generate a constant of extended or untyped type.
6255 // For big endian targets, we need to adjust the offset to the pointer to
6256 // load the correct bytes.
6257 if (TLI.isBigEndian()) {
6258 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
6259 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
6260 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
6263 uint64_t PtrOff = ShAmt / 8;
6264 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
6265 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0),
6266 PtrType, LN0->getBasePtr(),
6267 DAG.getConstant(PtrOff, PtrType));
6268 AddToWorklist(NewPtr.getNode());
6271 if (ExtType == ISD::NON_EXTLOAD)
6272 Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr,
6273 LN0->getPointerInfo().getWithOffset(PtrOff),
6274 LN0->isVolatile(), LN0->isNonTemporal(),
6275 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
6277 Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(),NewPtr,
6278 LN0->getPointerInfo().getWithOffset(PtrOff),
6279 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
6280 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
6282 // Replace the old load's chain with the new load's chain.
6283 WorklistRemover DeadNodes(*this);
6284 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
6286 // Shift the result left, if we've swallowed a left shift.
6287 SDValue Result = Load;
6288 if (ShLeftAmt != 0) {
6289 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
6290 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
6292 // If the shift amount is as large as the result size (but, presumably,
6293 // no larger than the source) then the useful bits of the result are
6294 // zero; we can't simply return the shortened shift, because the result
6295 // of that operation is undefined.
6296 if (ShLeftAmt >= VT.getSizeInBits())
6297 Result = DAG.getConstant(0, VT);
6299 Result = DAG.getNode(ISD::SHL, SDLoc(N0), VT,
6300 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
6303 // Return the new loaded value.
6307 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
6308 SDValue N0 = N->getOperand(0);
6309 SDValue N1 = N->getOperand(1);
6310 EVT VT = N->getValueType(0);
6311 EVT EVT = cast<VTSDNode>(N1)->getVT();
6312 unsigned VTBits = VT.getScalarType().getSizeInBits();
6313 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
6315 // fold (sext_in_reg c1) -> c1
6316 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
6317 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1);
6319 // If the input is already sign extended, just drop the extension.
6320 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
6323 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
6324 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
6325 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT()))
6326 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
6327 N0.getOperand(0), N1);
6329 // fold (sext_in_reg (sext x)) -> (sext x)
6330 // fold (sext_in_reg (aext x)) -> (sext x)
6331 // if x is small enough.
6332 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
6333 SDValue N00 = N0.getOperand(0);
6334 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
6335 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
6336 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
6339 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
6340 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
6341 return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT);
6343 // fold operands of sext_in_reg based on knowledge that the top bits are not
6345 if (SimplifyDemandedBits(SDValue(N, 0)))
6346 return SDValue(N, 0);
6348 // fold (sext_in_reg (load x)) -> (smaller sextload x)
6349 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
6350 SDValue NarrowLoad = ReduceLoadWidth(N);
6351 if (NarrowLoad.getNode())
6354 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
6355 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
6356 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
6357 if (N0.getOpcode() == ISD::SRL) {
6358 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
6359 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
6360 // We can turn this into an SRA iff the input to the SRL is already sign
6362 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
6363 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
6364 return DAG.getNode(ISD::SRA, SDLoc(N), VT,
6365 N0.getOperand(0), N0.getOperand(1));
6369 // fold (sext_inreg (extload x)) -> (sextload x)
6370 if (ISD::isEXTLoad(N0.getNode()) &&
6371 ISD::isUNINDEXEDLoad(N0.getNode()) &&
6372 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
6373 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6374 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) {
6375 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6376 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
6378 LN0->getBasePtr(), EVT,
6379 LN0->getMemOperand());
6380 CombineTo(N, ExtLoad);
6381 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
6382 AddToWorklist(ExtLoad.getNode());
6383 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6385 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
6386 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
6388 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
6389 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6390 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) {
6391 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6392 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
6394 LN0->getBasePtr(), EVT,
6395 LN0->getMemOperand());
6396 CombineTo(N, ExtLoad);
6397 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
6398 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6401 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
6402 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
6403 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
6404 N0.getOperand(1), false);
6405 if (BSwap.getNode())
6406 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
6410 // Fold a sext_inreg of a build_vector of ConstantSDNodes or undefs
6411 // into a build_vector.
6412 if (ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
6413 SmallVector<SDValue, 8> Elts;
6414 unsigned NumElts = N0->getNumOperands();
6415 unsigned ShAmt = VTBits - EVTBits;
6417 for (unsigned i = 0; i != NumElts; ++i) {
6418 SDValue Op = N0->getOperand(i);
6419 if (Op->getOpcode() == ISD::UNDEF) {
6424 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
6425 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
6426 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
6427 Op.getValueType()));
6430 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Elts);
6436 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
6437 SDValue N0 = N->getOperand(0);
6438 EVT VT = N->getValueType(0);
6439 bool isLE = TLI.isLittleEndian();
6442 if (N0.getValueType() == N->getValueType(0))
6444 // fold (truncate c1) -> c1
6445 if (isa<ConstantSDNode>(N0))
6446 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0);
6447 // fold (truncate (truncate x)) -> (truncate x)
6448 if (N0.getOpcode() == ISD::TRUNCATE)
6449 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6450 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
6451 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
6452 N0.getOpcode() == ISD::SIGN_EXTEND ||
6453 N0.getOpcode() == ISD::ANY_EXTEND) {
6454 if (N0.getOperand(0).getValueType().bitsLT(VT))
6455 // if the source is smaller than the dest, we still need an extend
6456 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
6458 if (N0.getOperand(0).getValueType().bitsGT(VT))
6459 // if the source is larger than the dest, than we just need the truncate
6460 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6461 // if the source and dest are the same type, we can drop both the extend
6462 // and the truncate.
6463 return N0.getOperand(0);
6466 // Fold extract-and-trunc into a narrow extract. For example:
6467 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
6468 // i32 y = TRUNCATE(i64 x)
6470 // v16i8 b = BITCAST (v2i64 val)
6471 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
6473 // Note: We only run this optimization after type legalization (which often
6474 // creates this pattern) and before operation legalization after which
6475 // we need to be more careful about the vector instructions that we generate.
6476 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6477 LegalTypes && !LegalOperations && N0->hasOneUse() && VT != MVT::i1) {
6479 EVT VecTy = N0.getOperand(0).getValueType();
6480 EVT ExTy = N0.getValueType();
6481 EVT TrTy = N->getValueType(0);
6483 unsigned NumElem = VecTy.getVectorNumElements();
6484 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
6486 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
6487 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
6489 SDValue EltNo = N0->getOperand(1);
6490 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
6491 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
6492 EVT IndexTy = TLI.getVectorIdxTy();
6493 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
6495 SDValue V = DAG.getNode(ISD::BITCAST, SDLoc(N),
6496 NVT, N0.getOperand(0));
6498 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
6500 DAG.getConstant(Index, IndexTy));
6504 // trunc (select c, a, b) -> select c, (trunc a), (trunc b)
6505 if (N0.getOpcode() == ISD::SELECT) {
6506 EVT SrcVT = N0.getValueType();
6507 if ((!LegalOperations || TLI.isOperationLegal(ISD::SELECT, SrcVT)) &&
6508 TLI.isTruncateFree(SrcVT, VT)) {
6510 SDValue Cond = N0.getOperand(0);
6511 SDValue TruncOp0 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(1));
6512 SDValue TruncOp1 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(2));
6513 return DAG.getNode(ISD::SELECT, SDLoc(N), VT, Cond, TruncOp0, TruncOp1);
6517 // Fold a series of buildvector, bitcast, and truncate if possible.
6519 // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to
6520 // (2xi32 (buildvector x, y)).
6521 if (Level == AfterLegalizeVectorOps && VT.isVector() &&
6522 N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
6523 N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
6524 N0.getOperand(0).hasOneUse()) {
6526 SDValue BuildVect = N0.getOperand(0);
6527 EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType();
6528 EVT TruncVecEltTy = VT.getVectorElementType();
6530 // Check that the element types match.
6531 if (BuildVectEltTy == TruncVecEltTy) {
6532 // Now we only need to compute the offset of the truncated elements.
6533 unsigned BuildVecNumElts = BuildVect.getNumOperands();
6534 unsigned TruncVecNumElts = VT.getVectorNumElements();
6535 unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts;
6537 assert((BuildVecNumElts % TruncVecNumElts) == 0 &&
6538 "Invalid number of elements");
6540 SmallVector<SDValue, 8> Opnds;
6541 for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset)
6542 Opnds.push_back(BuildVect.getOperand(i));
6544 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
6548 // See if we can simplify the input to this truncate through knowledge that
6549 // only the low bits are being used.
6550 // For example "trunc (or (shl x, 8), y)" // -> trunc y
6551 // Currently we only perform this optimization on scalars because vectors
6552 // may have different active low bits.
6553 if (!VT.isVector()) {
6555 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
6556 VT.getSizeInBits()));
6557 if (Shorter.getNode())
6558 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter);
6560 // fold (truncate (load x)) -> (smaller load x)
6561 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
6562 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
6563 SDValue Reduced = ReduceLoadWidth(N);
6564 if (Reduced.getNode())
6566 // Handle the case where the load remains an extending load even
6567 // after truncation.
6568 if (N0.hasOneUse() && ISD::isUNINDEXEDLoad(N0.getNode())) {
6569 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6570 if (!LN0->isVolatile() &&
6571 LN0->getMemoryVT().getStoreSizeInBits() < VT.getSizeInBits()) {
6572 SDValue NewLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(LN0),
6573 VT, LN0->getChain(), LN0->getBasePtr(),
6575 LN0->getMemOperand());
6576 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLoad.getValue(1));
6581 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
6582 // where ... are all 'undef'.
6583 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
6584 SmallVector<EVT, 8> VTs;
6587 unsigned NumDefs = 0;
6589 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
6590 SDValue X = N0.getOperand(i);
6591 if (X.getOpcode() != ISD::UNDEF) {
6596 // Stop if more than one members are non-undef.
6599 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
6600 VT.getVectorElementType(),
6601 X.getValueType().getVectorNumElements()));
6605 return DAG.getUNDEF(VT);
6608 assert(V.getNode() && "The single defined operand is empty!");
6609 SmallVector<SDValue, 8> Opnds;
6610 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
6612 Opnds.push_back(DAG.getUNDEF(VTs[i]));
6615 SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V);
6616 AddToWorklist(NV.getNode());
6617 Opnds.push_back(NV);
6619 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Opnds);
6623 // Simplify the operands using demanded-bits information.
6624 if (!VT.isVector() &&
6625 SimplifyDemandedBits(SDValue(N, 0)))
6626 return SDValue(N, 0);
6631 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
6632 SDValue Elt = N->getOperand(i);
6633 if (Elt.getOpcode() != ISD::MERGE_VALUES)
6634 return Elt.getNode();
6635 return Elt.getOperand(Elt.getResNo()).getNode();
6638 /// build_pair (load, load) -> load
6639 /// if load locations are consecutive.
6640 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
6641 assert(N->getOpcode() == ISD::BUILD_PAIR);
6643 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
6644 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
6645 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
6646 LD1->getAddressSpace() != LD2->getAddressSpace())
6648 EVT LD1VT = LD1->getValueType(0);
6650 if (ISD::isNON_EXTLoad(LD2) &&
6652 // If both are volatile this would reduce the number of volatile loads.
6653 // If one is volatile it might be ok, but play conservative and bail out.
6654 !LD1->isVolatile() &&
6655 !LD2->isVolatile() &&
6656 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
6657 unsigned Align = LD1->getAlignment();
6658 unsigned NewAlign = TLI.getDataLayout()->
6659 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6661 if (NewAlign <= Align &&
6662 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
6663 return DAG.getLoad(VT, SDLoc(N), LD1->getChain(),
6664 LD1->getBasePtr(), LD1->getPointerInfo(),
6665 false, false, false, Align);
6671 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
6672 SDValue N0 = N->getOperand(0);
6673 EVT VT = N->getValueType(0);
6675 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
6676 // Only do this before legalize, since afterward the target may be depending
6677 // on the bitconvert.
6678 // First check to see if this is all constant.
6680 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
6682 bool isSimple = cast<BuildVectorSDNode>(N0)->isConstant();
6684 EVT DestEltVT = N->getValueType(0).getVectorElementType();
6685 assert(!DestEltVT.isVector() &&
6686 "Element type of vector ValueType must not be vector!");
6688 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
6691 // If the input is a constant, let getNode fold it.
6692 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
6693 // If we can't allow illegal operations, we need to check that this is just
6694 // a fp -> int or int -> conversion and that the resulting operation will
6696 if (!LegalOperations ||
6697 (isa<ConstantSDNode>(N0) && VT.isFloatingPoint() && !VT.isVector() &&
6698 TLI.isOperationLegal(ISD::ConstantFP, VT)) ||
6699 (isa<ConstantFPSDNode>(N0) && VT.isInteger() && !VT.isVector() &&
6700 TLI.isOperationLegal(ISD::Constant, VT)))
6701 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, N0);
6704 // (conv (conv x, t1), t2) -> (conv x, t2)
6705 if (N0.getOpcode() == ISD::BITCAST)
6706 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT,
6709 // fold (conv (load x)) -> (load (conv*)x)
6710 // If the resultant load doesn't need a higher alignment than the original!
6711 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6712 // Do not change the width of a volatile load.
6713 !cast<LoadSDNode>(N0)->isVolatile() &&
6714 // Do not remove the cast if the types differ in endian layout.
6715 TLI.hasBigEndianPartOrdering(N0.getValueType()) ==
6716 TLI.hasBigEndianPartOrdering(VT) &&
6717 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)) &&
6718 TLI.isLoadBitCastBeneficial(N0.getValueType(), VT)) {
6719 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6720 unsigned Align = TLI.getDataLayout()->
6721 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6722 unsigned OrigAlign = LN0->getAlignment();
6724 if (Align <= OrigAlign) {
6725 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(),
6726 LN0->getBasePtr(), LN0->getPointerInfo(),
6727 LN0->isVolatile(), LN0->isNonTemporal(),
6728 LN0->isInvariant(), OrigAlign,
6730 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
6735 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
6736 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
6737 // This often reduces constant pool loads.
6738 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) ||
6739 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) &&
6740 N0.getNode()->hasOneUse() && VT.isInteger() &&
6741 !VT.isVector() && !N0.getValueType().isVector()) {
6742 SDValue NewConv = DAG.getNode(ISD::BITCAST, SDLoc(N0), VT,
6744 AddToWorklist(NewConv.getNode());
6746 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6747 if (N0.getOpcode() == ISD::FNEG)
6748 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
6749 NewConv, DAG.getConstant(SignBit, VT));
6750 assert(N0.getOpcode() == ISD::FABS);
6751 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6752 NewConv, DAG.getConstant(~SignBit, VT));
6755 // fold (bitconvert (fcopysign cst, x)) ->
6756 // (or (and (bitconvert x), sign), (and cst, (not sign)))
6757 // Note that we don't handle (copysign x, cst) because this can always be
6758 // folded to an fneg or fabs.
6759 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
6760 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
6761 VT.isInteger() && !VT.isVector()) {
6762 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
6763 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
6764 if (isTypeLegal(IntXVT)) {
6765 SDValue X = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6766 IntXVT, N0.getOperand(1));
6767 AddToWorklist(X.getNode());
6769 // If X has a different width than the result/lhs, sext it or truncate it.
6770 unsigned VTWidth = VT.getSizeInBits();
6771 if (OrigXWidth < VTWidth) {
6772 X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X);
6773 AddToWorklist(X.getNode());
6774 } else if (OrigXWidth > VTWidth) {
6775 // To get the sign bit in the right place, we have to shift it right
6776 // before truncating.
6777 X = DAG.getNode(ISD::SRL, SDLoc(X),
6778 X.getValueType(), X,
6779 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
6780 AddToWorklist(X.getNode());
6781 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
6782 AddToWorklist(X.getNode());
6785 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6786 X = DAG.getNode(ISD::AND, SDLoc(X), VT,
6787 X, DAG.getConstant(SignBit, VT));
6788 AddToWorklist(X.getNode());
6790 SDValue Cst = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6791 VT, N0.getOperand(0));
6792 Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT,
6793 Cst, DAG.getConstant(~SignBit, VT));
6794 AddToWorklist(Cst.getNode());
6796 return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst);
6800 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
6801 if (N0.getOpcode() == ISD::BUILD_PAIR) {
6802 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
6803 if (CombineLD.getNode())
6810 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
6811 EVT VT = N->getValueType(0);
6812 return CombineConsecutiveLoads(N, VT);
6815 /// We know that BV is a build_vector node with Constant, ConstantFP or Undef
6816 /// operands. DstEltVT indicates the destination element value type.
6817 SDValue DAGCombiner::
6818 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
6819 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
6821 // If this is already the right type, we're done.
6822 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
6824 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
6825 unsigned DstBitSize = DstEltVT.getSizeInBits();
6827 // If this is a conversion of N elements of one type to N elements of another
6828 // type, convert each element. This handles FP<->INT cases.
6829 if (SrcBitSize == DstBitSize) {
6830 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6831 BV->getValueType(0).getVectorNumElements());
6833 // Due to the FP element handling below calling this routine recursively,
6834 // we can end up with a scalar-to-vector node here.
6835 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
6836 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6837 DAG.getNode(ISD::BITCAST, SDLoc(BV),
6838 DstEltVT, BV->getOperand(0)));
6840 SmallVector<SDValue, 8> Ops;
6841 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6842 SDValue Op = BV->getOperand(i);
6843 // If the vector element type is not legal, the BUILD_VECTOR operands
6844 // are promoted and implicitly truncated. Make that explicit here.
6845 if (Op.getValueType() != SrcEltVT)
6846 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op);
6847 Ops.push_back(DAG.getNode(ISD::BITCAST, SDLoc(BV),
6849 AddToWorklist(Ops.back().getNode());
6851 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6854 // Otherwise, we're growing or shrinking the elements. To avoid having to
6855 // handle annoying details of growing/shrinking FP values, we convert them to
6857 if (SrcEltVT.isFloatingPoint()) {
6858 // Convert the input float vector to a int vector where the elements are the
6860 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
6861 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
6865 // Now we know the input is an integer vector. If the output is a FP type,
6866 // convert to integer first, then to FP of the right size.
6867 if (DstEltVT.isFloatingPoint()) {
6868 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
6869 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
6871 // Next, convert to FP elements of the same size.
6872 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
6875 // Okay, we know the src/dst types are both integers of differing types.
6876 // Handling growing first.
6877 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
6878 if (SrcBitSize < DstBitSize) {
6879 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
6881 SmallVector<SDValue, 8> Ops;
6882 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
6883 i += NumInputsPerOutput) {
6884 bool isLE = TLI.isLittleEndian();
6885 APInt NewBits = APInt(DstBitSize, 0);
6886 bool EltIsUndef = true;
6887 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
6888 // Shift the previously computed bits over.
6889 NewBits <<= SrcBitSize;
6890 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
6891 if (Op.getOpcode() == ISD::UNDEF) continue;
6894 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
6895 zextOrTrunc(SrcBitSize).zext(DstBitSize);
6899 Ops.push_back(DAG.getUNDEF(DstEltVT));
6901 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
6904 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
6905 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6908 // Finally, this must be the case where we are shrinking elements: each input
6909 // turns into multiple outputs.
6910 bool isS2V = ISD::isScalarToVector(BV);
6911 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
6912 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6913 NumOutputsPerInput*BV->getNumOperands());
6914 SmallVector<SDValue, 8> Ops;
6916 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6917 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
6918 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
6919 Ops.push_back(DAG.getUNDEF(DstEltVT));
6923 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
6924 getAPIntValue().zextOrTrunc(SrcBitSize);
6926 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
6927 APInt ThisVal = OpVal.trunc(DstBitSize);
6928 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
6929 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
6930 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
6931 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6933 OpVal = OpVal.lshr(DstBitSize);
6936 // For big endian targets, swap the order of the pieces of each element.
6937 if (TLI.isBigEndian())
6938 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
6941 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6944 SDValue DAGCombiner::visitFADD(SDNode *N) {
6945 SDValue N0 = N->getOperand(0);
6946 SDValue N1 = N->getOperand(1);
6947 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6948 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6949 EVT VT = N->getValueType(0);
6950 const TargetOptions &Options = DAG.getTarget().Options;
6953 if (VT.isVector()) {
6954 SDValue FoldedVOp = SimplifyVBinOp(N);
6955 if (FoldedVOp.getNode()) return FoldedVOp;
6958 // fold (fadd c1, c2) -> c1 + c2
6960 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N1);
6962 // canonicalize constant to RHS
6963 if (N0CFP && !N1CFP)
6964 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N0);
6966 // fold (fadd A, (fneg B)) -> (fsub A, B)
6967 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6968 isNegatibleForFree(N1, LegalOperations, TLI, &Options) == 2)
6969 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0,
6970 GetNegatedExpression(N1, DAG, LegalOperations));
6972 // fold (fadd (fneg A), B) -> (fsub B, A)
6973 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6974 isNegatibleForFree(N0, LegalOperations, TLI, &Options) == 2)
6975 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N1,
6976 GetNegatedExpression(N0, DAG, LegalOperations));
6978 // If 'unsafe math' is enabled, fold lots of things.
6979 if (Options.UnsafeFPMath) {
6980 // No FP constant should be created after legalization as Instruction
6981 // Selection pass has a hard time dealing with FP constants.
6982 bool AllowNewConst = (Level < AfterLegalizeDAG);
6984 // fold (fadd A, 0) -> A
6985 if (N1CFP && N1CFP->getValueAPF().isZero())
6988 // fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
6989 if (N1CFP && N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
6990 isa<ConstantFPSDNode>(N0.getOperand(1)))
6991 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0.getOperand(0),
6992 DAG.getNode(ISD::FADD, SDLoc(N), VT,
6993 N0.getOperand(1), N1));
6995 // If allowed, fold (fadd (fneg x), x) -> 0.0
6996 if (AllowNewConst && N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1)
6997 return DAG.getConstantFP(0.0, VT);
6999 // If allowed, fold (fadd x, (fneg x)) -> 0.0
7000 if (AllowNewConst && N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0)
7001 return DAG.getConstantFP(0.0, VT);
7003 // We can fold chains of FADD's of the same value into multiplications.
7004 // This transform is not safe in general because we are reducing the number
7005 // of rounding steps.
7006 if (TLI.isOperationLegalOrCustom(ISD::FMUL, VT) && !N0CFP && !N1CFP) {
7007 if (N0.getOpcode() == ISD::FMUL) {
7008 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
7009 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
7011 // (fadd (fmul x, c), x) -> (fmul x, c+1)
7012 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
7013 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
7015 DAG.getConstantFP(1.0, VT));
7016 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, NewCFP);
7019 // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2)
7020 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
7021 N1.getOperand(0) == N1.getOperand(1) &&
7022 N0.getOperand(0) == N1.getOperand(0)) {
7023 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
7025 DAG.getConstantFP(2.0, VT));
7026 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7027 N0.getOperand(0), NewCFP);
7031 if (N1.getOpcode() == ISD::FMUL) {
7032 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
7033 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
7035 // (fadd x, (fmul x, c)) -> (fmul x, c+1)
7036 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
7037 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
7039 DAG.getConstantFP(1.0, VT));
7040 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, NewCFP);
7043 // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2)
7044 if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD &&
7045 N0.getOperand(0) == N0.getOperand(1) &&
7046 N1.getOperand(0) == N0.getOperand(0)) {
7047 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
7049 DAG.getConstantFP(2.0, VT));
7050 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1.getOperand(0), NewCFP);
7054 if (N0.getOpcode() == ISD::FADD && AllowNewConst) {
7055 ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
7056 // (fadd (fadd x, x), x) -> (fmul x, 3.0)
7057 if (!CFP && N0.getOperand(0) == N0.getOperand(1) &&
7058 (N0.getOperand(0) == N1))
7059 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7060 N1, DAG.getConstantFP(3.0, VT));
7063 if (N1.getOpcode() == ISD::FADD && AllowNewConst) {
7064 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
7065 // (fadd x, (fadd x, x)) -> (fmul x, 3.0)
7066 if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) &&
7067 N1.getOperand(0) == N0)
7068 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7069 N0, DAG.getConstantFP(3.0, VT));
7072 // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0)
7073 if (AllowNewConst &&
7074 N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
7075 N0.getOperand(0) == N0.getOperand(1) &&
7076 N1.getOperand(0) == N1.getOperand(1) &&
7077 N0.getOperand(0) == N1.getOperand(0))
7078 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7079 N0.getOperand(0), DAG.getConstantFP(4.0, VT));
7081 } // enable-unsafe-fp-math
7083 // FADD -> FMA combines:
7084 if ((Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath) &&
7085 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
7086 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
7088 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
7089 if (N0.getOpcode() == ISD::FMUL &&
7090 (N0->hasOneUse() || TLI.enableAggressiveFMAFusion(VT)))
7091 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7092 N0.getOperand(0), N0.getOperand(1), N1);
7094 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
7095 // Note: Commutes FADD operands.
7096 if (N1.getOpcode() == ISD::FMUL &&
7097 (N1->hasOneUse() || TLI.enableAggressiveFMAFusion(VT)))
7098 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7099 N1.getOperand(0), N1.getOperand(1), N0);
7101 // When FP_EXTEND nodes are free on the target, and there is an opportunity
7102 // to combine into FMA, arrange such nodes accordingly.
7103 if (TLI.isFPExtFree(VT)) {
7105 // fold (fadd (fpext (fmul x, y)), z) -> (fma (fpext x), (fpext y), z)
7106 if (N0.getOpcode() == ISD::FP_EXTEND) {
7107 SDValue N00 = N0.getOperand(0);
7108 if (N00.getOpcode() == ISD::FMUL)
7109 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7110 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7112 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7113 N00.getOperand(1)), N1);
7116 // fold (fadd x, (fpext (fmul y, z)), z) -> (fma (fpext y), (fpext z), x)
7117 // Note: Commutes FADD operands.
7118 if (N1.getOpcode() == ISD::FP_EXTEND) {
7119 SDValue N10 = N1.getOperand(0);
7120 if (N10.getOpcode() == ISD::FMUL)
7121 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7122 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7124 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7125 N10.getOperand(1)), N0);
7129 // More folding opportunities when target permits.
7130 if (TLI.enableAggressiveFMAFusion(VT)) {
7132 // fold (fadd (fma x, y, (fmul u, v)), z) -> (fma x, y (fma u, v, z))
7133 if (N0.getOpcode() == ISD::FMA &&
7134 N0.getOperand(2).getOpcode() == ISD::FMUL)
7135 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7136 N0.getOperand(0), N0.getOperand(1),
7137 DAG.getNode(ISD::FMA, SDLoc(N), VT,
7138 N0.getOperand(2).getOperand(0),
7139 N0.getOperand(2).getOperand(1),
7142 // fold (fadd x, (fma y, z, (fmul u, v)) -> (fma y, z (fma u, v, x))
7143 if (N1->getOpcode() == ISD::FMA &&
7144 N1.getOperand(2).getOpcode() == ISD::FMUL)
7145 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7146 N1.getOperand(0), N1.getOperand(1),
7147 DAG.getNode(ISD::FMA, SDLoc(N), VT,
7148 N1.getOperand(2).getOperand(0),
7149 N1.getOperand(2).getOperand(1),
7157 SDValue DAGCombiner::visitFSUB(SDNode *N) {
7158 SDValue N0 = N->getOperand(0);
7159 SDValue N1 = N->getOperand(1);
7160 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
7161 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
7162 EVT VT = N->getValueType(0);
7164 const TargetOptions &Options = DAG.getTarget().Options;
7167 if (VT.isVector()) {
7168 SDValue FoldedVOp = SimplifyVBinOp(N);
7169 if (FoldedVOp.getNode()) return FoldedVOp;
7172 // fold (fsub c1, c2) -> c1-c2
7174 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0, N1);
7176 // fold (fsub A, (fneg B)) -> (fadd A, B)
7177 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
7178 return DAG.getNode(ISD::FADD, dl, VT, N0,
7179 GetNegatedExpression(N1, DAG, LegalOperations));
7181 // If 'unsafe math' is enabled, fold lots of things.
7182 if (Options.UnsafeFPMath) {
7184 if (N1CFP && N1CFP->getValueAPF().isZero())
7187 // (fsub 0, B) -> -B
7188 if (N0CFP && N0CFP->getValueAPF().isZero()) {
7189 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
7190 return GetNegatedExpression(N1, DAG, LegalOperations);
7191 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7192 return DAG.getNode(ISD::FNEG, dl, VT, N1);
7195 // (fsub x, x) -> 0.0
7197 return DAG.getConstantFP(0.0f, VT);
7199 // (fsub x, (fadd x, y)) -> (fneg y)
7200 // (fsub x, (fadd y, x)) -> (fneg y)
7201 if (N1.getOpcode() == ISD::FADD) {
7202 SDValue N10 = N1->getOperand(0);
7203 SDValue N11 = N1->getOperand(1);
7205 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI, &Options))
7206 return GetNegatedExpression(N11, DAG, LegalOperations);
7208 if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI, &Options))
7209 return GetNegatedExpression(N10, DAG, LegalOperations);
7213 // FSUB -> FMA combines:
7214 if ((Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath) &&
7215 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
7216 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
7218 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
7219 if (N0.getOpcode() == ISD::FMUL &&
7220 (N0->hasOneUse() || TLI.enableAggressiveFMAFusion(VT)))
7221 return DAG.getNode(ISD::FMA, dl, VT,
7222 N0.getOperand(0), N0.getOperand(1),
7223 DAG.getNode(ISD::FNEG, dl, VT, N1));
7225 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
7226 // Note: Commutes FSUB operands.
7227 if (N1.getOpcode() == ISD::FMUL &&
7228 (N1->hasOneUse() || TLI.enableAggressiveFMAFusion(VT)))
7229 return DAG.getNode(ISD::FMA, dl, VT,
7230 DAG.getNode(ISD::FNEG, dl, VT,
7232 N1.getOperand(1), N0);
7234 // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
7235 if (N0.getOpcode() == ISD::FNEG &&
7236 N0.getOperand(0).getOpcode() == ISD::FMUL &&
7237 ((N0->hasOneUse() && N0.getOperand(0).hasOneUse()) ||
7238 TLI.enableAggressiveFMAFusion(VT))) {
7239 SDValue N00 = N0.getOperand(0).getOperand(0);
7240 SDValue N01 = N0.getOperand(0).getOperand(1);
7241 return DAG.getNode(ISD::FMA, dl, VT,
7242 DAG.getNode(ISD::FNEG, dl, VT, N00), N01,
7243 DAG.getNode(ISD::FNEG, dl, VT, N1));
7246 // When FP_EXTEND nodes are free on the target, and there is an opportunity
7247 // to combine into FMA, arrange such nodes accordingly.
7248 if (TLI.isFPExtFree(VT)) {
7250 // fold (fsub (fpext (fmul x, y)), z)
7251 // -> (fma (fpext x), (fpext y), (fneg z))
7252 if (N0.getOpcode() == ISD::FP_EXTEND) {
7253 SDValue N00 = N0.getOperand(0);
7254 if (N00.getOpcode() == ISD::FMUL)
7255 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7256 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7258 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7260 DAG.getNode(ISD::FNEG, SDLoc(N), VT, N1));
7263 // fold (fsub x, (fpext (fmul y, z)))
7264 // -> (fma (fneg (fpext y)), (fpext z), x)
7265 // Note: Commutes FSUB operands.
7266 if (N1.getOpcode() == ISD::FP_EXTEND) {
7267 SDValue N10 = N1.getOperand(0);
7268 if (N10.getOpcode() == ISD::FMUL)
7269 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7270 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7271 DAG.getNode(ISD::FP_EXTEND, SDLoc(N),
7272 VT, N10.getOperand(0))),
7273 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7278 // fold (fsub (fpext (fneg (fmul, x, y))), z)
7279 // -> (fma (fneg (fpext x)), (fpext y), (fneg z))
7280 if (N0.getOpcode() == ISD::FP_EXTEND) {
7281 SDValue N00 = N0.getOperand(0);
7282 if (N00.getOpcode() == ISD::FNEG) {
7283 SDValue N000 = N00.getOperand(0);
7284 if (N000.getOpcode() == ISD::FMUL) {
7285 return DAG.getNode(ISD::FMA, dl, VT,
7286 DAG.getNode(ISD::FNEG, dl, VT,
7287 DAG.getNode(ISD::FP_EXTEND, SDLoc(N),
7288 VT, N000.getOperand(0))),
7289 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7290 N000.getOperand(1)),
7291 DAG.getNode(ISD::FNEG, dl, VT, N1));
7296 // fold (fsub (fneg (fpext (fmul, x, y))), z)
7297 // -> (fma (fneg (fpext x)), (fpext y), (fneg z))
7298 if (N0.getOpcode() == ISD::FNEG) {
7299 SDValue N00 = N0.getOperand(0);
7300 if (N00.getOpcode() == ISD::FP_EXTEND) {
7301 SDValue N000 = N00.getOperand(0);
7302 if (N000.getOpcode() == ISD::FMUL) {
7303 return DAG.getNode(ISD::FMA, dl, VT,
7304 DAG.getNode(ISD::FNEG, dl, VT,
7305 DAG.getNode(ISD::FP_EXTEND, SDLoc(N),
7306 VT, N000.getOperand(0))),
7307 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7308 N000.getOperand(1)),
7309 DAG.getNode(ISD::FNEG, dl, VT, N1));
7315 // More folding opportunities when target permits.
7316 if (TLI.enableAggressiveFMAFusion(VT)) {
7318 // fold (fsub (fma x, y, (fmul u, v)), z)
7319 // -> (fma x, y (fma u, v, (fneg z)))
7320 if (N0.getOpcode() == ISD::FMA &&
7321 N0.getOperand(2).getOpcode() == ISD::FMUL)
7322 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7323 N0.getOperand(0), N0.getOperand(1),
7324 DAG.getNode(ISD::FMA, SDLoc(N), VT,
7325 N0.getOperand(2).getOperand(0),
7326 N0.getOperand(2).getOperand(1),
7327 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7330 // fold (fsub x, (fma y, z, (fmul u, v)))
7331 // -> (fma (fneg y), z, (fma (fneg u), v, x))
7332 if (N1.getOpcode() == ISD::FMA &&
7333 N1.getOperand(2).getOpcode() == ISD::FMUL) {
7334 SDValue N20 = N1.getOperand(2).getOperand(0);
7335 SDValue N21 = N1.getOperand(2).getOperand(1);
7336 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7337 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7340 DAG.getNode(ISD::FMA, SDLoc(N), VT,
7341 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7351 SDValue DAGCombiner::visitFMUL(SDNode *N) {
7352 SDValue N0 = N->getOperand(0);
7353 SDValue N1 = N->getOperand(1);
7354 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
7355 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
7356 EVT VT = N->getValueType(0);
7357 const TargetOptions &Options = DAG.getTarget().Options;
7360 if (VT.isVector()) {
7361 // This just handles C1 * C2 for vectors. Other vector folds are below.
7362 SDValue FoldedVOp = SimplifyVBinOp(N);
7363 if (FoldedVOp.getNode())
7365 // Canonicalize vector constant to RHS.
7366 if (N0.getOpcode() == ISD::BUILD_VECTOR &&
7367 N1.getOpcode() != ISD::BUILD_VECTOR)
7368 if (auto *BV0 = dyn_cast<BuildVectorSDNode>(N0))
7369 if (BV0->isConstant())
7370 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, N1, N0);
7373 // fold (fmul c1, c2) -> c1*c2
7375 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, N1);
7377 // canonicalize constant to RHS
7378 if (N0CFP && !N1CFP)
7379 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, N0);
7381 // fold (fmul A, 1.0) -> A
7382 if (N1CFP && N1CFP->isExactlyValue(1.0))
7385 if (Options.UnsafeFPMath) {
7386 // fold (fmul A, 0) -> 0
7387 if (N1CFP && N1CFP->getValueAPF().isZero())
7390 // fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
7391 if (N0.getOpcode() == ISD::FMUL) {
7392 // Fold scalars or any vector constants (not just splats).
7393 // This fold is done in general by InstCombine, but extra fmul insts
7394 // may have been generated during lowering.
7395 SDValue N01 = N0.getOperand(1);
7396 auto *BV1 = dyn_cast<BuildVectorSDNode>(N1);
7397 auto *BV01 = dyn_cast<BuildVectorSDNode>(N01);
7398 if ((N1CFP && isConstOrConstSplatFP(N01)) ||
7399 (BV1 && BV01 && BV1->isConstant() && BV01->isConstant())) {
7401 SDValue MulConsts = DAG.getNode(ISD::FMUL, SL, VT, N01, N1);
7402 return DAG.getNode(ISD::FMUL, SL, VT, N0.getOperand(0), MulConsts);
7406 // fold (fmul (fadd x, x), c) -> (fmul x, (fmul 2.0, c))
7407 // Undo the fmul 2.0, x -> fadd x, x transformation, since if it occurs
7408 // during an early run of DAGCombiner can prevent folding with fmuls
7409 // inserted during lowering.
7410 if (N0.getOpcode() == ISD::FADD && N0.getOperand(0) == N0.getOperand(1)) {
7412 const SDValue Two = DAG.getConstantFP(2.0, VT);
7413 SDValue MulConsts = DAG.getNode(ISD::FMUL, SL, VT, Two, N1);
7414 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0.getOperand(0), MulConsts);
7418 // fold (fmul X, 2.0) -> (fadd X, X)
7419 if (N1CFP && N1CFP->isExactlyValue(+2.0))
7420 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N0);
7422 // fold (fmul X, -1.0) -> (fneg X)
7423 if (N1CFP && N1CFP->isExactlyValue(-1.0))
7424 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7425 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
7427 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
7428 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
7429 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
7430 // Both can be negated for free, check to see if at least one is cheaper
7432 if (LHSNeg == 2 || RHSNeg == 2)
7433 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7434 GetNegatedExpression(N0, DAG, LegalOperations),
7435 GetNegatedExpression(N1, DAG, LegalOperations));
7442 SDValue DAGCombiner::visitFMA(SDNode *N) {
7443 SDValue N0 = N->getOperand(0);
7444 SDValue N1 = N->getOperand(1);
7445 SDValue N2 = N->getOperand(2);
7446 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7447 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7448 EVT VT = N->getValueType(0);
7450 const TargetOptions &Options = DAG.getTarget().Options;
7452 // Constant fold FMA.
7453 if (isa<ConstantFPSDNode>(N0) &&
7454 isa<ConstantFPSDNode>(N1) &&
7455 isa<ConstantFPSDNode>(N2)) {
7456 return DAG.getNode(ISD::FMA, dl, VT, N0, N1, N2);
7459 if (Options.UnsafeFPMath) {
7460 if (N0CFP && N0CFP->isZero())
7462 if (N1CFP && N1CFP->isZero())
7465 if (N0CFP && N0CFP->isExactlyValue(1.0))
7466 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2);
7467 if (N1CFP && N1CFP->isExactlyValue(1.0))
7468 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2);
7470 // Canonicalize (fma c, x, y) -> (fma x, c, y)
7471 if (N0CFP && !N1CFP)
7472 return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2);
7474 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
7475 if (Options.UnsafeFPMath && N1CFP &&
7476 N2.getOpcode() == ISD::FMUL &&
7477 N0 == N2.getOperand(0) &&
7478 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
7479 return DAG.getNode(ISD::FMUL, dl, VT, N0,
7480 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
7484 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
7485 if (Options.UnsafeFPMath &&
7486 N0.getOpcode() == ISD::FMUL && N1CFP &&
7487 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
7488 return DAG.getNode(ISD::FMA, dl, VT,
7490 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
7494 // (fma x, 1, y) -> (fadd x, y)
7495 // (fma x, -1, y) -> (fadd (fneg x), y)
7497 if (N1CFP->isExactlyValue(1.0))
7498 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
7500 if (N1CFP->isExactlyValue(-1.0) &&
7501 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
7502 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
7503 AddToWorklist(RHSNeg.getNode());
7504 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
7508 // (fma x, c, x) -> (fmul x, (c+1))
7509 if (Options.UnsafeFPMath && N1CFP && N0 == N2)
7510 return DAG.getNode(ISD::FMUL, dl, VT, N0,
7511 DAG.getNode(ISD::FADD, dl, VT,
7512 N1, DAG.getConstantFP(1.0, VT)));
7514 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
7515 if (Options.UnsafeFPMath && N1CFP &&
7516 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0)
7517 return DAG.getNode(ISD::FMUL, dl, VT, N0,
7518 DAG.getNode(ISD::FADD, dl, VT,
7519 N1, DAG.getConstantFP(-1.0, VT)));
7525 SDValue DAGCombiner::visitFDIV(SDNode *N) {
7526 SDValue N0 = N->getOperand(0);
7527 SDValue N1 = N->getOperand(1);
7528 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7529 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7530 EVT VT = N->getValueType(0);
7532 const TargetOptions &Options = DAG.getTarget().Options;
7535 if (VT.isVector()) {
7536 SDValue FoldedVOp = SimplifyVBinOp(N);
7537 if (FoldedVOp.getNode()) return FoldedVOp;
7540 // fold (fdiv c1, c2) -> c1/c2
7542 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1);
7544 if (Options.UnsafeFPMath) {
7545 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
7547 // Compute the reciprocal 1.0 / c2.
7548 APFloat N1APF = N1CFP->getValueAPF();
7549 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
7550 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
7551 // Only do the transform if the reciprocal is a legal fp immediate that
7552 // isn't too nasty (eg NaN, denormal, ...).
7553 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
7554 (!LegalOperations ||
7555 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
7556 // backend)... we should handle this gracefully after Legalize.
7557 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
7558 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
7559 TLI.isFPImmLegal(Recip, VT)))
7560 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0,
7561 DAG.getConstantFP(Recip, VT));
7564 // If this FDIV is part of a reciprocal square root, it may be folded
7565 // into a target-specific square root estimate instruction.
7566 if (N1.getOpcode() == ISD::FSQRT) {
7567 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0))) {
7568 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7570 } else if (N1.getOpcode() == ISD::FP_EXTEND &&
7571 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
7572 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0).getOperand(0))) {
7573 RV = DAG.getNode(ISD::FP_EXTEND, SDLoc(N1), VT, RV);
7574 AddToWorklist(RV.getNode());
7575 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7577 } else if (N1.getOpcode() == ISD::FP_ROUND &&
7578 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
7579 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0).getOperand(0))) {
7580 RV = DAG.getNode(ISD::FP_ROUND, SDLoc(N1), VT, RV, N1.getOperand(1));
7581 AddToWorklist(RV.getNode());
7582 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7584 } else if (N1.getOpcode() == ISD::FMUL) {
7585 // Look through an FMUL. Even though this won't remove the FDIV directly,
7586 // it's still worthwhile to get rid of the FSQRT if possible.
7589 if (N1.getOperand(0).getOpcode() == ISD::FSQRT) {
7590 SqrtOp = N1.getOperand(0);
7591 OtherOp = N1.getOperand(1);
7592 } else if (N1.getOperand(1).getOpcode() == ISD::FSQRT) {
7593 SqrtOp = N1.getOperand(1);
7594 OtherOp = N1.getOperand(0);
7596 if (SqrtOp.getNode()) {
7597 // We found a FSQRT, so try to make this fold:
7598 // x / (y * sqrt(z)) -> x * (rsqrt(z) / y)
7599 if (SDValue RV = BuildRsqrtEstimate(SqrtOp.getOperand(0))) {
7600 RV = DAG.getNode(ISD::FDIV, SDLoc(N1), VT, RV, OtherOp);
7601 AddToWorklist(RV.getNode());
7602 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7607 // Fold into a reciprocal estimate and multiply instead of a real divide.
7608 if (SDValue RV = BuildReciprocalEstimate(N1)) {
7609 AddToWorklist(RV.getNode());
7610 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7614 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
7615 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
7616 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
7617 // Both can be negated for free, check to see if at least one is cheaper
7619 if (LHSNeg == 2 || RHSNeg == 2)
7620 return DAG.getNode(ISD::FDIV, SDLoc(N), VT,
7621 GetNegatedExpression(N0, DAG, LegalOperations),
7622 GetNegatedExpression(N1, DAG, LegalOperations));
7626 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
7628 // E.g., (a / D; b / D;) -> (recip = 1.0 / D; a * recip; b * recip)
7629 // Notice that this is not always beneficial. One reason is different target
7630 // may have different costs for FDIV and FMUL, so sometimes the cost of two
7631 // FDIVs may be lower than the cost of one FDIV and two FMULs. Another reason
7632 // is the critical path is increased from "one FDIV" to "one FDIV + one FMUL".
7633 if (Options.UnsafeFPMath) {
7634 // Skip if current node is a reciprocal.
7635 if (N0CFP && N0CFP->isExactlyValue(1.0))
7638 SmallVector<SDNode *, 4> Users;
7639 // Find all FDIV users of the same divisor.
7640 for (SDNode::use_iterator UI = N1.getNode()->use_begin(),
7641 UE = N1.getNode()->use_end();
7643 SDNode *User = UI.getUse().getUser();
7644 if (User->getOpcode() == ISD::FDIV && User->getOperand(1) == N1)
7645 Users.push_back(User);
7648 if (TLI.combineRepeatedFPDivisors(Users.size())) {
7649 SDValue FPOne = DAG.getConstantFP(1.0, VT); // floating point 1.0
7650 SDValue Reciprocal = DAG.getNode(ISD::FDIV, SDLoc(N), VT, FPOne, N1);
7652 // Dividend / Divisor -> Dividend * Reciprocal
7653 for (auto I = Users.begin(), E = Users.end(); I != E; ++I) {
7654 if ((*I)->getOperand(0) != FPOne) {
7655 SDValue NewNode = DAG.getNode(ISD::FMUL, SDLoc(*I), VT,
7656 (*I)->getOperand(0), Reciprocal);
7657 DAG.ReplaceAllUsesWith(*I, NewNode.getNode());
7667 SDValue DAGCombiner::visitFREM(SDNode *N) {
7668 SDValue N0 = N->getOperand(0);
7669 SDValue N1 = N->getOperand(1);
7670 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7671 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7672 EVT VT = N->getValueType(0);
7674 // fold (frem c1, c2) -> fmod(c1,c2)
7676 return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1);
7681 SDValue DAGCombiner::visitFSQRT(SDNode *N) {
7682 if (DAG.getTarget().Options.UnsafeFPMath &&
7683 !TLI.isFsqrtCheap()) {
7684 // Compute this as X * (1/sqrt(X)) = X * (X ** -0.5)
7685 if (SDValue RV = BuildRsqrtEstimate(N->getOperand(0))) {
7686 EVT VT = RV.getValueType();
7687 RV = DAG.getNode(ISD::FMUL, SDLoc(N), VT, N->getOperand(0), RV);
7688 AddToWorklist(RV.getNode());
7690 // Unfortunately, RV is now NaN if the input was exactly 0.
7691 // Select out this case and force the answer to 0.
7692 SDValue Zero = DAG.getConstantFP(0.0, VT);
7694 DAG.getSetCC(SDLoc(N), TLI.getSetCCResultType(*DAG.getContext(), VT),
7695 N->getOperand(0), Zero, ISD::SETEQ);
7696 AddToWorklist(ZeroCmp.getNode());
7697 AddToWorklist(RV.getNode());
7699 RV = DAG.getNode(VT.isVector() ? ISD::VSELECT : ISD::SELECT,
7700 SDLoc(N), VT, ZeroCmp, Zero, RV);
7707 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
7708 SDValue N0 = N->getOperand(0);
7709 SDValue N1 = N->getOperand(1);
7710 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7711 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7712 EVT VT = N->getValueType(0);
7714 if (N0CFP && N1CFP) // Constant fold
7715 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1);
7718 const APFloat& V = N1CFP->getValueAPF();
7719 // copysign(x, c1) -> fabs(x) iff ispos(c1)
7720 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
7721 if (!V.isNegative()) {
7722 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
7723 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7725 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7726 return DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7727 DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0));
7731 // copysign(fabs(x), y) -> copysign(x, y)
7732 // copysign(fneg(x), y) -> copysign(x, y)
7733 // copysign(copysign(x,z), y) -> copysign(x, y)
7734 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
7735 N0.getOpcode() == ISD::FCOPYSIGN)
7736 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7737 N0.getOperand(0), N1);
7739 // copysign(x, abs(y)) -> abs(x)
7740 if (N1.getOpcode() == ISD::FABS)
7741 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7743 // copysign(x, copysign(y,z)) -> copysign(x, z)
7744 if (N1.getOpcode() == ISD::FCOPYSIGN)
7745 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7746 N0, N1.getOperand(1));
7748 // copysign(x, fp_extend(y)) -> copysign(x, y)
7749 // copysign(x, fp_round(y)) -> copysign(x, y)
7750 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
7751 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7752 N0, N1.getOperand(0));
7757 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
7758 SDValue N0 = N->getOperand(0);
7759 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
7760 EVT VT = N->getValueType(0);
7761 EVT OpVT = N0.getValueType();
7763 // fold (sint_to_fp c1) -> c1fp
7765 // ...but only if the target supports immediate floating-point values
7766 (!LegalOperations ||
7767 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
7768 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
7770 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
7771 // but UINT_TO_FP is legal on this target, try to convert.
7772 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
7773 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
7774 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
7775 if (DAG.SignBitIsZero(N0))
7776 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
7779 // The next optimizations are desirable only if SELECT_CC can be lowered.
7780 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
7781 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7782 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
7784 (!LegalOperations ||
7785 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7787 { N0.getOperand(0), N0.getOperand(1),
7788 DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT),
7790 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7793 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
7794 // (select_cc x, y, 1.0, 0.0,, cc)
7795 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
7796 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
7797 (!LegalOperations ||
7798 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7800 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
7801 DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT),
7802 N0.getOperand(0).getOperand(2) };
7803 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7810 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
7811 SDValue N0 = N->getOperand(0);
7812 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
7813 EVT VT = N->getValueType(0);
7814 EVT OpVT = N0.getValueType();
7816 // fold (uint_to_fp c1) -> c1fp
7818 // ...but only if the target supports immediate floating-point values
7819 (!LegalOperations ||
7820 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
7821 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
7823 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
7824 // but SINT_TO_FP is legal on this target, try to convert.
7825 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
7826 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
7827 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
7828 if (DAG.SignBitIsZero(N0))
7829 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
7832 // The next optimizations are desirable only if SELECT_CC can be lowered.
7833 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
7834 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7836 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
7837 (!LegalOperations ||
7838 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7840 { N0.getOperand(0), N0.getOperand(1),
7841 DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT),
7843 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7850 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
7851 SDValue N0 = N->getOperand(0);
7852 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7853 EVT VT = N->getValueType(0);
7855 // fold (fp_to_sint c1fp) -> c1
7857 return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0);
7862 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
7863 SDValue N0 = N->getOperand(0);
7864 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7865 EVT VT = N->getValueType(0);
7867 // fold (fp_to_uint c1fp) -> c1
7869 return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0);
7874 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
7875 SDValue N0 = N->getOperand(0);
7876 SDValue N1 = N->getOperand(1);
7877 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7878 EVT VT = N->getValueType(0);
7880 // fold (fp_round c1fp) -> c1fp
7882 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1);
7884 // fold (fp_round (fp_extend x)) -> x
7885 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
7886 return N0.getOperand(0);
7888 // fold (fp_round (fp_round x)) -> (fp_round x)
7889 if (N0.getOpcode() == ISD::FP_ROUND) {
7890 // This is a value preserving truncation if both round's are.
7891 bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
7892 N0.getNode()->getConstantOperandVal(1) == 1;
7893 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0.getOperand(0),
7894 DAG.getIntPtrConstant(IsTrunc));
7897 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
7898 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
7899 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT,
7900 N0.getOperand(0), N1);
7901 AddToWorklist(Tmp.getNode());
7902 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7903 Tmp, N0.getOperand(1));
7909 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
7910 SDValue N0 = N->getOperand(0);
7911 EVT VT = N->getValueType(0);
7912 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
7913 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7915 // fold (fp_round_inreg c1fp) -> c1fp
7916 if (N0CFP && isTypeLegal(EVT)) {
7917 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
7918 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, Round);
7924 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
7925 SDValue N0 = N->getOperand(0);
7926 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7927 EVT VT = N->getValueType(0);
7929 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
7930 if (N->hasOneUse() &&
7931 N->use_begin()->getOpcode() == ISD::FP_ROUND)
7934 // fold (fp_extend c1fp) -> c1fp
7936 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0);
7938 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
7940 if (N0.getOpcode() == ISD::FP_ROUND
7941 && N0.getNode()->getConstantOperandVal(1) == 1) {
7942 SDValue In = N0.getOperand(0);
7943 if (In.getValueType() == VT) return In;
7944 if (VT.bitsLT(In.getValueType()))
7945 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT,
7946 In, N0.getOperand(1));
7947 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In);
7950 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
7951 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7952 TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) {
7953 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7954 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
7956 LN0->getBasePtr(), N0.getValueType(),
7957 LN0->getMemOperand());
7958 CombineTo(N, ExtLoad);
7959 CombineTo(N0.getNode(),
7960 DAG.getNode(ISD::FP_ROUND, SDLoc(N0),
7961 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
7962 ExtLoad.getValue(1));
7963 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7969 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
7970 SDValue N0 = N->getOperand(0);
7971 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7972 EVT VT = N->getValueType(0);
7974 // fold (fceil c1) -> fceil(c1)
7976 return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0);
7981 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
7982 SDValue N0 = N->getOperand(0);
7983 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7984 EVT VT = N->getValueType(0);
7986 // fold (ftrunc c1) -> ftrunc(c1)
7988 return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0);
7993 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
7994 SDValue N0 = N->getOperand(0);
7995 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7996 EVT VT = N->getValueType(0);
7998 // fold (ffloor c1) -> ffloor(c1)
8000 return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0);
8005 // FIXME: FNEG and FABS have a lot in common; refactor.
8006 SDValue DAGCombiner::visitFNEG(SDNode *N) {
8007 SDValue N0 = N->getOperand(0);
8008 EVT VT = N->getValueType(0);
8010 if (VT.isVector()) {
8011 SDValue FoldedVOp = SimplifyVUnaryOp(N);
8012 if (FoldedVOp.getNode()) return FoldedVOp;
8015 // Constant fold FNEG.
8016 if (isa<ConstantFPSDNode>(N0))
8017 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N->getOperand(0));
8019 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
8020 &DAG.getTarget().Options))
8021 return GetNegatedExpression(N0, DAG, LegalOperations);
8023 // Transform fneg(bitconvert(x)) -> bitconvert(x ^ sign) to avoid loading
8024 // constant pool values.
8025 if (!TLI.isFNegFree(VT) &&
8026 N0.getOpcode() == ISD::BITCAST &&
8027 N0.getNode()->hasOneUse()) {
8028 SDValue Int = N0.getOperand(0);
8029 EVT IntVT = Int.getValueType();
8030 if (IntVT.isInteger() && !IntVT.isVector()) {
8032 if (N0.getValueType().isVector()) {
8033 // For a vector, get a mask such as 0x80... per scalar element
8035 SignMask = APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
8036 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
8038 // For a scalar, just generate 0x80...
8039 SignMask = APInt::getSignBit(IntVT.getSizeInBits());
8041 Int = DAG.getNode(ISD::XOR, SDLoc(N0), IntVT, Int,
8042 DAG.getConstant(SignMask, IntVT));
8043 AddToWorklist(Int.getNode());
8044 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Int);
8048 // (fneg (fmul c, x)) -> (fmul -c, x)
8049 if (N0.getOpcode() == ISD::FMUL) {
8050 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
8052 APFloat CVal = CFP1->getValueAPF();
8054 if (Level >= AfterLegalizeDAG &&
8055 (TLI.isFPImmLegal(CVal, N->getValueType(0)) ||
8056 TLI.isOperationLegal(ISD::ConstantFP, N->getValueType(0))))
8058 ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
8059 DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0.getOperand(1)));
8066 SDValue DAGCombiner::visitFMINNUM(SDNode *N) {
8067 SDValue N0 = N->getOperand(0);
8068 SDValue N1 = N->getOperand(1);
8069 const ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8070 const ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8072 if (N0CFP && N1CFP) {
8073 const APFloat &C0 = N0CFP->getValueAPF();
8074 const APFloat &C1 = N1CFP->getValueAPF();
8075 return DAG.getConstantFP(minnum(C0, C1), N->getValueType(0));
8079 EVT VT = N->getValueType(0);
8080 // Canonicalize to constant on RHS.
8081 return DAG.getNode(ISD::FMINNUM, SDLoc(N), VT, N1, N0);
8087 SDValue DAGCombiner::visitFMAXNUM(SDNode *N) {
8088 SDValue N0 = N->getOperand(0);
8089 SDValue N1 = N->getOperand(1);
8090 const ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8091 const ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8093 if (N0CFP && N1CFP) {
8094 const APFloat &C0 = N0CFP->getValueAPF();
8095 const APFloat &C1 = N1CFP->getValueAPF();
8096 return DAG.getConstantFP(maxnum(C0, C1), N->getValueType(0));
8100 EVT VT = N->getValueType(0);
8101 // Canonicalize to constant on RHS.
8102 return DAG.getNode(ISD::FMAXNUM, SDLoc(N), VT, N1, N0);
8108 SDValue DAGCombiner::visitFABS(SDNode *N) {
8109 SDValue N0 = N->getOperand(0);
8110 EVT VT = N->getValueType(0);
8112 if (VT.isVector()) {
8113 SDValue FoldedVOp = SimplifyVUnaryOp(N);
8114 if (FoldedVOp.getNode()) return FoldedVOp;
8117 // fold (fabs c1) -> fabs(c1)
8118 if (isa<ConstantFPSDNode>(N0))
8119 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
8121 // fold (fabs (fabs x)) -> (fabs x)
8122 if (N0.getOpcode() == ISD::FABS)
8123 return N->getOperand(0);
8125 // fold (fabs (fneg x)) -> (fabs x)
8126 // fold (fabs (fcopysign x, y)) -> (fabs x)
8127 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
8128 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0));
8130 // Transform fabs(bitconvert(x)) -> bitconvert(x & ~sign) to avoid loading
8131 // constant pool values.
8132 if (!TLI.isFAbsFree(VT) &&
8133 N0.getOpcode() == ISD::BITCAST &&
8134 N0.getNode()->hasOneUse()) {
8135 SDValue Int = N0.getOperand(0);
8136 EVT IntVT = Int.getValueType();
8137 if (IntVT.isInteger() && !IntVT.isVector()) {
8139 if (N0.getValueType().isVector()) {
8140 // For a vector, get a mask such as 0x7f... per scalar element
8142 SignMask = ~APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
8143 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
8145 // For a scalar, just generate 0x7f...
8146 SignMask = ~APInt::getSignBit(IntVT.getSizeInBits());
8148 Int = DAG.getNode(ISD::AND, SDLoc(N0), IntVT, Int,
8149 DAG.getConstant(SignMask, IntVT));
8150 AddToWorklist(Int.getNode());
8151 return DAG.getNode(ISD::BITCAST, SDLoc(N), N->getValueType(0), Int);
8158 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
8159 SDValue Chain = N->getOperand(0);
8160 SDValue N1 = N->getOperand(1);
8161 SDValue N2 = N->getOperand(2);
8163 // If N is a constant we could fold this into a fallthrough or unconditional
8164 // branch. However that doesn't happen very often in normal code, because
8165 // Instcombine/SimplifyCFG should have handled the available opportunities.
8166 // If we did this folding here, it would be necessary to update the
8167 // MachineBasicBlock CFG, which is awkward.
8169 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
8171 if (N1.getOpcode() == ISD::SETCC &&
8172 TLI.isOperationLegalOrCustom(ISD::BR_CC,
8173 N1.getOperand(0).getValueType())) {
8174 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
8175 Chain, N1.getOperand(2),
8176 N1.getOperand(0), N1.getOperand(1), N2);
8179 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
8180 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
8181 (N1.getOperand(0).hasOneUse() &&
8182 N1.getOperand(0).getOpcode() == ISD::SRL))) {
8183 SDNode *Trunc = nullptr;
8184 if (N1.getOpcode() == ISD::TRUNCATE) {
8185 // Look pass the truncate.
8186 Trunc = N1.getNode();
8187 N1 = N1.getOperand(0);
8190 // Match this pattern so that we can generate simpler code:
8193 // %b = and i32 %a, 2
8194 // %c = srl i32 %b, 1
8195 // brcond i32 %c ...
8200 // %b = and i32 %a, 2
8201 // %c = setcc eq %b, 0
8204 // This applies only when the AND constant value has one bit set and the
8205 // SRL constant is equal to the log2 of the AND constant. The back-end is
8206 // smart enough to convert the result into a TEST/JMP sequence.
8207 SDValue Op0 = N1.getOperand(0);
8208 SDValue Op1 = N1.getOperand(1);
8210 if (Op0.getOpcode() == ISD::AND &&
8211 Op1.getOpcode() == ISD::Constant) {
8212 SDValue AndOp1 = Op0.getOperand(1);
8214 if (AndOp1.getOpcode() == ISD::Constant) {
8215 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
8217 if (AndConst.isPowerOf2() &&
8218 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
8220 DAG.getSetCC(SDLoc(N),
8221 getSetCCResultType(Op0.getValueType()),
8222 Op0, DAG.getConstant(0, Op0.getValueType()),
8225 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, SDLoc(N),
8226 MVT::Other, Chain, SetCC, N2);
8227 // Don't add the new BRCond into the worklist or else SimplifySelectCC
8228 // will convert it back to (X & C1) >> C2.
8229 CombineTo(N, NewBRCond, false);
8230 // Truncate is dead.
8232 deleteAndRecombine(Trunc);
8233 // Replace the uses of SRL with SETCC
8234 WorklistRemover DeadNodes(*this);
8235 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
8236 deleteAndRecombine(N1.getNode());
8237 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8243 // Restore N1 if the above transformation doesn't match.
8244 N1 = N->getOperand(1);
8247 // Transform br(xor(x, y)) -> br(x != y)
8248 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
8249 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
8250 SDNode *TheXor = N1.getNode();
8251 SDValue Op0 = TheXor->getOperand(0);
8252 SDValue Op1 = TheXor->getOperand(1);
8253 if (Op0.getOpcode() == Op1.getOpcode()) {
8254 // Avoid missing important xor optimizations.
8255 SDValue Tmp = visitXOR(TheXor);
8256 if (Tmp.getNode()) {
8257 if (Tmp.getNode() != TheXor) {
8258 DEBUG(dbgs() << "\nReplacing.8 ";
8260 dbgs() << "\nWith: ";
8261 Tmp.getNode()->dump(&DAG);
8263 WorklistRemover DeadNodes(*this);
8264 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
8265 deleteAndRecombine(TheXor);
8266 return DAG.getNode(ISD::BRCOND, SDLoc(N),
8267 MVT::Other, Chain, Tmp, N2);
8270 // visitXOR has changed XOR's operands or replaced the XOR completely,
8272 return SDValue(N, 0);
8276 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
8278 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
8279 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
8280 Op0.getOpcode() == ISD::XOR) {
8281 TheXor = Op0.getNode();
8285 EVT SetCCVT = N1.getValueType();
8287 SetCCVT = getSetCCResultType(SetCCVT);
8288 SDValue SetCC = DAG.getSetCC(SDLoc(TheXor),
8291 Equal ? ISD::SETEQ : ISD::SETNE);
8292 // Replace the uses of XOR with SETCC
8293 WorklistRemover DeadNodes(*this);
8294 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
8295 deleteAndRecombine(N1.getNode());
8296 return DAG.getNode(ISD::BRCOND, SDLoc(N),
8297 MVT::Other, Chain, SetCC, N2);
8304 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
8306 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
8307 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
8308 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
8310 // If N is a constant we could fold this into a fallthrough or unconditional
8311 // branch. However that doesn't happen very often in normal code, because
8312 // Instcombine/SimplifyCFG should have handled the available opportunities.
8313 // If we did this folding here, it would be necessary to update the
8314 // MachineBasicBlock CFG, which is awkward.
8316 // Use SimplifySetCC to simplify SETCC's.
8317 SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()),
8318 CondLHS, CondRHS, CC->get(), SDLoc(N),
8320 if (Simp.getNode()) AddToWorklist(Simp.getNode());
8322 // fold to a simpler setcc
8323 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
8324 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
8325 N->getOperand(0), Simp.getOperand(2),
8326 Simp.getOperand(0), Simp.getOperand(1),
8332 /// Return true if 'Use' is a load or a store that uses N as its base pointer
8333 /// and that N may be folded in the load / store addressing mode.
8334 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
8336 const TargetLowering &TLI) {
8338 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
8339 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
8341 VT = Use->getValueType(0);
8342 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
8343 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
8345 VT = ST->getValue().getValueType();
8349 TargetLowering::AddrMode AM;
8350 if (N->getOpcode() == ISD::ADD) {
8351 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
8354 AM.BaseOffs = Offset->getSExtValue();
8358 } else if (N->getOpcode() == ISD::SUB) {
8359 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
8362 AM.BaseOffs = -Offset->getSExtValue();
8369 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
8372 /// Try turning a load/store into a pre-indexed load/store when the base
8373 /// pointer is an add or subtract and it has other uses besides the load/store.
8374 /// After the transformation, the new indexed load/store has effectively folded
8375 /// the add/subtract in and all of its other uses are redirected to the
8377 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
8378 if (Level < AfterLegalizeDAG)
8384 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8385 if (LD->isIndexed())
8387 VT = LD->getMemoryVT();
8388 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
8389 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
8391 Ptr = LD->getBasePtr();
8392 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8393 if (ST->isIndexed())
8395 VT = ST->getMemoryVT();
8396 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
8397 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
8399 Ptr = ST->getBasePtr();
8405 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
8406 // out. There is no reason to make this a preinc/predec.
8407 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
8408 Ptr.getNode()->hasOneUse())
8411 // Ask the target to do addressing mode selection.
8414 ISD::MemIndexedMode AM = ISD::UNINDEXED;
8415 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
8418 // Backends without true r+i pre-indexed forms may need to pass a
8419 // constant base with a variable offset so that constant coercion
8420 // will work with the patterns in canonical form.
8421 bool Swapped = false;
8422 if (isa<ConstantSDNode>(BasePtr)) {
8423 std::swap(BasePtr, Offset);
8427 // Don't create a indexed load / store with zero offset.
8428 if (isa<ConstantSDNode>(Offset) &&
8429 cast<ConstantSDNode>(Offset)->isNullValue())
8432 // Try turning it into a pre-indexed load / store except when:
8433 // 1) The new base ptr is a frame index.
8434 // 2) If N is a store and the new base ptr is either the same as or is a
8435 // predecessor of the value being stored.
8436 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
8437 // that would create a cycle.
8438 // 4) All uses are load / store ops that use it as old base ptr.
8440 // Check #1. Preinc'ing a frame index would require copying the stack pointer
8441 // (plus the implicit offset) to a register to preinc anyway.
8442 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
8447 SDValue Val = cast<StoreSDNode>(N)->getValue();
8448 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
8452 // If the offset is a constant, there may be other adds of constants that
8453 // can be folded with this one. We should do this to avoid having to keep
8454 // a copy of the original base pointer.
8455 SmallVector<SDNode *, 16> OtherUses;
8456 if (isa<ConstantSDNode>(Offset))
8457 for (SDNode *Use : BasePtr.getNode()->uses()) {
8458 if (Use == Ptr.getNode())
8461 if (Use->isPredecessorOf(N))
8464 if (Use->getOpcode() != ISD::ADD && Use->getOpcode() != ISD::SUB) {
8469 SDValue Op0 = Use->getOperand(0), Op1 = Use->getOperand(1);
8470 if (Op1.getNode() == BasePtr.getNode())
8471 std::swap(Op0, Op1);
8472 assert(Op0.getNode() == BasePtr.getNode() &&
8473 "Use of ADD/SUB but not an operand");
8475 if (!isa<ConstantSDNode>(Op1)) {
8480 // FIXME: In some cases, we can be smarter about this.
8481 if (Op1.getValueType() != Offset.getValueType()) {
8486 OtherUses.push_back(Use);
8490 std::swap(BasePtr, Offset);
8492 // Now check for #3 and #4.
8493 bool RealUse = false;
8495 // Caches for hasPredecessorHelper
8496 SmallPtrSet<const SDNode *, 32> Visited;
8497 SmallVector<const SDNode *, 16> Worklist;
8499 for (SDNode *Use : Ptr.getNode()->uses()) {
8502 if (N->hasPredecessorHelper(Use, Visited, Worklist))
8505 // If Ptr may be folded in addressing mode of other use, then it's
8506 // not profitable to do this transformation.
8507 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
8516 Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
8517 BasePtr, Offset, AM);
8519 Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
8520 BasePtr, Offset, AM);
8523 DEBUG(dbgs() << "\nReplacing.4 ";
8525 dbgs() << "\nWith: ";
8526 Result.getNode()->dump(&DAG);
8528 WorklistRemover DeadNodes(*this);
8530 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
8531 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
8533 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
8536 // Finally, since the node is now dead, remove it from the graph.
8537 deleteAndRecombine(N);
8540 std::swap(BasePtr, Offset);
8542 // Replace other uses of BasePtr that can be updated to use Ptr
8543 for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) {
8544 unsigned OffsetIdx = 1;
8545 if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode())
8547 assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() ==
8548 BasePtr.getNode() && "Expected BasePtr operand");
8550 // We need to replace ptr0 in the following expression:
8551 // x0 * offset0 + y0 * ptr0 = t0
8553 // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store)
8555 // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the
8556 // indexed load/store and the expresion that needs to be re-written.
8558 // Therefore, we have:
8559 // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1
8561 ConstantSDNode *CN =
8562 cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx));
8564 APInt Offset0 = CN->getAPIntValue();
8565 APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue();
8567 X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1;
8568 Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1;
8569 X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1;
8570 Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1;
8572 unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD;
8574 APInt CNV = Offset0;
8575 if (X0 < 0) CNV = -CNV;
8576 if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1;
8577 else CNV = CNV - Offset1;
8579 // We can now generate the new expression.
8580 SDValue NewOp1 = DAG.getConstant(CNV, CN->getValueType(0));
8581 SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0);
8583 SDValue NewUse = DAG.getNode(Opcode,
8584 SDLoc(OtherUses[i]),
8585 OtherUses[i]->getValueType(0), NewOp1, NewOp2);
8586 DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse);
8587 deleteAndRecombine(OtherUses[i]);
8590 // Replace the uses of Ptr with uses of the updated base value.
8591 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
8592 deleteAndRecombine(Ptr.getNode());
8597 /// Try to combine a load/store with a add/sub of the base pointer node into a
8598 /// post-indexed load/store. The transformation folded the add/subtract into the
8599 /// new indexed load/store effectively and all of its uses are redirected to the
8601 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
8602 if (Level < AfterLegalizeDAG)
8608 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8609 if (LD->isIndexed())
8611 VT = LD->getMemoryVT();
8612 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
8613 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
8615 Ptr = LD->getBasePtr();
8616 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8617 if (ST->isIndexed())
8619 VT = ST->getMemoryVT();
8620 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
8621 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
8623 Ptr = ST->getBasePtr();
8629 if (Ptr.getNode()->hasOneUse())
8632 for (SDNode *Op : Ptr.getNode()->uses()) {
8634 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
8639 ISD::MemIndexedMode AM = ISD::UNINDEXED;
8640 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
8641 // Don't create a indexed load / store with zero offset.
8642 if (isa<ConstantSDNode>(Offset) &&
8643 cast<ConstantSDNode>(Offset)->isNullValue())
8646 // Try turning it into a post-indexed load / store except when
8647 // 1) All uses are load / store ops that use it as base ptr (and
8648 // it may be folded as addressing mmode).
8649 // 2) Op must be independent of N, i.e. Op is neither a predecessor
8650 // nor a successor of N. Otherwise, if Op is folded that would
8653 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
8657 bool TryNext = false;
8658 for (SDNode *Use : BasePtr.getNode()->uses()) {
8659 if (Use == Ptr.getNode())
8662 // If all the uses are load / store addresses, then don't do the
8664 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
8665 bool RealUse = false;
8666 for (SDNode *UseUse : Use->uses()) {
8667 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
8682 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
8683 SDValue Result = isLoad
8684 ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
8685 BasePtr, Offset, AM)
8686 : DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
8687 BasePtr, Offset, AM);
8690 DEBUG(dbgs() << "\nReplacing.5 ";
8692 dbgs() << "\nWith: ";
8693 Result.getNode()->dump(&DAG);
8695 WorklistRemover DeadNodes(*this);
8697 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
8698 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
8700 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
8703 // Finally, since the node is now dead, remove it from the graph.
8704 deleteAndRecombine(N);
8706 // Replace the uses of Use with uses of the updated base value.
8707 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
8708 Result.getValue(isLoad ? 1 : 0));
8709 deleteAndRecombine(Op);
8718 /// \brief Return the base-pointer arithmetic from an indexed \p LD.
8719 SDValue DAGCombiner::SplitIndexingFromLoad(LoadSDNode *LD) {
8720 ISD::MemIndexedMode AM = LD->getAddressingMode();
8721 assert(AM != ISD::UNINDEXED);
8722 SDValue BP = LD->getOperand(1);
8723 SDValue Inc = LD->getOperand(2);
8725 // Some backends use TargetConstants for load offsets, but don't expect
8726 // TargetConstants in general ADD nodes. We can convert these constants into
8727 // regular Constants (if the constant is not opaque).
8728 assert((Inc.getOpcode() != ISD::TargetConstant ||
8729 !cast<ConstantSDNode>(Inc)->isOpaque()) &&
8730 "Cannot split out indexing using opaque target constants");
8731 if (Inc.getOpcode() == ISD::TargetConstant) {
8732 ConstantSDNode *ConstInc = cast<ConstantSDNode>(Inc);
8733 Inc = DAG.getConstant(*ConstInc->getConstantIntValue(),
8734 ConstInc->getValueType(0));
8738 (AM == ISD::PRE_INC || AM == ISD::POST_INC ? ISD::ADD : ISD::SUB);
8739 return DAG.getNode(Opc, SDLoc(LD), BP.getSimpleValueType(), BP, Inc);
8742 SDValue DAGCombiner::visitLOAD(SDNode *N) {
8743 LoadSDNode *LD = cast<LoadSDNode>(N);
8744 SDValue Chain = LD->getChain();
8745 SDValue Ptr = LD->getBasePtr();
8747 // If load is not volatile and there are no uses of the loaded value (and
8748 // the updated indexed value in case of indexed loads), change uses of the
8749 // chain value into uses of the chain input (i.e. delete the dead load).
8750 if (!LD->isVolatile()) {
8751 if (N->getValueType(1) == MVT::Other) {
8753 if (!N->hasAnyUseOfValue(0)) {
8754 // It's not safe to use the two value CombineTo variant here. e.g.
8755 // v1, chain2 = load chain1, loc
8756 // v2, chain3 = load chain2, loc
8758 // Now we replace use of chain2 with chain1. This makes the second load
8759 // isomorphic to the one we are deleting, and thus makes this load live.
8760 DEBUG(dbgs() << "\nReplacing.6 ";
8762 dbgs() << "\nWith chain: ";
8763 Chain.getNode()->dump(&DAG);
8765 WorklistRemover DeadNodes(*this);
8766 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
8769 deleteAndRecombine(N);
8771 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8775 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
8777 // If this load has an opaque TargetConstant offset, then we cannot split
8778 // the indexing into an add/sub directly (that TargetConstant may not be
8779 // valid for a different type of node, and we cannot convert an opaque
8780 // target constant into a regular constant).
8781 bool HasOTCInc = LD->getOperand(2).getOpcode() == ISD::TargetConstant &&
8782 cast<ConstantSDNode>(LD->getOperand(2))->isOpaque();
8784 if (!N->hasAnyUseOfValue(0) &&
8785 ((MaySplitLoadIndex && !HasOTCInc) || !N->hasAnyUseOfValue(1))) {
8786 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
8788 if (N->hasAnyUseOfValue(1) && MaySplitLoadIndex && !HasOTCInc) {
8789 Index = SplitIndexingFromLoad(LD);
8790 // Try to fold the base pointer arithmetic into subsequent loads and
8792 AddUsersToWorklist(N);
8794 Index = DAG.getUNDEF(N->getValueType(1));
8795 DEBUG(dbgs() << "\nReplacing.7 ";
8797 dbgs() << "\nWith: ";
8798 Undef.getNode()->dump(&DAG);
8799 dbgs() << " and 2 other values\n");
8800 WorklistRemover DeadNodes(*this);
8801 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
8802 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Index);
8803 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
8804 deleteAndRecombine(N);
8805 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8810 // If this load is directly stored, replace the load value with the stored
8812 // TODO: Handle store large -> read small portion.
8813 // TODO: Handle TRUNCSTORE/LOADEXT
8814 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
8815 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
8816 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
8817 if (PrevST->getBasePtr() == Ptr &&
8818 PrevST->getValue().getValueType() == N->getValueType(0))
8819 return CombineTo(N, Chain.getOperand(1), Chain);
8823 // Try to infer better alignment information than the load already has.
8824 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
8825 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
8826 if (Align > LD->getMemOperand()->getBaseAlignment()) {
8828 DAG.getExtLoad(LD->getExtensionType(), SDLoc(N),
8829 LD->getValueType(0),
8830 Chain, Ptr, LD->getPointerInfo(),
8832 LD->isVolatile(), LD->isNonTemporal(),
8833 LD->isInvariant(), Align, LD->getAAInfo());
8834 return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true);
8839 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
8840 : DAG.getSubtarget().useAA();
8842 if (CombinerAAOnlyFunc.getNumOccurrences() &&
8843 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
8846 if (UseAA && LD->isUnindexed()) {
8847 // Walk up chain skipping non-aliasing memory nodes.
8848 SDValue BetterChain = FindBetterChain(N, Chain);
8850 // If there is a better chain.
8851 if (Chain != BetterChain) {
8854 // Replace the chain to void dependency.
8855 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
8856 ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD),
8857 BetterChain, Ptr, LD->getMemOperand());
8859 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD),
8860 LD->getValueType(0),
8861 BetterChain, Ptr, LD->getMemoryVT(),
8862 LD->getMemOperand());
8865 // Create token factor to keep old chain connected.
8866 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
8867 MVT::Other, Chain, ReplLoad.getValue(1));
8869 // Make sure the new and old chains are cleaned up.
8870 AddToWorklist(Token.getNode());
8872 // Replace uses with load result and token factor. Don't add users
8874 return CombineTo(N, ReplLoad.getValue(0), Token, false);
8878 // Try transforming N to an indexed load.
8879 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
8880 return SDValue(N, 0);
8882 // Try to slice up N to more direct loads if the slices are mapped to
8883 // different register banks or pairing can take place.
8885 return SDValue(N, 0);
8891 /// \brief Helper structure used to slice a load in smaller loads.
8892 /// Basically a slice is obtained from the following sequence:
8893 /// Origin = load Ty1, Base
8894 /// Shift = srl Ty1 Origin, CstTy Amount
8895 /// Inst = trunc Shift to Ty2
8897 /// Then, it will be rewriten into:
8898 /// Slice = load SliceTy, Base + SliceOffset
8899 /// [Inst = zext Slice to Ty2], only if SliceTy <> Ty2
8901 /// SliceTy is deduced from the number of bits that are actually used to
8903 struct LoadedSlice {
8904 /// \brief Helper structure used to compute the cost of a slice.
8906 /// Are we optimizing for code size.
8911 unsigned CrossRegisterBanksCopies;
8915 Cost(bool ForCodeSize = false)
8916 : ForCodeSize(ForCodeSize), Loads(0), Truncates(0),
8917 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {}
8919 /// \brief Get the cost of one isolated slice.
8920 Cost(const LoadedSlice &LS, bool ForCodeSize = false)
8921 : ForCodeSize(ForCodeSize), Loads(1), Truncates(0),
8922 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {
8923 EVT TruncType = LS.Inst->getValueType(0);
8924 EVT LoadedType = LS.getLoadedType();
8925 if (TruncType != LoadedType &&
8926 !LS.DAG->getTargetLoweringInfo().isZExtFree(LoadedType, TruncType))
8930 /// \brief Account for slicing gain in the current cost.
8931 /// Slicing provide a few gains like removing a shift or a
8932 /// truncate. This method allows to grow the cost of the original
8933 /// load with the gain from this slice.
8934 void addSliceGain(const LoadedSlice &LS) {
8935 // Each slice saves a truncate.
8936 const TargetLowering &TLI = LS.DAG->getTargetLoweringInfo();
8937 if (!TLI.isTruncateFree(LS.Inst->getValueType(0),
8938 LS.Inst->getOperand(0).getValueType()))
8940 // If there is a shift amount, this slice gets rid of it.
8943 // If this slice can merge a cross register bank copy, account for it.
8944 if (LS.canMergeExpensiveCrossRegisterBankCopy())
8945 ++CrossRegisterBanksCopies;
8948 Cost &operator+=(const Cost &RHS) {
8950 Truncates += RHS.Truncates;
8951 CrossRegisterBanksCopies += RHS.CrossRegisterBanksCopies;
8957 bool operator==(const Cost &RHS) const {
8958 return Loads == RHS.Loads && Truncates == RHS.Truncates &&
8959 CrossRegisterBanksCopies == RHS.CrossRegisterBanksCopies &&
8960 ZExts == RHS.ZExts && Shift == RHS.Shift;
8963 bool operator!=(const Cost &RHS) const { return !(*this == RHS); }
8965 bool operator<(const Cost &RHS) const {
8966 // Assume cross register banks copies are as expensive as loads.
8967 // FIXME: Do we want some more target hooks?
8968 unsigned ExpensiveOpsLHS = Loads + CrossRegisterBanksCopies;
8969 unsigned ExpensiveOpsRHS = RHS.Loads + RHS.CrossRegisterBanksCopies;
8970 // Unless we are optimizing for code size, consider the
8971 // expensive operation first.
8972 if (!ForCodeSize && ExpensiveOpsLHS != ExpensiveOpsRHS)
8973 return ExpensiveOpsLHS < ExpensiveOpsRHS;
8974 return (Truncates + ZExts + Shift + ExpensiveOpsLHS) <
8975 (RHS.Truncates + RHS.ZExts + RHS.Shift + ExpensiveOpsRHS);
8978 bool operator>(const Cost &RHS) const { return RHS < *this; }
8980 bool operator<=(const Cost &RHS) const { return !(RHS < *this); }
8982 bool operator>=(const Cost &RHS) const { return !(*this < RHS); }
8984 // The last instruction that represent the slice. This should be a
8985 // truncate instruction.
8987 // The original load instruction.
8989 // The right shift amount in bits from the original load.
8991 // The DAG from which Origin came from.
8992 // This is used to get some contextual information about legal types, etc.
8995 LoadedSlice(SDNode *Inst = nullptr, LoadSDNode *Origin = nullptr,
8996 unsigned Shift = 0, SelectionDAG *DAG = nullptr)
8997 : Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {}
8999 LoadedSlice(const LoadedSlice &LS)
9000 : Inst(LS.Inst), Origin(LS.Origin), Shift(LS.Shift), DAG(LS.DAG) {}
9002 /// \brief Get the bits used in a chunk of bits \p BitWidth large.
9003 /// \return Result is \p BitWidth and has used bits set to 1 and
9004 /// not used bits set to 0.
9005 APInt getUsedBits() const {
9006 // Reproduce the trunc(lshr) sequence:
9007 // - Start from the truncated value.
9008 // - Zero extend to the desired bit width.
9010 assert(Origin && "No original load to compare against.");
9011 unsigned BitWidth = Origin->getValueSizeInBits(0);
9012 assert(Inst && "This slice is not bound to an instruction");
9013 assert(Inst->getValueSizeInBits(0) <= BitWidth &&
9014 "Extracted slice is bigger than the whole type!");
9015 APInt UsedBits(Inst->getValueSizeInBits(0), 0);
9016 UsedBits.setAllBits();
9017 UsedBits = UsedBits.zext(BitWidth);
9022 /// \brief Get the size of the slice to be loaded in bytes.
9023 unsigned getLoadedSize() const {
9024 unsigned SliceSize = getUsedBits().countPopulation();
9025 assert(!(SliceSize & 0x7) && "Size is not a multiple of a byte.");
9026 return SliceSize / 8;
9029 /// \brief Get the type that will be loaded for this slice.
9030 /// Note: This may not be the final type for the slice.
9031 EVT getLoadedType() const {
9032 assert(DAG && "Missing context");
9033 LLVMContext &Ctxt = *DAG->getContext();
9034 return EVT::getIntegerVT(Ctxt, getLoadedSize() * 8);
9037 /// \brief Get the alignment of the load used for this slice.
9038 unsigned getAlignment() const {
9039 unsigned Alignment = Origin->getAlignment();
9040 unsigned Offset = getOffsetFromBase();
9042 Alignment = MinAlign(Alignment, Alignment + Offset);
9046 /// \brief Check if this slice can be rewritten with legal operations.
9047 bool isLegal() const {
9048 // An invalid slice is not legal.
9049 if (!Origin || !Inst || !DAG)
9052 // Offsets are for indexed load only, we do not handle that.
9053 if (Origin->getOffset().getOpcode() != ISD::UNDEF)
9056 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
9058 // Check that the type is legal.
9059 EVT SliceType = getLoadedType();
9060 if (!TLI.isTypeLegal(SliceType))
9063 // Check that the load is legal for this type.
9064 if (!TLI.isOperationLegal(ISD::LOAD, SliceType))
9067 // Check that the offset can be computed.
9068 // 1. Check its type.
9069 EVT PtrType = Origin->getBasePtr().getValueType();
9070 if (PtrType == MVT::Untyped || PtrType.isExtended())
9073 // 2. Check that it fits in the immediate.
9074 if (!TLI.isLegalAddImmediate(getOffsetFromBase()))
9077 // 3. Check that the computation is legal.
9078 if (!TLI.isOperationLegal(ISD::ADD, PtrType))
9081 // Check that the zext is legal if it needs one.
9082 EVT TruncateType = Inst->getValueType(0);
9083 if (TruncateType != SliceType &&
9084 !TLI.isOperationLegal(ISD::ZERO_EXTEND, TruncateType))
9090 /// \brief Get the offset in bytes of this slice in the original chunk of
9092 /// \pre DAG != nullptr.
9093 uint64_t getOffsetFromBase() const {
9094 assert(DAG && "Missing context.");
9096 DAG->getTargetLoweringInfo().getDataLayout()->isBigEndian();
9097 assert(!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported.");
9098 uint64_t Offset = Shift / 8;
9099 unsigned TySizeInBytes = Origin->getValueSizeInBits(0) / 8;
9100 assert(!(Origin->getValueSizeInBits(0) & 0x7) &&
9101 "The size of the original loaded type is not a multiple of a"
9103 // If Offset is bigger than TySizeInBytes, it means we are loading all
9104 // zeros. This should have been optimized before in the process.
9105 assert(TySizeInBytes > Offset &&
9106 "Invalid shift amount for given loaded size");
9108 Offset = TySizeInBytes - Offset - getLoadedSize();
9112 /// \brief Generate the sequence of instructions to load the slice
9113 /// represented by this object and redirect the uses of this slice to
9114 /// this new sequence of instructions.
9115 /// \pre this->Inst && this->Origin are valid Instructions and this
9116 /// object passed the legal check: LoadedSlice::isLegal returned true.
9117 /// \return The last instruction of the sequence used to load the slice.
9118 SDValue loadSlice() const {
9119 assert(Inst && Origin && "Unable to replace a non-existing slice.");
9120 const SDValue &OldBaseAddr = Origin->getBasePtr();
9121 SDValue BaseAddr = OldBaseAddr;
9122 // Get the offset in that chunk of bytes w.r.t. the endianess.
9123 int64_t Offset = static_cast<int64_t>(getOffsetFromBase());
9124 assert(Offset >= 0 && "Offset too big to fit in int64_t!");
9126 // BaseAddr = BaseAddr + Offset.
9127 EVT ArithType = BaseAddr.getValueType();
9128 BaseAddr = DAG->getNode(ISD::ADD, SDLoc(Origin), ArithType, BaseAddr,
9129 DAG->getConstant(Offset, ArithType));
9132 // Create the type of the loaded slice according to its size.
9133 EVT SliceType = getLoadedType();
9135 // Create the load for the slice.
9136 SDValue LastInst = DAG->getLoad(
9137 SliceType, SDLoc(Origin), Origin->getChain(), BaseAddr,
9138 Origin->getPointerInfo().getWithOffset(Offset), Origin->isVolatile(),
9139 Origin->isNonTemporal(), Origin->isInvariant(), getAlignment());
9140 // If the final type is not the same as the loaded type, this means that
9141 // we have to pad with zero. Create a zero extend for that.
9142 EVT FinalType = Inst->getValueType(0);
9143 if (SliceType != FinalType)
9145 DAG->getNode(ISD::ZERO_EXTEND, SDLoc(LastInst), FinalType, LastInst);
9149 /// \brief Check if this slice can be merged with an expensive cross register
9150 /// bank copy. E.g.,
9152 /// f = bitcast i32 i to float
9153 bool canMergeExpensiveCrossRegisterBankCopy() const {
9154 if (!Inst || !Inst->hasOneUse())
9156 SDNode *Use = *Inst->use_begin();
9157 if (Use->getOpcode() != ISD::BITCAST)
9159 assert(DAG && "Missing context");
9160 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
9161 EVT ResVT = Use->getValueType(0);
9162 const TargetRegisterClass *ResRC = TLI.getRegClassFor(ResVT.getSimpleVT());
9163 const TargetRegisterClass *ArgRC =
9164 TLI.getRegClassFor(Use->getOperand(0).getValueType().getSimpleVT());
9165 if (ArgRC == ResRC || !TLI.isOperationLegal(ISD::LOAD, ResVT))
9168 // At this point, we know that we perform a cross-register-bank copy.
9169 // Check if it is expensive.
9170 const TargetRegisterInfo *TRI = DAG->getSubtarget().getRegisterInfo();
9171 // Assume bitcasts are cheap, unless both register classes do not
9172 // explicitly share a common sub class.
9173 if (!TRI || TRI->getCommonSubClass(ArgRC, ResRC))
9176 // Check if it will be merged with the load.
9177 // 1. Check the alignment constraint.
9178 unsigned RequiredAlignment = TLI.getDataLayout()->getABITypeAlignment(
9179 ResVT.getTypeForEVT(*DAG->getContext()));
9181 if (RequiredAlignment > getAlignment())
9184 // 2. Check that the load is a legal operation for that type.
9185 if (!TLI.isOperationLegal(ISD::LOAD, ResVT))
9188 // 3. Check that we do not have a zext in the way.
9189 if (Inst->getValueType(0) != getLoadedType())
9197 /// \brief Check that all bits set in \p UsedBits form a dense region, i.e.,
9198 /// \p UsedBits looks like 0..0 1..1 0..0.
9199 static bool areUsedBitsDense(const APInt &UsedBits) {
9200 // If all the bits are one, this is dense!
9201 if (UsedBits.isAllOnesValue())
9204 // Get rid of the unused bits on the right.
9205 APInt NarrowedUsedBits = UsedBits.lshr(UsedBits.countTrailingZeros());
9206 // Get rid of the unused bits on the left.
9207 if (NarrowedUsedBits.countLeadingZeros())
9208 NarrowedUsedBits = NarrowedUsedBits.trunc(NarrowedUsedBits.getActiveBits());
9209 // Check that the chunk of bits is completely used.
9210 return NarrowedUsedBits.isAllOnesValue();
9213 /// \brief Check whether or not \p First and \p Second are next to each other
9214 /// in memory. This means that there is no hole between the bits loaded
9215 /// by \p First and the bits loaded by \p Second.
9216 static bool areSlicesNextToEachOther(const LoadedSlice &First,
9217 const LoadedSlice &Second) {
9218 assert(First.Origin == Second.Origin && First.Origin &&
9219 "Unable to match different memory origins.");
9220 APInt UsedBits = First.getUsedBits();
9221 assert((UsedBits & Second.getUsedBits()) == 0 &&
9222 "Slices are not supposed to overlap.");
9223 UsedBits |= Second.getUsedBits();
9224 return areUsedBitsDense(UsedBits);
9227 /// \brief Adjust the \p GlobalLSCost according to the target
9228 /// paring capabilities and the layout of the slices.
9229 /// \pre \p GlobalLSCost should account for at least as many loads as
9230 /// there is in the slices in \p LoadedSlices.
9231 static void adjustCostForPairing(SmallVectorImpl<LoadedSlice> &LoadedSlices,
9232 LoadedSlice::Cost &GlobalLSCost) {
9233 unsigned NumberOfSlices = LoadedSlices.size();
9234 // If there is less than 2 elements, no pairing is possible.
9235 if (NumberOfSlices < 2)
9238 // Sort the slices so that elements that are likely to be next to each
9239 // other in memory are next to each other in the list.
9240 std::sort(LoadedSlices.begin(), LoadedSlices.end(),
9241 [](const LoadedSlice &LHS, const LoadedSlice &RHS) {
9242 assert(LHS.Origin == RHS.Origin && "Different bases not implemented.");
9243 return LHS.getOffsetFromBase() < RHS.getOffsetFromBase();
9245 const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo();
9246 // First (resp. Second) is the first (resp. Second) potentially candidate
9247 // to be placed in a paired load.
9248 const LoadedSlice *First = nullptr;
9249 const LoadedSlice *Second = nullptr;
9250 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice,
9251 // Set the beginning of the pair.
9254 Second = &LoadedSlices[CurrSlice];
9256 // If First is NULL, it means we start a new pair.
9257 // Get to the next slice.
9261 EVT LoadedType = First->getLoadedType();
9263 // If the types of the slices are different, we cannot pair them.
9264 if (LoadedType != Second->getLoadedType())
9267 // Check if the target supplies paired loads for this type.
9268 unsigned RequiredAlignment = 0;
9269 if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) {
9270 // move to the next pair, this type is hopeless.
9274 // Check if we meet the alignment requirement.
9275 if (RequiredAlignment > First->getAlignment())
9278 // Check that both loads are next to each other in memory.
9279 if (!areSlicesNextToEachOther(*First, *Second))
9282 assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!");
9283 --GlobalLSCost.Loads;
9284 // Move to the next pair.
9289 /// \brief Check the profitability of all involved LoadedSlice.
9290 /// Currently, it is considered profitable if there is exactly two
9291 /// involved slices (1) which are (2) next to each other in memory, and
9292 /// whose cost (\see LoadedSlice::Cost) is smaller than the original load (3).
9294 /// Note: The order of the elements in \p LoadedSlices may be modified, but not
9295 /// the elements themselves.
9297 /// FIXME: When the cost model will be mature enough, we can relax
9298 /// constraints (1) and (2).
9299 static bool isSlicingProfitable(SmallVectorImpl<LoadedSlice> &LoadedSlices,
9300 const APInt &UsedBits, bool ForCodeSize) {
9301 unsigned NumberOfSlices = LoadedSlices.size();
9302 if (StressLoadSlicing)
9303 return NumberOfSlices > 1;
9306 if (NumberOfSlices != 2)
9310 if (!areUsedBitsDense(UsedBits))
9314 LoadedSlice::Cost OrigCost(ForCodeSize), GlobalSlicingCost(ForCodeSize);
9315 // The original code has one big load.
9317 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice) {
9318 const LoadedSlice &LS = LoadedSlices[CurrSlice];
9319 // Accumulate the cost of all the slices.
9320 LoadedSlice::Cost SliceCost(LS, ForCodeSize);
9321 GlobalSlicingCost += SliceCost;
9323 // Account as cost in the original configuration the gain obtained
9324 // with the current slices.
9325 OrigCost.addSliceGain(LS);
9328 // If the target supports paired load, adjust the cost accordingly.
9329 adjustCostForPairing(LoadedSlices, GlobalSlicingCost);
9330 return OrigCost > GlobalSlicingCost;
9333 /// \brief If the given load, \p LI, is used only by trunc or trunc(lshr)
9334 /// operations, split it in the various pieces being extracted.
9336 /// This sort of thing is introduced by SROA.
9337 /// This slicing takes care not to insert overlapping loads.
9338 /// \pre LI is a simple load (i.e., not an atomic or volatile load).
9339 bool DAGCombiner::SliceUpLoad(SDNode *N) {
9340 if (Level < AfterLegalizeDAG)
9343 LoadSDNode *LD = cast<LoadSDNode>(N);
9344 if (LD->isVolatile() || !ISD::isNormalLoad(LD) ||
9345 !LD->getValueType(0).isInteger())
9348 // Keep track of already used bits to detect overlapping values.
9349 // In that case, we will just abort the transformation.
9350 APInt UsedBits(LD->getValueSizeInBits(0), 0);
9352 SmallVector<LoadedSlice, 4> LoadedSlices;
9354 // Check if this load is used as several smaller chunks of bits.
9355 // Basically, look for uses in trunc or trunc(lshr) and record a new chain
9356 // of computation for each trunc.
9357 for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end();
9358 UI != UIEnd; ++UI) {
9359 // Skip the uses of the chain.
9360 if (UI.getUse().getResNo() != 0)
9366 // Check if this is a trunc(lshr).
9367 if (User->getOpcode() == ISD::SRL && User->hasOneUse() &&
9368 isa<ConstantSDNode>(User->getOperand(1))) {
9369 Shift = cast<ConstantSDNode>(User->getOperand(1))->getZExtValue();
9370 User = *User->use_begin();
9373 // At this point, User is a Truncate, iff we encountered, trunc or
9375 if (User->getOpcode() != ISD::TRUNCATE)
9378 // The width of the type must be a power of 2 and greater than 8-bits.
9379 // Otherwise the load cannot be represented in LLVM IR.
9380 // Moreover, if we shifted with a non-8-bits multiple, the slice
9381 // will be across several bytes. We do not support that.
9382 unsigned Width = User->getValueSizeInBits(0);
9383 if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7))
9386 // Build the slice for this chain of computations.
9387 LoadedSlice LS(User, LD, Shift, &DAG);
9388 APInt CurrentUsedBits = LS.getUsedBits();
9390 // Check if this slice overlaps with another.
9391 if ((CurrentUsedBits & UsedBits) != 0)
9393 // Update the bits used globally.
9394 UsedBits |= CurrentUsedBits;
9396 // Check if the new slice would be legal.
9400 // Record the slice.
9401 LoadedSlices.push_back(LS);
9404 // Abort slicing if it does not seem to be profitable.
9405 if (!isSlicingProfitable(LoadedSlices, UsedBits, ForCodeSize))
9410 // Rewrite each chain to use an independent load.
9411 // By construction, each chain can be represented by a unique load.
9413 // Prepare the argument for the new token factor for all the slices.
9414 SmallVector<SDValue, 8> ArgChains;
9415 for (SmallVectorImpl<LoadedSlice>::const_iterator
9416 LSIt = LoadedSlices.begin(),
9417 LSItEnd = LoadedSlices.end();
9418 LSIt != LSItEnd; ++LSIt) {
9419 SDValue SliceInst = LSIt->loadSlice();
9420 CombineTo(LSIt->Inst, SliceInst, true);
9421 if (SliceInst.getNode()->getOpcode() != ISD::LOAD)
9422 SliceInst = SliceInst.getOperand(0);
9423 assert(SliceInst->getOpcode() == ISD::LOAD &&
9424 "It takes more than a zext to get to the loaded slice!!");
9425 ArgChains.push_back(SliceInst.getValue(1));
9428 SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other,
9430 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
9434 /// Check to see if V is (and load (ptr), imm), where the load is having
9435 /// specific bytes cleared out. If so, return the byte size being masked out
9436 /// and the shift amount.
9437 static std::pair<unsigned, unsigned>
9438 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
9439 std::pair<unsigned, unsigned> Result(0, 0);
9441 // Check for the structure we're looking for.
9442 if (V->getOpcode() != ISD::AND ||
9443 !isa<ConstantSDNode>(V->getOperand(1)) ||
9444 !ISD::isNormalLoad(V->getOperand(0).getNode()))
9447 // Check the chain and pointer.
9448 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
9449 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
9451 // The store should be chained directly to the load or be an operand of a
9453 if (LD == Chain.getNode())
9455 else if (Chain->getOpcode() != ISD::TokenFactor)
9456 return Result; // Fail.
9459 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
9460 if (Chain->getOperand(i).getNode() == LD) {
9464 if (!isOk) return Result;
9467 // This only handles simple types.
9468 if (V.getValueType() != MVT::i16 &&
9469 V.getValueType() != MVT::i32 &&
9470 V.getValueType() != MVT::i64)
9473 // Check the constant mask. Invert it so that the bits being masked out are
9474 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
9475 // follow the sign bit for uniformity.
9476 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
9477 unsigned NotMaskLZ = countLeadingZeros(NotMask);
9478 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
9479 unsigned NotMaskTZ = countTrailingZeros(NotMask);
9480 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
9481 if (NotMaskLZ == 64) return Result; // All zero mask.
9483 // See if we have a continuous run of bits. If so, we have 0*1+0*
9484 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
9487 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
9488 if (V.getValueType() != MVT::i64 && NotMaskLZ)
9489 NotMaskLZ -= 64-V.getValueSizeInBits();
9491 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
9492 switch (MaskedBytes) {
9496 default: return Result; // All one mask, or 5-byte mask.
9499 // Verify that the first bit starts at a multiple of mask so that the access
9500 // is aligned the same as the access width.
9501 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
9503 Result.first = MaskedBytes;
9504 Result.second = NotMaskTZ/8;
9509 /// Check to see if IVal is something that provides a value as specified by
9510 /// MaskInfo. If so, replace the specified store with a narrower store of
9513 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
9514 SDValue IVal, StoreSDNode *St,
9516 unsigned NumBytes = MaskInfo.first;
9517 unsigned ByteShift = MaskInfo.second;
9518 SelectionDAG &DAG = DC->getDAG();
9520 // Check to see if IVal is all zeros in the part being masked in by the 'or'
9521 // that uses this. If not, this is not a replacement.
9522 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
9523 ByteShift*8, (ByteShift+NumBytes)*8);
9524 if (!DAG.MaskedValueIsZero(IVal, Mask)) return nullptr;
9526 // Check that it is legal on the target to do this. It is legal if the new
9527 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
9529 MVT VT = MVT::getIntegerVT(NumBytes*8);
9530 if (!DC->isTypeLegal(VT))
9533 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
9534 // shifted by ByteShift and truncated down to NumBytes.
9536 IVal = DAG.getNode(ISD::SRL, SDLoc(IVal), IVal.getValueType(), IVal,
9537 DAG.getConstant(ByteShift*8,
9538 DC->getShiftAmountTy(IVal.getValueType())));
9540 // Figure out the offset for the store and the alignment of the access.
9542 unsigned NewAlign = St->getAlignment();
9544 if (DAG.getTargetLoweringInfo().isLittleEndian())
9545 StOffset = ByteShift;
9547 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
9549 SDValue Ptr = St->getBasePtr();
9551 Ptr = DAG.getNode(ISD::ADD, SDLoc(IVal), Ptr.getValueType(),
9552 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
9553 NewAlign = MinAlign(NewAlign, StOffset);
9556 // Truncate down to the new size.
9557 IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal);
9560 return DAG.getStore(St->getChain(), SDLoc(St), IVal, Ptr,
9561 St->getPointerInfo().getWithOffset(StOffset),
9562 false, false, NewAlign).getNode();
9566 /// Look for sequence of load / op / store where op is one of 'or', 'xor', and
9567 /// 'and' of immediates. If 'op' is only touching some of the loaded bits, try
9568 /// narrowing the load and store if it would end up being a win for performance
9570 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
9571 StoreSDNode *ST = cast<StoreSDNode>(N);
9572 if (ST->isVolatile())
9575 SDValue Chain = ST->getChain();
9576 SDValue Value = ST->getValue();
9577 SDValue Ptr = ST->getBasePtr();
9578 EVT VT = Value.getValueType();
9580 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
9583 unsigned Opc = Value.getOpcode();
9585 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
9586 // is a byte mask indicating a consecutive number of bytes, check to see if
9587 // Y is known to provide just those bytes. If so, we try to replace the
9588 // load + replace + store sequence with a single (narrower) store, which makes
9590 if (Opc == ISD::OR) {
9591 std::pair<unsigned, unsigned> MaskedLoad;
9592 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
9593 if (MaskedLoad.first)
9594 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
9595 Value.getOperand(1), ST,this))
9596 return SDValue(NewST, 0);
9598 // Or is commutative, so try swapping X and Y.
9599 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
9600 if (MaskedLoad.first)
9601 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
9602 Value.getOperand(0), ST,this))
9603 return SDValue(NewST, 0);
9606 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
9607 Value.getOperand(1).getOpcode() != ISD::Constant)
9610 SDValue N0 = Value.getOperand(0);
9611 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
9612 Chain == SDValue(N0.getNode(), 1)) {
9613 LoadSDNode *LD = cast<LoadSDNode>(N0);
9614 if (LD->getBasePtr() != Ptr ||
9615 LD->getPointerInfo().getAddrSpace() !=
9616 ST->getPointerInfo().getAddrSpace())
9619 // Find the type to narrow it the load / op / store to.
9620 SDValue N1 = Value.getOperand(1);
9621 unsigned BitWidth = N1.getValueSizeInBits();
9622 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
9623 if (Opc == ISD::AND)
9624 Imm ^= APInt::getAllOnesValue(BitWidth);
9625 if (Imm == 0 || Imm.isAllOnesValue())
9627 unsigned ShAmt = Imm.countTrailingZeros();
9628 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
9629 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
9630 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
9631 // The narrowing should be profitable, the load/store operation should be
9632 // legal (or custom) and the store size should be equal to the NewVT width.
9633 while (NewBW < BitWidth &&
9634 (NewVT.getStoreSizeInBits() != NewBW ||
9635 !TLI.isOperationLegalOrCustom(Opc, NewVT) ||
9636 !TLI.isNarrowingProfitable(VT, NewVT))) {
9637 NewBW = NextPowerOf2(NewBW);
9638 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
9640 if (NewBW >= BitWidth)
9643 // If the lsb changed does not start at the type bitwidth boundary,
9644 // start at the previous one.
9646 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
9647 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt,
9648 std::min(BitWidth, ShAmt + NewBW));
9649 if ((Imm & Mask) == Imm) {
9650 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
9651 if (Opc == ISD::AND)
9652 NewImm ^= APInt::getAllOnesValue(NewBW);
9653 uint64_t PtrOff = ShAmt / 8;
9654 // For big endian targets, we need to adjust the offset to the pointer to
9655 // load the correct bytes.
9656 if (TLI.isBigEndian())
9657 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
9659 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
9660 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
9661 if (NewAlign < TLI.getDataLayout()->getABITypeAlignment(NewVTTy))
9664 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD),
9665 Ptr.getValueType(), Ptr,
9666 DAG.getConstant(PtrOff, Ptr.getValueType()));
9667 SDValue NewLD = DAG.getLoad(NewVT, SDLoc(N0),
9668 LD->getChain(), NewPtr,
9669 LD->getPointerInfo().getWithOffset(PtrOff),
9670 LD->isVolatile(), LD->isNonTemporal(),
9671 LD->isInvariant(), NewAlign,
9673 SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD,
9674 DAG.getConstant(NewImm, NewVT));
9675 SDValue NewST = DAG.getStore(Chain, SDLoc(N),
9677 ST->getPointerInfo().getWithOffset(PtrOff),
9678 false, false, NewAlign);
9680 AddToWorklist(NewPtr.getNode());
9681 AddToWorklist(NewLD.getNode());
9682 AddToWorklist(NewVal.getNode());
9683 WorklistRemover DeadNodes(*this);
9684 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
9693 /// For a given floating point load / store pair, if the load value isn't used
9694 /// by any other operations, then consider transforming the pair to integer
9695 /// load / store operations if the target deems the transformation profitable.
9696 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
9697 StoreSDNode *ST = cast<StoreSDNode>(N);
9698 SDValue Chain = ST->getChain();
9699 SDValue Value = ST->getValue();
9700 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
9701 Value.hasOneUse() &&
9702 Chain == SDValue(Value.getNode(), 1)) {
9703 LoadSDNode *LD = cast<LoadSDNode>(Value);
9704 EVT VT = LD->getMemoryVT();
9705 if (!VT.isFloatingPoint() ||
9706 VT != ST->getMemoryVT() ||
9707 LD->isNonTemporal() ||
9708 ST->isNonTemporal() ||
9709 LD->getPointerInfo().getAddrSpace() != 0 ||
9710 ST->getPointerInfo().getAddrSpace() != 0)
9713 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
9714 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
9715 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
9716 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
9717 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
9720 unsigned LDAlign = LD->getAlignment();
9721 unsigned STAlign = ST->getAlignment();
9722 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
9723 unsigned ABIAlign = TLI.getDataLayout()->getABITypeAlignment(IntVTTy);
9724 if (LDAlign < ABIAlign || STAlign < ABIAlign)
9727 SDValue NewLD = DAG.getLoad(IntVT, SDLoc(Value),
9728 LD->getChain(), LD->getBasePtr(),
9729 LD->getPointerInfo(),
9730 false, false, false, LDAlign);
9732 SDValue NewST = DAG.getStore(NewLD.getValue(1), SDLoc(N),
9733 NewLD, ST->getBasePtr(),
9734 ST->getPointerInfo(),
9735 false, false, STAlign);
9737 AddToWorklist(NewLD.getNode());
9738 AddToWorklist(NewST.getNode());
9739 WorklistRemover DeadNodes(*this);
9740 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
9748 /// Helper struct to parse and store a memory address as base + index + offset.
9749 /// We ignore sign extensions when it is safe to do so.
9750 /// The following two expressions are not equivalent. To differentiate we need
9751 /// to store whether there was a sign extension involved in the index
9753 /// (load (i64 add (i64 copyfromreg %c)
9754 /// (i64 signextend (add (i8 load %index)
9758 /// (load (i64 add (i64 copyfromreg %c)
9759 /// (i64 signextend (i32 add (i32 signextend (i8 load %index))
9761 struct BaseIndexOffset {
9765 bool IsIndexSignExt;
9767 BaseIndexOffset() : Offset(0), IsIndexSignExt(false) {}
9769 BaseIndexOffset(SDValue Base, SDValue Index, int64_t Offset,
9770 bool IsIndexSignExt) :
9771 Base(Base), Index(Index), Offset(Offset), IsIndexSignExt(IsIndexSignExt) {}
9773 bool equalBaseIndex(const BaseIndexOffset &Other) {
9774 return Other.Base == Base && Other.Index == Index &&
9775 Other.IsIndexSignExt == IsIndexSignExt;
9778 /// Parses tree in Ptr for base, index, offset addresses.
9779 static BaseIndexOffset match(SDValue Ptr) {
9780 bool IsIndexSignExt = false;
9782 // We only can pattern match BASE + INDEX + OFFSET. If Ptr is not an ADD
9783 // instruction, then it could be just the BASE or everything else we don't
9784 // know how to handle. Just use Ptr as BASE and give up.
9785 if (Ptr->getOpcode() != ISD::ADD)
9786 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
9788 // We know that we have at least an ADD instruction. Try to pattern match
9789 // the simple case of BASE + OFFSET.
9790 if (isa<ConstantSDNode>(Ptr->getOperand(1))) {
9791 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
9792 return BaseIndexOffset(Ptr->getOperand(0), SDValue(), Offset,
9796 // Inside a loop the current BASE pointer is calculated using an ADD and a
9797 // MUL instruction. In this case Ptr is the actual BASE pointer.
9798 // (i64 add (i64 %array_ptr)
9799 // (i64 mul (i64 %induction_var)
9800 // (i64 %element_size)))
9801 if (Ptr->getOperand(1)->getOpcode() == ISD::MUL)
9802 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
9804 // Look at Base + Index + Offset cases.
9805 SDValue Base = Ptr->getOperand(0);
9806 SDValue IndexOffset = Ptr->getOperand(1);
9808 // Skip signextends.
9809 if (IndexOffset->getOpcode() == ISD::SIGN_EXTEND) {
9810 IndexOffset = IndexOffset->getOperand(0);
9811 IsIndexSignExt = true;
9814 // Either the case of Base + Index (no offset) or something else.
9815 if (IndexOffset->getOpcode() != ISD::ADD)
9816 return BaseIndexOffset(Base, IndexOffset, 0, IsIndexSignExt);
9818 // Now we have the case of Base + Index + offset.
9819 SDValue Index = IndexOffset->getOperand(0);
9820 SDValue Offset = IndexOffset->getOperand(1);
9822 if (!isa<ConstantSDNode>(Offset))
9823 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
9825 // Ignore signextends.
9826 if (Index->getOpcode() == ISD::SIGN_EXTEND) {
9827 Index = Index->getOperand(0);
9828 IsIndexSignExt = true;
9829 } else IsIndexSignExt = false;
9831 int64_t Off = cast<ConstantSDNode>(Offset)->getSExtValue();
9832 return BaseIndexOffset(Base, Index, Off, IsIndexSignExt);
9836 bool DAGCombiner::MergeStoresOfConstantsOrVecElts(
9837 SmallVectorImpl<MemOpLink> &StoreNodes, EVT MemVT,
9838 unsigned NumElem, bool IsConstantSrc, bool UseVector) {
9839 // Make sure we have something to merge.
9843 int64_t ElementSizeBytes = MemVT.getSizeInBits() / 8;
9844 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
9845 unsigned EarliestNodeUsed = 0;
9847 for (unsigned i=0; i < NumElem; ++i) {
9848 // Find a chain for the new wide-store operand. Notice that some
9849 // of the store nodes that we found may not be selected for inclusion
9850 // in the wide store. The chain we use needs to be the chain of the
9851 // earliest store node which is *used* and replaced by the wide store.
9852 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
9853 EarliestNodeUsed = i;
9856 // The earliest Node in the DAG.
9857 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
9858 SDLoc DL(StoreNodes[0].MemNode);
9862 // Find a legal type for the vector store.
9863 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
9864 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
9865 if (IsConstantSrc) {
9866 // A vector store with a constant source implies that the constant is
9867 // zero; we only handle merging stores of constant zeros because the zero
9868 // can be materialized without a load.
9869 // It may be beneficial to loosen this restriction to allow non-zero
9871 StoredVal = DAG.getConstant(0, Ty);
9873 SmallVector<SDValue, 8> Ops;
9874 for (unsigned i = 0; i < NumElem ; ++i) {
9875 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9876 SDValue Val = St->getValue();
9877 // All of the operands of a BUILD_VECTOR must have the same type.
9878 if (Val.getValueType() != MemVT)
9883 // Build the extracted vector elements back into a vector.
9884 StoredVal = DAG.getNode(ISD::BUILD_VECTOR, DL, Ty, Ops);
9887 // We should always use a vector store when merging extracted vector
9888 // elements, so this path implies a store of constants.
9889 assert(IsConstantSrc && "Merged vector elements should use vector store");
9891 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
9892 APInt StoreInt(StoreBW, 0);
9894 // Construct a single integer constant which is made of the smaller
9896 bool IsLE = TLI.isLittleEndian();
9897 for (unsigned i = 0; i < NumElem ; ++i) {
9898 unsigned Idx = IsLE ? (NumElem - 1 - i) : i;
9899 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
9900 SDValue Val = St->getValue();
9901 StoreInt <<= ElementSizeBytes*8;
9902 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
9903 StoreInt |= C->getAPIntValue().zext(StoreBW);
9904 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
9905 StoreInt |= C->getValueAPF().bitcastToAPInt().zext(StoreBW);
9907 llvm_unreachable("Invalid constant element type");
9911 // Create the new Load and Store operations.
9912 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9913 StoredVal = DAG.getConstant(StoreInt, StoreTy);
9916 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), DL, StoredVal,
9917 FirstInChain->getBasePtr(),
9918 FirstInChain->getPointerInfo(),
9920 FirstInChain->getAlignment());
9922 // Replace the first store with the new store
9923 CombineTo(EarliestOp, NewStore);
9924 // Erase all other stores.
9925 for (unsigned i = 0; i < NumElem ; ++i) {
9926 if (StoreNodes[i].MemNode == EarliestOp)
9928 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9929 // ReplaceAllUsesWith will replace all uses that existed when it was
9930 // called, but graph optimizations may cause new ones to appear. For
9931 // example, the case in pr14333 looks like
9933 // St's chain -> St -> another store -> X
9935 // And the only difference from St to the other store is the chain.
9936 // When we change it's chain to be St's chain they become identical,
9937 // get CSEed and the net result is that X is now a use of St.
9938 // Since we know that St is redundant, just iterate.
9939 while (!St->use_empty())
9940 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain());
9941 deleteAndRecombine(St);
9947 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
9948 EVT MemVT = St->getMemoryVT();
9949 int64_t ElementSizeBytes = MemVT.getSizeInBits()/8;
9950 bool NoVectors = DAG.getMachineFunction().getFunction()->getAttributes().
9951 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
9953 // Don't merge vectors into wider inputs.
9954 if (MemVT.isVector() || !MemVT.isSimple())
9957 // Perform an early exit check. Do not bother looking at stored values that
9958 // are not constants, loads, or extracted vector elements.
9959 SDValue StoredVal = St->getValue();
9960 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
9961 bool IsConstantSrc = isa<ConstantSDNode>(StoredVal) ||
9962 isa<ConstantFPSDNode>(StoredVal);
9963 bool IsExtractVecEltSrc = (StoredVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT);
9965 if (!IsConstantSrc && !IsLoadSrc && !IsExtractVecEltSrc)
9968 // Only look at ends of store sequences.
9969 SDValue Chain = SDValue(St, 0);
9970 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
9973 // This holds the base pointer, index, and the offset in bytes from the base
9975 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr());
9977 // We must have a base and an offset.
9978 if (!BasePtr.Base.getNode())
9981 // Do not handle stores to undef base pointers.
9982 if (BasePtr.Base.getOpcode() == ISD::UNDEF)
9985 // Save the LoadSDNodes that we find in the chain.
9986 // We need to make sure that these nodes do not interfere with
9987 // any of the store nodes.
9988 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
9990 // Save the StoreSDNodes that we find in the chain.
9991 SmallVector<MemOpLink, 8> StoreNodes;
9993 // Walk up the chain and look for nodes with offsets from the same
9994 // base pointer. Stop when reaching an instruction with a different kind
9995 // or instruction which has a different base pointer.
9997 StoreSDNode *Index = St;
9999 // If the chain has more than one use, then we can't reorder the mem ops.
10000 if (Index != St && !SDValue(Index, 0)->hasOneUse())
10003 // Find the base pointer and offset for this memory node.
10004 BaseIndexOffset Ptr = BaseIndexOffset::match(Index->getBasePtr());
10006 // Check that the base pointer is the same as the original one.
10007 if (!Ptr.equalBaseIndex(BasePtr))
10010 // Check that the alignment is the same.
10011 if (Index->getAlignment() != St->getAlignment())
10014 // The memory operands must not be volatile.
10015 if (Index->isVolatile() || Index->isIndexed())
10019 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
10020 if (St->isTruncatingStore())
10023 // The stored memory type must be the same.
10024 if (Index->getMemoryVT() != MemVT)
10027 // We do not allow unaligned stores because we want to prevent overriding
10029 if (Index->getAlignment()*8 != MemVT.getSizeInBits())
10032 // We found a potential memory operand to merge.
10033 StoreNodes.push_back(MemOpLink(Index, Ptr.Offset, Seq++));
10035 // Find the next memory operand in the chain. If the next operand in the
10036 // chain is a store then move up and continue the scan with the next
10037 // memory operand. If the next operand is a load save it and use alias
10038 // information to check if it interferes with anything.
10039 SDNode *NextInChain = Index->getChain().getNode();
10041 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
10042 // We found a store node. Use it for the next iteration.
10045 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
10046 if (Ldn->isVolatile()) {
10051 // Save the load node for later. Continue the scan.
10052 AliasLoadNodes.push_back(Ldn);
10053 NextInChain = Ldn->getChain().getNode();
10062 // Check if there is anything to merge.
10063 if (StoreNodes.size() < 2)
10066 // Sort the memory operands according to their distance from the base pointer.
10067 std::sort(StoreNodes.begin(), StoreNodes.end(),
10068 [](MemOpLink LHS, MemOpLink RHS) {
10069 return LHS.OffsetFromBase < RHS.OffsetFromBase ||
10070 (LHS.OffsetFromBase == RHS.OffsetFromBase &&
10071 LHS.SequenceNum > RHS.SequenceNum);
10074 // Scan the memory operations on the chain and find the first non-consecutive
10075 // store memory address.
10076 unsigned LastConsecutiveStore = 0;
10077 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
10078 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) {
10080 // Check that the addresses are consecutive starting from the second
10081 // element in the list of stores.
10083 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
10084 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
10088 bool Alias = false;
10089 // Check if this store interferes with any of the loads that we found.
10090 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld)
10091 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) {
10095 // We found a load that alias with this store. Stop the sequence.
10099 // Mark this node as useful.
10100 LastConsecutiveStore = i;
10103 // The node with the lowest store address.
10104 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
10106 // Store the constants into memory as one consecutive store.
10107 if (IsConstantSrc) {
10108 unsigned LastLegalType = 0;
10109 unsigned LastLegalVectorType = 0;
10110 bool NonZero = false;
10111 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
10112 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10113 SDValue StoredVal = St->getValue();
10115 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
10116 NonZero |= !C->isNullValue();
10117 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) {
10118 NonZero |= !C->getConstantFPValue()->isNullValue();
10124 // Find a legal type for the constant store.
10125 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
10126 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
10127 if (TLI.isTypeLegal(StoreTy))
10128 LastLegalType = i+1;
10129 // Or check whether a truncstore is legal.
10130 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
10131 TargetLowering::TypePromoteInteger) {
10132 EVT LegalizedStoredValueTy =
10133 TLI.getTypeToTransformTo(*DAG.getContext(), StoredVal.getValueType());
10134 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy))
10135 LastLegalType = i+1;
10138 // Find a legal type for the vector store.
10139 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
10140 if (TLI.isTypeLegal(Ty))
10141 LastLegalVectorType = i + 1;
10144 // We only use vectors if the constant is known to be zero and the
10145 // function is not marked with the noimplicitfloat attribute.
10146 if (NonZero || NoVectors)
10147 LastLegalVectorType = 0;
10149 // Check if we found a legal integer type to store.
10150 if (LastLegalType == 0 && LastLegalVectorType == 0)
10153 bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors;
10154 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType;
10156 return MergeStoresOfConstantsOrVecElts(StoreNodes, MemVT, NumElem,
10160 // When extracting multiple vector elements, try to store them
10161 // in one vector store rather than a sequence of scalar stores.
10162 if (IsExtractVecEltSrc) {
10163 unsigned NumElem = 0;
10164 for (unsigned i = 0; i < LastConsecutiveStore + 1; ++i) {
10165 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10166 SDValue StoredVal = St->getValue();
10167 // This restriction could be loosened.
10168 // Bail out if any stored values are not elements extracted from a vector.
10169 // It should be possible to handle mixed sources, but load sources need
10170 // more careful handling (see the block of code below that handles
10171 // consecutive loads).
10172 if (StoredVal.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
10175 // Find a legal type for the vector store.
10176 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
10177 if (TLI.isTypeLegal(Ty))
10181 return MergeStoresOfConstantsOrVecElts(StoreNodes, MemVT, NumElem,
10185 // Below we handle the case of multiple consecutive stores that
10186 // come from multiple consecutive loads. We merge them into a single
10187 // wide load and a single wide store.
10189 // Look for load nodes which are used by the stored values.
10190 SmallVector<MemOpLink, 8> LoadNodes;
10192 // Find acceptable loads. Loads need to have the same chain (token factor),
10193 // must not be zext, volatile, indexed, and they must be consecutive.
10194 BaseIndexOffset LdBasePtr;
10195 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
10196 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10197 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
10200 // Loads must only have one use.
10201 if (!Ld->hasNUsesOfValue(1, 0))
10204 // Check that the alignment is the same as the stores.
10205 if (Ld->getAlignment() != St->getAlignment())
10208 // The memory operands must not be volatile.
10209 if (Ld->isVolatile() || Ld->isIndexed())
10212 // We do not accept ext loads.
10213 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
10216 // The stored memory type must be the same.
10217 if (Ld->getMemoryVT() != MemVT)
10220 BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld->getBasePtr());
10221 // If this is not the first ptr that we check.
10222 if (LdBasePtr.Base.getNode()) {
10223 // The base ptr must be the same.
10224 if (!LdPtr.equalBaseIndex(LdBasePtr))
10227 // Check that all other base pointers are the same as this one.
10231 // We found a potential memory operand to merge.
10232 LoadNodes.push_back(MemOpLink(Ld, LdPtr.Offset, 0));
10235 if (LoadNodes.size() < 2)
10238 // If we have load/store pair instructions and we only have two values,
10240 unsigned RequiredAlignment;
10241 if (LoadNodes.size() == 2 && TLI.hasPairedLoad(MemVT, RequiredAlignment) &&
10242 St->getAlignment() >= RequiredAlignment)
10245 // Scan the memory operations on the chain and find the first non-consecutive
10246 // load memory address. These variables hold the index in the store node
10248 unsigned LastConsecutiveLoad = 0;
10249 // This variable refers to the size and not index in the array.
10250 unsigned LastLegalVectorType = 0;
10251 unsigned LastLegalIntegerType = 0;
10252 StartAddress = LoadNodes[0].OffsetFromBase;
10253 SDValue FirstChain = LoadNodes[0].MemNode->getChain();
10254 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
10255 // All loads much share the same chain.
10256 if (LoadNodes[i].MemNode->getChain() != FirstChain)
10259 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
10260 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
10262 LastConsecutiveLoad = i;
10264 // Find a legal type for the vector store.
10265 EVT StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
10266 if (TLI.isTypeLegal(StoreTy))
10267 LastLegalVectorType = i + 1;
10269 // Find a legal type for the integer store.
10270 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
10271 StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
10272 if (TLI.isTypeLegal(StoreTy))
10273 LastLegalIntegerType = i + 1;
10274 // Or check whether a truncstore and extload is legal.
10275 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
10276 TargetLowering::TypePromoteInteger) {
10277 EVT LegalizedStoredValueTy =
10278 TLI.getTypeToTransformTo(*DAG.getContext(), StoreTy);
10279 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
10280 TLI.isLoadExtLegal(ISD::ZEXTLOAD, LegalizedStoredValueTy, StoreTy) &&
10281 TLI.isLoadExtLegal(ISD::SEXTLOAD, LegalizedStoredValueTy, StoreTy) &&
10282 TLI.isLoadExtLegal(ISD::EXTLOAD, LegalizedStoredValueTy, StoreTy))
10283 LastLegalIntegerType = i+1;
10287 // Only use vector types if the vector type is larger than the integer type.
10288 // If they are the same, use integers.
10289 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors;
10290 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType);
10292 // We add +1 here because the LastXXX variables refer to location while
10293 // the NumElem refers to array/index size.
10294 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1;
10295 NumElem = std::min(LastLegalType, NumElem);
10300 // The earliest Node in the DAG.
10301 unsigned EarliestNodeUsed = 0;
10302 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
10303 for (unsigned i=1; i<NumElem; ++i) {
10304 // Find a chain for the new wide-store operand. Notice that some
10305 // of the store nodes that we found may not be selected for inclusion
10306 // in the wide store. The chain we use needs to be the chain of the
10307 // earliest store node which is *used* and replaced by the wide store.
10308 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
10309 EarliestNodeUsed = i;
10312 // Find if it is better to use vectors or integers to load and store
10316 JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
10318 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
10319 JointMemOpVT = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
10322 SDLoc LoadDL(LoadNodes[0].MemNode);
10323 SDLoc StoreDL(StoreNodes[0].MemNode);
10325 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
10326 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL,
10327 FirstLoad->getChain(),
10328 FirstLoad->getBasePtr(),
10329 FirstLoad->getPointerInfo(),
10330 false, false, false,
10331 FirstLoad->getAlignment());
10333 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), StoreDL, NewLoad,
10334 FirstInChain->getBasePtr(),
10335 FirstInChain->getPointerInfo(), false, false,
10336 FirstInChain->getAlignment());
10338 // Replace one of the loads with the new load.
10339 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode);
10340 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
10341 SDValue(NewLoad.getNode(), 1));
10343 // Remove the rest of the load chains.
10344 for (unsigned i = 1; i < NumElem ; ++i) {
10345 // Replace all chain users of the old load nodes with the chain of the new
10347 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
10348 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain());
10351 // Replace the first store with the new store.
10352 CombineTo(EarliestOp, NewStore);
10353 // Erase all other stores.
10354 for (unsigned i = 0; i < NumElem ; ++i) {
10355 // Remove all Store nodes.
10356 if (StoreNodes[i].MemNode == EarliestOp)
10358 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10359 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain());
10360 deleteAndRecombine(St);
10366 SDValue DAGCombiner::visitSTORE(SDNode *N) {
10367 StoreSDNode *ST = cast<StoreSDNode>(N);
10368 SDValue Chain = ST->getChain();
10369 SDValue Value = ST->getValue();
10370 SDValue Ptr = ST->getBasePtr();
10372 // If this is a store of a bit convert, store the input value if the
10373 // resultant store does not need a higher alignment than the original.
10374 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
10375 ST->isUnindexed()) {
10376 unsigned OrigAlign = ST->getAlignment();
10377 EVT SVT = Value.getOperand(0).getValueType();
10378 unsigned Align = TLI.getDataLayout()->
10379 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
10380 if (Align <= OrigAlign &&
10381 ((!LegalOperations && !ST->isVolatile()) ||
10382 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
10383 return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0),
10384 Ptr, ST->getPointerInfo(), ST->isVolatile(),
10385 ST->isNonTemporal(), OrigAlign,
10389 // Turn 'store undef, Ptr' -> nothing.
10390 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
10393 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
10394 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
10395 // NOTE: If the original store is volatile, this transform must not increase
10396 // the number of stores. For example, on x86-32 an f64 can be stored in one
10397 // processor operation but an i64 (which is not legal) requires two. So the
10398 // transform should not be done in this case.
10399 if (Value.getOpcode() != ISD::TargetConstantFP) {
10401 switch (CFP->getSimpleValueType(0).SimpleTy) {
10402 default: llvm_unreachable("Unknown FP type");
10403 case MVT::f16: // We don't do this for these yet.
10409 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
10410 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
10411 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
10412 bitcastToAPInt().getZExtValue(), MVT::i32);
10413 return DAG.getStore(Chain, SDLoc(N), Tmp,
10414 Ptr, ST->getMemOperand());
10418 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
10419 !ST->isVolatile()) ||
10420 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
10421 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
10422 getZExtValue(), MVT::i64);
10423 return DAG.getStore(Chain, SDLoc(N), Tmp,
10424 Ptr, ST->getMemOperand());
10427 if (!ST->isVolatile() &&
10428 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
10429 // Many FP stores are not made apparent until after legalize, e.g. for
10430 // argument passing. Since this is so common, custom legalize the
10431 // 64-bit integer store into two 32-bit stores.
10432 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
10433 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
10434 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
10435 if (TLI.isBigEndian()) std::swap(Lo, Hi);
10437 unsigned Alignment = ST->getAlignment();
10438 bool isVolatile = ST->isVolatile();
10439 bool isNonTemporal = ST->isNonTemporal();
10440 AAMDNodes AAInfo = ST->getAAInfo();
10442 SDValue St0 = DAG.getStore(Chain, SDLoc(ST), Lo,
10443 Ptr, ST->getPointerInfo(),
10444 isVolatile, isNonTemporal,
10445 ST->getAlignment(), AAInfo);
10446 Ptr = DAG.getNode(ISD::ADD, SDLoc(N), Ptr.getValueType(), Ptr,
10447 DAG.getConstant(4, Ptr.getValueType()));
10448 Alignment = MinAlign(Alignment, 4U);
10449 SDValue St1 = DAG.getStore(Chain, SDLoc(ST), Hi,
10450 Ptr, ST->getPointerInfo().getWithOffset(4),
10451 isVolatile, isNonTemporal,
10452 Alignment, AAInfo);
10453 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other,
10462 // Try to infer better alignment information than the store already has.
10463 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
10464 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
10465 if (Align > ST->getAlignment())
10466 return DAG.getTruncStore(Chain, SDLoc(N), Value,
10467 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
10468 ST->isVolatile(), ST->isNonTemporal(), Align,
10473 // Try transforming a pair floating point load / store ops to integer
10474 // load / store ops.
10475 SDValue NewST = TransformFPLoadStorePair(N);
10476 if (NewST.getNode())
10479 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
10480 : DAG.getSubtarget().useAA();
10482 if (CombinerAAOnlyFunc.getNumOccurrences() &&
10483 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
10486 if (UseAA && ST->isUnindexed()) {
10487 // Walk up chain skipping non-aliasing memory nodes.
10488 SDValue BetterChain = FindBetterChain(N, Chain);
10490 // If there is a better chain.
10491 if (Chain != BetterChain) {
10494 // Replace the chain to avoid dependency.
10495 if (ST->isTruncatingStore()) {
10496 ReplStore = DAG.getTruncStore(BetterChain, SDLoc(N), Value, Ptr,
10497 ST->getMemoryVT(), ST->getMemOperand());
10499 ReplStore = DAG.getStore(BetterChain, SDLoc(N), Value, Ptr,
10500 ST->getMemOperand());
10503 // Create token to keep both nodes around.
10504 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
10505 MVT::Other, Chain, ReplStore);
10507 // Make sure the new and old chains are cleaned up.
10508 AddToWorklist(Token.getNode());
10510 // Don't add users to work list.
10511 return CombineTo(N, Token, false);
10515 // Try transforming N to an indexed store.
10516 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
10517 return SDValue(N, 0);
10519 // FIXME: is there such a thing as a truncating indexed store?
10520 if (ST->isTruncatingStore() && ST->isUnindexed() &&
10521 Value.getValueType().isInteger()) {
10522 // See if we can simplify the input to this truncstore with knowledge that
10523 // only the low bits are being used. For example:
10524 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
10526 GetDemandedBits(Value,
10527 APInt::getLowBitsSet(
10528 Value.getValueType().getScalarType().getSizeInBits(),
10529 ST->getMemoryVT().getScalarType().getSizeInBits()));
10530 AddToWorklist(Value.getNode());
10531 if (Shorter.getNode())
10532 return DAG.getTruncStore(Chain, SDLoc(N), Shorter,
10533 Ptr, ST->getMemoryVT(), ST->getMemOperand());
10535 // Otherwise, see if we can simplify the operation with
10536 // SimplifyDemandedBits, which only works if the value has a single use.
10537 if (SimplifyDemandedBits(Value,
10538 APInt::getLowBitsSet(
10539 Value.getValueType().getScalarType().getSizeInBits(),
10540 ST->getMemoryVT().getScalarType().getSizeInBits())))
10541 return SDValue(N, 0);
10544 // If this is a load followed by a store to the same location, then the store
10546 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
10547 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
10548 ST->isUnindexed() && !ST->isVolatile() &&
10549 // There can't be any side effects between the load and store, such as
10550 // a call or store.
10551 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
10552 // The store is dead, remove it.
10557 // If this is a store followed by a store with the same value to the same
10558 // location, then the store is dead/noop.
10559 if (StoreSDNode *ST1 = dyn_cast<StoreSDNode>(Chain)) {
10560 if (ST1->getBasePtr() == Ptr && ST->getMemoryVT() == ST1->getMemoryVT() &&
10561 ST1->getValue() == Value && ST->isUnindexed() && !ST->isVolatile() &&
10562 ST1->isUnindexed() && !ST1->isVolatile()) {
10563 // The store is dead, remove it.
10568 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
10569 // truncating store. We can do this even if this is already a truncstore.
10570 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
10571 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
10572 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
10573 ST->getMemoryVT())) {
10574 return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0),
10575 Ptr, ST->getMemoryVT(), ST->getMemOperand());
10578 // Only perform this optimization before the types are legal, because we
10579 // don't want to perform this optimization on every DAGCombine invocation.
10581 bool EverChanged = false;
10584 // There can be multiple store sequences on the same chain.
10585 // Keep trying to merge store sequences until we are unable to do so
10586 // or until we merge the last store on the chain.
10587 bool Changed = MergeConsecutiveStores(ST);
10588 EverChanged |= Changed;
10589 if (!Changed) break;
10590 } while (ST->getOpcode() != ISD::DELETED_NODE);
10593 return SDValue(N, 0);
10596 return ReduceLoadOpStoreWidth(N);
10599 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
10600 SDValue InVec = N->getOperand(0);
10601 SDValue InVal = N->getOperand(1);
10602 SDValue EltNo = N->getOperand(2);
10605 // If the inserted element is an UNDEF, just use the input vector.
10606 if (InVal.getOpcode() == ISD::UNDEF)
10609 EVT VT = InVec.getValueType();
10611 // If we can't generate a legal BUILD_VECTOR, exit
10612 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
10615 // Check that we know which element is being inserted
10616 if (!isa<ConstantSDNode>(EltNo))
10618 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
10620 // Canonicalize insert_vector_elt dag nodes.
10622 // (insert_vector_elt (insert_vector_elt A, Idx0), Idx1)
10623 // -> (insert_vector_elt (insert_vector_elt A, Idx1), Idx0)
10625 // Do this only if the child insert_vector node has one use; also
10626 // do this only if indices are both constants and Idx1 < Idx0.
10627 if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT && InVec.hasOneUse()
10628 && isa<ConstantSDNode>(InVec.getOperand(2))) {
10629 unsigned OtherElt =
10630 cast<ConstantSDNode>(InVec.getOperand(2))->getZExtValue();
10631 if (Elt < OtherElt) {
10633 SDValue NewOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(N), VT,
10634 InVec.getOperand(0), InVal, EltNo);
10635 AddToWorklist(NewOp.getNode());
10636 return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(InVec.getNode()),
10637 VT, NewOp, InVec.getOperand(1), InVec.getOperand(2));
10641 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
10642 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
10643 // vector elements.
10644 SmallVector<SDValue, 8> Ops;
10645 // Do not combine these two vectors if the output vector will not replace
10646 // the input vector.
10647 if (InVec.getOpcode() == ISD::BUILD_VECTOR && InVec.hasOneUse()) {
10648 Ops.append(InVec.getNode()->op_begin(),
10649 InVec.getNode()->op_end());
10650 } else if (InVec.getOpcode() == ISD::UNDEF) {
10651 unsigned NElts = VT.getVectorNumElements();
10652 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
10657 // Insert the element
10658 if (Elt < Ops.size()) {
10659 // All the operands of BUILD_VECTOR must have the same type;
10660 // we enforce that here.
10661 EVT OpVT = Ops[0].getValueType();
10662 if (InVal.getValueType() != OpVT)
10663 InVal = OpVT.bitsGT(InVal.getValueType()) ?
10664 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
10665 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
10669 // Return the new vector
10670 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
10673 SDValue DAGCombiner::ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
10674 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad) {
10675 EVT ResultVT = EVE->getValueType(0);
10676 EVT VecEltVT = InVecVT.getVectorElementType();
10677 unsigned Align = OriginalLoad->getAlignment();
10678 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
10679 VecEltVT.getTypeForEVT(*DAG.getContext()));
10681 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VecEltVT))
10686 SDValue NewPtr = OriginalLoad->getBasePtr();
10688 EVT PtrType = NewPtr.getValueType();
10689 MachinePointerInfo MPI;
10690 if (auto *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo)) {
10691 int Elt = ConstEltNo->getZExtValue();
10692 unsigned PtrOff = VecEltVT.getSizeInBits() * Elt / 8;
10693 if (TLI.isBigEndian())
10694 PtrOff = InVecVT.getSizeInBits() / 8 - PtrOff;
10695 Offset = DAG.getConstant(PtrOff, PtrType);
10696 MPI = OriginalLoad->getPointerInfo().getWithOffset(PtrOff);
10698 Offset = DAG.getNode(
10699 ISD::MUL, SDLoc(EVE), EltNo.getValueType(), EltNo,
10700 DAG.getConstant(VecEltVT.getStoreSize(), EltNo.getValueType()));
10701 if (TLI.isBigEndian())
10702 Offset = DAG.getNode(
10703 ISD::SUB, SDLoc(EVE), EltNo.getValueType(),
10704 DAG.getConstant(InVecVT.getStoreSize(), EltNo.getValueType()), Offset);
10705 MPI = OriginalLoad->getPointerInfo();
10707 NewPtr = DAG.getNode(ISD::ADD, SDLoc(EVE), PtrType, NewPtr, Offset);
10709 // The replacement we need to do here is a little tricky: we need to
10710 // replace an extractelement of a load with a load.
10711 // Use ReplaceAllUsesOfValuesWith to do the replacement.
10712 // Note that this replacement assumes that the extractvalue is the only
10713 // use of the load; that's okay because we don't want to perform this
10714 // transformation in other cases anyway.
10717 if (ResultVT.bitsGT(VecEltVT)) {
10718 // If the result type of vextract is wider than the load, then issue an
10719 // extending load instead.
10720 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, ResultVT,
10724 Load = DAG.getExtLoad(
10725 ExtType, SDLoc(EVE), ResultVT, OriginalLoad->getChain(), NewPtr, MPI,
10726 VecEltVT, OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
10727 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
10728 Chain = Load.getValue(1);
10730 Load = DAG.getLoad(
10731 VecEltVT, SDLoc(EVE), OriginalLoad->getChain(), NewPtr, MPI,
10732 OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
10733 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
10734 Chain = Load.getValue(1);
10735 if (ResultVT.bitsLT(VecEltVT))
10736 Load = DAG.getNode(ISD::TRUNCATE, SDLoc(EVE), ResultVT, Load);
10738 Load = DAG.getNode(ISD::BITCAST, SDLoc(EVE), ResultVT, Load);
10740 WorklistRemover DeadNodes(*this);
10741 SDValue From[] = { SDValue(EVE, 0), SDValue(OriginalLoad, 1) };
10742 SDValue To[] = { Load, Chain };
10743 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
10744 // Since we're explicitly calling ReplaceAllUses, add the new node to the
10745 // worklist explicitly as well.
10746 AddToWorklist(Load.getNode());
10747 AddUsersToWorklist(Load.getNode()); // Add users too
10748 // Make sure to revisit this node to clean it up; it will usually be dead.
10749 AddToWorklist(EVE);
10751 return SDValue(EVE, 0);
10754 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
10755 // (vextract (scalar_to_vector val, 0) -> val
10756 SDValue InVec = N->getOperand(0);
10757 EVT VT = InVec.getValueType();
10758 EVT NVT = N->getValueType(0);
10760 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
10761 // Check if the result type doesn't match the inserted element type. A
10762 // SCALAR_TO_VECTOR may truncate the inserted element and the
10763 // EXTRACT_VECTOR_ELT may widen the extracted vector.
10764 SDValue InOp = InVec.getOperand(0);
10765 if (InOp.getValueType() != NVT) {
10766 assert(InOp.getValueType().isInteger() && NVT.isInteger());
10767 return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT);
10772 SDValue EltNo = N->getOperand(1);
10773 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
10775 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
10776 // We only perform this optimization before the op legalization phase because
10777 // we may introduce new vector instructions which are not backed by TD
10778 // patterns. For example on AVX, extracting elements from a wide vector
10779 // without using extract_subvector. However, if we can find an underlying
10780 // scalar value, then we can always use that.
10781 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
10783 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
10784 int NumElem = VT.getVectorNumElements();
10785 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
10786 // Find the new index to extract from.
10787 int OrigElt = SVOp->getMaskElt(Elt);
10789 // Extracting an undef index is undef.
10791 return DAG.getUNDEF(NVT);
10793 // Select the right vector half to extract from.
10795 if (OrigElt < NumElem) {
10796 SVInVec = InVec->getOperand(0);
10798 SVInVec = InVec->getOperand(1);
10799 OrigElt -= NumElem;
10802 if (SVInVec.getOpcode() == ISD::BUILD_VECTOR) {
10803 SDValue InOp = SVInVec.getOperand(OrigElt);
10804 if (InOp.getValueType() != NVT) {
10805 assert(InOp.getValueType().isInteger() && NVT.isInteger());
10806 InOp = DAG.getSExtOrTrunc(InOp, SDLoc(SVInVec), NVT);
10812 // FIXME: We should handle recursing on other vector shuffles and
10813 // scalar_to_vector here as well.
10815 if (!LegalOperations) {
10816 EVT IndexTy = TLI.getVectorIdxTy();
10817 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT,
10818 SVInVec, DAG.getConstant(OrigElt, IndexTy));
10822 bool BCNumEltsChanged = false;
10823 EVT ExtVT = VT.getVectorElementType();
10826 // If the result of load has to be truncated, then it's not necessarily
10828 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
10831 if (InVec.getOpcode() == ISD::BITCAST) {
10832 // Don't duplicate a load with other uses.
10833 if (!InVec.hasOneUse())
10836 EVT BCVT = InVec.getOperand(0).getValueType();
10837 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
10839 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
10840 BCNumEltsChanged = true;
10841 InVec = InVec.getOperand(0);
10842 ExtVT = BCVT.getVectorElementType();
10845 // (vextract (vN[if]M load $addr), i) -> ([if]M load $addr + i * size)
10846 if (!LegalOperations && !ConstEltNo && InVec.hasOneUse() &&
10847 ISD::isNormalLoad(InVec.getNode()) &&
10848 !N->getOperand(1)->hasPredecessor(InVec.getNode())) {
10849 SDValue Index = N->getOperand(1);
10850 if (LoadSDNode *OrigLoad = dyn_cast<LoadSDNode>(InVec))
10851 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, Index,
10855 // Perform only after legalization to ensure build_vector / vector_shuffle
10856 // optimizations have already been done.
10857 if (!LegalOperations) return SDValue();
10859 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
10860 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
10861 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
10864 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
10866 LoadSDNode *LN0 = nullptr;
10867 const ShuffleVectorSDNode *SVN = nullptr;
10868 if (ISD::isNormalLoad(InVec.getNode())) {
10869 LN0 = cast<LoadSDNode>(InVec);
10870 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
10871 InVec.getOperand(0).getValueType() == ExtVT &&
10872 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
10873 // Don't duplicate a load with other uses.
10874 if (!InVec.hasOneUse())
10877 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
10878 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
10879 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
10881 // (load $addr+1*size)
10883 // Don't duplicate a load with other uses.
10884 if (!InVec.hasOneUse())
10887 // If the bit convert changed the number of elements, it is unsafe
10888 // to examine the mask.
10889 if (BCNumEltsChanged)
10892 // Select the input vector, guarding against out of range extract vector.
10893 unsigned NumElems = VT.getVectorNumElements();
10894 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
10895 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
10897 if (InVec.getOpcode() == ISD::BITCAST) {
10898 // Don't duplicate a load with other uses.
10899 if (!InVec.hasOneUse())
10902 InVec = InVec.getOperand(0);
10904 if (ISD::isNormalLoad(InVec.getNode())) {
10905 LN0 = cast<LoadSDNode>(InVec);
10906 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
10907 EltNo = DAG.getConstant(Elt, EltNo.getValueType());
10911 // Make sure we found a non-volatile load and the extractelement is
10913 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
10916 // If Idx was -1 above, Elt is going to be -1, so just return undef.
10918 return DAG.getUNDEF(LVT);
10920 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, EltNo, LN0);
10926 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
10927 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
10928 // We perform this optimization post type-legalization because
10929 // the type-legalizer often scalarizes integer-promoted vectors.
10930 // Performing this optimization before may create bit-casts which
10931 // will be type-legalized to complex code sequences.
10932 // We perform this optimization only before the operation legalizer because we
10933 // may introduce illegal operations.
10934 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
10937 unsigned NumInScalars = N->getNumOperands();
10939 EVT VT = N->getValueType(0);
10941 // Check to see if this is a BUILD_VECTOR of a bunch of values
10942 // which come from any_extend or zero_extend nodes. If so, we can create
10943 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
10944 // optimizations. We do not handle sign-extend because we can't fill the sign
10946 EVT SourceType = MVT::Other;
10947 bool AllAnyExt = true;
10949 for (unsigned i = 0; i != NumInScalars; ++i) {
10950 SDValue In = N->getOperand(i);
10951 // Ignore undef inputs.
10952 if (In.getOpcode() == ISD::UNDEF) continue;
10954 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
10955 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
10957 // Abort if the element is not an extension.
10958 if (!ZeroExt && !AnyExt) {
10959 SourceType = MVT::Other;
10963 // The input is a ZeroExt or AnyExt. Check the original type.
10964 EVT InTy = In.getOperand(0).getValueType();
10966 // Check that all of the widened source types are the same.
10967 if (SourceType == MVT::Other)
10970 else if (InTy != SourceType) {
10971 // Multiple income types. Abort.
10972 SourceType = MVT::Other;
10976 // Check if all of the extends are ANY_EXTENDs.
10977 AllAnyExt &= AnyExt;
10980 // In order to have valid types, all of the inputs must be extended from the
10981 // same source type and all of the inputs must be any or zero extend.
10982 // Scalar sizes must be a power of two.
10983 EVT OutScalarTy = VT.getScalarType();
10984 bool ValidTypes = SourceType != MVT::Other &&
10985 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
10986 isPowerOf2_32(SourceType.getSizeInBits());
10988 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
10989 // turn into a single shuffle instruction.
10993 bool isLE = TLI.isLittleEndian();
10994 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
10995 assert(ElemRatio > 1 && "Invalid element size ratio");
10996 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
10997 DAG.getConstant(0, SourceType);
10999 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
11000 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
11002 // Populate the new build_vector
11003 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
11004 SDValue Cast = N->getOperand(i);
11005 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
11006 Cast.getOpcode() == ISD::ZERO_EXTEND ||
11007 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
11009 if (Cast.getOpcode() == ISD::UNDEF)
11010 In = DAG.getUNDEF(SourceType);
11012 In = Cast->getOperand(0);
11013 unsigned Index = isLE ? (i * ElemRatio) :
11014 (i * ElemRatio + (ElemRatio - 1));
11016 assert(Index < Ops.size() && "Invalid index");
11020 // The type of the new BUILD_VECTOR node.
11021 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
11022 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
11023 "Invalid vector size");
11024 // Check if the new vector type is legal.
11025 if (!isTypeLegal(VecVT)) return SDValue();
11027 // Make the new BUILD_VECTOR.
11028 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
11030 // The new BUILD_VECTOR node has the potential to be further optimized.
11031 AddToWorklist(BV.getNode());
11032 // Bitcast to the desired type.
11033 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
11036 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
11037 EVT VT = N->getValueType(0);
11039 unsigned NumInScalars = N->getNumOperands();
11042 EVT SrcVT = MVT::Other;
11043 unsigned Opcode = ISD::DELETED_NODE;
11044 unsigned NumDefs = 0;
11046 for (unsigned i = 0; i != NumInScalars; ++i) {
11047 SDValue In = N->getOperand(i);
11048 unsigned Opc = In.getOpcode();
11050 if (Opc == ISD::UNDEF)
11053 // If all scalar values are floats and converted from integers.
11054 if (Opcode == ISD::DELETED_NODE &&
11055 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
11062 EVT InVT = In.getOperand(0).getValueType();
11064 // If all scalar values are typed differently, bail out. It's chosen to
11065 // simplify BUILD_VECTOR of integer types.
11066 if (SrcVT == MVT::Other)
11073 // If the vector has just one element defined, it's not worth to fold it into
11074 // a vectorized one.
11078 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
11079 && "Should only handle conversion from integer to float.");
11080 assert(SrcVT != MVT::Other && "Cannot determine source type!");
11082 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
11084 if (!TLI.isOperationLegalOrCustom(Opcode, NVT))
11087 SmallVector<SDValue, 8> Opnds;
11088 for (unsigned i = 0; i != NumInScalars; ++i) {
11089 SDValue In = N->getOperand(i);
11091 if (In.getOpcode() == ISD::UNDEF)
11092 Opnds.push_back(DAG.getUNDEF(SrcVT));
11094 Opnds.push_back(In.getOperand(0));
11096 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Opnds);
11097 AddToWorklist(BV.getNode());
11099 return DAG.getNode(Opcode, dl, VT, BV);
11102 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
11103 unsigned NumInScalars = N->getNumOperands();
11105 EVT VT = N->getValueType(0);
11107 // A vector built entirely of undefs is undef.
11108 if (ISD::allOperandsUndef(N))
11109 return DAG.getUNDEF(VT);
11111 SDValue V = reduceBuildVecExtToExtBuildVec(N);
11115 V = reduceBuildVecConvertToConvertBuildVec(N);
11119 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
11120 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
11121 // at most two distinct vectors, turn this into a shuffle node.
11123 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
11124 if (!isTypeLegal(VT))
11127 // May only combine to shuffle after legalize if shuffle is legal.
11128 if (LegalOperations && !TLI.isOperationLegal(ISD::VECTOR_SHUFFLE, VT))
11131 SDValue VecIn1, VecIn2;
11132 bool UsesZeroVector = false;
11133 for (unsigned i = 0; i != NumInScalars; ++i) {
11134 SDValue Op = N->getOperand(i);
11135 // Ignore undef inputs.
11136 if (Op.getOpcode() == ISD::UNDEF) continue;
11138 // See if we can combine this build_vector into a blend with a zero vector.
11139 if (!VecIn2.getNode() && ((Op.getOpcode() == ISD::Constant &&
11140 cast<ConstantSDNode>(Op.getNode())->isNullValue()) ||
11141 (Op.getOpcode() == ISD::ConstantFP &&
11142 cast<ConstantFPSDNode>(Op.getNode())->getValueAPF().isZero()))) {
11143 UsesZeroVector = true;
11147 // If this input is something other than a EXTRACT_VECTOR_ELT with a
11148 // constant index, bail out.
11149 if (Op.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
11150 !isa<ConstantSDNode>(Op.getOperand(1))) {
11151 VecIn1 = VecIn2 = SDValue(nullptr, 0);
11155 // We allow up to two distinct input vectors.
11156 SDValue ExtractedFromVec = Op.getOperand(0);
11157 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
11160 if (!VecIn1.getNode()) {
11161 VecIn1 = ExtractedFromVec;
11162 } else if (!VecIn2.getNode() && !UsesZeroVector) {
11163 VecIn2 = ExtractedFromVec;
11165 // Too many inputs.
11166 VecIn1 = VecIn2 = SDValue(nullptr, 0);
11171 // If everything is good, we can make a shuffle operation.
11172 if (VecIn1.getNode()) {
11173 unsigned InNumElements = VecIn1.getValueType().getVectorNumElements();
11174 SmallVector<int, 8> Mask;
11175 for (unsigned i = 0; i != NumInScalars; ++i) {
11176 unsigned Opcode = N->getOperand(i).getOpcode();
11177 if (Opcode == ISD::UNDEF) {
11178 Mask.push_back(-1);
11182 // Operands can also be zero.
11183 if (Opcode != ISD::EXTRACT_VECTOR_ELT) {
11184 assert(UsesZeroVector &&
11185 (Opcode == ISD::Constant || Opcode == ISD::ConstantFP) &&
11186 "Unexpected node found!");
11187 Mask.push_back(NumInScalars+i);
11191 // If extracting from the first vector, just use the index directly.
11192 SDValue Extract = N->getOperand(i);
11193 SDValue ExtVal = Extract.getOperand(1);
11194 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
11195 if (Extract.getOperand(0) == VecIn1) {
11196 Mask.push_back(ExtIndex);
11200 // Otherwise, use InIdx + InputVecSize
11201 Mask.push_back(InNumElements + ExtIndex);
11204 // Avoid introducing illegal shuffles with zero.
11205 if (UsesZeroVector && !TLI.isVectorClearMaskLegal(Mask, VT))
11208 // We can't generate a shuffle node with mismatched input and output types.
11209 // Attempt to transform a single input vector to the correct type.
11210 if ((VT != VecIn1.getValueType())) {
11211 // If the input vector type has a different base type to the output
11212 // vector type, bail out.
11213 EVT VTElemType = VT.getVectorElementType();
11214 if ((VecIn1.getValueType().getVectorElementType() != VTElemType) ||
11215 (VecIn2.getNode() &&
11216 (VecIn2.getValueType().getVectorElementType() != VTElemType)))
11219 // If the input vector is too small, widen it.
11220 // We only support widening of vectors which are half the size of the
11221 // output registers. For example XMM->YMM widening on X86 with AVX.
11222 EVT VecInT = VecIn1.getValueType();
11223 if (VecInT.getSizeInBits() * 2 == VT.getSizeInBits()) {
11224 // If we only have one small input, widen it by adding undef values.
11225 if (!VecIn2.getNode())
11226 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, VecIn1,
11227 DAG.getUNDEF(VecIn1.getValueType()));
11228 else if (VecIn1.getValueType() == VecIn2.getValueType()) {
11229 // If we have two small inputs of the same type, try to concat them.
11230 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, VecIn1, VecIn2);
11231 VecIn2 = SDValue(nullptr, 0);
11234 } else if (VecInT.getSizeInBits() == VT.getSizeInBits() * 2) {
11235 // If the input vector is too large, try to split it.
11236 // We don't support having two input vectors that are too large.
11237 if (VecIn2.getNode())
11240 if (!TLI.isExtractSubvectorCheap(VT, VT.getVectorNumElements()))
11243 // Try to replace VecIn1 with two extract_subvectors
11244 // No need to update the masks, they should still be correct.
11245 VecIn2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, VecIn1,
11246 DAG.getConstant(VT.getVectorNumElements(), TLI.getVectorIdxTy()));
11247 VecIn1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, VecIn1,
11248 DAG.getConstant(0, TLI.getVectorIdxTy()));
11249 UsesZeroVector = false;
11254 if (UsesZeroVector)
11255 VecIn2 = VT.isInteger() ? DAG.getConstant(0, VT) :
11256 DAG.getConstantFP(0.0, VT);
11258 // If VecIn2 is unused then change it to undef.
11259 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
11261 // Check that we were able to transform all incoming values to the same
11263 if (VecIn2.getValueType() != VecIn1.getValueType() ||
11264 VecIn1.getValueType() != VT)
11267 // Return the new VECTOR_SHUFFLE node.
11271 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]);
11277 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
11278 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
11279 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
11280 // inputs come from at most two distinct vectors, turn this into a shuffle
11283 // If we only have one input vector, we don't need to do any concatenation.
11284 if (N->getNumOperands() == 1)
11285 return N->getOperand(0);
11287 // Check if all of the operands are undefs.
11288 EVT VT = N->getValueType(0);
11289 if (ISD::allOperandsUndef(N))
11290 return DAG.getUNDEF(VT);
11292 // Optimize concat_vectors where one of the vectors is undef.
11293 if (N->getNumOperands() == 2 &&
11294 N->getOperand(1)->getOpcode() == ISD::UNDEF) {
11295 SDValue In = N->getOperand(0);
11296 assert(In.getValueType().isVector() && "Must concat vectors");
11298 // Transform: concat_vectors(scalar, undef) -> scalar_to_vector(sclr).
11299 if (In->getOpcode() == ISD::BITCAST &&
11300 !In->getOperand(0)->getValueType(0).isVector()) {
11301 SDValue Scalar = In->getOperand(0);
11302 EVT SclTy = Scalar->getValueType(0);
11304 if (!SclTy.isFloatingPoint() && !SclTy.isInteger())
11307 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SclTy,
11308 VT.getSizeInBits() / SclTy.getSizeInBits());
11309 if (!TLI.isTypeLegal(NVT) || !TLI.isTypeLegal(Scalar.getValueType()))
11312 SDLoc dl = SDLoc(N);
11313 SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NVT, Scalar);
11314 return DAG.getNode(ISD::BITCAST, dl, VT, Res);
11318 // fold (concat_vectors (BUILD_VECTOR A, B, ...), (BUILD_VECTOR C, D, ...))
11319 // -> (BUILD_VECTOR A, B, ..., C, D, ...)
11320 if (N->getNumOperands() == 2 &&
11321 N->getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
11322 N->getOperand(1).getOpcode() == ISD::BUILD_VECTOR) {
11323 EVT VT = N->getValueType(0);
11324 SDValue N0 = N->getOperand(0);
11325 SDValue N1 = N->getOperand(1);
11326 SmallVector<SDValue, 8> Opnds;
11327 unsigned BuildVecNumElts = N0.getNumOperands();
11329 EVT SclTy0 = N0.getOperand(0)->getValueType(0);
11330 EVT SclTy1 = N1.getOperand(0)->getValueType(0);
11331 if (SclTy0.isFloatingPoint()) {
11332 for (unsigned i = 0; i != BuildVecNumElts; ++i)
11333 Opnds.push_back(N0.getOperand(i));
11334 for (unsigned i = 0; i != BuildVecNumElts; ++i)
11335 Opnds.push_back(N1.getOperand(i));
11337 // If BUILD_VECTOR are from built from integer, they may have different
11338 // operand types. Get the smaller type and truncate all operands to it.
11339 EVT MinTy = SclTy0.bitsLE(SclTy1) ? SclTy0 : SclTy1;
11340 for (unsigned i = 0; i != BuildVecNumElts; ++i)
11341 Opnds.push_back(DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinTy,
11342 N0.getOperand(i)));
11343 for (unsigned i = 0; i != BuildVecNumElts; ++i)
11344 Opnds.push_back(DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinTy,
11345 N1.getOperand(i)));
11348 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
11351 // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
11352 // nodes often generate nop CONCAT_VECTOR nodes.
11353 // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that
11354 // place the incoming vectors at the exact same location.
11355 SDValue SingleSource = SDValue();
11356 unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements();
11358 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
11359 SDValue Op = N->getOperand(i);
11361 if (Op.getOpcode() == ISD::UNDEF)
11364 // Check if this is the identity extract:
11365 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
11368 // Find the single incoming vector for the extract_subvector.
11369 if (SingleSource.getNode()) {
11370 if (Op.getOperand(0) != SingleSource)
11373 SingleSource = Op.getOperand(0);
11375 // Check the source type is the same as the type of the result.
11376 // If not, this concat may extend the vector, so we can not
11377 // optimize it away.
11378 if (SingleSource.getValueType() != N->getValueType(0))
11382 unsigned IdentityIndex = i * PartNumElem;
11383 ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1));
11384 // The extract index must be constant.
11388 // Check that we are reading from the identity index.
11389 if (CS->getZExtValue() != IdentityIndex)
11393 if (SingleSource.getNode())
11394 return SingleSource;
11399 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
11400 EVT NVT = N->getValueType(0);
11401 SDValue V = N->getOperand(0);
11403 if (V->getOpcode() == ISD::CONCAT_VECTORS) {
11405 // (extract_subvec (concat V1, V2, ...), i)
11408 // Only operand 0 is checked as 'concat' assumes all inputs of the same
11410 if (V->getOperand(0).getValueType() != NVT)
11412 unsigned Idx = dyn_cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
11413 unsigned NumElems = NVT.getVectorNumElements();
11414 assert((Idx % NumElems) == 0 &&
11415 "IDX in concat is not a multiple of the result vector length.");
11416 return V->getOperand(Idx / NumElems);
11420 if (V->getOpcode() == ISD::BITCAST)
11421 V = V.getOperand(0);
11423 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
11425 // Handle only simple case where vector being inserted and vector
11426 // being extracted are of same type, and are half size of larger vectors.
11427 EVT BigVT = V->getOperand(0).getValueType();
11428 EVT SmallVT = V->getOperand(1).getValueType();
11429 if (!NVT.bitsEq(SmallVT) || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
11432 // Only handle cases where both indexes are constants with the same type.
11433 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
11434 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
11436 if (InsIdx && ExtIdx &&
11437 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
11438 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
11440 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
11442 // indices are equal or bit offsets are equal => V1
11443 // otherwise => (extract_subvec V1, ExtIdx)
11444 if (InsIdx->getZExtValue() * SmallVT.getScalarType().getSizeInBits() ==
11445 ExtIdx->getZExtValue() * NVT.getScalarType().getSizeInBits())
11446 return DAG.getNode(ISD::BITCAST, dl, NVT, V->getOperand(1));
11447 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT,
11448 DAG.getNode(ISD::BITCAST, dl,
11449 N->getOperand(0).getValueType(),
11450 V->getOperand(0)), N->getOperand(1));
11457 static SDValue simplifyShuffleOperandRecursively(SmallBitVector &UsedElements,
11458 SDValue V, SelectionDAG &DAG) {
11460 EVT VT = V.getValueType();
11462 switch (V.getOpcode()) {
11466 case ISD::CONCAT_VECTORS: {
11467 EVT OpVT = V->getOperand(0).getValueType();
11468 int OpSize = OpVT.getVectorNumElements();
11469 SmallBitVector OpUsedElements(OpSize, false);
11470 bool FoundSimplification = false;
11471 SmallVector<SDValue, 4> NewOps;
11472 NewOps.reserve(V->getNumOperands());
11473 for (int i = 0, NumOps = V->getNumOperands(); i < NumOps; ++i) {
11474 SDValue Op = V->getOperand(i);
11475 bool OpUsed = false;
11476 for (int j = 0; j < OpSize; ++j)
11477 if (UsedElements[i * OpSize + j]) {
11478 OpUsedElements[j] = true;
11482 OpUsed ? simplifyShuffleOperandRecursively(OpUsedElements, Op, DAG)
11483 : DAG.getUNDEF(OpVT));
11484 FoundSimplification |= Op == NewOps.back();
11485 OpUsedElements.reset();
11487 if (FoundSimplification)
11488 V = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, NewOps);
11492 case ISD::INSERT_SUBVECTOR: {
11493 SDValue BaseV = V->getOperand(0);
11494 SDValue SubV = V->getOperand(1);
11495 auto *IdxN = dyn_cast<ConstantSDNode>(V->getOperand(2));
11499 int SubSize = SubV.getValueType().getVectorNumElements();
11500 int Idx = IdxN->getZExtValue();
11501 bool SubVectorUsed = false;
11502 SmallBitVector SubUsedElements(SubSize, false);
11503 for (int i = 0; i < SubSize; ++i)
11504 if (UsedElements[i + Idx]) {
11505 SubVectorUsed = true;
11506 SubUsedElements[i] = true;
11507 UsedElements[i + Idx] = false;
11510 // Now recurse on both the base and sub vectors.
11511 SDValue SimplifiedSubV =
11513 ? simplifyShuffleOperandRecursively(SubUsedElements, SubV, DAG)
11514 : DAG.getUNDEF(SubV.getValueType());
11515 SDValue SimplifiedBaseV = simplifyShuffleOperandRecursively(UsedElements, BaseV, DAG);
11516 if (SimplifiedSubV != SubV || SimplifiedBaseV != BaseV)
11517 V = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT,
11518 SimplifiedBaseV, SimplifiedSubV, V->getOperand(2));
11524 static SDValue simplifyShuffleOperands(ShuffleVectorSDNode *SVN, SDValue N0,
11525 SDValue N1, SelectionDAG &DAG) {
11526 EVT VT = SVN->getValueType(0);
11527 int NumElts = VT.getVectorNumElements();
11528 SmallBitVector N0UsedElements(NumElts, false), N1UsedElements(NumElts, false);
11529 for (int M : SVN->getMask())
11530 if (M >= 0 && M < NumElts)
11531 N0UsedElements[M] = true;
11532 else if (M >= NumElts)
11533 N1UsedElements[M - NumElts] = true;
11535 SDValue S0 = simplifyShuffleOperandRecursively(N0UsedElements, N0, DAG);
11536 SDValue S1 = simplifyShuffleOperandRecursively(N1UsedElements, N1, DAG);
11537 if (S0 == N0 && S1 == N1)
11540 return DAG.getVectorShuffle(VT, SDLoc(SVN), S0, S1, SVN->getMask());
11543 // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat,
11544 // or turn a shuffle of a single concat into simpler shuffle then concat.
11545 static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
11546 EVT VT = N->getValueType(0);
11547 unsigned NumElts = VT.getVectorNumElements();
11549 SDValue N0 = N->getOperand(0);
11550 SDValue N1 = N->getOperand(1);
11551 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
11553 SmallVector<SDValue, 4> Ops;
11554 EVT ConcatVT = N0.getOperand(0).getValueType();
11555 unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements();
11556 unsigned NumConcats = NumElts / NumElemsPerConcat;
11558 // Special case: shuffle(concat(A,B)) can be more efficiently represented
11559 // as concat(shuffle(A,B),UNDEF) if the shuffle doesn't set any of the high
11560 // half vector elements.
11561 if (NumElemsPerConcat * 2 == NumElts && N1.getOpcode() == ISD::UNDEF &&
11562 std::all_of(SVN->getMask().begin() + NumElemsPerConcat,
11563 SVN->getMask().end(), [](int i) { return i == -1; })) {
11564 N0 = DAG.getVectorShuffle(ConcatVT, SDLoc(N), N0.getOperand(0), N0.getOperand(1),
11565 ArrayRef<int>(SVN->getMask().begin(), NumElemsPerConcat));
11566 N1 = DAG.getUNDEF(ConcatVT);
11567 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, N0, N1);
11570 // Look at every vector that's inserted. We're looking for exact
11571 // subvector-sized copies from a concatenated vector
11572 for (unsigned I = 0; I != NumConcats; ++I) {
11573 // Make sure we're dealing with a copy.
11574 unsigned Begin = I * NumElemsPerConcat;
11575 bool AllUndef = true, NoUndef = true;
11576 for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) {
11577 if (SVN->getMaskElt(J) >= 0)
11584 if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0)
11587 for (unsigned J = 1; J != NumElemsPerConcat; ++J)
11588 if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J))
11591 unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat;
11592 if (FirstElt < N0.getNumOperands())
11593 Ops.push_back(N0.getOperand(FirstElt));
11595 Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands()));
11597 } else if (AllUndef) {
11598 Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType()));
11599 } else { // Mixed with general masks and undefs, can't do optimization.
11604 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops);
11607 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
11608 EVT VT = N->getValueType(0);
11609 unsigned NumElts = VT.getVectorNumElements();
11611 SDValue N0 = N->getOperand(0);
11612 SDValue N1 = N->getOperand(1);
11614 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
11616 // Canonicalize shuffle undef, undef -> undef
11617 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
11618 return DAG.getUNDEF(VT);
11620 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
11622 // Canonicalize shuffle v, v -> v, undef
11624 SmallVector<int, 8> NewMask;
11625 for (unsigned i = 0; i != NumElts; ++i) {
11626 int Idx = SVN->getMaskElt(i);
11627 if (Idx >= (int)NumElts) Idx -= NumElts;
11628 NewMask.push_back(Idx);
11630 return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT),
11634 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
11635 if (N0.getOpcode() == ISD::UNDEF) {
11636 SmallVector<int, 8> NewMask;
11637 for (unsigned i = 0; i != NumElts; ++i) {
11638 int Idx = SVN->getMaskElt(i);
11640 if (Idx >= (int)NumElts)
11643 Idx = -1; // remove reference to lhs
11645 NewMask.push_back(Idx);
11647 return DAG.getVectorShuffle(VT, SDLoc(N), N1, DAG.getUNDEF(VT),
11651 // Remove references to rhs if it is undef
11652 if (N1.getOpcode() == ISD::UNDEF) {
11653 bool Changed = false;
11654 SmallVector<int, 8> NewMask;
11655 for (unsigned i = 0; i != NumElts; ++i) {
11656 int Idx = SVN->getMaskElt(i);
11657 if (Idx >= (int)NumElts) {
11661 NewMask.push_back(Idx);
11664 return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, &NewMask[0]);
11667 // If it is a splat, check if the argument vector is another splat or a
11669 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
11670 SDNode *V = N0.getNode();
11672 // If this is a bit convert that changes the element type of the vector but
11673 // not the number of vector elements, look through it. Be careful not to
11674 // look though conversions that change things like v4f32 to v2f64.
11675 if (V->getOpcode() == ISD::BITCAST) {
11676 SDValue ConvInput = V->getOperand(0);
11677 if (ConvInput.getValueType().isVector() &&
11678 ConvInput.getValueType().getVectorNumElements() == NumElts)
11679 V = ConvInput.getNode();
11682 if (V->getOpcode() == ISD::BUILD_VECTOR) {
11683 assert(V->getNumOperands() == NumElts &&
11684 "BUILD_VECTOR has wrong number of operands");
11686 bool AllSame = true;
11687 for (unsigned i = 0; i != NumElts; ++i) {
11688 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
11689 Base = V->getOperand(i);
11693 // Splat of <u, u, u, u>, return <u, u, u, u>
11694 if (!Base.getNode())
11696 for (unsigned i = 0; i != NumElts; ++i) {
11697 if (V->getOperand(i) != Base) {
11702 // Splat of <x, x, x, x>, return <x, x, x, x>
11706 // If the splatted element is a constant, just build the vector out of
11707 // constants directly.
11708 const SDValue &Splatted = V->getOperand(SVN->getSplatIndex());
11709 if (isa<ConstantSDNode>(Splatted) || isa<ConstantFPSDNode>(Splatted)) {
11710 SmallVector<SDValue, 8> Ops;
11711 for (unsigned i = 0; i != NumElts; ++i) {
11712 Ops.push_back(Splatted);
11714 SDValue NewBV = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
11715 V->getValueType(0), Ops);
11717 // We may have jumped through bitcasts, so the type of the
11718 // BUILD_VECTOR may not match the type of the shuffle.
11719 if (V->getValueType(0) != VT)
11720 NewBV = DAG.getNode(ISD::BITCAST, SDLoc(N), VT, NewBV);
11726 // There are various patterns used to build up a vector from smaller vectors,
11727 // subvectors, or elements. Scan chains of these and replace unused insertions
11728 // or components with undef.
11729 if (SDValue S = simplifyShuffleOperands(SVN, N0, N1, DAG))
11732 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
11733 Level < AfterLegalizeVectorOps &&
11734 (N1.getOpcode() == ISD::UNDEF ||
11735 (N1.getOpcode() == ISD::CONCAT_VECTORS &&
11736 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) {
11737 SDValue V = partitionShuffleOfConcats(N, DAG);
11743 // Canonicalize shuffles according to rules:
11744 // shuffle(A, shuffle(A, B)) -> shuffle(shuffle(A,B), A)
11745 // shuffle(B, shuffle(A, B)) -> shuffle(shuffle(A,B), B)
11746 // shuffle(B, shuffle(A, Undef)) -> shuffle(shuffle(A, Undef), B)
11747 if (N1.getOpcode() == ISD::VECTOR_SHUFFLE &&
11748 N0.getOpcode() != ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
11749 TLI.isTypeLegal(VT)) {
11750 // The incoming shuffle must be of the same type as the result of the
11751 // current shuffle.
11752 assert(N1->getOperand(0).getValueType() == VT &&
11753 "Shuffle types don't match");
11755 SDValue SV0 = N1->getOperand(0);
11756 SDValue SV1 = N1->getOperand(1);
11757 bool HasSameOp0 = N0 == SV0;
11758 bool IsSV1Undef = SV1.getOpcode() == ISD::UNDEF;
11759 if (HasSameOp0 || IsSV1Undef || N0 == SV1)
11760 // Commute the operands of this shuffle so that next rule
11762 return DAG.getCommutedVectorShuffle(*SVN);
11765 // Try to fold according to rules:
11766 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2)
11767 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2)
11768 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2)
11769 // Don't try to fold shuffles with illegal type.
11770 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
11771 TLI.isTypeLegal(VT)) {
11772 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
11774 // The incoming shuffle must be of the same type as the result of the
11775 // current shuffle.
11776 assert(OtherSV->getOperand(0).getValueType() == VT &&
11777 "Shuffle types don't match");
11780 SmallVector<int, 4> Mask;
11781 // Compute the combined shuffle mask for a shuffle with SV0 as the first
11782 // operand, and SV1 as the second operand.
11783 for (unsigned i = 0; i != NumElts; ++i) {
11784 int Idx = SVN->getMaskElt(i);
11786 // Propagate Undef.
11787 Mask.push_back(Idx);
11791 SDValue CurrentVec;
11792 if (Idx < (int)NumElts) {
11793 // This shuffle index refers to the inner shuffle N0. Lookup the inner
11794 // shuffle mask to identify which vector is actually referenced.
11795 Idx = OtherSV->getMaskElt(Idx);
11797 // Propagate Undef.
11798 Mask.push_back(Idx);
11802 CurrentVec = (Idx < (int) NumElts) ? OtherSV->getOperand(0)
11803 : OtherSV->getOperand(1);
11805 // This shuffle index references an element within N1.
11809 // Simple case where 'CurrentVec' is UNDEF.
11810 if (CurrentVec.getOpcode() == ISD::UNDEF) {
11811 Mask.push_back(-1);
11815 // Canonicalize the shuffle index. We don't know yet if CurrentVec
11816 // will be the first or second operand of the combined shuffle.
11817 Idx = Idx % NumElts;
11818 if (!SV0.getNode() || SV0 == CurrentVec) {
11819 // Ok. CurrentVec is the left hand side.
11820 // Update the mask accordingly.
11822 Mask.push_back(Idx);
11826 // Bail out if we cannot convert the shuffle pair into a single shuffle.
11827 if (SV1.getNode() && SV1 != CurrentVec)
11830 // Ok. CurrentVec is the right hand side.
11831 // Update the mask accordingly.
11833 Mask.push_back(Idx + NumElts);
11836 // Check if all indices in Mask are Undef. In case, propagate Undef.
11837 bool isUndefMask = true;
11838 for (unsigned i = 0; i != NumElts && isUndefMask; ++i)
11839 isUndefMask &= Mask[i] < 0;
11842 return DAG.getUNDEF(VT);
11844 if (!SV0.getNode())
11845 SV0 = DAG.getUNDEF(VT);
11846 if (!SV1.getNode())
11847 SV1 = DAG.getUNDEF(VT);
11849 // Avoid introducing shuffles with illegal mask.
11850 if (!TLI.isShuffleMaskLegal(Mask, VT)) {
11851 // Compute the commuted shuffle mask and test again.
11852 for (unsigned i = 0; i != NumElts; ++i) {
11856 else if (idx < (int)NumElts)
11857 Mask[i] = idx + NumElts;
11859 Mask[i] = idx - NumElts;
11862 if (!TLI.isShuffleMaskLegal(Mask, VT))
11865 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, A, M2)
11866 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, A, M2)
11867 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, B, M2)
11868 std::swap(SV0, SV1);
11871 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2)
11872 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2)
11873 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2)
11874 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, &Mask[0]);
11880 SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
11881 SDValue N0 = N->getOperand(0);
11882 SDValue N2 = N->getOperand(2);
11884 // If the input vector is a concatenation, and the insert replaces
11885 // one of the halves, we can optimize into a single concat_vectors.
11886 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
11887 N0->getNumOperands() == 2 && N2.getOpcode() == ISD::Constant) {
11888 APInt InsIdx = cast<ConstantSDNode>(N2)->getAPIntValue();
11889 EVT VT = N->getValueType(0);
11891 // Lower half: fold (insert_subvector (concat_vectors X, Y), Z) ->
11892 // (concat_vectors Z, Y)
11894 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
11895 N->getOperand(1), N0.getOperand(1));
11897 // Upper half: fold (insert_subvector (concat_vectors X, Y), Z) ->
11898 // (concat_vectors X, Z)
11899 if (InsIdx == VT.getVectorNumElements()/2)
11900 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
11901 N0.getOperand(0), N->getOperand(1));
11907 /// Returns a vector_shuffle if it able to transform an AND to a vector_shuffle
11908 /// with the destination vector and a zero vector.
11909 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
11910 /// vector_shuffle V, Zero, <0, 4, 2, 4>
11911 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
11912 EVT VT = N->getValueType(0);
11914 SDValue LHS = N->getOperand(0);
11915 SDValue RHS = N->getOperand(1);
11916 if (N->getOpcode() == ISD::AND) {
11917 if (RHS.getOpcode() == ISD::BITCAST)
11918 RHS = RHS.getOperand(0);
11919 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
11920 SmallVector<int, 8> Indices;
11921 unsigned NumElts = RHS.getNumOperands();
11922 for (unsigned i = 0; i != NumElts; ++i) {
11923 SDValue Elt = RHS.getOperand(i);
11924 if (!isa<ConstantSDNode>(Elt))
11927 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
11928 Indices.push_back(i);
11929 else if (cast<ConstantSDNode>(Elt)->isNullValue())
11930 Indices.push_back(NumElts+i);
11935 // Let's see if the target supports this vector_shuffle.
11936 EVT RVT = RHS.getValueType();
11937 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
11940 // Return the new VECTOR_SHUFFLE node.
11941 EVT EltVT = RVT.getVectorElementType();
11942 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
11943 DAG.getConstant(0, EltVT));
11944 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), RVT, ZeroOps);
11945 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
11946 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
11947 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
11954 /// Visit a binary vector operation, like ADD.
11955 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
11956 assert(N->getValueType(0).isVector() &&
11957 "SimplifyVBinOp only works on vectors!");
11959 SDValue LHS = N->getOperand(0);
11960 SDValue RHS = N->getOperand(1);
11961 SDValue Shuffle = XformToShuffleWithZero(N);
11962 if (Shuffle.getNode()) return Shuffle;
11964 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
11966 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
11967 RHS.getOpcode() == ISD::BUILD_VECTOR) {
11968 // Check if both vectors are constants. If not bail out.
11969 if (!(cast<BuildVectorSDNode>(LHS)->isConstant() &&
11970 cast<BuildVectorSDNode>(RHS)->isConstant()))
11973 SmallVector<SDValue, 8> Ops;
11974 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
11975 SDValue LHSOp = LHS.getOperand(i);
11976 SDValue RHSOp = RHS.getOperand(i);
11978 // Can't fold divide by zero.
11979 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
11980 N->getOpcode() == ISD::FDIV) {
11981 if ((RHSOp.getOpcode() == ISD::Constant &&
11982 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
11983 (RHSOp.getOpcode() == ISD::ConstantFP &&
11984 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
11988 EVT VT = LHSOp.getValueType();
11989 EVT RVT = RHSOp.getValueType();
11991 // Integer BUILD_VECTOR operands may have types larger than the element
11992 // size (e.g., when the element type is not legal). Prior to type
11993 // legalization, the types may not match between the two BUILD_VECTORS.
11994 // Truncate one of the operands to make them match.
11995 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
11996 RHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, RHSOp);
11998 LHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), RVT, LHSOp);
12002 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(LHS), VT,
12004 if (FoldOp.getOpcode() != ISD::UNDEF &&
12005 FoldOp.getOpcode() != ISD::Constant &&
12006 FoldOp.getOpcode() != ISD::ConstantFP)
12008 Ops.push_back(FoldOp);
12009 AddToWorklist(FoldOp.getNode());
12012 if (Ops.size() == LHS.getNumOperands())
12013 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), LHS.getValueType(), Ops);
12016 // Type legalization might introduce new shuffles in the DAG.
12017 // Fold (VBinOp (shuffle (A, Undef, Mask)), (shuffle (B, Undef, Mask)))
12018 // -> (shuffle (VBinOp (A, B)), Undef, Mask).
12019 if (LegalTypes && isa<ShuffleVectorSDNode>(LHS) &&
12020 isa<ShuffleVectorSDNode>(RHS) && LHS.hasOneUse() && RHS.hasOneUse() &&
12021 LHS.getOperand(1).getOpcode() == ISD::UNDEF &&
12022 RHS.getOperand(1).getOpcode() == ISD::UNDEF) {
12023 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(LHS);
12024 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(RHS);
12026 if (SVN0->getMask().equals(SVN1->getMask())) {
12027 EVT VT = N->getValueType(0);
12028 SDValue UndefVector = LHS.getOperand(1);
12029 SDValue NewBinOp = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
12030 LHS.getOperand(0), RHS.getOperand(0));
12031 AddUsersToWorklist(N);
12032 return DAG.getVectorShuffle(VT, SDLoc(N), NewBinOp, UndefVector,
12033 &SVN0->getMask()[0]);
12040 /// Visit a binary vector operation, like FABS/FNEG.
12041 SDValue DAGCombiner::SimplifyVUnaryOp(SDNode *N) {
12042 assert(N->getValueType(0).isVector() &&
12043 "SimplifyVUnaryOp only works on vectors!");
12045 SDValue N0 = N->getOperand(0);
12047 if (N0.getOpcode() != ISD::BUILD_VECTOR)
12050 // Operand is a BUILD_VECTOR node, see if we can constant fold it.
12051 SmallVector<SDValue, 8> Ops;
12052 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
12053 SDValue Op = N0.getOperand(i);
12054 if (Op.getOpcode() != ISD::UNDEF &&
12055 Op.getOpcode() != ISD::ConstantFP)
12057 EVT EltVT = Op.getValueType();
12058 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(N0), EltVT, Op);
12059 if (FoldOp.getOpcode() != ISD::UNDEF &&
12060 FoldOp.getOpcode() != ISD::ConstantFP)
12062 Ops.push_back(FoldOp);
12063 AddToWorklist(FoldOp.getNode());
12066 if (Ops.size() != N0.getNumOperands())
12069 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), N0.getValueType(), Ops);
12072 SDValue DAGCombiner::SimplifySelect(SDLoc DL, SDValue N0,
12073 SDValue N1, SDValue N2){
12074 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
12076 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
12077 cast<CondCodeSDNode>(N0.getOperand(2))->get());
12079 // If we got a simplified select_cc node back from SimplifySelectCC, then
12080 // break it down into a new SETCC node, and a new SELECT node, and then return
12081 // the SELECT node, since we were called with a SELECT node.
12082 if (SCC.getNode()) {
12083 // Check to see if we got a select_cc back (to turn into setcc/select).
12084 // Otherwise, just return whatever node we got back, like fabs.
12085 if (SCC.getOpcode() == ISD::SELECT_CC) {
12086 SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0),
12088 SCC.getOperand(0), SCC.getOperand(1),
12089 SCC.getOperand(4));
12090 AddToWorklist(SETCC.getNode());
12091 return DAG.getSelect(SDLoc(SCC), SCC.getValueType(), SETCC,
12092 SCC.getOperand(2), SCC.getOperand(3));
12100 /// Given a SELECT or a SELECT_CC node, where LHS and RHS are the two values
12101 /// being selected between, see if we can simplify the select. Callers of this
12102 /// should assume that TheSelect is deleted if this returns true. As such, they
12103 /// should return the appropriate thing (e.g. the node) back to the top-level of
12104 /// the DAG combiner loop to avoid it being looked at.
12105 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
12108 // Cannot simplify select with vector condition
12109 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
12111 // If this is a select from two identical things, try to pull the operation
12112 // through the select.
12113 if (LHS.getOpcode() != RHS.getOpcode() ||
12114 !LHS.hasOneUse() || !RHS.hasOneUse())
12117 // If this is a load and the token chain is identical, replace the select
12118 // of two loads with a load through a select of the address to load from.
12119 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
12120 // constants have been dropped into the constant pool.
12121 if (LHS.getOpcode() == ISD::LOAD) {
12122 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
12123 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
12125 // Token chains must be identical.
12126 if (LHS.getOperand(0) != RHS.getOperand(0) ||
12127 // Do not let this transformation reduce the number of volatile loads.
12128 LLD->isVolatile() || RLD->isVolatile() ||
12129 // If this is an EXTLOAD, the VT's must match.
12130 LLD->getMemoryVT() != RLD->getMemoryVT() ||
12131 // If this is an EXTLOAD, the kind of extension must match.
12132 (LLD->getExtensionType() != RLD->getExtensionType() &&
12133 // The only exception is if one of the extensions is anyext.
12134 LLD->getExtensionType() != ISD::EXTLOAD &&
12135 RLD->getExtensionType() != ISD::EXTLOAD) ||
12136 // FIXME: this discards src value information. This is
12137 // over-conservative. It would be beneficial to be able to remember
12138 // both potential memory locations. Since we are discarding
12139 // src value info, don't do the transformation if the memory
12140 // locations are not in the default address space.
12141 LLD->getPointerInfo().getAddrSpace() != 0 ||
12142 RLD->getPointerInfo().getAddrSpace() != 0 ||
12143 !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(),
12144 LLD->getBasePtr().getValueType()))
12147 // Check that the select condition doesn't reach either load. If so,
12148 // folding this will induce a cycle into the DAG. If not, this is safe to
12149 // xform, so create a select of the addresses.
12151 if (TheSelect->getOpcode() == ISD::SELECT) {
12152 SDNode *CondNode = TheSelect->getOperand(0).getNode();
12153 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
12154 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
12156 // The loads must not depend on one another.
12157 if (LLD->isPredecessorOf(RLD) ||
12158 RLD->isPredecessorOf(LLD))
12160 Addr = DAG.getSelect(SDLoc(TheSelect),
12161 LLD->getBasePtr().getValueType(),
12162 TheSelect->getOperand(0), LLD->getBasePtr(),
12163 RLD->getBasePtr());
12164 } else { // Otherwise SELECT_CC
12165 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
12166 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
12168 if ((LLD->hasAnyUseOfValue(1) &&
12169 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
12170 (RLD->hasAnyUseOfValue(1) &&
12171 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
12174 Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect),
12175 LLD->getBasePtr().getValueType(),
12176 TheSelect->getOperand(0),
12177 TheSelect->getOperand(1),
12178 LLD->getBasePtr(), RLD->getBasePtr(),
12179 TheSelect->getOperand(4));
12183 // It is safe to replace the two loads if they have different alignments,
12184 // but the new load must be the minimum (most restrictive) alignment of the
12186 bool isInvariant = LLD->isInvariant() & RLD->isInvariant();
12187 unsigned Alignment = std::min(LLD->getAlignment(), RLD->getAlignment());
12188 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
12189 Load = DAG.getLoad(TheSelect->getValueType(0),
12191 // FIXME: Discards pointer and AA info.
12192 LLD->getChain(), Addr, MachinePointerInfo(),
12193 LLD->isVolatile(), LLD->isNonTemporal(),
12194 isInvariant, Alignment);
12196 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
12197 RLD->getExtensionType() : LLD->getExtensionType(),
12199 TheSelect->getValueType(0),
12200 // FIXME: Discards pointer and AA info.
12201 LLD->getChain(), Addr, MachinePointerInfo(),
12202 LLD->getMemoryVT(), LLD->isVolatile(),
12203 LLD->isNonTemporal(), isInvariant, Alignment);
12206 // Users of the select now use the result of the load.
12207 CombineTo(TheSelect, Load);
12209 // Users of the old loads now use the new load's chain. We know the
12210 // old-load value is dead now.
12211 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
12212 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
12219 /// Simplify an expression of the form (N0 cond N1) ? N2 : N3
12220 /// where 'cond' is the comparison specified by CC.
12221 SDValue DAGCombiner::SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1,
12222 SDValue N2, SDValue N3,
12223 ISD::CondCode CC, bool NotExtCompare) {
12224 // (x ? y : y) -> y.
12225 if (N2 == N3) return N2;
12227 EVT VT = N2.getValueType();
12228 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
12229 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
12230 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
12232 // Determine if the condition we're dealing with is constant
12233 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
12234 N0, N1, CC, DL, false);
12235 if (SCC.getNode()) AddToWorklist(SCC.getNode());
12236 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
12238 // fold select_cc true, x, y -> x
12239 if (SCCC && !SCCC->isNullValue())
12241 // fold select_cc false, x, y -> y
12242 if (SCCC && SCCC->isNullValue())
12245 // Check to see if we can simplify the select into an fabs node
12246 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
12247 // Allow either -0.0 or 0.0
12248 if (CFP->getValueAPF().isZero()) {
12249 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
12250 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
12251 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
12252 N2 == N3.getOperand(0))
12253 return DAG.getNode(ISD::FABS, DL, VT, N0);
12255 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
12256 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
12257 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
12258 N2.getOperand(0) == N3)
12259 return DAG.getNode(ISD::FABS, DL, VT, N3);
12263 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
12264 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
12265 // in it. This is a win when the constant is not otherwise available because
12266 // it replaces two constant pool loads with one. We only do this if the FP
12267 // type is known to be legal, because if it isn't, then we are before legalize
12268 // types an we want the other legalization to happen first (e.g. to avoid
12269 // messing with soft float) and if the ConstantFP is not legal, because if
12270 // it is legal, we may not need to store the FP constant in a constant pool.
12271 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
12272 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
12273 if (TLI.isTypeLegal(N2.getValueType()) &&
12274 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
12275 TargetLowering::Legal &&
12276 !TLI.isFPImmLegal(TV->getValueAPF(), TV->getValueType(0)) &&
12277 !TLI.isFPImmLegal(FV->getValueAPF(), FV->getValueType(0))) &&
12278 // If both constants have multiple uses, then we won't need to do an
12279 // extra load, they are likely around in registers for other users.
12280 (TV->hasOneUse() || FV->hasOneUse())) {
12281 Constant *Elts[] = {
12282 const_cast<ConstantFP*>(FV->getConstantFPValue()),
12283 const_cast<ConstantFP*>(TV->getConstantFPValue())
12285 Type *FPTy = Elts[0]->getType();
12286 const DataLayout &TD = *TLI.getDataLayout();
12288 // Create a ConstantArray of the two constants.
12289 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
12290 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
12291 TD.getPrefTypeAlignment(FPTy));
12292 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
12294 // Get the offsets to the 0 and 1 element of the array so that we can
12295 // select between them.
12296 SDValue Zero = DAG.getIntPtrConstant(0);
12297 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
12298 SDValue One = DAG.getIntPtrConstant(EltSize);
12300 SDValue Cond = DAG.getSetCC(DL,
12301 getSetCCResultType(N0.getValueType()),
12303 AddToWorklist(Cond.getNode());
12304 SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(),
12306 AddToWorklist(CstOffset.getNode());
12307 CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx,
12309 AddToWorklist(CPIdx.getNode());
12310 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
12311 MachinePointerInfo::getConstantPool(), false,
12312 false, false, Alignment);
12317 // Check to see if we can perform the "gzip trick", transforming
12318 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
12319 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
12320 (N1C->isNullValue() || // (a < 0) ? b : 0
12321 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
12322 EVT XType = N0.getValueType();
12323 EVT AType = N2.getValueType();
12324 if (XType.bitsGE(AType)) {
12325 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
12326 // single-bit constant.
12327 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
12328 unsigned ShCtV = N2C->getAPIntValue().logBase2();
12329 ShCtV = XType.getSizeInBits()-ShCtV-1;
12330 SDValue ShCt = DAG.getConstant(ShCtV,
12331 getShiftAmountTy(N0.getValueType()));
12332 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0),
12334 AddToWorklist(Shift.getNode());
12336 if (XType.bitsGT(AType)) {
12337 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
12338 AddToWorklist(Shift.getNode());
12341 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
12344 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0),
12346 DAG.getConstant(XType.getSizeInBits()-1,
12347 getShiftAmountTy(N0.getValueType())));
12348 AddToWorklist(Shift.getNode());
12350 if (XType.bitsGT(AType)) {
12351 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
12352 AddToWorklist(Shift.getNode());
12355 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
12359 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
12360 // where y is has a single bit set.
12361 // A plaintext description would be, we can turn the SELECT_CC into an AND
12362 // when the condition can be materialized as an all-ones register. Any
12363 // single bit-test can be materialized as an all-ones register with
12364 // shift-left and shift-right-arith.
12365 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
12366 N0->getValueType(0) == VT &&
12367 N1C && N1C->isNullValue() &&
12368 N2C && N2C->isNullValue()) {
12369 SDValue AndLHS = N0->getOperand(0);
12370 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
12371 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
12372 // Shift the tested bit over the sign bit.
12373 APInt AndMask = ConstAndRHS->getAPIntValue();
12375 DAG.getConstant(AndMask.countLeadingZeros(),
12376 getShiftAmountTy(AndLHS.getValueType()));
12377 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt);
12379 // Now arithmetic right shift it all the way over, so the result is either
12380 // all-ones, or zero.
12382 DAG.getConstant(AndMask.getBitWidth()-1,
12383 getShiftAmountTy(Shl.getValueType()));
12384 SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt);
12386 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
12390 // fold select C, 16, 0 -> shl C, 4
12391 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
12392 TLI.getBooleanContents(N0.getValueType()) ==
12393 TargetLowering::ZeroOrOneBooleanContent) {
12395 // If the caller doesn't want us to simplify this into a zext of a compare,
12397 if (NotExtCompare && N2C->getAPIntValue() == 1)
12400 // Get a SetCC of the condition
12401 // NOTE: Don't create a SETCC if it's not legal on this target.
12402 if (!LegalOperations ||
12403 TLI.isOperationLegal(ISD::SETCC,
12404 LegalTypes ? getSetCCResultType(N0.getValueType()) : MVT::i1)) {
12406 // cast from setcc result type to select result type
12408 SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()),
12410 if (N2.getValueType().bitsLT(SCC.getValueType()))
12411 Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2),
12412 N2.getValueType());
12414 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
12415 N2.getValueType(), SCC);
12417 SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC);
12418 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
12419 N2.getValueType(), SCC);
12422 AddToWorklist(SCC.getNode());
12423 AddToWorklist(Temp.getNode());
12425 if (N2C->getAPIntValue() == 1)
12428 // shl setcc result by log2 n2c
12429 return DAG.getNode(
12430 ISD::SHL, DL, N2.getValueType(), Temp,
12431 DAG.getConstant(N2C->getAPIntValue().logBase2(),
12432 getShiftAmountTy(Temp.getValueType())));
12436 // Check to see if this is the equivalent of setcc
12437 // FIXME: Turn all of these into setcc if setcc if setcc is legal
12438 // otherwise, go ahead with the folds.
12439 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
12440 EVT XType = N0.getValueType();
12441 if (!LegalOperations ||
12442 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(XType))) {
12443 SDValue Res = DAG.getSetCC(DL, getSetCCResultType(XType), N0, N1, CC);
12444 if (Res.getValueType() != VT)
12445 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
12449 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
12450 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
12451 (!LegalOperations ||
12452 TLI.isOperationLegal(ISD::CTLZ, XType))) {
12453 SDValue Ctlz = DAG.getNode(ISD::CTLZ, SDLoc(N0), XType, N0);
12454 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
12455 DAG.getConstant(Log2_32(XType.getSizeInBits()),
12456 getShiftAmountTy(Ctlz.getValueType())));
12458 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
12459 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
12460 SDValue NegN0 = DAG.getNode(ISD::SUB, SDLoc(N0),
12461 XType, DAG.getConstant(0, XType), N0);
12462 SDValue NotN0 = DAG.getNOT(SDLoc(N0), N0, XType);
12463 return DAG.getNode(ISD::SRL, DL, XType,
12464 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
12465 DAG.getConstant(XType.getSizeInBits()-1,
12466 getShiftAmountTy(XType)));
12468 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
12469 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
12470 SDValue Sign = DAG.getNode(ISD::SRL, SDLoc(N0), XType, N0,
12471 DAG.getConstant(XType.getSizeInBits()-1,
12472 getShiftAmountTy(N0.getValueType())));
12473 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
12477 // Check to see if this is an integer abs.
12478 // select_cc setg[te] X, 0, X, -X ->
12479 // select_cc setgt X, -1, X, -X ->
12480 // select_cc setl[te] X, 0, -X, X ->
12481 // select_cc setlt X, 1, -X, X ->
12482 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
12484 ConstantSDNode *SubC = nullptr;
12485 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
12486 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
12487 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
12488 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
12489 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
12490 (N1C->isOne() && CC == ISD::SETLT)) &&
12491 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
12492 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
12494 EVT XType = N0.getValueType();
12495 if (SubC && SubC->isNullValue() && XType.isInteger()) {
12496 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0), XType,
12498 DAG.getConstant(XType.getSizeInBits()-1,
12499 getShiftAmountTy(N0.getValueType())));
12500 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0),
12502 AddToWorklist(Shift.getNode());
12503 AddToWorklist(Add.getNode());
12504 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
12511 /// This is a stub for TargetLowering::SimplifySetCC.
12512 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
12513 SDValue N1, ISD::CondCode Cond,
12514 SDLoc DL, bool foldBooleans) {
12515 TargetLowering::DAGCombinerInfo
12516 DagCombineInfo(DAG, Level, false, this);
12517 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
12520 /// Given an ISD::SDIV node expressing a divide by constant, return
12521 /// a DAG expression to select that will generate the same value by multiplying
12522 /// by a magic number.
12523 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
12524 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
12525 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
12529 // Avoid division by zero.
12530 if (!C->getAPIntValue())
12533 std::vector<SDNode*> Built;
12535 TLI.BuildSDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
12537 for (SDNode *N : Built)
12542 /// Given an ISD::SDIV node expressing a divide by constant power of 2, return a
12543 /// DAG expression that will generate the same value by right shifting.
12544 SDValue DAGCombiner::BuildSDIVPow2(SDNode *N) {
12545 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
12549 // Avoid division by zero.
12550 if (!C->getAPIntValue())
12553 std::vector<SDNode *> Built;
12554 SDValue S = TLI.BuildSDIVPow2(N, C->getAPIntValue(), DAG, &Built);
12556 for (SDNode *N : Built)
12561 /// Given an ISD::UDIV node expressing a divide by constant, return a DAG
12562 /// expression that will generate the same value by multiplying by a magic
12564 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
12565 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
12566 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
12570 // Avoid division by zero.
12571 if (!C->getAPIntValue())
12574 std::vector<SDNode*> Built;
12576 TLI.BuildUDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
12578 for (SDNode *N : Built)
12583 SDValue DAGCombiner::BuildReciprocalEstimate(SDValue Op) {
12584 if (Level >= AfterLegalizeDAG)
12587 // Expose the DAG combiner to the target combiner implementations.
12588 TargetLowering::DAGCombinerInfo DCI(DAG, Level, false, this);
12590 unsigned Iterations = 0;
12591 if (SDValue Est = TLI.getRecipEstimate(Op, DCI, Iterations)) {
12593 // Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
12594 // For the reciprocal, we need to find the zero of the function:
12595 // F(X) = A X - 1 [which has a zero at X = 1/A]
12597 // X_{i+1} = X_i (2 - A X_i) = X_i + X_i (1 - A X_i) [this second form
12598 // does not require additional intermediate precision]
12599 EVT VT = Op.getValueType();
12601 SDValue FPOne = DAG.getConstantFP(1.0, VT);
12603 AddToWorklist(Est.getNode());
12605 // Newton iterations: Est = Est + Est (1 - Arg * Est)
12606 for (unsigned i = 0; i < Iterations; ++i) {
12607 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Op, Est);
12608 AddToWorklist(NewEst.getNode());
12610 NewEst = DAG.getNode(ISD::FSUB, DL, VT, FPOne, NewEst);
12611 AddToWorklist(NewEst.getNode());
12613 NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst);
12614 AddToWorklist(NewEst.getNode());
12616 Est = DAG.getNode(ISD::FADD, DL, VT, Est, NewEst);
12617 AddToWorklist(Est.getNode());
12626 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
12627 /// For the reciprocal sqrt, we need to find the zero of the function:
12628 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
12630 /// X_{i+1} = X_i (1.5 - A X_i^2 / 2)
12631 /// As a result, we precompute A/2 prior to the iteration loop.
12632 SDValue DAGCombiner::BuildRsqrtNROneConst(SDValue Arg, SDValue Est,
12633 unsigned Iterations) {
12634 EVT VT = Arg.getValueType();
12636 SDValue ThreeHalves = DAG.getConstantFP(1.5, VT);
12638 // We now need 0.5 * Arg which we can write as (1.5 * Arg - Arg) so that
12639 // this entire sequence requires only one FP constant.
12640 SDValue HalfArg = DAG.getNode(ISD::FMUL, DL, VT, ThreeHalves, Arg);
12641 AddToWorklist(HalfArg.getNode());
12643 HalfArg = DAG.getNode(ISD::FSUB, DL, VT, HalfArg, Arg);
12644 AddToWorklist(HalfArg.getNode());
12646 // Newton iterations: Est = Est * (1.5 - HalfArg * Est * Est)
12647 for (unsigned i = 0; i < Iterations; ++i) {
12648 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, Est);
12649 AddToWorklist(NewEst.getNode());
12651 NewEst = DAG.getNode(ISD::FMUL, DL, VT, HalfArg, NewEst);
12652 AddToWorklist(NewEst.getNode());
12654 NewEst = DAG.getNode(ISD::FSUB, DL, VT, ThreeHalves, NewEst);
12655 AddToWorklist(NewEst.getNode());
12657 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst);
12658 AddToWorklist(Est.getNode());
12663 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
12664 /// For the reciprocal sqrt, we need to find the zero of the function:
12665 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
12667 /// X_{i+1} = (-0.5 * X_i) * (A * X_i * X_i + (-3.0))
12668 SDValue DAGCombiner::BuildRsqrtNRTwoConst(SDValue Arg, SDValue Est,
12669 unsigned Iterations) {
12670 EVT VT = Arg.getValueType();
12672 SDValue MinusThree = DAG.getConstantFP(-3.0, VT);
12673 SDValue MinusHalf = DAG.getConstantFP(-0.5, VT);
12675 // Newton iterations: Est = -0.5 * Est * (-3.0 + Arg * Est * Est)
12676 for (unsigned i = 0; i < Iterations; ++i) {
12677 SDValue HalfEst = DAG.getNode(ISD::FMUL, DL, VT, Est, MinusHalf);
12678 AddToWorklist(HalfEst.getNode());
12680 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Est);
12681 AddToWorklist(Est.getNode());
12683 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Arg);
12684 AddToWorklist(Est.getNode());
12686 Est = DAG.getNode(ISD::FADD, DL, VT, Est, MinusThree);
12687 AddToWorklist(Est.getNode());
12689 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, HalfEst);
12690 AddToWorklist(Est.getNode());
12695 SDValue DAGCombiner::BuildRsqrtEstimate(SDValue Op) {
12696 if (Level >= AfterLegalizeDAG)
12699 // Expose the DAG combiner to the target combiner implementations.
12700 TargetLowering::DAGCombinerInfo DCI(DAG, Level, false, this);
12701 unsigned Iterations = 0;
12702 bool UseOneConstNR = false;
12703 if (SDValue Est = TLI.getRsqrtEstimate(Op, DCI, Iterations, UseOneConstNR)) {
12704 AddToWorklist(Est.getNode());
12706 Est = UseOneConstNR ?
12707 BuildRsqrtNROneConst(Op, Est, Iterations) :
12708 BuildRsqrtNRTwoConst(Op, Est, Iterations);
12716 /// Return true if base is a frame index, which is known not to alias with
12717 /// anything but itself. Provides base object and offset as results.
12718 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
12719 const GlobalValue *&GV, const void *&CV) {
12720 // Assume it is a primitive operation.
12721 Base = Ptr; Offset = 0; GV = nullptr; CV = nullptr;
12723 // If it's an adding a simple constant then integrate the offset.
12724 if (Base.getOpcode() == ISD::ADD) {
12725 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
12726 Base = Base.getOperand(0);
12727 Offset += C->getZExtValue();
12731 // Return the underlying GlobalValue, and update the Offset. Return false
12732 // for GlobalAddressSDNode since the same GlobalAddress may be represented
12733 // by multiple nodes with different offsets.
12734 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
12735 GV = G->getGlobal();
12736 Offset += G->getOffset();
12740 // Return the underlying Constant value, and update the Offset. Return false
12741 // for ConstantSDNodes since the same constant pool entry may be represented
12742 // by multiple nodes with different offsets.
12743 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
12744 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
12745 : (const void *)C->getConstVal();
12746 Offset += C->getOffset();
12749 // If it's any of the following then it can't alias with anything but itself.
12750 return isa<FrameIndexSDNode>(Base);
12753 /// Return true if there is any possibility that the two addresses overlap.
12754 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const {
12755 // If they are the same then they must be aliases.
12756 if (Op0->getBasePtr() == Op1->getBasePtr()) return true;
12758 // If they are both volatile then they cannot be reordered.
12759 if (Op0->isVolatile() && Op1->isVolatile()) return true;
12761 // Gather base node and offset information.
12762 SDValue Base1, Base2;
12763 int64_t Offset1, Offset2;
12764 const GlobalValue *GV1, *GV2;
12765 const void *CV1, *CV2;
12766 bool isFrameIndex1 = FindBaseOffset(Op0->getBasePtr(),
12767 Base1, Offset1, GV1, CV1);
12768 bool isFrameIndex2 = FindBaseOffset(Op1->getBasePtr(),
12769 Base2, Offset2, GV2, CV2);
12771 // If they have a same base address then check to see if they overlap.
12772 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
12773 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
12774 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
12776 // It is possible for different frame indices to alias each other, mostly
12777 // when tail call optimization reuses return address slots for arguments.
12778 // To catch this case, look up the actual index of frame indices to compute
12779 // the real alias relationship.
12780 if (isFrameIndex1 && isFrameIndex2) {
12781 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12782 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
12783 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
12784 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
12785 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
12788 // Otherwise, if we know what the bases are, and they aren't identical, then
12789 // we know they cannot alias.
12790 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
12793 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
12794 // compared to the size and offset of the access, we may be able to prove they
12795 // do not alias. This check is conservative for now to catch cases created by
12796 // splitting vector types.
12797 if ((Op0->getOriginalAlignment() == Op1->getOriginalAlignment()) &&
12798 (Op0->getSrcValueOffset() != Op1->getSrcValueOffset()) &&
12799 (Op0->getMemoryVT().getSizeInBits() >> 3 ==
12800 Op1->getMemoryVT().getSizeInBits() >> 3) &&
12801 (Op0->getOriginalAlignment() > Op0->getMemoryVT().getSizeInBits()) >> 3) {
12802 int64_t OffAlign1 = Op0->getSrcValueOffset() % Op0->getOriginalAlignment();
12803 int64_t OffAlign2 = Op1->getSrcValueOffset() % Op1->getOriginalAlignment();
12805 // There is no overlap between these relatively aligned accesses of similar
12806 // size, return no alias.
12807 if ((OffAlign1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign2 ||
12808 (OffAlign2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign1)
12812 bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0
12814 : DAG.getSubtarget().useAA();
12816 if (CombinerAAOnlyFunc.getNumOccurrences() &&
12817 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
12821 Op0->getMemOperand()->getValue() && Op1->getMemOperand()->getValue()) {
12822 // Use alias analysis information.
12823 int64_t MinOffset = std::min(Op0->getSrcValueOffset(),
12824 Op1->getSrcValueOffset());
12825 int64_t Overlap1 = (Op0->getMemoryVT().getSizeInBits() >> 3) +
12826 Op0->getSrcValueOffset() - MinOffset;
12827 int64_t Overlap2 = (Op1->getMemoryVT().getSizeInBits() >> 3) +
12828 Op1->getSrcValueOffset() - MinOffset;
12829 AliasAnalysis::AliasResult AAResult =
12830 AA.alias(AliasAnalysis::Location(Op0->getMemOperand()->getValue(),
12832 UseTBAA ? Op0->getAAInfo() : AAMDNodes()),
12833 AliasAnalysis::Location(Op1->getMemOperand()->getValue(),
12835 UseTBAA ? Op1->getAAInfo() : AAMDNodes()));
12836 if (AAResult == AliasAnalysis::NoAlias)
12840 // Otherwise we have to assume they alias.
12844 /// Walk up chain skipping non-aliasing memory nodes,
12845 /// looking for aliasing nodes and adding them to the Aliases vector.
12846 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
12847 SmallVectorImpl<SDValue> &Aliases) {
12848 SmallVector<SDValue, 8> Chains; // List of chains to visit.
12849 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
12851 // Get alias information for node.
12852 bool IsLoad = isa<LoadSDNode>(N) && !cast<LSBaseSDNode>(N)->isVolatile();
12855 Chains.push_back(OriginalChain);
12856 unsigned Depth = 0;
12858 // Look at each chain and determine if it is an alias. If so, add it to the
12859 // aliases list. If not, then continue up the chain looking for the next
12861 while (!Chains.empty()) {
12862 SDValue Chain = Chains.back();
12865 // For TokenFactor nodes, look at each operand and only continue up the
12866 // chain until we find two aliases. If we've seen two aliases, assume we'll
12867 // find more and revert to original chain since the xform is unlikely to be
12870 // FIXME: The depth check could be made to return the last non-aliasing
12871 // chain we found before we hit a tokenfactor rather than the original
12873 if (Depth > 6 || Aliases.size() == 2) {
12875 Aliases.push_back(OriginalChain);
12879 // Don't bother if we've been before.
12880 if (!Visited.insert(Chain.getNode()).second)
12883 switch (Chain.getOpcode()) {
12884 case ISD::EntryToken:
12885 // Entry token is ideal chain operand, but handled in FindBetterChain.
12890 // Get alias information for Chain.
12891 bool IsOpLoad = isa<LoadSDNode>(Chain.getNode()) &&
12892 !cast<LSBaseSDNode>(Chain.getNode())->isVolatile();
12894 // If chain is alias then stop here.
12895 if (!(IsLoad && IsOpLoad) &&
12896 isAlias(cast<LSBaseSDNode>(N), cast<LSBaseSDNode>(Chain.getNode()))) {
12897 Aliases.push_back(Chain);
12899 // Look further up the chain.
12900 Chains.push_back(Chain.getOperand(0));
12906 case ISD::TokenFactor:
12907 // We have to check each of the operands of the token factor for "small"
12908 // token factors, so we queue them up. Adding the operands to the queue
12909 // (stack) in reverse order maintains the original order and increases the
12910 // likelihood that getNode will find a matching token factor (CSE.)
12911 if (Chain.getNumOperands() > 16) {
12912 Aliases.push_back(Chain);
12915 for (unsigned n = Chain.getNumOperands(); n;)
12916 Chains.push_back(Chain.getOperand(--n));
12921 // For all other instructions we will just have to take what we can get.
12922 Aliases.push_back(Chain);
12927 // We need to be careful here to also search for aliases through the
12928 // value operand of a store, etc. Consider the following situation:
12930 // L1 = load Token1, %52
12931 // S1 = store Token1, L1, %51
12932 // L2 = load Token1, %52+8
12933 // S2 = store Token1, L2, %51+8
12934 // Token2 = Token(S1, S2)
12935 // L3 = load Token2, %53
12936 // S3 = store Token2, L3, %52
12937 // L4 = load Token2, %53+8
12938 // S4 = store Token2, L4, %52+8
12939 // If we search for aliases of S3 (which loads address %52), and we look
12940 // only through the chain, then we'll miss the trivial dependence on L1
12941 // (which also loads from %52). We then might change all loads and
12942 // stores to use Token1 as their chain operand, which could result in
12943 // copying %53 into %52 before copying %52 into %51 (which should
12946 // The problem is, however, that searching for such data dependencies
12947 // can become expensive, and the cost is not directly related to the
12948 // chain depth. Instead, we'll rule out such configurations here by
12949 // insisting that we've visited all chain users (except for users
12950 // of the original chain, which is not necessary). When doing this,
12951 // we need to look through nodes we don't care about (otherwise, things
12952 // like register copies will interfere with trivial cases).
12954 SmallVector<const SDNode *, 16> Worklist;
12955 for (const SDNode *N : Visited)
12956 if (N != OriginalChain.getNode())
12957 Worklist.push_back(N);
12959 while (!Worklist.empty()) {
12960 const SDNode *M = Worklist.pop_back_val();
12962 // We have already visited M, and want to make sure we've visited any uses
12963 // of M that we care about. For uses that we've not visisted, and don't
12964 // care about, queue them to the worklist.
12966 for (SDNode::use_iterator UI = M->use_begin(),
12967 UIE = M->use_end(); UI != UIE; ++UI)
12968 if (UI.getUse().getValueType() == MVT::Other &&
12969 Visited.insert(*UI).second) {
12970 if (isa<MemIntrinsicSDNode>(*UI) || isa<MemSDNode>(*UI)) {
12971 // We've not visited this use, and we care about it (it could have an
12972 // ordering dependency with the original node).
12974 Aliases.push_back(OriginalChain);
12978 // We've not visited this use, but we don't care about it. Mark it as
12979 // visited and enqueue it to the worklist.
12980 Worklist.push_back(*UI);
12985 /// Walk up chain skipping non-aliasing memory nodes, looking for a better chain
12986 /// (aliasing node.)
12987 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
12988 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
12990 // Accumulate all the aliases to this node.
12991 GatherAllAliases(N, OldChain, Aliases);
12993 // If no operands then chain to entry token.
12994 if (Aliases.size() == 0)
12995 return DAG.getEntryNode();
12997 // If a single operand then chain to it. We don't need to revisit it.
12998 if (Aliases.size() == 1)
13001 // Construct a custom tailored token factor.
13002 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Aliases);
13005 /// This is the entry point for the file.
13006 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
13007 CodeGenOpt::Level OptLevel) {
13008 /// This is the main entry point to this class.
13009 DAGCombiner(*this, AA, OptLevel).Run(Level);