1 //===-- DAGCombiner.cpp - Implement a DAG node combiner -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass combines dag nodes to form fewer, simpler DAG nodes. It can be run
11 // both before and after the DAG is legalized.
13 // This pass is not a substitute for the LLVM IR instcombine pass. This pass is
14 // primarily intended to handle simplification opportunities that are implicit
15 // in the LLVM IR and exposed by the various codegen lowering phases.
17 //===----------------------------------------------------------------------===//
19 #include "llvm/CodeGen/SelectionDAG.h"
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/ADT/SmallBitVector.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/Analysis/AliasAnalysis.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/DerivedTypes.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/LLVMContext.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Target/TargetLowering.h"
37 #include "llvm/Target/TargetOptions.h"
38 #include "llvm/Target/TargetRegisterInfo.h"
39 #include "llvm/Target/TargetSubtargetInfo.h"
43 #define DEBUG_TYPE "dagcombine"
45 STATISTIC(NodesCombined , "Number of dag nodes combined");
46 STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created");
47 STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created");
48 STATISTIC(OpsNarrowed , "Number of load/op/store narrowed");
49 STATISTIC(LdStFP2Int , "Number of fp load/store pairs transformed to int");
50 STATISTIC(SlicedLoads, "Number of load sliced");
54 CombinerAA("combiner-alias-analysis", cl::Hidden,
55 cl::desc("Enable DAG combiner alias-analysis heuristics"));
58 CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden,
59 cl::desc("Enable DAG combiner's use of IR alias analysis"));
62 UseTBAA("combiner-use-tbaa", cl::Hidden, cl::init(true),
63 cl::desc("Enable DAG combiner's use of TBAA"));
66 static cl::opt<std::string>
67 CombinerAAOnlyFunc("combiner-aa-only-func", cl::Hidden,
68 cl::desc("Only use DAG-combiner alias analysis in this"
72 /// Hidden option to stress test load slicing, i.e., when this option
73 /// is enabled, load slicing bypasses most of its profitability guards.
75 StressLoadSlicing("combiner-stress-load-slicing", cl::Hidden,
76 cl::desc("Bypass the profitability model of load "
81 MaySplitLoadIndex("combiner-split-load-index", cl::Hidden, cl::init(true),
82 cl::desc("DAG combiner may split indexing from loads"));
84 //------------------------------ DAGCombiner ---------------------------------//
88 const TargetLowering &TLI;
90 CodeGenOpt::Level OptLevel;
95 /// \brief Worklist of all of the nodes that need to be simplified.
97 /// This must behave as a stack -- new nodes to process are pushed onto the
98 /// back and when processing we pop off of the back.
100 /// The worklist will not contain duplicates but may contain null entries
101 /// due to nodes being deleted from the underlying DAG.
102 SmallVector<SDNode *, 64> Worklist;
104 /// \brief Mapping from an SDNode to its position on the worklist.
106 /// This is used to find and remove nodes from the worklist (by nulling
107 /// them) when they are deleted from the underlying DAG. It relies on
108 /// stable indices of nodes within the worklist.
109 DenseMap<SDNode *, unsigned> WorklistMap;
111 /// \brief Set of nodes which have been combined (at least once).
113 /// This is used to allow us to reliably add any operands of a DAG node
114 /// which have not yet been combined to the worklist.
115 SmallPtrSet<SDNode *, 64> CombinedNodes;
117 // AA - Used for DAG load/store alias analysis.
120 /// When an instruction is simplified, add all users of the instruction to
121 /// the work lists because they might get more simplified now.
122 void AddUsersToWorklist(SDNode *N) {
123 for (SDNode *Node : N->uses())
127 /// Call the node-specific routine that folds each particular type of node.
128 SDValue visit(SDNode *N);
131 /// Add to the worklist making sure its instance is at the back (next to be
133 void AddToWorklist(SDNode *N) {
134 // Skip handle nodes as they can't usefully be combined and confuse the
135 // zero-use deletion strategy.
136 if (N->getOpcode() == ISD::HANDLENODE)
139 if (WorklistMap.insert(std::make_pair(N, Worklist.size())).second)
140 Worklist.push_back(N);
143 /// Remove all instances of N from the worklist.
144 void removeFromWorklist(SDNode *N) {
145 CombinedNodes.erase(N);
147 auto It = WorklistMap.find(N);
148 if (It == WorklistMap.end())
149 return; // Not in the worklist.
151 // Null out the entry rather than erasing it to avoid a linear operation.
152 Worklist[It->second] = nullptr;
153 WorklistMap.erase(It);
156 void deleteAndRecombine(SDNode *N);
157 bool recursivelyDeleteUnusedNodes(SDNode *N);
159 SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
162 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) {
163 return CombineTo(N, &Res, 1, AddTo);
166 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1,
168 SDValue To[] = { Res0, Res1 };
169 return CombineTo(N, To, 2, AddTo);
172 void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO);
176 /// Check the specified integer node value to see if it can be simplified or
177 /// if things it uses can be simplified by bit propagation.
178 /// If so, return true.
179 bool SimplifyDemandedBits(SDValue Op) {
180 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
181 APInt Demanded = APInt::getAllOnesValue(BitWidth);
182 return SimplifyDemandedBits(Op, Demanded);
185 bool SimplifyDemandedBits(SDValue Op, const APInt &Demanded);
187 bool CombineToPreIndexedLoadStore(SDNode *N);
188 bool CombineToPostIndexedLoadStore(SDNode *N);
189 SDValue SplitIndexingFromLoad(LoadSDNode *LD);
190 bool SliceUpLoad(SDNode *N);
192 /// \brief Replace an ISD::EXTRACT_VECTOR_ELT of a load with a narrowed
195 /// \param EVE ISD::EXTRACT_VECTOR_ELT to be replaced.
196 /// \param InVecVT type of the input vector to EVE with bitcasts resolved.
197 /// \param EltNo index of the vector element to load.
198 /// \param OriginalLoad load that EVE came from to be replaced.
199 /// \returns EVE on success SDValue() on failure.
200 SDValue ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
201 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad);
202 void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad);
203 SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace);
204 SDValue SExtPromoteOperand(SDValue Op, EVT PVT);
205 SDValue ZExtPromoteOperand(SDValue Op, EVT PVT);
206 SDValue PromoteIntBinOp(SDValue Op);
207 SDValue PromoteIntShiftOp(SDValue Op);
208 SDValue PromoteExtend(SDValue Op);
209 bool PromoteLoad(SDValue Op);
211 void ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
212 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
213 ISD::NodeType ExtType);
215 /// Call the node-specific routine that knows how to fold each
216 /// particular type of node. If that doesn't do anything, try the
217 /// target-specific DAG combines.
218 SDValue combine(SDNode *N);
220 // Visitation implementation - Implement dag node combining for different
221 // node types. The semantics are as follows:
223 // SDValue.getNode() == 0 - No change was made
224 // SDValue.getNode() == N - N was replaced, is dead and has been handled.
225 // otherwise - N should be replaced by the returned Operand.
227 SDValue visitTokenFactor(SDNode *N);
228 SDValue visitMERGE_VALUES(SDNode *N);
229 SDValue visitADD(SDNode *N);
230 SDValue visitSUB(SDNode *N);
231 SDValue visitADDC(SDNode *N);
232 SDValue visitSUBC(SDNode *N);
233 SDValue visitADDE(SDNode *N);
234 SDValue visitSUBE(SDNode *N);
235 SDValue visitMUL(SDNode *N);
236 SDValue visitSDIV(SDNode *N);
237 SDValue visitUDIV(SDNode *N);
238 SDValue visitSREM(SDNode *N);
239 SDValue visitUREM(SDNode *N);
240 SDValue visitMULHU(SDNode *N);
241 SDValue visitMULHS(SDNode *N);
242 SDValue visitSMUL_LOHI(SDNode *N);
243 SDValue visitUMUL_LOHI(SDNode *N);
244 SDValue visitSMULO(SDNode *N);
245 SDValue visitUMULO(SDNode *N);
246 SDValue visitSDIVREM(SDNode *N);
247 SDValue visitUDIVREM(SDNode *N);
248 SDValue visitAND(SDNode *N);
249 SDValue visitANDLike(SDValue N0, SDValue N1, SDNode *LocReference);
250 SDValue visitOR(SDNode *N);
251 SDValue visitORLike(SDValue N0, SDValue N1, SDNode *LocReference);
252 SDValue visitXOR(SDNode *N);
253 SDValue SimplifyVBinOp(SDNode *N);
254 SDValue visitSHL(SDNode *N);
255 SDValue visitSRA(SDNode *N);
256 SDValue visitSRL(SDNode *N);
257 SDValue visitRotate(SDNode *N);
258 SDValue visitCTLZ(SDNode *N);
259 SDValue visitCTLZ_ZERO_UNDEF(SDNode *N);
260 SDValue visitCTTZ(SDNode *N);
261 SDValue visitCTTZ_ZERO_UNDEF(SDNode *N);
262 SDValue visitCTPOP(SDNode *N);
263 SDValue visitSELECT(SDNode *N);
264 SDValue visitVSELECT(SDNode *N);
265 SDValue visitSELECT_CC(SDNode *N);
266 SDValue visitSETCC(SDNode *N);
267 SDValue visitSIGN_EXTEND(SDNode *N);
268 SDValue visitZERO_EXTEND(SDNode *N);
269 SDValue visitANY_EXTEND(SDNode *N);
270 SDValue visitSIGN_EXTEND_INREG(SDNode *N);
271 SDValue visitTRUNCATE(SDNode *N);
272 SDValue visitBITCAST(SDNode *N);
273 SDValue visitBUILD_PAIR(SDNode *N);
274 SDValue visitFADD(SDNode *N);
275 SDValue visitFSUB(SDNode *N);
276 SDValue visitFMUL(SDNode *N);
277 SDValue visitFMA(SDNode *N);
278 SDValue visitFDIV(SDNode *N);
279 SDValue visitFREM(SDNode *N);
280 SDValue visitFSQRT(SDNode *N);
281 SDValue visitFCOPYSIGN(SDNode *N);
282 SDValue visitSINT_TO_FP(SDNode *N);
283 SDValue visitUINT_TO_FP(SDNode *N);
284 SDValue visitFP_TO_SINT(SDNode *N);
285 SDValue visitFP_TO_UINT(SDNode *N);
286 SDValue visitFP_ROUND(SDNode *N);
287 SDValue visitFP_ROUND_INREG(SDNode *N);
288 SDValue visitFP_EXTEND(SDNode *N);
289 SDValue visitFNEG(SDNode *N);
290 SDValue visitFABS(SDNode *N);
291 SDValue visitFCEIL(SDNode *N);
292 SDValue visitFTRUNC(SDNode *N);
293 SDValue visitFFLOOR(SDNode *N);
294 SDValue visitFMINNUM(SDNode *N);
295 SDValue visitFMAXNUM(SDNode *N);
296 SDValue visitBRCOND(SDNode *N);
297 SDValue visitBR_CC(SDNode *N);
298 SDValue visitLOAD(SDNode *N);
299 SDValue visitSTORE(SDNode *N);
300 SDValue visitINSERT_VECTOR_ELT(SDNode *N);
301 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
302 SDValue visitBUILD_VECTOR(SDNode *N);
303 SDValue visitCONCAT_VECTORS(SDNode *N);
304 SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
305 SDValue visitVECTOR_SHUFFLE(SDNode *N);
306 SDValue visitSCALAR_TO_VECTOR(SDNode *N);
307 SDValue visitINSERT_SUBVECTOR(SDNode *N);
308 SDValue visitMLOAD(SDNode *N);
309 SDValue visitMSTORE(SDNode *N);
310 SDValue visitFP_TO_FP16(SDNode *N);
312 SDValue XformToShuffleWithZero(SDNode *N);
313 SDValue ReassociateOps(unsigned Opc, SDLoc DL, SDValue LHS, SDValue RHS);
315 SDValue visitShiftByConstant(SDNode *N, ConstantSDNode *Amt);
317 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
318 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
319 SDValue SimplifySelect(SDLoc DL, SDValue N0, SDValue N1, SDValue N2);
320 SDValue SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1, SDValue N2,
321 SDValue N3, ISD::CondCode CC,
322 bool NotExtCompare = false);
323 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
324 SDLoc DL, bool foldBooleans = true);
326 bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
328 bool isOneUseSetCC(SDValue N) const;
330 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
332 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
333 SDValue CombineExtLoad(SDNode *N);
334 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
335 SDValue BuildSDIV(SDNode *N);
336 SDValue BuildSDIVPow2(SDNode *N);
337 SDValue BuildUDIV(SDNode *N);
338 SDValue BuildReciprocalEstimate(SDValue Op);
339 SDValue BuildRsqrtEstimate(SDValue Op);
340 SDValue BuildRsqrtNROneConst(SDValue Op, SDValue Est, unsigned Iterations);
341 SDValue BuildRsqrtNRTwoConst(SDValue Op, SDValue Est, unsigned Iterations);
342 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
343 bool DemandHighBits = true);
344 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
345 SDNode *MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg,
346 SDValue InnerPos, SDValue InnerNeg,
347 unsigned PosOpcode, unsigned NegOpcode,
349 SDNode *MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL);
350 SDValue ReduceLoadWidth(SDNode *N);
351 SDValue ReduceLoadOpStoreWidth(SDNode *N);
352 SDValue TransformFPLoadStorePair(SDNode *N);
353 SDValue reduceBuildVecExtToExtBuildVec(SDNode *N);
354 SDValue reduceBuildVecConvertToConvertBuildVec(SDNode *N);
356 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
358 /// Walk up chain skipping non-aliasing memory nodes,
359 /// looking for aliasing nodes and adding them to the Aliases vector.
360 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
361 SmallVectorImpl<SDValue> &Aliases);
363 /// Return true if there is any possibility that the two addresses overlap.
364 bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const;
366 /// Walk up chain skipping non-aliasing memory nodes, looking for a better
367 /// chain (aliasing node.)
368 SDValue FindBetterChain(SDNode *N, SDValue Chain);
370 /// Holds a pointer to an LSBaseSDNode as well as information on where it
371 /// is located in a sequence of memory operations connected by a chain.
373 MemOpLink (LSBaseSDNode *N, int64_t Offset, unsigned Seq):
374 MemNode(N), OffsetFromBase(Offset), SequenceNum(Seq) { }
375 // Ptr to the mem node.
376 LSBaseSDNode *MemNode;
377 // Offset from the base ptr.
378 int64_t OffsetFromBase;
379 // What is the sequence number of this mem node.
380 // Lowest mem operand in the DAG starts at zero.
381 unsigned SequenceNum;
384 /// This is a helper function for MergeConsecutiveStores. When the source
385 /// elements of the consecutive stores are all constants or all extracted
386 /// vector elements, try to merge them into one larger store.
387 /// \return True if a merged store was created.
388 bool MergeStoresOfConstantsOrVecElts(SmallVectorImpl<MemOpLink> &StoreNodes,
389 EVT MemVT, unsigned NumElem,
390 bool IsConstantSrc, bool UseVector);
392 /// Merge consecutive store operations into a wide store.
393 /// This optimization uses wide integers or vectors when possible.
394 /// \return True if some memory operations were changed.
395 bool MergeConsecutiveStores(StoreSDNode *N);
397 /// \brief Try to transform a truncation where C is a constant:
398 /// (trunc (and X, C)) -> (and (trunc X), (trunc C))
400 /// \p N needs to be a truncation and its first operand an AND. Other
401 /// requirements are checked by the function (e.g. that trunc is
402 /// single-use) and if missed an empty SDValue is returned.
403 SDValue distributeTruncateThroughAnd(SDNode *N);
406 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
407 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
408 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {
409 auto *F = DAG.getMachineFunction().getFunction();
410 ForCodeSize = F->hasFnAttribute(Attribute::OptimizeForSize) ||
411 F->hasFnAttribute(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, ArrayRef<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 integer BuildVector
710 // or constant integer.
711 static SDNode *isConstantIntBuildVectorOrConstantInt(SDValue N) {
712 if (isa<ConstantSDNode>(N))
714 if (ISD::isBuildVectorOfConstantSDNodes(N.getNode()))
719 // \brief Returns the SDNode if it is a constant float BuildVector
720 // or constant float.
721 static SDNode *isConstantFPBuildVectorOrConstantFP(SDValue N) {
722 if (isa<ConstantFPSDNode>(N))
724 if (ISD::isBuildVectorOfConstantFPSDNodes(N.getNode()))
729 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
731 static ConstantSDNode *isConstOrConstSplat(SDValue N) {
732 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N))
735 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
736 BitVector UndefElements;
737 ConstantSDNode *CN = BV->getConstantSplatNode(&UndefElements);
739 // BuildVectors can truncate their operands. Ignore that case here.
740 // FIXME: We blindly ignore splats which include undef which is overly
742 if (CN && UndefElements.none() &&
743 CN->getValueType(0) == N.getValueType().getScalarType())
750 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
752 static ConstantFPSDNode *isConstOrConstSplatFP(SDValue N) {
753 if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N))
756 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
757 BitVector UndefElements;
758 ConstantFPSDNode *CN = BV->getConstantFPSplatNode(&UndefElements);
760 if (CN && UndefElements.none())
767 SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL,
768 SDValue N0, SDValue N1) {
769 EVT VT = N0.getValueType();
770 if (N0.getOpcode() == Opc) {
771 if (SDNode *L = isConstantIntBuildVectorOrConstantInt(N0.getOperand(1))) {
772 if (SDNode *R = isConstantIntBuildVectorOrConstantInt(N1)) {
773 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
774 if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, L, R))
775 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
778 if (N0.hasOneUse()) {
779 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one
781 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1);
782 if (!OpNode.getNode())
784 AddToWorklist(OpNode.getNode());
785 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
790 if (N1.getOpcode() == Opc) {
791 if (SDNode *R = isConstantIntBuildVectorOrConstantInt(N1.getOperand(1))) {
792 if (SDNode *L = isConstantIntBuildVectorOrConstantInt(N0)) {
793 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
794 if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, R, L))
795 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
798 if (N1.hasOneUse()) {
799 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one
801 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N1.getOperand(0), N0);
802 if (!OpNode.getNode())
804 AddToWorklist(OpNode.getNode());
805 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
813 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
815 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
817 DEBUG(dbgs() << "\nReplacing.1 ";
819 dbgs() << "\nWith: ";
820 To[0].getNode()->dump(&DAG);
821 dbgs() << " and " << NumTo-1 << " other values\n");
822 for (unsigned i = 0, e = NumTo; i != e; ++i)
823 assert((!To[i].getNode() ||
824 N->getValueType(i) == To[i].getValueType()) &&
825 "Cannot combine value to value of different type!");
827 WorklistRemover DeadNodes(*this);
828 DAG.ReplaceAllUsesWith(N, To);
830 // Push the new nodes and any users onto the worklist
831 for (unsigned i = 0, e = NumTo; i != e; ++i) {
832 if (To[i].getNode()) {
833 AddToWorklist(To[i].getNode());
834 AddUsersToWorklist(To[i].getNode());
839 // Finally, if the node is now dead, remove it from the graph. The node
840 // may not be dead if the replacement process recursively simplified to
841 // something else needing this node.
843 deleteAndRecombine(N);
844 return SDValue(N, 0);
848 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
849 // Replace all uses. If any nodes become isomorphic to other nodes and
850 // are deleted, make sure to remove them from our worklist.
851 WorklistRemover DeadNodes(*this);
852 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New);
854 // Push the new node and any (possibly new) users onto the worklist.
855 AddToWorklist(TLO.New.getNode());
856 AddUsersToWorklist(TLO.New.getNode());
858 // Finally, if the node is now dead, remove it from the graph. The node
859 // may not be dead if the replacement process recursively simplified to
860 // something else needing this node.
861 if (TLO.Old.getNode()->use_empty())
862 deleteAndRecombine(TLO.Old.getNode());
865 /// Check the specified integer node value to see if it can be simplified or if
866 /// things it uses can be simplified by bit propagation. If so, return true.
867 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
868 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
869 APInt KnownZero, KnownOne;
870 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
874 AddToWorklist(Op.getNode());
876 // Replace the old value with the new one.
878 DEBUG(dbgs() << "\nReplacing.2 ";
879 TLO.Old.getNode()->dump(&DAG);
880 dbgs() << "\nWith: ";
881 TLO.New.getNode()->dump(&DAG);
884 CommitTargetLoweringOpt(TLO);
888 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
890 EVT VT = Load->getValueType(0);
891 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
893 DEBUG(dbgs() << "\nReplacing.9 ";
895 dbgs() << "\nWith: ";
896 Trunc.getNode()->dump(&DAG);
898 WorklistRemover DeadNodes(*this);
899 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc);
900 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1));
901 deleteAndRecombine(Load);
902 AddToWorklist(Trunc.getNode());
905 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
908 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
909 EVT MemVT = LD->getMemoryVT();
910 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
911 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, PVT, MemVT) ? ISD::ZEXTLOAD
913 : LD->getExtensionType();
915 return DAG.getExtLoad(ExtType, dl, PVT,
916 LD->getChain(), LD->getBasePtr(),
917 MemVT, LD->getMemOperand());
920 unsigned Opc = Op.getOpcode();
923 case ISD::AssertSext:
924 return DAG.getNode(ISD::AssertSext, dl, PVT,
925 SExtPromoteOperand(Op.getOperand(0), PVT),
927 case ISD::AssertZext:
928 return DAG.getNode(ISD::AssertZext, dl, PVT,
929 ZExtPromoteOperand(Op.getOperand(0), PVT),
931 case ISD::Constant: {
933 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
934 return DAG.getNode(ExtOpc, dl, PVT, Op);
938 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
940 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
943 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
944 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
946 EVT OldVT = Op.getValueType();
948 bool Replace = false;
949 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
950 if (!NewOp.getNode())
952 AddToWorklist(NewOp.getNode());
955 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
956 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
957 DAG.getValueType(OldVT));
960 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
961 EVT OldVT = Op.getValueType();
963 bool Replace = false;
964 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
965 if (!NewOp.getNode())
967 AddToWorklist(NewOp.getNode());
970 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
971 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
974 /// Promote the specified integer binary operation if the target indicates it is
975 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
976 /// i32 since i16 instructions are longer.
977 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
978 if (!LegalOperations)
981 EVT VT = Op.getValueType();
982 if (VT.isVector() || !VT.isInteger())
985 // If operation type is 'undesirable', e.g. i16 on x86, consider
987 unsigned Opc = Op.getOpcode();
988 if (TLI.isTypeDesirableForOp(Opc, VT))
992 // Consult target whether it is a good idea to promote this operation and
993 // what's the right type to promote it to.
994 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
995 assert(PVT != VT && "Don't know what type to promote to!");
997 bool Replace0 = false;
998 SDValue N0 = Op.getOperand(0);
999 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
1003 bool Replace1 = false;
1004 SDValue N1 = Op.getOperand(1);
1009 NN1 = PromoteOperand(N1, PVT, Replace1);
1014 AddToWorklist(NN0.getNode());
1016 AddToWorklist(NN1.getNode());
1019 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
1021 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
1023 DEBUG(dbgs() << "\nPromoting ";
1024 Op.getNode()->dump(&DAG));
1026 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1027 DAG.getNode(Opc, dl, PVT, NN0, NN1));
1032 /// Promote the specified integer shift operation if the target indicates it is
1033 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
1034 /// i32 since i16 instructions are longer.
1035 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
1036 if (!LegalOperations)
1039 EVT VT = Op.getValueType();
1040 if (VT.isVector() || !VT.isInteger())
1043 // If operation type is 'undesirable', e.g. i16 on x86, consider
1045 unsigned Opc = Op.getOpcode();
1046 if (TLI.isTypeDesirableForOp(Opc, VT))
1050 // Consult target whether it is a good idea to promote this operation and
1051 // what's the right type to promote it to.
1052 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1053 assert(PVT != VT && "Don't know what type to promote to!");
1055 bool Replace = false;
1056 SDValue N0 = Op.getOperand(0);
1057 if (Opc == ISD::SRA)
1058 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
1059 else if (Opc == ISD::SRL)
1060 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
1062 N0 = PromoteOperand(N0, PVT, Replace);
1066 AddToWorklist(N0.getNode());
1068 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
1070 DEBUG(dbgs() << "\nPromoting ";
1071 Op.getNode()->dump(&DAG));
1073 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1074 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
1079 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
1080 if (!LegalOperations)
1083 EVT VT = Op.getValueType();
1084 if (VT.isVector() || !VT.isInteger())
1087 // If operation type is 'undesirable', e.g. i16 on x86, consider
1089 unsigned Opc = Op.getOpcode();
1090 if (TLI.isTypeDesirableForOp(Opc, VT))
1094 // Consult target whether it is a good idea to promote this operation and
1095 // what's the right type to promote it to.
1096 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1097 assert(PVT != VT && "Don't know what type to promote to!");
1098 // fold (aext (aext x)) -> (aext x)
1099 // fold (aext (zext x)) -> (zext x)
1100 // fold (aext (sext x)) -> (sext x)
1101 DEBUG(dbgs() << "\nPromoting ";
1102 Op.getNode()->dump(&DAG));
1103 return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0));
1108 bool DAGCombiner::PromoteLoad(SDValue Op) {
1109 if (!LegalOperations)
1112 EVT VT = Op.getValueType();
1113 if (VT.isVector() || !VT.isInteger())
1116 // If operation type is 'undesirable', e.g. i16 on x86, consider
1118 unsigned Opc = Op.getOpcode();
1119 if (TLI.isTypeDesirableForOp(Opc, VT))
1123 // Consult target whether it is a good idea to promote this operation and
1124 // what's the right type to promote it to.
1125 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1126 assert(PVT != VT && "Don't know what type to promote to!");
1129 SDNode *N = Op.getNode();
1130 LoadSDNode *LD = cast<LoadSDNode>(N);
1131 EVT MemVT = LD->getMemoryVT();
1132 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
1133 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, PVT, MemVT) ? ISD::ZEXTLOAD
1135 : LD->getExtensionType();
1136 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
1137 LD->getChain(), LD->getBasePtr(),
1138 MemVT, LD->getMemOperand());
1139 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
1141 DEBUG(dbgs() << "\nPromoting ";
1144 Result.getNode()->dump(&DAG);
1146 WorklistRemover DeadNodes(*this);
1147 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
1148 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1));
1149 deleteAndRecombine(N);
1150 AddToWorklist(Result.getNode());
1156 /// \brief Recursively delete a node which has no uses and any operands for
1157 /// which it is the only use.
1159 /// Note that this both deletes the nodes and removes them from the worklist.
1160 /// It also adds any nodes who have had a user deleted to the worklist as they
1161 /// may now have only one use and subject to other combines.
1162 bool DAGCombiner::recursivelyDeleteUnusedNodes(SDNode *N) {
1163 if (!N->use_empty())
1166 SmallSetVector<SDNode *, 16> Nodes;
1169 N = Nodes.pop_back_val();
1173 if (N->use_empty()) {
1174 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1175 Nodes.insert(N->getOperand(i).getNode());
1177 removeFromWorklist(N);
1182 } while (!Nodes.empty());
1186 //===----------------------------------------------------------------------===//
1187 // Main DAG Combiner implementation
1188 //===----------------------------------------------------------------------===//
1190 void DAGCombiner::Run(CombineLevel AtLevel) {
1191 // set the instance variables, so that the various visit routines may use it.
1193 LegalOperations = Level >= AfterLegalizeVectorOps;
1194 LegalTypes = Level >= AfterLegalizeTypes;
1196 // Add all the dag nodes to the worklist.
1197 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
1198 E = DAG.allnodes_end(); I != E; ++I)
1201 // Create a dummy node (which is not added to allnodes), that adds a reference
1202 // to the root node, preventing it from being deleted, and tracking any
1203 // changes of the root.
1204 HandleSDNode Dummy(DAG.getRoot());
1206 // while the worklist isn't empty, find a node and
1207 // try and combine it.
1208 while (!WorklistMap.empty()) {
1210 // The Worklist holds the SDNodes in order, but it may contain null entries.
1212 N = Worklist.pop_back_val();
1215 bool GoodWorklistEntry = WorklistMap.erase(N);
1216 (void)GoodWorklistEntry;
1217 assert(GoodWorklistEntry &&
1218 "Found a worklist entry without a corresponding map entry!");
1220 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1221 // N is deleted from the DAG, since they too may now be dead or may have a
1222 // reduced number of uses, allowing other xforms.
1223 if (recursivelyDeleteUnusedNodes(N))
1226 WorklistRemover DeadNodes(*this);
1228 // If this combine is running after legalizing the DAG, re-legalize any
1229 // nodes pulled off the worklist.
1230 if (Level == AfterLegalizeDAG) {
1231 SmallSetVector<SDNode *, 16> UpdatedNodes;
1232 bool NIsValid = DAG.LegalizeOp(N, UpdatedNodes);
1234 for (SDNode *LN : UpdatedNodes) {
1236 AddUsersToWorklist(LN);
1242 DEBUG(dbgs() << "\nCombining: "; N->dump(&DAG));
1244 // Add any operands of the new node which have not yet been combined to the
1245 // worklist as well. Because the worklist uniques things already, this
1246 // won't repeatedly process the same operand.
1247 CombinedNodes.insert(N);
1248 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1249 if (!CombinedNodes.count(N->getOperand(i).getNode()))
1250 AddToWorklist(N->getOperand(i).getNode());
1252 SDValue RV = combine(N);
1259 // If we get back the same node we passed in, rather than a new node or
1260 // zero, we know that the node must have defined multiple values and
1261 // CombineTo was used. Since CombineTo takes care of the worklist
1262 // mechanics for us, we have no work to do in this case.
1263 if (RV.getNode() == N)
1266 assert(N->getOpcode() != ISD::DELETED_NODE &&
1267 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1268 "Node was deleted but visit returned new node!");
1270 DEBUG(dbgs() << " ... into: ";
1271 RV.getNode()->dump(&DAG));
1273 // Transfer debug value.
1274 DAG.TransferDbgValues(SDValue(N, 0), RV);
1275 if (N->getNumValues() == RV.getNode()->getNumValues())
1276 DAG.ReplaceAllUsesWith(N, RV.getNode());
1278 assert(N->getValueType(0) == RV.getValueType() &&
1279 N->getNumValues() == 1 && "Type mismatch");
1281 DAG.ReplaceAllUsesWith(N, &OpV);
1284 // Push the new node and any users onto the worklist
1285 AddToWorklist(RV.getNode());
1286 AddUsersToWorklist(RV.getNode());
1288 // Finally, if the node is now dead, remove it from the graph. The node
1289 // may not be dead if the replacement process recursively simplified to
1290 // something else needing this node. This will also take care of adding any
1291 // operands which have lost a user to the worklist.
1292 recursivelyDeleteUnusedNodes(N);
1295 // If the root changed (e.g. it was a dead load, update the root).
1296 DAG.setRoot(Dummy.getValue());
1297 DAG.RemoveDeadNodes();
1300 SDValue DAGCombiner::visit(SDNode *N) {
1301 switch (N->getOpcode()) {
1303 case ISD::TokenFactor: return visitTokenFactor(N);
1304 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1305 case ISD::ADD: return visitADD(N);
1306 case ISD::SUB: return visitSUB(N);
1307 case ISD::ADDC: return visitADDC(N);
1308 case ISD::SUBC: return visitSUBC(N);
1309 case ISD::ADDE: return visitADDE(N);
1310 case ISD::SUBE: return visitSUBE(N);
1311 case ISD::MUL: return visitMUL(N);
1312 case ISD::SDIV: return visitSDIV(N);
1313 case ISD::UDIV: return visitUDIV(N);
1314 case ISD::SREM: return visitSREM(N);
1315 case ISD::UREM: return visitUREM(N);
1316 case ISD::MULHU: return visitMULHU(N);
1317 case ISD::MULHS: return visitMULHS(N);
1318 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1319 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1320 case ISD::SMULO: return visitSMULO(N);
1321 case ISD::UMULO: return visitUMULO(N);
1322 case ISD::SDIVREM: return visitSDIVREM(N);
1323 case ISD::UDIVREM: return visitUDIVREM(N);
1324 case ISD::AND: return visitAND(N);
1325 case ISD::OR: return visitOR(N);
1326 case ISD::XOR: return visitXOR(N);
1327 case ISD::SHL: return visitSHL(N);
1328 case ISD::SRA: return visitSRA(N);
1329 case ISD::SRL: return visitSRL(N);
1331 case ISD::ROTL: return visitRotate(N);
1332 case ISD::CTLZ: return visitCTLZ(N);
1333 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1334 case ISD::CTTZ: return visitCTTZ(N);
1335 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1336 case ISD::CTPOP: return visitCTPOP(N);
1337 case ISD::SELECT: return visitSELECT(N);
1338 case ISD::VSELECT: return visitVSELECT(N);
1339 case ISD::SELECT_CC: return visitSELECT_CC(N);
1340 case ISD::SETCC: return visitSETCC(N);
1341 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1342 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1343 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1344 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1345 case ISD::TRUNCATE: return visitTRUNCATE(N);
1346 case ISD::BITCAST: return visitBITCAST(N);
1347 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1348 case ISD::FADD: return visitFADD(N);
1349 case ISD::FSUB: return visitFSUB(N);
1350 case ISD::FMUL: return visitFMUL(N);
1351 case ISD::FMA: return visitFMA(N);
1352 case ISD::FDIV: return visitFDIV(N);
1353 case ISD::FREM: return visitFREM(N);
1354 case ISD::FSQRT: return visitFSQRT(N);
1355 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1356 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1357 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1358 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1359 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1360 case ISD::FP_ROUND: return visitFP_ROUND(N);
1361 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1362 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1363 case ISD::FNEG: return visitFNEG(N);
1364 case ISD::FABS: return visitFABS(N);
1365 case ISD::FFLOOR: return visitFFLOOR(N);
1366 case ISD::FMINNUM: return visitFMINNUM(N);
1367 case ISD::FMAXNUM: return visitFMAXNUM(N);
1368 case ISD::FCEIL: return visitFCEIL(N);
1369 case ISD::FTRUNC: return visitFTRUNC(N);
1370 case ISD::BRCOND: return visitBRCOND(N);
1371 case ISD::BR_CC: return visitBR_CC(N);
1372 case ISD::LOAD: return visitLOAD(N);
1373 case ISD::STORE: return visitSTORE(N);
1374 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1375 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1376 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1377 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1378 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1379 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1380 case ISD::SCALAR_TO_VECTOR: return visitSCALAR_TO_VECTOR(N);
1381 case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N);
1382 case ISD::MLOAD: return visitMLOAD(N);
1383 case ISD::MSTORE: return visitMSTORE(N);
1384 case ISD::FP_TO_FP16: return visitFP_TO_FP16(N);
1389 SDValue DAGCombiner::combine(SDNode *N) {
1390 SDValue RV = visit(N);
1392 // If nothing happened, try a target-specific DAG combine.
1393 if (!RV.getNode()) {
1394 assert(N->getOpcode() != ISD::DELETED_NODE &&
1395 "Node was deleted but visit returned NULL!");
1397 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1398 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1400 // Expose the DAG combiner to the target combiner impls.
1401 TargetLowering::DAGCombinerInfo
1402 DagCombineInfo(DAG, Level, false, this);
1404 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1408 // If nothing happened still, try promoting the operation.
1409 if (!RV.getNode()) {
1410 switch (N->getOpcode()) {
1418 RV = PromoteIntBinOp(SDValue(N, 0));
1423 RV = PromoteIntShiftOp(SDValue(N, 0));
1425 case ISD::SIGN_EXTEND:
1426 case ISD::ZERO_EXTEND:
1427 case ISD::ANY_EXTEND:
1428 RV = PromoteExtend(SDValue(N, 0));
1431 if (PromoteLoad(SDValue(N, 0)))
1437 // If N is a commutative binary node, try commuting it to enable more
1439 if (!RV.getNode() && SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1440 N->getNumValues() == 1) {
1441 SDValue N0 = N->getOperand(0);
1442 SDValue N1 = N->getOperand(1);
1444 // Constant operands are canonicalized to RHS.
1445 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1446 SDValue Ops[] = {N1, N0};
1448 if (const BinaryWithFlagsSDNode *BinNode =
1449 dyn_cast<BinaryWithFlagsSDNode>(N)) {
1450 CSENode = DAG.getNodeIfExists(
1451 N->getOpcode(), N->getVTList(), Ops, BinNode->hasNoUnsignedWrap(),
1452 BinNode->hasNoSignedWrap(), BinNode->isExact());
1454 CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops);
1457 return SDValue(CSENode, 0);
1464 /// Given a node, return its input chain if it has one, otherwise return a null
1466 static SDValue getInputChainForNode(SDNode *N) {
1467 if (unsigned NumOps = N->getNumOperands()) {
1468 if (N->getOperand(0).getValueType() == MVT::Other)
1469 return N->getOperand(0);
1470 if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1471 return N->getOperand(NumOps-1);
1472 for (unsigned i = 1; i < NumOps-1; ++i)
1473 if (N->getOperand(i).getValueType() == MVT::Other)
1474 return N->getOperand(i);
1479 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1480 // If N has two operands, where one has an input chain equal to the other,
1481 // the 'other' chain is redundant.
1482 if (N->getNumOperands() == 2) {
1483 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1484 return N->getOperand(0);
1485 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1486 return N->getOperand(1);
1489 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1490 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1491 SmallPtrSet<SDNode*, 16> SeenOps;
1492 bool Changed = false; // If we should replace this token factor.
1494 // Start out with this token factor.
1497 // Iterate through token factors. The TFs grows when new token factors are
1499 for (unsigned i = 0; i < TFs.size(); ++i) {
1500 SDNode *TF = TFs[i];
1502 // Check each of the operands.
1503 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1504 SDValue Op = TF->getOperand(i);
1506 switch (Op.getOpcode()) {
1507 case ISD::EntryToken:
1508 // Entry tokens don't need to be added to the list. They are
1513 case ISD::TokenFactor:
1514 if (Op.hasOneUse() &&
1515 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1516 // Queue up for processing.
1517 TFs.push_back(Op.getNode());
1518 // Clean up in case the token factor is removed.
1519 AddToWorklist(Op.getNode());
1526 // Only add if it isn't already in the list.
1527 if (SeenOps.insert(Op.getNode()).second)
1538 // If we've changed things around then replace token factor.
1541 // The entry token is the only possible outcome.
1542 Result = DAG.getEntryNode();
1544 // New and improved token factor.
1545 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Ops);
1548 // Add users to worklist if AA is enabled, since it may introduce
1549 // a lot of new chained token factors while removing memory deps.
1550 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
1551 : DAG.getSubtarget().useAA();
1552 return CombineTo(N, Result, UseAA /*add to worklist*/);
1558 /// MERGE_VALUES can always be eliminated.
1559 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1560 WorklistRemover DeadNodes(*this);
1561 // Replacing results may cause a different MERGE_VALUES to suddenly
1562 // be CSE'd with N, and carry its uses with it. Iterate until no
1563 // uses remain, to ensure that the node can be safely deleted.
1564 // First add the users of this node to the work list so that they
1565 // can be tried again once they have new operands.
1566 AddUsersToWorklist(N);
1568 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1569 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1570 } while (!N->use_empty());
1571 deleteAndRecombine(N);
1572 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1575 SDValue DAGCombiner::visitADD(SDNode *N) {
1576 SDValue N0 = N->getOperand(0);
1577 SDValue N1 = N->getOperand(1);
1578 EVT VT = N0.getValueType();
1581 if (VT.isVector()) {
1582 if (SDValue FoldedVOp = SimplifyVBinOp(N))
1585 // fold (add x, 0) -> x, vector edition
1586 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1588 if (ISD::isBuildVectorAllZeros(N0.getNode()))
1592 // fold (add x, undef) -> undef
1593 if (N0.getOpcode() == ISD::UNDEF)
1595 if (N1.getOpcode() == ISD::UNDEF)
1597 // fold (add c1, c2) -> c1+c2
1598 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1599 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1601 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1602 // canonicalize constant to RHS
1603 if (isConstantIntBuildVectorOrConstantInt(N0) &&
1604 !isConstantIntBuildVectorOrConstantInt(N1))
1605 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, N0);
1606 // fold (add x, 0) -> x
1607 if (N1C && N1C->isNullValue())
1609 // fold (add Sym, c) -> Sym+c
1610 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1611 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1612 GA->getOpcode() == ISD::GlobalAddress)
1613 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1615 (uint64_t)N1C->getSExtValue());
1616 // fold ((c1-A)+c2) -> (c1+c2)-A
1617 if (N1C && N0.getOpcode() == ISD::SUB)
1618 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1619 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1620 DAG.getConstant(N1C->getAPIntValue()+
1621 N0C->getAPIntValue(), VT),
1624 if (SDValue RADD = ReassociateOps(ISD::ADD, SDLoc(N), N0, N1))
1626 // fold ((0-A) + B) -> B-A
1627 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1628 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1629 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, N0.getOperand(1));
1630 // fold (A + (0-B)) -> A-B
1631 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1632 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1633 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1.getOperand(1));
1634 // fold (A+(B-A)) -> B
1635 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1636 return N1.getOperand(0);
1637 // fold ((B-A)+A) -> B
1638 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1639 return N0.getOperand(0);
1640 // fold (A+(B-(A+C))) to (B-C)
1641 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1642 N0 == N1.getOperand(1).getOperand(0))
1643 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1644 N1.getOperand(1).getOperand(1));
1645 // fold (A+(B-(C+A))) to (B-C)
1646 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1647 N0 == N1.getOperand(1).getOperand(1))
1648 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1649 N1.getOperand(1).getOperand(0));
1650 // fold (A+((B-A)+or-C)) to (B+or-C)
1651 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1652 N1.getOperand(0).getOpcode() == ISD::SUB &&
1653 N0 == N1.getOperand(0).getOperand(1))
1654 return DAG.getNode(N1.getOpcode(), SDLoc(N), VT,
1655 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1657 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1658 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1659 SDValue N00 = N0.getOperand(0);
1660 SDValue N01 = N0.getOperand(1);
1661 SDValue N10 = N1.getOperand(0);
1662 SDValue N11 = N1.getOperand(1);
1664 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1665 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1666 DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10),
1667 DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11));
1670 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1671 return SDValue(N, 0);
1673 // fold (a+b) -> (a|b) iff a and b share no bits.
1674 if (VT.isInteger() && !VT.isVector()) {
1675 APInt LHSZero, LHSOne;
1676 APInt RHSZero, RHSOne;
1677 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1679 if (LHSZero.getBoolValue()) {
1680 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1682 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1683 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1684 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero){
1685 if (!LegalOperations || TLI.isOperationLegal(ISD::OR, VT))
1686 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1);
1691 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1692 if (N1.getOpcode() == ISD::SHL &&
1693 N1.getOperand(0).getOpcode() == ISD::SUB)
1694 if (ConstantSDNode *C =
1695 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1696 if (C->getAPIntValue() == 0)
1697 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0,
1698 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1699 N1.getOperand(0).getOperand(1),
1701 if (N0.getOpcode() == ISD::SHL &&
1702 N0.getOperand(0).getOpcode() == ISD::SUB)
1703 if (ConstantSDNode *C =
1704 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1705 if (C->getAPIntValue() == 0)
1706 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1,
1707 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1708 N0.getOperand(0).getOperand(1),
1711 if (N1.getOpcode() == ISD::AND) {
1712 SDValue AndOp0 = N1.getOperand(0);
1713 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1714 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1715 unsigned DestBits = VT.getScalarType().getSizeInBits();
1717 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1718 // and similar xforms where the inner op is either ~0 or 0.
1719 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1721 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1725 // add (sext i1), X -> sub X, (zext i1)
1726 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1727 N0.getOperand(0).getValueType() == MVT::i1 &&
1728 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1730 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1731 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1734 // add X, (sextinreg Y i1) -> sub X, (and Y 1)
1735 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
1736 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
1737 if (TN->getVT() == MVT::i1) {
1739 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
1740 DAG.getConstant(1, VT));
1741 return DAG.getNode(ISD::SUB, DL, VT, N0, ZExt);
1748 SDValue DAGCombiner::visitADDC(SDNode *N) {
1749 SDValue N0 = N->getOperand(0);
1750 SDValue N1 = N->getOperand(1);
1751 EVT VT = N0.getValueType();
1753 // If the flag result is dead, turn this into an ADD.
1754 if (!N->hasAnyUseOfValue(1))
1755 return CombineTo(N, DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N1),
1756 DAG.getNode(ISD::CARRY_FALSE,
1757 SDLoc(N), MVT::Glue));
1759 // canonicalize constant to RHS.
1760 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1761 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1763 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N1, N0);
1765 // fold (addc x, 0) -> x + no carry out
1766 if (N1C && N1C->isNullValue())
1767 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1768 SDLoc(N), MVT::Glue));
1770 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1771 APInt LHSZero, LHSOne;
1772 APInt RHSZero, RHSOne;
1773 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1775 if (LHSZero.getBoolValue()) {
1776 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1778 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1779 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1780 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1781 return CombineTo(N, DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1),
1782 DAG.getNode(ISD::CARRY_FALSE,
1783 SDLoc(N), MVT::Glue));
1789 SDValue DAGCombiner::visitADDE(SDNode *N) {
1790 SDValue N0 = N->getOperand(0);
1791 SDValue N1 = N->getOperand(1);
1792 SDValue CarryIn = N->getOperand(2);
1794 // canonicalize constant to RHS
1795 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1796 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1798 return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(),
1801 // fold (adde x, y, false) -> (addc x, y)
1802 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1803 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1);
1808 // Since it may not be valid to emit a fold to zero for vector initializers
1809 // check if we can before folding.
1810 static SDValue tryFoldToZero(SDLoc DL, const TargetLowering &TLI, EVT VT,
1812 bool LegalOperations, bool LegalTypes) {
1814 return DAG.getConstant(0, VT);
1815 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
1816 return DAG.getConstant(0, VT);
1820 SDValue DAGCombiner::visitSUB(SDNode *N) {
1821 SDValue N0 = N->getOperand(0);
1822 SDValue N1 = N->getOperand(1);
1823 EVT VT = N0.getValueType();
1826 if (VT.isVector()) {
1827 if (SDValue FoldedVOp = SimplifyVBinOp(N))
1830 // fold (sub x, 0) -> x, vector edition
1831 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1835 // fold (sub x, x) -> 0
1836 // FIXME: Refactor this and xor and other similar operations together.
1838 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
1839 // fold (sub c1, c2) -> c1-c2
1840 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1841 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1843 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1844 // fold (sub x, c) -> (add x, -c)
1846 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0,
1847 DAG.getConstant(-N1C->getAPIntValue(), VT));
1848 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1849 if (N0C && N0C->isAllOnesValue())
1850 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
1851 // fold A-(A-B) -> B
1852 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1853 return N1.getOperand(1);
1854 // fold (A+B)-A -> B
1855 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1856 return N0.getOperand(1);
1857 // fold (A+B)-B -> A
1858 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1859 return N0.getOperand(0);
1860 // fold C2-(A+C1) -> (C2-C1)-A
1861 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? nullptr :
1862 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1863 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1864 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(),
1866 return DAG.getNode(ISD::SUB, SDLoc(N), VT, NewC,
1869 // fold ((A+(B+or-C))-B) -> A+or-C
1870 if (N0.getOpcode() == ISD::ADD &&
1871 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1872 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1873 N0.getOperand(1).getOperand(0) == N1)
1874 return DAG.getNode(N0.getOperand(1).getOpcode(), SDLoc(N), VT,
1875 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1876 // fold ((A+(C+B))-B) -> A+C
1877 if (N0.getOpcode() == ISD::ADD &&
1878 N0.getOperand(1).getOpcode() == ISD::ADD &&
1879 N0.getOperand(1).getOperand(1) == N1)
1880 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1881 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1882 // fold ((A-(B-C))-C) -> A-B
1883 if (N0.getOpcode() == ISD::SUB &&
1884 N0.getOperand(1).getOpcode() == ISD::SUB &&
1885 N0.getOperand(1).getOperand(1) == N1)
1886 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1887 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1889 // If either operand of a sub is undef, the result is undef
1890 if (N0.getOpcode() == ISD::UNDEF)
1892 if (N1.getOpcode() == ISD::UNDEF)
1895 // If the relocation model supports it, consider symbol offsets.
1896 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1897 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1898 // fold (sub Sym, c) -> Sym-c
1899 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1900 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1902 (uint64_t)N1C->getSExtValue());
1903 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1904 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1905 if (GA->getGlobal() == GB->getGlobal())
1906 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1910 // sub X, (sextinreg Y i1) -> add X, (and Y 1)
1911 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
1912 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
1913 if (TN->getVT() == MVT::i1) {
1915 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
1916 DAG.getConstant(1, VT));
1917 return DAG.getNode(ISD::ADD, DL, VT, N0, ZExt);
1924 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1925 SDValue N0 = N->getOperand(0);
1926 SDValue N1 = N->getOperand(1);
1927 EVT VT = N0.getValueType();
1929 // If the flag result is dead, turn this into an SUB.
1930 if (!N->hasAnyUseOfValue(1))
1931 return CombineTo(N, DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1),
1932 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1935 // fold (subc x, x) -> 0 + no borrow
1937 return CombineTo(N, DAG.getConstant(0, VT),
1938 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1941 // fold (subc x, 0) -> x + no borrow
1942 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1943 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1944 if (N1C && N1C->isNullValue())
1945 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1948 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1949 if (N0C && N0C->isAllOnesValue())
1950 return CombineTo(N, DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0),
1951 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1957 SDValue DAGCombiner::visitSUBE(SDNode *N) {
1958 SDValue N0 = N->getOperand(0);
1959 SDValue N1 = N->getOperand(1);
1960 SDValue CarryIn = N->getOperand(2);
1962 // fold (sube x, y, false) -> (subc x, y)
1963 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1964 return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1);
1969 SDValue DAGCombiner::visitMUL(SDNode *N) {
1970 SDValue N0 = N->getOperand(0);
1971 SDValue N1 = N->getOperand(1);
1972 EVT VT = N0.getValueType();
1974 // fold (mul x, undef) -> 0
1975 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1976 return DAG.getConstant(0, VT);
1978 bool N0IsConst = false;
1979 bool N1IsConst = false;
1980 APInt ConstValue0, ConstValue1;
1982 if (VT.isVector()) {
1983 if (SDValue FoldedVOp = SimplifyVBinOp(N))
1986 N0IsConst = isConstantSplatVector(N0.getNode(), ConstValue0);
1987 N1IsConst = isConstantSplatVector(N1.getNode(), ConstValue1);
1989 N0IsConst = isa<ConstantSDNode>(N0);
1991 ConstValue0 = cast<ConstantSDNode>(N0)->getAPIntValue();
1992 N1IsConst = isa<ConstantSDNode>(N1);
1994 ConstValue1 = cast<ConstantSDNode>(N1)->getAPIntValue();
1997 // fold (mul c1, c2) -> c1*c2
1998 if (N0IsConst && N1IsConst)
1999 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0.getNode(), N1.getNode());
2001 // canonicalize constant to RHS (vector doesn't have to splat)
2002 if (isConstantIntBuildVectorOrConstantInt(N0) &&
2003 !isConstantIntBuildVectorOrConstantInt(N1))
2004 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0);
2005 // fold (mul x, 0) -> 0
2006 if (N1IsConst && ConstValue1 == 0)
2008 // We require a splat of the entire scalar bit width for non-contiguous
2011 ConstValue1.getBitWidth() == VT.getScalarType().getSizeInBits();
2012 // fold (mul x, 1) -> x
2013 if (N1IsConst && ConstValue1 == 1 && IsFullSplat)
2015 // fold (mul x, -1) -> 0-x
2016 if (N1IsConst && ConstValue1.isAllOnesValue())
2017 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2018 DAG.getConstant(0, VT), N0);
2019 // fold (mul x, (1 << c)) -> x << c
2020 if (N1IsConst && ConstValue1.isPowerOf2() && IsFullSplat)
2021 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
2022 DAG.getConstant(ConstValue1.logBase2(),
2023 getShiftAmountTy(N0.getValueType())));
2024 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
2025 if (N1IsConst && (-ConstValue1).isPowerOf2() && IsFullSplat) {
2026 unsigned Log2Val = (-ConstValue1).logBase2();
2027 // FIXME: If the input is something that is easily negated (e.g. a
2028 // single-use add), we should put the negate there.
2029 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2030 DAG.getConstant(0, VT),
2031 DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
2032 DAG.getConstant(Log2Val,
2033 getShiftAmountTy(N0.getValueType()))));
2037 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
2038 if (N1IsConst && N0.getOpcode() == ISD::SHL &&
2039 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2040 isa<ConstantSDNode>(N0.getOperand(1)))) {
2041 SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT,
2042 N1, N0.getOperand(1));
2043 AddToWorklist(C3.getNode());
2044 return DAG.getNode(ISD::MUL, SDLoc(N), VT,
2045 N0.getOperand(0), C3);
2048 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
2051 SDValue Sh(nullptr,0), Y(nullptr,0);
2052 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
2053 if (N0.getOpcode() == ISD::SHL &&
2054 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2055 isa<ConstantSDNode>(N0.getOperand(1))) &&
2056 N0.getNode()->hasOneUse()) {
2058 } else if (N1.getOpcode() == ISD::SHL &&
2059 isa<ConstantSDNode>(N1.getOperand(1)) &&
2060 N1.getNode()->hasOneUse()) {
2065 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2066 Sh.getOperand(0), Y);
2067 return DAG.getNode(ISD::SHL, SDLoc(N), VT,
2068 Mul, Sh.getOperand(1));
2072 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
2073 if (N1IsConst && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
2074 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2075 isa<ConstantSDNode>(N0.getOperand(1))))
2076 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
2077 DAG.getNode(ISD::MUL, SDLoc(N0), VT,
2078 N0.getOperand(0), N1),
2079 DAG.getNode(ISD::MUL, SDLoc(N1), VT,
2080 N0.getOperand(1), N1));
2083 if (SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1))
2089 SDValue DAGCombiner::visitSDIV(SDNode *N) {
2090 SDValue N0 = N->getOperand(0);
2091 SDValue N1 = N->getOperand(1);
2092 EVT VT = N->getValueType(0);
2096 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2099 // fold (sdiv c1, c2) -> c1/c2
2100 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2101 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2102 if (N0C && N1C && !N1C->isNullValue())
2103 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
2104 // fold (sdiv X, 1) -> X
2105 if (N1C && N1C->getAPIntValue() == 1LL)
2107 // fold (sdiv X, -1) -> 0-X
2108 if (N1C && N1C->isAllOnesValue())
2109 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2110 DAG.getConstant(0, VT), N0);
2111 // If we know the sign bits of both operands are zero, strength reduce to a
2112 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
2113 if (!VT.isVector()) {
2114 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2115 return DAG.getNode(ISD::UDIV, SDLoc(N), N1.getValueType(),
2119 // fold (sdiv X, pow2) -> simple ops after legalize
2120 if (N1C && !N1C->isNullValue() && (N1C->getAPIntValue().isPowerOf2() ||
2121 (-N1C->getAPIntValue()).isPowerOf2())) {
2122 // If dividing by powers of two is cheap, then don't perform the following
2124 if (TLI.isPow2SDivCheap())
2127 // Target-specific implementation of sdiv x, pow2.
2128 SDValue Res = BuildSDIVPow2(N);
2132 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
2134 // Splat the sign bit into the register
2136 DAG.getNode(ISD::SRA, SDLoc(N), VT, N0,
2137 DAG.getConstant(VT.getScalarSizeInBits() - 1,
2138 getShiftAmountTy(N0.getValueType())));
2139 AddToWorklist(SGN.getNode());
2141 // Add (N0 < 0) ? abs2 - 1 : 0;
2143 DAG.getNode(ISD::SRL, SDLoc(N), VT, SGN,
2144 DAG.getConstant(VT.getScalarSizeInBits() - lg2,
2145 getShiftAmountTy(SGN.getValueType())));
2146 SDValue ADD = DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, SRL);
2147 AddToWorklist(SRL.getNode());
2148 AddToWorklist(ADD.getNode()); // Divide by pow2
2149 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), VT, ADD,
2150 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType())));
2152 // If we're dividing by a positive value, we're done. Otherwise, we must
2153 // negate the result.
2154 if (N1C->getAPIntValue().isNonNegative())
2157 AddToWorklist(SRA.getNode());
2158 return DAG.getNode(ISD::SUB, SDLoc(N), VT, DAG.getConstant(0, VT), SRA);
2161 // If integer divide is expensive and we satisfy the requirements, emit an
2162 // alternate sequence.
2163 if (N1C && !TLI.isIntDivCheap()) {
2164 SDValue Op = BuildSDIV(N);
2165 if (Op.getNode()) return Op;
2169 if (N0.getOpcode() == ISD::UNDEF)
2170 return DAG.getConstant(0, VT);
2171 // X / undef -> undef
2172 if (N1.getOpcode() == ISD::UNDEF)
2178 SDValue DAGCombiner::visitUDIV(SDNode *N) {
2179 SDValue N0 = N->getOperand(0);
2180 SDValue N1 = N->getOperand(1);
2181 EVT VT = N->getValueType(0);
2185 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2188 // fold (udiv c1, c2) -> c1/c2
2189 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2190 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2191 if (N0C && N1C && !N1C->isNullValue())
2192 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
2193 // fold (udiv x, (1 << c)) -> x >>u c
2194 if (N1C && N1C->getAPIntValue().isPowerOf2())
2195 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0,
2196 DAG.getConstant(N1C->getAPIntValue().logBase2(),
2197 getShiftAmountTy(N0.getValueType())));
2198 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
2199 if (N1.getOpcode() == ISD::SHL) {
2200 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2201 if (SHC->getAPIntValue().isPowerOf2()) {
2202 EVT ADDVT = N1.getOperand(1).getValueType();
2203 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N), ADDVT,
2205 DAG.getConstant(SHC->getAPIntValue()
2208 AddToWorklist(Add.getNode());
2209 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, Add);
2213 // fold (udiv x, c) -> alternate
2214 if (N1C && !TLI.isIntDivCheap()) {
2215 SDValue Op = BuildUDIV(N);
2216 if (Op.getNode()) return Op;
2220 if (N0.getOpcode() == ISD::UNDEF)
2221 return DAG.getConstant(0, VT);
2222 // X / undef -> undef
2223 if (N1.getOpcode() == ISD::UNDEF)
2229 SDValue DAGCombiner::visitSREM(SDNode *N) {
2230 SDValue N0 = N->getOperand(0);
2231 SDValue N1 = N->getOperand(1);
2232 EVT VT = N->getValueType(0);
2234 // fold (srem c1, c2) -> c1%c2
2235 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2236 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2237 if (N0C && N1C && !N1C->isNullValue())
2238 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
2239 // If we know the sign bits of both operands are zero, strength reduce to a
2240 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
2241 if (!VT.isVector()) {
2242 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2243 return DAG.getNode(ISD::UREM, SDLoc(N), VT, N0, N1);
2246 // If X/C can be simplified by the division-by-constant logic, lower
2247 // X%C to the equivalent of X-X/C*C.
2248 if (N1C && !N1C->isNullValue()) {
2249 SDValue Div = DAG.getNode(ISD::SDIV, SDLoc(N), VT, N0, N1);
2250 AddToWorklist(Div.getNode());
2251 SDValue OptimizedDiv = combine(Div.getNode());
2252 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2253 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2255 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2256 AddToWorklist(Mul.getNode());
2262 if (N0.getOpcode() == ISD::UNDEF)
2263 return DAG.getConstant(0, VT);
2264 // X % undef -> undef
2265 if (N1.getOpcode() == ISD::UNDEF)
2271 SDValue DAGCombiner::visitUREM(SDNode *N) {
2272 SDValue N0 = N->getOperand(0);
2273 SDValue N1 = N->getOperand(1);
2274 EVT VT = N->getValueType(0);
2276 // fold (urem c1, c2) -> c1%c2
2277 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2278 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2279 if (N0C && N1C && !N1C->isNullValue())
2280 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
2281 // fold (urem x, pow2) -> (and x, pow2-1)
2282 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
2283 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0,
2284 DAG.getConstant(N1C->getAPIntValue()-1,VT));
2285 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2286 if (N1.getOpcode() == ISD::SHL) {
2287 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2288 if (SHC->getAPIntValue().isPowerOf2()) {
2290 DAG.getNode(ISD::ADD, SDLoc(N), VT, N1,
2291 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
2293 AddToWorklist(Add.getNode());
2294 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, Add);
2299 // If X/C can be simplified by the division-by-constant logic, lower
2300 // X%C to the equivalent of X-X/C*C.
2301 if (N1C && !N1C->isNullValue()) {
2302 SDValue Div = DAG.getNode(ISD::UDIV, SDLoc(N), VT, N0, N1);
2303 AddToWorklist(Div.getNode());
2304 SDValue OptimizedDiv = combine(Div.getNode());
2305 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2306 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2308 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2309 AddToWorklist(Mul.getNode());
2315 if (N0.getOpcode() == ISD::UNDEF)
2316 return DAG.getConstant(0, VT);
2317 // X % undef -> undef
2318 if (N1.getOpcode() == ISD::UNDEF)
2324 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2325 SDValue N0 = N->getOperand(0);
2326 SDValue N1 = N->getOperand(1);
2327 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2328 EVT VT = N->getValueType(0);
2331 // fold (mulhs x, 0) -> 0
2332 if (N1C && N1C->isNullValue())
2334 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2335 if (N1C && N1C->getAPIntValue() == 1)
2336 return DAG.getNode(ISD::SRA, SDLoc(N), N0.getValueType(), N0,
2337 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2338 getShiftAmountTy(N0.getValueType())));
2339 // fold (mulhs x, undef) -> 0
2340 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2341 return DAG.getConstant(0, VT);
2343 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2345 if (VT.isSimple() && !VT.isVector()) {
2346 MVT Simple = VT.getSimpleVT();
2347 unsigned SimpleSize = Simple.getSizeInBits();
2348 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2349 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2350 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2351 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2352 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2353 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2354 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2355 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2362 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2363 SDValue N0 = N->getOperand(0);
2364 SDValue N1 = N->getOperand(1);
2365 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2366 EVT VT = N->getValueType(0);
2369 // fold (mulhu x, 0) -> 0
2370 if (N1C && N1C->isNullValue())
2372 // fold (mulhu x, 1) -> 0
2373 if (N1C && N1C->getAPIntValue() == 1)
2374 return DAG.getConstant(0, N0.getValueType());
2375 // fold (mulhu x, undef) -> 0
2376 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2377 return DAG.getConstant(0, VT);
2379 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2381 if (VT.isSimple() && !VT.isVector()) {
2382 MVT Simple = VT.getSimpleVT();
2383 unsigned SimpleSize = Simple.getSizeInBits();
2384 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2385 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2386 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2387 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2388 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2389 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2390 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2391 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2398 /// Perform optimizations common to nodes that compute two values. LoOp and HiOp
2399 /// give the opcodes for the two computations that are being performed. Return
2400 /// true if a simplification was made.
2401 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2403 // If the high half is not needed, just compute the low half.
2404 bool HiExists = N->hasAnyUseOfValue(1);
2406 (!LegalOperations ||
2407 TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) {
2408 SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2409 return CombineTo(N, Res, Res);
2412 // If the low half is not needed, just compute the high half.
2413 bool LoExists = N->hasAnyUseOfValue(0);
2415 (!LegalOperations ||
2416 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2417 SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2418 return CombineTo(N, Res, Res);
2421 // If both halves are used, return as it is.
2422 if (LoExists && HiExists)
2425 // If the two computed results can be simplified separately, separate them.
2427 SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2428 AddToWorklist(Lo.getNode());
2429 SDValue LoOpt = combine(Lo.getNode());
2430 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2431 (!LegalOperations ||
2432 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2433 return CombineTo(N, LoOpt, LoOpt);
2437 SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2438 AddToWorklist(Hi.getNode());
2439 SDValue HiOpt = combine(Hi.getNode());
2440 if (HiOpt.getNode() && HiOpt != Hi &&
2441 (!LegalOperations ||
2442 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2443 return CombineTo(N, HiOpt, HiOpt);
2449 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2450 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2451 if (Res.getNode()) return Res;
2453 EVT VT = N->getValueType(0);
2456 // If the type is twice as wide is legal, transform the mulhu to a wider
2457 // multiply plus a shift.
2458 if (VT.isSimple() && !VT.isVector()) {
2459 MVT Simple = VT.getSimpleVT();
2460 unsigned SimpleSize = Simple.getSizeInBits();
2461 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2462 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2463 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2464 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2465 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2466 // Compute the high part as N1.
2467 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2468 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2469 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2470 // Compute the low part as N0.
2471 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2472 return CombineTo(N, Lo, Hi);
2479 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2480 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2481 if (Res.getNode()) return Res;
2483 EVT VT = N->getValueType(0);
2486 // If the type is twice as wide is legal, transform the mulhu to a wider
2487 // multiply plus a shift.
2488 if (VT.isSimple() && !VT.isVector()) {
2489 MVT Simple = VT.getSimpleVT();
2490 unsigned SimpleSize = Simple.getSizeInBits();
2491 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2492 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2493 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2494 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2495 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2496 // Compute the high part as N1.
2497 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2498 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2499 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2500 // Compute the low part as N0.
2501 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2502 return CombineTo(N, Lo, Hi);
2509 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2510 // (smulo x, 2) -> (saddo x, x)
2511 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2512 if (C2->getAPIntValue() == 2)
2513 return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(),
2514 N->getOperand(0), N->getOperand(0));
2519 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2520 // (umulo x, 2) -> (uaddo x, x)
2521 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2522 if (C2->getAPIntValue() == 2)
2523 return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(),
2524 N->getOperand(0), N->getOperand(0));
2529 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2530 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2531 if (Res.getNode()) return Res;
2536 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2537 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2538 if (Res.getNode()) return Res;
2543 /// If this is a binary operator with two operands of the same opcode, try to
2545 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2546 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2547 EVT VT = N0.getValueType();
2548 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2550 // Bail early if none of these transforms apply.
2551 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2553 // For each of OP in AND/OR/XOR:
2554 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2555 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2556 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2557 // fold (OP (bswap x), (bswap y)) -> (bswap (OP x, y))
2558 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2560 // do not sink logical op inside of a vector extend, since it may combine
2562 EVT Op0VT = N0.getOperand(0).getValueType();
2563 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2564 N0.getOpcode() == ISD::SIGN_EXTEND ||
2565 N0.getOpcode() == ISD::BSWAP ||
2566 // Avoid infinite looping with PromoteIntBinOp.
2567 (N0.getOpcode() == ISD::ANY_EXTEND &&
2568 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2569 (N0.getOpcode() == ISD::TRUNCATE &&
2570 (!TLI.isZExtFree(VT, Op0VT) ||
2571 !TLI.isTruncateFree(Op0VT, VT)) &&
2572 TLI.isTypeLegal(Op0VT))) &&
2574 Op0VT == N1.getOperand(0).getValueType() &&
2575 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2576 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2577 N0.getOperand(0).getValueType(),
2578 N0.getOperand(0), N1.getOperand(0));
2579 AddToWorklist(ORNode.getNode());
2580 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode);
2583 // For each of OP in SHL/SRL/SRA/AND...
2584 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2585 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2586 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2587 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2588 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2589 N0.getOperand(1) == N1.getOperand(1)) {
2590 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2591 N0.getOperand(0).getValueType(),
2592 N0.getOperand(0), N1.getOperand(0));
2593 AddToWorklist(ORNode.getNode());
2594 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
2595 ORNode, N0.getOperand(1));
2598 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2599 // Only perform this optimization after type legalization and before
2600 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2601 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2602 // we don't want to undo this promotion.
2603 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2605 if ((N0.getOpcode() == ISD::BITCAST ||
2606 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) &&
2607 Level == AfterLegalizeTypes) {
2608 SDValue In0 = N0.getOperand(0);
2609 SDValue In1 = N1.getOperand(0);
2610 EVT In0Ty = In0.getValueType();
2611 EVT In1Ty = In1.getValueType();
2613 // If both incoming values are integers, and the original types are the
2615 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2616 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1);
2617 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op);
2618 AddToWorklist(Op.getNode());
2623 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2624 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2625 // If both shuffles use the same mask, and both shuffle within a single
2626 // vector, then it is worthwhile to move the swizzle after the operation.
2627 // The type-legalizer generates this pattern when loading illegal
2628 // vector types from memory. In many cases this allows additional shuffle
2630 // There are other cases where moving the shuffle after the xor/and/or
2631 // is profitable even if shuffles don't perform a swizzle.
2632 // If both shuffles use the same mask, and both shuffles have the same first
2633 // or second operand, then it might still be profitable to move the shuffle
2634 // after the xor/and/or operation.
2635 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) {
2636 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2637 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2639 assert(N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() &&
2640 "Inputs to shuffles are not the same type");
2642 // Check that both shuffles use the same mask. The masks are known to be of
2643 // the same length because the result vector type is the same.
2644 // Check also that shuffles have only one use to avoid introducing extra
2646 if (SVN0->hasOneUse() && SVN1->hasOneUse() &&
2647 SVN0->getMask().equals(SVN1->getMask())) {
2648 SDValue ShOp = N0->getOperand(1);
2650 // Don't try to fold this node if it requires introducing a
2651 // build vector of all zeros that might be illegal at this stage.
2652 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2654 ShOp = DAG.getConstant(0, VT);
2659 // (AND (shuf (A, C), shuf (B, C)) -> shuf (AND (A, B), C)
2660 // (OR (shuf (A, C), shuf (B, C)) -> shuf (OR (A, B), C)
2661 // (XOR (shuf (A, C), shuf (B, C)) -> shuf (XOR (A, B), V_0)
2662 if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) {
2663 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2664 N0->getOperand(0), N1->getOperand(0));
2665 AddToWorklist(NewNode.getNode());
2666 return DAG.getVectorShuffle(VT, SDLoc(N), NewNode, ShOp,
2667 &SVN0->getMask()[0]);
2670 // Don't try to fold this node if it requires introducing a
2671 // build vector of all zeros that might be illegal at this stage.
2672 ShOp = N0->getOperand(0);
2673 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2675 ShOp = DAG.getConstant(0, VT);
2680 // (AND (shuf (C, A), shuf (C, B)) -> shuf (C, AND (A, B))
2681 // (OR (shuf (C, A), shuf (C, B)) -> shuf (C, OR (A, B))
2682 // (XOR (shuf (C, A), shuf (C, B)) -> shuf (V_0, XOR (A, B))
2683 if (N0->getOperand(0) == N1->getOperand(0) && ShOp.getNode()) {
2684 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2685 N0->getOperand(1), N1->getOperand(1));
2686 AddToWorklist(NewNode.getNode());
2687 return DAG.getVectorShuffle(VT, SDLoc(N), ShOp, NewNode,
2688 &SVN0->getMask()[0]);
2696 /// This contains all DAGCombine rules which reduce two values combined by
2697 /// an And operation to a single value. This makes them reusable in the context
2698 /// of visitSELECT(). Rules involving constants are not included as
2699 /// visitSELECT() already handles those cases.
2700 SDValue DAGCombiner::visitANDLike(SDValue N0, SDValue N1,
2701 SDNode *LocReference) {
2702 EVT VT = N1.getValueType();
2704 // fold (and x, undef) -> 0
2705 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2706 return DAG.getConstant(0, VT);
2707 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2708 SDValue LL, LR, RL, RR, CC0, CC1;
2709 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2710 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2711 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2713 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2714 LL.getValueType().isInteger()) {
2715 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2716 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
2717 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2718 LR.getValueType(), LL, RL);
2719 AddToWorklist(ORNode.getNode());
2720 return DAG.getSetCC(SDLoc(LocReference), VT, ORNode, LR, Op1);
2722 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2723 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2724 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(N0),
2725 LR.getValueType(), LL, RL);
2726 AddToWorklist(ANDNode.getNode());
2727 return DAG.getSetCC(SDLoc(LocReference), VT, ANDNode, LR, Op1);
2729 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2730 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2731 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2732 LR.getValueType(), LL, RL);
2733 AddToWorklist(ORNode.getNode());
2734 return DAG.getSetCC(SDLoc(LocReference), VT, ORNode, LR, Op1);
2737 // Simplify (and (setne X, 0), (setne X, -1)) -> (setuge (add X, 1), 2)
2738 if (LL == RL && isa<ConstantSDNode>(LR) && isa<ConstantSDNode>(RR) &&
2739 Op0 == Op1 && LL.getValueType().isInteger() &&
2740 Op0 == ISD::SETNE && ((cast<ConstantSDNode>(LR)->isNullValue() &&
2741 cast<ConstantSDNode>(RR)->isAllOnesValue()) ||
2742 (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
2743 cast<ConstantSDNode>(RR)->isNullValue()))) {
2744 SDValue ADDNode = DAG.getNode(ISD::ADD, SDLoc(N0), LL.getValueType(),
2745 LL, DAG.getConstant(1, LL.getValueType()));
2746 AddToWorklist(ADDNode.getNode());
2747 return DAG.getSetCC(SDLoc(LocReference), VT, ADDNode,
2748 DAG.getConstant(2, LL.getValueType()), ISD::SETUGE);
2750 // canonicalize equivalent to ll == rl
2751 if (LL == RR && LR == RL) {
2752 Op1 = ISD::getSetCCSwappedOperands(Op1);
2755 if (LL == RL && LR == RR) {
2756 bool isInteger = LL.getValueType().isInteger();
2757 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2758 if (Result != ISD::SETCC_INVALID &&
2759 (!LegalOperations ||
2760 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
2761 TLI.isOperationLegal(ISD::SETCC,
2762 getSetCCResultType(N0.getSimpleValueType())))))
2763 return DAG.getSetCC(SDLoc(LocReference), N0.getValueType(),
2768 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
2769 VT.getSizeInBits() <= 64) {
2770 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2771 APInt ADDC = ADDI->getAPIntValue();
2772 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2773 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
2774 // immediate for an add, but it is legal if its top c2 bits are set,
2775 // transform the ADD so the immediate doesn't need to be materialized
2777 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
2778 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
2779 SRLI->getZExtValue());
2780 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
2782 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2784 DAG.getNode(ISD::ADD, SDLoc(N0), VT,
2785 N0.getOperand(0), DAG.getConstant(ADDC, VT));
2786 CombineTo(N0.getNode(), NewAdd);
2787 // Return N so it doesn't get rechecked!
2788 return SDValue(LocReference, 0);
2799 SDValue DAGCombiner::visitAND(SDNode *N) {
2800 SDValue N0 = N->getOperand(0);
2801 SDValue N1 = N->getOperand(1);
2802 EVT VT = N1.getValueType();
2805 if (VT.isVector()) {
2806 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2809 // fold (and x, 0) -> 0, vector edition
2810 if (ISD::isBuildVectorAllZeros(N0.getNode()))
2811 // do not return N0, because undef node may exist in N0
2812 return DAG.getConstant(
2813 APInt::getNullValue(
2814 N0.getValueType().getScalarType().getSizeInBits()),
2816 if (ISD::isBuildVectorAllZeros(N1.getNode()))
2817 // do not return N1, because undef node may exist in N1
2818 return DAG.getConstant(
2819 APInt::getNullValue(
2820 N1.getValueType().getScalarType().getSizeInBits()),
2823 // fold (and x, -1) -> x, vector edition
2824 if (ISD::isBuildVectorAllOnes(N0.getNode()))
2826 if (ISD::isBuildVectorAllOnes(N1.getNode()))
2830 // fold (and c1, c2) -> c1&c2
2831 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2832 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2834 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2835 // canonicalize constant to RHS
2836 if (isConstantIntBuildVectorOrConstantInt(N0) &&
2837 !isConstantIntBuildVectorOrConstantInt(N1))
2838 return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0);
2839 // fold (and x, -1) -> x
2840 if (N1C && N1C->isAllOnesValue())
2842 // if (and x, c) is known to be zero, return 0
2843 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2844 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2845 APInt::getAllOnesValue(BitWidth)))
2846 return DAG.getConstant(0, VT);
2848 if (SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1))
2850 // fold (and (or x, C), D) -> D if (C & D) == D
2851 if (N1C && N0.getOpcode() == ISD::OR)
2852 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2853 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2855 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2856 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2857 SDValue N0Op0 = N0.getOperand(0);
2858 APInt Mask = ~N1C->getAPIntValue();
2859 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2860 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2861 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
2862 N0.getValueType(), N0Op0);
2864 // Replace uses of the AND with uses of the Zero extend node.
2867 // We actually want to replace all uses of the any_extend with the
2868 // zero_extend, to avoid duplicating things. This will later cause this
2869 // AND to be folded.
2870 CombineTo(N0.getNode(), Zext);
2871 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2874 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2875 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2876 // already be zero by virtue of the width of the base type of the load.
2878 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2880 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2881 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2882 N0.getOpcode() == ISD::LOAD) {
2883 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2884 N0 : N0.getOperand(0) );
2886 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2887 // This can be a pure constant or a vector splat, in which case we treat the
2888 // vector as a scalar and use the splat value.
2889 APInt Constant = APInt::getNullValue(1);
2890 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2891 Constant = C->getAPIntValue();
2892 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2893 APInt SplatValue, SplatUndef;
2894 unsigned SplatBitSize;
2896 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2897 SplatBitSize, HasAnyUndefs);
2899 // Undef bits can contribute to a possible optimisation if set, so
2901 SplatValue |= SplatUndef;
2903 // The splat value may be something like "0x00FFFFFF", which means 0 for
2904 // the first vector value and FF for the rest, repeating. We need a mask
2905 // that will apply equally to all members of the vector, so AND all the
2906 // lanes of the constant together.
2907 EVT VT = Vector->getValueType(0);
2908 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2910 // If the splat value has been compressed to a bitlength lower
2911 // than the size of the vector lane, we need to re-expand it to
2913 if (BitWidth > SplatBitSize)
2914 for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
2915 SplatBitSize < BitWidth;
2916 SplatBitSize = SplatBitSize * 2)
2917 SplatValue |= SplatValue.shl(SplatBitSize);
2919 // Make sure that variable 'Constant' is only set if 'SplatBitSize' is a
2920 // multiple of 'BitWidth'. Otherwise, we could propagate a wrong value.
2921 if (SplatBitSize % BitWidth == 0) {
2922 Constant = APInt::getAllOnesValue(BitWidth);
2923 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
2924 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
2929 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
2930 // actually legal and isn't going to get expanded, else this is a false
2932 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
2933 Load->getValueType(0),
2934 Load->getMemoryVT());
2936 // Resize the constant to the same size as the original memory access before
2937 // extension. If it is still the AllOnesValue then this AND is completely
2940 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
2943 switch (Load->getExtensionType()) {
2944 default: B = false; break;
2945 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
2947 case ISD::NON_EXTLOAD: B = true; break;
2950 if (B && Constant.isAllOnesValue()) {
2951 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
2952 // preserve semantics once we get rid of the AND.
2953 SDValue NewLoad(Load, 0);
2954 if (Load->getExtensionType() == ISD::EXTLOAD) {
2955 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
2956 Load->getValueType(0), SDLoc(Load),
2957 Load->getChain(), Load->getBasePtr(),
2958 Load->getOffset(), Load->getMemoryVT(),
2959 Load->getMemOperand());
2960 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
2961 if (Load->getNumValues() == 3) {
2962 // PRE/POST_INC loads have 3 values.
2963 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
2964 NewLoad.getValue(2) };
2965 CombineTo(Load, To, 3, true);
2967 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
2971 // Fold the AND away, taking care not to fold to the old load node if we
2973 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
2975 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2979 // fold (and (load x), 255) -> (zextload x, i8)
2980 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2981 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2982 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2983 (N0.getOpcode() == ISD::ANY_EXTEND &&
2984 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2985 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2986 LoadSDNode *LN0 = HasAnyExt
2987 ? cast<LoadSDNode>(N0.getOperand(0))
2988 : cast<LoadSDNode>(N0);
2989 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2990 LN0->isUnindexed() && N0.hasOneUse() && SDValue(LN0, 0).hasOneUse()) {
2991 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2992 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2993 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2994 EVT LoadedVT = LN0->getMemoryVT();
2995 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2997 if (ExtVT == LoadedVT &&
2998 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy,
3002 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
3003 LN0->getChain(), LN0->getBasePtr(), ExtVT,
3004 LN0->getMemOperand());
3006 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
3007 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3010 // Do not change the width of a volatile load.
3011 // Do not generate loads of non-round integer types since these can
3012 // be expensive (and would be wrong if the type is not byte sized).
3013 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
3014 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy,
3016 EVT PtrType = LN0->getOperand(1).getValueType();
3018 unsigned Alignment = LN0->getAlignment();
3019 SDValue NewPtr = LN0->getBasePtr();
3021 // For big endian targets, we need to add an offset to the pointer
3022 // to load the correct bytes. For little endian systems, we merely
3023 // need to read fewer bytes from the same pointer.
3024 if (TLI.isBigEndian()) {
3025 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
3026 unsigned EVTStoreBytes = ExtVT.getStoreSize();
3027 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
3028 NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0), PtrType,
3029 NewPtr, DAG.getConstant(PtrOff, PtrType));
3030 Alignment = MinAlign(Alignment, PtrOff);
3033 AddToWorklist(NewPtr.getNode());
3036 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
3037 LN0->getChain(), NewPtr,
3038 LN0->getPointerInfo(),
3039 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
3040 LN0->isInvariant(), Alignment, LN0->getAAInfo());
3042 CombineTo(LN0, Load, Load.getValue(1));
3043 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3049 if (SDValue Combined = visitANDLike(N0, N1, N))
3052 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
3053 if (N0.getOpcode() == N1.getOpcode()) {
3054 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3055 if (Tmp.getNode()) return Tmp;
3058 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
3059 // fold (and (sra)) -> (and (srl)) when possible.
3060 if (!VT.isVector() &&
3061 SimplifyDemandedBits(SDValue(N, 0)))
3062 return SDValue(N, 0);
3064 // fold (zext_inreg (extload x)) -> (zextload x)
3065 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
3066 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3067 EVT MemVT = LN0->getMemoryVT();
3068 // If we zero all the possible extended bits, then we can turn this into
3069 // a zextload if we are running before legalize or the operation is legal.
3070 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
3071 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
3072 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
3073 ((!LegalOperations && !LN0->isVolatile()) ||
3074 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) {
3075 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
3076 LN0->getChain(), LN0->getBasePtr(),
3077 MemVT, LN0->getMemOperand());
3079 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
3080 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3083 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
3084 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
3086 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3087 EVT MemVT = LN0->getMemoryVT();
3088 // If we zero all the possible extended bits, then we can turn this into
3089 // a zextload if we are running before legalize or the operation is legal.
3090 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
3091 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
3092 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
3093 ((!LegalOperations && !LN0->isVolatile()) ||
3094 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) {
3095 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
3096 LN0->getChain(), LN0->getBasePtr(),
3097 MemVT, LN0->getMemOperand());
3099 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
3100 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3103 // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const)
3104 if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) {
3105 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
3106 N0.getOperand(1), false);
3107 if (BSwap.getNode())
3114 /// Match (a >> 8) | (a << 8) as (bswap a) >> 16.
3115 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
3116 bool DemandHighBits) {
3117 if (!LegalOperations)
3120 EVT VT = N->getValueType(0);
3121 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
3123 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3126 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
3127 bool LookPassAnd0 = false;
3128 bool LookPassAnd1 = false;
3129 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
3131 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
3133 if (N0.getOpcode() == ISD::AND) {
3134 if (!N0.getNode()->hasOneUse())
3136 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3137 if (!N01C || N01C->getZExtValue() != 0xFF00)
3139 N0 = N0.getOperand(0);
3140 LookPassAnd0 = true;
3143 if (N1.getOpcode() == ISD::AND) {
3144 if (!N1.getNode()->hasOneUse())
3146 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3147 if (!N11C || N11C->getZExtValue() != 0xFF)
3149 N1 = N1.getOperand(0);
3150 LookPassAnd1 = true;
3153 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
3155 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
3157 if (!N0.getNode()->hasOneUse() ||
3158 !N1.getNode()->hasOneUse())
3161 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3162 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3165 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
3168 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
3169 SDValue N00 = N0->getOperand(0);
3170 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
3171 if (!N00.getNode()->hasOneUse())
3173 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
3174 if (!N001C || N001C->getZExtValue() != 0xFF)
3176 N00 = N00.getOperand(0);
3177 LookPassAnd0 = true;
3180 SDValue N10 = N1->getOperand(0);
3181 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
3182 if (!N10.getNode()->hasOneUse())
3184 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
3185 if (!N101C || N101C->getZExtValue() != 0xFF00)
3187 N10 = N10.getOperand(0);
3188 LookPassAnd1 = true;
3194 // Make sure everything beyond the low halfword gets set to zero since the SRL
3195 // 16 will clear the top bits.
3196 unsigned OpSizeInBits = VT.getSizeInBits();
3197 if (DemandHighBits && OpSizeInBits > 16) {
3198 // If the left-shift isn't masked out then the only way this is a bswap is
3199 // if all bits beyond the low 8 are 0. In that case the entire pattern
3200 // reduces to a left shift anyway: leave it for other parts of the combiner.
3204 // However, if the right shift isn't masked out then it might be because
3205 // it's not needed. See if we can spot that too.
3206 if (!LookPassAnd1 &&
3207 !DAG.MaskedValueIsZero(
3208 N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16)))
3212 SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00);
3213 if (OpSizeInBits > 16)
3214 Res = DAG.getNode(ISD::SRL, SDLoc(N), VT, Res,
3215 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
3219 /// Return true if the specified node is an element that makes up a 32-bit
3220 /// packed halfword byteswap.
3221 /// ((x & 0x000000ff) << 8) |
3222 /// ((x & 0x0000ff00) >> 8) |
3223 /// ((x & 0x00ff0000) << 8) |
3224 /// ((x & 0xff000000) >> 8)
3225 static bool isBSwapHWordElement(SDValue N, MutableArrayRef<SDNode *> Parts) {
3226 if (!N.getNode()->hasOneUse())
3229 unsigned Opc = N.getOpcode();
3230 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
3233 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3238 switch (N1C->getZExtValue()) {
3241 case 0xFF: Num = 0; break;
3242 case 0xFF00: Num = 1; break;
3243 case 0xFF0000: Num = 2; break;
3244 case 0xFF000000: Num = 3; break;
3247 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
3248 SDValue N0 = N.getOperand(0);
3249 if (Opc == ISD::AND) {
3250 if (Num == 0 || Num == 2) {
3252 // (x >> 8) & 0xff0000
3253 if (N0.getOpcode() != ISD::SRL)
3255 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3256 if (!C || C->getZExtValue() != 8)
3259 // (x << 8) & 0xff00
3260 // (x << 8) & 0xff000000
3261 if (N0.getOpcode() != ISD::SHL)
3263 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3264 if (!C || C->getZExtValue() != 8)
3267 } else if (Opc == ISD::SHL) {
3269 // (x & 0xff0000) << 8
3270 if (Num != 0 && Num != 2)
3272 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3273 if (!C || C->getZExtValue() != 8)
3275 } else { // Opc == ISD::SRL
3276 // (x & 0xff00) >> 8
3277 // (x & 0xff000000) >> 8
3278 if (Num != 1 && Num != 3)
3280 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3281 if (!C || C->getZExtValue() != 8)
3288 Parts[Num] = N0.getOperand(0).getNode();
3292 /// Match a 32-bit packed halfword bswap. That is
3293 /// ((x & 0x000000ff) << 8) |
3294 /// ((x & 0x0000ff00) >> 8) |
3295 /// ((x & 0x00ff0000) << 8) |
3296 /// ((x & 0xff000000) >> 8)
3297 /// => (rotl (bswap x), 16)
3298 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
3299 if (!LegalOperations)
3302 EVT VT = N->getValueType(0);
3305 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3309 // (or (or (and), (and)), (or (and), (and)))
3310 // (or (or (or (and), (and)), (and)), (and))
3311 if (N0.getOpcode() != ISD::OR)
3313 SDValue N00 = N0.getOperand(0);
3314 SDValue N01 = N0.getOperand(1);
3315 SDNode *Parts[4] = {};
3317 if (N1.getOpcode() == ISD::OR &&
3318 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) {
3319 // (or (or (and), (and)), (or (and), (and)))
3320 SDValue N000 = N00.getOperand(0);
3321 if (!isBSwapHWordElement(N000, Parts))
3324 SDValue N001 = N00.getOperand(1);
3325 if (!isBSwapHWordElement(N001, Parts))
3327 SDValue N010 = N01.getOperand(0);
3328 if (!isBSwapHWordElement(N010, Parts))
3330 SDValue N011 = N01.getOperand(1);
3331 if (!isBSwapHWordElement(N011, Parts))
3334 // (or (or (or (and), (and)), (and)), (and))
3335 if (!isBSwapHWordElement(N1, Parts))
3337 if (!isBSwapHWordElement(N01, Parts))
3339 if (N00.getOpcode() != ISD::OR)
3341 SDValue N000 = N00.getOperand(0);
3342 if (!isBSwapHWordElement(N000, Parts))
3344 SDValue N001 = N00.getOperand(1);
3345 if (!isBSwapHWordElement(N001, Parts))
3349 // Make sure the parts are all coming from the same node.
3350 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
3353 SDValue BSwap = DAG.getNode(ISD::BSWAP, SDLoc(N), VT,
3354 SDValue(Parts[0],0));
3356 // Result of the bswap should be rotated by 16. If it's not legal, then
3357 // do (x << 16) | (x >> 16).
3358 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
3359 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3360 return DAG.getNode(ISD::ROTL, SDLoc(N), VT, BSwap, ShAmt);
3361 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3362 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, BSwap, ShAmt);
3363 return DAG.getNode(ISD::OR, SDLoc(N), VT,
3364 DAG.getNode(ISD::SHL, SDLoc(N), VT, BSwap, ShAmt),
3365 DAG.getNode(ISD::SRL, SDLoc(N), VT, BSwap, ShAmt));
3368 /// This contains all DAGCombine rules which reduce two values combined by
3369 /// an Or operation to a single value \see visitANDLike().
3370 SDValue DAGCombiner::visitORLike(SDValue N0, SDValue N1, SDNode *LocReference) {
3371 EVT VT = N1.getValueType();
3372 // fold (or x, undef) -> -1
3373 if (!LegalOperations &&
3374 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3375 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3376 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
3378 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3379 SDValue LL, LR, RL, RR, CC0, CC1;
3380 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3381 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3382 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3384 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3385 LL.getValueType().isInteger()) {
3386 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3387 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3388 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3389 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3390 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(LR),
3391 LR.getValueType(), LL, RL);
3392 AddToWorklist(ORNode.getNode());
3393 return DAG.getSetCC(SDLoc(LocReference), VT, ORNode, LR, Op1);
3395 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3396 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3397 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3398 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3399 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(LR),
3400 LR.getValueType(), LL, RL);
3401 AddToWorklist(ANDNode.getNode());
3402 return DAG.getSetCC(SDLoc(LocReference), VT, ANDNode, LR, Op1);
3405 // canonicalize equivalent to ll == rl
3406 if (LL == RR && LR == RL) {
3407 Op1 = ISD::getSetCCSwappedOperands(Op1);
3410 if (LL == RL && LR == RR) {
3411 bool isInteger = LL.getValueType().isInteger();
3412 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3413 if (Result != ISD::SETCC_INVALID &&
3414 (!LegalOperations ||
3415 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
3416 TLI.isOperationLegal(ISD::SETCC,
3417 getSetCCResultType(N0.getValueType())))))
3418 return DAG.getSetCC(SDLoc(LocReference), N0.getValueType(),
3423 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3424 if (N0.getOpcode() == ISD::AND &&
3425 N1.getOpcode() == ISD::AND &&
3426 N0.getOperand(1).getOpcode() == ISD::Constant &&
3427 N1.getOperand(1).getOpcode() == ISD::Constant &&
3428 // Don't increase # computations.
3429 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3430 // We can only do this xform if we know that bits from X that are set in C2
3431 // but not in C1 are already zero. Likewise for Y.
3432 const APInt &LHSMask =
3433 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3434 const APInt &RHSMask =
3435 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3437 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3438 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3439 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3440 N0.getOperand(0), N1.getOperand(0));
3441 return DAG.getNode(ISD::AND, SDLoc(LocReference), VT, X,
3442 DAG.getConstant(LHSMask | RHSMask, VT));
3446 // (or (and X, M), (and X, N)) -> (and X, (or M, N))
3447 if (N0.getOpcode() == ISD::AND &&
3448 N1.getOpcode() == ISD::AND &&
3449 N0.getOperand(0) == N1.getOperand(0) &&
3450 // Don't increase # computations.
3451 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3452 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3453 N0.getOperand(1), N1.getOperand(1));
3454 return DAG.getNode(ISD::AND, SDLoc(LocReference), VT, N0.getOperand(0), X);
3460 SDValue DAGCombiner::visitOR(SDNode *N) {
3461 SDValue N0 = N->getOperand(0);
3462 SDValue N1 = N->getOperand(1);
3463 EVT VT = N1.getValueType();
3466 if (VT.isVector()) {
3467 if (SDValue FoldedVOp = SimplifyVBinOp(N))
3470 // fold (or x, 0) -> x, vector edition
3471 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3473 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3476 // fold (or x, -1) -> -1, vector edition
3477 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3478 // do not return N0, because undef node may exist in N0
3479 return DAG.getConstant(
3480 APInt::getAllOnesValue(
3481 N0.getValueType().getScalarType().getSizeInBits()),
3483 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3484 // do not return N1, because undef node may exist in N1
3485 return DAG.getConstant(
3486 APInt::getAllOnesValue(
3487 N1.getValueType().getScalarType().getSizeInBits()),
3490 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask1)
3491 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf B, A, Mask2)
3492 // Do this only if the resulting shuffle is legal.
3493 if (isa<ShuffleVectorSDNode>(N0) &&
3494 isa<ShuffleVectorSDNode>(N1) &&
3495 // Avoid folding a node with illegal type.
3496 TLI.isTypeLegal(VT) &&
3497 N0->getOperand(1) == N1->getOperand(1) &&
3498 ISD::isBuildVectorAllZeros(N0.getOperand(1).getNode())) {
3499 bool CanFold = true;
3500 unsigned NumElts = VT.getVectorNumElements();
3501 const ShuffleVectorSDNode *SV0 = cast<ShuffleVectorSDNode>(N0);
3502 const ShuffleVectorSDNode *SV1 = cast<ShuffleVectorSDNode>(N1);
3503 // We construct two shuffle masks:
3504 // - Mask1 is a shuffle mask for a shuffle with N0 as the first operand
3505 // and N1 as the second operand.
3506 // - Mask2 is a shuffle mask for a shuffle with N1 as the first operand
3507 // and N0 as the second operand.
3508 // We do this because OR is commutable and therefore there might be
3509 // two ways to fold this node into a shuffle.
3510 SmallVector<int,4> Mask1;
3511 SmallVector<int,4> Mask2;
3513 for (unsigned i = 0; i != NumElts && CanFold; ++i) {
3514 int M0 = SV0->getMaskElt(i);
3515 int M1 = SV1->getMaskElt(i);
3517 // Both shuffle indexes are undef. Propagate Undef.
3518 if (M0 < 0 && M1 < 0) {
3519 Mask1.push_back(M0);
3520 Mask2.push_back(M0);
3524 if (M0 < 0 || M1 < 0 ||
3525 (M0 < (int)NumElts && M1 < (int)NumElts) ||
3526 (M0 >= (int)NumElts && M1 >= (int)NumElts)) {
3531 Mask1.push_back(M0 < (int)NumElts ? M0 : M1 + NumElts);
3532 Mask2.push_back(M1 < (int)NumElts ? M1 : M0 + NumElts);
3536 // Fold this sequence only if the resulting shuffle is 'legal'.
3537 if (TLI.isShuffleMaskLegal(Mask1, VT))
3538 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0),
3539 N1->getOperand(0), &Mask1[0]);
3540 if (TLI.isShuffleMaskLegal(Mask2, VT))
3541 return DAG.getVectorShuffle(VT, SDLoc(N), N1->getOperand(0),
3542 N0->getOperand(0), &Mask2[0]);
3547 // fold (or c1, c2) -> c1|c2
3548 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3549 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3551 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
3552 // canonicalize constant to RHS
3553 if (isConstantIntBuildVectorOrConstantInt(N0) &&
3554 !isConstantIntBuildVectorOrConstantInt(N1))
3555 return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0);
3556 // fold (or x, 0) -> x
3557 if (N1C && N1C->isNullValue())
3559 // fold (or x, -1) -> -1
3560 if (N1C && N1C->isAllOnesValue())
3562 // fold (or x, c) -> c iff (x & ~c) == 0
3563 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3566 if (SDValue Combined = visitORLike(N0, N1, N))
3569 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3570 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
3571 if (BSwap.getNode())
3573 BSwap = MatchBSwapHWordLow(N, N0, N1);
3574 if (BSwap.getNode())
3578 if (SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1))
3580 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3581 // iff (c1 & c2) == 0.
3582 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3583 isa<ConstantSDNode>(N0.getOperand(1))) {
3584 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3585 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) {
3586 if (SDValue COR = DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1))
3588 ISD::AND, SDLoc(N), VT,
3589 DAG.getNode(ISD::OR, SDLoc(N0), VT, N0.getOperand(0), N1), COR);
3593 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3594 if (N0.getOpcode() == N1.getOpcode()) {
3595 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3596 if (Tmp.getNode()) return Tmp;
3599 // See if this is some rotate idiom.
3600 if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N)))
3601 return SDValue(Rot, 0);
3603 // Simplify the operands using demanded-bits information.
3604 if (!VT.isVector() &&
3605 SimplifyDemandedBits(SDValue(N, 0)))
3606 return SDValue(N, 0);
3611 /// Match "(X shl/srl V1) & V2" where V2 may not be present.
3612 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3613 if (Op.getOpcode() == ISD::AND) {
3614 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3615 Mask = Op.getOperand(1);
3616 Op = Op.getOperand(0);
3622 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3630 // Return true if we can prove that, whenever Neg and Pos are both in the
3631 // range [0, OpSize), Neg == (Pos == 0 ? 0 : OpSize - Pos). This means that
3632 // for two opposing shifts shift1 and shift2 and a value X with OpBits bits:
3634 // (or (shift1 X, Neg), (shift2 X, Pos))
3636 // reduces to a rotate in direction shift2 by Pos or (equivalently) a rotate
3637 // in direction shift1 by Neg. The range [0, OpSize) means that we only need
3638 // to consider shift amounts with defined behavior.
3639 static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned OpSize) {
3640 // If OpSize is a power of 2 then:
3642 // (a) (Pos == 0 ? 0 : OpSize - Pos) == (OpSize - Pos) & (OpSize - 1)
3643 // (b) Neg == Neg & (OpSize - 1) whenever Neg is in [0, OpSize).
3645 // So if OpSize is a power of 2 and Neg is (and Neg', OpSize-1), we check
3646 // for the stronger condition:
3648 // Neg & (OpSize - 1) == (OpSize - Pos) & (OpSize - 1) [A]
3650 // for all Neg and Pos. Since Neg & (OpSize - 1) == Neg' & (OpSize - 1)
3651 // we can just replace Neg with Neg' for the rest of the function.
3653 // In other cases we check for the even stronger condition:
3655 // Neg == OpSize - Pos [B]
3657 // for all Neg and Pos. Note that the (or ...) then invokes undefined
3658 // behavior if Pos == 0 (and consequently Neg == OpSize).
3660 // We could actually use [A] whenever OpSize is a power of 2, but the
3661 // only extra cases that it would match are those uninteresting ones
3662 // where Neg and Pos are never in range at the same time. E.g. for
3663 // OpSize == 32, using [A] would allow a Neg of the form (sub 64, Pos)
3664 // as well as (sub 32, Pos), but:
3666 // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos))
3668 // always invokes undefined behavior for 32-bit X.
3670 // Below, Mask == OpSize - 1 when using [A] and is all-ones otherwise.
3671 unsigned MaskLoBits = 0;
3672 if (Neg.getOpcode() == ISD::AND &&
3673 isPowerOf2_64(OpSize) &&
3674 Neg.getOperand(1).getOpcode() == ISD::Constant &&
3675 cast<ConstantSDNode>(Neg.getOperand(1))->getAPIntValue() == OpSize - 1) {
3676 Neg = Neg.getOperand(0);
3677 MaskLoBits = Log2_64(OpSize);
3680 // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1.
3681 if (Neg.getOpcode() != ISD::SUB)
3683 ConstantSDNode *NegC = dyn_cast<ConstantSDNode>(Neg.getOperand(0));
3686 SDValue NegOp1 = Neg.getOperand(1);
3688 // On the RHS of [A], if Pos is Pos' & (OpSize - 1), just replace Pos with
3689 // Pos'. The truncation is redundant for the purpose of the equality.
3691 Pos.getOpcode() == ISD::AND &&
3692 Pos.getOperand(1).getOpcode() == ISD::Constant &&
3693 cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() == OpSize - 1)
3694 Pos = Pos.getOperand(0);
3696 // The condition we need is now:
3698 // (NegC - NegOp1) & Mask == (OpSize - Pos) & Mask
3700 // If NegOp1 == Pos then we need:
3702 // OpSize & Mask == NegC & Mask
3704 // (because "x & Mask" is a truncation and distributes through subtraction).
3707 Width = NegC->getAPIntValue();
3708 // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC.
3709 // Then the condition we want to prove becomes:
3711 // (NegC - NegOp1) & Mask == (OpSize - (NegOp1 + PosC)) & Mask
3713 // which, again because "x & Mask" is a truncation, becomes:
3715 // NegC & Mask == (OpSize - PosC) & Mask
3716 // OpSize & Mask == (NegC + PosC) & Mask
3717 else if (Pos.getOpcode() == ISD::ADD &&
3718 Pos.getOperand(0) == NegOp1 &&
3719 Pos.getOperand(1).getOpcode() == ISD::Constant)
3720 Width = (cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() +
3721 NegC->getAPIntValue());
3725 // Now we just need to check that OpSize & Mask == Width & Mask.
3727 // Opsize & Mask is 0 since Mask is Opsize - 1.
3728 return Width.getLoBits(MaskLoBits) == 0;
3729 return Width == OpSize;
3732 // A subroutine of MatchRotate used once we have found an OR of two opposite
3733 // shifts of Shifted. If Neg == <operand size> - Pos then the OR reduces
3734 // to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the
3735 // former being preferred if supported. InnerPos and InnerNeg are Pos and
3736 // Neg with outer conversions stripped away.
3737 SDNode *DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos,
3738 SDValue Neg, SDValue InnerPos,
3739 SDValue InnerNeg, unsigned PosOpcode,
3740 unsigned NegOpcode, SDLoc DL) {
3741 // fold (or (shl x, (*ext y)),
3742 // (srl x, (*ext (sub 32, y)))) ->
3743 // (rotl x, y) or (rotr x, (sub 32, y))
3745 // fold (or (shl x, (*ext (sub 32, y))),
3746 // (srl x, (*ext y))) ->
3747 // (rotr x, y) or (rotl x, (sub 32, y))
3748 EVT VT = Shifted.getValueType();
3749 if (matchRotateSub(InnerPos, InnerNeg, VT.getSizeInBits())) {
3750 bool HasPos = TLI.isOperationLegalOrCustom(PosOpcode, VT);
3751 return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted,
3752 HasPos ? Pos : Neg).getNode();
3758 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3759 // idioms for rotate, and if the target supports rotation instructions, generate
3761 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL) {
3762 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3763 EVT VT = LHS.getValueType();
3764 if (!TLI.isTypeLegal(VT)) return nullptr;
3766 // The target must have at least one rotate flavor.
3767 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3768 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3769 if (!HasROTL && !HasROTR) return nullptr;
3771 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3772 SDValue LHSShift; // The shift.
3773 SDValue LHSMask; // AND value if any.
3774 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3775 return nullptr; // Not part of a rotate.
3777 SDValue RHSShift; // The shift.
3778 SDValue RHSMask; // AND value if any.
3779 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3780 return nullptr; // Not part of a rotate.
3782 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3783 return nullptr; // Not shifting the same value.
3785 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3786 return nullptr; // Shifts must disagree.
3788 // Canonicalize shl to left side in a shl/srl pair.
3789 if (RHSShift.getOpcode() == ISD::SHL) {
3790 std::swap(LHS, RHS);
3791 std::swap(LHSShift, RHSShift);
3792 std::swap(LHSMask , RHSMask );
3795 unsigned OpSizeInBits = VT.getSizeInBits();
3796 SDValue LHSShiftArg = LHSShift.getOperand(0);
3797 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3798 SDValue RHSShiftArg = RHSShift.getOperand(0);
3799 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3801 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3802 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3803 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3804 RHSShiftAmt.getOpcode() == ISD::Constant) {
3805 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3806 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3807 if ((LShVal + RShVal) != OpSizeInBits)
3810 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3811 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
3813 // If there is an AND of either shifted operand, apply it to the result.
3814 if (LHSMask.getNode() || RHSMask.getNode()) {
3815 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3817 if (LHSMask.getNode()) {
3818 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3819 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3821 if (RHSMask.getNode()) {
3822 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3823 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3826 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3829 return Rot.getNode();
3832 // If there is a mask here, and we have a variable shift, we can't be sure
3833 // that we're masking out the right stuff.
3834 if (LHSMask.getNode() || RHSMask.getNode())
3837 // If the shift amount is sign/zext/any-extended just peel it off.
3838 SDValue LExtOp0 = LHSShiftAmt;
3839 SDValue RExtOp0 = RHSShiftAmt;
3840 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3841 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3842 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3843 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3844 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3845 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3846 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3847 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3848 LExtOp0 = LHSShiftAmt.getOperand(0);
3849 RExtOp0 = RHSShiftAmt.getOperand(0);
3852 SDNode *TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt,
3853 LExtOp0, RExtOp0, ISD::ROTL, ISD::ROTR, DL);
3857 SDNode *TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt,
3858 RExtOp0, LExtOp0, ISD::ROTR, ISD::ROTL, DL);
3865 SDValue DAGCombiner::visitXOR(SDNode *N) {
3866 SDValue N0 = N->getOperand(0);
3867 SDValue N1 = N->getOperand(1);
3868 EVT VT = N0.getValueType();
3871 if (VT.isVector()) {
3872 if (SDValue FoldedVOp = SimplifyVBinOp(N))
3875 // fold (xor x, 0) -> x, vector edition
3876 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3878 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3882 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3883 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3884 return DAG.getConstant(0, VT);
3885 // fold (xor x, undef) -> undef
3886 if (N0.getOpcode() == ISD::UNDEF)
3888 if (N1.getOpcode() == ISD::UNDEF)
3890 // fold (xor c1, c2) -> c1^c2
3891 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3892 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3894 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3895 // canonicalize constant to RHS
3896 if (isConstantIntBuildVectorOrConstantInt(N0) &&
3897 !isConstantIntBuildVectorOrConstantInt(N1))
3898 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
3899 // fold (xor x, 0) -> x
3900 if (N1C && N1C->isNullValue())
3903 if (SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1))
3906 // fold !(x cc y) -> (x !cc y)
3907 SDValue LHS, RHS, CC;
3908 if (TLI.isConstTrueVal(N1.getNode()) && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3909 bool isInt = LHS.getValueType().isInteger();
3910 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3913 if (!LegalOperations ||
3914 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) {
3915 switch (N0.getOpcode()) {
3917 llvm_unreachable("Unhandled SetCC Equivalent!");
3919 return DAG.getSetCC(SDLoc(N), VT, LHS, RHS, NotCC);
3920 case ISD::SELECT_CC:
3921 return DAG.getSelectCC(SDLoc(N), LHS, RHS, N0.getOperand(2),
3922 N0.getOperand(3), NotCC);
3927 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3928 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3929 N0.getNode()->hasOneUse() &&
3930 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3931 SDValue V = N0.getOperand(0);
3932 V = DAG.getNode(ISD::XOR, SDLoc(N0), V.getValueType(), V,
3933 DAG.getConstant(1, V.getValueType()));
3934 AddToWorklist(V.getNode());
3935 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V);
3938 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3939 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3940 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3941 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3942 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3943 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3944 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3945 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3946 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
3947 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3950 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3951 if (N1C && N1C->isAllOnesValue() &&
3952 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3953 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3954 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3955 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3956 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3957 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3958 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
3959 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3962 // fold (xor (and x, y), y) -> (and (not x), y)
3963 if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3964 N0->getOperand(1) == N1) {
3965 SDValue X = N0->getOperand(0);
3966 SDValue NotX = DAG.getNOT(SDLoc(X), X, VT);
3967 AddToWorklist(NotX.getNode());
3968 return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1);
3970 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3971 if (N1C && N0.getOpcode() == ISD::XOR) {
3972 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3973 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3975 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(1),
3976 DAG.getConstant(N1C->getAPIntValue() ^
3977 N00C->getAPIntValue(), VT));
3979 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(0),
3980 DAG.getConstant(N1C->getAPIntValue() ^
3981 N01C->getAPIntValue(), VT));
3983 // fold (xor x, x) -> 0
3985 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
3987 // fold (xor (shl 1, x), -1) -> (rotl ~1, x)
3988 // Here is a concrete example of this equivalence:
3990 // i16 shl == 1 << 14 == 16384 == 0b0100000000000000
3991 // i16 xor == ~(1 << 14) == 49151 == 0b1011111111111111
3995 // i16 ~1 == 0b1111111111111110
3996 // i16 rol(~1, 14) == 0b1011111111111111
3998 // Some additional tips to help conceptualize this transform:
3999 // - Try to see the operation as placing a single zero in a value of all ones.
4000 // - There exists no value for x which would allow the result to contain zero.
4001 // - Values of x larger than the bitwidth are undefined and do not require a
4002 // consistent result.
4003 // - Pushing the zero left requires shifting one bits in from the right.
4004 // A rotate left of ~1 is a nice way of achieving the desired result.
4005 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
4006 if (auto *N1C = dyn_cast<ConstantSDNode>(N1.getNode()))
4007 if (N0.getOpcode() == ISD::SHL)
4008 if (auto *ShlLHS = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
4009 if (N1C->isAllOnesValue() && ShlLHS->isOne())
4010 return DAG.getNode(ISD::ROTL, SDLoc(N), VT, DAG.getConstant(~1, VT),
4013 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
4014 if (N0.getOpcode() == N1.getOpcode()) {
4015 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
4016 if (Tmp.getNode()) return Tmp;
4019 // Simplify the expression using non-local knowledge.
4020 if (!VT.isVector() &&
4021 SimplifyDemandedBits(SDValue(N, 0)))
4022 return SDValue(N, 0);
4027 /// Handle transforms common to the three shifts, when the shift amount is a
4029 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, ConstantSDNode *Amt) {
4030 // We can't and shouldn't fold opaque constants.
4031 if (Amt->isOpaque())
4034 SDNode *LHS = N->getOperand(0).getNode();
4035 if (!LHS->hasOneUse()) return SDValue();
4037 // We want to pull some binops through shifts, so that we have (and (shift))
4038 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
4039 // thing happens with address calculations, so it's important to canonicalize
4041 bool HighBitSet = false; // Can we transform this if the high bit is set?
4043 switch (LHS->getOpcode()) {
4044 default: return SDValue();
4047 HighBitSet = false; // We can only transform sra if the high bit is clear.
4050 HighBitSet = true; // We can only transform sra if the high bit is set.
4053 if (N->getOpcode() != ISD::SHL)
4054 return SDValue(); // only shl(add) not sr[al](add).
4055 HighBitSet = false; // We can only transform sra if the high bit is clear.
4059 // We require the RHS of the binop to be a constant and not opaque as well.
4060 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
4061 if (!BinOpCst || BinOpCst->isOpaque()) return SDValue();
4063 // FIXME: disable this unless the input to the binop is a shift by a constant.
4064 // If it is not a shift, it pessimizes some common cases like:
4066 // void foo(int *X, int i) { X[i & 1235] = 1; }
4067 // int bar(int *X, int i) { return X[i & 255]; }
4068 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
4069 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
4070 BinOpLHSVal->getOpcode() != ISD::SRA &&
4071 BinOpLHSVal->getOpcode() != ISD::SRL) ||
4072 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
4075 EVT VT = N->getValueType(0);
4077 // If this is a signed shift right, and the high bit is modified by the
4078 // logical operation, do not perform the transformation. The highBitSet
4079 // boolean indicates the value of the high bit of the constant which would
4080 // cause it to be modified for this operation.
4081 if (N->getOpcode() == ISD::SRA) {
4082 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
4083 if (BinOpRHSSignSet != HighBitSet)
4087 if (!TLI.isDesirableToCommuteWithShift(LHS))
4090 // Fold the constants, shifting the binop RHS by the shift amount.
4091 SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)),
4093 LHS->getOperand(1), N->getOperand(1));
4094 assert(isa<ConstantSDNode>(NewRHS) && "Folding was not successful!");
4096 // Create the new shift.
4097 SDValue NewShift = DAG.getNode(N->getOpcode(),
4098 SDLoc(LHS->getOperand(0)),
4099 VT, LHS->getOperand(0), N->getOperand(1));
4101 // Create the new binop.
4102 return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS);
4105 SDValue DAGCombiner::distributeTruncateThroughAnd(SDNode *N) {
4106 assert(N->getOpcode() == ISD::TRUNCATE);
4107 assert(N->getOperand(0).getOpcode() == ISD::AND);
4109 // (truncate:TruncVT (and N00, N01C)) -> (and (truncate:TruncVT N00), TruncC)
4110 if (N->hasOneUse() && N->getOperand(0).hasOneUse()) {
4111 SDValue N01 = N->getOperand(0).getOperand(1);
4113 if (ConstantSDNode *N01C = isConstOrConstSplat(N01)) {
4114 EVT TruncVT = N->getValueType(0);
4115 SDValue N00 = N->getOperand(0).getOperand(0);
4116 APInt TruncC = N01C->getAPIntValue();
4117 TruncC = TruncC.trunc(TruncVT.getScalarSizeInBits());
4119 return DAG.getNode(ISD::AND, SDLoc(N), TruncVT,
4120 DAG.getNode(ISD::TRUNCATE, SDLoc(N), TruncVT, N00),
4121 DAG.getConstant(TruncC, TruncVT));
4128 SDValue DAGCombiner::visitRotate(SDNode *N) {
4129 // fold (rot* x, (trunc (and y, c))) -> (rot* x, (and (trunc y), (trunc c))).
4130 if (N->getOperand(1).getOpcode() == ISD::TRUNCATE &&
4131 N->getOperand(1).getOperand(0).getOpcode() == ISD::AND) {
4132 SDValue NewOp1 = distributeTruncateThroughAnd(N->getOperand(1).getNode());
4133 if (NewOp1.getNode())
4134 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
4135 N->getOperand(0), NewOp1);
4140 SDValue DAGCombiner::visitSHL(SDNode *N) {
4141 SDValue N0 = N->getOperand(0);
4142 SDValue N1 = N->getOperand(1);
4143 EVT VT = N0.getValueType();
4144 unsigned OpSizeInBits = VT.getScalarSizeInBits();
4147 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4148 if (VT.isVector()) {
4149 if (SDValue FoldedVOp = SimplifyVBinOp(N))
4152 BuildVectorSDNode *N1CV = dyn_cast<BuildVectorSDNode>(N1);
4153 // If setcc produces all-one true value then:
4154 // (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<<N1CV)
4155 if (N1CV && N1CV->isConstant()) {
4156 if (N0.getOpcode() == ISD::AND) {
4157 SDValue N00 = N0->getOperand(0);
4158 SDValue N01 = N0->getOperand(1);
4159 BuildVectorSDNode *N01CV = dyn_cast<BuildVectorSDNode>(N01);
4161 if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC &&
4162 TLI.getBooleanContents(N00.getOperand(0).getValueType()) ==
4163 TargetLowering::ZeroOrNegativeOneBooleanContent) {
4164 if (SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, VT, N01CV, N1CV))
4165 return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C);
4168 N1C = isConstOrConstSplat(N1);
4173 // fold (shl c1, c2) -> c1<<c2
4174 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4176 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
4177 // fold (shl 0, x) -> 0
4178 if (N0C && N0C->isNullValue())
4180 // fold (shl x, c >= size(x)) -> undef
4181 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4182 return DAG.getUNDEF(VT);
4183 // fold (shl x, 0) -> x
4184 if (N1C && N1C->isNullValue())
4186 // fold (shl undef, x) -> 0
4187 if (N0.getOpcode() == ISD::UNDEF)
4188 return DAG.getConstant(0, VT);
4189 // if (shl x, c) is known to be zero, return 0
4190 if (DAG.MaskedValueIsZero(SDValue(N, 0),
4191 APInt::getAllOnesValue(OpSizeInBits)))
4192 return DAG.getConstant(0, VT);
4193 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
4194 if (N1.getOpcode() == ISD::TRUNCATE &&
4195 N1.getOperand(0).getOpcode() == ISD::AND) {
4196 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4197 if (NewOp1.getNode())
4198 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, NewOp1);
4201 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4202 return SDValue(N, 0);
4204 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
4205 if (N1C && N0.getOpcode() == ISD::SHL) {
4206 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4207 uint64_t c1 = N0C1->getZExtValue();
4208 uint64_t c2 = N1C->getZExtValue();
4209 if (c1 + c2 >= OpSizeInBits)
4210 return DAG.getConstant(0, VT);
4211 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4212 DAG.getConstant(c1 + c2, N1.getValueType()));
4216 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
4217 // For this to be valid, the second form must not preserve any of the bits
4218 // that are shifted out by the inner shift in the first form. This means
4219 // the outer shift size must be >= the number of bits added by the ext.
4220 // As a corollary, we don't care what kind of ext it is.
4221 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
4222 N0.getOpcode() == ISD::ANY_EXTEND ||
4223 N0.getOpcode() == ISD::SIGN_EXTEND) &&
4224 N0.getOperand(0).getOpcode() == ISD::SHL) {
4225 SDValue N0Op0 = N0.getOperand(0);
4226 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4227 uint64_t c1 = N0Op0C1->getZExtValue();
4228 uint64_t c2 = N1C->getZExtValue();
4229 EVT InnerShiftVT = N0Op0.getValueType();
4230 uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits();
4231 if (c2 >= OpSizeInBits - InnerShiftSize) {
4232 if (c1 + c2 >= OpSizeInBits)
4233 return DAG.getConstant(0, VT);
4234 return DAG.getNode(ISD::SHL, SDLoc(N0), VT,
4235 DAG.getNode(N0.getOpcode(), SDLoc(N0), VT,
4236 N0Op0->getOperand(0)),
4237 DAG.getConstant(c1 + c2, N1.getValueType()));
4242 // fold (shl (zext (srl x, C)), C) -> (zext (shl (srl x, C), C))
4243 // Only fold this if the inner zext has no other uses to avoid increasing
4244 // the total number of instructions.
4245 if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() &&
4246 N0.getOperand(0).getOpcode() == ISD::SRL) {
4247 SDValue N0Op0 = N0.getOperand(0);
4248 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4249 uint64_t c1 = N0Op0C1->getZExtValue();
4250 if (c1 < VT.getScalarSizeInBits()) {
4251 uint64_t c2 = N1C->getZExtValue();
4253 SDValue NewOp0 = N0.getOperand(0);
4254 EVT CountVT = NewOp0.getOperand(1).getValueType();
4255 SDValue NewSHL = DAG.getNode(ISD::SHL, SDLoc(N), NewOp0.getValueType(),
4256 NewOp0, DAG.getConstant(c2, CountVT));
4257 AddToWorklist(NewSHL.getNode());
4258 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
4264 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
4265 // (and (srl x, (sub c1, c2), MASK)
4266 // Only fold this if the inner shift has no other uses -- if it does, folding
4267 // this will increase the total number of instructions.
4268 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
4269 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4270 uint64_t c1 = N0C1->getZExtValue();
4271 if (c1 < OpSizeInBits) {
4272 uint64_t c2 = N1C->getZExtValue();
4273 APInt Mask = APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - c1);
4276 Mask = Mask.shl(c2 - c1);
4277 Shift = DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4278 DAG.getConstant(c2 - c1, N1.getValueType()));
4280 Mask = Mask.lshr(c1 - c2);
4281 Shift = DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4282 DAG.getConstant(c1 - c2, N1.getValueType()));
4284 return DAG.getNode(ISD::AND, SDLoc(N0), VT, Shift,
4285 DAG.getConstant(Mask, VT));
4289 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
4290 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
4291 unsigned BitSize = VT.getScalarSizeInBits();
4292 SDValue HiBitsMask =
4293 DAG.getConstant(APInt::getHighBitsSet(BitSize,
4294 BitSize - N1C->getZExtValue()), VT);
4295 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4299 // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
4300 // Variant of version done on multiply, except mul by a power of 2 is turned
4303 if (N1C && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
4304 (isa<ConstantSDNode>(N0.getOperand(1)) ||
4305 isConstantSplatVector(N0.getOperand(1).getNode(), Val))) {
4306 SDValue Shl0 = DAG.getNode(ISD::SHL, SDLoc(N0), VT, N0.getOperand(0), N1);
4307 SDValue Shl1 = DAG.getNode(ISD::SHL, SDLoc(N1), VT, N0.getOperand(1), N1);
4308 return DAG.getNode(ISD::ADD, SDLoc(N), VT, Shl0, Shl1);
4312 SDValue NewSHL = visitShiftByConstant(N, N1C);
4313 if (NewSHL.getNode())
4320 SDValue DAGCombiner::visitSRA(SDNode *N) {
4321 SDValue N0 = N->getOperand(0);
4322 SDValue N1 = N->getOperand(1);
4323 EVT VT = N0.getValueType();
4324 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4327 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4328 if (VT.isVector()) {
4329 if (SDValue FoldedVOp = SimplifyVBinOp(N))
4332 N1C = isConstOrConstSplat(N1);
4335 // fold (sra c1, c2) -> (sra c1, c2)
4336 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4338 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
4339 // fold (sra 0, x) -> 0
4340 if (N0C && N0C->isNullValue())
4342 // fold (sra -1, x) -> -1
4343 if (N0C && N0C->isAllOnesValue())
4345 // fold (sra x, (setge c, size(x))) -> undef
4346 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4347 return DAG.getUNDEF(VT);
4348 // fold (sra x, 0) -> x
4349 if (N1C && N1C->isNullValue())
4351 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
4353 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
4354 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
4355 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
4357 ExtVT = EVT::getVectorVT(*DAG.getContext(),
4358 ExtVT, VT.getVectorNumElements());
4359 if ((!LegalOperations ||
4360 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
4361 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
4362 N0.getOperand(0), DAG.getValueType(ExtVT));
4365 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
4366 if (N1C && N0.getOpcode() == ISD::SRA) {
4367 if (ConstantSDNode *C1 = isConstOrConstSplat(N0.getOperand(1))) {
4368 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
4369 if (Sum >= OpSizeInBits)
4370 Sum = OpSizeInBits - 1;
4371 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0.getOperand(0),
4372 DAG.getConstant(Sum, N1.getValueType()));
4376 // fold (sra (shl X, m), (sub result_size, n))
4377 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
4378 // result_size - n != m.
4379 // If truncate is free for the target sext(shl) is likely to result in better
4381 if (N0.getOpcode() == ISD::SHL && N1C) {
4382 // Get the two constanst of the shifts, CN0 = m, CN = n.
4383 const ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1));
4385 LLVMContext &Ctx = *DAG.getContext();
4386 // Determine what the truncate's result bitsize and type would be.
4387 EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - N1C->getZExtValue());
4390 TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorNumElements());
4392 // Determine the residual right-shift amount.
4393 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
4395 // If the shift is not a no-op (in which case this should be just a sign
4396 // extend already), the truncated to type is legal, sign_extend is legal
4397 // on that type, and the truncate to that type is both legal and free,
4398 // perform the transform.
4399 if ((ShiftAmt > 0) &&
4400 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
4401 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
4402 TLI.isTruncateFree(VT, TruncVT)) {
4404 SDValue Amt = DAG.getConstant(ShiftAmt,
4405 getShiftAmountTy(N0.getOperand(0).getValueType()));
4406 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0), VT,
4407 N0.getOperand(0), Amt);
4408 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), TruncVT,
4410 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N),
4411 N->getValueType(0), Trunc);
4416 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
4417 if (N1.getOpcode() == ISD::TRUNCATE &&
4418 N1.getOperand(0).getOpcode() == ISD::AND) {
4419 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4420 if (NewOp1.getNode())
4421 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, NewOp1);
4424 // fold (sra (trunc (srl x, c1)), c2) -> (trunc (sra x, c1 + c2))
4425 // if c1 is equal to the number of bits the trunc removes
4426 if (N0.getOpcode() == ISD::TRUNCATE &&
4427 (N0.getOperand(0).getOpcode() == ISD::SRL ||
4428 N0.getOperand(0).getOpcode() == ISD::SRA) &&
4429 N0.getOperand(0).hasOneUse() &&
4430 N0.getOperand(0).getOperand(1).hasOneUse() &&
4432 SDValue N0Op0 = N0.getOperand(0);
4433 if (ConstantSDNode *LargeShift = isConstOrConstSplat(N0Op0.getOperand(1))) {
4434 unsigned LargeShiftVal = LargeShift->getZExtValue();
4435 EVT LargeVT = N0Op0.getValueType();
4437 if (LargeVT.getScalarSizeInBits() - OpSizeInBits == LargeShiftVal) {
4439 DAG.getConstant(LargeShiftVal + N1C->getZExtValue(),
4440 getShiftAmountTy(N0Op0.getOperand(0).getValueType()));
4441 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), LargeVT,
4442 N0Op0.getOperand(0), Amt);
4443 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, SRA);
4448 // Simplify, based on bits shifted out of the LHS.
4449 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4450 return SDValue(N, 0);
4453 // If the sign bit is known to be zero, switch this to a SRL.
4454 if (DAG.SignBitIsZero(N0))
4455 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1);
4458 SDValue NewSRA = visitShiftByConstant(N, N1C);
4459 if (NewSRA.getNode())
4466 SDValue DAGCombiner::visitSRL(SDNode *N) {
4467 SDValue N0 = N->getOperand(0);
4468 SDValue N1 = N->getOperand(1);
4469 EVT VT = N0.getValueType();
4470 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4473 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4474 if (VT.isVector()) {
4475 if (SDValue FoldedVOp = SimplifyVBinOp(N))
4478 N1C = isConstOrConstSplat(N1);
4481 // fold (srl c1, c2) -> c1 >>u c2
4482 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4484 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
4485 // fold (srl 0, x) -> 0
4486 if (N0C && N0C->isNullValue())
4488 // fold (srl x, c >= size(x)) -> undef
4489 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4490 return DAG.getUNDEF(VT);
4491 // fold (srl x, 0) -> x
4492 if (N1C && N1C->isNullValue())
4494 // if (srl x, c) is known to be zero, return 0
4495 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
4496 APInt::getAllOnesValue(OpSizeInBits)))
4497 return DAG.getConstant(0, VT);
4499 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
4500 if (N1C && N0.getOpcode() == ISD::SRL) {
4501 if (ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1))) {
4502 uint64_t c1 = N01C->getZExtValue();
4503 uint64_t c2 = N1C->getZExtValue();
4504 if (c1 + c2 >= OpSizeInBits)
4505 return DAG.getConstant(0, VT);
4506 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4507 DAG.getConstant(c1 + c2, N1.getValueType()));
4511 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
4512 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
4513 N0.getOperand(0).getOpcode() == ISD::SRL &&
4514 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
4516 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
4517 uint64_t c2 = N1C->getZExtValue();
4518 EVT InnerShiftVT = N0.getOperand(0).getValueType();
4519 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
4520 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
4521 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
4522 if (c1 + OpSizeInBits == InnerShiftSize) {
4523 if (c1 + c2 >= InnerShiftSize)
4524 return DAG.getConstant(0, VT);
4525 return DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT,
4526 DAG.getNode(ISD::SRL, SDLoc(N0), InnerShiftVT,
4527 N0.getOperand(0)->getOperand(0),
4528 DAG.getConstant(c1 + c2, ShiftCountVT)));
4532 // fold (srl (shl x, c), c) -> (and x, cst2)
4533 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1) {
4534 unsigned BitSize = N0.getScalarValueSizeInBits();
4535 if (BitSize <= 64) {
4536 uint64_t ShAmt = N1C->getZExtValue() + 64 - BitSize;
4537 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4538 DAG.getConstant(~0ULL >> ShAmt, VT));
4542 // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask)
4543 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
4544 // Shifting in all undef bits?
4545 EVT SmallVT = N0.getOperand(0).getValueType();
4546 unsigned BitSize = SmallVT.getScalarSizeInBits();
4547 if (N1C->getZExtValue() >= BitSize)
4548 return DAG.getUNDEF(VT);
4550 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
4551 uint64_t ShiftAmt = N1C->getZExtValue();
4552 SDValue SmallShift = DAG.getNode(ISD::SRL, SDLoc(N0), SmallVT,
4554 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
4555 AddToWorklist(SmallShift.getNode());
4556 APInt Mask = APInt::getAllOnesValue(OpSizeInBits).lshr(ShiftAmt);
4557 return DAG.getNode(ISD::AND, SDLoc(N), VT,
4558 DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, SmallShift),
4559 DAG.getConstant(Mask, VT));
4563 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
4564 // bit, which is unmodified by sra.
4565 if (N1C && N1C->getZExtValue() + 1 == OpSizeInBits) {
4566 if (N0.getOpcode() == ISD::SRA)
4567 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1);
4570 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
4571 if (N1C && N0.getOpcode() == ISD::CTLZ &&
4572 N1C->getAPIntValue() == Log2_32(OpSizeInBits)) {
4573 APInt KnownZero, KnownOne;
4574 DAG.computeKnownBits(N0.getOperand(0), KnownZero, KnownOne);
4576 // If any of the input bits are KnownOne, then the input couldn't be all
4577 // zeros, thus the result of the srl will always be zero.
4578 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
4580 // If all of the bits input the to ctlz node are known to be zero, then
4581 // the result of the ctlz is "32" and the result of the shift is one.
4582 APInt UnknownBits = ~KnownZero;
4583 if (UnknownBits == 0) return DAG.getConstant(1, VT);
4585 // Otherwise, check to see if there is exactly one bit input to the ctlz.
4586 if ((UnknownBits & (UnknownBits - 1)) == 0) {
4587 // Okay, we know that only that the single bit specified by UnknownBits
4588 // could be set on input to the CTLZ node. If this bit is set, the SRL
4589 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
4590 // to an SRL/XOR pair, which is likely to simplify more.
4591 unsigned ShAmt = UnknownBits.countTrailingZeros();
4592 SDValue Op = N0.getOperand(0);
4595 Op = DAG.getNode(ISD::SRL, SDLoc(N0), VT, Op,
4596 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
4597 AddToWorklist(Op.getNode());
4600 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
4601 Op, DAG.getConstant(1, VT));
4605 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
4606 if (N1.getOpcode() == ISD::TRUNCATE &&
4607 N1.getOperand(0).getOpcode() == ISD::AND) {
4608 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4609 if (NewOp1.getNode())
4610 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, NewOp1);
4613 // fold operands of srl based on knowledge that the low bits are not
4615 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4616 return SDValue(N, 0);
4619 SDValue NewSRL = visitShiftByConstant(N, N1C);
4620 if (NewSRL.getNode())
4624 // Attempt to convert a srl of a load into a narrower zero-extending load.
4625 SDValue NarrowLoad = ReduceLoadWidth(N);
4626 if (NarrowLoad.getNode())
4629 // Here is a common situation. We want to optimize:
4632 // %b = and i32 %a, 2
4633 // %c = srl i32 %b, 1
4634 // brcond i32 %c ...
4640 // %c = setcc eq %b, 0
4643 // However when after the source operand of SRL is optimized into AND, the SRL
4644 // itself may not be optimized further. Look for it and add the BRCOND into
4646 if (N->hasOneUse()) {
4647 SDNode *Use = *N->use_begin();
4648 if (Use->getOpcode() == ISD::BRCOND)
4650 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
4651 // Also look pass the truncate.
4652 Use = *Use->use_begin();
4653 if (Use->getOpcode() == ISD::BRCOND)
4661 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
4662 SDValue N0 = N->getOperand(0);
4663 EVT VT = N->getValueType(0);
4665 // fold (ctlz c1) -> c2
4666 if (isa<ConstantSDNode>(N0))
4667 return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0);
4671 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4672 SDValue N0 = N->getOperand(0);
4673 EVT VT = N->getValueType(0);
4675 // fold (ctlz_zero_undef c1) -> c2
4676 if (isa<ConstantSDNode>(N0))
4677 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4681 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4682 SDValue N0 = N->getOperand(0);
4683 EVT VT = N->getValueType(0);
4685 // fold (cttz c1) -> c2
4686 if (isa<ConstantSDNode>(N0))
4687 return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0);
4691 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4692 SDValue N0 = N->getOperand(0);
4693 EVT VT = N->getValueType(0);
4695 // fold (cttz_zero_undef c1) -> c2
4696 if (isa<ConstantSDNode>(N0))
4697 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4701 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4702 SDValue N0 = N->getOperand(0);
4703 EVT VT = N->getValueType(0);
4705 // fold (ctpop c1) -> c2
4706 if (isa<ConstantSDNode>(N0))
4707 return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0);
4712 /// \brief Generate Min/Max node
4713 static SDValue combineMinNumMaxNum(SDLoc DL, EVT VT, SDValue LHS, SDValue RHS,
4714 SDValue True, SDValue False,
4715 ISD::CondCode CC, const TargetLowering &TLI,
4716 SelectionDAG &DAG) {
4717 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
4727 unsigned Opcode = (LHS == True) ? ISD::FMINNUM : ISD::FMAXNUM;
4728 if (TLI.isOperationLegal(Opcode, VT))
4729 return DAG.getNode(Opcode, DL, VT, LHS, RHS);
4738 unsigned Opcode = (LHS == True) ? ISD::FMAXNUM : ISD::FMINNUM;
4739 if (TLI.isOperationLegal(Opcode, VT))
4740 return DAG.getNode(Opcode, DL, VT, LHS, RHS);
4748 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4749 SDValue N0 = N->getOperand(0);
4750 SDValue N1 = N->getOperand(1);
4751 SDValue N2 = N->getOperand(2);
4752 EVT VT = N->getValueType(0);
4753 EVT VT0 = N0.getValueType();
4755 // fold (select C, X, X) -> X
4758 // fold (select true, X, Y) -> X
4759 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4760 if (N0C && !N0C->isNullValue())
4762 // fold (select false, X, Y) -> Y
4763 if (N0C && N0C->isNullValue())
4765 // fold (select C, 1, X) -> (or C, X)
4766 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4767 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4768 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4769 // fold (select C, 0, 1) -> (xor C, 1)
4770 // We can't do this reliably if integer based booleans have different contents
4771 // to floating point based booleans. This is because we can't tell whether we
4772 // have an integer-based boolean or a floating-point-based boolean unless we
4773 // can find the SETCC that produced it and inspect its operands. This is
4774 // fairly easy if C is the SETCC node, but it can potentially be
4775 // undiscoverable (or not reasonably discoverable). For example, it could be
4776 // in another basic block or it could require searching a complicated
4778 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
4779 if (VT.isInteger() &&
4780 (VT0 == MVT::i1 || (VT0.isInteger() &&
4781 TLI.getBooleanContents(false, false) ==
4782 TLI.getBooleanContents(false, true) &&
4783 TLI.getBooleanContents(false, false) ==
4784 TargetLowering::ZeroOrOneBooleanContent)) &&
4785 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4788 return DAG.getNode(ISD::XOR, SDLoc(N), VT0,
4789 N0, DAG.getConstant(1, VT0));
4790 XORNode = DAG.getNode(ISD::XOR, SDLoc(N0), VT0,
4791 N0, DAG.getConstant(1, VT0));
4792 AddToWorklist(XORNode.getNode());
4794 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, XORNode);
4795 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, XORNode);
4797 // fold (select C, 0, X) -> (and (not C), X)
4798 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4799 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4800 AddToWorklist(NOTNode.getNode());
4801 return DAG.getNode(ISD::AND, SDLoc(N), VT, NOTNode, N2);
4803 // fold (select C, X, 1) -> (or (not C), X)
4804 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4805 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4806 AddToWorklist(NOTNode.getNode());
4807 return DAG.getNode(ISD::OR, SDLoc(N), VT, NOTNode, N1);
4809 // fold (select C, X, 0) -> (and C, X)
4810 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4811 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4812 // fold (select X, X, Y) -> (or X, Y)
4813 // fold (select X, 1, Y) -> (or X, Y)
4814 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4815 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4816 // fold (select X, Y, X) -> (and X, Y)
4817 // fold (select X, Y, 0) -> (and X, Y)
4818 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4819 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4821 // If we can fold this based on the true/false value, do so.
4822 if (SimplifySelectOps(N, N1, N2))
4823 return SDValue(N, 0); // Don't revisit N.
4825 // fold selects based on a setcc into other things, such as min/max/abs
4826 if (N0.getOpcode() == ISD::SETCC) {
4827 // select x, y (fcmp lt x, y) -> fminnum x, y
4828 // select x, y (fcmp gt x, y) -> fmaxnum x, y
4830 // This is OK if we don't care about what happens if either operand is a
4834 // FIXME: Instead of testing for UnsafeFPMath, this should be checking for
4835 // no signed zeros as well as no nans.
4836 const TargetOptions &Options = DAG.getTarget().Options;
4837 if (Options.UnsafeFPMath &&
4838 VT.isFloatingPoint() && N0.hasOneUse() &&
4839 DAG.isKnownNeverNaN(N1) && DAG.isKnownNeverNaN(N2)) {
4840 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
4843 combineMinNumMaxNum(SDLoc(N), VT, N0.getOperand(0), N0.getOperand(1),
4844 N1, N2, CC, TLI, DAG);
4849 if ((!LegalOperations &&
4850 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT)) ||
4851 TLI.isOperationLegal(ISD::SELECT_CC, VT))
4852 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT,
4853 N0.getOperand(0), N0.getOperand(1),
4854 N1, N2, N0.getOperand(2));
4855 return SimplifySelect(SDLoc(N), N0, N1, N2);
4858 if (VT0 == MVT::i1) {
4859 if (TLI.shouldNormalizeToSelectSequence(*DAG.getContext(), VT)) {
4860 // select (and Cond0, Cond1), X, Y
4861 // -> select Cond0, (select Cond1, X, Y), Y
4862 if (N0->getOpcode() == ISD::AND && N0->hasOneUse()) {
4863 SDValue Cond0 = N0->getOperand(0);
4864 SDValue Cond1 = N0->getOperand(1);
4865 SDValue InnerSelect = DAG.getNode(ISD::SELECT, SDLoc(N),
4866 N1.getValueType(), Cond1, N1, N2);
4867 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Cond0,
4870 // select (or Cond0, Cond1), X, Y -> select Cond0, X, (select Cond1, X, Y)
4871 if (N0->getOpcode() == ISD::OR && N0->hasOneUse()) {
4872 SDValue Cond0 = N0->getOperand(0);
4873 SDValue Cond1 = N0->getOperand(1);
4874 SDValue InnerSelect = DAG.getNode(ISD::SELECT, SDLoc(N),
4875 N1.getValueType(), Cond1, N1, N2);
4876 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Cond0, N1,
4881 // select Cond0, (select Cond1, X, Y), Y -> select (and Cond0, Cond1), X, Y
4882 if (N1->getOpcode() == ISD::SELECT) {
4883 SDValue N1_0 = N1->getOperand(0);
4884 SDValue N1_1 = N1->getOperand(1);
4885 SDValue N1_2 = N1->getOperand(2);
4886 if (N1_2 == N2 && N0.getValueType() == N1_0.getValueType()) {
4887 // Create the actual and node if we can generate good code for it.
4888 if (!TLI.shouldNormalizeToSelectSequence(*DAG.getContext(), VT)) {
4889 SDValue And = DAG.getNode(ISD::AND, SDLoc(N), N0.getValueType(),
4891 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), And,
4894 // Otherwise see if we can optimize the "and" to a better pattern.
4895 if (SDValue Combined = visitANDLike(N0, N1_0, N))
4896 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Combined,
4900 // select Cond0, X, (select Cond1, X, Y) -> select (or Cond0, Cond1), X, Y
4901 if (N2->getOpcode() == ISD::SELECT) {
4902 SDValue N2_0 = N2->getOperand(0);
4903 SDValue N2_1 = N2->getOperand(1);
4904 SDValue N2_2 = N2->getOperand(2);
4905 if (N2_1 == N1 && N0.getValueType() == N2_0.getValueType()) {
4906 // Create the actual or node if we can generate good code for it.
4907 if (!TLI.shouldNormalizeToSelectSequence(*DAG.getContext(), VT)) {
4908 SDValue Or = DAG.getNode(ISD::OR, SDLoc(N), N0.getValueType(),
4910 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Or,
4913 // Otherwise see if we can optimize to a better pattern.
4914 if (SDValue Combined = visitORLike(N0, N2_0, N))
4915 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Combined,
4925 std::pair<SDValue, SDValue> SplitVSETCC(const SDNode *N, SelectionDAG &DAG) {
4928 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
4930 // Split the inputs.
4931 SDValue Lo, Hi, LL, LH, RL, RH;
4932 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
4933 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
4935 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));
4936 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));
4938 return std::make_pair(Lo, Hi);
4941 // This function assumes all the vselect's arguments are CONCAT_VECTOR
4942 // nodes and that the condition is a BV of ConstantSDNodes (or undefs).
4943 static SDValue ConvertSelectToConcatVector(SDNode *N, SelectionDAG &DAG) {
4945 SDValue Cond = N->getOperand(0);
4946 SDValue LHS = N->getOperand(1);
4947 SDValue RHS = N->getOperand(2);
4948 EVT VT = N->getValueType(0);
4949 int NumElems = VT.getVectorNumElements();
4950 assert(LHS.getOpcode() == ISD::CONCAT_VECTORS &&
4951 RHS.getOpcode() == ISD::CONCAT_VECTORS &&
4952 Cond.getOpcode() == ISD::BUILD_VECTOR);
4954 // CONCAT_VECTOR can take an arbitrary number of arguments. We only care about
4955 // binary ones here.
4956 if (LHS->getNumOperands() != 2 || RHS->getNumOperands() != 2)
4959 // We're sure we have an even number of elements due to the
4960 // concat_vectors we have as arguments to vselect.
4961 // Skip BV elements until we find one that's not an UNDEF
4962 // After we find an UNDEF element, keep looping until we get to half the
4963 // length of the BV and see if all the non-undef nodes are the same.
4964 ConstantSDNode *BottomHalf = nullptr;
4965 for (int i = 0; i < NumElems / 2; ++i) {
4966 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4969 if (BottomHalf == nullptr)
4970 BottomHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4971 else if (Cond->getOperand(i).getNode() != BottomHalf)
4975 // Do the same for the second half of the BuildVector
4976 ConstantSDNode *TopHalf = nullptr;
4977 for (int i = NumElems / 2; i < NumElems; ++i) {
4978 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4981 if (TopHalf == nullptr)
4982 TopHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4983 else if (Cond->getOperand(i).getNode() != TopHalf)
4987 assert(TopHalf && BottomHalf &&
4988 "One half of the selector was all UNDEFs and the other was all the "
4989 "same value. This should have been addressed before this function.");
4991 ISD::CONCAT_VECTORS, dl, VT,
4992 BottomHalf->isNullValue() ? RHS->getOperand(0) : LHS->getOperand(0),
4993 TopHalf->isNullValue() ? RHS->getOperand(1) : LHS->getOperand(1));
4996 SDValue DAGCombiner::visitMSTORE(SDNode *N) {
4998 if (Level >= AfterLegalizeTypes)
5001 MaskedStoreSDNode *MST = dyn_cast<MaskedStoreSDNode>(N);
5002 SDValue Mask = MST->getMask();
5003 SDValue Data = MST->getValue();
5006 // If the MSTORE data type requires splitting and the mask is provided by a
5007 // SETCC, then split both nodes and its operands before legalization. This
5008 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5009 // and enables future optimizations (e.g. min/max pattern matching on X86).
5010 if (Mask.getOpcode() == ISD::SETCC) {
5012 // Check if any splitting is required.
5013 if (TLI.getTypeAction(*DAG.getContext(), Data.getValueType()) !=
5014 TargetLowering::TypeSplitVector)
5017 SDValue MaskLo, MaskHi, Lo, Hi;
5018 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
5021 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MST->getValueType(0));
5023 SDValue Chain = MST->getChain();
5024 SDValue Ptr = MST->getBasePtr();
5026 EVT MemoryVT = MST->getMemoryVT();
5027 unsigned Alignment = MST->getOriginalAlignment();
5029 // if Alignment is equal to the vector size,
5030 // take the half of it for the second part
5031 unsigned SecondHalfAlignment =
5032 (Alignment == Data->getValueType(0).getSizeInBits()/8) ?
5033 Alignment/2 : Alignment;
5035 EVT LoMemVT, HiMemVT;
5036 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
5038 SDValue DataLo, DataHi;
5039 std::tie(DataLo, DataHi) = DAG.SplitVector(Data, DL);
5041 MachineMemOperand *MMO = DAG.getMachineFunction().
5042 getMachineMemOperand(MST->getPointerInfo(),
5043 MachineMemOperand::MOStore, LoMemVT.getStoreSize(),
5044 Alignment, MST->getAAInfo(), MST->getRanges());
5046 Lo = DAG.getMaskedStore(Chain, DL, DataLo, Ptr, MaskLo, LoMemVT, MMO,
5047 MST->isTruncatingStore());
5049 unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
5050 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5051 DAG.getConstant(IncrementSize, Ptr.getValueType()));
5053 MMO = DAG.getMachineFunction().
5054 getMachineMemOperand(MST->getPointerInfo(),
5055 MachineMemOperand::MOStore, HiMemVT.getStoreSize(),
5056 SecondHalfAlignment, MST->getAAInfo(),
5059 Hi = DAG.getMaskedStore(Chain, DL, DataHi, Ptr, MaskHi, HiMemVT, MMO,
5060 MST->isTruncatingStore());
5062 AddToWorklist(Lo.getNode());
5063 AddToWorklist(Hi.getNode());
5065 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);
5070 SDValue DAGCombiner::visitMLOAD(SDNode *N) {
5072 if (Level >= AfterLegalizeTypes)
5075 MaskedLoadSDNode *MLD = dyn_cast<MaskedLoadSDNode>(N);
5076 SDValue Mask = MLD->getMask();
5079 // If the MLOAD result requires splitting and the mask is provided by a
5080 // SETCC, then split both nodes and its operands before legalization. This
5081 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5082 // and enables future optimizations (e.g. min/max pattern matching on X86).
5084 if (Mask.getOpcode() == ISD::SETCC) {
5085 EVT VT = N->getValueType(0);
5087 // Check if any splitting is required.
5088 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
5089 TargetLowering::TypeSplitVector)
5092 SDValue MaskLo, MaskHi, Lo, Hi;
5093 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
5095 SDValue Src0 = MLD->getSrc0();
5096 SDValue Src0Lo, Src0Hi;
5097 std::tie(Src0Lo, Src0Hi) = DAG.SplitVector(Src0, DL);
5100 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MLD->getValueType(0));
5102 SDValue Chain = MLD->getChain();
5103 SDValue Ptr = MLD->getBasePtr();
5104 EVT MemoryVT = MLD->getMemoryVT();
5105 unsigned Alignment = MLD->getOriginalAlignment();
5107 // if Alignment is equal to the vector size,
5108 // take the half of it for the second part
5109 unsigned SecondHalfAlignment =
5110 (Alignment == MLD->getValueType(0).getSizeInBits()/8) ?
5111 Alignment/2 : Alignment;
5113 EVT LoMemVT, HiMemVT;
5114 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
5116 MachineMemOperand *MMO = DAG.getMachineFunction().
5117 getMachineMemOperand(MLD->getPointerInfo(),
5118 MachineMemOperand::MOLoad, LoMemVT.getStoreSize(),
5119 Alignment, MLD->getAAInfo(), MLD->getRanges());
5121 Lo = DAG.getMaskedLoad(LoVT, DL, Chain, Ptr, MaskLo, Src0Lo, LoMemVT, MMO,
5124 unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
5125 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5126 DAG.getConstant(IncrementSize, Ptr.getValueType()));
5128 MMO = DAG.getMachineFunction().
5129 getMachineMemOperand(MLD->getPointerInfo(),
5130 MachineMemOperand::MOLoad, HiMemVT.getStoreSize(),
5131 SecondHalfAlignment, MLD->getAAInfo(), MLD->getRanges());
5133 Hi = DAG.getMaskedLoad(HiVT, DL, Chain, Ptr, MaskHi, Src0Hi, HiMemVT, MMO,
5136 AddToWorklist(Lo.getNode());
5137 AddToWorklist(Hi.getNode());
5139 // Build a factor node to remember that this load is independent of the
5141 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo.getValue(1),
5144 // Legalized the chain result - switch anything that used the old chain to
5146 DAG.ReplaceAllUsesOfValueWith(SDValue(MLD, 1), Chain);
5148 SDValue LoadRes = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
5150 SDValue RetOps[] = { LoadRes, Chain };
5151 return DAG.getMergeValues(RetOps, DL);
5156 SDValue DAGCombiner::visitVSELECT(SDNode *N) {
5157 SDValue N0 = N->getOperand(0);
5158 SDValue N1 = N->getOperand(1);
5159 SDValue N2 = N->getOperand(2);
5162 // Canonicalize integer abs.
5163 // vselect (setg[te] X, 0), X, -X ->
5164 // vselect (setgt X, -1), X, -X ->
5165 // vselect (setl[te] X, 0), -X, X ->
5166 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
5167 if (N0.getOpcode() == ISD::SETCC) {
5168 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
5169 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5171 bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
5173 if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
5174 (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) &&
5175 N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1))
5176 isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode());
5177 else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) &&
5178 N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1))
5179 isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode());
5182 EVT VT = LHS.getValueType();
5183 SDValue Shift = DAG.getNode(
5184 ISD::SRA, DL, VT, LHS,
5185 DAG.getConstant(VT.getScalarType().getSizeInBits() - 1, VT));
5186 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift);
5187 AddToWorklist(Shift.getNode());
5188 AddToWorklist(Add.getNode());
5189 return DAG.getNode(ISD::XOR, DL, VT, Add, Shift);
5193 // If the VSELECT result requires splitting and the mask is provided by a
5194 // SETCC, then split both nodes and its operands before legalization. This
5195 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5196 // and enables future optimizations (e.g. min/max pattern matching on X86).
5197 if (N0.getOpcode() == ISD::SETCC) {
5198 EVT VT = N->getValueType(0);
5200 // Check if any splitting is required.
5201 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
5202 TargetLowering::TypeSplitVector)
5205 SDValue Lo, Hi, CCLo, CCHi, LL, LH, RL, RH;
5206 std::tie(CCLo, CCHi) = SplitVSETCC(N0.getNode(), DAG);
5207 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 1);
5208 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 2);
5210 Lo = DAG.getNode(N->getOpcode(), DL, LL.getValueType(), CCLo, LL, RL);
5211 Hi = DAG.getNode(N->getOpcode(), DL, LH.getValueType(), CCHi, LH, RH);
5213 // Add the new VSELECT nodes to the work list in case they need to be split
5215 AddToWorklist(Lo.getNode());
5216 AddToWorklist(Hi.getNode());
5218 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
5221 // Fold (vselect (build_vector all_ones), N1, N2) -> N1
5222 if (ISD::isBuildVectorAllOnes(N0.getNode()))
5224 // Fold (vselect (build_vector all_zeros), N1, N2) -> N2
5225 if (ISD::isBuildVectorAllZeros(N0.getNode()))
5228 // The ConvertSelectToConcatVector function is assuming both the above
5229 // checks for (vselect (build_vector all{ones,zeros) ...) have been made
5231 if (N1.getOpcode() == ISD::CONCAT_VECTORS &&
5232 N2.getOpcode() == ISD::CONCAT_VECTORS &&
5233 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
5234 SDValue CV = ConvertSelectToConcatVector(N, DAG);
5242 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
5243 SDValue N0 = N->getOperand(0);
5244 SDValue N1 = N->getOperand(1);
5245 SDValue N2 = N->getOperand(2);
5246 SDValue N3 = N->getOperand(3);
5247 SDValue N4 = N->getOperand(4);
5248 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
5250 // fold select_cc lhs, rhs, x, x, cc -> x
5254 // Determine if the condition we're dealing with is constant
5255 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
5256 N0, N1, CC, SDLoc(N), false);
5257 if (SCC.getNode()) {
5258 AddToWorklist(SCC.getNode());
5260 if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) {
5261 if (!SCCC->isNullValue())
5262 return N2; // cond always true -> true val
5264 return N3; // cond always false -> false val
5265 } else if (SCC->getOpcode() == ISD::UNDEF) {
5266 // When the condition is UNDEF, just return the first operand. This is
5267 // coherent the DAG creation, no setcc node is created in this case
5269 } else if (SCC.getOpcode() == ISD::SETCC) {
5270 // Fold to a simpler select_cc
5271 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(),
5272 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
5277 // If we can fold this based on the true/false value, do so.
5278 if (SimplifySelectOps(N, N2, N3))
5279 return SDValue(N, 0); // Don't revisit N.
5281 // fold select_cc into other things, such as min/max/abs
5282 return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC);
5285 SDValue DAGCombiner::visitSETCC(SDNode *N) {
5286 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
5287 cast<CondCodeSDNode>(N->getOperand(2))->get(),
5291 // tryToFoldExtendOfConstant - Try to fold a sext/zext/aext
5292 // dag node into a ConstantSDNode or a build_vector of constants.
5293 // This function is called by the DAGCombiner when visiting sext/zext/aext
5294 // dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND).
5295 // Vector extends are not folded if operations are legal; this is to
5296 // avoid introducing illegal build_vector dag nodes.
5297 static SDNode *tryToFoldExtendOfConstant(SDNode *N, const TargetLowering &TLI,
5298 SelectionDAG &DAG, bool LegalTypes,
5299 bool LegalOperations) {
5300 unsigned Opcode = N->getOpcode();
5301 SDValue N0 = N->getOperand(0);
5302 EVT VT = N->getValueType(0);
5304 assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND ||
5305 Opcode == ISD::ANY_EXTEND) && "Expected EXTEND dag node in input!");
5307 // fold (sext c1) -> c1
5308 // fold (zext c1) -> c1
5309 // fold (aext c1) -> c1
5310 if (isa<ConstantSDNode>(N0))
5311 return DAG.getNode(Opcode, SDLoc(N), VT, N0).getNode();
5313 // fold (sext (build_vector AllConstants) -> (build_vector AllConstants)
5314 // fold (zext (build_vector AllConstants) -> (build_vector AllConstants)
5315 // fold (aext (build_vector AllConstants) -> (build_vector AllConstants)
5316 EVT SVT = VT.getScalarType();
5317 if (!(VT.isVector() &&
5318 (!LegalTypes || (!LegalOperations && TLI.isTypeLegal(SVT))) &&
5319 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())))
5322 // We can fold this node into a build_vector.
5323 unsigned VTBits = SVT.getSizeInBits();
5324 unsigned EVTBits = N0->getValueType(0).getScalarType().getSizeInBits();
5325 unsigned ShAmt = VTBits - EVTBits;
5326 SmallVector<SDValue, 8> Elts;
5327 unsigned NumElts = N0->getNumOperands();
5330 for (unsigned i=0; i != NumElts; ++i) {
5331 SDValue Op = N0->getOperand(i);
5332 if (Op->getOpcode() == ISD::UNDEF) {
5333 Elts.push_back(DAG.getUNDEF(SVT));
5337 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
5338 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
5339 if (Opcode == ISD::SIGN_EXTEND)
5340 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
5343 Elts.push_back(DAG.getConstant(C.shl(ShAmt).lshr(ShAmt).getZExtValue(),
5347 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Elts).getNode();
5350 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
5351 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
5352 // transformation. Returns true if extension are possible and the above
5353 // mentioned transformation is profitable.
5354 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
5356 SmallVectorImpl<SDNode *> &ExtendNodes,
5357 const TargetLowering &TLI) {
5358 bool HasCopyToRegUses = false;
5359 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
5360 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
5361 UE = N0.getNode()->use_end();
5366 if (UI.getUse().getResNo() != N0.getResNo())
5368 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
5369 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
5370 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
5371 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
5372 // Sign bits will be lost after a zext.
5375 for (unsigned i = 0; i != 2; ++i) {
5376 SDValue UseOp = User->getOperand(i);
5379 if (!isa<ConstantSDNode>(UseOp))
5384 ExtendNodes.push_back(User);
5387 // If truncates aren't free and there are users we can't
5388 // extend, it isn't worthwhile.
5391 // Remember if this value is live-out.
5392 if (User->getOpcode() == ISD::CopyToReg)
5393 HasCopyToRegUses = true;
5396 if (HasCopyToRegUses) {
5397 bool BothLiveOut = false;
5398 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
5400 SDUse &Use = UI.getUse();
5401 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
5407 // Both unextended and extended values are live out. There had better be
5408 // a good reason for the transformation.
5409 return ExtendNodes.size();
5414 void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
5415 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
5416 ISD::NodeType ExtType) {
5417 // Extend SetCC uses if necessary.
5418 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
5419 SDNode *SetCC = SetCCs[i];
5420 SmallVector<SDValue, 4> Ops;
5422 for (unsigned j = 0; j != 2; ++j) {
5423 SDValue SOp = SetCC->getOperand(j);
5425 Ops.push_back(ExtLoad);
5427 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
5430 Ops.push_back(SetCC->getOperand(2));
5431 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), Ops));
5435 // FIXME: Bring more similar combines here, common to sext/zext (maybe aext?).
5436 SDValue DAGCombiner::CombineExtLoad(SDNode *N) {
5437 SDValue N0 = N->getOperand(0);
5438 EVT DstVT = N->getValueType(0);
5439 EVT SrcVT = N0.getValueType();
5441 assert((N->getOpcode() == ISD::SIGN_EXTEND ||
5442 N->getOpcode() == ISD::ZERO_EXTEND) &&
5443 "Unexpected node type (not an extend)!");
5445 // fold (sext (load x)) to multiple smaller sextloads; same for zext.
5446 // For example, on a target with legal v4i32, but illegal v8i32, turn:
5447 // (v8i32 (sext (v8i16 (load x))))
5449 // (v8i32 (concat_vectors (v4i32 (sextload x)),
5450 // (v4i32 (sextload (x + 16)))))
5451 // Where uses of the original load, i.e.:
5453 // are replaced with:
5455 // (v8i32 (concat_vectors (v4i32 (sextload x)),
5456 // (v4i32 (sextload (x + 16)))))))
5458 // This combine is only applicable to illegal, but splittable, vectors.
5459 // All legal types, and illegal non-vector types, are handled elsewhere.
5460 // This combine is controlled by TargetLowering::isVectorLoadExtDesirable.
5462 if (N0->getOpcode() != ISD::LOAD)
5465 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5467 if (!ISD::isNON_EXTLoad(LN0) || !ISD::isUNINDEXEDLoad(LN0) ||
5468 !N0.hasOneUse() || LN0->isVolatile() || !DstVT.isVector() ||
5469 !DstVT.isPow2VectorType() || !TLI.isVectorLoadExtDesirable(SDValue(N, 0)))
5472 SmallVector<SDNode *, 4> SetCCs;
5473 if (!ExtendUsesToFormExtLoad(N, N0, N->getOpcode(), SetCCs, TLI))
5476 ISD::LoadExtType ExtType =
5477 N->getOpcode() == ISD::SIGN_EXTEND ? ISD::SEXTLOAD : ISD::ZEXTLOAD;
5479 // Try to split the vector types to get down to legal types.
5480 EVT SplitSrcVT = SrcVT;
5481 EVT SplitDstVT = DstVT;
5482 while (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT) &&
5483 SplitSrcVT.getVectorNumElements() > 1) {
5484 SplitDstVT = DAG.GetSplitDestVTs(SplitDstVT).first;
5485 SplitSrcVT = DAG.GetSplitDestVTs(SplitSrcVT).first;
5488 if (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT))
5492 const unsigned NumSplits =
5493 DstVT.getVectorNumElements() / SplitDstVT.getVectorNumElements();
5494 const unsigned Stride = SplitSrcVT.getStoreSize();
5495 SmallVector<SDValue, 4> Loads;
5496 SmallVector<SDValue, 4> Chains;
5498 SDValue BasePtr = LN0->getBasePtr();
5499 for (unsigned Idx = 0; Idx < NumSplits; Idx++) {
5500 const unsigned Offset = Idx * Stride;
5501 const unsigned Align = MinAlign(LN0->getAlignment(), Offset);
5503 SDValue SplitLoad = DAG.getExtLoad(
5504 ExtType, DL, SplitDstVT, LN0->getChain(), BasePtr,
5505 LN0->getPointerInfo().getWithOffset(Offset), SplitSrcVT,
5506 LN0->isVolatile(), LN0->isNonTemporal(), LN0->isInvariant(),
5507 Align, LN0->getAAInfo());
5509 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
5510 DAG.getConstant(Stride, BasePtr.getValueType()));
5512 Loads.push_back(SplitLoad.getValue(0));
5513 Chains.push_back(SplitLoad.getValue(1));
5516 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
5517 SDValue NewValue = DAG.getNode(ISD::CONCAT_VECTORS, DL, DstVT, Loads);
5519 CombineTo(N, NewValue);
5521 // Replace uses of the original load (before extension)
5522 // with a truncate of the concatenated sextloaded vectors.
5524 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(), NewValue);
5525 CombineTo(N0.getNode(), Trunc, NewChain);
5526 ExtendSetCCUses(SetCCs, Trunc, NewValue, DL,
5527 (ISD::NodeType)N->getOpcode());
5528 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5531 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
5532 SDValue N0 = N->getOperand(0);
5533 EVT VT = N->getValueType(0);
5535 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5537 return SDValue(Res, 0);
5539 // fold (sext (sext x)) -> (sext x)
5540 // fold (sext (aext x)) -> (sext x)
5541 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5542 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT,
5545 if (N0.getOpcode() == ISD::TRUNCATE) {
5546 // fold (sext (truncate (load x))) -> (sext (smaller load x))
5547 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
5548 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5549 if (NarrowLoad.getNode()) {
5550 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5551 if (NarrowLoad.getNode() != N0.getNode()) {
5552 CombineTo(N0.getNode(), NarrowLoad);
5553 // CombineTo deleted the truncate, if needed, but not what's under it.
5556 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5559 // See if the value being truncated is already sign extended. If so, just
5560 // eliminate the trunc/sext pair.
5561 SDValue Op = N0.getOperand(0);
5562 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
5563 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
5564 unsigned DestBits = VT.getScalarType().getSizeInBits();
5565 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
5567 if (OpBits == DestBits) {
5568 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
5569 // bits, it is already ready.
5570 if (NumSignBits > DestBits-MidBits)
5572 } else if (OpBits < DestBits) {
5573 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
5574 // bits, just sext from i32.
5575 if (NumSignBits > OpBits-MidBits)
5576 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, Op);
5578 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
5579 // bits, just truncate to i32.
5580 if (NumSignBits > OpBits-MidBits)
5581 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5584 // fold (sext (truncate x)) -> (sextinreg x).
5585 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
5586 N0.getValueType())) {
5587 if (OpBits < DestBits)
5588 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op);
5589 else if (OpBits > DestBits)
5590 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op);
5591 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, Op,
5592 DAG.getValueType(N0.getValueType()));
5596 // fold (sext (load x)) -> (sext (truncate (sextload x)))
5597 // Only generate vector extloads when 1) they're legal, and 2) they are
5598 // deemed desirable by the target.
5599 if (ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5600 ((!LegalOperations && !VT.isVector() &&
5601 !cast<LoadSDNode>(N0)->isVolatile()) ||
5602 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, N0.getValueType()))) {
5603 bool DoXform = true;
5604 SmallVector<SDNode*, 4> SetCCs;
5605 if (!N0.hasOneUse())
5606 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
5608 DoXform &= TLI.isVectorLoadExtDesirable(SDValue(N, 0));
5610 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5611 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5613 LN0->getBasePtr(), N0.getValueType(),
5614 LN0->getMemOperand());
5615 CombineTo(N, ExtLoad);
5616 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5617 N0.getValueType(), ExtLoad);
5618 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5619 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5621 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5625 // fold (sext (load x)) to multiple smaller sextloads.
5626 // Only on illegal but splittable vectors.
5627 if (SDValue ExtLoad = CombineExtLoad(N))
5630 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
5631 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
5632 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5633 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5634 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5635 EVT MemVT = LN0->getMemoryVT();
5636 if ((!LegalOperations && !LN0->isVolatile()) ||
5637 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, MemVT)) {
5638 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5640 LN0->getBasePtr(), MemVT,
5641 LN0->getMemOperand());
5642 CombineTo(N, ExtLoad);
5643 CombineTo(N0.getNode(),
5644 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5645 N0.getValueType(), ExtLoad),
5646 ExtLoad.getValue(1));
5647 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5651 // fold (sext (and/or/xor (load x), cst)) ->
5652 // (and/or/xor (sextload x), (sext cst))
5653 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5654 N0.getOpcode() == ISD::XOR) &&
5655 isa<LoadSDNode>(N0.getOperand(0)) &&
5656 N0.getOperand(1).getOpcode() == ISD::Constant &&
5657 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, N0.getValueType()) &&
5658 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5659 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5660 if (LN0->getExtensionType() != ISD::ZEXTLOAD && LN0->isUnindexed()) {
5661 bool DoXform = true;
5662 SmallVector<SDNode*, 4> SetCCs;
5663 if (!N0.hasOneUse())
5664 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
5667 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN0), VT,
5668 LN0->getChain(), LN0->getBasePtr(),
5670 LN0->getMemOperand());
5671 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5672 Mask = Mask.sext(VT.getSizeInBits());
5673 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5674 ExtLoad, DAG.getConstant(Mask, VT));
5675 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5676 SDLoc(N0.getOperand(0)),
5677 N0.getOperand(0).getValueType(), ExtLoad);
5679 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5680 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5682 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5687 if (N0.getOpcode() == ISD::SETCC) {
5688 EVT N0VT = N0.getOperand(0).getValueType();
5689 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
5690 // Only do this before legalize for now.
5691 if (VT.isVector() && !LegalOperations &&
5692 TLI.getBooleanContents(N0VT) ==
5693 TargetLowering::ZeroOrNegativeOneBooleanContent) {
5694 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
5695 // of the same size as the compared operands. Only optimize sext(setcc())
5696 // if this is the case.
5697 EVT SVT = getSetCCResultType(N0VT);
5699 // We know that the # elements of the results is the same as the
5700 // # elements of the compare (and the # elements of the compare result
5701 // for that matter). Check to see that they are the same size. If so,
5702 // we know that the element size of the sext'd result matches the
5703 // element size of the compare operands.
5704 if (VT.getSizeInBits() == SVT.getSizeInBits())
5705 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5707 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5709 // If the desired elements are smaller or larger than the source
5710 // elements we can use a matching integer vector type and then
5711 // truncate/sign extend
5712 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5713 if (SVT == MatchingVectorType) {
5714 SDValue VsetCC = DAG.getSetCC(SDLoc(N), MatchingVectorType,
5715 N0.getOperand(0), N0.getOperand(1),
5716 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5717 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
5721 // sext(setcc x, y, cc) -> (select (setcc x, y, cc), -1, 0)
5722 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
5724 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
5726 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5727 NegOne, DAG.getConstant(0, VT),
5728 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5729 if (SCC.getNode()) return SCC;
5731 if (!VT.isVector()) {
5732 EVT SetCCVT = getSetCCResultType(N0.getOperand(0).getValueType());
5733 if (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, SetCCVT)) {
5735 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5736 SDValue SetCC = DAG.getSetCC(DL, SetCCVT,
5737 N0.getOperand(0), N0.getOperand(1), CC);
5738 return DAG.getSelect(DL, VT, SetCC,
5739 NegOne, DAG.getConstant(0, VT));
5744 // fold (sext x) -> (zext x) if the sign bit is known zero.
5745 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
5746 DAG.SignBitIsZero(N0))
5747 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0);
5752 // isTruncateOf - If N is a truncate of some other value, return true, record
5753 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
5754 // This function computes KnownZero to avoid a duplicated call to
5755 // computeKnownBits in the caller.
5756 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
5759 if (N->getOpcode() == ISD::TRUNCATE) {
5760 Op = N->getOperand(0);
5761 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5765 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
5766 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
5769 SDValue Op0 = N->getOperand(0);
5770 SDValue Op1 = N->getOperand(1);
5771 assert(Op0.getValueType() == Op1.getValueType());
5773 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
5774 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
5775 if (COp0 && COp0->isNullValue())
5777 else if (COp1 && COp1->isNullValue())
5782 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5784 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
5790 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
5791 SDValue N0 = N->getOperand(0);
5792 EVT VT = N->getValueType(0);
5794 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5796 return SDValue(Res, 0);
5798 // fold (zext (zext x)) -> (zext x)
5799 // fold (zext (aext x)) -> (zext x)
5800 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5801 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT,
5804 // fold (zext (truncate x)) -> (zext x) or
5805 // (zext (truncate x)) -> (truncate x)
5806 // This is valid when the truncated bits of x are already zero.
5807 // FIXME: We should extend this to work for vectors too.
5810 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
5811 APInt TruncatedBits =
5812 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
5813 APInt(Op.getValueSizeInBits(), 0) :
5814 APInt::getBitsSet(Op.getValueSizeInBits(),
5815 N0.getValueSizeInBits(),
5816 std::min(Op.getValueSizeInBits(),
5817 VT.getSizeInBits()));
5818 if (TruncatedBits == (KnownZero & TruncatedBits)) {
5819 if (VT.bitsGT(Op.getValueType()))
5820 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, Op);
5821 if (VT.bitsLT(Op.getValueType()))
5822 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5828 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5829 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
5830 if (N0.getOpcode() == ISD::TRUNCATE) {
5831 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5832 if (NarrowLoad.getNode()) {
5833 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5834 if (NarrowLoad.getNode() != N0.getNode()) {
5835 CombineTo(N0.getNode(), NarrowLoad);
5836 // CombineTo deleted the truncate, if needed, but not what's under it.
5839 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5843 // fold (zext (truncate x)) -> (and x, mask)
5844 if (N0.getOpcode() == ISD::TRUNCATE &&
5845 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
5847 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5848 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
5849 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5850 if (NarrowLoad.getNode()) {
5851 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5852 if (NarrowLoad.getNode() != N0.getNode()) {
5853 CombineTo(N0.getNode(), NarrowLoad);
5854 // CombineTo deleted the truncate, if needed, but not what's under it.
5857 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5860 SDValue Op = N0.getOperand(0);
5861 if (Op.getValueType().bitsLT(VT)) {
5862 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, Op);
5863 AddToWorklist(Op.getNode());
5864 } else if (Op.getValueType().bitsGT(VT)) {
5865 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5866 AddToWorklist(Op.getNode());
5868 return DAG.getZeroExtendInReg(Op, SDLoc(N),
5869 N0.getValueType().getScalarType());
5872 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
5873 // if either of the casts is not free.
5874 if (N0.getOpcode() == ISD::AND &&
5875 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5876 N0.getOperand(1).getOpcode() == ISD::Constant &&
5877 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5878 N0.getValueType()) ||
5879 !TLI.isZExtFree(N0.getValueType(), VT))) {
5880 SDValue X = N0.getOperand(0).getOperand(0);
5881 if (X.getValueType().bitsLT(VT)) {
5882 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(X), VT, X);
5883 } else if (X.getValueType().bitsGT(VT)) {
5884 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
5886 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5887 Mask = Mask.zext(VT.getSizeInBits());
5888 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5889 X, DAG.getConstant(Mask, VT));
5892 // fold (zext (load x)) -> (zext (truncate (zextload x)))
5893 // Only generate vector extloads when 1) they're legal, and 2) they are
5894 // deemed desirable by the target.
5895 if (ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5896 ((!LegalOperations && !VT.isVector() &&
5897 !cast<LoadSDNode>(N0)->isVolatile()) ||
5898 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, N0.getValueType()))) {
5899 bool DoXform = true;
5900 SmallVector<SDNode*, 4> SetCCs;
5901 if (!N0.hasOneUse())
5902 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
5904 DoXform &= TLI.isVectorLoadExtDesirable(SDValue(N, 0));
5906 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5907 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5909 LN0->getBasePtr(), N0.getValueType(),
5910 LN0->getMemOperand());
5911 CombineTo(N, ExtLoad);
5912 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5913 N0.getValueType(), ExtLoad);
5914 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5916 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5918 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5922 // fold (zext (load x)) to multiple smaller zextloads.
5923 // Only on illegal but splittable vectors.
5924 if (SDValue ExtLoad = CombineExtLoad(N))
5927 // fold (zext (and/or/xor (load x), cst)) ->
5928 // (and/or/xor (zextload x), (zext cst))
5929 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5930 N0.getOpcode() == ISD::XOR) &&
5931 isa<LoadSDNode>(N0.getOperand(0)) &&
5932 N0.getOperand(1).getOpcode() == ISD::Constant &&
5933 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, N0.getValueType()) &&
5934 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5935 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5936 if (LN0->getExtensionType() != ISD::SEXTLOAD && LN0->isUnindexed()) {
5937 bool DoXform = true;
5938 SmallVector<SDNode*, 4> SetCCs;
5939 if (!N0.hasOneUse())
5940 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
5943 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), VT,
5944 LN0->getChain(), LN0->getBasePtr(),
5946 LN0->getMemOperand());
5947 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5948 Mask = Mask.zext(VT.getSizeInBits());
5949 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5950 ExtLoad, DAG.getConstant(Mask, VT));
5951 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5952 SDLoc(N0.getOperand(0)),
5953 N0.getOperand(0).getValueType(), ExtLoad);
5955 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5956 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5958 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5963 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
5964 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
5965 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5966 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5967 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5968 EVT MemVT = LN0->getMemoryVT();
5969 if ((!LegalOperations && !LN0->isVolatile()) ||
5970 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT)) {
5971 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5973 LN0->getBasePtr(), MemVT,
5974 LN0->getMemOperand());
5975 CombineTo(N, ExtLoad);
5976 CombineTo(N0.getNode(),
5977 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(),
5979 ExtLoad.getValue(1));
5980 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5984 if (N0.getOpcode() == ISD::SETCC) {
5985 if (!LegalOperations && VT.isVector() &&
5986 N0.getValueType().getVectorElementType() == MVT::i1) {
5987 EVT N0VT = N0.getOperand(0).getValueType();
5988 if (getSetCCResultType(N0VT) == N0.getValueType())
5991 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
5992 // Only do this before legalize for now.
5993 EVT EltVT = VT.getVectorElementType();
5994 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
5995 DAG.getConstant(1, EltVT));
5996 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5997 // We know that the # elements of the results is the same as the
5998 // # elements of the compare (and the # elements of the compare result
5999 // for that matter). Check to see that they are the same size. If so,
6000 // we know that the element size of the sext'd result matches the
6001 // element size of the compare operands.
6002 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6003 DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
6005 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
6006 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT,
6009 // If the desired elements are smaller or larger than the source
6010 // elements we can use a matching integer vector type and then
6011 // truncate/sign extend
6012 EVT MatchingElementType =
6013 EVT::getIntegerVT(*DAG.getContext(),
6014 N0VT.getScalarType().getSizeInBits());
6015 EVT MatchingVectorType =
6016 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
6017 N0VT.getVectorNumElements());
6019 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
6021 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6022 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6023 DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT),
6024 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, OneOps));
6027 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
6029 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
6030 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
6031 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
6032 if (SCC.getNode()) return SCC;
6035 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
6036 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
6037 isa<ConstantSDNode>(N0.getOperand(1)) &&
6038 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
6040 SDValue ShAmt = N0.getOperand(1);
6041 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
6042 if (N0.getOpcode() == ISD::SHL) {
6043 SDValue InnerZExt = N0.getOperand(0);
6044 // If the original shl may be shifting out bits, do not perform this
6046 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
6047 InnerZExt.getOperand(0).getValueType().getSizeInBits();
6048 if (ShAmtVal > KnownZeroBits)
6054 // Ensure that the shift amount is wide enough for the shifted value.
6055 if (VT.getSizeInBits() >= 256)
6056 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
6058 return DAG.getNode(N0.getOpcode(), DL, VT,
6059 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
6066 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
6067 SDValue N0 = N->getOperand(0);
6068 EVT VT = N->getValueType(0);
6070 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
6072 return SDValue(Res, 0);
6074 // fold (aext (aext x)) -> (aext x)
6075 // fold (aext (zext x)) -> (zext x)
6076 // fold (aext (sext x)) -> (sext x)
6077 if (N0.getOpcode() == ISD::ANY_EXTEND ||
6078 N0.getOpcode() == ISD::ZERO_EXTEND ||
6079 N0.getOpcode() == ISD::SIGN_EXTEND)
6080 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0));
6082 // fold (aext (truncate (load x))) -> (aext (smaller load x))
6083 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
6084 if (N0.getOpcode() == ISD::TRUNCATE) {
6085 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
6086 if (NarrowLoad.getNode()) {
6087 SDNode* oye = N0.getNode()->getOperand(0).getNode();
6088 if (NarrowLoad.getNode() != N0.getNode()) {
6089 CombineTo(N0.getNode(), NarrowLoad);
6090 // CombineTo deleted the truncate, if needed, but not what's under it.
6093 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6097 // fold (aext (truncate x))
6098 if (N0.getOpcode() == ISD::TRUNCATE) {
6099 SDValue TruncOp = N0.getOperand(0);
6100 if (TruncOp.getValueType() == VT)
6101 return TruncOp; // x iff x size == zext size.
6102 if (TruncOp.getValueType().bitsGT(VT))
6103 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, TruncOp);
6104 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, TruncOp);
6107 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
6108 // if the trunc is not free.
6109 if (N0.getOpcode() == ISD::AND &&
6110 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
6111 N0.getOperand(1).getOpcode() == ISD::Constant &&
6112 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
6113 N0.getValueType())) {
6114 SDValue X = N0.getOperand(0).getOperand(0);
6115 if (X.getValueType().bitsLT(VT)) {
6116 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, X);
6117 } else if (X.getValueType().bitsGT(VT)) {
6118 X = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, X);
6120 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
6121 Mask = Mask.zext(VT.getSizeInBits());
6122 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6123 X, DAG.getConstant(Mask, VT));
6126 // fold (aext (load x)) -> (aext (truncate (extload x)))
6127 // None of the supported targets knows how to perform load and any_ext
6128 // on vectors in one instruction. We only perform this transformation on
6130 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
6131 ISD::isUNINDEXEDLoad(N0.getNode()) &&
6132 TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) {
6133 bool DoXform = true;
6134 SmallVector<SDNode*, 4> SetCCs;
6135 if (!N0.hasOneUse())
6136 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
6138 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6139 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
6141 LN0->getBasePtr(), N0.getValueType(),
6142 LN0->getMemOperand());
6143 CombineTo(N, ExtLoad);
6144 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
6145 N0.getValueType(), ExtLoad);
6146 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
6147 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
6149 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6153 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
6154 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
6155 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
6156 if (N0.getOpcode() == ISD::LOAD &&
6157 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
6159 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6160 ISD::LoadExtType ExtType = LN0->getExtensionType();
6161 EVT MemVT = LN0->getMemoryVT();
6162 if (!LegalOperations || TLI.isLoadExtLegal(ExtType, VT, MemVT)) {
6163 SDValue ExtLoad = DAG.getExtLoad(ExtType, SDLoc(N),
6164 VT, LN0->getChain(), LN0->getBasePtr(),
6165 MemVT, LN0->getMemOperand());
6166 CombineTo(N, ExtLoad);
6167 CombineTo(N0.getNode(),
6168 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
6169 N0.getValueType(), ExtLoad),
6170 ExtLoad.getValue(1));
6171 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6175 if (N0.getOpcode() == ISD::SETCC) {
6177 // aext(setcc) -> vsetcc
6178 // aext(setcc) -> truncate(vsetcc)
6179 // aext(setcc) -> aext(vsetcc)
6180 // Only do this before legalize for now.
6181 if (VT.isVector() && !LegalOperations) {
6182 EVT N0VT = N0.getOperand(0).getValueType();
6183 // We know that the # elements of the results is the same as the
6184 // # elements of the compare (and the # elements of the compare result
6185 // for that matter). Check to see that they are the same size. If so,
6186 // we know that the element size of the sext'd result matches the
6187 // element size of the compare operands.
6188 if (VT.getSizeInBits() == N0VT.getSizeInBits())
6189 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
6191 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6192 // If the desired elements are smaller or larger than the source
6193 // elements we can use a matching integer vector type and then
6194 // truncate/any extend
6196 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
6198 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
6200 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6201 return DAG.getAnyExtOrTrunc(VsetCC, SDLoc(N), VT);
6205 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
6207 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
6208 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
6209 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
6217 /// See if the specified operand can be simplified with the knowledge that only
6218 /// the bits specified by Mask are used. If so, return the simpler operand,
6219 /// otherwise return a null SDValue.
6220 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
6221 switch (V.getOpcode()) {
6223 case ISD::Constant: {
6224 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
6225 assert(CV && "Const value should be ConstSDNode.");
6226 const APInt &CVal = CV->getAPIntValue();
6227 APInt NewVal = CVal & Mask;
6229 return DAG.getConstant(NewVal, V.getValueType());
6234 // If the LHS or RHS don't contribute bits to the or, drop them.
6235 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
6236 return V.getOperand(1);
6237 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
6238 return V.getOperand(0);
6241 // Only look at single-use SRLs.
6242 if (!V.getNode()->hasOneUse())
6244 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
6245 // See if we can recursively simplify the LHS.
6246 unsigned Amt = RHSC->getZExtValue();
6248 // Watch out for shift count overflow though.
6249 if (Amt >= Mask.getBitWidth()) break;
6250 APInt NewMask = Mask << Amt;
6251 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
6252 if (SimplifyLHS.getNode())
6253 return DAG.getNode(ISD::SRL, SDLoc(V), V.getValueType(),
6254 SimplifyLHS, V.getOperand(1));
6260 /// If the result of a wider load is shifted to right of N bits and then
6261 /// truncated to a narrower type and where N is a multiple of number of bits of
6262 /// the narrower type, transform it to a narrower load from address + N / num of
6263 /// bits of new type. If the result is to be extended, also fold the extension
6264 /// to form a extending load.
6265 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
6266 unsigned Opc = N->getOpcode();
6268 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
6269 SDValue N0 = N->getOperand(0);
6270 EVT VT = N->getValueType(0);
6273 // This transformation isn't valid for vector loads.
6277 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
6279 if (Opc == ISD::SIGN_EXTEND_INREG) {
6280 ExtType = ISD::SEXTLOAD;
6281 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
6282 } else if (Opc == ISD::SRL) {
6283 // Another special-case: SRL is basically zero-extending a narrower value.
6284 ExtType = ISD::ZEXTLOAD;
6286 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
6287 if (!N01) return SDValue();
6288 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
6289 VT.getSizeInBits() - N01->getZExtValue());
6291 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, VT, ExtVT))
6294 unsigned EVTBits = ExtVT.getSizeInBits();
6296 // Do not generate loads of non-round integer types since these can
6297 // be expensive (and would be wrong if the type is not byte sized).
6298 if (!ExtVT.isRound())
6302 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
6303 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
6304 ShAmt = N01->getZExtValue();
6305 // Is the shift amount a multiple of size of VT?
6306 if ((ShAmt & (EVTBits-1)) == 0) {
6307 N0 = N0.getOperand(0);
6308 // Is the load width a multiple of size of VT?
6309 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
6313 // At this point, we must have a load or else we can't do the transform.
6314 if (!isa<LoadSDNode>(N0)) return SDValue();
6316 // Because a SRL must be assumed to *need* to zero-extend the high bits
6317 // (as opposed to anyext the high bits), we can't combine the zextload
6318 // lowering of SRL and an sextload.
6319 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
6322 // If the shift amount is larger than the input type then we're not
6323 // accessing any of the loaded bytes. If the load was a zextload/extload
6324 // then the result of the shift+trunc is zero/undef (handled elsewhere).
6325 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
6330 // If the load is shifted left (and the result isn't shifted back right),
6331 // we can fold the truncate through the shift.
6332 unsigned ShLeftAmt = 0;
6333 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
6334 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
6335 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
6336 ShLeftAmt = N01->getZExtValue();
6337 N0 = N0.getOperand(0);
6341 // If we haven't found a load, we can't narrow it. Don't transform one with
6342 // multiple uses, this would require adding a new load.
6343 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse())
6346 // Don't change the width of a volatile load.
6347 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6348 if (LN0->isVolatile())
6351 // Verify that we are actually reducing a load width here.
6352 if (LN0->getMemoryVT().getSizeInBits() < EVTBits)
6355 // For the transform to be legal, the load must produce only two values
6356 // (the value loaded and the chain). Don't transform a pre-increment
6357 // load, for example, which produces an extra value. Otherwise the
6358 // transformation is not equivalent, and the downstream logic to replace
6359 // uses gets things wrong.
6360 if (LN0->getNumValues() > 2)
6363 // If the load that we're shrinking is an extload and we're not just
6364 // discarding the extension we can't simply shrink the load. Bail.
6365 // TODO: It would be possible to merge the extensions in some cases.
6366 if (LN0->getExtensionType() != ISD::NON_EXTLOAD &&
6367 LN0->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits() + ShAmt)
6370 if (!TLI.shouldReduceLoadWidth(LN0, ExtType, ExtVT))
6373 EVT PtrType = N0.getOperand(1).getValueType();
6375 if (PtrType == MVT::Untyped || PtrType.isExtended())
6376 // It's not possible to generate a constant of extended or untyped type.
6379 // For big endian targets, we need to adjust the offset to the pointer to
6380 // load the correct bytes.
6381 if (TLI.isBigEndian()) {
6382 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
6383 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
6384 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
6387 uint64_t PtrOff = ShAmt / 8;
6388 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
6389 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0),
6390 PtrType, LN0->getBasePtr(),
6391 DAG.getConstant(PtrOff, PtrType));
6392 AddToWorklist(NewPtr.getNode());
6395 if (ExtType == ISD::NON_EXTLOAD)
6396 Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr,
6397 LN0->getPointerInfo().getWithOffset(PtrOff),
6398 LN0->isVolatile(), LN0->isNonTemporal(),
6399 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
6401 Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(),NewPtr,
6402 LN0->getPointerInfo().getWithOffset(PtrOff),
6403 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
6404 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
6406 // Replace the old load's chain with the new load's chain.
6407 WorklistRemover DeadNodes(*this);
6408 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
6410 // Shift the result left, if we've swallowed a left shift.
6411 SDValue Result = Load;
6412 if (ShLeftAmt != 0) {
6413 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
6414 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
6416 // If the shift amount is as large as the result size (but, presumably,
6417 // no larger than the source) then the useful bits of the result are
6418 // zero; we can't simply return the shortened shift, because the result
6419 // of that operation is undefined.
6420 if (ShLeftAmt >= VT.getSizeInBits())
6421 Result = DAG.getConstant(0, VT);
6423 Result = DAG.getNode(ISD::SHL, SDLoc(N0), VT,
6424 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
6427 // Return the new loaded value.
6431 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
6432 SDValue N0 = N->getOperand(0);
6433 SDValue N1 = N->getOperand(1);
6434 EVT VT = N->getValueType(0);
6435 EVT EVT = cast<VTSDNode>(N1)->getVT();
6436 unsigned VTBits = VT.getScalarType().getSizeInBits();
6437 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
6439 // fold (sext_in_reg c1) -> c1
6440 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
6441 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1);
6443 // If the input is already sign extended, just drop the extension.
6444 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
6447 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
6448 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
6449 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT()))
6450 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
6451 N0.getOperand(0), N1);
6453 // fold (sext_in_reg (sext x)) -> (sext x)
6454 // fold (sext_in_reg (aext x)) -> (sext x)
6455 // if x is small enough.
6456 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
6457 SDValue N00 = N0.getOperand(0);
6458 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
6459 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
6460 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
6463 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
6464 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
6465 return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT);
6467 // fold operands of sext_in_reg based on knowledge that the top bits are not
6469 if (SimplifyDemandedBits(SDValue(N, 0)))
6470 return SDValue(N, 0);
6472 // fold (sext_in_reg (load x)) -> (smaller sextload x)
6473 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
6474 SDValue NarrowLoad = ReduceLoadWidth(N);
6475 if (NarrowLoad.getNode())
6478 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
6479 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
6480 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
6481 if (N0.getOpcode() == ISD::SRL) {
6482 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
6483 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
6484 // We can turn this into an SRA iff the input to the SRL is already sign
6486 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
6487 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
6488 return DAG.getNode(ISD::SRA, SDLoc(N), VT,
6489 N0.getOperand(0), N0.getOperand(1));
6493 // fold (sext_inreg (extload x)) -> (sextload x)
6494 if (ISD::isEXTLoad(N0.getNode()) &&
6495 ISD::isUNINDEXEDLoad(N0.getNode()) &&
6496 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
6497 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6498 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) {
6499 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6500 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
6502 LN0->getBasePtr(), EVT,
6503 LN0->getMemOperand());
6504 CombineTo(N, ExtLoad);
6505 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
6506 AddToWorklist(ExtLoad.getNode());
6507 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6509 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
6510 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
6512 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
6513 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6514 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) {
6515 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6516 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
6518 LN0->getBasePtr(), EVT,
6519 LN0->getMemOperand());
6520 CombineTo(N, ExtLoad);
6521 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
6522 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6525 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
6526 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
6527 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
6528 N0.getOperand(1), false);
6529 if (BSwap.getNode())
6530 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
6534 // Fold a sext_inreg of a build_vector of ConstantSDNodes or undefs
6535 // into a build_vector.
6536 if (ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
6537 SmallVector<SDValue, 8> Elts;
6538 unsigned NumElts = N0->getNumOperands();
6539 unsigned ShAmt = VTBits - EVTBits;
6541 for (unsigned i = 0; i != NumElts; ++i) {
6542 SDValue Op = N0->getOperand(i);
6543 if (Op->getOpcode() == ISD::UNDEF) {
6548 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
6549 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
6550 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
6551 Op.getValueType()));
6554 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Elts);
6560 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
6561 SDValue N0 = N->getOperand(0);
6562 EVT VT = N->getValueType(0);
6563 bool isLE = TLI.isLittleEndian();
6566 if (N0.getValueType() == N->getValueType(0))
6568 // fold (truncate c1) -> c1
6569 if (isConstantIntBuildVectorOrConstantInt(N0))
6570 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0);
6571 // fold (truncate (truncate x)) -> (truncate x)
6572 if (N0.getOpcode() == ISD::TRUNCATE)
6573 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6574 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
6575 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
6576 N0.getOpcode() == ISD::SIGN_EXTEND ||
6577 N0.getOpcode() == ISD::ANY_EXTEND) {
6578 if (N0.getOperand(0).getValueType().bitsLT(VT))
6579 // if the source is smaller than the dest, we still need an extend
6580 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
6582 if (N0.getOperand(0).getValueType().bitsGT(VT))
6583 // if the source is larger than the dest, than we just need the truncate
6584 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6585 // if the source and dest are the same type, we can drop both the extend
6586 // and the truncate.
6587 return N0.getOperand(0);
6590 // Fold extract-and-trunc into a narrow extract. For example:
6591 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
6592 // i32 y = TRUNCATE(i64 x)
6594 // v16i8 b = BITCAST (v2i64 val)
6595 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
6597 // Note: We only run this optimization after type legalization (which often
6598 // creates this pattern) and before operation legalization after which
6599 // we need to be more careful about the vector instructions that we generate.
6600 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6601 LegalTypes && !LegalOperations && N0->hasOneUse() && VT != MVT::i1) {
6603 EVT VecTy = N0.getOperand(0).getValueType();
6604 EVT ExTy = N0.getValueType();
6605 EVT TrTy = N->getValueType(0);
6607 unsigned NumElem = VecTy.getVectorNumElements();
6608 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
6610 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
6611 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
6613 SDValue EltNo = N0->getOperand(1);
6614 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
6615 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
6616 EVT IndexTy = TLI.getVectorIdxTy();
6617 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
6619 SDValue V = DAG.getNode(ISD::BITCAST, SDLoc(N),
6620 NVT, N0.getOperand(0));
6622 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
6624 DAG.getConstant(Index, IndexTy));
6628 // trunc (select c, a, b) -> select c, (trunc a), (trunc b)
6629 if (N0.getOpcode() == ISD::SELECT) {
6630 EVT SrcVT = N0.getValueType();
6631 if ((!LegalOperations || TLI.isOperationLegal(ISD::SELECT, SrcVT)) &&
6632 TLI.isTruncateFree(SrcVT, VT)) {
6634 SDValue Cond = N0.getOperand(0);
6635 SDValue TruncOp0 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(1));
6636 SDValue TruncOp1 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(2));
6637 return DAG.getNode(ISD::SELECT, SDLoc(N), VT, Cond, TruncOp0, TruncOp1);
6641 // Fold a series of buildvector, bitcast, and truncate if possible.
6643 // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to
6644 // (2xi32 (buildvector x, y)).
6645 if (Level == AfterLegalizeVectorOps && VT.isVector() &&
6646 N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
6647 N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
6648 N0.getOperand(0).hasOneUse()) {
6650 SDValue BuildVect = N0.getOperand(0);
6651 EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType();
6652 EVT TruncVecEltTy = VT.getVectorElementType();
6654 // Check that the element types match.
6655 if (BuildVectEltTy == TruncVecEltTy) {
6656 // Now we only need to compute the offset of the truncated elements.
6657 unsigned BuildVecNumElts = BuildVect.getNumOperands();
6658 unsigned TruncVecNumElts = VT.getVectorNumElements();
6659 unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts;
6661 assert((BuildVecNumElts % TruncVecNumElts) == 0 &&
6662 "Invalid number of elements");
6664 SmallVector<SDValue, 8> Opnds;
6665 for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset)
6666 Opnds.push_back(BuildVect.getOperand(i));
6668 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
6672 // See if we can simplify the input to this truncate through knowledge that
6673 // only the low bits are being used.
6674 // For example "trunc (or (shl x, 8), y)" // -> trunc y
6675 // Currently we only perform this optimization on scalars because vectors
6676 // may have different active low bits.
6677 if (!VT.isVector()) {
6679 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
6680 VT.getSizeInBits()));
6681 if (Shorter.getNode())
6682 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter);
6684 // fold (truncate (load x)) -> (smaller load x)
6685 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
6686 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
6687 SDValue Reduced = ReduceLoadWidth(N);
6688 if (Reduced.getNode())
6690 // Handle the case where the load remains an extending load even
6691 // after truncation.
6692 if (N0.hasOneUse() && ISD::isUNINDEXEDLoad(N0.getNode())) {
6693 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6694 if (!LN0->isVolatile() &&
6695 LN0->getMemoryVT().getStoreSizeInBits() < VT.getSizeInBits()) {
6696 SDValue NewLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(LN0),
6697 VT, LN0->getChain(), LN0->getBasePtr(),
6699 LN0->getMemOperand());
6700 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLoad.getValue(1));
6705 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
6706 // where ... are all 'undef'.
6707 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
6708 SmallVector<EVT, 8> VTs;
6711 unsigned NumDefs = 0;
6713 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
6714 SDValue X = N0.getOperand(i);
6715 if (X.getOpcode() != ISD::UNDEF) {
6720 // Stop if more than one members are non-undef.
6723 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
6724 VT.getVectorElementType(),
6725 X.getValueType().getVectorNumElements()));
6729 return DAG.getUNDEF(VT);
6732 assert(V.getNode() && "The single defined operand is empty!");
6733 SmallVector<SDValue, 8> Opnds;
6734 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
6736 Opnds.push_back(DAG.getUNDEF(VTs[i]));
6739 SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V);
6740 AddToWorklist(NV.getNode());
6741 Opnds.push_back(NV);
6743 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Opnds);
6747 // Simplify the operands using demanded-bits information.
6748 if (!VT.isVector() &&
6749 SimplifyDemandedBits(SDValue(N, 0)))
6750 return SDValue(N, 0);
6755 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
6756 SDValue Elt = N->getOperand(i);
6757 if (Elt.getOpcode() != ISD::MERGE_VALUES)
6758 return Elt.getNode();
6759 return Elt.getOperand(Elt.getResNo()).getNode();
6762 /// build_pair (load, load) -> load
6763 /// if load locations are consecutive.
6764 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
6765 assert(N->getOpcode() == ISD::BUILD_PAIR);
6767 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
6768 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
6769 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
6770 LD1->getAddressSpace() != LD2->getAddressSpace())
6772 EVT LD1VT = LD1->getValueType(0);
6774 if (ISD::isNON_EXTLoad(LD2) &&
6776 // If both are volatile this would reduce the number of volatile loads.
6777 // If one is volatile it might be ok, but play conservative and bail out.
6778 !LD1->isVolatile() &&
6779 !LD2->isVolatile() &&
6780 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
6781 unsigned Align = LD1->getAlignment();
6782 unsigned NewAlign = TLI.getDataLayout()->
6783 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6785 if (NewAlign <= Align &&
6786 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
6787 return DAG.getLoad(VT, SDLoc(N), LD1->getChain(),
6788 LD1->getBasePtr(), LD1->getPointerInfo(),
6789 false, false, false, Align);
6795 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
6796 SDValue N0 = N->getOperand(0);
6797 EVT VT = N->getValueType(0);
6799 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
6800 // Only do this before legalize, since afterward the target may be depending
6801 // on the bitconvert.
6802 // First check to see if this is all constant.
6804 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
6806 bool isSimple = cast<BuildVectorSDNode>(N0)->isConstant();
6808 EVT DestEltVT = N->getValueType(0).getVectorElementType();
6809 assert(!DestEltVT.isVector() &&
6810 "Element type of vector ValueType must not be vector!");
6812 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
6815 // If the input is a constant, let getNode fold it.
6816 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
6817 // If we can't allow illegal operations, we need to check that this is just
6818 // a fp -> int or int -> conversion and that the resulting operation will
6820 if (!LegalOperations ||
6821 (isa<ConstantSDNode>(N0) && VT.isFloatingPoint() && !VT.isVector() &&
6822 TLI.isOperationLegal(ISD::ConstantFP, VT)) ||
6823 (isa<ConstantFPSDNode>(N0) && VT.isInteger() && !VT.isVector() &&
6824 TLI.isOperationLegal(ISD::Constant, VT)))
6825 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, N0);
6828 // (conv (conv x, t1), t2) -> (conv x, t2)
6829 if (N0.getOpcode() == ISD::BITCAST)
6830 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT,
6833 // fold (conv (load x)) -> (load (conv*)x)
6834 // If the resultant load doesn't need a higher alignment than the original!
6835 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6836 // Do not change the width of a volatile load.
6837 !cast<LoadSDNode>(N0)->isVolatile() &&
6838 // Do not remove the cast if the types differ in endian layout.
6839 TLI.hasBigEndianPartOrdering(N0.getValueType()) ==
6840 TLI.hasBigEndianPartOrdering(VT) &&
6841 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)) &&
6842 TLI.isLoadBitCastBeneficial(N0.getValueType(), VT)) {
6843 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6844 unsigned Align = TLI.getDataLayout()->
6845 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6846 unsigned OrigAlign = LN0->getAlignment();
6848 if (Align <= OrigAlign) {
6849 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(),
6850 LN0->getBasePtr(), LN0->getPointerInfo(),
6851 LN0->isVolatile(), LN0->isNonTemporal(),
6852 LN0->isInvariant(), OrigAlign,
6854 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
6859 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
6860 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
6861 // This often reduces constant pool loads.
6862 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) ||
6863 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) &&
6864 N0.getNode()->hasOneUse() && VT.isInteger() &&
6865 !VT.isVector() && !N0.getValueType().isVector()) {
6866 SDValue NewConv = DAG.getNode(ISD::BITCAST, SDLoc(N0), VT,
6868 AddToWorklist(NewConv.getNode());
6870 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6871 if (N0.getOpcode() == ISD::FNEG)
6872 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
6873 NewConv, DAG.getConstant(SignBit, VT));
6874 assert(N0.getOpcode() == ISD::FABS);
6875 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6876 NewConv, DAG.getConstant(~SignBit, VT));
6879 // fold (bitconvert (fcopysign cst, x)) ->
6880 // (or (and (bitconvert x), sign), (and cst, (not sign)))
6881 // Note that we don't handle (copysign x, cst) because this can always be
6882 // folded to an fneg or fabs.
6883 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
6884 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
6885 VT.isInteger() && !VT.isVector()) {
6886 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
6887 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
6888 if (isTypeLegal(IntXVT)) {
6889 SDValue X = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6890 IntXVT, N0.getOperand(1));
6891 AddToWorklist(X.getNode());
6893 // If X has a different width than the result/lhs, sext it or truncate it.
6894 unsigned VTWidth = VT.getSizeInBits();
6895 if (OrigXWidth < VTWidth) {
6896 X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X);
6897 AddToWorklist(X.getNode());
6898 } else if (OrigXWidth > VTWidth) {
6899 // To get the sign bit in the right place, we have to shift it right
6900 // before truncating.
6901 X = DAG.getNode(ISD::SRL, SDLoc(X),
6902 X.getValueType(), X,
6903 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
6904 AddToWorklist(X.getNode());
6905 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
6906 AddToWorklist(X.getNode());
6909 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6910 X = DAG.getNode(ISD::AND, SDLoc(X), VT,
6911 X, DAG.getConstant(SignBit, VT));
6912 AddToWorklist(X.getNode());
6914 SDValue Cst = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6915 VT, N0.getOperand(0));
6916 Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT,
6917 Cst, DAG.getConstant(~SignBit, VT));
6918 AddToWorklist(Cst.getNode());
6920 return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst);
6924 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
6925 if (N0.getOpcode() == ISD::BUILD_PAIR) {
6926 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
6927 if (CombineLD.getNode())
6934 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
6935 EVT VT = N->getValueType(0);
6936 return CombineConsecutiveLoads(N, VT);
6939 /// We know that BV is a build_vector node with Constant, ConstantFP or Undef
6940 /// operands. DstEltVT indicates the destination element value type.
6941 SDValue DAGCombiner::
6942 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
6943 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
6945 // If this is already the right type, we're done.
6946 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
6948 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
6949 unsigned DstBitSize = DstEltVT.getSizeInBits();
6951 // If this is a conversion of N elements of one type to N elements of another
6952 // type, convert each element. This handles FP<->INT cases.
6953 if (SrcBitSize == DstBitSize) {
6954 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6955 BV->getValueType(0).getVectorNumElements());
6957 // Due to the FP element handling below calling this routine recursively,
6958 // we can end up with a scalar-to-vector node here.
6959 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
6960 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6961 DAG.getNode(ISD::BITCAST, SDLoc(BV),
6962 DstEltVT, BV->getOperand(0)));
6964 SmallVector<SDValue, 8> Ops;
6965 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6966 SDValue Op = BV->getOperand(i);
6967 // If the vector element type is not legal, the BUILD_VECTOR operands
6968 // are promoted and implicitly truncated. Make that explicit here.
6969 if (Op.getValueType() != SrcEltVT)
6970 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op);
6971 Ops.push_back(DAG.getNode(ISD::BITCAST, SDLoc(BV),
6973 AddToWorklist(Ops.back().getNode());
6975 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6978 // Otherwise, we're growing or shrinking the elements. To avoid having to
6979 // handle annoying details of growing/shrinking FP values, we convert them to
6981 if (SrcEltVT.isFloatingPoint()) {
6982 // Convert the input float vector to a int vector where the elements are the
6984 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
6985 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
6989 // Now we know the input is an integer vector. If the output is a FP type,
6990 // convert to integer first, then to FP of the right size.
6991 if (DstEltVT.isFloatingPoint()) {
6992 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
6993 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
6995 // Next, convert to FP elements of the same size.
6996 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
6999 // Okay, we know the src/dst types are both integers of differing types.
7000 // Handling growing first.
7001 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
7002 if (SrcBitSize < DstBitSize) {
7003 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
7005 SmallVector<SDValue, 8> Ops;
7006 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
7007 i += NumInputsPerOutput) {
7008 bool isLE = TLI.isLittleEndian();
7009 APInt NewBits = APInt(DstBitSize, 0);
7010 bool EltIsUndef = true;
7011 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
7012 // Shift the previously computed bits over.
7013 NewBits <<= SrcBitSize;
7014 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
7015 if (Op.getOpcode() == ISD::UNDEF) continue;
7018 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
7019 zextOrTrunc(SrcBitSize).zext(DstBitSize);
7023 Ops.push_back(DAG.getUNDEF(DstEltVT));
7025 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
7028 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
7029 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
7032 // Finally, this must be the case where we are shrinking elements: each input
7033 // turns into multiple outputs.
7034 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
7035 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
7036 NumOutputsPerInput*BV->getNumOperands());
7037 SmallVector<SDValue, 8> Ops;
7039 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
7040 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
7041 Ops.append(NumOutputsPerInput, DAG.getUNDEF(DstEltVT));
7045 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
7046 getAPIntValue().zextOrTrunc(SrcBitSize);
7048 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
7049 APInt ThisVal = OpVal.trunc(DstBitSize);
7050 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
7051 OpVal = OpVal.lshr(DstBitSize);
7054 // For big endian targets, swap the order of the pieces of each element.
7055 if (TLI.isBigEndian())
7056 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
7059 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
7062 // Attempt different variants of (fadd (fmul a, b), c) -> fma or fmad
7063 static SDValue performFaddFmulCombines(unsigned FusedOpcode,
7066 const TargetLowering &TLI,
7067 SelectionDAG &DAG) {
7068 SDValue N0 = N->getOperand(0);
7069 SDValue N1 = N->getOperand(1);
7070 EVT VT = N->getValueType(0);
7072 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
7073 if (N0.getOpcode() == ISD::FMUL &&
7074 (Aggressive || N0->hasOneUse())) {
7075 return DAG.getNode(FusedOpcode, SDLoc(N), VT,
7076 N0.getOperand(0), N0.getOperand(1), N1);
7079 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
7080 // Note: Commutes FADD operands.
7081 if (N1.getOpcode() == ISD::FMUL &&
7082 (Aggressive || N1->hasOneUse())) {
7083 return DAG.getNode(FusedOpcode, SDLoc(N), VT,
7084 N1.getOperand(0), N1.getOperand(1), N0);
7087 // More folding opportunities when target permits.
7089 // fold (fadd (fma x, y, (fmul u, v)), z) -> (fma x, y (fma u, v, z))
7090 if (N0.getOpcode() == ISD::FMA &&
7091 N0.getOperand(2).getOpcode() == ISD::FMUL) {
7092 return DAG.getNode(FusedOpcode, SDLoc(N), VT,
7093 N0.getOperand(0), N0.getOperand(1),
7094 DAG.getNode(FusedOpcode, SDLoc(N), VT,
7095 N0.getOperand(2).getOperand(0),
7096 N0.getOperand(2).getOperand(1),
7100 // fold (fadd x, (fma y, z, (fmul u, v)) -> (fma y, z (fma u, v, x))
7101 if (N1->getOpcode() == ISD::FMA &&
7102 N1.getOperand(2).getOpcode() == ISD::FMUL) {
7103 return DAG.getNode(FusedOpcode, SDLoc(N), VT,
7104 N1.getOperand(0), N1.getOperand(1),
7105 DAG.getNode(FusedOpcode, SDLoc(N), VT,
7106 N1.getOperand(2).getOperand(0),
7107 N1.getOperand(2).getOperand(1),
7115 static SDValue performFsubFmulCombines(unsigned FusedOpcode,
7118 const TargetLowering &TLI,
7119 SelectionDAG &DAG) {
7120 SDValue N0 = N->getOperand(0);
7121 SDValue N1 = N->getOperand(1);
7122 EVT VT = N->getValueType(0);
7126 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
7127 if (N0.getOpcode() == ISD::FMUL &&
7128 (Aggressive || N0->hasOneUse())) {
7129 return DAG.getNode(FusedOpcode, SL, VT,
7130 N0.getOperand(0), N0.getOperand(1),
7131 DAG.getNode(ISD::FNEG, SL, VT, N1));
7134 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
7135 // Note: Commutes FSUB operands.
7136 if (N1.getOpcode() == ISD::FMUL &&
7137 (Aggressive || N1->hasOneUse()))
7138 return DAG.getNode(FusedOpcode, SL, VT,
7139 DAG.getNode(ISD::FNEG, SL, VT,
7141 N1.getOperand(1), N0);
7143 // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
7144 if (N0.getOpcode() == ISD::FNEG &&
7145 N0.getOperand(0).getOpcode() == ISD::FMUL &&
7146 (Aggressive || (N0->hasOneUse() && N0.getOperand(0).hasOneUse()))) {
7147 SDValue N00 = N0.getOperand(0).getOperand(0);
7148 SDValue N01 = N0.getOperand(0).getOperand(1);
7149 return DAG.getNode(FusedOpcode, SL, VT,
7150 DAG.getNode(ISD::FNEG, SL, VT, N00), N01,
7151 DAG.getNode(ISD::FNEG, SL, VT, N1));
7154 // More folding opportunities when target permits.
7156 // fold (fsub (fma x, y, (fmul u, v)), z)
7157 // -> (fma x, y (fma u, v, (fneg z)))
7158 if (N0.getOpcode() == FusedOpcode &&
7159 N0.getOperand(2).getOpcode() == ISD::FMUL) {
7160 return DAG.getNode(FusedOpcode, SDLoc(N), VT,
7161 N0.getOperand(0), N0.getOperand(1),
7162 DAG.getNode(FusedOpcode, SDLoc(N), VT,
7163 N0.getOperand(2).getOperand(0),
7164 N0.getOperand(2).getOperand(1),
7165 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7169 // fold (fsub x, (fma y, z, (fmul u, v)))
7170 // -> (fma (fneg y), z, (fma (fneg u), v, x))
7171 if (N1.getOpcode() == FusedOpcode &&
7172 N1.getOperand(2).getOpcode() == ISD::FMUL) {
7173 SDValue N20 = N1.getOperand(2).getOperand(0);
7174 SDValue N21 = N1.getOperand(2).getOperand(1);
7175 return DAG.getNode(FusedOpcode, SDLoc(N), VT,
7176 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7179 DAG.getNode(FusedOpcode, SDLoc(N), VT,
7180 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7189 SDValue DAGCombiner::visitFADD(SDNode *N) {
7190 SDValue N0 = N->getOperand(0);
7191 SDValue N1 = N->getOperand(1);
7192 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7193 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7194 EVT VT = N->getValueType(0);
7195 const TargetOptions &Options = DAG.getTarget().Options;
7199 if (SDValue FoldedVOp = SimplifyVBinOp(N))
7202 // fold (fadd c1, c2) -> c1 + c2
7204 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N1);
7206 // canonicalize constant to RHS
7207 if (N0CFP && !N1CFP)
7208 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N0);
7210 // fold (fadd A, (fneg B)) -> (fsub A, B)
7211 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
7212 isNegatibleForFree(N1, LegalOperations, TLI, &Options) == 2)
7213 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0,
7214 GetNegatedExpression(N1, DAG, LegalOperations));
7216 // fold (fadd (fneg A), B) -> (fsub B, A)
7217 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
7218 isNegatibleForFree(N0, LegalOperations, TLI, &Options) == 2)
7219 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N1,
7220 GetNegatedExpression(N0, DAG, LegalOperations));
7222 // If 'unsafe math' is enabled, fold lots of things.
7223 if (Options.UnsafeFPMath) {
7224 // No FP constant should be created after legalization as Instruction
7225 // Selection pass has a hard time dealing with FP constants.
7226 bool AllowNewConst = (Level < AfterLegalizeDAG);
7228 // fold (fadd A, 0) -> A
7229 if (N1CFP && N1CFP->getValueAPF().isZero())
7232 // fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
7233 if (N1CFP && N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
7234 isa<ConstantFPSDNode>(N0.getOperand(1)))
7235 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0.getOperand(0),
7236 DAG.getNode(ISD::FADD, SDLoc(N), VT,
7237 N0.getOperand(1), N1));
7239 // If allowed, fold (fadd (fneg x), x) -> 0.0
7240 if (AllowNewConst && N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1)
7241 return DAG.getConstantFP(0.0, VT);
7243 // If allowed, fold (fadd x, (fneg x)) -> 0.0
7244 if (AllowNewConst && N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0)
7245 return DAG.getConstantFP(0.0, VT);
7247 // We can fold chains of FADD's of the same value into multiplications.
7248 // This transform is not safe in general because we are reducing the number
7249 // of rounding steps.
7250 if (TLI.isOperationLegalOrCustom(ISD::FMUL, VT) && !N0CFP && !N1CFP) {
7251 if (N0.getOpcode() == ISD::FMUL) {
7252 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
7253 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
7255 // (fadd (fmul x, c), x) -> (fmul x, c+1)
7256 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
7257 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
7259 DAG.getConstantFP(1.0, VT));
7260 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, NewCFP);
7263 // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2)
7264 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
7265 N1.getOperand(0) == N1.getOperand(1) &&
7266 N0.getOperand(0) == N1.getOperand(0)) {
7267 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
7269 DAG.getConstantFP(2.0, VT));
7270 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7271 N0.getOperand(0), NewCFP);
7275 if (N1.getOpcode() == ISD::FMUL) {
7276 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
7277 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
7279 // (fadd x, (fmul x, c)) -> (fmul x, c+1)
7280 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
7281 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
7283 DAG.getConstantFP(1.0, VT));
7284 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, NewCFP);
7287 // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2)
7288 if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD &&
7289 N0.getOperand(0) == N0.getOperand(1) &&
7290 N1.getOperand(0) == N0.getOperand(0)) {
7291 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
7293 DAG.getConstantFP(2.0, VT));
7294 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1.getOperand(0), NewCFP);
7298 if (N0.getOpcode() == ISD::FADD && AllowNewConst) {
7299 ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
7300 // (fadd (fadd x, x), x) -> (fmul x, 3.0)
7301 if (!CFP && N0.getOperand(0) == N0.getOperand(1) &&
7302 (N0.getOperand(0) == N1))
7303 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7304 N1, DAG.getConstantFP(3.0, VT));
7307 if (N1.getOpcode() == ISD::FADD && AllowNewConst) {
7308 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
7309 // (fadd x, (fadd x, x)) -> (fmul x, 3.0)
7310 if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) &&
7311 N1.getOperand(0) == N0)
7312 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7313 N0, DAG.getConstantFP(3.0, VT));
7316 // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0)
7317 if (AllowNewConst &&
7318 N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
7319 N0.getOperand(0) == N0.getOperand(1) &&
7320 N1.getOperand(0) == N1.getOperand(1) &&
7321 N0.getOperand(0) == N1.getOperand(0))
7322 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7323 N0.getOperand(0), DAG.getConstantFP(4.0, VT));
7325 } // enable-unsafe-fp-math
7327 if (LegalOperations && TLI.isOperationLegal(ISD::FMAD, VT)) {
7328 // Assume if there is an fmad instruction that it should be aggressively
7330 if (SDValue Fused = performFaddFmulCombines(ISD::FMAD, true, N, TLI, DAG))
7334 // FADD -> FMA combines:
7335 if ((Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath) &&
7336 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
7337 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
7339 if (!TLI.isOperationLegal(ISD::FMAD, VT)) {
7340 // Don't form FMA if we are preferring FMAD.
7342 = performFaddFmulCombines(ISD::FMA,
7343 TLI.enableAggressiveFMAFusion(VT),
7349 // When FP_EXTEND nodes are free on the target, and there is an opportunity
7350 // to combine into FMA, arrange such nodes accordingly.
7351 if (TLI.isFPExtFree(VT)) {
7353 // fold (fadd (fpext (fmul x, y)), z) -> (fma (fpext x), (fpext y), z)
7354 if (N0.getOpcode() == ISD::FP_EXTEND) {
7355 SDValue N00 = N0.getOperand(0);
7356 if (N00.getOpcode() == ISD::FMUL)
7357 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7358 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7360 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7361 N00.getOperand(1)), N1);
7364 // fold (fadd x, (fpext (fmul y, z)), z) -> (fma (fpext y), (fpext z), x)
7365 // Note: Commutes FADD operands.
7366 if (N1.getOpcode() == ISD::FP_EXTEND) {
7367 SDValue N10 = N1.getOperand(0);
7368 if (N10.getOpcode() == ISD::FMUL)
7369 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7370 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7372 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7373 N10.getOperand(1)), N0);
7381 SDValue DAGCombiner::visitFSUB(SDNode *N) {
7382 SDValue N0 = N->getOperand(0);
7383 SDValue N1 = N->getOperand(1);
7384 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
7385 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
7386 EVT VT = N->getValueType(0);
7388 const TargetOptions &Options = DAG.getTarget().Options;
7392 if (SDValue FoldedVOp = SimplifyVBinOp(N))
7395 // fold (fsub c1, c2) -> c1-c2
7397 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0, N1);
7399 // fold (fsub A, (fneg B)) -> (fadd A, B)
7400 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
7401 return DAG.getNode(ISD::FADD, dl, VT, N0,
7402 GetNegatedExpression(N1, DAG, LegalOperations));
7404 // If 'unsafe math' is enabled, fold lots of things.
7405 if (Options.UnsafeFPMath) {
7407 if (N1CFP && N1CFP->getValueAPF().isZero())
7410 // (fsub 0, B) -> -B
7411 if (N0CFP && N0CFP->getValueAPF().isZero()) {
7412 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
7413 return GetNegatedExpression(N1, DAG, LegalOperations);
7414 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7415 return DAG.getNode(ISD::FNEG, dl, VT, N1);
7418 // (fsub x, x) -> 0.0
7420 return DAG.getConstantFP(0.0f, VT);
7422 // (fsub x, (fadd x, y)) -> (fneg y)
7423 // (fsub x, (fadd y, x)) -> (fneg y)
7424 if (N1.getOpcode() == ISD::FADD) {
7425 SDValue N10 = N1->getOperand(0);
7426 SDValue N11 = N1->getOperand(1);
7428 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI, &Options))
7429 return GetNegatedExpression(N11, DAG, LegalOperations);
7431 if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI, &Options))
7432 return GetNegatedExpression(N10, DAG, LegalOperations);
7436 if (LegalOperations && TLI.isOperationLegal(ISD::FMAD, VT)) {
7437 // Assume if there is an fmad instruction that it should be aggressively
7439 if (SDValue Fused = performFsubFmulCombines(ISD::FMAD, true, N, TLI, DAG))
7443 // FSUB -> FMA combines:
7444 if ((Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath) &&
7445 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
7446 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
7448 if (!TLI.isOperationLegal(ISD::FMAD, VT)) {
7449 // Don't form FMA if we are preferring FMAD.
7452 = performFsubFmulCombines(ISD::FMA,
7453 TLI.enableAggressiveFMAFusion(VT),
7459 // When FP_EXTEND nodes are free on the target, and there is an opportunity
7460 // to combine into FMA, arrange such nodes accordingly.
7461 if (TLI.isFPExtFree(VT)) {
7462 // fold (fsub (fpext (fmul x, y)), z)
7463 // -> (fma (fpext x), (fpext y), (fneg z))
7464 if (N0.getOpcode() == ISD::FP_EXTEND) {
7465 SDValue N00 = N0.getOperand(0);
7466 if (N00.getOpcode() == ISD::FMUL)
7467 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7468 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7470 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7472 DAG.getNode(ISD::FNEG, SDLoc(N), VT, N1));
7475 // fold (fsub x, (fpext (fmul y, z)))
7476 // -> (fma (fneg (fpext y)), (fpext z), x)
7477 // Note: Commutes FSUB operands.
7478 if (N1.getOpcode() == ISD::FP_EXTEND) {
7479 SDValue N10 = N1.getOperand(0);
7480 if (N10.getOpcode() == ISD::FMUL)
7481 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7482 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7483 DAG.getNode(ISD::FP_EXTEND, SDLoc(N),
7484 VT, N10.getOperand(0))),
7485 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7490 // fold (fsub (fpext (fneg (fmul, x, y))), z)
7491 // -> (fma (fneg (fpext x)), (fpext y), (fneg z))
7492 if (N0.getOpcode() == ISD::FP_EXTEND) {
7493 SDValue N00 = N0.getOperand(0);
7494 if (N00.getOpcode() == ISD::FNEG) {
7495 SDValue N000 = N00.getOperand(0);
7496 if (N000.getOpcode() == ISD::FMUL) {
7497 return DAG.getNode(ISD::FMA, dl, VT,
7498 DAG.getNode(ISD::FNEG, dl, VT,
7499 DAG.getNode(ISD::FP_EXTEND, SDLoc(N),
7500 VT, N000.getOperand(0))),
7501 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7502 N000.getOperand(1)),
7503 DAG.getNode(ISD::FNEG, dl, VT, N1));
7508 // fold (fsub (fneg (fpext (fmul, x, y))), z)
7509 // -> (fma (fneg (fpext x)), (fpext y), (fneg z))
7510 if (N0.getOpcode() == ISD::FNEG) {
7511 SDValue N00 = N0.getOperand(0);
7512 if (N00.getOpcode() == ISD::FP_EXTEND) {
7513 SDValue N000 = N00.getOperand(0);
7514 if (N000.getOpcode() == ISD::FMUL) {
7515 return DAG.getNode(ISD::FMA, dl, VT,
7516 DAG.getNode(ISD::FNEG, dl, VT,
7517 DAG.getNode(ISD::FP_EXTEND, SDLoc(N),
7518 VT, N000.getOperand(0))),
7519 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7520 N000.getOperand(1)),
7521 DAG.getNode(ISD::FNEG, dl, VT, N1));
7531 SDValue DAGCombiner::visitFMUL(SDNode *N) {
7532 SDValue N0 = N->getOperand(0);
7533 SDValue N1 = N->getOperand(1);
7534 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
7535 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
7536 EVT VT = N->getValueType(0);
7537 const TargetOptions &Options = DAG.getTarget().Options;
7540 if (VT.isVector()) {
7541 // This just handles C1 * C2 for vectors. Other vector folds are below.
7542 if (SDValue FoldedVOp = SimplifyVBinOp(N))
7546 // fold (fmul c1, c2) -> c1*c2
7548 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, N1);
7550 // canonicalize constant to RHS
7551 if (isConstantFPBuildVectorOrConstantFP(N0) &&
7552 !isConstantFPBuildVectorOrConstantFP(N1))
7553 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, N0);
7555 // fold (fmul A, 1.0) -> A
7556 if (N1CFP && N1CFP->isExactlyValue(1.0))
7559 if (Options.UnsafeFPMath) {
7560 // fold (fmul A, 0) -> 0
7561 if (N1CFP && N1CFP->getValueAPF().isZero())
7564 // fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
7565 if (N0.getOpcode() == ISD::FMUL) {
7566 // Fold scalars or any vector constants (not just splats).
7567 // This fold is done in general by InstCombine, but extra fmul insts
7568 // may have been generated during lowering.
7569 SDValue N00 = N0.getOperand(0);
7570 SDValue N01 = N0.getOperand(1);
7571 auto *BV1 = dyn_cast<BuildVectorSDNode>(N1);
7572 auto *BV00 = dyn_cast<BuildVectorSDNode>(N00);
7573 auto *BV01 = dyn_cast<BuildVectorSDNode>(N01);
7575 // Check 1: Make sure that the first operand of the inner multiply is NOT
7576 // a constant. Otherwise, we may induce infinite looping.
7577 if (!(isConstOrConstSplatFP(N00) || (BV00 && BV00->isConstant()))) {
7578 // Check 2: Make sure that the second operand of the inner multiply and
7579 // the second operand of the outer multiply are constants.
7580 if ((N1CFP && isConstOrConstSplatFP(N01)) ||
7581 (BV1 && BV01 && BV1->isConstant() && BV01->isConstant())) {
7583 SDValue MulConsts = DAG.getNode(ISD::FMUL, SL, VT, N01, N1);
7584 return DAG.getNode(ISD::FMUL, SL, VT, N00, MulConsts);
7589 // fold (fmul (fadd x, x), c) -> (fmul x, (fmul 2.0, c))
7590 // Undo the fmul 2.0, x -> fadd x, x transformation, since if it occurs
7591 // during an early run of DAGCombiner can prevent folding with fmuls
7592 // inserted during lowering.
7593 if (N0.getOpcode() == ISD::FADD && N0.getOperand(0) == N0.getOperand(1)) {
7595 const SDValue Two = DAG.getConstantFP(2.0, VT);
7596 SDValue MulConsts = DAG.getNode(ISD::FMUL, SL, VT, Two, N1);
7597 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0.getOperand(0), MulConsts);
7601 // fold (fmul X, 2.0) -> (fadd X, X)
7602 if (N1CFP && N1CFP->isExactlyValue(+2.0))
7603 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N0);
7605 // fold (fmul X, -1.0) -> (fneg X)
7606 if (N1CFP && N1CFP->isExactlyValue(-1.0))
7607 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7608 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
7610 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
7611 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
7612 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
7613 // Both can be negated for free, check to see if at least one is cheaper
7615 if (LHSNeg == 2 || RHSNeg == 2)
7616 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7617 GetNegatedExpression(N0, DAG, LegalOperations),
7618 GetNegatedExpression(N1, DAG, LegalOperations));
7625 SDValue DAGCombiner::visitFMA(SDNode *N) {
7626 SDValue N0 = N->getOperand(0);
7627 SDValue N1 = N->getOperand(1);
7628 SDValue N2 = N->getOperand(2);
7629 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7630 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7631 EVT VT = N->getValueType(0);
7633 const TargetOptions &Options = DAG.getTarget().Options;
7635 // Constant fold FMA.
7636 if (isa<ConstantFPSDNode>(N0) &&
7637 isa<ConstantFPSDNode>(N1) &&
7638 isa<ConstantFPSDNode>(N2)) {
7639 return DAG.getNode(ISD::FMA, dl, VT, N0, N1, N2);
7642 if (Options.UnsafeFPMath) {
7643 if (N0CFP && N0CFP->isZero())
7645 if (N1CFP && N1CFP->isZero())
7648 if (N0CFP && N0CFP->isExactlyValue(1.0))
7649 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2);
7650 if (N1CFP && N1CFP->isExactlyValue(1.0))
7651 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2);
7653 // Canonicalize (fma c, x, y) -> (fma x, c, y)
7654 if (N0CFP && !N1CFP)
7655 return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2);
7657 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
7658 if (Options.UnsafeFPMath && N1CFP &&
7659 N2.getOpcode() == ISD::FMUL &&
7660 N0 == N2.getOperand(0) &&
7661 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
7662 return DAG.getNode(ISD::FMUL, dl, VT, N0,
7663 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
7667 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
7668 if (Options.UnsafeFPMath &&
7669 N0.getOpcode() == ISD::FMUL && N1CFP &&
7670 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
7671 return DAG.getNode(ISD::FMA, dl, VT,
7673 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
7677 // (fma x, 1, y) -> (fadd x, y)
7678 // (fma x, -1, y) -> (fadd (fneg x), y)
7680 if (N1CFP->isExactlyValue(1.0))
7681 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
7683 if (N1CFP->isExactlyValue(-1.0) &&
7684 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
7685 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
7686 AddToWorklist(RHSNeg.getNode());
7687 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
7691 // (fma x, c, x) -> (fmul x, (c+1))
7692 if (Options.UnsafeFPMath && N1CFP && N0 == N2)
7693 return DAG.getNode(ISD::FMUL, dl, VT, N0,
7694 DAG.getNode(ISD::FADD, dl, VT,
7695 N1, DAG.getConstantFP(1.0, VT)));
7697 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
7698 if (Options.UnsafeFPMath && N1CFP &&
7699 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0)
7700 return DAG.getNode(ISD::FMUL, dl, VT, N0,
7701 DAG.getNode(ISD::FADD, dl, VT,
7702 N1, DAG.getConstantFP(-1.0, VT)));
7708 SDValue DAGCombiner::visitFDIV(SDNode *N) {
7709 SDValue N0 = N->getOperand(0);
7710 SDValue N1 = N->getOperand(1);
7711 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7712 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7713 EVT VT = N->getValueType(0);
7715 const TargetOptions &Options = DAG.getTarget().Options;
7719 if (SDValue FoldedVOp = SimplifyVBinOp(N))
7722 // fold (fdiv c1, c2) -> c1/c2
7724 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1);
7726 if (Options.UnsafeFPMath) {
7727 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
7729 // Compute the reciprocal 1.0 / c2.
7730 APFloat N1APF = N1CFP->getValueAPF();
7731 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
7732 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
7733 // Only do the transform if the reciprocal is a legal fp immediate that
7734 // isn't too nasty (eg NaN, denormal, ...).
7735 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
7736 (!LegalOperations ||
7737 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
7738 // backend)... we should handle this gracefully after Legalize.
7739 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
7740 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
7741 TLI.isFPImmLegal(Recip, VT)))
7742 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0,
7743 DAG.getConstantFP(Recip, VT));
7746 // If this FDIV is part of a reciprocal square root, it may be folded
7747 // into a target-specific square root estimate instruction.
7748 if (N1.getOpcode() == ISD::FSQRT) {
7749 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0))) {
7750 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7752 } else if (N1.getOpcode() == ISD::FP_EXTEND &&
7753 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
7754 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0).getOperand(0))) {
7755 RV = DAG.getNode(ISD::FP_EXTEND, SDLoc(N1), VT, RV);
7756 AddToWorklist(RV.getNode());
7757 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7759 } else if (N1.getOpcode() == ISD::FP_ROUND &&
7760 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
7761 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0).getOperand(0))) {
7762 RV = DAG.getNode(ISD::FP_ROUND, SDLoc(N1), VT, RV, N1.getOperand(1));
7763 AddToWorklist(RV.getNode());
7764 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7766 } else if (N1.getOpcode() == ISD::FMUL) {
7767 // Look through an FMUL. Even though this won't remove the FDIV directly,
7768 // it's still worthwhile to get rid of the FSQRT if possible.
7771 if (N1.getOperand(0).getOpcode() == ISD::FSQRT) {
7772 SqrtOp = N1.getOperand(0);
7773 OtherOp = N1.getOperand(1);
7774 } else if (N1.getOperand(1).getOpcode() == ISD::FSQRT) {
7775 SqrtOp = N1.getOperand(1);
7776 OtherOp = N1.getOperand(0);
7778 if (SqrtOp.getNode()) {
7779 // We found a FSQRT, so try to make this fold:
7780 // x / (y * sqrt(z)) -> x * (rsqrt(z) / y)
7781 if (SDValue RV = BuildRsqrtEstimate(SqrtOp.getOperand(0))) {
7782 RV = DAG.getNode(ISD::FDIV, SDLoc(N1), VT, RV, OtherOp);
7783 AddToWorklist(RV.getNode());
7784 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7789 // Fold into a reciprocal estimate and multiply instead of a real divide.
7790 if (SDValue RV = BuildReciprocalEstimate(N1)) {
7791 AddToWorklist(RV.getNode());
7792 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7796 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
7797 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
7798 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
7799 // Both can be negated for free, check to see if at least one is cheaper
7801 if (LHSNeg == 2 || RHSNeg == 2)
7802 return DAG.getNode(ISD::FDIV, SDLoc(N), VT,
7803 GetNegatedExpression(N0, DAG, LegalOperations),
7804 GetNegatedExpression(N1, DAG, LegalOperations));
7808 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
7810 // E.g., (a / D; b / D;) -> (recip = 1.0 / D; a * recip; b * recip)
7811 // Notice that this is not always beneficial. One reason is different target
7812 // may have different costs for FDIV and FMUL, so sometimes the cost of two
7813 // FDIVs may be lower than the cost of one FDIV and two FMULs. Another reason
7814 // is the critical path is increased from "one FDIV" to "one FDIV + one FMUL".
7815 if (Options.UnsafeFPMath) {
7816 // Skip if current node is a reciprocal.
7817 if (N0CFP && N0CFP->isExactlyValue(1.0))
7820 SmallVector<SDNode *, 4> Users;
7821 // Find all FDIV users of the same divisor.
7822 for (SDNode::use_iterator UI = N1.getNode()->use_begin(),
7823 UE = N1.getNode()->use_end();
7825 SDNode *User = UI.getUse().getUser();
7826 if (User->getOpcode() == ISD::FDIV && User->getOperand(1) == N1)
7827 Users.push_back(User);
7830 if (TLI.combineRepeatedFPDivisors(Users.size())) {
7831 SDValue FPOne = DAG.getConstantFP(1.0, VT); // floating point 1.0
7832 SDValue Reciprocal = DAG.getNode(ISD::FDIV, SDLoc(N), VT, FPOne, N1);
7834 // Dividend / Divisor -> Dividend * Reciprocal
7835 for (auto I = Users.begin(), E = Users.end(); I != E; ++I) {
7836 if ((*I)->getOperand(0) != FPOne) {
7837 SDValue NewNode = DAG.getNode(ISD::FMUL, SDLoc(*I), VT,
7838 (*I)->getOperand(0), Reciprocal);
7839 DAG.ReplaceAllUsesWith(*I, NewNode.getNode());
7849 SDValue DAGCombiner::visitFREM(SDNode *N) {
7850 SDValue N0 = N->getOperand(0);
7851 SDValue N1 = N->getOperand(1);
7852 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7853 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7854 EVT VT = N->getValueType(0);
7856 // fold (frem c1, c2) -> fmod(c1,c2)
7858 return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1);
7863 SDValue DAGCombiner::visitFSQRT(SDNode *N) {
7864 if (DAG.getTarget().Options.UnsafeFPMath &&
7865 !TLI.isFsqrtCheap()) {
7866 // Compute this as X * (1/sqrt(X)) = X * (X ** -0.5)
7867 if (SDValue RV = BuildRsqrtEstimate(N->getOperand(0))) {
7868 EVT VT = RV.getValueType();
7869 RV = DAG.getNode(ISD::FMUL, SDLoc(N), VT, N->getOperand(0), RV);
7870 AddToWorklist(RV.getNode());
7872 // Unfortunately, RV is now NaN if the input was exactly 0.
7873 // Select out this case and force the answer to 0.
7874 SDValue Zero = DAG.getConstantFP(0.0, VT);
7876 DAG.getSetCC(SDLoc(N), TLI.getSetCCResultType(*DAG.getContext(), VT),
7877 N->getOperand(0), Zero, ISD::SETEQ);
7878 AddToWorklist(ZeroCmp.getNode());
7879 AddToWorklist(RV.getNode());
7881 RV = DAG.getNode(VT.isVector() ? ISD::VSELECT : ISD::SELECT,
7882 SDLoc(N), VT, ZeroCmp, Zero, RV);
7889 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
7890 SDValue N0 = N->getOperand(0);
7891 SDValue N1 = N->getOperand(1);
7892 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7893 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7894 EVT VT = N->getValueType(0);
7896 if (N0CFP && N1CFP) // Constant fold
7897 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1);
7900 const APFloat& V = N1CFP->getValueAPF();
7901 // copysign(x, c1) -> fabs(x) iff ispos(c1)
7902 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
7903 if (!V.isNegative()) {
7904 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
7905 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7907 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7908 return DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7909 DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0));
7913 // copysign(fabs(x), y) -> copysign(x, y)
7914 // copysign(fneg(x), y) -> copysign(x, y)
7915 // copysign(copysign(x,z), y) -> copysign(x, y)
7916 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
7917 N0.getOpcode() == ISD::FCOPYSIGN)
7918 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7919 N0.getOperand(0), N1);
7921 // copysign(x, abs(y)) -> abs(x)
7922 if (N1.getOpcode() == ISD::FABS)
7923 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7925 // copysign(x, copysign(y,z)) -> copysign(x, z)
7926 if (N1.getOpcode() == ISD::FCOPYSIGN)
7927 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7928 N0, N1.getOperand(1));
7930 // copysign(x, fp_extend(y)) -> copysign(x, y)
7931 // copysign(x, fp_round(y)) -> copysign(x, y)
7932 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
7933 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7934 N0, N1.getOperand(0));
7939 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
7940 SDValue N0 = N->getOperand(0);
7941 EVT VT = N->getValueType(0);
7942 EVT OpVT = N0.getValueType();
7944 // fold (sint_to_fp c1) -> c1fp
7945 if (isConstantIntBuildVectorOrConstantInt(N0) &&
7946 // ...but only if the target supports immediate floating-point values
7947 (!LegalOperations ||
7948 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
7949 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
7951 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
7952 // but UINT_TO_FP is legal on this target, try to convert.
7953 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
7954 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
7955 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
7956 if (DAG.SignBitIsZero(N0))
7957 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
7960 // The next optimizations are desirable only if SELECT_CC can be lowered.
7961 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
7962 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7963 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
7965 (!LegalOperations ||
7966 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7968 { N0.getOperand(0), N0.getOperand(1),
7969 DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT),
7971 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7974 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
7975 // (select_cc x, y, 1.0, 0.0,, cc)
7976 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
7977 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
7978 (!LegalOperations ||
7979 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7981 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
7982 DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT),
7983 N0.getOperand(0).getOperand(2) };
7984 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7991 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
7992 SDValue N0 = N->getOperand(0);
7993 EVT VT = N->getValueType(0);
7994 EVT OpVT = N0.getValueType();
7996 // fold (uint_to_fp c1) -> c1fp
7997 if (isConstantIntBuildVectorOrConstantInt(N0) &&
7998 // ...but only if the target supports immediate floating-point values
7999 (!LegalOperations ||
8000 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
8001 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
8003 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
8004 // but SINT_TO_FP is legal on this target, try to convert.
8005 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
8006 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
8007 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
8008 if (DAG.SignBitIsZero(N0))
8009 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
8012 // The next optimizations are desirable only if SELECT_CC can be lowered.
8013 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
8014 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
8016 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
8017 (!LegalOperations ||
8018 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
8020 { N0.getOperand(0), N0.getOperand(1),
8021 DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT),
8023 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
8030 // Fold (fp_to_{s/u}int ({s/u}int_to_fpx)) -> zext x, sext x, trunc x, or x
8031 static SDValue FoldIntToFPToInt(SDNode *N, SelectionDAG &DAG) {
8032 SDValue N0 = N->getOperand(0);
8033 EVT VT = N->getValueType(0);
8035 if (N0.getOpcode() != ISD::UINT_TO_FP && N0.getOpcode() != ISD::SINT_TO_FP)
8038 SDValue Src = N0.getOperand(0);
8039 EVT SrcVT = Src.getValueType();
8040 bool IsInputSigned = N0.getOpcode() == ISD::SINT_TO_FP;
8041 bool IsOutputSigned = N->getOpcode() == ISD::FP_TO_SINT;
8043 // We can safely assume the conversion won't overflow the output range,
8044 // because (for example) (uint8_t)18293.f is undefined behavior.
8046 // Since we can assume the conversion won't overflow, our decision as to
8047 // whether the input will fit in the float should depend on the minimum
8048 // of the input range and output range.
8050 // This means this is also safe for a signed input and unsigned output, since
8051 // a negative input would lead to undefined behavior.
8052 unsigned InputSize = (int)SrcVT.getScalarSizeInBits() - IsInputSigned;
8053 unsigned OutputSize = (int)VT.getScalarSizeInBits() - IsOutputSigned;
8054 unsigned ActualSize = std::min(InputSize, OutputSize);
8055 const fltSemantics &sem = DAG.EVTToAPFloatSemantics(N0.getValueType());
8057 // We can only fold away the float conversion if the input range can be
8058 // represented exactly in the float range.
8059 if (APFloat::semanticsPrecision(sem) >= ActualSize) {
8060 if (VT.getScalarSizeInBits() > SrcVT.getScalarSizeInBits()) {
8061 unsigned ExtOp = IsInputSigned && IsOutputSigned ? ISD::SIGN_EXTEND
8063 return DAG.getNode(ExtOp, SDLoc(N), VT, Src);
8065 if (VT.getScalarSizeInBits() < SrcVT.getScalarSizeInBits())
8066 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Src);
8069 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Src);
8074 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
8075 SDValue N0 = N->getOperand(0);
8076 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8077 EVT VT = N->getValueType(0);
8079 // fold (fp_to_sint c1fp) -> c1
8081 return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0);
8083 return FoldIntToFPToInt(N, DAG);
8086 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
8087 SDValue N0 = N->getOperand(0);
8088 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8089 EVT VT = N->getValueType(0);
8091 // fold (fp_to_uint c1fp) -> c1
8093 return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0);
8095 return FoldIntToFPToInt(N, DAG);
8098 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
8099 SDValue N0 = N->getOperand(0);
8100 SDValue N1 = N->getOperand(1);
8101 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8102 EVT VT = N->getValueType(0);
8104 // fold (fp_round c1fp) -> c1fp
8106 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1);
8108 // fold (fp_round (fp_extend x)) -> x
8109 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
8110 return N0.getOperand(0);
8112 // fold (fp_round (fp_round x)) -> (fp_round x)
8113 if (N0.getOpcode() == ISD::FP_ROUND) {
8114 const bool NIsTrunc = N->getConstantOperandVal(1) == 1;
8115 const bool N0IsTrunc = N0.getNode()->getConstantOperandVal(1) == 1;
8116 // If the first fp_round isn't a value preserving truncation, it might
8117 // introduce a tie in the second fp_round, that wouldn't occur in the
8118 // single-step fp_round we want to fold to.
8119 // In other words, double rounding isn't the same as rounding.
8120 // Also, this is a value preserving truncation iff both fp_round's are.
8121 if (DAG.getTarget().Options.UnsafeFPMath || N0IsTrunc)
8122 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0.getOperand(0),
8123 DAG.getIntPtrConstant(NIsTrunc && N0IsTrunc));
8126 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
8127 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
8128 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT,
8129 N0.getOperand(0), N1);
8130 AddToWorklist(Tmp.getNode());
8131 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
8132 Tmp, N0.getOperand(1));
8138 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
8139 SDValue N0 = N->getOperand(0);
8140 EVT VT = N->getValueType(0);
8141 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
8142 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8144 // fold (fp_round_inreg c1fp) -> c1fp
8145 if (N0CFP && isTypeLegal(EVT)) {
8146 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
8147 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, Round);
8153 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
8154 SDValue N0 = N->getOperand(0);
8155 EVT VT = N->getValueType(0);
8157 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
8158 if (N->hasOneUse() &&
8159 N->use_begin()->getOpcode() == ISD::FP_ROUND)
8162 // fold (fp_extend c1fp) -> c1fp
8163 if (isConstantFPBuildVectorOrConstantFP(N0))
8164 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0);
8166 // fold (fp_extend (fp16_to_fp op)) -> (fp16_to_fp op)
8167 if (N0.getOpcode() == ISD::FP16_TO_FP &&
8168 TLI.getOperationAction(ISD::FP16_TO_FP, VT) == TargetLowering::Legal)
8169 return DAG.getNode(ISD::FP16_TO_FP, SDLoc(N), VT, N0.getOperand(0));
8171 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
8173 if (N0.getOpcode() == ISD::FP_ROUND
8174 && N0.getNode()->getConstantOperandVal(1) == 1) {
8175 SDValue In = N0.getOperand(0);
8176 if (In.getValueType() == VT) return In;
8177 if (VT.bitsLT(In.getValueType()))
8178 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT,
8179 In, N0.getOperand(1));
8180 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In);
8183 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
8184 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
8185 TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) {
8186 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
8187 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
8189 LN0->getBasePtr(), N0.getValueType(),
8190 LN0->getMemOperand());
8191 CombineTo(N, ExtLoad);
8192 CombineTo(N0.getNode(),
8193 DAG.getNode(ISD::FP_ROUND, SDLoc(N0),
8194 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
8195 ExtLoad.getValue(1));
8196 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8202 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
8203 SDValue N0 = N->getOperand(0);
8204 EVT VT = N->getValueType(0);
8206 // fold (fceil c1) -> fceil(c1)
8207 if (isConstantFPBuildVectorOrConstantFP(N0))
8208 return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0);
8213 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
8214 SDValue N0 = N->getOperand(0);
8215 EVT VT = N->getValueType(0);
8217 // fold (ftrunc c1) -> ftrunc(c1)
8218 if (isConstantFPBuildVectorOrConstantFP(N0))
8219 return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0);
8224 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
8225 SDValue N0 = N->getOperand(0);
8226 EVT VT = N->getValueType(0);
8228 // fold (ffloor c1) -> ffloor(c1)
8229 if (isConstantFPBuildVectorOrConstantFP(N0))
8230 return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0);
8235 // FIXME: FNEG and FABS have a lot in common; refactor.
8236 SDValue DAGCombiner::visitFNEG(SDNode *N) {
8237 SDValue N0 = N->getOperand(0);
8238 EVT VT = N->getValueType(0);
8240 // Constant fold FNEG.
8241 if (isConstantFPBuildVectorOrConstantFP(N0))
8242 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
8244 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
8245 &DAG.getTarget().Options))
8246 return GetNegatedExpression(N0, DAG, LegalOperations);
8248 // Transform fneg(bitconvert(x)) -> bitconvert(x ^ sign) to avoid loading
8249 // constant pool values.
8250 if (!TLI.isFNegFree(VT) &&
8251 N0.getOpcode() == ISD::BITCAST &&
8252 N0.getNode()->hasOneUse()) {
8253 SDValue Int = N0.getOperand(0);
8254 EVT IntVT = Int.getValueType();
8255 if (IntVT.isInteger() && !IntVT.isVector()) {
8257 if (N0.getValueType().isVector()) {
8258 // For a vector, get a mask such as 0x80... per scalar element
8260 SignMask = APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
8261 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
8263 // For a scalar, just generate 0x80...
8264 SignMask = APInt::getSignBit(IntVT.getSizeInBits());
8266 Int = DAG.getNode(ISD::XOR, SDLoc(N0), IntVT, Int,
8267 DAG.getConstant(SignMask, IntVT));
8268 AddToWorklist(Int.getNode());
8269 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Int);
8273 // (fneg (fmul c, x)) -> (fmul -c, x)
8274 if (N0.getOpcode() == ISD::FMUL) {
8275 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
8277 APFloat CVal = CFP1->getValueAPF();
8279 if (Level >= AfterLegalizeDAG &&
8280 (TLI.isFPImmLegal(CVal, N->getValueType(0)) ||
8281 TLI.isOperationLegal(ISD::ConstantFP, N->getValueType(0))))
8283 ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
8284 DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0.getOperand(1)));
8291 SDValue DAGCombiner::visitFMINNUM(SDNode *N) {
8292 SDValue N0 = N->getOperand(0);
8293 SDValue N1 = N->getOperand(1);
8294 const ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8295 const ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8297 if (N0CFP && N1CFP) {
8298 const APFloat &C0 = N0CFP->getValueAPF();
8299 const APFloat &C1 = N1CFP->getValueAPF();
8300 return DAG.getConstantFP(minnum(C0, C1), N->getValueType(0));
8304 EVT VT = N->getValueType(0);
8305 // Canonicalize to constant on RHS.
8306 return DAG.getNode(ISD::FMINNUM, SDLoc(N), VT, N1, N0);
8312 SDValue DAGCombiner::visitFMAXNUM(SDNode *N) {
8313 SDValue N0 = N->getOperand(0);
8314 SDValue N1 = N->getOperand(1);
8315 const ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8316 const ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8318 if (N0CFP && N1CFP) {
8319 const APFloat &C0 = N0CFP->getValueAPF();
8320 const APFloat &C1 = N1CFP->getValueAPF();
8321 return DAG.getConstantFP(maxnum(C0, C1), N->getValueType(0));
8325 EVT VT = N->getValueType(0);
8326 // Canonicalize to constant on RHS.
8327 return DAG.getNode(ISD::FMAXNUM, SDLoc(N), VT, N1, N0);
8333 SDValue DAGCombiner::visitFABS(SDNode *N) {
8334 SDValue N0 = N->getOperand(0);
8335 EVT VT = N->getValueType(0);
8337 // fold (fabs c1) -> fabs(c1)
8338 if (isConstantFPBuildVectorOrConstantFP(N0))
8339 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
8341 // fold (fabs (fabs x)) -> (fabs x)
8342 if (N0.getOpcode() == ISD::FABS)
8343 return N->getOperand(0);
8345 // fold (fabs (fneg x)) -> (fabs x)
8346 // fold (fabs (fcopysign x, y)) -> (fabs x)
8347 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
8348 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0));
8350 // Transform fabs(bitconvert(x)) -> bitconvert(x & ~sign) to avoid loading
8351 // constant pool values.
8352 if (!TLI.isFAbsFree(VT) &&
8353 N0.getOpcode() == ISD::BITCAST &&
8354 N0.getNode()->hasOneUse()) {
8355 SDValue Int = N0.getOperand(0);
8356 EVT IntVT = Int.getValueType();
8357 if (IntVT.isInteger() && !IntVT.isVector()) {
8359 if (N0.getValueType().isVector()) {
8360 // For a vector, get a mask such as 0x7f... per scalar element
8362 SignMask = ~APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
8363 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
8365 // For a scalar, just generate 0x7f...
8366 SignMask = ~APInt::getSignBit(IntVT.getSizeInBits());
8368 Int = DAG.getNode(ISD::AND, SDLoc(N0), IntVT, Int,
8369 DAG.getConstant(SignMask, IntVT));
8370 AddToWorklist(Int.getNode());
8371 return DAG.getNode(ISD::BITCAST, SDLoc(N), N->getValueType(0), Int);
8378 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
8379 SDValue Chain = N->getOperand(0);
8380 SDValue N1 = N->getOperand(1);
8381 SDValue N2 = N->getOperand(2);
8383 // If N is a constant we could fold this into a fallthrough or unconditional
8384 // branch. However that doesn't happen very often in normal code, because
8385 // Instcombine/SimplifyCFG should have handled the available opportunities.
8386 // If we did this folding here, it would be necessary to update the
8387 // MachineBasicBlock CFG, which is awkward.
8389 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
8391 if (N1.getOpcode() == ISD::SETCC &&
8392 TLI.isOperationLegalOrCustom(ISD::BR_CC,
8393 N1.getOperand(0).getValueType())) {
8394 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
8395 Chain, N1.getOperand(2),
8396 N1.getOperand(0), N1.getOperand(1), N2);
8399 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
8400 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
8401 (N1.getOperand(0).hasOneUse() &&
8402 N1.getOperand(0).getOpcode() == ISD::SRL))) {
8403 SDNode *Trunc = nullptr;
8404 if (N1.getOpcode() == ISD::TRUNCATE) {
8405 // Look pass the truncate.
8406 Trunc = N1.getNode();
8407 N1 = N1.getOperand(0);
8410 // Match this pattern so that we can generate simpler code:
8413 // %b = and i32 %a, 2
8414 // %c = srl i32 %b, 1
8415 // brcond i32 %c ...
8420 // %b = and i32 %a, 2
8421 // %c = setcc eq %b, 0
8424 // This applies only when the AND constant value has one bit set and the
8425 // SRL constant is equal to the log2 of the AND constant. The back-end is
8426 // smart enough to convert the result into a TEST/JMP sequence.
8427 SDValue Op0 = N1.getOperand(0);
8428 SDValue Op1 = N1.getOperand(1);
8430 if (Op0.getOpcode() == ISD::AND &&
8431 Op1.getOpcode() == ISD::Constant) {
8432 SDValue AndOp1 = Op0.getOperand(1);
8434 if (AndOp1.getOpcode() == ISD::Constant) {
8435 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
8437 if (AndConst.isPowerOf2() &&
8438 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
8440 DAG.getSetCC(SDLoc(N),
8441 getSetCCResultType(Op0.getValueType()),
8442 Op0, DAG.getConstant(0, Op0.getValueType()),
8445 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, SDLoc(N),
8446 MVT::Other, Chain, SetCC, N2);
8447 // Don't add the new BRCond into the worklist or else SimplifySelectCC
8448 // will convert it back to (X & C1) >> C2.
8449 CombineTo(N, NewBRCond, false);
8450 // Truncate is dead.
8452 deleteAndRecombine(Trunc);
8453 // Replace the uses of SRL with SETCC
8454 WorklistRemover DeadNodes(*this);
8455 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
8456 deleteAndRecombine(N1.getNode());
8457 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8463 // Restore N1 if the above transformation doesn't match.
8464 N1 = N->getOperand(1);
8467 // Transform br(xor(x, y)) -> br(x != y)
8468 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
8469 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
8470 SDNode *TheXor = N1.getNode();
8471 SDValue Op0 = TheXor->getOperand(0);
8472 SDValue Op1 = TheXor->getOperand(1);
8473 if (Op0.getOpcode() == Op1.getOpcode()) {
8474 // Avoid missing important xor optimizations.
8475 SDValue Tmp = visitXOR(TheXor);
8476 if (Tmp.getNode()) {
8477 if (Tmp.getNode() != TheXor) {
8478 DEBUG(dbgs() << "\nReplacing.8 ";
8480 dbgs() << "\nWith: ";
8481 Tmp.getNode()->dump(&DAG);
8483 WorklistRemover DeadNodes(*this);
8484 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
8485 deleteAndRecombine(TheXor);
8486 return DAG.getNode(ISD::BRCOND, SDLoc(N),
8487 MVT::Other, Chain, Tmp, N2);
8490 // visitXOR has changed XOR's operands or replaced the XOR completely,
8492 return SDValue(N, 0);
8496 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
8498 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
8499 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
8500 Op0.getOpcode() == ISD::XOR) {
8501 TheXor = Op0.getNode();
8505 EVT SetCCVT = N1.getValueType();
8507 SetCCVT = getSetCCResultType(SetCCVT);
8508 SDValue SetCC = DAG.getSetCC(SDLoc(TheXor),
8511 Equal ? ISD::SETEQ : ISD::SETNE);
8512 // Replace the uses of XOR with SETCC
8513 WorklistRemover DeadNodes(*this);
8514 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
8515 deleteAndRecombine(N1.getNode());
8516 return DAG.getNode(ISD::BRCOND, SDLoc(N),
8517 MVT::Other, Chain, SetCC, N2);
8524 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
8526 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
8527 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
8528 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
8530 // If N is a constant we could fold this into a fallthrough or unconditional
8531 // branch. However that doesn't happen very often in normal code, because
8532 // Instcombine/SimplifyCFG should have handled the available opportunities.
8533 // If we did this folding here, it would be necessary to update the
8534 // MachineBasicBlock CFG, which is awkward.
8536 // Use SimplifySetCC to simplify SETCC's.
8537 SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()),
8538 CondLHS, CondRHS, CC->get(), SDLoc(N),
8540 if (Simp.getNode()) AddToWorklist(Simp.getNode());
8542 // fold to a simpler setcc
8543 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
8544 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
8545 N->getOperand(0), Simp.getOperand(2),
8546 Simp.getOperand(0), Simp.getOperand(1),
8552 /// Return true if 'Use' is a load or a store that uses N as its base pointer
8553 /// and that N may be folded in the load / store addressing mode.
8554 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
8556 const TargetLowering &TLI) {
8558 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
8559 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
8561 VT = Use->getValueType(0);
8562 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
8563 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
8565 VT = ST->getValue().getValueType();
8569 TargetLowering::AddrMode AM;
8570 if (N->getOpcode() == ISD::ADD) {
8571 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
8574 AM.BaseOffs = Offset->getSExtValue();
8578 } else if (N->getOpcode() == ISD::SUB) {
8579 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
8582 AM.BaseOffs = -Offset->getSExtValue();
8589 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
8592 /// Try turning a load/store into a pre-indexed load/store when the base
8593 /// pointer is an add or subtract and it has other uses besides the load/store.
8594 /// After the transformation, the new indexed load/store has effectively folded
8595 /// the add/subtract in and all of its other uses are redirected to the
8597 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
8598 if (Level < AfterLegalizeDAG)
8604 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8605 if (LD->isIndexed())
8607 VT = LD->getMemoryVT();
8608 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
8609 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
8611 Ptr = LD->getBasePtr();
8612 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8613 if (ST->isIndexed())
8615 VT = ST->getMemoryVT();
8616 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
8617 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
8619 Ptr = ST->getBasePtr();
8625 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
8626 // out. There is no reason to make this a preinc/predec.
8627 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
8628 Ptr.getNode()->hasOneUse())
8631 // Ask the target to do addressing mode selection.
8634 ISD::MemIndexedMode AM = ISD::UNINDEXED;
8635 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
8638 // Backends without true r+i pre-indexed forms may need to pass a
8639 // constant base with a variable offset so that constant coercion
8640 // will work with the patterns in canonical form.
8641 bool Swapped = false;
8642 if (isa<ConstantSDNode>(BasePtr)) {
8643 std::swap(BasePtr, Offset);
8647 // Don't create a indexed load / store with zero offset.
8648 if (isa<ConstantSDNode>(Offset) &&
8649 cast<ConstantSDNode>(Offset)->isNullValue())
8652 // Try turning it into a pre-indexed load / store except when:
8653 // 1) The new base ptr is a frame index.
8654 // 2) If N is a store and the new base ptr is either the same as or is a
8655 // predecessor of the value being stored.
8656 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
8657 // that would create a cycle.
8658 // 4) All uses are load / store ops that use it as old base ptr.
8660 // Check #1. Preinc'ing a frame index would require copying the stack pointer
8661 // (plus the implicit offset) to a register to preinc anyway.
8662 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
8667 SDValue Val = cast<StoreSDNode>(N)->getValue();
8668 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
8672 // If the offset is a constant, there may be other adds of constants that
8673 // can be folded with this one. We should do this to avoid having to keep
8674 // a copy of the original base pointer.
8675 SmallVector<SDNode *, 16> OtherUses;
8676 if (isa<ConstantSDNode>(Offset))
8677 for (SDNode *Use : BasePtr.getNode()->uses()) {
8678 if (Use == Ptr.getNode())
8681 if (Use->isPredecessorOf(N))
8684 if (Use->getOpcode() != ISD::ADD && Use->getOpcode() != ISD::SUB) {
8689 SDValue Op0 = Use->getOperand(0), Op1 = Use->getOperand(1);
8690 if (Op1.getNode() == BasePtr.getNode())
8691 std::swap(Op0, Op1);
8692 assert(Op0.getNode() == BasePtr.getNode() &&
8693 "Use of ADD/SUB but not an operand");
8695 if (!isa<ConstantSDNode>(Op1)) {
8700 // FIXME: In some cases, we can be smarter about this.
8701 if (Op1.getValueType() != Offset.getValueType()) {
8706 OtherUses.push_back(Use);
8710 std::swap(BasePtr, Offset);
8712 // Now check for #3 and #4.
8713 bool RealUse = false;
8715 // Caches for hasPredecessorHelper
8716 SmallPtrSet<const SDNode *, 32> Visited;
8717 SmallVector<const SDNode *, 16> Worklist;
8719 for (SDNode *Use : Ptr.getNode()->uses()) {
8722 if (N->hasPredecessorHelper(Use, Visited, Worklist))
8725 // If Ptr may be folded in addressing mode of other use, then it's
8726 // not profitable to do this transformation.
8727 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
8736 Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
8737 BasePtr, Offset, AM);
8739 Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
8740 BasePtr, Offset, AM);
8743 DEBUG(dbgs() << "\nReplacing.4 ";
8745 dbgs() << "\nWith: ";
8746 Result.getNode()->dump(&DAG);
8748 WorklistRemover DeadNodes(*this);
8750 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
8751 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
8753 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
8756 // Finally, since the node is now dead, remove it from the graph.
8757 deleteAndRecombine(N);
8760 std::swap(BasePtr, Offset);
8762 // Replace other uses of BasePtr that can be updated to use Ptr
8763 for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) {
8764 unsigned OffsetIdx = 1;
8765 if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode())
8767 assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() ==
8768 BasePtr.getNode() && "Expected BasePtr operand");
8770 // We need to replace ptr0 in the following expression:
8771 // x0 * offset0 + y0 * ptr0 = t0
8773 // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store)
8775 // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the
8776 // indexed load/store and the expresion that needs to be re-written.
8778 // Therefore, we have:
8779 // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1
8781 ConstantSDNode *CN =
8782 cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx));
8784 APInt Offset0 = CN->getAPIntValue();
8785 APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue();
8787 X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1;
8788 Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1;
8789 X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1;
8790 Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1;
8792 unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD;
8794 APInt CNV = Offset0;
8795 if (X0 < 0) CNV = -CNV;
8796 if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1;
8797 else CNV = CNV - Offset1;
8799 // We can now generate the new expression.
8800 SDValue NewOp1 = DAG.getConstant(CNV, CN->getValueType(0));
8801 SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0);
8803 SDValue NewUse = DAG.getNode(Opcode,
8804 SDLoc(OtherUses[i]),
8805 OtherUses[i]->getValueType(0), NewOp1, NewOp2);
8806 DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse);
8807 deleteAndRecombine(OtherUses[i]);
8810 // Replace the uses of Ptr with uses of the updated base value.
8811 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
8812 deleteAndRecombine(Ptr.getNode());
8817 /// Try to combine a load/store with a add/sub of the base pointer node into a
8818 /// post-indexed load/store. The transformation folded the add/subtract into the
8819 /// new indexed load/store effectively and all of its uses are redirected to the
8821 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
8822 if (Level < AfterLegalizeDAG)
8828 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8829 if (LD->isIndexed())
8831 VT = LD->getMemoryVT();
8832 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
8833 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
8835 Ptr = LD->getBasePtr();
8836 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8837 if (ST->isIndexed())
8839 VT = ST->getMemoryVT();
8840 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
8841 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
8843 Ptr = ST->getBasePtr();
8849 if (Ptr.getNode()->hasOneUse())
8852 for (SDNode *Op : Ptr.getNode()->uses()) {
8854 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
8859 ISD::MemIndexedMode AM = ISD::UNINDEXED;
8860 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
8861 // Don't create a indexed load / store with zero offset.
8862 if (isa<ConstantSDNode>(Offset) &&
8863 cast<ConstantSDNode>(Offset)->isNullValue())
8866 // Try turning it into a post-indexed load / store except when
8867 // 1) All uses are load / store ops that use it as base ptr (and
8868 // it may be folded as addressing mmode).
8869 // 2) Op must be independent of N, i.e. Op is neither a predecessor
8870 // nor a successor of N. Otherwise, if Op is folded that would
8873 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
8877 bool TryNext = false;
8878 for (SDNode *Use : BasePtr.getNode()->uses()) {
8879 if (Use == Ptr.getNode())
8882 // If all the uses are load / store addresses, then don't do the
8884 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
8885 bool RealUse = false;
8886 for (SDNode *UseUse : Use->uses()) {
8887 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
8902 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
8903 SDValue Result = isLoad
8904 ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
8905 BasePtr, Offset, AM)
8906 : DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
8907 BasePtr, Offset, AM);
8910 DEBUG(dbgs() << "\nReplacing.5 ";
8912 dbgs() << "\nWith: ";
8913 Result.getNode()->dump(&DAG);
8915 WorklistRemover DeadNodes(*this);
8917 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
8918 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
8920 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
8923 // Finally, since the node is now dead, remove it from the graph.
8924 deleteAndRecombine(N);
8926 // Replace the uses of Use with uses of the updated base value.
8927 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
8928 Result.getValue(isLoad ? 1 : 0));
8929 deleteAndRecombine(Op);
8938 /// \brief Return the base-pointer arithmetic from an indexed \p LD.
8939 SDValue DAGCombiner::SplitIndexingFromLoad(LoadSDNode *LD) {
8940 ISD::MemIndexedMode AM = LD->getAddressingMode();
8941 assert(AM != ISD::UNINDEXED);
8942 SDValue BP = LD->getOperand(1);
8943 SDValue Inc = LD->getOperand(2);
8945 // Some backends use TargetConstants for load offsets, but don't expect
8946 // TargetConstants in general ADD nodes. We can convert these constants into
8947 // regular Constants (if the constant is not opaque).
8948 assert((Inc.getOpcode() != ISD::TargetConstant ||
8949 !cast<ConstantSDNode>(Inc)->isOpaque()) &&
8950 "Cannot split out indexing using opaque target constants");
8951 if (Inc.getOpcode() == ISD::TargetConstant) {
8952 ConstantSDNode *ConstInc = cast<ConstantSDNode>(Inc);
8953 Inc = DAG.getConstant(*ConstInc->getConstantIntValue(),
8954 ConstInc->getValueType(0));
8958 (AM == ISD::PRE_INC || AM == ISD::POST_INC ? ISD::ADD : ISD::SUB);
8959 return DAG.getNode(Opc, SDLoc(LD), BP.getSimpleValueType(), BP, Inc);
8962 SDValue DAGCombiner::visitLOAD(SDNode *N) {
8963 LoadSDNode *LD = cast<LoadSDNode>(N);
8964 SDValue Chain = LD->getChain();
8965 SDValue Ptr = LD->getBasePtr();
8967 // If load is not volatile and there are no uses of the loaded value (and
8968 // the updated indexed value in case of indexed loads), change uses of the
8969 // chain value into uses of the chain input (i.e. delete the dead load).
8970 if (!LD->isVolatile()) {
8971 if (N->getValueType(1) == MVT::Other) {
8973 if (!N->hasAnyUseOfValue(0)) {
8974 // It's not safe to use the two value CombineTo variant here. e.g.
8975 // v1, chain2 = load chain1, loc
8976 // v2, chain3 = load chain2, loc
8978 // Now we replace use of chain2 with chain1. This makes the second load
8979 // isomorphic to the one we are deleting, and thus makes this load live.
8980 DEBUG(dbgs() << "\nReplacing.6 ";
8982 dbgs() << "\nWith chain: ";
8983 Chain.getNode()->dump(&DAG);
8985 WorklistRemover DeadNodes(*this);
8986 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
8989 deleteAndRecombine(N);
8991 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8995 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
8997 // If this load has an opaque TargetConstant offset, then we cannot split
8998 // the indexing into an add/sub directly (that TargetConstant may not be
8999 // valid for a different type of node, and we cannot convert an opaque
9000 // target constant into a regular constant).
9001 bool HasOTCInc = LD->getOperand(2).getOpcode() == ISD::TargetConstant &&
9002 cast<ConstantSDNode>(LD->getOperand(2))->isOpaque();
9004 if (!N->hasAnyUseOfValue(0) &&
9005 ((MaySplitLoadIndex && !HasOTCInc) || !N->hasAnyUseOfValue(1))) {
9006 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
9008 if (N->hasAnyUseOfValue(1) && MaySplitLoadIndex && !HasOTCInc) {
9009 Index = SplitIndexingFromLoad(LD);
9010 // Try to fold the base pointer arithmetic into subsequent loads and
9012 AddUsersToWorklist(N);
9014 Index = DAG.getUNDEF(N->getValueType(1));
9015 DEBUG(dbgs() << "\nReplacing.7 ";
9017 dbgs() << "\nWith: ";
9018 Undef.getNode()->dump(&DAG);
9019 dbgs() << " and 2 other values\n");
9020 WorklistRemover DeadNodes(*this);
9021 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
9022 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Index);
9023 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
9024 deleteAndRecombine(N);
9025 return SDValue(N, 0); // Return N so it doesn't get rechecked!
9030 // If this load is directly stored, replace the load value with the stored
9032 // TODO: Handle store large -> read small portion.
9033 // TODO: Handle TRUNCSTORE/LOADEXT
9034 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
9035 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
9036 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
9037 if (PrevST->getBasePtr() == Ptr &&
9038 PrevST->getValue().getValueType() == N->getValueType(0))
9039 return CombineTo(N, Chain.getOperand(1), Chain);
9043 // Try to infer better alignment information than the load already has.
9044 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
9045 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
9046 if (Align > LD->getMemOperand()->getBaseAlignment()) {
9048 DAG.getExtLoad(LD->getExtensionType(), SDLoc(N),
9049 LD->getValueType(0),
9050 Chain, Ptr, LD->getPointerInfo(),
9052 LD->isVolatile(), LD->isNonTemporal(),
9053 LD->isInvariant(), Align, LD->getAAInfo());
9054 if (NewLoad.getNode() != N)
9055 return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true);
9060 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
9061 : DAG.getSubtarget().useAA();
9063 if (CombinerAAOnlyFunc.getNumOccurrences() &&
9064 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
9067 if (UseAA && LD->isUnindexed()) {
9068 // Walk up chain skipping non-aliasing memory nodes.
9069 SDValue BetterChain = FindBetterChain(N, Chain);
9071 // If there is a better chain.
9072 if (Chain != BetterChain) {
9075 // Replace the chain to void dependency.
9076 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
9077 ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD),
9078 BetterChain, Ptr, LD->getMemOperand());
9080 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD),
9081 LD->getValueType(0),
9082 BetterChain, Ptr, LD->getMemoryVT(),
9083 LD->getMemOperand());
9086 // Create token factor to keep old chain connected.
9087 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
9088 MVT::Other, Chain, ReplLoad.getValue(1));
9090 // Make sure the new and old chains are cleaned up.
9091 AddToWorklist(Token.getNode());
9093 // Replace uses with load result and token factor. Don't add users
9095 return CombineTo(N, ReplLoad.getValue(0), Token, false);
9099 // Try transforming N to an indexed load.
9100 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
9101 return SDValue(N, 0);
9103 // Try to slice up N to more direct loads if the slices are mapped to
9104 // different register banks or pairing can take place.
9106 return SDValue(N, 0);
9112 /// \brief Helper structure used to slice a load in smaller loads.
9113 /// Basically a slice is obtained from the following sequence:
9114 /// Origin = load Ty1, Base
9115 /// Shift = srl Ty1 Origin, CstTy Amount
9116 /// Inst = trunc Shift to Ty2
9118 /// Then, it will be rewriten into:
9119 /// Slice = load SliceTy, Base + SliceOffset
9120 /// [Inst = zext Slice to Ty2], only if SliceTy <> Ty2
9122 /// SliceTy is deduced from the number of bits that are actually used to
9124 struct LoadedSlice {
9125 /// \brief Helper structure used to compute the cost of a slice.
9127 /// Are we optimizing for code size.
9132 unsigned CrossRegisterBanksCopies;
9136 Cost(bool ForCodeSize = false)
9137 : ForCodeSize(ForCodeSize), Loads(0), Truncates(0),
9138 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {}
9140 /// \brief Get the cost of one isolated slice.
9141 Cost(const LoadedSlice &LS, bool ForCodeSize = false)
9142 : ForCodeSize(ForCodeSize), Loads(1), Truncates(0),
9143 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {
9144 EVT TruncType = LS.Inst->getValueType(0);
9145 EVT LoadedType = LS.getLoadedType();
9146 if (TruncType != LoadedType &&
9147 !LS.DAG->getTargetLoweringInfo().isZExtFree(LoadedType, TruncType))
9151 /// \brief Account for slicing gain in the current cost.
9152 /// Slicing provide a few gains like removing a shift or a
9153 /// truncate. This method allows to grow the cost of the original
9154 /// load with the gain from this slice.
9155 void addSliceGain(const LoadedSlice &LS) {
9156 // Each slice saves a truncate.
9157 const TargetLowering &TLI = LS.DAG->getTargetLoweringInfo();
9158 if (!TLI.isTruncateFree(LS.Inst->getValueType(0),
9159 LS.Inst->getOperand(0).getValueType()))
9161 // If there is a shift amount, this slice gets rid of it.
9164 // If this slice can merge a cross register bank copy, account for it.
9165 if (LS.canMergeExpensiveCrossRegisterBankCopy())
9166 ++CrossRegisterBanksCopies;
9169 Cost &operator+=(const Cost &RHS) {
9171 Truncates += RHS.Truncates;
9172 CrossRegisterBanksCopies += RHS.CrossRegisterBanksCopies;
9178 bool operator==(const Cost &RHS) const {
9179 return Loads == RHS.Loads && Truncates == RHS.Truncates &&
9180 CrossRegisterBanksCopies == RHS.CrossRegisterBanksCopies &&
9181 ZExts == RHS.ZExts && Shift == RHS.Shift;
9184 bool operator!=(const Cost &RHS) const { return !(*this == RHS); }
9186 bool operator<(const Cost &RHS) const {
9187 // Assume cross register banks copies are as expensive as loads.
9188 // FIXME: Do we want some more target hooks?
9189 unsigned ExpensiveOpsLHS = Loads + CrossRegisterBanksCopies;
9190 unsigned ExpensiveOpsRHS = RHS.Loads + RHS.CrossRegisterBanksCopies;
9191 // Unless we are optimizing for code size, consider the
9192 // expensive operation first.
9193 if (!ForCodeSize && ExpensiveOpsLHS != ExpensiveOpsRHS)
9194 return ExpensiveOpsLHS < ExpensiveOpsRHS;
9195 return (Truncates + ZExts + Shift + ExpensiveOpsLHS) <
9196 (RHS.Truncates + RHS.ZExts + RHS.Shift + ExpensiveOpsRHS);
9199 bool operator>(const Cost &RHS) const { return RHS < *this; }
9201 bool operator<=(const Cost &RHS) const { return !(RHS < *this); }
9203 bool operator>=(const Cost &RHS) const { return !(*this < RHS); }
9205 // The last instruction that represent the slice. This should be a
9206 // truncate instruction.
9208 // The original load instruction.
9210 // The right shift amount in bits from the original load.
9212 // The DAG from which Origin came from.
9213 // This is used to get some contextual information about legal types, etc.
9216 LoadedSlice(SDNode *Inst = nullptr, LoadSDNode *Origin = nullptr,
9217 unsigned Shift = 0, SelectionDAG *DAG = nullptr)
9218 : Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {}
9220 /// \brief Get the bits used in a chunk of bits \p BitWidth large.
9221 /// \return Result is \p BitWidth and has used bits set to 1 and
9222 /// not used bits set to 0.
9223 APInt getUsedBits() const {
9224 // Reproduce the trunc(lshr) sequence:
9225 // - Start from the truncated value.
9226 // - Zero extend to the desired bit width.
9228 assert(Origin && "No original load to compare against.");
9229 unsigned BitWidth = Origin->getValueSizeInBits(0);
9230 assert(Inst && "This slice is not bound to an instruction");
9231 assert(Inst->getValueSizeInBits(0) <= BitWidth &&
9232 "Extracted slice is bigger than the whole type!");
9233 APInt UsedBits(Inst->getValueSizeInBits(0), 0);
9234 UsedBits.setAllBits();
9235 UsedBits = UsedBits.zext(BitWidth);
9240 /// \brief Get the size of the slice to be loaded in bytes.
9241 unsigned getLoadedSize() const {
9242 unsigned SliceSize = getUsedBits().countPopulation();
9243 assert(!(SliceSize & 0x7) && "Size is not a multiple of a byte.");
9244 return SliceSize / 8;
9247 /// \brief Get the type that will be loaded for this slice.
9248 /// Note: This may not be the final type for the slice.
9249 EVT getLoadedType() const {
9250 assert(DAG && "Missing context");
9251 LLVMContext &Ctxt = *DAG->getContext();
9252 return EVT::getIntegerVT(Ctxt, getLoadedSize() * 8);
9255 /// \brief Get the alignment of the load used for this slice.
9256 unsigned getAlignment() const {
9257 unsigned Alignment = Origin->getAlignment();
9258 unsigned Offset = getOffsetFromBase();
9260 Alignment = MinAlign(Alignment, Alignment + Offset);
9264 /// \brief Check if this slice can be rewritten with legal operations.
9265 bool isLegal() const {
9266 // An invalid slice is not legal.
9267 if (!Origin || !Inst || !DAG)
9270 // Offsets are for indexed load only, we do not handle that.
9271 if (Origin->getOffset().getOpcode() != ISD::UNDEF)
9274 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
9276 // Check that the type is legal.
9277 EVT SliceType = getLoadedType();
9278 if (!TLI.isTypeLegal(SliceType))
9281 // Check that the load is legal for this type.
9282 if (!TLI.isOperationLegal(ISD::LOAD, SliceType))
9285 // Check that the offset can be computed.
9286 // 1. Check its type.
9287 EVT PtrType = Origin->getBasePtr().getValueType();
9288 if (PtrType == MVT::Untyped || PtrType.isExtended())
9291 // 2. Check that it fits in the immediate.
9292 if (!TLI.isLegalAddImmediate(getOffsetFromBase()))
9295 // 3. Check that the computation is legal.
9296 if (!TLI.isOperationLegal(ISD::ADD, PtrType))
9299 // Check that the zext is legal if it needs one.
9300 EVT TruncateType = Inst->getValueType(0);
9301 if (TruncateType != SliceType &&
9302 !TLI.isOperationLegal(ISD::ZERO_EXTEND, TruncateType))
9308 /// \brief Get the offset in bytes of this slice in the original chunk of
9310 /// \pre DAG != nullptr.
9311 uint64_t getOffsetFromBase() const {
9312 assert(DAG && "Missing context.");
9314 DAG->getTargetLoweringInfo().getDataLayout()->isBigEndian();
9315 assert(!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported.");
9316 uint64_t Offset = Shift / 8;
9317 unsigned TySizeInBytes = Origin->getValueSizeInBits(0) / 8;
9318 assert(!(Origin->getValueSizeInBits(0) & 0x7) &&
9319 "The size of the original loaded type is not a multiple of a"
9321 // If Offset is bigger than TySizeInBytes, it means we are loading all
9322 // zeros. This should have been optimized before in the process.
9323 assert(TySizeInBytes > Offset &&
9324 "Invalid shift amount for given loaded size");
9326 Offset = TySizeInBytes - Offset - getLoadedSize();
9330 /// \brief Generate the sequence of instructions to load the slice
9331 /// represented by this object and redirect the uses of this slice to
9332 /// this new sequence of instructions.
9333 /// \pre this->Inst && this->Origin are valid Instructions and this
9334 /// object passed the legal check: LoadedSlice::isLegal returned true.
9335 /// \return The last instruction of the sequence used to load the slice.
9336 SDValue loadSlice() const {
9337 assert(Inst && Origin && "Unable to replace a non-existing slice.");
9338 const SDValue &OldBaseAddr = Origin->getBasePtr();
9339 SDValue BaseAddr = OldBaseAddr;
9340 // Get the offset in that chunk of bytes w.r.t. the endianess.
9341 int64_t Offset = static_cast<int64_t>(getOffsetFromBase());
9342 assert(Offset >= 0 && "Offset too big to fit in int64_t!");
9344 // BaseAddr = BaseAddr + Offset.
9345 EVT ArithType = BaseAddr.getValueType();
9346 BaseAddr = DAG->getNode(ISD::ADD, SDLoc(Origin), ArithType, BaseAddr,
9347 DAG->getConstant(Offset, ArithType));
9350 // Create the type of the loaded slice according to its size.
9351 EVT SliceType = getLoadedType();
9353 // Create the load for the slice.
9354 SDValue LastInst = DAG->getLoad(
9355 SliceType, SDLoc(Origin), Origin->getChain(), BaseAddr,
9356 Origin->getPointerInfo().getWithOffset(Offset), Origin->isVolatile(),
9357 Origin->isNonTemporal(), Origin->isInvariant(), getAlignment());
9358 // If the final type is not the same as the loaded type, this means that
9359 // we have to pad with zero. Create a zero extend for that.
9360 EVT FinalType = Inst->getValueType(0);
9361 if (SliceType != FinalType)
9363 DAG->getNode(ISD::ZERO_EXTEND, SDLoc(LastInst), FinalType, LastInst);
9367 /// \brief Check if this slice can be merged with an expensive cross register
9368 /// bank copy. E.g.,
9370 /// f = bitcast i32 i to float
9371 bool canMergeExpensiveCrossRegisterBankCopy() const {
9372 if (!Inst || !Inst->hasOneUse())
9374 SDNode *Use = *Inst->use_begin();
9375 if (Use->getOpcode() != ISD::BITCAST)
9377 assert(DAG && "Missing context");
9378 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
9379 EVT ResVT = Use->getValueType(0);
9380 const TargetRegisterClass *ResRC = TLI.getRegClassFor(ResVT.getSimpleVT());
9381 const TargetRegisterClass *ArgRC =
9382 TLI.getRegClassFor(Use->getOperand(0).getValueType().getSimpleVT());
9383 if (ArgRC == ResRC || !TLI.isOperationLegal(ISD::LOAD, ResVT))
9386 // At this point, we know that we perform a cross-register-bank copy.
9387 // Check if it is expensive.
9388 const TargetRegisterInfo *TRI = DAG->getSubtarget().getRegisterInfo();
9389 // Assume bitcasts are cheap, unless both register classes do not
9390 // explicitly share a common sub class.
9391 if (!TRI || TRI->getCommonSubClass(ArgRC, ResRC))
9394 // Check if it will be merged with the load.
9395 // 1. Check the alignment constraint.
9396 unsigned RequiredAlignment = TLI.getDataLayout()->getABITypeAlignment(
9397 ResVT.getTypeForEVT(*DAG->getContext()));
9399 if (RequiredAlignment > getAlignment())
9402 // 2. Check that the load is a legal operation for that type.
9403 if (!TLI.isOperationLegal(ISD::LOAD, ResVT))
9406 // 3. Check that we do not have a zext in the way.
9407 if (Inst->getValueType(0) != getLoadedType())
9415 /// \brief Check that all bits set in \p UsedBits form a dense region, i.e.,
9416 /// \p UsedBits looks like 0..0 1..1 0..0.
9417 static bool areUsedBitsDense(const APInt &UsedBits) {
9418 // If all the bits are one, this is dense!
9419 if (UsedBits.isAllOnesValue())
9422 // Get rid of the unused bits on the right.
9423 APInt NarrowedUsedBits = UsedBits.lshr(UsedBits.countTrailingZeros());
9424 // Get rid of the unused bits on the left.
9425 if (NarrowedUsedBits.countLeadingZeros())
9426 NarrowedUsedBits = NarrowedUsedBits.trunc(NarrowedUsedBits.getActiveBits());
9427 // Check that the chunk of bits is completely used.
9428 return NarrowedUsedBits.isAllOnesValue();
9431 /// \brief Check whether or not \p First and \p Second are next to each other
9432 /// in memory. This means that there is no hole between the bits loaded
9433 /// by \p First and the bits loaded by \p Second.
9434 static bool areSlicesNextToEachOther(const LoadedSlice &First,
9435 const LoadedSlice &Second) {
9436 assert(First.Origin == Second.Origin && First.Origin &&
9437 "Unable to match different memory origins.");
9438 APInt UsedBits = First.getUsedBits();
9439 assert((UsedBits & Second.getUsedBits()) == 0 &&
9440 "Slices are not supposed to overlap.");
9441 UsedBits |= Second.getUsedBits();
9442 return areUsedBitsDense(UsedBits);
9445 /// \brief Adjust the \p GlobalLSCost according to the target
9446 /// paring capabilities and the layout of the slices.
9447 /// \pre \p GlobalLSCost should account for at least as many loads as
9448 /// there is in the slices in \p LoadedSlices.
9449 static void adjustCostForPairing(SmallVectorImpl<LoadedSlice> &LoadedSlices,
9450 LoadedSlice::Cost &GlobalLSCost) {
9451 unsigned NumberOfSlices = LoadedSlices.size();
9452 // If there is less than 2 elements, no pairing is possible.
9453 if (NumberOfSlices < 2)
9456 // Sort the slices so that elements that are likely to be next to each
9457 // other in memory are next to each other in the list.
9458 std::sort(LoadedSlices.begin(), LoadedSlices.end(),
9459 [](const LoadedSlice &LHS, const LoadedSlice &RHS) {
9460 assert(LHS.Origin == RHS.Origin && "Different bases not implemented.");
9461 return LHS.getOffsetFromBase() < RHS.getOffsetFromBase();
9463 const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo();
9464 // First (resp. Second) is the first (resp. Second) potentially candidate
9465 // to be placed in a paired load.
9466 const LoadedSlice *First = nullptr;
9467 const LoadedSlice *Second = nullptr;
9468 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice,
9469 // Set the beginning of the pair.
9472 Second = &LoadedSlices[CurrSlice];
9474 // If First is NULL, it means we start a new pair.
9475 // Get to the next slice.
9479 EVT LoadedType = First->getLoadedType();
9481 // If the types of the slices are different, we cannot pair them.
9482 if (LoadedType != Second->getLoadedType())
9485 // Check if the target supplies paired loads for this type.
9486 unsigned RequiredAlignment = 0;
9487 if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) {
9488 // move to the next pair, this type is hopeless.
9492 // Check if we meet the alignment requirement.
9493 if (RequiredAlignment > First->getAlignment())
9496 // Check that both loads are next to each other in memory.
9497 if (!areSlicesNextToEachOther(*First, *Second))
9500 assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!");
9501 --GlobalLSCost.Loads;
9502 // Move to the next pair.
9507 /// \brief Check the profitability of all involved LoadedSlice.
9508 /// Currently, it is considered profitable if there is exactly two
9509 /// involved slices (1) which are (2) next to each other in memory, and
9510 /// whose cost (\see LoadedSlice::Cost) is smaller than the original load (3).
9512 /// Note: The order of the elements in \p LoadedSlices may be modified, but not
9513 /// the elements themselves.
9515 /// FIXME: When the cost model will be mature enough, we can relax
9516 /// constraints (1) and (2).
9517 static bool isSlicingProfitable(SmallVectorImpl<LoadedSlice> &LoadedSlices,
9518 const APInt &UsedBits, bool ForCodeSize) {
9519 unsigned NumberOfSlices = LoadedSlices.size();
9520 if (StressLoadSlicing)
9521 return NumberOfSlices > 1;
9524 if (NumberOfSlices != 2)
9528 if (!areUsedBitsDense(UsedBits))
9532 LoadedSlice::Cost OrigCost(ForCodeSize), GlobalSlicingCost(ForCodeSize);
9533 // The original code has one big load.
9535 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice) {
9536 const LoadedSlice &LS = LoadedSlices[CurrSlice];
9537 // Accumulate the cost of all the slices.
9538 LoadedSlice::Cost SliceCost(LS, ForCodeSize);
9539 GlobalSlicingCost += SliceCost;
9541 // Account as cost in the original configuration the gain obtained
9542 // with the current slices.
9543 OrigCost.addSliceGain(LS);
9546 // If the target supports paired load, adjust the cost accordingly.
9547 adjustCostForPairing(LoadedSlices, GlobalSlicingCost);
9548 return OrigCost > GlobalSlicingCost;
9551 /// \brief If the given load, \p LI, is used only by trunc or trunc(lshr)
9552 /// operations, split it in the various pieces being extracted.
9554 /// This sort of thing is introduced by SROA.
9555 /// This slicing takes care not to insert overlapping loads.
9556 /// \pre LI is a simple load (i.e., not an atomic or volatile load).
9557 bool DAGCombiner::SliceUpLoad(SDNode *N) {
9558 if (Level < AfterLegalizeDAG)
9561 LoadSDNode *LD = cast<LoadSDNode>(N);
9562 if (LD->isVolatile() || !ISD::isNormalLoad(LD) ||
9563 !LD->getValueType(0).isInteger())
9566 // Keep track of already used bits to detect overlapping values.
9567 // In that case, we will just abort the transformation.
9568 APInt UsedBits(LD->getValueSizeInBits(0), 0);
9570 SmallVector<LoadedSlice, 4> LoadedSlices;
9572 // Check if this load is used as several smaller chunks of bits.
9573 // Basically, look for uses in trunc or trunc(lshr) and record a new chain
9574 // of computation for each trunc.
9575 for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end();
9576 UI != UIEnd; ++UI) {
9577 // Skip the uses of the chain.
9578 if (UI.getUse().getResNo() != 0)
9584 // Check if this is a trunc(lshr).
9585 if (User->getOpcode() == ISD::SRL && User->hasOneUse() &&
9586 isa<ConstantSDNode>(User->getOperand(1))) {
9587 Shift = cast<ConstantSDNode>(User->getOperand(1))->getZExtValue();
9588 User = *User->use_begin();
9591 // At this point, User is a Truncate, iff we encountered, trunc or
9593 if (User->getOpcode() != ISD::TRUNCATE)
9596 // The width of the type must be a power of 2 and greater than 8-bits.
9597 // Otherwise the load cannot be represented in LLVM IR.
9598 // Moreover, if we shifted with a non-8-bits multiple, the slice
9599 // will be across several bytes. We do not support that.
9600 unsigned Width = User->getValueSizeInBits(0);
9601 if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7))
9604 // Build the slice for this chain of computations.
9605 LoadedSlice LS(User, LD, Shift, &DAG);
9606 APInt CurrentUsedBits = LS.getUsedBits();
9608 // Check if this slice overlaps with another.
9609 if ((CurrentUsedBits & UsedBits) != 0)
9611 // Update the bits used globally.
9612 UsedBits |= CurrentUsedBits;
9614 // Check if the new slice would be legal.
9618 // Record the slice.
9619 LoadedSlices.push_back(LS);
9622 // Abort slicing if it does not seem to be profitable.
9623 if (!isSlicingProfitable(LoadedSlices, UsedBits, ForCodeSize))
9628 // Rewrite each chain to use an independent load.
9629 // By construction, each chain can be represented by a unique load.
9631 // Prepare the argument for the new token factor for all the slices.
9632 SmallVector<SDValue, 8> ArgChains;
9633 for (SmallVectorImpl<LoadedSlice>::const_iterator
9634 LSIt = LoadedSlices.begin(),
9635 LSItEnd = LoadedSlices.end();
9636 LSIt != LSItEnd; ++LSIt) {
9637 SDValue SliceInst = LSIt->loadSlice();
9638 CombineTo(LSIt->Inst, SliceInst, true);
9639 if (SliceInst.getNode()->getOpcode() != ISD::LOAD)
9640 SliceInst = SliceInst.getOperand(0);
9641 assert(SliceInst->getOpcode() == ISD::LOAD &&
9642 "It takes more than a zext to get to the loaded slice!!");
9643 ArgChains.push_back(SliceInst.getValue(1));
9646 SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other,
9648 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
9652 /// Check to see if V is (and load (ptr), imm), where the load is having
9653 /// specific bytes cleared out. If so, return the byte size being masked out
9654 /// and the shift amount.
9655 static std::pair<unsigned, unsigned>
9656 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
9657 std::pair<unsigned, unsigned> Result(0, 0);
9659 // Check for the structure we're looking for.
9660 if (V->getOpcode() != ISD::AND ||
9661 !isa<ConstantSDNode>(V->getOperand(1)) ||
9662 !ISD::isNormalLoad(V->getOperand(0).getNode()))
9665 // Check the chain and pointer.
9666 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
9667 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
9669 // The store should be chained directly to the load or be an operand of a
9671 if (LD == Chain.getNode())
9673 else if (Chain->getOpcode() != ISD::TokenFactor)
9674 return Result; // Fail.
9677 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
9678 if (Chain->getOperand(i).getNode() == LD) {
9682 if (!isOk) return Result;
9685 // This only handles simple types.
9686 if (V.getValueType() != MVT::i16 &&
9687 V.getValueType() != MVT::i32 &&
9688 V.getValueType() != MVT::i64)
9691 // Check the constant mask. Invert it so that the bits being masked out are
9692 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
9693 // follow the sign bit for uniformity.
9694 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
9695 unsigned NotMaskLZ = countLeadingZeros(NotMask);
9696 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
9697 unsigned NotMaskTZ = countTrailingZeros(NotMask);
9698 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
9699 if (NotMaskLZ == 64) return Result; // All zero mask.
9701 // See if we have a continuous run of bits. If so, we have 0*1+0*
9702 if (countTrailingOnes(NotMask >> NotMaskTZ) + NotMaskTZ + NotMaskLZ != 64)
9705 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
9706 if (V.getValueType() != MVT::i64 && NotMaskLZ)
9707 NotMaskLZ -= 64-V.getValueSizeInBits();
9709 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
9710 switch (MaskedBytes) {
9714 default: return Result; // All one mask, or 5-byte mask.
9717 // Verify that the first bit starts at a multiple of mask so that the access
9718 // is aligned the same as the access width.
9719 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
9721 Result.first = MaskedBytes;
9722 Result.second = NotMaskTZ/8;
9727 /// Check to see if IVal is something that provides a value as specified by
9728 /// MaskInfo. If so, replace the specified store with a narrower store of
9731 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
9732 SDValue IVal, StoreSDNode *St,
9734 unsigned NumBytes = MaskInfo.first;
9735 unsigned ByteShift = MaskInfo.second;
9736 SelectionDAG &DAG = DC->getDAG();
9738 // Check to see if IVal is all zeros in the part being masked in by the 'or'
9739 // that uses this. If not, this is not a replacement.
9740 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
9741 ByteShift*8, (ByteShift+NumBytes)*8);
9742 if (!DAG.MaskedValueIsZero(IVal, Mask)) return nullptr;
9744 // Check that it is legal on the target to do this. It is legal if the new
9745 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
9747 MVT VT = MVT::getIntegerVT(NumBytes*8);
9748 if (!DC->isTypeLegal(VT))
9751 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
9752 // shifted by ByteShift and truncated down to NumBytes.
9754 IVal = DAG.getNode(ISD::SRL, SDLoc(IVal), IVal.getValueType(), IVal,
9755 DAG.getConstant(ByteShift*8,
9756 DC->getShiftAmountTy(IVal.getValueType())));
9758 // Figure out the offset for the store and the alignment of the access.
9760 unsigned NewAlign = St->getAlignment();
9762 if (DAG.getTargetLoweringInfo().isLittleEndian())
9763 StOffset = ByteShift;
9765 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
9767 SDValue Ptr = St->getBasePtr();
9769 Ptr = DAG.getNode(ISD::ADD, SDLoc(IVal), Ptr.getValueType(),
9770 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
9771 NewAlign = MinAlign(NewAlign, StOffset);
9774 // Truncate down to the new size.
9775 IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal);
9778 return DAG.getStore(St->getChain(), SDLoc(St), IVal, Ptr,
9779 St->getPointerInfo().getWithOffset(StOffset),
9780 false, false, NewAlign).getNode();
9784 /// Look for sequence of load / op / store where op is one of 'or', 'xor', and
9785 /// 'and' of immediates. If 'op' is only touching some of the loaded bits, try
9786 /// narrowing the load and store if it would end up being a win for performance
9788 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
9789 StoreSDNode *ST = cast<StoreSDNode>(N);
9790 if (ST->isVolatile())
9793 SDValue Chain = ST->getChain();
9794 SDValue Value = ST->getValue();
9795 SDValue Ptr = ST->getBasePtr();
9796 EVT VT = Value.getValueType();
9798 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
9801 unsigned Opc = Value.getOpcode();
9803 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
9804 // is a byte mask indicating a consecutive number of bytes, check to see if
9805 // Y is known to provide just those bytes. If so, we try to replace the
9806 // load + replace + store sequence with a single (narrower) store, which makes
9808 if (Opc == ISD::OR) {
9809 std::pair<unsigned, unsigned> MaskedLoad;
9810 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
9811 if (MaskedLoad.first)
9812 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
9813 Value.getOperand(1), ST,this))
9814 return SDValue(NewST, 0);
9816 // Or is commutative, so try swapping X and Y.
9817 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
9818 if (MaskedLoad.first)
9819 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
9820 Value.getOperand(0), ST,this))
9821 return SDValue(NewST, 0);
9824 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
9825 Value.getOperand(1).getOpcode() != ISD::Constant)
9828 SDValue N0 = Value.getOperand(0);
9829 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
9830 Chain == SDValue(N0.getNode(), 1)) {
9831 LoadSDNode *LD = cast<LoadSDNode>(N0);
9832 if (LD->getBasePtr() != Ptr ||
9833 LD->getPointerInfo().getAddrSpace() !=
9834 ST->getPointerInfo().getAddrSpace())
9837 // Find the type to narrow it the load / op / store to.
9838 SDValue N1 = Value.getOperand(1);
9839 unsigned BitWidth = N1.getValueSizeInBits();
9840 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
9841 if (Opc == ISD::AND)
9842 Imm ^= APInt::getAllOnesValue(BitWidth);
9843 if (Imm == 0 || Imm.isAllOnesValue())
9845 unsigned ShAmt = Imm.countTrailingZeros();
9846 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
9847 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
9848 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
9849 // The narrowing should be profitable, the load/store operation should be
9850 // legal (or custom) and the store size should be equal to the NewVT width.
9851 while (NewBW < BitWidth &&
9852 (NewVT.getStoreSizeInBits() != NewBW ||
9853 !TLI.isOperationLegalOrCustom(Opc, NewVT) ||
9854 !TLI.isNarrowingProfitable(VT, NewVT))) {
9855 NewBW = NextPowerOf2(NewBW);
9856 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
9858 if (NewBW >= BitWidth)
9861 // If the lsb changed does not start at the type bitwidth boundary,
9862 // start at the previous one.
9864 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
9865 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt,
9866 std::min(BitWidth, ShAmt + NewBW));
9867 if ((Imm & Mask) == Imm) {
9868 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
9869 if (Opc == ISD::AND)
9870 NewImm ^= APInt::getAllOnesValue(NewBW);
9871 uint64_t PtrOff = ShAmt / 8;
9872 // For big endian targets, we need to adjust the offset to the pointer to
9873 // load the correct bytes.
9874 if (TLI.isBigEndian())
9875 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
9877 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
9878 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
9879 if (NewAlign < TLI.getDataLayout()->getABITypeAlignment(NewVTTy))
9882 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD),
9883 Ptr.getValueType(), Ptr,
9884 DAG.getConstant(PtrOff, Ptr.getValueType()));
9885 SDValue NewLD = DAG.getLoad(NewVT, SDLoc(N0),
9886 LD->getChain(), NewPtr,
9887 LD->getPointerInfo().getWithOffset(PtrOff),
9888 LD->isVolatile(), LD->isNonTemporal(),
9889 LD->isInvariant(), NewAlign,
9891 SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD,
9892 DAG.getConstant(NewImm, NewVT));
9893 SDValue NewST = DAG.getStore(Chain, SDLoc(N),
9895 ST->getPointerInfo().getWithOffset(PtrOff),
9896 false, false, NewAlign);
9898 AddToWorklist(NewPtr.getNode());
9899 AddToWorklist(NewLD.getNode());
9900 AddToWorklist(NewVal.getNode());
9901 WorklistRemover DeadNodes(*this);
9902 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
9911 /// For a given floating point load / store pair, if the load value isn't used
9912 /// by any other operations, then consider transforming the pair to integer
9913 /// load / store operations if the target deems the transformation profitable.
9914 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
9915 StoreSDNode *ST = cast<StoreSDNode>(N);
9916 SDValue Chain = ST->getChain();
9917 SDValue Value = ST->getValue();
9918 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
9919 Value.hasOneUse() &&
9920 Chain == SDValue(Value.getNode(), 1)) {
9921 LoadSDNode *LD = cast<LoadSDNode>(Value);
9922 EVT VT = LD->getMemoryVT();
9923 if (!VT.isFloatingPoint() ||
9924 VT != ST->getMemoryVT() ||
9925 LD->isNonTemporal() ||
9926 ST->isNonTemporal() ||
9927 LD->getPointerInfo().getAddrSpace() != 0 ||
9928 ST->getPointerInfo().getAddrSpace() != 0)
9931 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
9932 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
9933 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
9934 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
9935 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
9938 unsigned LDAlign = LD->getAlignment();
9939 unsigned STAlign = ST->getAlignment();
9940 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
9941 unsigned ABIAlign = TLI.getDataLayout()->getABITypeAlignment(IntVTTy);
9942 if (LDAlign < ABIAlign || STAlign < ABIAlign)
9945 SDValue NewLD = DAG.getLoad(IntVT, SDLoc(Value),
9946 LD->getChain(), LD->getBasePtr(),
9947 LD->getPointerInfo(),
9948 false, false, false, LDAlign);
9950 SDValue NewST = DAG.getStore(NewLD.getValue(1), SDLoc(N),
9951 NewLD, ST->getBasePtr(),
9952 ST->getPointerInfo(),
9953 false, false, STAlign);
9955 AddToWorklist(NewLD.getNode());
9956 AddToWorklist(NewST.getNode());
9957 WorklistRemover DeadNodes(*this);
9958 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
9967 /// Helper struct to parse and store a memory address as base + index + offset.
9968 /// We ignore sign extensions when it is safe to do so.
9969 /// The following two expressions are not equivalent. To differentiate we need
9970 /// to store whether there was a sign extension involved in the index
9972 /// (load (i64 add (i64 copyfromreg %c)
9973 /// (i64 signextend (add (i8 load %index)
9977 /// (load (i64 add (i64 copyfromreg %c)
9978 /// (i64 signextend (i32 add (i32 signextend (i8 load %index))
9980 struct BaseIndexOffset {
9984 bool IsIndexSignExt;
9986 BaseIndexOffset() : Offset(0), IsIndexSignExt(false) {}
9988 BaseIndexOffset(SDValue Base, SDValue Index, int64_t Offset,
9989 bool IsIndexSignExt) :
9990 Base(Base), Index(Index), Offset(Offset), IsIndexSignExt(IsIndexSignExt) {}
9992 bool equalBaseIndex(const BaseIndexOffset &Other) {
9993 return Other.Base == Base && Other.Index == Index &&
9994 Other.IsIndexSignExt == IsIndexSignExt;
9997 /// Parses tree in Ptr for base, index, offset addresses.
9998 static BaseIndexOffset match(SDValue Ptr) {
9999 bool IsIndexSignExt = false;
10001 // We only can pattern match BASE + INDEX + OFFSET. If Ptr is not an ADD
10002 // instruction, then it could be just the BASE or everything else we don't
10003 // know how to handle. Just use Ptr as BASE and give up.
10004 if (Ptr->getOpcode() != ISD::ADD)
10005 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
10007 // We know that we have at least an ADD instruction. Try to pattern match
10008 // the simple case of BASE + OFFSET.
10009 if (isa<ConstantSDNode>(Ptr->getOperand(1))) {
10010 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
10011 return BaseIndexOffset(Ptr->getOperand(0), SDValue(), Offset,
10015 // Inside a loop the current BASE pointer is calculated using an ADD and a
10016 // MUL instruction. In this case Ptr is the actual BASE pointer.
10017 // (i64 add (i64 %array_ptr)
10018 // (i64 mul (i64 %induction_var)
10019 // (i64 %element_size)))
10020 if (Ptr->getOperand(1)->getOpcode() == ISD::MUL)
10021 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
10023 // Look at Base + Index + Offset cases.
10024 SDValue Base = Ptr->getOperand(0);
10025 SDValue IndexOffset = Ptr->getOperand(1);
10027 // Skip signextends.
10028 if (IndexOffset->getOpcode() == ISD::SIGN_EXTEND) {
10029 IndexOffset = IndexOffset->getOperand(0);
10030 IsIndexSignExt = true;
10033 // Either the case of Base + Index (no offset) or something else.
10034 if (IndexOffset->getOpcode() != ISD::ADD)
10035 return BaseIndexOffset(Base, IndexOffset, 0, IsIndexSignExt);
10037 // Now we have the case of Base + Index + offset.
10038 SDValue Index = IndexOffset->getOperand(0);
10039 SDValue Offset = IndexOffset->getOperand(1);
10041 if (!isa<ConstantSDNode>(Offset))
10042 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
10044 // Ignore signextends.
10045 if (Index->getOpcode() == ISD::SIGN_EXTEND) {
10046 Index = Index->getOperand(0);
10047 IsIndexSignExt = true;
10048 } else IsIndexSignExt = false;
10050 int64_t Off = cast<ConstantSDNode>(Offset)->getSExtValue();
10051 return BaseIndexOffset(Base, Index, Off, IsIndexSignExt);
10056 bool DAGCombiner::MergeStoresOfConstantsOrVecElts(
10057 SmallVectorImpl<MemOpLink> &StoreNodes, EVT MemVT,
10058 unsigned NumElem, bool IsConstantSrc, bool UseVector) {
10059 // Make sure we have something to merge.
10063 int64_t ElementSizeBytes = MemVT.getSizeInBits() / 8;
10064 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
10065 unsigned LatestNodeUsed = 0;
10067 for (unsigned i=0; i < NumElem; ++i) {
10068 // Find a chain for the new wide-store operand. Notice that some
10069 // of the store nodes that we found may not be selected for inclusion
10070 // in the wide store. The chain we use needs to be the chain of the
10071 // latest store node which is *used* and replaced by the wide store.
10072 if (StoreNodes[i].SequenceNum < StoreNodes[LatestNodeUsed].SequenceNum)
10073 LatestNodeUsed = i;
10076 // The latest Node in the DAG.
10077 LSBaseSDNode *LatestOp = StoreNodes[LatestNodeUsed].MemNode;
10078 SDLoc DL(StoreNodes[0].MemNode);
10082 // Find a legal type for the vector store.
10083 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
10084 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
10085 if (IsConstantSrc) {
10086 // A vector store with a constant source implies that the constant is
10087 // zero; we only handle merging stores of constant zeros because the zero
10088 // can be materialized without a load.
10089 // It may be beneficial to loosen this restriction to allow non-zero
10091 StoredVal = DAG.getConstant(0, Ty);
10093 SmallVector<SDValue, 8> Ops;
10094 for (unsigned i = 0; i < NumElem ; ++i) {
10095 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10096 SDValue Val = St->getValue();
10097 // All of the operands of a BUILD_VECTOR must have the same type.
10098 if (Val.getValueType() != MemVT)
10100 Ops.push_back(Val);
10103 // Build the extracted vector elements back into a vector.
10104 StoredVal = DAG.getNode(ISD::BUILD_VECTOR, DL, Ty, Ops);
10107 // We should always use a vector store when merging extracted vector
10108 // elements, so this path implies a store of constants.
10109 assert(IsConstantSrc && "Merged vector elements should use vector store");
10111 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
10112 APInt StoreInt(StoreBW, 0);
10114 // Construct a single integer constant which is made of the smaller
10115 // constant inputs.
10116 bool IsLE = TLI.isLittleEndian();
10117 for (unsigned i = 0; i < NumElem ; ++i) {
10118 unsigned Idx = IsLE ? (NumElem - 1 - i) : i;
10119 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
10120 SDValue Val = St->getValue();
10121 StoreInt <<= ElementSizeBytes*8;
10122 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
10123 StoreInt |= C->getAPIntValue().zext(StoreBW);
10124 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
10125 StoreInt |= C->getValueAPF().bitcastToAPInt().zext(StoreBW);
10127 llvm_unreachable("Invalid constant element type");
10131 // Create the new Load and Store operations.
10132 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
10133 StoredVal = DAG.getConstant(StoreInt, StoreTy);
10136 SDValue NewStore = DAG.getStore(LatestOp->getChain(), DL, StoredVal,
10137 FirstInChain->getBasePtr(),
10138 FirstInChain->getPointerInfo(),
10140 FirstInChain->getAlignment());
10142 // Replace the last store with the new store
10143 CombineTo(LatestOp, NewStore);
10144 // Erase all other stores.
10145 for (unsigned i = 0; i < NumElem ; ++i) {
10146 if (StoreNodes[i].MemNode == LatestOp)
10148 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10149 // ReplaceAllUsesWith will replace all uses that existed when it was
10150 // called, but graph optimizations may cause new ones to appear. For
10151 // example, the case in pr14333 looks like
10153 // St's chain -> St -> another store -> X
10155 // And the only difference from St to the other store is the chain.
10156 // When we change it's chain to be St's chain they become identical,
10157 // get CSEed and the net result is that X is now a use of St.
10158 // Since we know that St is redundant, just iterate.
10159 while (!St->use_empty())
10160 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain());
10161 deleteAndRecombine(St);
10167 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
10168 if (OptLevel == CodeGenOpt::None)
10171 EVT MemVT = St->getMemoryVT();
10172 int64_t ElementSizeBytes = MemVT.getSizeInBits()/8;
10173 bool NoVectors = DAG.getMachineFunction().getFunction()->hasFnAttribute(
10174 Attribute::NoImplicitFloat);
10176 // Don't merge vectors into wider inputs.
10177 if (MemVT.isVector() || !MemVT.isSimple())
10180 // Perform an early exit check. Do not bother looking at stored values that
10181 // are not constants, loads, or extracted vector elements.
10182 SDValue StoredVal = St->getValue();
10183 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
10184 bool IsConstantSrc = isa<ConstantSDNode>(StoredVal) ||
10185 isa<ConstantFPSDNode>(StoredVal);
10186 bool IsExtractVecEltSrc = (StoredVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT);
10188 if (!IsConstantSrc && !IsLoadSrc && !IsExtractVecEltSrc)
10191 // Only look at ends of store sequences.
10192 SDValue Chain = SDValue(St, 0);
10193 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
10196 // This holds the base pointer, index, and the offset in bytes from the base
10198 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr());
10200 // We must have a base and an offset.
10201 if (!BasePtr.Base.getNode())
10204 // Do not handle stores to undef base pointers.
10205 if (BasePtr.Base.getOpcode() == ISD::UNDEF)
10208 // Save the LoadSDNodes that we find in the chain.
10209 // We need to make sure that these nodes do not interfere with
10210 // any of the store nodes.
10211 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
10213 // Save the StoreSDNodes that we find in the chain.
10214 SmallVector<MemOpLink, 8> StoreNodes;
10216 // Walk up the chain and look for nodes with offsets from the same
10217 // base pointer. Stop when reaching an instruction with a different kind
10218 // or instruction which has a different base pointer.
10220 StoreSDNode *Index = St;
10222 // If the chain has more than one use, then we can't reorder the mem ops.
10223 if (Index != St && !SDValue(Index, 0)->hasOneUse())
10226 // Find the base pointer and offset for this memory node.
10227 BaseIndexOffset Ptr = BaseIndexOffset::match(Index->getBasePtr());
10229 // Check that the base pointer is the same as the original one.
10230 if (!Ptr.equalBaseIndex(BasePtr))
10233 // Check that the alignment is the same.
10234 if (Index->getAlignment() != St->getAlignment())
10237 // The memory operands must not be volatile.
10238 if (Index->isVolatile() || Index->isIndexed())
10242 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
10243 if (St->isTruncatingStore())
10246 // The stored memory type must be the same.
10247 if (Index->getMemoryVT() != MemVT)
10250 // We do not allow unaligned stores because we want to prevent overriding
10252 if (Index->getAlignment()*8 != MemVT.getSizeInBits())
10255 // We found a potential memory operand to merge.
10256 StoreNodes.push_back(MemOpLink(Index, Ptr.Offset, Seq++));
10258 // Find the next memory operand in the chain. If the next operand in the
10259 // chain is a store then move up and continue the scan with the next
10260 // memory operand. If the next operand is a load save it and use alias
10261 // information to check if it interferes with anything.
10262 SDNode *NextInChain = Index->getChain().getNode();
10264 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
10265 // We found a store node. Use it for the next iteration.
10268 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
10269 if (Ldn->isVolatile()) {
10274 // Save the load node for later. Continue the scan.
10275 AliasLoadNodes.push_back(Ldn);
10276 NextInChain = Ldn->getChain().getNode();
10285 // Check if there is anything to merge.
10286 if (StoreNodes.size() < 2)
10289 // Sort the memory operands according to their distance from the base pointer.
10290 std::sort(StoreNodes.begin(), StoreNodes.end(),
10291 [](MemOpLink LHS, MemOpLink RHS) {
10292 return LHS.OffsetFromBase < RHS.OffsetFromBase ||
10293 (LHS.OffsetFromBase == RHS.OffsetFromBase &&
10294 LHS.SequenceNum > RHS.SequenceNum);
10297 // Scan the memory operations on the chain and find the first non-consecutive
10298 // store memory address.
10299 unsigned LastConsecutiveStore = 0;
10300 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
10301 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) {
10303 // Check that the addresses are consecutive starting from the second
10304 // element in the list of stores.
10306 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
10307 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
10311 bool Alias = false;
10312 // Check if this store interferes with any of the loads that we found.
10313 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld)
10314 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) {
10318 // We found a load that alias with this store. Stop the sequence.
10322 // Mark this node as useful.
10323 LastConsecutiveStore = i;
10326 // The node with the lowest store address.
10327 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
10329 // Store the constants into memory as one consecutive store.
10330 if (IsConstantSrc) {
10331 unsigned LastLegalType = 0;
10332 unsigned LastLegalVectorType = 0;
10333 bool NonZero = false;
10334 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
10335 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10336 SDValue StoredVal = St->getValue();
10338 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
10339 NonZero |= !C->isNullValue();
10340 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) {
10341 NonZero |= !C->getConstantFPValue()->isNullValue();
10347 // Find a legal type for the constant store.
10348 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
10349 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
10350 if (TLI.isTypeLegal(StoreTy))
10351 LastLegalType = i+1;
10352 // Or check whether a truncstore is legal.
10353 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
10354 TargetLowering::TypePromoteInteger) {
10355 EVT LegalizedStoredValueTy =
10356 TLI.getTypeToTransformTo(*DAG.getContext(), StoredVal.getValueType());
10357 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy))
10358 LastLegalType = i+1;
10361 // Find a legal type for the vector store.
10362 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
10363 if (TLI.isTypeLegal(Ty))
10364 LastLegalVectorType = i + 1;
10367 // We only use vectors if the constant is known to be zero and the
10368 // function is not marked with the noimplicitfloat attribute.
10369 if (NonZero || NoVectors)
10370 LastLegalVectorType = 0;
10372 // Check if we found a legal integer type to store.
10373 if (LastLegalType == 0 && LastLegalVectorType == 0)
10376 bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors;
10377 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType;
10379 return MergeStoresOfConstantsOrVecElts(StoreNodes, MemVT, NumElem,
10383 // When extracting multiple vector elements, try to store them
10384 // in one vector store rather than a sequence of scalar stores.
10385 if (IsExtractVecEltSrc) {
10386 unsigned NumElem = 0;
10387 for (unsigned i = 0; i < LastConsecutiveStore + 1; ++i) {
10388 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10389 SDValue StoredVal = St->getValue();
10390 // This restriction could be loosened.
10391 // Bail out if any stored values are not elements extracted from a vector.
10392 // It should be possible to handle mixed sources, but load sources need
10393 // more careful handling (see the block of code below that handles
10394 // consecutive loads).
10395 if (StoredVal.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
10398 // Find a legal type for the vector store.
10399 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
10400 if (TLI.isTypeLegal(Ty))
10404 return MergeStoresOfConstantsOrVecElts(StoreNodes, MemVT, NumElem,
10408 // Below we handle the case of multiple consecutive stores that
10409 // come from multiple consecutive loads. We merge them into a single
10410 // wide load and a single wide store.
10412 // Look for load nodes which are used by the stored values.
10413 SmallVector<MemOpLink, 8> LoadNodes;
10415 // Find acceptable loads. Loads need to have the same chain (token factor),
10416 // must not be zext, volatile, indexed, and they must be consecutive.
10417 BaseIndexOffset LdBasePtr;
10418 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
10419 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10420 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
10423 // Loads must only have one use.
10424 if (!Ld->hasNUsesOfValue(1, 0))
10427 // Check that the alignment is the same as the stores.
10428 if (Ld->getAlignment() != St->getAlignment())
10431 // The memory operands must not be volatile.
10432 if (Ld->isVolatile() || Ld->isIndexed())
10435 // We do not accept ext loads.
10436 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
10439 // The stored memory type must be the same.
10440 if (Ld->getMemoryVT() != MemVT)
10443 BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld->getBasePtr());
10444 // If this is not the first ptr that we check.
10445 if (LdBasePtr.Base.getNode()) {
10446 // The base ptr must be the same.
10447 if (!LdPtr.equalBaseIndex(LdBasePtr))
10450 // Check that all other base pointers are the same as this one.
10454 // We found a potential memory operand to merge.
10455 LoadNodes.push_back(MemOpLink(Ld, LdPtr.Offset, 0));
10458 if (LoadNodes.size() < 2)
10461 // If we have load/store pair instructions and we only have two values,
10463 unsigned RequiredAlignment;
10464 if (LoadNodes.size() == 2 && TLI.hasPairedLoad(MemVT, RequiredAlignment) &&
10465 St->getAlignment() >= RequiredAlignment)
10468 // Scan the memory operations on the chain and find the first non-consecutive
10469 // load memory address. These variables hold the index in the store node
10471 unsigned LastConsecutiveLoad = 0;
10472 // This variable refers to the size and not index in the array.
10473 unsigned LastLegalVectorType = 0;
10474 unsigned LastLegalIntegerType = 0;
10475 StartAddress = LoadNodes[0].OffsetFromBase;
10476 SDValue FirstChain = LoadNodes[0].MemNode->getChain();
10477 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
10478 // All loads much share the same chain.
10479 if (LoadNodes[i].MemNode->getChain() != FirstChain)
10482 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
10483 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
10485 LastConsecutiveLoad = i;
10487 // Find a legal type for the vector store.
10488 EVT StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
10489 if (TLI.isTypeLegal(StoreTy))
10490 LastLegalVectorType = i + 1;
10492 // Find a legal type for the integer store.
10493 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
10494 StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
10495 if (TLI.isTypeLegal(StoreTy))
10496 LastLegalIntegerType = i + 1;
10497 // Or check whether a truncstore and extload is legal.
10498 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
10499 TargetLowering::TypePromoteInteger) {
10500 EVT LegalizedStoredValueTy =
10501 TLI.getTypeToTransformTo(*DAG.getContext(), StoreTy);
10502 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
10503 TLI.isLoadExtLegal(ISD::ZEXTLOAD, LegalizedStoredValueTy, StoreTy) &&
10504 TLI.isLoadExtLegal(ISD::SEXTLOAD, LegalizedStoredValueTy, StoreTy) &&
10505 TLI.isLoadExtLegal(ISD::EXTLOAD, LegalizedStoredValueTy, StoreTy))
10506 LastLegalIntegerType = i+1;
10510 // Only use vector types if the vector type is larger than the integer type.
10511 // If they are the same, use integers.
10512 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors;
10513 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType);
10515 // We add +1 here because the LastXXX variables refer to location while
10516 // the NumElem refers to array/index size.
10517 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1;
10518 NumElem = std::min(LastLegalType, NumElem);
10523 // The latest Node in the DAG.
10524 unsigned LatestNodeUsed = 0;
10525 for (unsigned i=1; i<NumElem; ++i) {
10526 // Find a chain for the new wide-store operand. Notice that some
10527 // of the store nodes that we found may not be selected for inclusion
10528 // in the wide store. The chain we use needs to be the chain of the
10529 // latest store node which is *used* and replaced by the wide store.
10530 if (StoreNodes[i].SequenceNum < StoreNodes[LatestNodeUsed].SequenceNum)
10531 LatestNodeUsed = i;
10534 LSBaseSDNode *LatestOp = StoreNodes[LatestNodeUsed].MemNode;
10536 // Find if it is better to use vectors or integers to load and store
10540 JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
10542 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
10543 JointMemOpVT = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
10546 SDLoc LoadDL(LoadNodes[0].MemNode);
10547 SDLoc StoreDL(StoreNodes[0].MemNode);
10549 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
10550 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL,
10551 FirstLoad->getChain(),
10552 FirstLoad->getBasePtr(),
10553 FirstLoad->getPointerInfo(),
10554 false, false, false,
10555 FirstLoad->getAlignment());
10557 SDValue NewStore = DAG.getStore(LatestOp->getChain(), StoreDL, NewLoad,
10558 FirstInChain->getBasePtr(),
10559 FirstInChain->getPointerInfo(), false, false,
10560 FirstInChain->getAlignment());
10562 // Replace one of the loads with the new load.
10563 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode);
10564 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
10565 SDValue(NewLoad.getNode(), 1));
10567 // Remove the rest of the load chains.
10568 for (unsigned i = 1; i < NumElem ; ++i) {
10569 // Replace all chain users of the old load nodes with the chain of the new
10571 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
10572 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain());
10575 // Replace the last store with the new store.
10576 CombineTo(LatestOp, NewStore);
10577 // Erase all other stores.
10578 for (unsigned i = 0; i < NumElem ; ++i) {
10579 // Remove all Store nodes.
10580 if (StoreNodes[i].MemNode == LatestOp)
10582 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10583 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain());
10584 deleteAndRecombine(St);
10590 SDValue DAGCombiner::visitSTORE(SDNode *N) {
10591 StoreSDNode *ST = cast<StoreSDNode>(N);
10592 SDValue Chain = ST->getChain();
10593 SDValue Value = ST->getValue();
10594 SDValue Ptr = ST->getBasePtr();
10596 // If this is a store of a bit convert, store the input value if the
10597 // resultant store does not need a higher alignment than the original.
10598 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
10599 ST->isUnindexed()) {
10600 unsigned OrigAlign = ST->getAlignment();
10601 EVT SVT = Value.getOperand(0).getValueType();
10602 unsigned Align = TLI.getDataLayout()->
10603 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
10604 if (Align <= OrigAlign &&
10605 ((!LegalOperations && !ST->isVolatile()) ||
10606 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
10607 return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0),
10608 Ptr, ST->getPointerInfo(), ST->isVolatile(),
10609 ST->isNonTemporal(), OrigAlign,
10613 // Turn 'store undef, Ptr' -> nothing.
10614 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
10617 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
10618 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
10619 // NOTE: If the original store is volatile, this transform must not increase
10620 // the number of stores. For example, on x86-32 an f64 can be stored in one
10621 // processor operation but an i64 (which is not legal) requires two. So the
10622 // transform should not be done in this case.
10623 if (Value.getOpcode() != ISD::TargetConstantFP) {
10625 switch (CFP->getSimpleValueType(0).SimpleTy) {
10626 default: llvm_unreachable("Unknown FP type");
10627 case MVT::f16: // We don't do this for these yet.
10633 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
10634 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
10635 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
10636 bitcastToAPInt().getZExtValue(), MVT::i32);
10637 return DAG.getStore(Chain, SDLoc(N), Tmp,
10638 Ptr, ST->getMemOperand());
10642 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
10643 !ST->isVolatile()) ||
10644 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
10645 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
10646 getZExtValue(), MVT::i64);
10647 return DAG.getStore(Chain, SDLoc(N), Tmp,
10648 Ptr, ST->getMemOperand());
10651 if (!ST->isVolatile() &&
10652 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
10653 // Many FP stores are not made apparent until after legalize, e.g. for
10654 // argument passing. Since this is so common, custom legalize the
10655 // 64-bit integer store into two 32-bit stores.
10656 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
10657 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
10658 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
10659 if (TLI.isBigEndian()) std::swap(Lo, Hi);
10661 unsigned Alignment = ST->getAlignment();
10662 bool isVolatile = ST->isVolatile();
10663 bool isNonTemporal = ST->isNonTemporal();
10664 AAMDNodes AAInfo = ST->getAAInfo();
10666 SDValue St0 = DAG.getStore(Chain, SDLoc(ST), Lo,
10667 Ptr, ST->getPointerInfo(),
10668 isVolatile, isNonTemporal,
10669 ST->getAlignment(), AAInfo);
10670 Ptr = DAG.getNode(ISD::ADD, SDLoc(N), Ptr.getValueType(), Ptr,
10671 DAG.getConstant(4, Ptr.getValueType()));
10672 Alignment = MinAlign(Alignment, 4U);
10673 SDValue St1 = DAG.getStore(Chain, SDLoc(ST), Hi,
10674 Ptr, ST->getPointerInfo().getWithOffset(4),
10675 isVolatile, isNonTemporal,
10676 Alignment, AAInfo);
10677 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other,
10686 // Try to infer better alignment information than the store already has.
10687 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
10688 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
10689 if (Align > ST->getAlignment()) {
10691 DAG.getTruncStore(Chain, SDLoc(N), Value,
10692 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
10693 ST->isVolatile(), ST->isNonTemporal(), Align,
10695 if (NewStore.getNode() != N)
10696 return CombineTo(ST, NewStore, true);
10701 // Try transforming a pair floating point load / store ops to integer
10702 // load / store ops.
10703 SDValue NewST = TransformFPLoadStorePair(N);
10704 if (NewST.getNode())
10707 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
10708 : DAG.getSubtarget().useAA();
10710 if (CombinerAAOnlyFunc.getNumOccurrences() &&
10711 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
10714 if (UseAA && ST->isUnindexed()) {
10715 // Walk up chain skipping non-aliasing memory nodes.
10716 SDValue BetterChain = FindBetterChain(N, Chain);
10718 // If there is a better chain.
10719 if (Chain != BetterChain) {
10722 // Replace the chain to avoid dependency.
10723 if (ST->isTruncatingStore()) {
10724 ReplStore = DAG.getTruncStore(BetterChain, SDLoc(N), Value, Ptr,
10725 ST->getMemoryVT(), ST->getMemOperand());
10727 ReplStore = DAG.getStore(BetterChain, SDLoc(N), Value, Ptr,
10728 ST->getMemOperand());
10731 // Create token to keep both nodes around.
10732 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
10733 MVT::Other, Chain, ReplStore);
10735 // Make sure the new and old chains are cleaned up.
10736 AddToWorklist(Token.getNode());
10738 // Don't add users to work list.
10739 return CombineTo(N, Token, false);
10743 // Try transforming N to an indexed store.
10744 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
10745 return SDValue(N, 0);
10747 // FIXME: is there such a thing as a truncating indexed store?
10748 if (ST->isTruncatingStore() && ST->isUnindexed() &&
10749 Value.getValueType().isInteger()) {
10750 // See if we can simplify the input to this truncstore with knowledge that
10751 // only the low bits are being used. For example:
10752 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
10754 GetDemandedBits(Value,
10755 APInt::getLowBitsSet(
10756 Value.getValueType().getScalarType().getSizeInBits(),
10757 ST->getMemoryVT().getScalarType().getSizeInBits()));
10758 AddToWorklist(Value.getNode());
10759 if (Shorter.getNode())
10760 return DAG.getTruncStore(Chain, SDLoc(N), Shorter,
10761 Ptr, ST->getMemoryVT(), ST->getMemOperand());
10763 // Otherwise, see if we can simplify the operation with
10764 // SimplifyDemandedBits, which only works if the value has a single use.
10765 if (SimplifyDemandedBits(Value,
10766 APInt::getLowBitsSet(
10767 Value.getValueType().getScalarType().getSizeInBits(),
10768 ST->getMemoryVT().getScalarType().getSizeInBits())))
10769 return SDValue(N, 0);
10772 // If this is a load followed by a store to the same location, then the store
10774 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
10775 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
10776 ST->isUnindexed() && !ST->isVolatile() &&
10777 // There can't be any side effects between the load and store, such as
10778 // a call or store.
10779 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
10780 // The store is dead, remove it.
10785 // If this is a store followed by a store with the same value to the same
10786 // location, then the store is dead/noop.
10787 if (StoreSDNode *ST1 = dyn_cast<StoreSDNode>(Chain)) {
10788 if (ST1->getBasePtr() == Ptr && ST->getMemoryVT() == ST1->getMemoryVT() &&
10789 ST1->getValue() == Value && ST->isUnindexed() && !ST->isVolatile() &&
10790 ST1->isUnindexed() && !ST1->isVolatile()) {
10791 // The store is dead, remove it.
10796 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
10797 // truncating store. We can do this even if this is already a truncstore.
10798 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
10799 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
10800 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
10801 ST->getMemoryVT())) {
10802 return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0),
10803 Ptr, ST->getMemoryVT(), ST->getMemOperand());
10806 // Only perform this optimization before the types are legal, because we
10807 // don't want to perform this optimization on every DAGCombine invocation.
10809 bool EverChanged = false;
10812 // There can be multiple store sequences on the same chain.
10813 // Keep trying to merge store sequences until we are unable to do so
10814 // or until we merge the last store on the chain.
10815 bool Changed = MergeConsecutiveStores(ST);
10816 EverChanged |= Changed;
10817 if (!Changed) break;
10818 } while (ST->getOpcode() != ISD::DELETED_NODE);
10821 return SDValue(N, 0);
10824 return ReduceLoadOpStoreWidth(N);
10827 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
10828 SDValue InVec = N->getOperand(0);
10829 SDValue InVal = N->getOperand(1);
10830 SDValue EltNo = N->getOperand(2);
10833 // If the inserted element is an UNDEF, just use the input vector.
10834 if (InVal.getOpcode() == ISD::UNDEF)
10837 EVT VT = InVec.getValueType();
10839 // If we can't generate a legal BUILD_VECTOR, exit
10840 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
10843 // Check that we know which element is being inserted
10844 if (!isa<ConstantSDNode>(EltNo))
10846 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
10848 // Canonicalize insert_vector_elt dag nodes.
10850 // (insert_vector_elt (insert_vector_elt A, Idx0), Idx1)
10851 // -> (insert_vector_elt (insert_vector_elt A, Idx1), Idx0)
10853 // Do this only if the child insert_vector node has one use; also
10854 // do this only if indices are both constants and Idx1 < Idx0.
10855 if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT && InVec.hasOneUse()
10856 && isa<ConstantSDNode>(InVec.getOperand(2))) {
10857 unsigned OtherElt =
10858 cast<ConstantSDNode>(InVec.getOperand(2))->getZExtValue();
10859 if (Elt < OtherElt) {
10861 SDValue NewOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(N), VT,
10862 InVec.getOperand(0), InVal, EltNo);
10863 AddToWorklist(NewOp.getNode());
10864 return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(InVec.getNode()),
10865 VT, NewOp, InVec.getOperand(1), InVec.getOperand(2));
10869 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
10870 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
10871 // vector elements.
10872 SmallVector<SDValue, 8> Ops;
10873 // Do not combine these two vectors if the output vector will not replace
10874 // the input vector.
10875 if (InVec.getOpcode() == ISD::BUILD_VECTOR && InVec.hasOneUse()) {
10876 Ops.append(InVec.getNode()->op_begin(),
10877 InVec.getNode()->op_end());
10878 } else if (InVec.getOpcode() == ISD::UNDEF) {
10879 unsigned NElts = VT.getVectorNumElements();
10880 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
10885 // Insert the element
10886 if (Elt < Ops.size()) {
10887 // All the operands of BUILD_VECTOR must have the same type;
10888 // we enforce that here.
10889 EVT OpVT = Ops[0].getValueType();
10890 if (InVal.getValueType() != OpVT)
10891 InVal = OpVT.bitsGT(InVal.getValueType()) ?
10892 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
10893 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
10897 // Return the new vector
10898 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
10901 SDValue DAGCombiner::ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
10902 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad) {
10903 EVT ResultVT = EVE->getValueType(0);
10904 EVT VecEltVT = InVecVT.getVectorElementType();
10905 unsigned Align = OriginalLoad->getAlignment();
10906 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
10907 VecEltVT.getTypeForEVT(*DAG.getContext()));
10909 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VecEltVT))
10914 SDValue NewPtr = OriginalLoad->getBasePtr();
10916 EVT PtrType = NewPtr.getValueType();
10917 MachinePointerInfo MPI;
10918 if (auto *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo)) {
10919 int Elt = ConstEltNo->getZExtValue();
10920 unsigned PtrOff = VecEltVT.getSizeInBits() * Elt / 8;
10921 if (TLI.isBigEndian())
10922 PtrOff = InVecVT.getSizeInBits() / 8 - PtrOff;
10923 Offset = DAG.getConstant(PtrOff, PtrType);
10924 MPI = OriginalLoad->getPointerInfo().getWithOffset(PtrOff);
10926 Offset = DAG.getNode(
10927 ISD::MUL, SDLoc(EVE), EltNo.getValueType(), EltNo,
10928 DAG.getConstant(VecEltVT.getStoreSize(), EltNo.getValueType()));
10929 if (TLI.isBigEndian())
10930 Offset = DAG.getNode(
10931 ISD::SUB, SDLoc(EVE), EltNo.getValueType(),
10932 DAG.getConstant(InVecVT.getStoreSize(), EltNo.getValueType()), Offset);
10933 MPI = OriginalLoad->getPointerInfo();
10935 NewPtr = DAG.getNode(ISD::ADD, SDLoc(EVE), PtrType, NewPtr, Offset);
10937 // The replacement we need to do here is a little tricky: we need to
10938 // replace an extractelement of a load with a load.
10939 // Use ReplaceAllUsesOfValuesWith to do the replacement.
10940 // Note that this replacement assumes that the extractvalue is the only
10941 // use of the load; that's okay because we don't want to perform this
10942 // transformation in other cases anyway.
10945 if (ResultVT.bitsGT(VecEltVT)) {
10946 // If the result type of vextract is wider than the load, then issue an
10947 // extending load instead.
10948 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, ResultVT,
10952 Load = DAG.getExtLoad(
10953 ExtType, SDLoc(EVE), ResultVT, OriginalLoad->getChain(), NewPtr, MPI,
10954 VecEltVT, OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
10955 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
10956 Chain = Load.getValue(1);
10958 Load = DAG.getLoad(
10959 VecEltVT, SDLoc(EVE), OriginalLoad->getChain(), NewPtr, MPI,
10960 OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
10961 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
10962 Chain = Load.getValue(1);
10963 if (ResultVT.bitsLT(VecEltVT))
10964 Load = DAG.getNode(ISD::TRUNCATE, SDLoc(EVE), ResultVT, Load);
10966 Load = DAG.getNode(ISD::BITCAST, SDLoc(EVE), ResultVT, Load);
10968 WorklistRemover DeadNodes(*this);
10969 SDValue From[] = { SDValue(EVE, 0), SDValue(OriginalLoad, 1) };
10970 SDValue To[] = { Load, Chain };
10971 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
10972 // Since we're explicitly calling ReplaceAllUses, add the new node to the
10973 // worklist explicitly as well.
10974 AddToWorklist(Load.getNode());
10975 AddUsersToWorklist(Load.getNode()); // Add users too
10976 // Make sure to revisit this node to clean it up; it will usually be dead.
10977 AddToWorklist(EVE);
10979 return SDValue(EVE, 0);
10982 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
10983 // (vextract (scalar_to_vector val, 0) -> val
10984 SDValue InVec = N->getOperand(0);
10985 EVT VT = InVec.getValueType();
10986 EVT NVT = N->getValueType(0);
10988 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
10989 // Check if the result type doesn't match the inserted element type. A
10990 // SCALAR_TO_VECTOR may truncate the inserted element and the
10991 // EXTRACT_VECTOR_ELT may widen the extracted vector.
10992 SDValue InOp = InVec.getOperand(0);
10993 if (InOp.getValueType() != NVT) {
10994 assert(InOp.getValueType().isInteger() && NVT.isInteger());
10995 return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT);
11000 SDValue EltNo = N->getOperand(1);
11001 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
11003 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
11004 // We only perform this optimization before the op legalization phase because
11005 // we may introduce new vector instructions which are not backed by TD
11006 // patterns. For example on AVX, extracting elements from a wide vector
11007 // without using extract_subvector. However, if we can find an underlying
11008 // scalar value, then we can always use that.
11009 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
11011 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
11012 int NumElem = VT.getVectorNumElements();
11013 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
11014 // Find the new index to extract from.
11015 int OrigElt = SVOp->getMaskElt(Elt);
11017 // Extracting an undef index is undef.
11019 return DAG.getUNDEF(NVT);
11021 // Select the right vector half to extract from.
11023 if (OrigElt < NumElem) {
11024 SVInVec = InVec->getOperand(0);
11026 SVInVec = InVec->getOperand(1);
11027 OrigElt -= NumElem;
11030 if (SVInVec.getOpcode() == ISD::BUILD_VECTOR) {
11031 SDValue InOp = SVInVec.getOperand(OrigElt);
11032 if (InOp.getValueType() != NVT) {
11033 assert(InOp.getValueType().isInteger() && NVT.isInteger());
11034 InOp = DAG.getSExtOrTrunc(InOp, SDLoc(SVInVec), NVT);
11040 // FIXME: We should handle recursing on other vector shuffles and
11041 // scalar_to_vector here as well.
11043 if (!LegalOperations) {
11044 EVT IndexTy = TLI.getVectorIdxTy();
11045 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT,
11046 SVInVec, DAG.getConstant(OrigElt, IndexTy));
11050 bool BCNumEltsChanged = false;
11051 EVT ExtVT = VT.getVectorElementType();
11054 // If the result of load has to be truncated, then it's not necessarily
11056 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
11059 if (InVec.getOpcode() == ISD::BITCAST) {
11060 // Don't duplicate a load with other uses.
11061 if (!InVec.hasOneUse())
11064 EVT BCVT = InVec.getOperand(0).getValueType();
11065 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
11067 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
11068 BCNumEltsChanged = true;
11069 InVec = InVec.getOperand(0);
11070 ExtVT = BCVT.getVectorElementType();
11073 // (vextract (vN[if]M load $addr), i) -> ([if]M load $addr + i * size)
11074 if (!LegalOperations && !ConstEltNo && InVec.hasOneUse() &&
11075 ISD::isNormalLoad(InVec.getNode()) &&
11076 !N->getOperand(1)->hasPredecessor(InVec.getNode())) {
11077 SDValue Index = N->getOperand(1);
11078 if (LoadSDNode *OrigLoad = dyn_cast<LoadSDNode>(InVec))
11079 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, Index,
11083 // Perform only after legalization to ensure build_vector / vector_shuffle
11084 // optimizations have already been done.
11085 if (!LegalOperations) return SDValue();
11087 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
11088 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
11089 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
11092 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
11094 LoadSDNode *LN0 = nullptr;
11095 const ShuffleVectorSDNode *SVN = nullptr;
11096 if (ISD::isNormalLoad(InVec.getNode())) {
11097 LN0 = cast<LoadSDNode>(InVec);
11098 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
11099 InVec.getOperand(0).getValueType() == ExtVT &&
11100 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
11101 // Don't duplicate a load with other uses.
11102 if (!InVec.hasOneUse())
11105 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
11106 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
11107 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
11109 // (load $addr+1*size)
11111 // Don't duplicate a load with other uses.
11112 if (!InVec.hasOneUse())
11115 // If the bit convert changed the number of elements, it is unsafe
11116 // to examine the mask.
11117 if (BCNumEltsChanged)
11120 // Select the input vector, guarding against out of range extract vector.
11121 unsigned NumElems = VT.getVectorNumElements();
11122 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
11123 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
11125 if (InVec.getOpcode() == ISD::BITCAST) {
11126 // Don't duplicate a load with other uses.
11127 if (!InVec.hasOneUse())
11130 InVec = InVec.getOperand(0);
11132 if (ISD::isNormalLoad(InVec.getNode())) {
11133 LN0 = cast<LoadSDNode>(InVec);
11134 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
11135 EltNo = DAG.getConstant(Elt, EltNo.getValueType());
11139 // Make sure we found a non-volatile load and the extractelement is
11141 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
11144 // If Idx was -1 above, Elt is going to be -1, so just return undef.
11146 return DAG.getUNDEF(LVT);
11148 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, EltNo, LN0);
11154 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
11155 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
11156 // We perform this optimization post type-legalization because
11157 // the type-legalizer often scalarizes integer-promoted vectors.
11158 // Performing this optimization before may create bit-casts which
11159 // will be type-legalized to complex code sequences.
11160 // We perform this optimization only before the operation legalizer because we
11161 // may introduce illegal operations.
11162 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
11165 unsigned NumInScalars = N->getNumOperands();
11167 EVT VT = N->getValueType(0);
11169 // Check to see if this is a BUILD_VECTOR of a bunch of values
11170 // which come from any_extend or zero_extend nodes. If so, we can create
11171 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
11172 // optimizations. We do not handle sign-extend because we can't fill the sign
11174 EVT SourceType = MVT::Other;
11175 bool AllAnyExt = true;
11177 for (unsigned i = 0; i != NumInScalars; ++i) {
11178 SDValue In = N->getOperand(i);
11179 // Ignore undef inputs.
11180 if (In.getOpcode() == ISD::UNDEF) continue;
11182 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
11183 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
11185 // Abort if the element is not an extension.
11186 if (!ZeroExt && !AnyExt) {
11187 SourceType = MVT::Other;
11191 // The input is a ZeroExt or AnyExt. Check the original type.
11192 EVT InTy = In.getOperand(0).getValueType();
11194 // Check that all of the widened source types are the same.
11195 if (SourceType == MVT::Other)
11198 else if (InTy != SourceType) {
11199 // Multiple income types. Abort.
11200 SourceType = MVT::Other;
11204 // Check if all of the extends are ANY_EXTENDs.
11205 AllAnyExt &= AnyExt;
11208 // In order to have valid types, all of the inputs must be extended from the
11209 // same source type and all of the inputs must be any or zero extend.
11210 // Scalar sizes must be a power of two.
11211 EVT OutScalarTy = VT.getScalarType();
11212 bool ValidTypes = SourceType != MVT::Other &&
11213 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
11214 isPowerOf2_32(SourceType.getSizeInBits());
11216 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
11217 // turn into a single shuffle instruction.
11221 bool isLE = TLI.isLittleEndian();
11222 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
11223 assert(ElemRatio > 1 && "Invalid element size ratio");
11224 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
11225 DAG.getConstant(0, SourceType);
11227 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
11228 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
11230 // Populate the new build_vector
11231 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
11232 SDValue Cast = N->getOperand(i);
11233 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
11234 Cast.getOpcode() == ISD::ZERO_EXTEND ||
11235 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
11237 if (Cast.getOpcode() == ISD::UNDEF)
11238 In = DAG.getUNDEF(SourceType);
11240 In = Cast->getOperand(0);
11241 unsigned Index = isLE ? (i * ElemRatio) :
11242 (i * ElemRatio + (ElemRatio - 1));
11244 assert(Index < Ops.size() && "Invalid index");
11248 // The type of the new BUILD_VECTOR node.
11249 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
11250 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
11251 "Invalid vector size");
11252 // Check if the new vector type is legal.
11253 if (!isTypeLegal(VecVT)) return SDValue();
11255 // Make the new BUILD_VECTOR.
11256 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
11258 // The new BUILD_VECTOR node has the potential to be further optimized.
11259 AddToWorklist(BV.getNode());
11260 // Bitcast to the desired type.
11261 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
11264 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
11265 EVT VT = N->getValueType(0);
11267 unsigned NumInScalars = N->getNumOperands();
11270 EVT SrcVT = MVT::Other;
11271 unsigned Opcode = ISD::DELETED_NODE;
11272 unsigned NumDefs = 0;
11274 for (unsigned i = 0; i != NumInScalars; ++i) {
11275 SDValue In = N->getOperand(i);
11276 unsigned Opc = In.getOpcode();
11278 if (Opc == ISD::UNDEF)
11281 // If all scalar values are floats and converted from integers.
11282 if (Opcode == ISD::DELETED_NODE &&
11283 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
11290 EVT InVT = In.getOperand(0).getValueType();
11292 // If all scalar values are typed differently, bail out. It's chosen to
11293 // simplify BUILD_VECTOR of integer types.
11294 if (SrcVT == MVT::Other)
11301 // If the vector has just one element defined, it's not worth to fold it into
11302 // a vectorized one.
11306 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
11307 && "Should only handle conversion from integer to float.");
11308 assert(SrcVT != MVT::Other && "Cannot determine source type!");
11310 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
11312 if (!TLI.isOperationLegalOrCustom(Opcode, NVT))
11315 // Just because the floating-point vector type is legal does not necessarily
11316 // mean that the corresponding integer vector type is.
11317 if (!isTypeLegal(NVT))
11320 SmallVector<SDValue, 8> Opnds;
11321 for (unsigned i = 0; i != NumInScalars; ++i) {
11322 SDValue In = N->getOperand(i);
11324 if (In.getOpcode() == ISD::UNDEF)
11325 Opnds.push_back(DAG.getUNDEF(SrcVT));
11327 Opnds.push_back(In.getOperand(0));
11329 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Opnds);
11330 AddToWorklist(BV.getNode());
11332 return DAG.getNode(Opcode, dl, VT, BV);
11335 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
11336 unsigned NumInScalars = N->getNumOperands();
11338 EVT VT = N->getValueType(0);
11340 // A vector built entirely of undefs is undef.
11341 if (ISD::allOperandsUndef(N))
11342 return DAG.getUNDEF(VT);
11344 if (SDValue V = reduceBuildVecExtToExtBuildVec(N))
11347 if (SDValue V = reduceBuildVecConvertToConvertBuildVec(N))
11350 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
11351 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
11352 // at most two distinct vectors, turn this into a shuffle node.
11354 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
11355 if (!isTypeLegal(VT))
11358 // May only combine to shuffle after legalize if shuffle is legal.
11359 if (LegalOperations && !TLI.isOperationLegal(ISD::VECTOR_SHUFFLE, VT))
11362 SDValue VecIn1, VecIn2;
11363 bool UsesZeroVector = false;
11364 for (unsigned i = 0; i != NumInScalars; ++i) {
11365 SDValue Op = N->getOperand(i);
11366 // Ignore undef inputs.
11367 if (Op.getOpcode() == ISD::UNDEF) continue;
11369 // See if we can combine this build_vector into a blend with a zero vector.
11370 if (!VecIn2.getNode() && ((Op.getOpcode() == ISD::Constant &&
11371 cast<ConstantSDNode>(Op.getNode())->isNullValue()) ||
11372 (Op.getOpcode() == ISD::ConstantFP &&
11373 cast<ConstantFPSDNode>(Op.getNode())->getValueAPF().isZero()))) {
11374 UsesZeroVector = true;
11378 // If this input is something other than a EXTRACT_VECTOR_ELT with a
11379 // constant index, bail out.
11380 if (Op.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
11381 !isa<ConstantSDNode>(Op.getOperand(1))) {
11382 VecIn1 = VecIn2 = SDValue(nullptr, 0);
11386 // We allow up to two distinct input vectors.
11387 SDValue ExtractedFromVec = Op.getOperand(0);
11388 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
11391 if (!VecIn1.getNode()) {
11392 VecIn1 = ExtractedFromVec;
11393 } else if (!VecIn2.getNode() && !UsesZeroVector) {
11394 VecIn2 = ExtractedFromVec;
11396 // Too many inputs.
11397 VecIn1 = VecIn2 = SDValue(nullptr, 0);
11402 // If everything is good, we can make a shuffle operation.
11403 if (VecIn1.getNode()) {
11404 unsigned InNumElements = VecIn1.getValueType().getVectorNumElements();
11405 SmallVector<int, 8> Mask;
11406 for (unsigned i = 0; i != NumInScalars; ++i) {
11407 unsigned Opcode = N->getOperand(i).getOpcode();
11408 if (Opcode == ISD::UNDEF) {
11409 Mask.push_back(-1);
11413 // Operands can also be zero.
11414 if (Opcode != ISD::EXTRACT_VECTOR_ELT) {
11415 assert(UsesZeroVector &&
11416 (Opcode == ISD::Constant || Opcode == ISD::ConstantFP) &&
11417 "Unexpected node found!");
11418 Mask.push_back(NumInScalars+i);
11422 // If extracting from the first vector, just use the index directly.
11423 SDValue Extract = N->getOperand(i);
11424 SDValue ExtVal = Extract.getOperand(1);
11425 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
11426 if (Extract.getOperand(0) == VecIn1) {
11427 Mask.push_back(ExtIndex);
11431 // Otherwise, use InIdx + InputVecSize
11432 Mask.push_back(InNumElements + ExtIndex);
11435 // Avoid introducing illegal shuffles with zero.
11436 if (UsesZeroVector && !TLI.isVectorClearMaskLegal(Mask, VT))
11439 // We can't generate a shuffle node with mismatched input and output types.
11440 // Attempt to transform a single input vector to the correct type.
11441 if ((VT != VecIn1.getValueType())) {
11442 // If the input vector type has a different base type to the output
11443 // vector type, bail out.
11444 EVT VTElemType = VT.getVectorElementType();
11445 if ((VecIn1.getValueType().getVectorElementType() != VTElemType) ||
11446 (VecIn2.getNode() &&
11447 (VecIn2.getValueType().getVectorElementType() != VTElemType)))
11450 // If the input vector is too small, widen it.
11451 // We only support widening of vectors which are half the size of the
11452 // output registers. For example XMM->YMM widening on X86 with AVX.
11453 EVT VecInT = VecIn1.getValueType();
11454 if (VecInT.getSizeInBits() * 2 == VT.getSizeInBits()) {
11455 // If we only have one small input, widen it by adding undef values.
11456 if (!VecIn2.getNode())
11457 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, VecIn1,
11458 DAG.getUNDEF(VecIn1.getValueType()));
11459 else if (VecIn1.getValueType() == VecIn2.getValueType()) {
11460 // If we have two small inputs of the same type, try to concat them.
11461 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, VecIn1, VecIn2);
11462 VecIn2 = SDValue(nullptr, 0);
11465 } else if (VecInT.getSizeInBits() == VT.getSizeInBits() * 2) {
11466 // If the input vector is too large, try to split it.
11467 // We don't support having two input vectors that are too large.
11468 // If the zero vector was used, we can not split the vector,
11469 // since we'd need 3 inputs.
11470 if (UsesZeroVector || VecIn2.getNode())
11473 if (!TLI.isExtractSubvectorCheap(VT, VT.getVectorNumElements()))
11476 // Try to replace VecIn1 with two extract_subvectors
11477 // No need to update the masks, they should still be correct.
11478 VecIn2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, VecIn1,
11479 DAG.getConstant(VT.getVectorNumElements(), TLI.getVectorIdxTy()));
11480 VecIn1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, VecIn1,
11481 DAG.getConstant(0, TLI.getVectorIdxTy()));
11486 if (UsesZeroVector)
11487 VecIn2 = VT.isInteger() ? DAG.getConstant(0, VT) :
11488 DAG.getConstantFP(0.0, VT);
11490 // If VecIn2 is unused then change it to undef.
11491 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
11493 // Check that we were able to transform all incoming values to the same
11495 if (VecIn2.getValueType() != VecIn1.getValueType() ||
11496 VecIn1.getValueType() != VT)
11499 // Return the new VECTOR_SHUFFLE node.
11503 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]);
11509 static SDValue combineConcatVectorOfScalars(SDNode *N, SelectionDAG &DAG) {
11510 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11511 EVT OpVT = N->getOperand(0).getValueType();
11513 // If the operands are legal vectors, leave them alone.
11514 if (TLI.isTypeLegal(OpVT))
11518 EVT VT = N->getValueType(0);
11519 SmallVector<SDValue, 8> Ops;
11521 EVT SVT = EVT::getIntegerVT(*DAG.getContext(), OpVT.getSizeInBits());
11522 SDValue ScalarUndef = DAG.getNode(ISD::UNDEF, DL, SVT);
11524 // Keep track of what we encounter.
11525 bool AnyInteger = false;
11526 bool AnyFP = false;
11527 for (const SDValue &Op : N->ops()) {
11528 if (ISD::BITCAST == Op.getOpcode() &&
11529 !Op.getOperand(0).getValueType().isVector())
11530 Ops.push_back(Op.getOperand(0));
11531 else if (ISD::UNDEF == Op.getOpcode())
11532 Ops.push_back(ScalarUndef);
11536 // Note whether we encounter an integer or floating point scalar.
11537 // If it's neither, bail out, it could be something weird like x86mmx.
11538 EVT LastOpVT = Ops.back().getValueType();
11539 if (LastOpVT.isFloatingPoint())
11541 else if (LastOpVT.isInteger())
11547 // If any of the operands is a floating point scalar bitcast to a vector,
11548 // use floating point types throughout, and bitcast everything.
11549 // Replace UNDEFs by another scalar UNDEF node, of the final desired type.
11551 SVT = EVT::getFloatingPointVT(OpVT.getSizeInBits());
11552 ScalarUndef = DAG.getNode(ISD::UNDEF, DL, SVT);
11554 for (SDValue &Op : Ops) {
11555 if (Op.getValueType() == SVT)
11557 if (Op.getOpcode() == ISD::UNDEF)
11560 Op = DAG.getNode(ISD::BITCAST, DL, SVT, Op);
11565 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SVT,
11566 VT.getSizeInBits() / SVT.getSizeInBits());
11567 return DAG.getNode(ISD::BITCAST, DL, VT,
11568 DAG.getNode(ISD::BUILD_VECTOR, DL, VecVT, Ops));
11571 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
11572 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
11573 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
11574 // inputs come from at most two distinct vectors, turn this into a shuffle
11577 // If we only have one input vector, we don't need to do any concatenation.
11578 if (N->getNumOperands() == 1)
11579 return N->getOperand(0);
11581 // Check if all of the operands are undefs.
11582 EVT VT = N->getValueType(0);
11583 if (ISD::allOperandsUndef(N))
11584 return DAG.getUNDEF(VT);
11586 // Optimize concat_vectors where all but the first of the vectors are undef.
11587 if (std::all_of(std::next(N->op_begin()), N->op_end(), [](const SDValue &Op) {
11588 return Op.getOpcode() == ISD::UNDEF;
11590 SDValue In = N->getOperand(0);
11591 assert(In.getValueType().isVector() && "Must concat vectors");
11593 // Transform: concat_vectors(scalar, undef) -> scalar_to_vector(sclr).
11594 if (In->getOpcode() == ISD::BITCAST &&
11595 !In->getOperand(0)->getValueType(0).isVector()) {
11596 SDValue Scalar = In->getOperand(0);
11598 // If the bitcast type isn't legal, it might be a trunc of a legal type;
11599 // look through the trunc so we can still do the transform:
11600 // concat_vectors(trunc(scalar), undef) -> scalar_to_vector(scalar)
11601 if (Scalar->getOpcode() == ISD::TRUNCATE &&
11602 !TLI.isTypeLegal(Scalar.getValueType()) &&
11603 TLI.isTypeLegal(Scalar->getOperand(0).getValueType()))
11604 Scalar = Scalar->getOperand(0);
11606 EVT SclTy = Scalar->getValueType(0);
11608 if (!SclTy.isFloatingPoint() && !SclTy.isInteger())
11611 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SclTy,
11612 VT.getSizeInBits() / SclTy.getSizeInBits());
11613 if (!TLI.isTypeLegal(NVT) || !TLI.isTypeLegal(Scalar.getValueType()))
11616 SDLoc dl = SDLoc(N);
11617 SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NVT, Scalar);
11618 return DAG.getNode(ISD::BITCAST, dl, VT, Res);
11622 // Fold any combination of BUILD_VECTOR or UNDEF nodes into one BUILD_VECTOR.
11623 // We have already tested above for an UNDEF only concatenation.
11624 // fold (concat_vectors (BUILD_VECTOR A, B, ...), (BUILD_VECTOR C, D, ...))
11625 // -> (BUILD_VECTOR A, B, ..., C, D, ...)
11626 auto IsBuildVectorOrUndef = [](const SDValue &Op) {
11627 return ISD::UNDEF == Op.getOpcode() || ISD::BUILD_VECTOR == Op.getOpcode();
11629 bool AllBuildVectorsOrUndefs =
11630 std::all_of(N->op_begin(), N->op_end(), IsBuildVectorOrUndef);
11631 if (AllBuildVectorsOrUndefs) {
11632 SmallVector<SDValue, 8> Opnds;
11633 EVT SVT = VT.getScalarType();
11636 if (!SVT.isFloatingPoint()) {
11637 // If BUILD_VECTOR are from built from integer, they may have different
11638 // operand types. Get the smallest type and truncate all operands to it.
11639 bool FoundMinVT = false;
11640 for (const SDValue &Op : N->ops())
11641 if (ISD::BUILD_VECTOR == Op.getOpcode()) {
11642 EVT OpSVT = Op.getOperand(0)->getValueType(0);
11643 MinVT = (!FoundMinVT || OpSVT.bitsLE(MinVT)) ? OpSVT : MinVT;
11646 assert(FoundMinVT && "Concat vector type mismatch");
11649 for (const SDValue &Op : N->ops()) {
11650 EVT OpVT = Op.getValueType();
11651 unsigned NumElts = OpVT.getVectorNumElements();
11653 if (ISD::UNDEF == Op.getOpcode())
11654 Opnds.append(NumElts, DAG.getUNDEF(MinVT));
11656 if (ISD::BUILD_VECTOR == Op.getOpcode()) {
11657 if (SVT.isFloatingPoint()) {
11658 assert(SVT == OpVT.getScalarType() && "Concat vector type mismatch");
11659 Opnds.append(Op->op_begin(), Op->op_begin() + NumElts);
11661 for (unsigned i = 0; i != NumElts; ++i)
11663 DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinVT, Op.getOperand(i)));
11668 assert(VT.getVectorNumElements() == Opnds.size() &&
11669 "Concat vector type mismatch");
11670 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
11673 // Fold CONCAT_VECTORS of only bitcast scalars (or undef) to BUILD_VECTOR.
11674 if (SDValue V = combineConcatVectorOfScalars(N, DAG))
11677 // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
11678 // nodes often generate nop CONCAT_VECTOR nodes.
11679 // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that
11680 // place the incoming vectors at the exact same location.
11681 SDValue SingleSource = SDValue();
11682 unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements();
11684 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
11685 SDValue Op = N->getOperand(i);
11687 if (Op.getOpcode() == ISD::UNDEF)
11690 // Check if this is the identity extract:
11691 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
11694 // Find the single incoming vector for the extract_subvector.
11695 if (SingleSource.getNode()) {
11696 if (Op.getOperand(0) != SingleSource)
11699 SingleSource = Op.getOperand(0);
11701 // Check the source type is the same as the type of the result.
11702 // If not, this concat may extend the vector, so we can not
11703 // optimize it away.
11704 if (SingleSource.getValueType() != N->getValueType(0))
11708 unsigned IdentityIndex = i * PartNumElem;
11709 ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1));
11710 // The extract index must be constant.
11714 // Check that we are reading from the identity index.
11715 if (CS->getZExtValue() != IdentityIndex)
11719 if (SingleSource.getNode())
11720 return SingleSource;
11725 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
11726 EVT NVT = N->getValueType(0);
11727 SDValue V = N->getOperand(0);
11729 if (V->getOpcode() == ISD::CONCAT_VECTORS) {
11731 // (extract_subvec (concat V1, V2, ...), i)
11734 // Only operand 0 is checked as 'concat' assumes all inputs of the same
11736 if (V->getOperand(0).getValueType() != NVT)
11738 unsigned Idx = N->getConstantOperandVal(1);
11739 unsigned NumElems = NVT.getVectorNumElements();
11740 assert((Idx % NumElems) == 0 &&
11741 "IDX in concat is not a multiple of the result vector length.");
11742 return V->getOperand(Idx / NumElems);
11746 if (V->getOpcode() == ISD::BITCAST)
11747 V = V.getOperand(0);
11749 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
11751 // Handle only simple case where vector being inserted and vector
11752 // being extracted are of same type, and are half size of larger vectors.
11753 EVT BigVT = V->getOperand(0).getValueType();
11754 EVT SmallVT = V->getOperand(1).getValueType();
11755 if (!NVT.bitsEq(SmallVT) || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
11758 // Only handle cases where both indexes are constants with the same type.
11759 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
11760 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
11762 if (InsIdx && ExtIdx &&
11763 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
11764 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
11766 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
11768 // indices are equal or bit offsets are equal => V1
11769 // otherwise => (extract_subvec V1, ExtIdx)
11770 if (InsIdx->getZExtValue() * SmallVT.getScalarType().getSizeInBits() ==
11771 ExtIdx->getZExtValue() * NVT.getScalarType().getSizeInBits())
11772 return DAG.getNode(ISD::BITCAST, dl, NVT, V->getOperand(1));
11773 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT,
11774 DAG.getNode(ISD::BITCAST, dl,
11775 N->getOperand(0).getValueType(),
11776 V->getOperand(0)), N->getOperand(1));
11783 static SDValue simplifyShuffleOperandRecursively(SmallBitVector &UsedElements,
11784 SDValue V, SelectionDAG &DAG) {
11786 EVT VT = V.getValueType();
11788 switch (V.getOpcode()) {
11792 case ISD::CONCAT_VECTORS: {
11793 EVT OpVT = V->getOperand(0).getValueType();
11794 int OpSize = OpVT.getVectorNumElements();
11795 SmallBitVector OpUsedElements(OpSize, false);
11796 bool FoundSimplification = false;
11797 SmallVector<SDValue, 4> NewOps;
11798 NewOps.reserve(V->getNumOperands());
11799 for (int i = 0, NumOps = V->getNumOperands(); i < NumOps; ++i) {
11800 SDValue Op = V->getOperand(i);
11801 bool OpUsed = false;
11802 for (int j = 0; j < OpSize; ++j)
11803 if (UsedElements[i * OpSize + j]) {
11804 OpUsedElements[j] = true;
11808 OpUsed ? simplifyShuffleOperandRecursively(OpUsedElements, Op, DAG)
11809 : DAG.getUNDEF(OpVT));
11810 FoundSimplification |= Op == NewOps.back();
11811 OpUsedElements.reset();
11813 if (FoundSimplification)
11814 V = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, NewOps);
11818 case ISD::INSERT_SUBVECTOR: {
11819 SDValue BaseV = V->getOperand(0);
11820 SDValue SubV = V->getOperand(1);
11821 auto *IdxN = dyn_cast<ConstantSDNode>(V->getOperand(2));
11825 int SubSize = SubV.getValueType().getVectorNumElements();
11826 int Idx = IdxN->getZExtValue();
11827 bool SubVectorUsed = false;
11828 SmallBitVector SubUsedElements(SubSize, false);
11829 for (int i = 0; i < SubSize; ++i)
11830 if (UsedElements[i + Idx]) {
11831 SubVectorUsed = true;
11832 SubUsedElements[i] = true;
11833 UsedElements[i + Idx] = false;
11836 // Now recurse on both the base and sub vectors.
11837 SDValue SimplifiedSubV =
11839 ? simplifyShuffleOperandRecursively(SubUsedElements, SubV, DAG)
11840 : DAG.getUNDEF(SubV.getValueType());
11841 SDValue SimplifiedBaseV = simplifyShuffleOperandRecursively(UsedElements, BaseV, DAG);
11842 if (SimplifiedSubV != SubV || SimplifiedBaseV != BaseV)
11843 V = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT,
11844 SimplifiedBaseV, SimplifiedSubV, V->getOperand(2));
11850 static SDValue simplifyShuffleOperands(ShuffleVectorSDNode *SVN, SDValue N0,
11851 SDValue N1, SelectionDAG &DAG) {
11852 EVT VT = SVN->getValueType(0);
11853 int NumElts = VT.getVectorNumElements();
11854 SmallBitVector N0UsedElements(NumElts, false), N1UsedElements(NumElts, false);
11855 for (int M : SVN->getMask())
11856 if (M >= 0 && M < NumElts)
11857 N0UsedElements[M] = true;
11858 else if (M >= NumElts)
11859 N1UsedElements[M - NumElts] = true;
11861 SDValue S0 = simplifyShuffleOperandRecursively(N0UsedElements, N0, DAG);
11862 SDValue S1 = simplifyShuffleOperandRecursively(N1UsedElements, N1, DAG);
11863 if (S0 == N0 && S1 == N1)
11866 return DAG.getVectorShuffle(VT, SDLoc(SVN), S0, S1, SVN->getMask());
11869 // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat,
11870 // or turn a shuffle of a single concat into simpler shuffle then concat.
11871 static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
11872 EVT VT = N->getValueType(0);
11873 unsigned NumElts = VT.getVectorNumElements();
11875 SDValue N0 = N->getOperand(0);
11876 SDValue N1 = N->getOperand(1);
11877 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
11879 SmallVector<SDValue, 4> Ops;
11880 EVT ConcatVT = N0.getOperand(0).getValueType();
11881 unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements();
11882 unsigned NumConcats = NumElts / NumElemsPerConcat;
11884 // Special case: shuffle(concat(A,B)) can be more efficiently represented
11885 // as concat(shuffle(A,B),UNDEF) if the shuffle doesn't set any of the high
11886 // half vector elements.
11887 if (NumElemsPerConcat * 2 == NumElts && N1.getOpcode() == ISD::UNDEF &&
11888 std::all_of(SVN->getMask().begin() + NumElemsPerConcat,
11889 SVN->getMask().end(), [](int i) { return i == -1; })) {
11890 N0 = DAG.getVectorShuffle(ConcatVT, SDLoc(N), N0.getOperand(0), N0.getOperand(1),
11891 ArrayRef<int>(SVN->getMask().begin(), NumElemsPerConcat));
11892 N1 = DAG.getUNDEF(ConcatVT);
11893 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, N0, N1);
11896 // Look at every vector that's inserted. We're looking for exact
11897 // subvector-sized copies from a concatenated vector
11898 for (unsigned I = 0; I != NumConcats; ++I) {
11899 // Make sure we're dealing with a copy.
11900 unsigned Begin = I * NumElemsPerConcat;
11901 bool AllUndef = true, NoUndef = true;
11902 for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) {
11903 if (SVN->getMaskElt(J) >= 0)
11910 if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0)
11913 for (unsigned J = 1; J != NumElemsPerConcat; ++J)
11914 if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J))
11917 unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat;
11918 if (FirstElt < N0.getNumOperands())
11919 Ops.push_back(N0.getOperand(FirstElt));
11921 Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands()));
11923 } else if (AllUndef) {
11924 Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType()));
11925 } else { // Mixed with general masks and undefs, can't do optimization.
11930 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops);
11933 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
11934 EVT VT = N->getValueType(0);
11935 unsigned NumElts = VT.getVectorNumElements();
11937 SDValue N0 = N->getOperand(0);
11938 SDValue N1 = N->getOperand(1);
11940 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
11942 // Canonicalize shuffle undef, undef -> undef
11943 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
11944 return DAG.getUNDEF(VT);
11946 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
11948 // Canonicalize shuffle v, v -> v, undef
11950 SmallVector<int, 8> NewMask;
11951 for (unsigned i = 0; i != NumElts; ++i) {
11952 int Idx = SVN->getMaskElt(i);
11953 if (Idx >= (int)NumElts) Idx -= NumElts;
11954 NewMask.push_back(Idx);
11956 return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT),
11960 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
11961 if (N0.getOpcode() == ISD::UNDEF) {
11962 SmallVector<int, 8> NewMask;
11963 for (unsigned i = 0; i != NumElts; ++i) {
11964 int Idx = SVN->getMaskElt(i);
11966 if (Idx >= (int)NumElts)
11969 Idx = -1; // remove reference to lhs
11971 NewMask.push_back(Idx);
11973 return DAG.getVectorShuffle(VT, SDLoc(N), N1, DAG.getUNDEF(VT),
11977 // Remove references to rhs if it is undef
11978 if (N1.getOpcode() == ISD::UNDEF) {
11979 bool Changed = false;
11980 SmallVector<int, 8> NewMask;
11981 for (unsigned i = 0; i != NumElts; ++i) {
11982 int Idx = SVN->getMaskElt(i);
11983 if (Idx >= (int)NumElts) {
11987 NewMask.push_back(Idx);
11990 return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, &NewMask[0]);
11993 // If it is a splat, check if the argument vector is another splat or a
11995 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
11996 SDNode *V = N0.getNode();
11998 // If this is a bit convert that changes the element type of the vector but
11999 // not the number of vector elements, look through it. Be careful not to
12000 // look though conversions that change things like v4f32 to v2f64.
12001 if (V->getOpcode() == ISD::BITCAST) {
12002 SDValue ConvInput = V->getOperand(0);
12003 if (ConvInput.getValueType().isVector() &&
12004 ConvInput.getValueType().getVectorNumElements() == NumElts)
12005 V = ConvInput.getNode();
12008 if (V->getOpcode() == ISD::BUILD_VECTOR) {
12009 assert(V->getNumOperands() == NumElts &&
12010 "BUILD_VECTOR has wrong number of operands");
12012 bool AllSame = true;
12013 for (unsigned i = 0; i != NumElts; ++i) {
12014 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
12015 Base = V->getOperand(i);
12019 // Splat of <u, u, u, u>, return <u, u, u, u>
12020 if (!Base.getNode())
12022 for (unsigned i = 0; i != NumElts; ++i) {
12023 if (V->getOperand(i) != Base) {
12028 // Splat of <x, x, x, x>, return <x, x, x, x>
12032 // Canonicalize any other splat as a build_vector.
12033 const SDValue &Splatted = V->getOperand(SVN->getSplatIndex());
12034 SmallVector<SDValue, 8> Ops(NumElts, Splatted);
12035 SDValue NewBV = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
12036 V->getValueType(0), Ops);
12038 // We may have jumped through bitcasts, so the type of the
12039 // BUILD_VECTOR may not match the type of the shuffle.
12040 if (V->getValueType(0) != VT)
12041 NewBV = DAG.getNode(ISD::BITCAST, SDLoc(N), VT, NewBV);
12046 // There are various patterns used to build up a vector from smaller vectors,
12047 // subvectors, or elements. Scan chains of these and replace unused insertions
12048 // or components with undef.
12049 if (SDValue S = simplifyShuffleOperands(SVN, N0, N1, DAG))
12052 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
12053 Level < AfterLegalizeVectorOps &&
12054 (N1.getOpcode() == ISD::UNDEF ||
12055 (N1.getOpcode() == ISD::CONCAT_VECTORS &&
12056 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) {
12057 SDValue V = partitionShuffleOfConcats(N, DAG);
12063 // Attempt to combine a shuffle of 2 inputs of 'scalar sources' -
12064 // BUILD_VECTOR or SCALAR_TO_VECTOR into a single BUILD_VECTOR.
12065 if (Level < AfterLegalizeVectorOps && TLI.isTypeLegal(VT)) {
12066 SmallVector<SDValue, 8> Ops;
12067 for (int M : SVN->getMask()) {
12068 SDValue Op = DAG.getUNDEF(VT.getScalarType());
12070 int Idx = M % NumElts;
12071 SDValue &S = (M < (int)NumElts ? N0 : N1);
12072 if (S.getOpcode() == ISD::BUILD_VECTOR && S.hasOneUse()) {
12073 Op = S.getOperand(Idx);
12074 } else if (S.getOpcode() == ISD::SCALAR_TO_VECTOR && S.hasOneUse()) {
12076 Op = S.getOperand(0);
12078 // Operand can't be combined - bail out.
12084 if (Ops.size() == VT.getVectorNumElements()) {
12085 // BUILD_VECTOR requires all inputs to be of the same type, find the
12086 // maximum type and extend them all.
12087 EVT SVT = VT.getScalarType();
12088 if (SVT.isInteger())
12089 for (SDValue &Op : Ops)
12090 SVT = (SVT.bitsLT(Op.getValueType()) ? Op.getValueType() : SVT);
12091 if (SVT != VT.getScalarType())
12092 for (SDValue &Op : Ops)
12093 Op = TLI.isZExtFree(Op.getValueType(), SVT)
12094 ? DAG.getZExtOrTrunc(Op, SDLoc(N), SVT)
12095 : DAG.getSExtOrTrunc(Op, SDLoc(N), SVT);
12096 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Ops);
12100 // If this shuffle only has a single input that is a bitcasted shuffle,
12101 // attempt to merge the 2 shuffles and suitably bitcast the inputs/output
12102 // back to their original types.
12103 if (N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
12104 N1.getOpcode() == ISD::UNDEF && Level < AfterLegalizeVectorOps &&
12105 TLI.isTypeLegal(VT)) {
12107 // Peek through the bitcast only if there is one user.
12109 while (BC0.getOpcode() == ISD::BITCAST) {
12110 if (!BC0.hasOneUse())
12112 BC0 = BC0.getOperand(0);
12115 auto ScaleShuffleMask = [](ArrayRef<int> Mask, int Scale) {
12117 return SmallVector<int, 8>(Mask.begin(), Mask.end());
12119 SmallVector<int, 8> NewMask;
12121 for (int s = 0; s != Scale; ++s)
12122 NewMask.push_back(M < 0 ? -1 : Scale * M + s);
12126 if (BC0.getOpcode() == ISD::VECTOR_SHUFFLE && BC0.hasOneUse()) {
12127 EVT SVT = VT.getScalarType();
12128 EVT InnerVT = BC0->getValueType(0);
12129 EVT InnerSVT = InnerVT.getScalarType();
12131 // Determine which shuffle works with the smaller scalar type.
12132 EVT ScaleVT = SVT.bitsLT(InnerSVT) ? VT : InnerVT;
12133 EVT ScaleSVT = ScaleVT.getScalarType();
12135 if (TLI.isTypeLegal(ScaleVT) &&
12136 0 == (InnerSVT.getSizeInBits() % ScaleSVT.getSizeInBits()) &&
12137 0 == (SVT.getSizeInBits() % ScaleSVT.getSizeInBits())) {
12139 int InnerScale = InnerSVT.getSizeInBits() / ScaleSVT.getSizeInBits();
12140 int OuterScale = SVT.getSizeInBits() / ScaleSVT.getSizeInBits();
12142 // Scale the shuffle masks to the smaller scalar type.
12143 ShuffleVectorSDNode *InnerSVN = cast<ShuffleVectorSDNode>(BC0);
12144 SmallVector<int, 8> InnerMask =
12145 ScaleShuffleMask(InnerSVN->getMask(), InnerScale);
12146 SmallVector<int, 8> OuterMask =
12147 ScaleShuffleMask(SVN->getMask(), OuterScale);
12149 // Merge the shuffle masks.
12150 SmallVector<int, 8> NewMask;
12151 for (int M : OuterMask)
12152 NewMask.push_back(M < 0 ? -1 : InnerMask[M]);
12154 // Test for shuffle mask legality over both commutations.
12155 SDValue SV0 = BC0->getOperand(0);
12156 SDValue SV1 = BC0->getOperand(1);
12157 bool LegalMask = TLI.isShuffleMaskLegal(NewMask, ScaleVT);
12159 std::swap(SV0, SV1);
12160 ShuffleVectorSDNode::commuteMask(NewMask);
12161 LegalMask = TLI.isShuffleMaskLegal(NewMask, ScaleVT);
12165 SV0 = DAG.getNode(ISD::BITCAST, SDLoc(N), ScaleVT, SV0);
12166 SV1 = DAG.getNode(ISD::BITCAST, SDLoc(N), ScaleVT, SV1);
12167 return DAG.getNode(
12168 ISD::BITCAST, SDLoc(N), VT,
12169 DAG.getVectorShuffle(ScaleVT, SDLoc(N), SV0, SV1, NewMask));
12175 // Canonicalize shuffles according to rules:
12176 // shuffle(A, shuffle(A, B)) -> shuffle(shuffle(A,B), A)
12177 // shuffle(B, shuffle(A, B)) -> shuffle(shuffle(A,B), B)
12178 // shuffle(B, shuffle(A, Undef)) -> shuffle(shuffle(A, Undef), B)
12179 if (N1.getOpcode() == ISD::VECTOR_SHUFFLE &&
12180 N0.getOpcode() != ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
12181 TLI.isTypeLegal(VT)) {
12182 // The incoming shuffle must be of the same type as the result of the
12183 // current shuffle.
12184 assert(N1->getOperand(0).getValueType() == VT &&
12185 "Shuffle types don't match");
12187 SDValue SV0 = N1->getOperand(0);
12188 SDValue SV1 = N1->getOperand(1);
12189 bool HasSameOp0 = N0 == SV0;
12190 bool IsSV1Undef = SV1.getOpcode() == ISD::UNDEF;
12191 if (HasSameOp0 || IsSV1Undef || N0 == SV1)
12192 // Commute the operands of this shuffle so that next rule
12194 return DAG.getCommutedVectorShuffle(*SVN);
12197 // Try to fold according to rules:
12198 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2)
12199 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2)
12200 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2)
12201 // Don't try to fold shuffles with illegal type.
12202 // Only fold if this shuffle is the only user of the other shuffle.
12203 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && N->isOnlyUserOf(N0.getNode()) &&
12204 Level < AfterLegalizeDAG && TLI.isTypeLegal(VT)) {
12205 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
12207 // The incoming shuffle must be of the same type as the result of the
12208 // current shuffle.
12209 assert(OtherSV->getOperand(0).getValueType() == VT &&
12210 "Shuffle types don't match");
12213 SmallVector<int, 4> Mask;
12214 // Compute the combined shuffle mask for a shuffle with SV0 as the first
12215 // operand, and SV1 as the second operand.
12216 for (unsigned i = 0; i != NumElts; ++i) {
12217 int Idx = SVN->getMaskElt(i);
12219 // Propagate Undef.
12220 Mask.push_back(Idx);
12224 SDValue CurrentVec;
12225 if (Idx < (int)NumElts) {
12226 // This shuffle index refers to the inner shuffle N0. Lookup the inner
12227 // shuffle mask to identify which vector is actually referenced.
12228 Idx = OtherSV->getMaskElt(Idx);
12230 // Propagate Undef.
12231 Mask.push_back(Idx);
12235 CurrentVec = (Idx < (int) NumElts) ? OtherSV->getOperand(0)
12236 : OtherSV->getOperand(1);
12238 // This shuffle index references an element within N1.
12242 // Simple case where 'CurrentVec' is UNDEF.
12243 if (CurrentVec.getOpcode() == ISD::UNDEF) {
12244 Mask.push_back(-1);
12248 // Canonicalize the shuffle index. We don't know yet if CurrentVec
12249 // will be the first or second operand of the combined shuffle.
12250 Idx = Idx % NumElts;
12251 if (!SV0.getNode() || SV0 == CurrentVec) {
12252 // Ok. CurrentVec is the left hand side.
12253 // Update the mask accordingly.
12255 Mask.push_back(Idx);
12259 // Bail out if we cannot convert the shuffle pair into a single shuffle.
12260 if (SV1.getNode() && SV1 != CurrentVec)
12263 // Ok. CurrentVec is the right hand side.
12264 // Update the mask accordingly.
12266 Mask.push_back(Idx + NumElts);
12269 // Check if all indices in Mask are Undef. In case, propagate Undef.
12270 bool isUndefMask = true;
12271 for (unsigned i = 0; i != NumElts && isUndefMask; ++i)
12272 isUndefMask &= Mask[i] < 0;
12275 return DAG.getUNDEF(VT);
12277 if (!SV0.getNode())
12278 SV0 = DAG.getUNDEF(VT);
12279 if (!SV1.getNode())
12280 SV1 = DAG.getUNDEF(VT);
12282 // Avoid introducing shuffles with illegal mask.
12283 if (!TLI.isShuffleMaskLegal(Mask, VT)) {
12284 ShuffleVectorSDNode::commuteMask(Mask);
12286 if (!TLI.isShuffleMaskLegal(Mask, VT))
12289 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, A, M2)
12290 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, A, M2)
12291 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, B, M2)
12292 std::swap(SV0, SV1);
12295 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2)
12296 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2)
12297 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2)
12298 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, &Mask[0]);
12304 SDValue DAGCombiner::visitSCALAR_TO_VECTOR(SDNode *N) {
12305 SDValue InVal = N->getOperand(0);
12306 EVT VT = N->getValueType(0);
12308 // Replace a SCALAR_TO_VECTOR(EXTRACT_VECTOR_ELT(V,C0)) pattern
12309 // with a VECTOR_SHUFFLE.
12310 if (InVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
12311 SDValue InVec = InVal->getOperand(0);
12312 SDValue EltNo = InVal->getOperand(1);
12314 // FIXME: We could support implicit truncation if the shuffle can be
12315 // scaled to a smaller vector scalar type.
12316 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(EltNo);
12317 if (C0 && VT == InVec.getValueType() &&
12318 VT.getScalarType() == InVal.getValueType()) {
12319 SmallVector<int, 8> NewMask(VT.getVectorNumElements(), -1);
12320 int Elt = C0->getZExtValue();
12323 if (TLI.isShuffleMaskLegal(NewMask, VT))
12324 return DAG.getVectorShuffle(VT, SDLoc(N), InVec, DAG.getUNDEF(VT),
12332 SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
12333 SDValue N0 = N->getOperand(0);
12334 SDValue N2 = N->getOperand(2);
12336 // If the input vector is a concatenation, and the insert replaces
12337 // one of the halves, we can optimize into a single concat_vectors.
12338 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
12339 N0->getNumOperands() == 2 && N2.getOpcode() == ISD::Constant) {
12340 APInt InsIdx = cast<ConstantSDNode>(N2)->getAPIntValue();
12341 EVT VT = N->getValueType(0);
12343 // Lower half: fold (insert_subvector (concat_vectors X, Y), Z) ->
12344 // (concat_vectors Z, Y)
12346 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
12347 N->getOperand(1), N0.getOperand(1));
12349 // Upper half: fold (insert_subvector (concat_vectors X, Y), Z) ->
12350 // (concat_vectors X, Z)
12351 if (InsIdx == VT.getVectorNumElements()/2)
12352 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
12353 N0.getOperand(0), N->getOperand(1));
12359 SDValue DAGCombiner::visitFP_TO_FP16(SDNode *N) {
12360 SDValue N0 = N->getOperand(0);
12362 // fold (fp_to_fp16 (fp16_to_fp op)) -> op
12363 if (N0->getOpcode() == ISD::FP16_TO_FP)
12364 return N0->getOperand(0);
12369 /// Returns a vector_shuffle if it able to transform an AND to a vector_shuffle
12370 /// with the destination vector and a zero vector.
12371 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
12372 /// vector_shuffle V, Zero, <0, 4, 2, 4>
12373 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
12374 EVT VT = N->getValueType(0);
12375 SDValue LHS = N->getOperand(0);
12376 SDValue RHS = N->getOperand(1);
12379 // Make sure we're not running after operation legalization where it
12380 // may have custom lowered the vector shuffles.
12381 if (LegalOperations)
12384 if (N->getOpcode() != ISD::AND)
12387 if (RHS.getOpcode() == ISD::BITCAST)
12388 RHS = RHS.getOperand(0);
12390 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
12391 SmallVector<int, 8> Indices;
12392 unsigned NumElts = RHS.getNumOperands();
12394 for (unsigned i = 0; i != NumElts; ++i) {
12395 SDValue Elt = RHS.getOperand(i);
12396 if (!isa<ConstantSDNode>(Elt))
12399 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
12400 Indices.push_back(i);
12401 else if (cast<ConstantSDNode>(Elt)->isNullValue())
12402 Indices.push_back(NumElts+i);
12407 // Let's see if the target supports this vector_shuffle.
12408 EVT RVT = RHS.getValueType();
12409 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
12412 // Return the new VECTOR_SHUFFLE node.
12413 EVT EltVT = RVT.getVectorElementType();
12414 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
12415 DAG.getConstant(0, EltVT));
12416 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), RVT, ZeroOps);
12417 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
12418 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
12419 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
12425 /// Visit a binary vector operation, like ADD.
12426 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
12427 assert(N->getValueType(0).isVector() &&
12428 "SimplifyVBinOp only works on vectors!");
12430 SDValue LHS = N->getOperand(0);
12431 SDValue RHS = N->getOperand(1);
12433 if (SDValue Shuffle = XformToShuffleWithZero(N))
12436 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
12438 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
12439 RHS.getOpcode() == ISD::BUILD_VECTOR) {
12440 // Check if both vectors are constants. If not bail out.
12441 if (!(cast<BuildVectorSDNode>(LHS)->isConstant() &&
12442 cast<BuildVectorSDNode>(RHS)->isConstant()))
12445 SmallVector<SDValue, 8> Ops;
12446 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
12447 SDValue LHSOp = LHS.getOperand(i);
12448 SDValue RHSOp = RHS.getOperand(i);
12450 // Can't fold divide by zero.
12451 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
12452 N->getOpcode() == ISD::FDIV) {
12453 if ((RHSOp.getOpcode() == ISD::Constant &&
12454 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
12455 (RHSOp.getOpcode() == ISD::ConstantFP &&
12456 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
12460 EVT VT = LHSOp.getValueType();
12461 EVT RVT = RHSOp.getValueType();
12463 // Integer BUILD_VECTOR operands may have types larger than the element
12464 // size (e.g., when the element type is not legal). Prior to type
12465 // legalization, the types may not match between the two BUILD_VECTORS.
12466 // Truncate one of the operands to make them match.
12467 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
12468 RHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, RHSOp);
12470 LHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), RVT, LHSOp);
12474 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(LHS), VT,
12476 if (FoldOp.getOpcode() != ISD::UNDEF &&
12477 FoldOp.getOpcode() != ISD::Constant &&
12478 FoldOp.getOpcode() != ISD::ConstantFP)
12480 Ops.push_back(FoldOp);
12481 AddToWorklist(FoldOp.getNode());
12484 if (Ops.size() == LHS.getNumOperands())
12485 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), LHS.getValueType(), Ops);
12488 // Type legalization might introduce new shuffles in the DAG.
12489 // Fold (VBinOp (shuffle (A, Undef, Mask)), (shuffle (B, Undef, Mask)))
12490 // -> (shuffle (VBinOp (A, B)), Undef, Mask).
12491 if (LegalTypes && isa<ShuffleVectorSDNode>(LHS) &&
12492 isa<ShuffleVectorSDNode>(RHS) && LHS.hasOneUse() && RHS.hasOneUse() &&
12493 LHS.getOperand(1).getOpcode() == ISD::UNDEF &&
12494 RHS.getOperand(1).getOpcode() == ISD::UNDEF) {
12495 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(LHS);
12496 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(RHS);
12498 if (SVN0->getMask().equals(SVN1->getMask())) {
12499 EVT VT = N->getValueType(0);
12500 SDValue UndefVector = LHS.getOperand(1);
12501 SDValue NewBinOp = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
12502 LHS.getOperand(0), RHS.getOperand(0));
12503 AddUsersToWorklist(N);
12504 return DAG.getVectorShuffle(VT, SDLoc(N), NewBinOp, UndefVector,
12505 &SVN0->getMask()[0]);
12512 SDValue DAGCombiner::SimplifySelect(SDLoc DL, SDValue N0,
12513 SDValue N1, SDValue N2){
12514 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
12516 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
12517 cast<CondCodeSDNode>(N0.getOperand(2))->get());
12519 // If we got a simplified select_cc node back from SimplifySelectCC, then
12520 // break it down into a new SETCC node, and a new SELECT node, and then return
12521 // the SELECT node, since we were called with a SELECT node.
12522 if (SCC.getNode()) {
12523 // Check to see if we got a select_cc back (to turn into setcc/select).
12524 // Otherwise, just return whatever node we got back, like fabs.
12525 if (SCC.getOpcode() == ISD::SELECT_CC) {
12526 SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0),
12528 SCC.getOperand(0), SCC.getOperand(1),
12529 SCC.getOperand(4));
12530 AddToWorklist(SETCC.getNode());
12531 return DAG.getSelect(SDLoc(SCC), SCC.getValueType(), SETCC,
12532 SCC.getOperand(2), SCC.getOperand(3));
12540 /// Given a SELECT or a SELECT_CC node, where LHS and RHS are the two values
12541 /// being selected between, see if we can simplify the select. Callers of this
12542 /// should assume that TheSelect is deleted if this returns true. As such, they
12543 /// should return the appropriate thing (e.g. the node) back to the top-level of
12544 /// the DAG combiner loop to avoid it being looked at.
12545 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
12548 // Cannot simplify select with vector condition
12549 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
12551 // If this is a select from two identical things, try to pull the operation
12552 // through the select.
12553 if (LHS.getOpcode() != RHS.getOpcode() ||
12554 !LHS.hasOneUse() || !RHS.hasOneUse())
12557 // If this is a load and the token chain is identical, replace the select
12558 // of two loads with a load through a select of the address to load from.
12559 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
12560 // constants have been dropped into the constant pool.
12561 if (LHS.getOpcode() == ISD::LOAD) {
12562 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
12563 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
12565 // Token chains must be identical.
12566 if (LHS.getOperand(0) != RHS.getOperand(0) ||
12567 // Do not let this transformation reduce the number of volatile loads.
12568 LLD->isVolatile() || RLD->isVolatile() ||
12569 // If this is an EXTLOAD, the VT's must match.
12570 LLD->getMemoryVT() != RLD->getMemoryVT() ||
12571 // If this is an EXTLOAD, the kind of extension must match.
12572 (LLD->getExtensionType() != RLD->getExtensionType() &&
12573 // The only exception is if one of the extensions is anyext.
12574 LLD->getExtensionType() != ISD::EXTLOAD &&
12575 RLD->getExtensionType() != ISD::EXTLOAD) ||
12576 // FIXME: this discards src value information. This is
12577 // over-conservative. It would be beneficial to be able to remember
12578 // both potential memory locations. Since we are discarding
12579 // src value info, don't do the transformation if the memory
12580 // locations are not in the default address space.
12581 LLD->getPointerInfo().getAddrSpace() != 0 ||
12582 RLD->getPointerInfo().getAddrSpace() != 0 ||
12583 !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(),
12584 LLD->getBasePtr().getValueType()))
12587 // Check that the select condition doesn't reach either load. If so,
12588 // folding this will induce a cycle into the DAG. If not, this is safe to
12589 // xform, so create a select of the addresses.
12591 if (TheSelect->getOpcode() == ISD::SELECT) {
12592 SDNode *CondNode = TheSelect->getOperand(0).getNode();
12593 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
12594 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
12596 // The loads must not depend on one another.
12597 if (LLD->isPredecessorOf(RLD) ||
12598 RLD->isPredecessorOf(LLD))
12600 Addr = DAG.getSelect(SDLoc(TheSelect),
12601 LLD->getBasePtr().getValueType(),
12602 TheSelect->getOperand(0), LLD->getBasePtr(),
12603 RLD->getBasePtr());
12604 } else { // Otherwise SELECT_CC
12605 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
12606 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
12608 if ((LLD->hasAnyUseOfValue(1) &&
12609 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
12610 (RLD->hasAnyUseOfValue(1) &&
12611 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
12614 Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect),
12615 LLD->getBasePtr().getValueType(),
12616 TheSelect->getOperand(0),
12617 TheSelect->getOperand(1),
12618 LLD->getBasePtr(), RLD->getBasePtr(),
12619 TheSelect->getOperand(4));
12623 // It is safe to replace the two loads if they have different alignments,
12624 // but the new load must be the minimum (most restrictive) alignment of the
12626 bool isInvariant = LLD->isInvariant() & RLD->isInvariant();
12627 unsigned Alignment = std::min(LLD->getAlignment(), RLD->getAlignment());
12628 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
12629 Load = DAG.getLoad(TheSelect->getValueType(0),
12631 // FIXME: Discards pointer and AA info.
12632 LLD->getChain(), Addr, MachinePointerInfo(),
12633 LLD->isVolatile(), LLD->isNonTemporal(),
12634 isInvariant, Alignment);
12636 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
12637 RLD->getExtensionType() : LLD->getExtensionType(),
12639 TheSelect->getValueType(0),
12640 // FIXME: Discards pointer and AA info.
12641 LLD->getChain(), Addr, MachinePointerInfo(),
12642 LLD->getMemoryVT(), LLD->isVolatile(),
12643 LLD->isNonTemporal(), isInvariant, Alignment);
12646 // Users of the select now use the result of the load.
12647 CombineTo(TheSelect, Load);
12649 // Users of the old loads now use the new load's chain. We know the
12650 // old-load value is dead now.
12651 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
12652 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
12659 /// Simplify an expression of the form (N0 cond N1) ? N2 : N3
12660 /// where 'cond' is the comparison specified by CC.
12661 SDValue DAGCombiner::SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1,
12662 SDValue N2, SDValue N3,
12663 ISD::CondCode CC, bool NotExtCompare) {
12664 // (x ? y : y) -> y.
12665 if (N2 == N3) return N2;
12667 EVT VT = N2.getValueType();
12668 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
12669 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
12670 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
12672 // Determine if the condition we're dealing with is constant
12673 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
12674 N0, N1, CC, DL, false);
12675 if (SCC.getNode()) AddToWorklist(SCC.getNode());
12676 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
12678 // fold select_cc true, x, y -> x
12679 if (SCCC && !SCCC->isNullValue())
12681 // fold select_cc false, x, y -> y
12682 if (SCCC && SCCC->isNullValue())
12685 // Check to see if we can simplify the select into an fabs node
12686 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
12687 // Allow either -0.0 or 0.0
12688 if (CFP->getValueAPF().isZero()) {
12689 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
12690 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
12691 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
12692 N2 == N3.getOperand(0))
12693 return DAG.getNode(ISD::FABS, DL, VT, N0);
12695 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
12696 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
12697 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
12698 N2.getOperand(0) == N3)
12699 return DAG.getNode(ISD::FABS, DL, VT, N3);
12703 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
12704 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
12705 // in it. This is a win when the constant is not otherwise available because
12706 // it replaces two constant pool loads with one. We only do this if the FP
12707 // type is known to be legal, because if it isn't, then we are before legalize
12708 // types an we want the other legalization to happen first (e.g. to avoid
12709 // messing with soft float) and if the ConstantFP is not legal, because if
12710 // it is legal, we may not need to store the FP constant in a constant pool.
12711 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
12712 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
12713 if (TLI.isTypeLegal(N2.getValueType()) &&
12714 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
12715 TargetLowering::Legal &&
12716 !TLI.isFPImmLegal(TV->getValueAPF(), TV->getValueType(0)) &&
12717 !TLI.isFPImmLegal(FV->getValueAPF(), FV->getValueType(0))) &&
12718 // If both constants have multiple uses, then we won't need to do an
12719 // extra load, they are likely around in registers for other users.
12720 (TV->hasOneUse() || FV->hasOneUse())) {
12721 Constant *Elts[] = {
12722 const_cast<ConstantFP*>(FV->getConstantFPValue()),
12723 const_cast<ConstantFP*>(TV->getConstantFPValue())
12725 Type *FPTy = Elts[0]->getType();
12726 const DataLayout &TD = *TLI.getDataLayout();
12728 // Create a ConstantArray of the two constants.
12729 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
12730 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
12731 TD.getPrefTypeAlignment(FPTy));
12732 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
12734 // Get the offsets to the 0 and 1 element of the array so that we can
12735 // select between them.
12736 SDValue Zero = DAG.getIntPtrConstant(0);
12737 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
12738 SDValue One = DAG.getIntPtrConstant(EltSize);
12740 SDValue Cond = DAG.getSetCC(DL,
12741 getSetCCResultType(N0.getValueType()),
12743 AddToWorklist(Cond.getNode());
12744 SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(),
12746 AddToWorklist(CstOffset.getNode());
12747 CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx,
12749 AddToWorklist(CPIdx.getNode());
12750 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
12751 MachinePointerInfo::getConstantPool(), false,
12752 false, false, Alignment);
12757 // Check to see if we can perform the "gzip trick", transforming
12758 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
12759 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
12760 (N1C->isNullValue() || // (a < 0) ? b : 0
12761 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
12762 EVT XType = N0.getValueType();
12763 EVT AType = N2.getValueType();
12764 if (XType.bitsGE(AType)) {
12765 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
12766 // single-bit constant.
12767 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
12768 unsigned ShCtV = N2C->getAPIntValue().logBase2();
12769 ShCtV = XType.getSizeInBits()-ShCtV-1;
12770 SDValue ShCt = DAG.getConstant(ShCtV,
12771 getShiftAmountTy(N0.getValueType()));
12772 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0),
12774 AddToWorklist(Shift.getNode());
12776 if (XType.bitsGT(AType)) {
12777 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
12778 AddToWorklist(Shift.getNode());
12781 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
12784 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0),
12786 DAG.getConstant(XType.getSizeInBits()-1,
12787 getShiftAmountTy(N0.getValueType())));
12788 AddToWorklist(Shift.getNode());
12790 if (XType.bitsGT(AType)) {
12791 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
12792 AddToWorklist(Shift.getNode());
12795 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
12799 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
12800 // where y is has a single bit set.
12801 // A plaintext description would be, we can turn the SELECT_CC into an AND
12802 // when the condition can be materialized as an all-ones register. Any
12803 // single bit-test can be materialized as an all-ones register with
12804 // shift-left and shift-right-arith.
12805 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
12806 N0->getValueType(0) == VT &&
12807 N1C && N1C->isNullValue() &&
12808 N2C && N2C->isNullValue()) {
12809 SDValue AndLHS = N0->getOperand(0);
12810 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
12811 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
12812 // Shift the tested bit over the sign bit.
12813 APInt AndMask = ConstAndRHS->getAPIntValue();
12815 DAG.getConstant(AndMask.countLeadingZeros(),
12816 getShiftAmountTy(AndLHS.getValueType()));
12817 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt);
12819 // Now arithmetic right shift it all the way over, so the result is either
12820 // all-ones, or zero.
12822 DAG.getConstant(AndMask.getBitWidth()-1,
12823 getShiftAmountTy(Shl.getValueType()));
12824 SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt);
12826 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
12830 // fold select C, 16, 0 -> shl C, 4
12831 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
12832 TLI.getBooleanContents(N0.getValueType()) ==
12833 TargetLowering::ZeroOrOneBooleanContent) {
12835 // If the caller doesn't want us to simplify this into a zext of a compare,
12837 if (NotExtCompare && N2C->getAPIntValue() == 1)
12840 // Get a SetCC of the condition
12841 // NOTE: Don't create a SETCC if it's not legal on this target.
12842 if (!LegalOperations ||
12843 TLI.isOperationLegal(ISD::SETCC,
12844 LegalTypes ? getSetCCResultType(N0.getValueType()) : MVT::i1)) {
12846 // cast from setcc result type to select result type
12848 SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()),
12850 if (N2.getValueType().bitsLT(SCC.getValueType()))
12851 Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2),
12852 N2.getValueType());
12854 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
12855 N2.getValueType(), SCC);
12857 SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC);
12858 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
12859 N2.getValueType(), SCC);
12862 AddToWorklist(SCC.getNode());
12863 AddToWorklist(Temp.getNode());
12865 if (N2C->getAPIntValue() == 1)
12868 // shl setcc result by log2 n2c
12869 return DAG.getNode(
12870 ISD::SHL, DL, N2.getValueType(), Temp,
12871 DAG.getConstant(N2C->getAPIntValue().logBase2(),
12872 getShiftAmountTy(Temp.getValueType())));
12876 // Check to see if this is the equivalent of setcc
12877 // FIXME: Turn all of these into setcc if setcc if setcc is legal
12878 // otherwise, go ahead with the folds.
12879 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
12880 EVT XType = N0.getValueType();
12881 if (!LegalOperations ||
12882 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(XType))) {
12883 SDValue Res = DAG.getSetCC(DL, getSetCCResultType(XType), N0, N1, CC);
12884 if (Res.getValueType() != VT)
12885 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
12889 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
12890 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
12891 (!LegalOperations ||
12892 TLI.isOperationLegal(ISD::CTLZ, XType))) {
12893 SDValue Ctlz = DAG.getNode(ISD::CTLZ, SDLoc(N0), XType, N0);
12894 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
12895 DAG.getConstant(Log2_32(XType.getSizeInBits()),
12896 getShiftAmountTy(Ctlz.getValueType())));
12898 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
12899 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
12900 SDValue NegN0 = DAG.getNode(ISD::SUB, SDLoc(N0),
12901 XType, DAG.getConstant(0, XType), N0);
12902 SDValue NotN0 = DAG.getNOT(SDLoc(N0), N0, XType);
12903 return DAG.getNode(ISD::SRL, DL, XType,
12904 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
12905 DAG.getConstant(XType.getSizeInBits()-1,
12906 getShiftAmountTy(XType)));
12908 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
12909 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
12910 SDValue Sign = DAG.getNode(ISD::SRL, SDLoc(N0), XType, N0,
12911 DAG.getConstant(XType.getSizeInBits()-1,
12912 getShiftAmountTy(N0.getValueType())));
12913 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
12917 // Check to see if this is an integer abs.
12918 // select_cc setg[te] X, 0, X, -X ->
12919 // select_cc setgt X, -1, X, -X ->
12920 // select_cc setl[te] X, 0, -X, X ->
12921 // select_cc setlt X, 1, -X, X ->
12922 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
12924 ConstantSDNode *SubC = nullptr;
12925 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
12926 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
12927 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
12928 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
12929 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
12930 (N1C->isOne() && CC == ISD::SETLT)) &&
12931 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
12932 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
12934 EVT XType = N0.getValueType();
12935 if (SubC && SubC->isNullValue() && XType.isInteger()) {
12936 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0), XType,
12938 DAG.getConstant(XType.getSizeInBits()-1,
12939 getShiftAmountTy(N0.getValueType())));
12940 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0),
12942 AddToWorklist(Shift.getNode());
12943 AddToWorklist(Add.getNode());
12944 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
12951 /// This is a stub for TargetLowering::SimplifySetCC.
12952 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
12953 SDValue N1, ISD::CondCode Cond,
12954 SDLoc DL, bool foldBooleans) {
12955 TargetLowering::DAGCombinerInfo
12956 DagCombineInfo(DAG, Level, false, this);
12957 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
12960 /// Given an ISD::SDIV node expressing a divide by constant, return
12961 /// a DAG expression to select that will generate the same value by multiplying
12962 /// by a magic number.
12963 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
12964 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
12965 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
12969 // Avoid division by zero.
12970 if (!C->getAPIntValue())
12973 std::vector<SDNode*> Built;
12975 TLI.BuildSDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
12977 for (SDNode *N : Built)
12982 /// Given an ISD::SDIV node expressing a divide by constant power of 2, return a
12983 /// DAG expression that will generate the same value by right shifting.
12984 SDValue DAGCombiner::BuildSDIVPow2(SDNode *N) {
12985 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
12989 // Avoid division by zero.
12990 if (!C->getAPIntValue())
12993 std::vector<SDNode *> Built;
12994 SDValue S = TLI.BuildSDIVPow2(N, C->getAPIntValue(), DAG, &Built);
12996 for (SDNode *N : Built)
13001 /// Given an ISD::UDIV node expressing a divide by constant, return a DAG
13002 /// expression that will generate the same value by multiplying by a magic
13004 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
13005 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
13006 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
13010 // Avoid division by zero.
13011 if (!C->getAPIntValue())
13014 std::vector<SDNode*> Built;
13016 TLI.BuildUDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
13018 for (SDNode *N : Built)
13023 SDValue DAGCombiner::BuildReciprocalEstimate(SDValue Op) {
13024 if (Level >= AfterLegalizeDAG)
13027 // Expose the DAG combiner to the target combiner implementations.
13028 TargetLowering::DAGCombinerInfo DCI(DAG, Level, false, this);
13030 unsigned Iterations = 0;
13031 if (SDValue Est = TLI.getRecipEstimate(Op, DCI, Iterations)) {
13033 // Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
13034 // For the reciprocal, we need to find the zero of the function:
13035 // F(X) = A X - 1 [which has a zero at X = 1/A]
13037 // X_{i+1} = X_i (2 - A X_i) = X_i + X_i (1 - A X_i) [this second form
13038 // does not require additional intermediate precision]
13039 EVT VT = Op.getValueType();
13041 SDValue FPOne = DAG.getConstantFP(1.0, VT);
13043 AddToWorklist(Est.getNode());
13045 // Newton iterations: Est = Est + Est (1 - Arg * Est)
13046 for (unsigned i = 0; i < Iterations; ++i) {
13047 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Op, Est);
13048 AddToWorklist(NewEst.getNode());
13050 NewEst = DAG.getNode(ISD::FSUB, DL, VT, FPOne, NewEst);
13051 AddToWorklist(NewEst.getNode());
13053 NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst);
13054 AddToWorklist(NewEst.getNode());
13056 Est = DAG.getNode(ISD::FADD, DL, VT, Est, NewEst);
13057 AddToWorklist(Est.getNode());
13066 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
13067 /// For the reciprocal sqrt, we need to find the zero of the function:
13068 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
13070 /// X_{i+1} = X_i (1.5 - A X_i^2 / 2)
13071 /// As a result, we precompute A/2 prior to the iteration loop.
13072 SDValue DAGCombiner::BuildRsqrtNROneConst(SDValue Arg, SDValue Est,
13073 unsigned Iterations) {
13074 EVT VT = Arg.getValueType();
13076 SDValue ThreeHalves = DAG.getConstantFP(1.5, VT);
13078 // We now need 0.5 * Arg which we can write as (1.5 * Arg - Arg) so that
13079 // this entire sequence requires only one FP constant.
13080 SDValue HalfArg = DAG.getNode(ISD::FMUL, DL, VT, ThreeHalves, Arg);
13081 AddToWorklist(HalfArg.getNode());
13083 HalfArg = DAG.getNode(ISD::FSUB, DL, VT, HalfArg, Arg);
13084 AddToWorklist(HalfArg.getNode());
13086 // Newton iterations: Est = Est * (1.5 - HalfArg * Est * Est)
13087 for (unsigned i = 0; i < Iterations; ++i) {
13088 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, Est);
13089 AddToWorklist(NewEst.getNode());
13091 NewEst = DAG.getNode(ISD::FMUL, DL, VT, HalfArg, NewEst);
13092 AddToWorklist(NewEst.getNode());
13094 NewEst = DAG.getNode(ISD::FSUB, DL, VT, ThreeHalves, NewEst);
13095 AddToWorklist(NewEst.getNode());
13097 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst);
13098 AddToWorklist(Est.getNode());
13103 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
13104 /// For the reciprocal sqrt, we need to find the zero of the function:
13105 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
13107 /// X_{i+1} = (-0.5 * X_i) * (A * X_i * X_i + (-3.0))
13108 SDValue DAGCombiner::BuildRsqrtNRTwoConst(SDValue Arg, SDValue Est,
13109 unsigned Iterations) {
13110 EVT VT = Arg.getValueType();
13112 SDValue MinusThree = DAG.getConstantFP(-3.0, VT);
13113 SDValue MinusHalf = DAG.getConstantFP(-0.5, VT);
13115 // Newton iterations: Est = -0.5 * Est * (-3.0 + Arg * Est * Est)
13116 for (unsigned i = 0; i < Iterations; ++i) {
13117 SDValue HalfEst = DAG.getNode(ISD::FMUL, DL, VT, Est, MinusHalf);
13118 AddToWorklist(HalfEst.getNode());
13120 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Est);
13121 AddToWorklist(Est.getNode());
13123 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Arg);
13124 AddToWorklist(Est.getNode());
13126 Est = DAG.getNode(ISD::FADD, DL, VT, Est, MinusThree);
13127 AddToWorklist(Est.getNode());
13129 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, HalfEst);
13130 AddToWorklist(Est.getNode());
13135 SDValue DAGCombiner::BuildRsqrtEstimate(SDValue Op) {
13136 if (Level >= AfterLegalizeDAG)
13139 // Expose the DAG combiner to the target combiner implementations.
13140 TargetLowering::DAGCombinerInfo DCI(DAG, Level, false, this);
13141 unsigned Iterations = 0;
13142 bool UseOneConstNR = false;
13143 if (SDValue Est = TLI.getRsqrtEstimate(Op, DCI, Iterations, UseOneConstNR)) {
13144 AddToWorklist(Est.getNode());
13146 Est = UseOneConstNR ?
13147 BuildRsqrtNROneConst(Op, Est, Iterations) :
13148 BuildRsqrtNRTwoConst(Op, Est, Iterations);
13156 /// Return true if base is a frame index, which is known not to alias with
13157 /// anything but itself. Provides base object and offset as results.
13158 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
13159 const GlobalValue *&GV, const void *&CV) {
13160 // Assume it is a primitive operation.
13161 Base = Ptr; Offset = 0; GV = nullptr; CV = nullptr;
13163 // If it's an adding a simple constant then integrate the offset.
13164 if (Base.getOpcode() == ISD::ADD) {
13165 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
13166 Base = Base.getOperand(0);
13167 Offset += C->getZExtValue();
13171 // Return the underlying GlobalValue, and update the Offset. Return false
13172 // for GlobalAddressSDNode since the same GlobalAddress may be represented
13173 // by multiple nodes with different offsets.
13174 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
13175 GV = G->getGlobal();
13176 Offset += G->getOffset();
13180 // Return the underlying Constant value, and update the Offset. Return false
13181 // for ConstantSDNodes since the same constant pool entry may be represented
13182 // by multiple nodes with different offsets.
13183 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
13184 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
13185 : (const void *)C->getConstVal();
13186 Offset += C->getOffset();
13189 // If it's any of the following then it can't alias with anything but itself.
13190 return isa<FrameIndexSDNode>(Base);
13193 /// Return true if there is any possibility that the two addresses overlap.
13194 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const {
13195 // If they are the same then they must be aliases.
13196 if (Op0->getBasePtr() == Op1->getBasePtr()) return true;
13198 // If they are both volatile then they cannot be reordered.
13199 if (Op0->isVolatile() && Op1->isVolatile()) return true;
13201 // Gather base node and offset information.
13202 SDValue Base1, Base2;
13203 int64_t Offset1, Offset2;
13204 const GlobalValue *GV1, *GV2;
13205 const void *CV1, *CV2;
13206 bool isFrameIndex1 = FindBaseOffset(Op0->getBasePtr(),
13207 Base1, Offset1, GV1, CV1);
13208 bool isFrameIndex2 = FindBaseOffset(Op1->getBasePtr(),
13209 Base2, Offset2, GV2, CV2);
13211 // If they have a same base address then check to see if they overlap.
13212 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
13213 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
13214 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
13216 // It is possible for different frame indices to alias each other, mostly
13217 // when tail call optimization reuses return address slots for arguments.
13218 // To catch this case, look up the actual index of frame indices to compute
13219 // the real alias relationship.
13220 if (isFrameIndex1 && isFrameIndex2) {
13221 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
13222 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
13223 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
13224 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
13225 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
13228 // Otherwise, if we know what the bases are, and they aren't identical, then
13229 // we know they cannot alias.
13230 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
13233 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
13234 // compared to the size and offset of the access, we may be able to prove they
13235 // do not alias. This check is conservative for now to catch cases created by
13236 // splitting vector types.
13237 if ((Op0->getOriginalAlignment() == Op1->getOriginalAlignment()) &&
13238 (Op0->getSrcValueOffset() != Op1->getSrcValueOffset()) &&
13239 (Op0->getMemoryVT().getSizeInBits() >> 3 ==
13240 Op1->getMemoryVT().getSizeInBits() >> 3) &&
13241 (Op0->getOriginalAlignment() > Op0->getMemoryVT().getSizeInBits()) >> 3) {
13242 int64_t OffAlign1 = Op0->getSrcValueOffset() % Op0->getOriginalAlignment();
13243 int64_t OffAlign2 = Op1->getSrcValueOffset() % Op1->getOriginalAlignment();
13245 // There is no overlap between these relatively aligned accesses of similar
13246 // size, return no alias.
13247 if ((OffAlign1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign2 ||
13248 (OffAlign2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign1)
13252 bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0
13254 : DAG.getSubtarget().useAA();
13256 if (CombinerAAOnlyFunc.getNumOccurrences() &&
13257 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
13261 Op0->getMemOperand()->getValue() && Op1->getMemOperand()->getValue()) {
13262 // Use alias analysis information.
13263 int64_t MinOffset = std::min(Op0->getSrcValueOffset(),
13264 Op1->getSrcValueOffset());
13265 int64_t Overlap1 = (Op0->getMemoryVT().getSizeInBits() >> 3) +
13266 Op0->getSrcValueOffset() - MinOffset;
13267 int64_t Overlap2 = (Op1->getMemoryVT().getSizeInBits() >> 3) +
13268 Op1->getSrcValueOffset() - MinOffset;
13269 AliasAnalysis::AliasResult AAResult =
13270 AA.alias(AliasAnalysis::Location(Op0->getMemOperand()->getValue(),
13272 UseTBAA ? Op0->getAAInfo() : AAMDNodes()),
13273 AliasAnalysis::Location(Op1->getMemOperand()->getValue(),
13275 UseTBAA ? Op1->getAAInfo() : AAMDNodes()));
13276 if (AAResult == AliasAnalysis::NoAlias)
13280 // Otherwise we have to assume they alias.
13284 /// Walk up chain skipping non-aliasing memory nodes,
13285 /// looking for aliasing nodes and adding them to the Aliases vector.
13286 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
13287 SmallVectorImpl<SDValue> &Aliases) {
13288 SmallVector<SDValue, 8> Chains; // List of chains to visit.
13289 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
13291 // Get alias information for node.
13292 bool IsLoad = isa<LoadSDNode>(N) && !cast<LSBaseSDNode>(N)->isVolatile();
13295 Chains.push_back(OriginalChain);
13296 unsigned Depth = 0;
13298 // Look at each chain and determine if it is an alias. If so, add it to the
13299 // aliases list. If not, then continue up the chain looking for the next
13301 while (!Chains.empty()) {
13302 SDValue Chain = Chains.back();
13305 // For TokenFactor nodes, look at each operand and only continue up the
13306 // chain until we find two aliases. If we've seen two aliases, assume we'll
13307 // find more and revert to original chain since the xform is unlikely to be
13310 // FIXME: The depth check could be made to return the last non-aliasing
13311 // chain we found before we hit a tokenfactor rather than the original
13313 if (Depth > 6 || Aliases.size() == 2) {
13315 Aliases.push_back(OriginalChain);
13319 // Don't bother if we've been before.
13320 if (!Visited.insert(Chain.getNode()).second)
13323 switch (Chain.getOpcode()) {
13324 case ISD::EntryToken:
13325 // Entry token is ideal chain operand, but handled in FindBetterChain.
13330 // Get alias information for Chain.
13331 bool IsOpLoad = isa<LoadSDNode>(Chain.getNode()) &&
13332 !cast<LSBaseSDNode>(Chain.getNode())->isVolatile();
13334 // If chain is alias then stop here.
13335 if (!(IsLoad && IsOpLoad) &&
13336 isAlias(cast<LSBaseSDNode>(N), cast<LSBaseSDNode>(Chain.getNode()))) {
13337 Aliases.push_back(Chain);
13339 // Look further up the chain.
13340 Chains.push_back(Chain.getOperand(0));
13346 case ISD::TokenFactor:
13347 // We have to check each of the operands of the token factor for "small"
13348 // token factors, so we queue them up. Adding the operands to the queue
13349 // (stack) in reverse order maintains the original order and increases the
13350 // likelihood that getNode will find a matching token factor (CSE.)
13351 if (Chain.getNumOperands() > 16) {
13352 Aliases.push_back(Chain);
13355 for (unsigned n = Chain.getNumOperands(); n;)
13356 Chains.push_back(Chain.getOperand(--n));
13361 // For all other instructions we will just have to take what we can get.
13362 Aliases.push_back(Chain);
13367 // We need to be careful here to also search for aliases through the
13368 // value operand of a store, etc. Consider the following situation:
13370 // L1 = load Token1, %52
13371 // S1 = store Token1, L1, %51
13372 // L2 = load Token1, %52+8
13373 // S2 = store Token1, L2, %51+8
13374 // Token2 = Token(S1, S2)
13375 // L3 = load Token2, %53
13376 // S3 = store Token2, L3, %52
13377 // L4 = load Token2, %53+8
13378 // S4 = store Token2, L4, %52+8
13379 // If we search for aliases of S3 (which loads address %52), and we look
13380 // only through the chain, then we'll miss the trivial dependence on L1
13381 // (which also loads from %52). We then might change all loads and
13382 // stores to use Token1 as their chain operand, which could result in
13383 // copying %53 into %52 before copying %52 into %51 (which should
13386 // The problem is, however, that searching for such data dependencies
13387 // can become expensive, and the cost is not directly related to the
13388 // chain depth. Instead, we'll rule out such configurations here by
13389 // insisting that we've visited all chain users (except for users
13390 // of the original chain, which is not necessary). When doing this,
13391 // we need to look through nodes we don't care about (otherwise, things
13392 // like register copies will interfere with trivial cases).
13394 SmallVector<const SDNode *, 16> Worklist;
13395 for (const SDNode *N : Visited)
13396 if (N != OriginalChain.getNode())
13397 Worklist.push_back(N);
13399 while (!Worklist.empty()) {
13400 const SDNode *M = Worklist.pop_back_val();
13402 // We have already visited M, and want to make sure we've visited any uses
13403 // of M that we care about. For uses that we've not visisted, and don't
13404 // care about, queue them to the worklist.
13406 for (SDNode::use_iterator UI = M->use_begin(),
13407 UIE = M->use_end(); UI != UIE; ++UI)
13408 if (UI.getUse().getValueType() == MVT::Other &&
13409 Visited.insert(*UI).second) {
13410 if (isa<MemIntrinsicSDNode>(*UI) || isa<MemSDNode>(*UI)) {
13411 // We've not visited this use, and we care about it (it could have an
13412 // ordering dependency with the original node).
13414 Aliases.push_back(OriginalChain);
13418 // We've not visited this use, but we don't care about it. Mark it as
13419 // visited and enqueue it to the worklist.
13420 Worklist.push_back(*UI);
13425 /// Walk up chain skipping non-aliasing memory nodes, looking for a better chain
13426 /// (aliasing node.)
13427 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
13428 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
13430 // Accumulate all the aliases to this node.
13431 GatherAllAliases(N, OldChain, Aliases);
13433 // If no operands then chain to entry token.
13434 if (Aliases.size() == 0)
13435 return DAG.getEntryNode();
13437 // If a single operand then chain to it. We don't need to revisit it.
13438 if (Aliases.size() == 1)
13441 // Construct a custom tailored token factor.
13442 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Aliases);
13445 /// This is the entry point for the file.
13446 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
13447 CodeGenOpt::Level OptLevel) {
13448 /// This is the main entry point to this class.
13449 DAGCombiner(*this, AA, OptLevel).Run(Level);