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);
311 SDValue XformToShuffleWithZero(SDNode *N);
312 SDValue ReassociateOps(unsigned Opc, SDLoc DL, SDValue LHS, SDValue RHS);
314 SDValue visitShiftByConstant(SDNode *N, ConstantSDNode *Amt);
316 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
317 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
318 SDValue SimplifySelect(SDLoc DL, SDValue N0, SDValue N1, SDValue N2);
319 SDValue SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1, SDValue N2,
320 SDValue N3, ISD::CondCode CC,
321 bool NotExtCompare = false);
322 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
323 SDLoc DL, bool foldBooleans = true);
325 bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
327 bool isOneUseSetCC(SDValue N) const;
329 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
331 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
332 SDValue CombineExtLoad(SDNode *N);
333 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
334 SDValue BuildSDIV(SDNode *N);
335 SDValue BuildSDIVPow2(SDNode *N);
336 SDValue BuildUDIV(SDNode *N);
337 SDValue BuildReciprocalEstimate(SDValue Op);
338 SDValue BuildRsqrtEstimate(SDValue Op);
339 SDValue BuildRsqrtNROneConst(SDValue Op, SDValue Est, unsigned Iterations);
340 SDValue BuildRsqrtNRTwoConst(SDValue Op, SDValue Est, unsigned Iterations);
341 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
342 bool DemandHighBits = true);
343 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
344 SDNode *MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg,
345 SDValue InnerPos, SDValue InnerNeg,
346 unsigned PosOpcode, unsigned NegOpcode,
348 SDNode *MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL);
349 SDValue ReduceLoadWidth(SDNode *N);
350 SDValue ReduceLoadOpStoreWidth(SDNode *N);
351 SDValue TransformFPLoadStorePair(SDNode *N);
352 SDValue reduceBuildVecExtToExtBuildVec(SDNode *N);
353 SDValue reduceBuildVecConvertToConvertBuildVec(SDNode *N);
355 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
357 /// Walk up chain skipping non-aliasing memory nodes,
358 /// looking for aliasing nodes and adding them to the Aliases vector.
359 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
360 SmallVectorImpl<SDValue> &Aliases);
362 /// Return true if there is any possibility that the two addresses overlap.
363 bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const;
365 /// Walk up chain skipping non-aliasing memory nodes, looking for a better
366 /// chain (aliasing node.)
367 SDValue FindBetterChain(SDNode *N, SDValue Chain);
369 /// Holds a pointer to an LSBaseSDNode as well as information on where it
370 /// is located in a sequence of memory operations connected by a chain.
372 MemOpLink (LSBaseSDNode *N, int64_t Offset, unsigned Seq):
373 MemNode(N), OffsetFromBase(Offset), SequenceNum(Seq) { }
374 // Ptr to the mem node.
375 LSBaseSDNode *MemNode;
376 // Offset from the base ptr.
377 int64_t OffsetFromBase;
378 // What is the sequence number of this mem node.
379 // Lowest mem operand in the DAG starts at zero.
380 unsigned SequenceNum;
383 /// This is a helper function for MergeConsecutiveStores. When the source
384 /// elements of the consecutive stores are all constants or all extracted
385 /// vector elements, try to merge them into one larger store.
386 /// \return True if a merged store was created.
387 bool MergeStoresOfConstantsOrVecElts(SmallVectorImpl<MemOpLink> &StoreNodes,
388 EVT MemVT, unsigned NumElem,
389 bool IsConstantSrc, bool UseVector);
391 /// Merge consecutive store operations into a wide store.
392 /// This optimization uses wide integers or vectors when possible.
393 /// \return True if some memory operations were changed.
394 bool MergeConsecutiveStores(StoreSDNode *N);
396 /// \brief Try to transform a truncation where C is a constant:
397 /// (trunc (and X, C)) -> (and (trunc X), (trunc C))
399 /// \p N needs to be a truncation and its first operand an AND. Other
400 /// requirements are checked by the function (e.g. that trunc is
401 /// single-use) and if missed an empty SDValue is returned.
402 SDValue distributeTruncateThroughAnd(SDNode *N);
405 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
406 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
407 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {
408 auto *F = DAG.getMachineFunction().getFunction();
409 ForCodeSize = F->hasFnAttribute(Attribute::OptimizeForSize) ||
410 F->hasFnAttribute(Attribute::MinSize);
413 /// Runs the dag combiner on all nodes in the work list
414 void Run(CombineLevel AtLevel);
416 SelectionDAG &getDAG() const { return DAG; }
418 /// Returns a type large enough to hold any valid shift amount - before type
419 /// legalization these can be huge.
420 EVT getShiftAmountTy(EVT LHSTy) {
421 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
422 if (LHSTy.isVector())
424 return LegalTypes ? TLI.getScalarShiftAmountTy(LHSTy)
425 : TLI.getPointerTy();
428 /// This method returns true if we are running before type legalization or
429 /// if the specified VT is legal.
430 bool isTypeLegal(const EVT &VT) {
431 if (!LegalTypes) return true;
432 return TLI.isTypeLegal(VT);
435 /// Convenience wrapper around TargetLowering::getSetCCResultType
436 EVT getSetCCResultType(EVT VT) const {
437 return TLI.getSetCCResultType(*DAG.getContext(), VT);
444 /// This class is a DAGUpdateListener that removes any deleted
445 /// nodes from the worklist.
446 class WorklistRemover : public SelectionDAG::DAGUpdateListener {
449 explicit WorklistRemover(DAGCombiner &dc)
450 : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {}
452 void NodeDeleted(SDNode *N, SDNode *E) override {
453 DC.removeFromWorklist(N);
458 //===----------------------------------------------------------------------===//
459 // TargetLowering::DAGCombinerInfo implementation
460 //===----------------------------------------------------------------------===//
462 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
463 ((DAGCombiner*)DC)->AddToWorklist(N);
466 void TargetLowering::DAGCombinerInfo::RemoveFromWorklist(SDNode *N) {
467 ((DAGCombiner*)DC)->removeFromWorklist(N);
470 SDValue TargetLowering::DAGCombinerInfo::
471 CombineTo(SDNode *N, ArrayRef<SDValue> To, bool AddTo) {
472 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
475 SDValue TargetLowering::DAGCombinerInfo::
476 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
477 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
481 SDValue TargetLowering::DAGCombinerInfo::
482 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
483 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
486 void TargetLowering::DAGCombinerInfo::
487 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
488 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
491 //===----------------------------------------------------------------------===//
493 //===----------------------------------------------------------------------===//
495 void DAGCombiner::deleteAndRecombine(SDNode *N) {
496 removeFromWorklist(N);
498 // If the operands of this node are only used by the node, they will now be
499 // dead. Make sure to re-visit them and recursively delete dead nodes.
500 for (const SDValue &Op : N->ops())
501 // For an operand generating multiple values, one of the values may
502 // become dead allowing further simplification (e.g. split index
503 // arithmetic from an indexed load).
504 if (Op->hasOneUse() || Op->getNumValues() > 1)
505 AddToWorklist(Op.getNode());
510 /// Return 1 if we can compute the negated form of the specified expression for
511 /// the same cost as the expression itself, or 2 if we can compute the negated
512 /// form more cheaply than the expression itself.
513 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
514 const TargetLowering &TLI,
515 const TargetOptions *Options,
516 unsigned Depth = 0) {
517 // fneg is removable even if it has multiple uses.
518 if (Op.getOpcode() == ISD::FNEG) return 2;
520 // Don't allow anything with multiple uses.
521 if (!Op.hasOneUse()) return 0;
523 // Don't recurse exponentially.
524 if (Depth > 6) return 0;
526 switch (Op.getOpcode()) {
527 default: return false;
528 case ISD::ConstantFP:
529 // Don't invert constant FP values after legalize. The negated constant
530 // isn't necessarily legal.
531 return LegalOperations ? 0 : 1;
533 // FIXME: determine better conditions for this xform.
534 if (!Options->UnsafeFPMath) return 0;
536 // After operation legalization, it might not be legal to create new FSUBs.
537 if (LegalOperations &&
538 !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType()))
541 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
542 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
545 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
546 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
549 // We can't turn -(A-B) into B-A when we honor signed zeros.
550 if (!Options->UnsafeFPMath) return 0;
552 // fold (fneg (fsub A, B)) -> (fsub B, A)
557 if (Options->HonorSignDependentRoundingFPMath()) return 0;
559 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
560 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
564 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
570 return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options,
575 /// If isNegatibleForFree returns true, return the newly negated expression.
576 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
577 bool LegalOperations, unsigned Depth = 0) {
578 const TargetOptions &Options = DAG.getTarget().Options;
579 // fneg is removable even if it has multiple uses.
580 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
582 // Don't allow anything with multiple uses.
583 assert(Op.hasOneUse() && "Unknown reuse!");
585 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
586 switch (Op.getOpcode()) {
587 default: llvm_unreachable("Unknown code");
588 case ISD::ConstantFP: {
589 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
591 return DAG.getConstantFP(V, Op.getValueType());
594 // FIXME: determine better conditions for this xform.
595 assert(Options.UnsafeFPMath);
597 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
598 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
599 DAG.getTargetLoweringInfo(), &Options, Depth+1))
600 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
601 GetNegatedExpression(Op.getOperand(0), DAG,
602 LegalOperations, Depth+1),
604 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
605 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
606 GetNegatedExpression(Op.getOperand(1), DAG,
607 LegalOperations, Depth+1),
610 // We can't turn -(A-B) into B-A when we honor signed zeros.
611 assert(Options.UnsafeFPMath);
613 // fold (fneg (fsub 0, B)) -> B
614 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
615 if (N0CFP->getValueAPF().isZero())
616 return Op.getOperand(1);
618 // fold (fneg (fsub A, B)) -> (fsub B, A)
619 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
620 Op.getOperand(1), Op.getOperand(0));
624 assert(!Options.HonorSignDependentRoundingFPMath());
626 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
627 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
628 DAG.getTargetLoweringInfo(), &Options, Depth+1))
629 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
630 GetNegatedExpression(Op.getOperand(0), DAG,
631 LegalOperations, Depth+1),
634 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
635 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
637 GetNegatedExpression(Op.getOperand(1), DAG,
638 LegalOperations, Depth+1));
642 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
643 GetNegatedExpression(Op.getOperand(0), DAG,
644 LegalOperations, Depth+1));
646 return DAG.getNode(ISD::FP_ROUND, SDLoc(Op), Op.getValueType(),
647 GetNegatedExpression(Op.getOperand(0), DAG,
648 LegalOperations, Depth+1),
653 // Return true if this node is a setcc, or is a select_cc
654 // that selects between the target values used for true and false, making it
655 // equivalent to a setcc. Also, set the incoming LHS, RHS, and CC references to
656 // the appropriate nodes based on the type of node we are checking. This
657 // simplifies life a bit for the callers.
658 bool DAGCombiner::isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
660 if (N.getOpcode() == ISD::SETCC) {
661 LHS = N.getOperand(0);
662 RHS = N.getOperand(1);
663 CC = N.getOperand(2);
667 if (N.getOpcode() != ISD::SELECT_CC ||
668 !TLI.isConstTrueVal(N.getOperand(2).getNode()) ||
669 !TLI.isConstFalseVal(N.getOperand(3).getNode()))
672 if (TLI.getBooleanContents(N.getValueType()) ==
673 TargetLowering::UndefinedBooleanContent)
676 LHS = N.getOperand(0);
677 RHS = N.getOperand(1);
678 CC = N.getOperand(4);
682 /// Return true if this is a SetCC-equivalent operation with only one use.
683 /// If this is true, it allows the users to invert the operation for free when
684 /// it is profitable to do so.
685 bool DAGCombiner::isOneUseSetCC(SDValue N) const {
687 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
692 /// Returns true if N is a BUILD_VECTOR node whose
693 /// elements are all the same constant or undefined.
694 static bool isConstantSplatVector(SDNode *N, APInt& SplatValue) {
695 BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(N);
700 unsigned SplatBitSize;
702 EVT EltVT = N->getValueType(0).getVectorElementType();
703 return (C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
705 EltVT.getSizeInBits() >= SplatBitSize);
708 // \brief Returns the SDNode if it is a constant integer BuildVector
709 // or constant integer.
710 static SDNode *isConstantIntBuildVectorOrConstantInt(SDValue N) {
711 if (isa<ConstantSDNode>(N))
713 if (ISD::isBuildVectorOfConstantSDNodes(N.getNode()))
718 // \brief Returns the SDNode if it is a constant float BuildVector
719 // or constant float.
720 static SDNode *isConstantFPBuildVectorOrConstantFP(SDValue N) {
721 if (isa<ConstantFPSDNode>(N))
723 if (ISD::isBuildVectorOfConstantFPSDNodes(N.getNode()))
728 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
730 static ConstantSDNode *isConstOrConstSplat(SDValue N) {
731 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N))
734 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
735 BitVector UndefElements;
736 ConstantSDNode *CN = BV->getConstantSplatNode(&UndefElements);
738 // BuildVectors can truncate their operands. Ignore that case here.
739 // FIXME: We blindly ignore splats which include undef which is overly
741 if (CN && UndefElements.none() &&
742 CN->getValueType(0) == N.getValueType().getScalarType())
749 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
751 static ConstantFPSDNode *isConstOrConstSplatFP(SDValue N) {
752 if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N))
755 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
756 BitVector UndefElements;
757 ConstantFPSDNode *CN = BV->getConstantFPSplatNode(&UndefElements);
759 if (CN && UndefElements.none())
766 SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL,
767 SDValue N0, SDValue N1) {
768 EVT VT = N0.getValueType();
769 if (N0.getOpcode() == Opc) {
770 if (SDNode *L = isConstantIntBuildVectorOrConstantInt(N0.getOperand(1))) {
771 if (SDNode *R = isConstantIntBuildVectorOrConstantInt(N1)) {
772 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
773 if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, L, R))
774 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
777 if (N0.hasOneUse()) {
778 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one
780 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1);
781 if (!OpNode.getNode())
783 AddToWorklist(OpNode.getNode());
784 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
789 if (N1.getOpcode() == Opc) {
790 if (SDNode *R = isConstantIntBuildVectorOrConstantInt(N1.getOperand(1))) {
791 if (SDNode *L = isConstantIntBuildVectorOrConstantInt(N0)) {
792 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
793 if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, R, L))
794 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
797 if (N1.hasOneUse()) {
798 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one
800 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N1.getOperand(0), N0);
801 if (!OpNode.getNode())
803 AddToWorklist(OpNode.getNode());
804 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
812 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
814 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
816 DEBUG(dbgs() << "\nReplacing.1 ";
818 dbgs() << "\nWith: ";
819 To[0].getNode()->dump(&DAG);
820 dbgs() << " and " << NumTo-1 << " other values\n");
821 for (unsigned i = 0, e = NumTo; i != e; ++i)
822 assert((!To[i].getNode() ||
823 N->getValueType(i) == To[i].getValueType()) &&
824 "Cannot combine value to value of different type!");
826 WorklistRemover DeadNodes(*this);
827 DAG.ReplaceAllUsesWith(N, To);
829 // Push the new nodes and any users onto the worklist
830 for (unsigned i = 0, e = NumTo; i != e; ++i) {
831 if (To[i].getNode()) {
832 AddToWorklist(To[i].getNode());
833 AddUsersToWorklist(To[i].getNode());
838 // Finally, if the node is now dead, remove it from the graph. The node
839 // may not be dead if the replacement process recursively simplified to
840 // something else needing this node.
842 deleteAndRecombine(N);
843 return SDValue(N, 0);
847 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
848 // Replace all uses. If any nodes become isomorphic to other nodes and
849 // are deleted, make sure to remove them from our worklist.
850 WorklistRemover DeadNodes(*this);
851 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New);
853 // Push the new node and any (possibly new) users onto the worklist.
854 AddToWorklist(TLO.New.getNode());
855 AddUsersToWorklist(TLO.New.getNode());
857 // Finally, if the node is now dead, remove it from the graph. The node
858 // may not be dead if the replacement process recursively simplified to
859 // something else needing this node.
860 if (TLO.Old.getNode()->use_empty())
861 deleteAndRecombine(TLO.Old.getNode());
864 /// Check the specified integer node value to see if it can be simplified or if
865 /// things it uses can be simplified by bit propagation. If so, return true.
866 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
867 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
868 APInt KnownZero, KnownOne;
869 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
873 AddToWorklist(Op.getNode());
875 // Replace the old value with the new one.
877 DEBUG(dbgs() << "\nReplacing.2 ";
878 TLO.Old.getNode()->dump(&DAG);
879 dbgs() << "\nWith: ";
880 TLO.New.getNode()->dump(&DAG);
883 CommitTargetLoweringOpt(TLO);
887 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
889 EVT VT = Load->getValueType(0);
890 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
892 DEBUG(dbgs() << "\nReplacing.9 ";
894 dbgs() << "\nWith: ";
895 Trunc.getNode()->dump(&DAG);
897 WorklistRemover DeadNodes(*this);
898 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc);
899 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1));
900 deleteAndRecombine(Load);
901 AddToWorklist(Trunc.getNode());
904 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
907 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
908 EVT MemVT = LD->getMemoryVT();
909 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
910 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, PVT, MemVT) ? ISD::ZEXTLOAD
912 : LD->getExtensionType();
914 return DAG.getExtLoad(ExtType, dl, PVT,
915 LD->getChain(), LD->getBasePtr(),
916 MemVT, LD->getMemOperand());
919 unsigned Opc = Op.getOpcode();
922 case ISD::AssertSext:
923 return DAG.getNode(ISD::AssertSext, dl, PVT,
924 SExtPromoteOperand(Op.getOperand(0), PVT),
926 case ISD::AssertZext:
927 return DAG.getNode(ISD::AssertZext, dl, PVT,
928 ZExtPromoteOperand(Op.getOperand(0), PVT),
930 case ISD::Constant: {
932 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
933 return DAG.getNode(ExtOpc, dl, PVT, Op);
937 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
939 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
942 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
943 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
945 EVT OldVT = Op.getValueType();
947 bool Replace = false;
948 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
949 if (!NewOp.getNode())
951 AddToWorklist(NewOp.getNode());
954 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
955 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
956 DAG.getValueType(OldVT));
959 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
960 EVT OldVT = Op.getValueType();
962 bool Replace = false;
963 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
964 if (!NewOp.getNode())
966 AddToWorklist(NewOp.getNode());
969 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
970 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
973 /// Promote the specified integer binary operation if the target indicates it is
974 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
975 /// i32 since i16 instructions are longer.
976 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
977 if (!LegalOperations)
980 EVT VT = Op.getValueType();
981 if (VT.isVector() || !VT.isInteger())
984 // If operation type is 'undesirable', e.g. i16 on x86, consider
986 unsigned Opc = Op.getOpcode();
987 if (TLI.isTypeDesirableForOp(Opc, VT))
991 // Consult target whether it is a good idea to promote this operation and
992 // what's the right type to promote it to.
993 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
994 assert(PVT != VT && "Don't know what type to promote to!");
996 bool Replace0 = false;
997 SDValue N0 = Op.getOperand(0);
998 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
1002 bool Replace1 = false;
1003 SDValue N1 = Op.getOperand(1);
1008 NN1 = PromoteOperand(N1, PVT, Replace1);
1013 AddToWorklist(NN0.getNode());
1015 AddToWorklist(NN1.getNode());
1018 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
1020 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
1022 DEBUG(dbgs() << "\nPromoting ";
1023 Op.getNode()->dump(&DAG));
1025 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1026 DAG.getNode(Opc, dl, PVT, NN0, NN1));
1031 /// Promote the specified integer shift operation if the target indicates it is
1032 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
1033 /// i32 since i16 instructions are longer.
1034 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
1035 if (!LegalOperations)
1038 EVT VT = Op.getValueType();
1039 if (VT.isVector() || !VT.isInteger())
1042 // If operation type is 'undesirable', e.g. i16 on x86, consider
1044 unsigned Opc = Op.getOpcode();
1045 if (TLI.isTypeDesirableForOp(Opc, VT))
1049 // Consult target whether it is a good idea to promote this operation and
1050 // what's the right type to promote it to.
1051 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1052 assert(PVT != VT && "Don't know what type to promote to!");
1054 bool Replace = false;
1055 SDValue N0 = Op.getOperand(0);
1056 if (Opc == ISD::SRA)
1057 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
1058 else if (Opc == ISD::SRL)
1059 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
1061 N0 = PromoteOperand(N0, PVT, Replace);
1065 AddToWorklist(N0.getNode());
1067 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
1069 DEBUG(dbgs() << "\nPromoting ";
1070 Op.getNode()->dump(&DAG));
1072 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1073 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
1078 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
1079 if (!LegalOperations)
1082 EVT VT = Op.getValueType();
1083 if (VT.isVector() || !VT.isInteger())
1086 // If operation type is 'undesirable', e.g. i16 on x86, consider
1088 unsigned Opc = Op.getOpcode();
1089 if (TLI.isTypeDesirableForOp(Opc, VT))
1093 // Consult target whether it is a good idea to promote this operation and
1094 // what's the right type to promote it to.
1095 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1096 assert(PVT != VT && "Don't know what type to promote to!");
1097 // fold (aext (aext x)) -> (aext x)
1098 // fold (aext (zext x)) -> (zext x)
1099 // fold (aext (sext x)) -> (sext x)
1100 DEBUG(dbgs() << "\nPromoting ";
1101 Op.getNode()->dump(&DAG));
1102 return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0));
1107 bool DAGCombiner::PromoteLoad(SDValue Op) {
1108 if (!LegalOperations)
1111 EVT VT = Op.getValueType();
1112 if (VT.isVector() || !VT.isInteger())
1115 // If operation type is 'undesirable', e.g. i16 on x86, consider
1117 unsigned Opc = Op.getOpcode();
1118 if (TLI.isTypeDesirableForOp(Opc, VT))
1122 // Consult target whether it is a good idea to promote this operation and
1123 // what's the right type to promote it to.
1124 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1125 assert(PVT != VT && "Don't know what type to promote to!");
1128 SDNode *N = Op.getNode();
1129 LoadSDNode *LD = cast<LoadSDNode>(N);
1130 EVT MemVT = LD->getMemoryVT();
1131 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
1132 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, PVT, MemVT) ? ISD::ZEXTLOAD
1134 : LD->getExtensionType();
1135 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
1136 LD->getChain(), LD->getBasePtr(),
1137 MemVT, LD->getMemOperand());
1138 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
1140 DEBUG(dbgs() << "\nPromoting ";
1143 Result.getNode()->dump(&DAG);
1145 WorklistRemover DeadNodes(*this);
1146 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
1147 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1));
1148 deleteAndRecombine(N);
1149 AddToWorklist(Result.getNode());
1155 /// \brief Recursively delete a node which has no uses and any operands for
1156 /// which it is the only use.
1158 /// Note that this both deletes the nodes and removes them from the worklist.
1159 /// It also adds any nodes who have had a user deleted to the worklist as they
1160 /// may now have only one use and subject to other combines.
1161 bool DAGCombiner::recursivelyDeleteUnusedNodes(SDNode *N) {
1162 if (!N->use_empty())
1165 SmallSetVector<SDNode *, 16> Nodes;
1168 N = Nodes.pop_back_val();
1172 if (N->use_empty()) {
1173 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1174 Nodes.insert(N->getOperand(i).getNode());
1176 removeFromWorklist(N);
1181 } while (!Nodes.empty());
1185 //===----------------------------------------------------------------------===//
1186 // Main DAG Combiner implementation
1187 //===----------------------------------------------------------------------===//
1189 void DAGCombiner::Run(CombineLevel AtLevel) {
1190 // set the instance variables, so that the various visit routines may use it.
1192 LegalOperations = Level >= AfterLegalizeVectorOps;
1193 LegalTypes = Level >= AfterLegalizeTypes;
1195 // Add all the dag nodes to the worklist.
1196 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
1197 E = DAG.allnodes_end(); I != E; ++I)
1200 // Create a dummy node (which is not added to allnodes), that adds a reference
1201 // to the root node, preventing it from being deleted, and tracking any
1202 // changes of the root.
1203 HandleSDNode Dummy(DAG.getRoot());
1205 // while the worklist isn't empty, find a node and
1206 // try and combine it.
1207 while (!WorklistMap.empty()) {
1209 // The Worklist holds the SDNodes in order, but it may contain null entries.
1211 N = Worklist.pop_back_val();
1214 bool GoodWorklistEntry = WorklistMap.erase(N);
1215 (void)GoodWorklistEntry;
1216 assert(GoodWorklistEntry &&
1217 "Found a worklist entry without a corresponding map entry!");
1219 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1220 // N is deleted from the DAG, since they too may now be dead or may have a
1221 // reduced number of uses, allowing other xforms.
1222 if (recursivelyDeleteUnusedNodes(N))
1225 WorklistRemover DeadNodes(*this);
1227 // If this combine is running after legalizing the DAG, re-legalize any
1228 // nodes pulled off the worklist.
1229 if (Level == AfterLegalizeDAG) {
1230 SmallSetVector<SDNode *, 16> UpdatedNodes;
1231 bool NIsValid = DAG.LegalizeOp(N, UpdatedNodes);
1233 for (SDNode *LN : UpdatedNodes) {
1235 AddUsersToWorklist(LN);
1241 DEBUG(dbgs() << "\nCombining: "; N->dump(&DAG));
1243 // Add any operands of the new node which have not yet been combined to the
1244 // worklist as well. Because the worklist uniques things already, this
1245 // won't repeatedly process the same operand.
1246 CombinedNodes.insert(N);
1247 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1248 if (!CombinedNodes.count(N->getOperand(i).getNode()))
1249 AddToWorklist(N->getOperand(i).getNode());
1251 SDValue RV = combine(N);
1258 // If we get back the same node we passed in, rather than a new node or
1259 // zero, we know that the node must have defined multiple values and
1260 // CombineTo was used. Since CombineTo takes care of the worklist
1261 // mechanics for us, we have no work to do in this case.
1262 if (RV.getNode() == N)
1265 assert(N->getOpcode() != ISD::DELETED_NODE &&
1266 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1267 "Node was deleted but visit returned new node!");
1269 DEBUG(dbgs() << " ... into: ";
1270 RV.getNode()->dump(&DAG));
1272 // Transfer debug value.
1273 DAG.TransferDbgValues(SDValue(N, 0), RV);
1274 if (N->getNumValues() == RV.getNode()->getNumValues())
1275 DAG.ReplaceAllUsesWith(N, RV.getNode());
1277 assert(N->getValueType(0) == RV.getValueType() &&
1278 N->getNumValues() == 1 && "Type mismatch");
1280 DAG.ReplaceAllUsesWith(N, &OpV);
1283 // Push the new node and any users onto the worklist
1284 AddToWorklist(RV.getNode());
1285 AddUsersToWorklist(RV.getNode());
1287 // Finally, if the node is now dead, remove it from the graph. The node
1288 // may not be dead if the replacement process recursively simplified to
1289 // something else needing this node. This will also take care of adding any
1290 // operands which have lost a user to the worklist.
1291 recursivelyDeleteUnusedNodes(N);
1294 // If the root changed (e.g. it was a dead load, update the root).
1295 DAG.setRoot(Dummy.getValue());
1296 DAG.RemoveDeadNodes();
1299 SDValue DAGCombiner::visit(SDNode *N) {
1300 switch (N->getOpcode()) {
1302 case ISD::TokenFactor: return visitTokenFactor(N);
1303 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1304 case ISD::ADD: return visitADD(N);
1305 case ISD::SUB: return visitSUB(N);
1306 case ISD::ADDC: return visitADDC(N);
1307 case ISD::SUBC: return visitSUBC(N);
1308 case ISD::ADDE: return visitADDE(N);
1309 case ISD::SUBE: return visitSUBE(N);
1310 case ISD::MUL: return visitMUL(N);
1311 case ISD::SDIV: return visitSDIV(N);
1312 case ISD::UDIV: return visitUDIV(N);
1313 case ISD::SREM: return visitSREM(N);
1314 case ISD::UREM: return visitUREM(N);
1315 case ISD::MULHU: return visitMULHU(N);
1316 case ISD::MULHS: return visitMULHS(N);
1317 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1318 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1319 case ISD::SMULO: return visitSMULO(N);
1320 case ISD::UMULO: return visitUMULO(N);
1321 case ISD::SDIVREM: return visitSDIVREM(N);
1322 case ISD::UDIVREM: return visitUDIVREM(N);
1323 case ISD::AND: return visitAND(N);
1324 case ISD::OR: return visitOR(N);
1325 case ISD::XOR: return visitXOR(N);
1326 case ISD::SHL: return visitSHL(N);
1327 case ISD::SRA: return visitSRA(N);
1328 case ISD::SRL: return visitSRL(N);
1330 case ISD::ROTL: return visitRotate(N);
1331 case ISD::CTLZ: return visitCTLZ(N);
1332 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1333 case ISD::CTTZ: return visitCTTZ(N);
1334 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1335 case ISD::CTPOP: return visitCTPOP(N);
1336 case ISD::SELECT: return visitSELECT(N);
1337 case ISD::VSELECT: return visitVSELECT(N);
1338 case ISD::SELECT_CC: return visitSELECT_CC(N);
1339 case ISD::SETCC: return visitSETCC(N);
1340 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1341 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1342 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1343 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1344 case ISD::TRUNCATE: return visitTRUNCATE(N);
1345 case ISD::BITCAST: return visitBITCAST(N);
1346 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1347 case ISD::FADD: return visitFADD(N);
1348 case ISD::FSUB: return visitFSUB(N);
1349 case ISD::FMUL: return visitFMUL(N);
1350 case ISD::FMA: return visitFMA(N);
1351 case ISD::FDIV: return visitFDIV(N);
1352 case ISD::FREM: return visitFREM(N);
1353 case ISD::FSQRT: return visitFSQRT(N);
1354 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1355 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1356 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1357 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1358 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1359 case ISD::FP_ROUND: return visitFP_ROUND(N);
1360 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1361 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1362 case ISD::FNEG: return visitFNEG(N);
1363 case ISD::FABS: return visitFABS(N);
1364 case ISD::FFLOOR: return visitFFLOOR(N);
1365 case ISD::FMINNUM: return visitFMINNUM(N);
1366 case ISD::FMAXNUM: return visitFMAXNUM(N);
1367 case ISD::FCEIL: return visitFCEIL(N);
1368 case ISD::FTRUNC: return visitFTRUNC(N);
1369 case ISD::BRCOND: return visitBRCOND(N);
1370 case ISD::BR_CC: return visitBR_CC(N);
1371 case ISD::LOAD: return visitLOAD(N);
1372 case ISD::STORE: return visitSTORE(N);
1373 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1374 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1375 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1376 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1377 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1378 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1379 case ISD::SCALAR_TO_VECTOR: return visitSCALAR_TO_VECTOR(N);
1380 case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N);
1381 case ISD::MLOAD: return visitMLOAD(N);
1382 case ISD::MSTORE: return visitMSTORE(N);
1387 SDValue DAGCombiner::combine(SDNode *N) {
1388 SDValue RV = visit(N);
1390 // If nothing happened, try a target-specific DAG combine.
1391 if (!RV.getNode()) {
1392 assert(N->getOpcode() != ISD::DELETED_NODE &&
1393 "Node was deleted but visit returned NULL!");
1395 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1396 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1398 // Expose the DAG combiner to the target combiner impls.
1399 TargetLowering::DAGCombinerInfo
1400 DagCombineInfo(DAG, Level, false, this);
1402 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1406 // If nothing happened still, try promoting the operation.
1407 if (!RV.getNode()) {
1408 switch (N->getOpcode()) {
1416 RV = PromoteIntBinOp(SDValue(N, 0));
1421 RV = PromoteIntShiftOp(SDValue(N, 0));
1423 case ISD::SIGN_EXTEND:
1424 case ISD::ZERO_EXTEND:
1425 case ISD::ANY_EXTEND:
1426 RV = PromoteExtend(SDValue(N, 0));
1429 if (PromoteLoad(SDValue(N, 0)))
1435 // If N is a commutative binary node, try commuting it to enable more
1437 if (!RV.getNode() && SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1438 N->getNumValues() == 1) {
1439 SDValue N0 = N->getOperand(0);
1440 SDValue N1 = N->getOperand(1);
1442 // Constant operands are canonicalized to RHS.
1443 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1444 SDValue Ops[] = {N1, N0};
1446 if (const BinaryWithFlagsSDNode *BinNode =
1447 dyn_cast<BinaryWithFlagsSDNode>(N)) {
1448 CSENode = DAG.getNodeIfExists(
1449 N->getOpcode(), N->getVTList(), Ops, BinNode->hasNoUnsignedWrap(),
1450 BinNode->hasNoSignedWrap(), BinNode->isExact());
1452 CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops);
1455 return SDValue(CSENode, 0);
1462 /// Given a node, return its input chain if it has one, otherwise return a null
1464 static SDValue getInputChainForNode(SDNode *N) {
1465 if (unsigned NumOps = N->getNumOperands()) {
1466 if (N->getOperand(0).getValueType() == MVT::Other)
1467 return N->getOperand(0);
1468 if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1469 return N->getOperand(NumOps-1);
1470 for (unsigned i = 1; i < NumOps-1; ++i)
1471 if (N->getOperand(i).getValueType() == MVT::Other)
1472 return N->getOperand(i);
1477 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1478 // If N has two operands, where one has an input chain equal to the other,
1479 // the 'other' chain is redundant.
1480 if (N->getNumOperands() == 2) {
1481 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1482 return N->getOperand(0);
1483 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1484 return N->getOperand(1);
1487 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1488 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1489 SmallPtrSet<SDNode*, 16> SeenOps;
1490 bool Changed = false; // If we should replace this token factor.
1492 // Start out with this token factor.
1495 // Iterate through token factors. The TFs grows when new token factors are
1497 for (unsigned i = 0; i < TFs.size(); ++i) {
1498 SDNode *TF = TFs[i];
1500 // Check each of the operands.
1501 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1502 SDValue Op = TF->getOperand(i);
1504 switch (Op.getOpcode()) {
1505 case ISD::EntryToken:
1506 // Entry tokens don't need to be added to the list. They are
1511 case ISD::TokenFactor:
1512 if (Op.hasOneUse() &&
1513 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1514 // Queue up for processing.
1515 TFs.push_back(Op.getNode());
1516 // Clean up in case the token factor is removed.
1517 AddToWorklist(Op.getNode());
1524 // Only add if it isn't already in the list.
1525 if (SeenOps.insert(Op.getNode()).second)
1536 // If we've changed things around then replace token factor.
1539 // The entry token is the only possible outcome.
1540 Result = DAG.getEntryNode();
1542 // New and improved token factor.
1543 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Ops);
1546 // Add users to worklist if AA is enabled, since it may introduce
1547 // a lot of new chained token factors while removing memory deps.
1548 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
1549 : DAG.getSubtarget().useAA();
1550 return CombineTo(N, Result, UseAA /*add to worklist*/);
1556 /// MERGE_VALUES can always be eliminated.
1557 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1558 WorklistRemover DeadNodes(*this);
1559 // Replacing results may cause a different MERGE_VALUES to suddenly
1560 // be CSE'd with N, and carry its uses with it. Iterate until no
1561 // uses remain, to ensure that the node can be safely deleted.
1562 // First add the users of this node to the work list so that they
1563 // can be tried again once they have new operands.
1564 AddUsersToWorklist(N);
1566 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1567 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1568 } while (!N->use_empty());
1569 deleteAndRecombine(N);
1570 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1573 SDValue DAGCombiner::visitADD(SDNode *N) {
1574 SDValue N0 = N->getOperand(0);
1575 SDValue N1 = N->getOperand(1);
1576 EVT VT = N0.getValueType();
1579 if (VT.isVector()) {
1580 if (SDValue FoldedVOp = SimplifyVBinOp(N))
1583 // fold (add x, 0) -> x, vector edition
1584 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1586 if (ISD::isBuildVectorAllZeros(N0.getNode()))
1590 // fold (add x, undef) -> undef
1591 if (N0.getOpcode() == ISD::UNDEF)
1593 if (N1.getOpcode() == ISD::UNDEF)
1595 // fold (add c1, c2) -> c1+c2
1596 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1597 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1599 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1600 // canonicalize constant to RHS
1602 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, N0);
1603 // fold (add x, 0) -> x
1604 if (N1C && N1C->isNullValue())
1606 // fold (add Sym, c) -> Sym+c
1607 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1608 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1609 GA->getOpcode() == ISD::GlobalAddress)
1610 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1612 (uint64_t)N1C->getSExtValue());
1613 // fold ((c1-A)+c2) -> (c1+c2)-A
1614 if (N1C && N0.getOpcode() == ISD::SUB)
1615 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1616 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1617 DAG.getConstant(N1C->getAPIntValue()+
1618 N0C->getAPIntValue(), VT),
1621 if (SDValue RADD = ReassociateOps(ISD::ADD, SDLoc(N), N0, N1))
1623 // fold ((0-A) + B) -> B-A
1624 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1625 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1626 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, N0.getOperand(1));
1627 // fold (A + (0-B)) -> A-B
1628 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1629 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1630 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1.getOperand(1));
1631 // fold (A+(B-A)) -> B
1632 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1633 return N1.getOperand(0);
1634 // fold ((B-A)+A) -> B
1635 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1636 return N0.getOperand(0);
1637 // fold (A+(B-(A+C))) to (B-C)
1638 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1639 N0 == N1.getOperand(1).getOperand(0))
1640 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1641 N1.getOperand(1).getOperand(1));
1642 // fold (A+(B-(C+A))) to (B-C)
1643 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1644 N0 == N1.getOperand(1).getOperand(1))
1645 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1646 N1.getOperand(1).getOperand(0));
1647 // fold (A+((B-A)+or-C)) to (B+or-C)
1648 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1649 N1.getOperand(0).getOpcode() == ISD::SUB &&
1650 N0 == N1.getOperand(0).getOperand(1))
1651 return DAG.getNode(N1.getOpcode(), SDLoc(N), VT,
1652 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1654 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1655 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1656 SDValue N00 = N0.getOperand(0);
1657 SDValue N01 = N0.getOperand(1);
1658 SDValue N10 = N1.getOperand(0);
1659 SDValue N11 = N1.getOperand(1);
1661 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1662 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1663 DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10),
1664 DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11));
1667 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1668 return SDValue(N, 0);
1670 // fold (a+b) -> (a|b) iff a and b share no bits.
1671 if (VT.isInteger() && !VT.isVector()) {
1672 APInt LHSZero, LHSOne;
1673 APInt RHSZero, RHSOne;
1674 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1676 if (LHSZero.getBoolValue()) {
1677 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1679 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1680 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1681 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero){
1682 if (!LegalOperations || TLI.isOperationLegal(ISD::OR, VT))
1683 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1);
1688 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1689 if (N1.getOpcode() == ISD::SHL &&
1690 N1.getOperand(0).getOpcode() == ISD::SUB)
1691 if (ConstantSDNode *C =
1692 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1693 if (C->getAPIntValue() == 0)
1694 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0,
1695 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1696 N1.getOperand(0).getOperand(1),
1698 if (N0.getOpcode() == ISD::SHL &&
1699 N0.getOperand(0).getOpcode() == ISD::SUB)
1700 if (ConstantSDNode *C =
1701 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1702 if (C->getAPIntValue() == 0)
1703 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1,
1704 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1705 N0.getOperand(0).getOperand(1),
1708 if (N1.getOpcode() == ISD::AND) {
1709 SDValue AndOp0 = N1.getOperand(0);
1710 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1711 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1712 unsigned DestBits = VT.getScalarType().getSizeInBits();
1714 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1715 // and similar xforms where the inner op is either ~0 or 0.
1716 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1718 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1722 // add (sext i1), X -> sub X, (zext i1)
1723 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1724 N0.getOperand(0).getValueType() == MVT::i1 &&
1725 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1727 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1728 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1731 // add X, (sextinreg Y i1) -> sub X, (and Y 1)
1732 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
1733 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
1734 if (TN->getVT() == MVT::i1) {
1736 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
1737 DAG.getConstant(1, VT));
1738 return DAG.getNode(ISD::SUB, DL, VT, N0, ZExt);
1745 SDValue DAGCombiner::visitADDC(SDNode *N) {
1746 SDValue N0 = N->getOperand(0);
1747 SDValue N1 = N->getOperand(1);
1748 EVT VT = N0.getValueType();
1750 // If the flag result is dead, turn this into an ADD.
1751 if (!N->hasAnyUseOfValue(1))
1752 return CombineTo(N, DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N1),
1753 DAG.getNode(ISD::CARRY_FALSE,
1754 SDLoc(N), MVT::Glue));
1756 // canonicalize constant to RHS.
1757 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1758 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1760 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N1, N0);
1762 // fold (addc x, 0) -> x + no carry out
1763 if (N1C && N1C->isNullValue())
1764 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1765 SDLoc(N), MVT::Glue));
1767 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1768 APInt LHSZero, LHSOne;
1769 APInt RHSZero, RHSOne;
1770 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1772 if (LHSZero.getBoolValue()) {
1773 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1775 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1776 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1777 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1778 return CombineTo(N, DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1),
1779 DAG.getNode(ISD::CARRY_FALSE,
1780 SDLoc(N), MVT::Glue));
1786 SDValue DAGCombiner::visitADDE(SDNode *N) {
1787 SDValue N0 = N->getOperand(0);
1788 SDValue N1 = N->getOperand(1);
1789 SDValue CarryIn = N->getOperand(2);
1791 // canonicalize constant to RHS
1792 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1793 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1795 return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(),
1798 // fold (adde x, y, false) -> (addc x, y)
1799 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1800 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1);
1805 // Since it may not be valid to emit a fold to zero for vector initializers
1806 // check if we can before folding.
1807 static SDValue tryFoldToZero(SDLoc DL, const TargetLowering &TLI, EVT VT,
1809 bool LegalOperations, bool LegalTypes) {
1811 return DAG.getConstant(0, VT);
1812 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
1813 return DAG.getConstant(0, VT);
1817 SDValue DAGCombiner::visitSUB(SDNode *N) {
1818 SDValue N0 = N->getOperand(0);
1819 SDValue N1 = N->getOperand(1);
1820 EVT VT = N0.getValueType();
1823 if (VT.isVector()) {
1824 if (SDValue FoldedVOp = SimplifyVBinOp(N))
1827 // fold (sub x, 0) -> x, vector edition
1828 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1832 // fold (sub x, x) -> 0
1833 // FIXME: Refactor this and xor and other similar operations together.
1835 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
1836 // fold (sub c1, c2) -> c1-c2
1837 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1838 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1840 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1841 // fold (sub x, c) -> (add x, -c)
1843 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0,
1844 DAG.getConstant(-N1C->getAPIntValue(), VT));
1845 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1846 if (N0C && N0C->isAllOnesValue())
1847 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
1848 // fold A-(A-B) -> B
1849 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1850 return N1.getOperand(1);
1851 // fold (A+B)-A -> B
1852 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1853 return N0.getOperand(1);
1854 // fold (A+B)-B -> A
1855 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1856 return N0.getOperand(0);
1857 // fold C2-(A+C1) -> (C2-C1)-A
1858 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? nullptr :
1859 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1860 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1861 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(),
1863 return DAG.getNode(ISD::SUB, SDLoc(N), VT, NewC,
1866 // fold ((A+(B+or-C))-B) -> A+or-C
1867 if (N0.getOpcode() == ISD::ADD &&
1868 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1869 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1870 N0.getOperand(1).getOperand(0) == N1)
1871 return DAG.getNode(N0.getOperand(1).getOpcode(), SDLoc(N), VT,
1872 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1873 // fold ((A+(C+B))-B) -> A+C
1874 if (N0.getOpcode() == ISD::ADD &&
1875 N0.getOperand(1).getOpcode() == ISD::ADD &&
1876 N0.getOperand(1).getOperand(1) == N1)
1877 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1878 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1879 // fold ((A-(B-C))-C) -> A-B
1880 if (N0.getOpcode() == ISD::SUB &&
1881 N0.getOperand(1).getOpcode() == ISD::SUB &&
1882 N0.getOperand(1).getOperand(1) == N1)
1883 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1884 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1886 // If either operand of a sub is undef, the result is undef
1887 if (N0.getOpcode() == ISD::UNDEF)
1889 if (N1.getOpcode() == ISD::UNDEF)
1892 // If the relocation model supports it, consider symbol offsets.
1893 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1894 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1895 // fold (sub Sym, c) -> Sym-c
1896 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1897 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1899 (uint64_t)N1C->getSExtValue());
1900 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1901 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1902 if (GA->getGlobal() == GB->getGlobal())
1903 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1907 // sub X, (sextinreg Y i1) -> add X, (and Y 1)
1908 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
1909 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
1910 if (TN->getVT() == MVT::i1) {
1912 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
1913 DAG.getConstant(1, VT));
1914 return DAG.getNode(ISD::ADD, DL, VT, N0, ZExt);
1921 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1922 SDValue N0 = N->getOperand(0);
1923 SDValue N1 = N->getOperand(1);
1924 EVT VT = N0.getValueType();
1926 // If the flag result is dead, turn this into an SUB.
1927 if (!N->hasAnyUseOfValue(1))
1928 return CombineTo(N, DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1),
1929 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1932 // fold (subc x, x) -> 0 + no borrow
1934 return CombineTo(N, DAG.getConstant(0, VT),
1935 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1938 // fold (subc x, 0) -> x + no borrow
1939 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1940 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1941 if (N1C && N1C->isNullValue())
1942 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1945 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1946 if (N0C && N0C->isAllOnesValue())
1947 return CombineTo(N, DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0),
1948 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1954 SDValue DAGCombiner::visitSUBE(SDNode *N) {
1955 SDValue N0 = N->getOperand(0);
1956 SDValue N1 = N->getOperand(1);
1957 SDValue CarryIn = N->getOperand(2);
1959 // fold (sube x, y, false) -> (subc x, y)
1960 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1961 return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1);
1966 SDValue DAGCombiner::visitMUL(SDNode *N) {
1967 SDValue N0 = N->getOperand(0);
1968 SDValue N1 = N->getOperand(1);
1969 EVT VT = N0.getValueType();
1971 // fold (mul x, undef) -> 0
1972 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1973 return DAG.getConstant(0, VT);
1975 bool N0IsConst = false;
1976 bool N1IsConst = false;
1977 APInt ConstValue0, ConstValue1;
1979 if (VT.isVector()) {
1980 if (SDValue FoldedVOp = SimplifyVBinOp(N))
1983 N0IsConst = isConstantSplatVector(N0.getNode(), ConstValue0);
1984 N1IsConst = isConstantSplatVector(N1.getNode(), ConstValue1);
1986 N0IsConst = dyn_cast<ConstantSDNode>(N0) != nullptr;
1987 ConstValue0 = N0IsConst ? (dyn_cast<ConstantSDNode>(N0))->getAPIntValue()
1989 N1IsConst = dyn_cast<ConstantSDNode>(N1) != nullptr;
1990 ConstValue1 = N1IsConst ? (dyn_cast<ConstantSDNode>(N1))->getAPIntValue()
1994 // fold (mul c1, c2) -> c1*c2
1995 if (N0IsConst && N1IsConst)
1996 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0.getNode(), N1.getNode());
1998 // canonicalize constant to RHS
1999 if (N0IsConst && !N1IsConst)
2000 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0);
2001 // fold (mul x, 0) -> 0
2002 if (N1IsConst && ConstValue1 == 0)
2004 // We require a splat of the entire scalar bit width for non-contiguous
2007 ConstValue1.getBitWidth() == VT.getScalarType().getSizeInBits();
2008 // fold (mul x, 1) -> x
2009 if (N1IsConst && ConstValue1 == 1 && IsFullSplat)
2011 // fold (mul x, -1) -> 0-x
2012 if (N1IsConst && ConstValue1.isAllOnesValue())
2013 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2014 DAG.getConstant(0, VT), N0);
2015 // fold (mul x, (1 << c)) -> x << c
2016 if (N1IsConst && ConstValue1.isPowerOf2() && IsFullSplat)
2017 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
2018 DAG.getConstant(ConstValue1.logBase2(),
2019 getShiftAmountTy(N0.getValueType())));
2020 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
2021 if (N1IsConst && (-ConstValue1).isPowerOf2() && IsFullSplat) {
2022 unsigned Log2Val = (-ConstValue1).logBase2();
2023 // FIXME: If the input is something that is easily negated (e.g. a
2024 // single-use add), we should put the negate there.
2025 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2026 DAG.getConstant(0, VT),
2027 DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
2028 DAG.getConstant(Log2Val,
2029 getShiftAmountTy(N0.getValueType()))));
2033 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
2034 if (N1IsConst && N0.getOpcode() == ISD::SHL &&
2035 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2036 isa<ConstantSDNode>(N0.getOperand(1)))) {
2037 SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT,
2038 N1, N0.getOperand(1));
2039 AddToWorklist(C3.getNode());
2040 return DAG.getNode(ISD::MUL, SDLoc(N), VT,
2041 N0.getOperand(0), C3);
2044 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
2047 SDValue Sh(nullptr,0), Y(nullptr,0);
2048 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
2049 if (N0.getOpcode() == ISD::SHL &&
2050 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2051 isa<ConstantSDNode>(N0.getOperand(1))) &&
2052 N0.getNode()->hasOneUse()) {
2054 } else if (N1.getOpcode() == ISD::SHL &&
2055 isa<ConstantSDNode>(N1.getOperand(1)) &&
2056 N1.getNode()->hasOneUse()) {
2061 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2062 Sh.getOperand(0), Y);
2063 return DAG.getNode(ISD::SHL, SDLoc(N), VT,
2064 Mul, Sh.getOperand(1));
2068 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
2069 if (N1IsConst && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
2070 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2071 isa<ConstantSDNode>(N0.getOperand(1))))
2072 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
2073 DAG.getNode(ISD::MUL, SDLoc(N0), VT,
2074 N0.getOperand(0), N1),
2075 DAG.getNode(ISD::MUL, SDLoc(N1), VT,
2076 N0.getOperand(1), N1));
2079 if (SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1))
2085 SDValue DAGCombiner::visitSDIV(SDNode *N) {
2086 SDValue N0 = N->getOperand(0);
2087 SDValue N1 = N->getOperand(1);
2088 EVT VT = N->getValueType(0);
2092 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2095 // fold (sdiv c1, c2) -> c1/c2
2096 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2097 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2098 if (N0C && N1C && !N1C->isNullValue())
2099 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
2100 // fold (sdiv X, 1) -> X
2101 if (N1C && N1C->getAPIntValue() == 1LL)
2103 // fold (sdiv X, -1) -> 0-X
2104 if (N1C && N1C->isAllOnesValue())
2105 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2106 DAG.getConstant(0, VT), N0);
2107 // If we know the sign bits of both operands are zero, strength reduce to a
2108 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
2109 if (!VT.isVector()) {
2110 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2111 return DAG.getNode(ISD::UDIV, SDLoc(N), N1.getValueType(),
2115 // fold (sdiv X, pow2) -> simple ops after legalize
2116 if (N1C && !N1C->isNullValue() && (N1C->getAPIntValue().isPowerOf2() ||
2117 (-N1C->getAPIntValue()).isPowerOf2())) {
2118 // If dividing by powers of two is cheap, then don't perform the following
2120 if (TLI.isPow2SDivCheap())
2123 // Target-specific implementation of sdiv x, pow2.
2124 SDValue Res = BuildSDIVPow2(N);
2128 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
2130 // Splat the sign bit into the register
2132 DAG.getNode(ISD::SRA, SDLoc(N), VT, N0,
2133 DAG.getConstant(VT.getScalarSizeInBits() - 1,
2134 getShiftAmountTy(N0.getValueType())));
2135 AddToWorklist(SGN.getNode());
2137 // Add (N0 < 0) ? abs2 - 1 : 0;
2139 DAG.getNode(ISD::SRL, SDLoc(N), VT, SGN,
2140 DAG.getConstant(VT.getScalarSizeInBits() - lg2,
2141 getShiftAmountTy(SGN.getValueType())));
2142 SDValue ADD = DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, SRL);
2143 AddToWorklist(SRL.getNode());
2144 AddToWorklist(ADD.getNode()); // Divide by pow2
2145 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), VT, ADD,
2146 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType())));
2148 // If we're dividing by a positive value, we're done. Otherwise, we must
2149 // negate the result.
2150 if (N1C->getAPIntValue().isNonNegative())
2153 AddToWorklist(SRA.getNode());
2154 return DAG.getNode(ISD::SUB, SDLoc(N), VT, DAG.getConstant(0, VT), SRA);
2157 // if integer divide is expensive and we satisfy the requirements, emit an
2158 // alternate sequence.
2159 if (N1C && !TLI.isIntDivCheap()) {
2160 SDValue Op = BuildSDIV(N);
2161 if (Op.getNode()) return Op;
2165 if (N0.getOpcode() == ISD::UNDEF)
2166 return DAG.getConstant(0, VT);
2167 // X / undef -> undef
2168 if (N1.getOpcode() == ISD::UNDEF)
2174 SDValue DAGCombiner::visitUDIV(SDNode *N) {
2175 SDValue N0 = N->getOperand(0);
2176 SDValue N1 = N->getOperand(1);
2177 EVT VT = N->getValueType(0);
2181 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2184 // fold (udiv c1, c2) -> c1/c2
2185 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2186 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2187 if (N0C && N1C && !N1C->isNullValue())
2188 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
2189 // fold (udiv x, (1 << c)) -> x >>u c
2190 if (N1C && N1C->getAPIntValue().isPowerOf2())
2191 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0,
2192 DAG.getConstant(N1C->getAPIntValue().logBase2(),
2193 getShiftAmountTy(N0.getValueType())));
2194 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
2195 if (N1.getOpcode() == ISD::SHL) {
2196 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2197 if (SHC->getAPIntValue().isPowerOf2()) {
2198 EVT ADDVT = N1.getOperand(1).getValueType();
2199 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N), ADDVT,
2201 DAG.getConstant(SHC->getAPIntValue()
2204 AddToWorklist(Add.getNode());
2205 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, Add);
2209 // fold (udiv x, c) -> alternate
2210 if (N1C && !TLI.isIntDivCheap()) {
2211 SDValue Op = BuildUDIV(N);
2212 if (Op.getNode()) return Op;
2216 if (N0.getOpcode() == ISD::UNDEF)
2217 return DAG.getConstant(0, VT);
2218 // X / undef -> undef
2219 if (N1.getOpcode() == ISD::UNDEF)
2225 SDValue DAGCombiner::visitSREM(SDNode *N) {
2226 SDValue N0 = N->getOperand(0);
2227 SDValue N1 = N->getOperand(1);
2228 EVT VT = N->getValueType(0);
2230 // fold (srem c1, c2) -> c1%c2
2231 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2232 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2233 if (N0C && N1C && !N1C->isNullValue())
2234 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
2235 // If we know the sign bits of both operands are zero, strength reduce to a
2236 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
2237 if (!VT.isVector()) {
2238 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2239 return DAG.getNode(ISD::UREM, SDLoc(N), VT, N0, N1);
2242 // If X/C can be simplified by the division-by-constant logic, lower
2243 // X%C to the equivalent of X-X/C*C.
2244 if (N1C && !N1C->isNullValue()) {
2245 SDValue Div = DAG.getNode(ISD::SDIV, SDLoc(N), VT, N0, N1);
2246 AddToWorklist(Div.getNode());
2247 SDValue OptimizedDiv = combine(Div.getNode());
2248 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2249 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2251 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2252 AddToWorklist(Mul.getNode());
2258 if (N0.getOpcode() == ISD::UNDEF)
2259 return DAG.getConstant(0, VT);
2260 // X % undef -> undef
2261 if (N1.getOpcode() == ISD::UNDEF)
2267 SDValue DAGCombiner::visitUREM(SDNode *N) {
2268 SDValue N0 = N->getOperand(0);
2269 SDValue N1 = N->getOperand(1);
2270 EVT VT = N->getValueType(0);
2272 // fold (urem c1, c2) -> c1%c2
2273 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2274 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2275 if (N0C && N1C && !N1C->isNullValue())
2276 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
2277 // fold (urem x, pow2) -> (and x, pow2-1)
2278 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
2279 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0,
2280 DAG.getConstant(N1C->getAPIntValue()-1,VT));
2281 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2282 if (N1.getOpcode() == ISD::SHL) {
2283 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2284 if (SHC->getAPIntValue().isPowerOf2()) {
2286 DAG.getNode(ISD::ADD, SDLoc(N), VT, N1,
2287 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
2289 AddToWorklist(Add.getNode());
2290 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, Add);
2295 // If X/C can be simplified by the division-by-constant logic, lower
2296 // X%C to the equivalent of X-X/C*C.
2297 if (N1C && !N1C->isNullValue()) {
2298 SDValue Div = DAG.getNode(ISD::UDIV, SDLoc(N), VT, N0, N1);
2299 AddToWorklist(Div.getNode());
2300 SDValue OptimizedDiv = combine(Div.getNode());
2301 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2302 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2304 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2305 AddToWorklist(Mul.getNode());
2311 if (N0.getOpcode() == ISD::UNDEF)
2312 return DAG.getConstant(0, VT);
2313 // X % undef -> undef
2314 if (N1.getOpcode() == ISD::UNDEF)
2320 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2321 SDValue N0 = N->getOperand(0);
2322 SDValue N1 = N->getOperand(1);
2323 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2324 EVT VT = N->getValueType(0);
2327 // fold (mulhs x, 0) -> 0
2328 if (N1C && N1C->isNullValue())
2330 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2331 if (N1C && N1C->getAPIntValue() == 1)
2332 return DAG.getNode(ISD::SRA, SDLoc(N), N0.getValueType(), N0,
2333 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2334 getShiftAmountTy(N0.getValueType())));
2335 // fold (mulhs x, undef) -> 0
2336 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2337 return DAG.getConstant(0, VT);
2339 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2341 if (VT.isSimple() && !VT.isVector()) {
2342 MVT Simple = VT.getSimpleVT();
2343 unsigned SimpleSize = Simple.getSizeInBits();
2344 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2345 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2346 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2347 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2348 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2349 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2350 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2351 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2358 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2359 SDValue N0 = N->getOperand(0);
2360 SDValue N1 = N->getOperand(1);
2361 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2362 EVT VT = N->getValueType(0);
2365 // fold (mulhu x, 0) -> 0
2366 if (N1C && N1C->isNullValue())
2368 // fold (mulhu x, 1) -> 0
2369 if (N1C && N1C->getAPIntValue() == 1)
2370 return DAG.getConstant(0, N0.getValueType());
2371 // fold (mulhu x, undef) -> 0
2372 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2373 return DAG.getConstant(0, VT);
2375 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2377 if (VT.isSimple() && !VT.isVector()) {
2378 MVT Simple = VT.getSimpleVT();
2379 unsigned SimpleSize = Simple.getSizeInBits();
2380 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2381 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2382 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2383 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2384 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2385 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2386 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2387 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2394 /// Perform optimizations common to nodes that compute two values. LoOp and HiOp
2395 /// give the opcodes for the two computations that are being performed. Return
2396 /// true if a simplification was made.
2397 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2399 // If the high half is not needed, just compute the low half.
2400 bool HiExists = N->hasAnyUseOfValue(1);
2402 (!LegalOperations ||
2403 TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) {
2404 SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2405 return CombineTo(N, Res, Res);
2408 // If the low half is not needed, just compute the high half.
2409 bool LoExists = N->hasAnyUseOfValue(0);
2411 (!LegalOperations ||
2412 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2413 SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2414 return CombineTo(N, Res, Res);
2417 // If both halves are used, return as it is.
2418 if (LoExists && HiExists)
2421 // If the two computed results can be simplified separately, separate them.
2423 SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2424 AddToWorklist(Lo.getNode());
2425 SDValue LoOpt = combine(Lo.getNode());
2426 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2427 (!LegalOperations ||
2428 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2429 return CombineTo(N, LoOpt, LoOpt);
2433 SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2434 AddToWorklist(Hi.getNode());
2435 SDValue HiOpt = combine(Hi.getNode());
2436 if (HiOpt.getNode() && HiOpt != Hi &&
2437 (!LegalOperations ||
2438 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2439 return CombineTo(N, HiOpt, HiOpt);
2445 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2446 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2447 if (Res.getNode()) return Res;
2449 EVT VT = N->getValueType(0);
2452 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2454 if (VT.isSimple() && !VT.isVector()) {
2455 MVT Simple = VT.getSimpleVT();
2456 unsigned SimpleSize = Simple.getSizeInBits();
2457 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2458 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2459 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2460 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2461 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2462 // Compute the high part as N1.
2463 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2464 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2465 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2466 // Compute the low part as N0.
2467 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2468 return CombineTo(N, Lo, Hi);
2475 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2476 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2477 if (Res.getNode()) return Res;
2479 EVT VT = N->getValueType(0);
2482 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2484 if (VT.isSimple() && !VT.isVector()) {
2485 MVT Simple = VT.getSimpleVT();
2486 unsigned SimpleSize = Simple.getSizeInBits();
2487 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2488 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2489 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2490 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2491 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2492 // Compute the high part as N1.
2493 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2494 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2495 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2496 // Compute the low part as N0.
2497 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2498 return CombineTo(N, Lo, Hi);
2505 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2506 // (smulo x, 2) -> (saddo x, x)
2507 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2508 if (C2->getAPIntValue() == 2)
2509 return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(),
2510 N->getOperand(0), N->getOperand(0));
2515 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2516 // (umulo x, 2) -> (uaddo x, x)
2517 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2518 if (C2->getAPIntValue() == 2)
2519 return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(),
2520 N->getOperand(0), N->getOperand(0));
2525 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2526 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2527 if (Res.getNode()) return Res;
2532 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2533 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2534 if (Res.getNode()) return Res;
2539 /// If this is a binary operator with two operands of the same opcode, try to
2541 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2542 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2543 EVT VT = N0.getValueType();
2544 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2546 // Bail early if none of these transforms apply.
2547 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2549 // For each of OP in AND/OR/XOR:
2550 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2551 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2552 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2553 // fold (OP (bswap x), (bswap y)) -> (bswap (OP x, y))
2554 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2556 // do not sink logical op inside of a vector extend, since it may combine
2558 EVT Op0VT = N0.getOperand(0).getValueType();
2559 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2560 N0.getOpcode() == ISD::SIGN_EXTEND ||
2561 N0.getOpcode() == ISD::BSWAP ||
2562 // Avoid infinite looping with PromoteIntBinOp.
2563 (N0.getOpcode() == ISD::ANY_EXTEND &&
2564 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2565 (N0.getOpcode() == ISD::TRUNCATE &&
2566 (!TLI.isZExtFree(VT, Op0VT) ||
2567 !TLI.isTruncateFree(Op0VT, VT)) &&
2568 TLI.isTypeLegal(Op0VT))) &&
2570 Op0VT == N1.getOperand(0).getValueType() &&
2571 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2572 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2573 N0.getOperand(0).getValueType(),
2574 N0.getOperand(0), N1.getOperand(0));
2575 AddToWorklist(ORNode.getNode());
2576 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode);
2579 // For each of OP in SHL/SRL/SRA/AND...
2580 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2581 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2582 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2583 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2584 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2585 N0.getOperand(1) == N1.getOperand(1)) {
2586 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2587 N0.getOperand(0).getValueType(),
2588 N0.getOperand(0), N1.getOperand(0));
2589 AddToWorklist(ORNode.getNode());
2590 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
2591 ORNode, N0.getOperand(1));
2594 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2595 // Only perform this optimization after type legalization and before
2596 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2597 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2598 // we don't want to undo this promotion.
2599 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2601 if ((N0.getOpcode() == ISD::BITCAST ||
2602 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) &&
2603 Level == AfterLegalizeTypes) {
2604 SDValue In0 = N0.getOperand(0);
2605 SDValue In1 = N1.getOperand(0);
2606 EVT In0Ty = In0.getValueType();
2607 EVT In1Ty = In1.getValueType();
2609 // If both incoming values are integers, and the original types are the
2611 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2612 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1);
2613 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op);
2614 AddToWorklist(Op.getNode());
2619 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2620 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2621 // If both shuffles use the same mask, and both shuffle within a single
2622 // vector, then it is worthwhile to move the swizzle after the operation.
2623 // The type-legalizer generates this pattern when loading illegal
2624 // vector types from memory. In many cases this allows additional shuffle
2626 // There are other cases where moving the shuffle after the xor/and/or
2627 // is profitable even if shuffles don't perform a swizzle.
2628 // If both shuffles use the same mask, and both shuffles have the same first
2629 // or second operand, then it might still be profitable to move the shuffle
2630 // after the xor/and/or operation.
2631 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) {
2632 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2633 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2635 assert(N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() &&
2636 "Inputs to shuffles are not the same type");
2638 // Check that both shuffles use the same mask. The masks are known to be of
2639 // the same length because the result vector type is the same.
2640 // Check also that shuffles have only one use to avoid introducing extra
2642 if (SVN0->hasOneUse() && SVN1->hasOneUse() &&
2643 SVN0->getMask().equals(SVN1->getMask())) {
2644 SDValue ShOp = N0->getOperand(1);
2646 // Don't try to fold this node if it requires introducing a
2647 // build vector of all zeros that might be illegal at this stage.
2648 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2650 ShOp = DAG.getConstant(0, VT);
2655 // (AND (shuf (A, C), shuf (B, C)) -> shuf (AND (A, B), C)
2656 // (OR (shuf (A, C), shuf (B, C)) -> shuf (OR (A, B), C)
2657 // (XOR (shuf (A, C), shuf (B, C)) -> shuf (XOR (A, B), V_0)
2658 if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) {
2659 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2660 N0->getOperand(0), N1->getOperand(0));
2661 AddToWorklist(NewNode.getNode());
2662 return DAG.getVectorShuffle(VT, SDLoc(N), NewNode, ShOp,
2663 &SVN0->getMask()[0]);
2666 // Don't try to fold this node if it requires introducing a
2667 // build vector of all zeros that might be illegal at this stage.
2668 ShOp = N0->getOperand(0);
2669 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2671 ShOp = DAG.getConstant(0, VT);
2676 // (AND (shuf (C, A), shuf (C, B)) -> shuf (C, AND (A, B))
2677 // (OR (shuf (C, A), shuf (C, B)) -> shuf (C, OR (A, B))
2678 // (XOR (shuf (C, A), shuf (C, B)) -> shuf (V_0, XOR (A, B))
2679 if (N0->getOperand(0) == N1->getOperand(0) && ShOp.getNode()) {
2680 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2681 N0->getOperand(1), N1->getOperand(1));
2682 AddToWorklist(NewNode.getNode());
2683 return DAG.getVectorShuffle(VT, SDLoc(N), ShOp, NewNode,
2684 &SVN0->getMask()[0]);
2692 /// This contains all DAGCombine rules which reduce two values combined by
2693 /// an And operation to a single value. This makes them reusable in the context
2694 /// of visitSELECT(). Rules involving constants are not included as
2695 /// visitSELECT() already handles those cases.
2696 SDValue DAGCombiner::visitANDLike(SDValue N0, SDValue N1,
2697 SDNode *LocReference) {
2698 EVT VT = N1.getValueType();
2700 // fold (and x, undef) -> 0
2701 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2702 return DAG.getConstant(0, VT);
2703 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2704 SDValue LL, LR, RL, RR, CC0, CC1;
2705 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2706 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2707 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2709 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2710 LL.getValueType().isInteger()) {
2711 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2712 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
2713 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2714 LR.getValueType(), LL, RL);
2715 AddToWorklist(ORNode.getNode());
2716 return DAG.getSetCC(SDLoc(LocReference), VT, ORNode, LR, Op1);
2718 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2719 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2720 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(N0),
2721 LR.getValueType(), LL, RL);
2722 AddToWorklist(ANDNode.getNode());
2723 return DAG.getSetCC(SDLoc(LocReference), VT, ANDNode, LR, Op1);
2725 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2726 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2727 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2728 LR.getValueType(), LL, RL);
2729 AddToWorklist(ORNode.getNode());
2730 return DAG.getSetCC(SDLoc(LocReference), VT, ORNode, LR, Op1);
2733 // Simplify (and (setne X, 0), (setne X, -1)) -> (setuge (add X, 1), 2)
2734 if (LL == RL && isa<ConstantSDNode>(LR) && isa<ConstantSDNode>(RR) &&
2735 Op0 == Op1 && LL.getValueType().isInteger() &&
2736 Op0 == ISD::SETNE && ((cast<ConstantSDNode>(LR)->isNullValue() &&
2737 cast<ConstantSDNode>(RR)->isAllOnesValue()) ||
2738 (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
2739 cast<ConstantSDNode>(RR)->isNullValue()))) {
2740 SDValue ADDNode = DAG.getNode(ISD::ADD, SDLoc(N0), LL.getValueType(),
2741 LL, DAG.getConstant(1, LL.getValueType()));
2742 AddToWorklist(ADDNode.getNode());
2743 return DAG.getSetCC(SDLoc(LocReference), VT, ADDNode,
2744 DAG.getConstant(2, LL.getValueType()), ISD::SETUGE);
2746 // canonicalize equivalent to ll == rl
2747 if (LL == RR && LR == RL) {
2748 Op1 = ISD::getSetCCSwappedOperands(Op1);
2751 if (LL == RL && LR == RR) {
2752 bool isInteger = LL.getValueType().isInteger();
2753 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2754 if (Result != ISD::SETCC_INVALID &&
2755 (!LegalOperations ||
2756 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
2757 TLI.isOperationLegal(ISD::SETCC,
2758 getSetCCResultType(N0.getSimpleValueType())))))
2759 return DAG.getSetCC(SDLoc(LocReference), N0.getValueType(),
2764 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
2765 VT.getSizeInBits() <= 64) {
2766 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2767 APInt ADDC = ADDI->getAPIntValue();
2768 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2769 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
2770 // immediate for an add, but it is legal if its top c2 bits are set,
2771 // transform the ADD so the immediate doesn't need to be materialized
2773 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
2774 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
2775 SRLI->getZExtValue());
2776 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
2778 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2780 DAG.getNode(ISD::ADD, SDLoc(N0), VT,
2781 N0.getOperand(0), DAG.getConstant(ADDC, VT));
2782 CombineTo(N0.getNode(), NewAdd);
2783 // Return N so it doesn't get rechecked!
2784 return SDValue(LocReference, 0);
2795 SDValue DAGCombiner::visitAND(SDNode *N) {
2796 SDValue N0 = N->getOperand(0);
2797 SDValue N1 = N->getOperand(1);
2798 EVT VT = N1.getValueType();
2801 if (VT.isVector()) {
2802 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2805 // fold (and x, 0) -> 0, vector edition
2806 if (ISD::isBuildVectorAllZeros(N0.getNode()))
2807 // do not return N0, because undef node may exist in N0
2808 return DAG.getConstant(
2809 APInt::getNullValue(
2810 N0.getValueType().getScalarType().getSizeInBits()),
2812 if (ISD::isBuildVectorAllZeros(N1.getNode()))
2813 // do not return N1, because undef node may exist in N1
2814 return DAG.getConstant(
2815 APInt::getNullValue(
2816 N1.getValueType().getScalarType().getSizeInBits()),
2819 // fold (and x, -1) -> x, vector edition
2820 if (ISD::isBuildVectorAllOnes(N0.getNode()))
2822 if (ISD::isBuildVectorAllOnes(N1.getNode()))
2826 // fold (and c1, c2) -> c1&c2
2827 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2828 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2830 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2831 // canonicalize constant to RHS
2833 return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0);
2834 // fold (and x, -1) -> x
2835 if (N1C && N1C->isAllOnesValue())
2837 // if (and x, c) is known to be zero, return 0
2838 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2839 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2840 APInt::getAllOnesValue(BitWidth)))
2841 return DAG.getConstant(0, VT);
2843 if (SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1))
2845 // fold (and (or x, C), D) -> D if (C & D) == D
2846 if (N1C && N0.getOpcode() == ISD::OR)
2847 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2848 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2850 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2851 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2852 SDValue N0Op0 = N0.getOperand(0);
2853 APInt Mask = ~N1C->getAPIntValue();
2854 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2855 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2856 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
2857 N0.getValueType(), N0Op0);
2859 // Replace uses of the AND with uses of the Zero extend node.
2862 // We actually want to replace all uses of the any_extend with the
2863 // zero_extend, to avoid duplicating things. This will later cause this
2864 // AND to be folded.
2865 CombineTo(N0.getNode(), Zext);
2866 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2869 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2870 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2871 // already be zero by virtue of the width of the base type of the load.
2873 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2875 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2876 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2877 N0.getOpcode() == ISD::LOAD) {
2878 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2879 N0 : N0.getOperand(0) );
2881 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2882 // This can be a pure constant or a vector splat, in which case we treat the
2883 // vector as a scalar and use the splat value.
2884 APInt Constant = APInt::getNullValue(1);
2885 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2886 Constant = C->getAPIntValue();
2887 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2888 APInt SplatValue, SplatUndef;
2889 unsigned SplatBitSize;
2891 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2892 SplatBitSize, HasAnyUndefs);
2894 // Undef bits can contribute to a possible optimisation if set, so
2896 SplatValue |= SplatUndef;
2898 // The splat value may be something like "0x00FFFFFF", which means 0 for
2899 // the first vector value and FF for the rest, repeating. We need a mask
2900 // that will apply equally to all members of the vector, so AND all the
2901 // lanes of the constant together.
2902 EVT VT = Vector->getValueType(0);
2903 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2905 // If the splat value has been compressed to a bitlength lower
2906 // than the size of the vector lane, we need to re-expand it to
2908 if (BitWidth > SplatBitSize)
2909 for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
2910 SplatBitSize < BitWidth;
2911 SplatBitSize = SplatBitSize * 2)
2912 SplatValue |= SplatValue.shl(SplatBitSize);
2914 // Make sure that variable 'Constant' is only set if 'SplatBitSize' is a
2915 // multiple of 'BitWidth'. Otherwise, we could propagate a wrong value.
2916 if (SplatBitSize % BitWidth == 0) {
2917 Constant = APInt::getAllOnesValue(BitWidth);
2918 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
2919 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
2924 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
2925 // actually legal and isn't going to get expanded, else this is a false
2927 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
2928 Load->getValueType(0),
2929 Load->getMemoryVT());
2931 // Resize the constant to the same size as the original memory access before
2932 // extension. If it is still the AllOnesValue then this AND is completely
2935 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
2938 switch (Load->getExtensionType()) {
2939 default: B = false; break;
2940 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
2942 case ISD::NON_EXTLOAD: B = true; break;
2945 if (B && Constant.isAllOnesValue()) {
2946 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
2947 // preserve semantics once we get rid of the AND.
2948 SDValue NewLoad(Load, 0);
2949 if (Load->getExtensionType() == ISD::EXTLOAD) {
2950 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
2951 Load->getValueType(0), SDLoc(Load),
2952 Load->getChain(), Load->getBasePtr(),
2953 Load->getOffset(), Load->getMemoryVT(),
2954 Load->getMemOperand());
2955 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
2956 if (Load->getNumValues() == 3) {
2957 // PRE/POST_INC loads have 3 values.
2958 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
2959 NewLoad.getValue(2) };
2960 CombineTo(Load, To, 3, true);
2962 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
2966 // Fold the AND away, taking care not to fold to the old load node if we
2968 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
2970 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2974 // fold (and (load x), 255) -> (zextload x, i8)
2975 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2976 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2977 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2978 (N0.getOpcode() == ISD::ANY_EXTEND &&
2979 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2980 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2981 LoadSDNode *LN0 = HasAnyExt
2982 ? cast<LoadSDNode>(N0.getOperand(0))
2983 : cast<LoadSDNode>(N0);
2984 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2985 LN0->isUnindexed() && N0.hasOneUse() && SDValue(LN0, 0).hasOneUse()) {
2986 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2987 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2988 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2989 EVT LoadedVT = LN0->getMemoryVT();
2990 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2992 if (ExtVT == LoadedVT &&
2993 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy,
2997 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
2998 LN0->getChain(), LN0->getBasePtr(), ExtVT,
2999 LN0->getMemOperand());
3001 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
3002 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3005 // Do not change the width of a volatile load.
3006 // Do not generate loads of non-round integer types since these can
3007 // be expensive (and would be wrong if the type is not byte sized).
3008 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
3009 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy,
3011 EVT PtrType = LN0->getOperand(1).getValueType();
3013 unsigned Alignment = LN0->getAlignment();
3014 SDValue NewPtr = LN0->getBasePtr();
3016 // For big endian targets, we need to add an offset to the pointer
3017 // to load the correct bytes. For little endian systems, we merely
3018 // need to read fewer bytes from the same pointer.
3019 if (TLI.isBigEndian()) {
3020 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
3021 unsigned EVTStoreBytes = ExtVT.getStoreSize();
3022 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
3023 NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0), PtrType,
3024 NewPtr, DAG.getConstant(PtrOff, PtrType));
3025 Alignment = MinAlign(Alignment, PtrOff);
3028 AddToWorklist(NewPtr.getNode());
3031 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
3032 LN0->getChain(), NewPtr,
3033 LN0->getPointerInfo(),
3034 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
3035 LN0->isInvariant(), Alignment, LN0->getAAInfo());
3037 CombineTo(LN0, Load, Load.getValue(1));
3038 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3044 if (SDValue Combined = visitANDLike(N0, N1, N))
3047 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
3048 if (N0.getOpcode() == N1.getOpcode()) {
3049 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3050 if (Tmp.getNode()) return Tmp;
3053 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
3054 // fold (and (sra)) -> (and (srl)) when possible.
3055 if (!VT.isVector() &&
3056 SimplifyDemandedBits(SDValue(N, 0)))
3057 return SDValue(N, 0);
3059 // fold (zext_inreg (extload x)) -> (zextload x)
3060 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
3061 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3062 EVT MemVT = LN0->getMemoryVT();
3063 // If we zero all the possible extended bits, then we can turn this into
3064 // a zextload if we are running before legalize or the operation is legal.
3065 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
3066 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
3067 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
3068 ((!LegalOperations && !LN0->isVolatile()) ||
3069 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) {
3070 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
3071 LN0->getChain(), LN0->getBasePtr(),
3072 MemVT, LN0->getMemOperand());
3074 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
3075 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3078 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
3079 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
3081 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3082 EVT MemVT = LN0->getMemoryVT();
3083 // If we zero all the possible extended bits, then we can turn this into
3084 // a zextload if we are running before legalize or the operation is legal.
3085 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
3086 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
3087 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
3088 ((!LegalOperations && !LN0->isVolatile()) ||
3089 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) {
3090 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
3091 LN0->getChain(), LN0->getBasePtr(),
3092 MemVT, LN0->getMemOperand());
3094 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
3095 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3098 // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const)
3099 if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) {
3100 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
3101 N0.getOperand(1), false);
3102 if (BSwap.getNode())
3109 /// Match (a >> 8) | (a << 8) as (bswap a) >> 16.
3110 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
3111 bool DemandHighBits) {
3112 if (!LegalOperations)
3115 EVT VT = N->getValueType(0);
3116 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
3118 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3121 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
3122 bool LookPassAnd0 = false;
3123 bool LookPassAnd1 = false;
3124 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
3126 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
3128 if (N0.getOpcode() == ISD::AND) {
3129 if (!N0.getNode()->hasOneUse())
3131 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3132 if (!N01C || N01C->getZExtValue() != 0xFF00)
3134 N0 = N0.getOperand(0);
3135 LookPassAnd0 = true;
3138 if (N1.getOpcode() == ISD::AND) {
3139 if (!N1.getNode()->hasOneUse())
3141 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3142 if (!N11C || N11C->getZExtValue() != 0xFF)
3144 N1 = N1.getOperand(0);
3145 LookPassAnd1 = true;
3148 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
3150 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
3152 if (!N0.getNode()->hasOneUse() ||
3153 !N1.getNode()->hasOneUse())
3156 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3157 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3160 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
3163 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
3164 SDValue N00 = N0->getOperand(0);
3165 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
3166 if (!N00.getNode()->hasOneUse())
3168 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
3169 if (!N001C || N001C->getZExtValue() != 0xFF)
3171 N00 = N00.getOperand(0);
3172 LookPassAnd0 = true;
3175 SDValue N10 = N1->getOperand(0);
3176 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
3177 if (!N10.getNode()->hasOneUse())
3179 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
3180 if (!N101C || N101C->getZExtValue() != 0xFF00)
3182 N10 = N10.getOperand(0);
3183 LookPassAnd1 = true;
3189 // Make sure everything beyond the low halfword gets set to zero since the SRL
3190 // 16 will clear the top bits.
3191 unsigned OpSizeInBits = VT.getSizeInBits();
3192 if (DemandHighBits && OpSizeInBits > 16) {
3193 // If the left-shift isn't masked out then the only way this is a bswap is
3194 // if all bits beyond the low 8 are 0. In that case the entire pattern
3195 // reduces to a left shift anyway: leave it for other parts of the combiner.
3199 // However, if the right shift isn't masked out then it might be because
3200 // it's not needed. See if we can spot that too.
3201 if (!LookPassAnd1 &&
3202 !DAG.MaskedValueIsZero(
3203 N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16)))
3207 SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00);
3208 if (OpSizeInBits > 16)
3209 Res = DAG.getNode(ISD::SRL, SDLoc(N), VT, Res,
3210 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
3214 /// Return true if the specified node is an element that makes up a 32-bit
3215 /// packed halfword byteswap.
3216 /// ((x & 0x000000ff) << 8) |
3217 /// ((x & 0x0000ff00) >> 8) |
3218 /// ((x & 0x00ff0000) << 8) |
3219 /// ((x & 0xff000000) >> 8)
3220 static bool isBSwapHWordElement(SDValue N, MutableArrayRef<SDNode *> Parts) {
3221 if (!N.getNode()->hasOneUse())
3224 unsigned Opc = N.getOpcode();
3225 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
3228 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3233 switch (N1C->getZExtValue()) {
3236 case 0xFF: Num = 0; break;
3237 case 0xFF00: Num = 1; break;
3238 case 0xFF0000: Num = 2; break;
3239 case 0xFF000000: Num = 3; break;
3242 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
3243 SDValue N0 = N.getOperand(0);
3244 if (Opc == ISD::AND) {
3245 if (Num == 0 || Num == 2) {
3247 // (x >> 8) & 0xff0000
3248 if (N0.getOpcode() != ISD::SRL)
3250 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3251 if (!C || C->getZExtValue() != 8)
3254 // (x << 8) & 0xff00
3255 // (x << 8) & 0xff000000
3256 if (N0.getOpcode() != ISD::SHL)
3258 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3259 if (!C || C->getZExtValue() != 8)
3262 } else if (Opc == ISD::SHL) {
3264 // (x & 0xff0000) << 8
3265 if (Num != 0 && Num != 2)
3267 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3268 if (!C || C->getZExtValue() != 8)
3270 } else { // Opc == ISD::SRL
3271 // (x & 0xff00) >> 8
3272 // (x & 0xff000000) >> 8
3273 if (Num != 1 && Num != 3)
3275 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3276 if (!C || C->getZExtValue() != 8)
3283 Parts[Num] = N0.getOperand(0).getNode();
3287 /// Match a 32-bit packed halfword bswap. That is
3288 /// ((x & 0x000000ff) << 8) |
3289 /// ((x & 0x0000ff00) >> 8) |
3290 /// ((x & 0x00ff0000) << 8) |
3291 /// ((x & 0xff000000) >> 8)
3292 /// => (rotl (bswap x), 16)
3293 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
3294 if (!LegalOperations)
3297 EVT VT = N->getValueType(0);
3300 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3304 // (or (or (and), (and)), (or (and), (and)))
3305 // (or (or (or (and), (and)), (and)), (and))
3306 if (N0.getOpcode() != ISD::OR)
3308 SDValue N00 = N0.getOperand(0);
3309 SDValue N01 = N0.getOperand(1);
3310 SDNode *Parts[4] = {};
3312 if (N1.getOpcode() == ISD::OR &&
3313 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) {
3314 // (or (or (and), (and)), (or (and), (and)))
3315 SDValue N000 = N00.getOperand(0);
3316 if (!isBSwapHWordElement(N000, Parts))
3319 SDValue N001 = N00.getOperand(1);
3320 if (!isBSwapHWordElement(N001, Parts))
3322 SDValue N010 = N01.getOperand(0);
3323 if (!isBSwapHWordElement(N010, Parts))
3325 SDValue N011 = N01.getOperand(1);
3326 if (!isBSwapHWordElement(N011, Parts))
3329 // (or (or (or (and), (and)), (and)), (and))
3330 if (!isBSwapHWordElement(N1, Parts))
3332 if (!isBSwapHWordElement(N01, Parts))
3334 if (N00.getOpcode() != ISD::OR)
3336 SDValue N000 = N00.getOperand(0);
3337 if (!isBSwapHWordElement(N000, Parts))
3339 SDValue N001 = N00.getOperand(1);
3340 if (!isBSwapHWordElement(N001, Parts))
3344 // Make sure the parts are all coming from the same node.
3345 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
3348 SDValue BSwap = DAG.getNode(ISD::BSWAP, SDLoc(N), VT,
3349 SDValue(Parts[0],0));
3351 // Result of the bswap should be rotated by 16. If it's not legal, then
3352 // do (x << 16) | (x >> 16).
3353 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
3354 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3355 return DAG.getNode(ISD::ROTL, SDLoc(N), VT, BSwap, ShAmt);
3356 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3357 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, BSwap, ShAmt);
3358 return DAG.getNode(ISD::OR, SDLoc(N), VT,
3359 DAG.getNode(ISD::SHL, SDLoc(N), VT, BSwap, ShAmt),
3360 DAG.getNode(ISD::SRL, SDLoc(N), VT, BSwap, ShAmt));
3363 /// This contains all DAGCombine rules which reduce two values combined by
3364 /// an Or operation to a single value \see visitANDLike().
3365 SDValue DAGCombiner::visitORLike(SDValue N0, SDValue N1, SDNode *LocReference) {
3366 EVT VT = N1.getValueType();
3367 // fold (or x, undef) -> -1
3368 if (!LegalOperations &&
3369 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3370 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3371 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
3373 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3374 SDValue LL, LR, RL, RR, CC0, CC1;
3375 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3376 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3377 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3379 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3380 LL.getValueType().isInteger()) {
3381 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3382 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3383 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3384 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3385 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(LR),
3386 LR.getValueType(), LL, RL);
3387 AddToWorklist(ORNode.getNode());
3388 return DAG.getSetCC(SDLoc(LocReference), VT, ORNode, LR, Op1);
3390 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3391 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3392 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3393 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3394 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(LR),
3395 LR.getValueType(), LL, RL);
3396 AddToWorklist(ANDNode.getNode());
3397 return DAG.getSetCC(SDLoc(LocReference), VT, ANDNode, LR, Op1);
3400 // canonicalize equivalent to ll == rl
3401 if (LL == RR && LR == RL) {
3402 Op1 = ISD::getSetCCSwappedOperands(Op1);
3405 if (LL == RL && LR == RR) {
3406 bool isInteger = LL.getValueType().isInteger();
3407 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3408 if (Result != ISD::SETCC_INVALID &&
3409 (!LegalOperations ||
3410 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
3411 TLI.isOperationLegal(ISD::SETCC,
3412 getSetCCResultType(N0.getValueType())))))
3413 return DAG.getSetCC(SDLoc(LocReference), N0.getValueType(),
3418 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3419 if (N0.getOpcode() == ISD::AND &&
3420 N1.getOpcode() == ISD::AND &&
3421 N0.getOperand(1).getOpcode() == ISD::Constant &&
3422 N1.getOperand(1).getOpcode() == ISD::Constant &&
3423 // Don't increase # computations.
3424 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3425 // We can only do this xform if we know that bits from X that are set in C2
3426 // but not in C1 are already zero. Likewise for Y.
3427 const APInt &LHSMask =
3428 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3429 const APInt &RHSMask =
3430 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3432 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3433 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3434 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3435 N0.getOperand(0), N1.getOperand(0));
3436 return DAG.getNode(ISD::AND, SDLoc(LocReference), VT, X,
3437 DAG.getConstant(LHSMask | RHSMask, VT));
3441 // (or (and X, M), (and X, N)) -> (and X, (or M, N))
3442 if (N0.getOpcode() == ISD::AND &&
3443 N1.getOpcode() == ISD::AND &&
3444 N0.getOperand(0) == N1.getOperand(0) &&
3445 // Don't increase # computations.
3446 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3447 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3448 N0.getOperand(1), N1.getOperand(1));
3449 return DAG.getNode(ISD::AND, SDLoc(LocReference), VT, N0.getOperand(0), X);
3455 SDValue DAGCombiner::visitOR(SDNode *N) {
3456 SDValue N0 = N->getOperand(0);
3457 SDValue N1 = N->getOperand(1);
3458 EVT VT = N1.getValueType();
3461 if (VT.isVector()) {
3462 if (SDValue FoldedVOp = SimplifyVBinOp(N))
3465 // fold (or x, 0) -> x, vector edition
3466 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3468 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3471 // fold (or x, -1) -> -1, vector edition
3472 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3473 // do not return N0, because undef node may exist in N0
3474 return DAG.getConstant(
3475 APInt::getAllOnesValue(
3476 N0.getValueType().getScalarType().getSizeInBits()),
3478 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3479 // do not return N1, because undef node may exist in N1
3480 return DAG.getConstant(
3481 APInt::getAllOnesValue(
3482 N1.getValueType().getScalarType().getSizeInBits()),
3485 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask1)
3486 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf B, A, Mask2)
3487 // Do this only if the resulting shuffle is legal.
3488 if (isa<ShuffleVectorSDNode>(N0) &&
3489 isa<ShuffleVectorSDNode>(N1) &&
3490 // Avoid folding a node with illegal type.
3491 TLI.isTypeLegal(VT) &&
3492 N0->getOperand(1) == N1->getOperand(1) &&
3493 ISD::isBuildVectorAllZeros(N0.getOperand(1).getNode())) {
3494 bool CanFold = true;
3495 unsigned NumElts = VT.getVectorNumElements();
3496 const ShuffleVectorSDNode *SV0 = cast<ShuffleVectorSDNode>(N0);
3497 const ShuffleVectorSDNode *SV1 = cast<ShuffleVectorSDNode>(N1);
3498 // We construct two shuffle masks:
3499 // - Mask1 is a shuffle mask for a shuffle with N0 as the first operand
3500 // and N1 as the second operand.
3501 // - Mask2 is a shuffle mask for a shuffle with N1 as the first operand
3502 // and N0 as the second operand.
3503 // We do this because OR is commutable and therefore there might be
3504 // two ways to fold this node into a shuffle.
3505 SmallVector<int,4> Mask1;
3506 SmallVector<int,4> Mask2;
3508 for (unsigned i = 0; i != NumElts && CanFold; ++i) {
3509 int M0 = SV0->getMaskElt(i);
3510 int M1 = SV1->getMaskElt(i);
3512 // Both shuffle indexes are undef. Propagate Undef.
3513 if (M0 < 0 && M1 < 0) {
3514 Mask1.push_back(M0);
3515 Mask2.push_back(M0);
3519 if (M0 < 0 || M1 < 0 ||
3520 (M0 < (int)NumElts && M1 < (int)NumElts) ||
3521 (M0 >= (int)NumElts && M1 >= (int)NumElts)) {
3526 Mask1.push_back(M0 < (int)NumElts ? M0 : M1 + NumElts);
3527 Mask2.push_back(M1 < (int)NumElts ? M1 : M0 + NumElts);
3531 // Fold this sequence only if the resulting shuffle is 'legal'.
3532 if (TLI.isShuffleMaskLegal(Mask1, VT))
3533 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0),
3534 N1->getOperand(0), &Mask1[0]);
3535 if (TLI.isShuffleMaskLegal(Mask2, VT))
3536 return DAG.getVectorShuffle(VT, SDLoc(N), N1->getOperand(0),
3537 N0->getOperand(0), &Mask2[0]);
3542 // fold (or c1, c2) -> c1|c2
3543 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3544 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3546 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
3547 // canonicalize constant to RHS
3549 return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0);
3550 // fold (or x, 0) -> x
3551 if (N1C && N1C->isNullValue())
3553 // fold (or x, -1) -> -1
3554 if (N1C && N1C->isAllOnesValue())
3556 // fold (or x, c) -> c iff (x & ~c) == 0
3557 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3560 if (SDValue Combined = visitORLike(N0, N1, N))
3563 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3564 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
3565 if (BSwap.getNode())
3567 BSwap = MatchBSwapHWordLow(N, N0, N1);
3568 if (BSwap.getNode())
3572 if (SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1))
3574 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3575 // iff (c1 & c2) == 0.
3576 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3577 isa<ConstantSDNode>(N0.getOperand(1))) {
3578 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3579 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) {
3580 if (SDValue COR = DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1))
3582 ISD::AND, SDLoc(N), VT,
3583 DAG.getNode(ISD::OR, SDLoc(N0), VT, N0.getOperand(0), N1), COR);
3587 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3588 if (N0.getOpcode() == N1.getOpcode()) {
3589 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3590 if (Tmp.getNode()) return Tmp;
3593 // See if this is some rotate idiom.
3594 if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N)))
3595 return SDValue(Rot, 0);
3597 // Simplify the operands using demanded-bits information.
3598 if (!VT.isVector() &&
3599 SimplifyDemandedBits(SDValue(N, 0)))
3600 return SDValue(N, 0);
3605 /// Match "(X shl/srl V1) & V2" where V2 may not be present.
3606 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3607 if (Op.getOpcode() == ISD::AND) {
3608 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3609 Mask = Op.getOperand(1);
3610 Op = Op.getOperand(0);
3616 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3624 // Return true if we can prove that, whenever Neg and Pos are both in the
3625 // range [0, OpSize), Neg == (Pos == 0 ? 0 : OpSize - Pos). This means that
3626 // for two opposing shifts shift1 and shift2 and a value X with OpBits bits:
3628 // (or (shift1 X, Neg), (shift2 X, Pos))
3630 // reduces to a rotate in direction shift2 by Pos or (equivalently) a rotate
3631 // in direction shift1 by Neg. The range [0, OpSize) means that we only need
3632 // to consider shift amounts with defined behavior.
3633 static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned OpSize) {
3634 // If OpSize is a power of 2 then:
3636 // (a) (Pos == 0 ? 0 : OpSize - Pos) == (OpSize - Pos) & (OpSize - 1)
3637 // (b) Neg == Neg & (OpSize - 1) whenever Neg is in [0, OpSize).
3639 // So if OpSize is a power of 2 and Neg is (and Neg', OpSize-1), we check
3640 // for the stronger condition:
3642 // Neg & (OpSize - 1) == (OpSize - Pos) & (OpSize - 1) [A]
3644 // for all Neg and Pos. Since Neg & (OpSize - 1) == Neg' & (OpSize - 1)
3645 // we can just replace Neg with Neg' for the rest of the function.
3647 // In other cases we check for the even stronger condition:
3649 // Neg == OpSize - Pos [B]
3651 // for all Neg and Pos. Note that the (or ...) then invokes undefined
3652 // behavior if Pos == 0 (and consequently Neg == OpSize).
3654 // We could actually use [A] whenever OpSize is a power of 2, but the
3655 // only extra cases that it would match are those uninteresting ones
3656 // where Neg and Pos are never in range at the same time. E.g. for
3657 // OpSize == 32, using [A] would allow a Neg of the form (sub 64, Pos)
3658 // as well as (sub 32, Pos), but:
3660 // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos))
3662 // always invokes undefined behavior for 32-bit X.
3664 // Below, Mask == OpSize - 1 when using [A] and is all-ones otherwise.
3665 unsigned MaskLoBits = 0;
3666 if (Neg.getOpcode() == ISD::AND &&
3667 isPowerOf2_64(OpSize) &&
3668 Neg.getOperand(1).getOpcode() == ISD::Constant &&
3669 cast<ConstantSDNode>(Neg.getOperand(1))->getAPIntValue() == OpSize - 1) {
3670 Neg = Neg.getOperand(0);
3671 MaskLoBits = Log2_64(OpSize);
3674 // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1.
3675 if (Neg.getOpcode() != ISD::SUB)
3677 ConstantSDNode *NegC = dyn_cast<ConstantSDNode>(Neg.getOperand(0));
3680 SDValue NegOp1 = Neg.getOperand(1);
3682 // On the RHS of [A], if Pos is Pos' & (OpSize - 1), just replace Pos with
3683 // Pos'. The truncation is redundant for the purpose of the equality.
3685 Pos.getOpcode() == ISD::AND &&
3686 Pos.getOperand(1).getOpcode() == ISD::Constant &&
3687 cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() == OpSize - 1)
3688 Pos = Pos.getOperand(0);
3690 // The condition we need is now:
3692 // (NegC - NegOp1) & Mask == (OpSize - Pos) & Mask
3694 // If NegOp1 == Pos then we need:
3696 // OpSize & Mask == NegC & Mask
3698 // (because "x & Mask" is a truncation and distributes through subtraction).
3701 Width = NegC->getAPIntValue();
3702 // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC.
3703 // Then the condition we want to prove becomes:
3705 // (NegC - NegOp1) & Mask == (OpSize - (NegOp1 + PosC)) & Mask
3707 // which, again because "x & Mask" is a truncation, becomes:
3709 // NegC & Mask == (OpSize - PosC) & Mask
3710 // OpSize & Mask == (NegC + PosC) & Mask
3711 else if (Pos.getOpcode() == ISD::ADD &&
3712 Pos.getOperand(0) == NegOp1 &&
3713 Pos.getOperand(1).getOpcode() == ISD::Constant)
3714 Width = (cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() +
3715 NegC->getAPIntValue());
3719 // Now we just need to check that OpSize & Mask == Width & Mask.
3721 // Opsize & Mask is 0 since Mask is Opsize - 1.
3722 return Width.getLoBits(MaskLoBits) == 0;
3723 return Width == OpSize;
3726 // A subroutine of MatchRotate used once we have found an OR of two opposite
3727 // shifts of Shifted. If Neg == <operand size> - Pos then the OR reduces
3728 // to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the
3729 // former being preferred if supported. InnerPos and InnerNeg are Pos and
3730 // Neg with outer conversions stripped away.
3731 SDNode *DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos,
3732 SDValue Neg, SDValue InnerPos,
3733 SDValue InnerNeg, unsigned PosOpcode,
3734 unsigned NegOpcode, SDLoc DL) {
3735 // fold (or (shl x, (*ext y)),
3736 // (srl x, (*ext (sub 32, y)))) ->
3737 // (rotl x, y) or (rotr x, (sub 32, y))
3739 // fold (or (shl x, (*ext (sub 32, y))),
3740 // (srl x, (*ext y))) ->
3741 // (rotr x, y) or (rotl x, (sub 32, y))
3742 EVT VT = Shifted.getValueType();
3743 if (matchRotateSub(InnerPos, InnerNeg, VT.getSizeInBits())) {
3744 bool HasPos = TLI.isOperationLegalOrCustom(PosOpcode, VT);
3745 return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted,
3746 HasPos ? Pos : Neg).getNode();
3752 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3753 // idioms for rotate, and if the target supports rotation instructions, generate
3755 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL) {
3756 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3757 EVT VT = LHS.getValueType();
3758 if (!TLI.isTypeLegal(VT)) return nullptr;
3760 // The target must have at least one rotate flavor.
3761 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3762 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3763 if (!HasROTL && !HasROTR) return nullptr;
3765 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3766 SDValue LHSShift; // The shift.
3767 SDValue LHSMask; // AND value if any.
3768 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3769 return nullptr; // Not part of a rotate.
3771 SDValue RHSShift; // The shift.
3772 SDValue RHSMask; // AND value if any.
3773 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3774 return nullptr; // Not part of a rotate.
3776 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3777 return nullptr; // Not shifting the same value.
3779 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3780 return nullptr; // Shifts must disagree.
3782 // Canonicalize shl to left side in a shl/srl pair.
3783 if (RHSShift.getOpcode() == ISD::SHL) {
3784 std::swap(LHS, RHS);
3785 std::swap(LHSShift, RHSShift);
3786 std::swap(LHSMask , RHSMask );
3789 unsigned OpSizeInBits = VT.getSizeInBits();
3790 SDValue LHSShiftArg = LHSShift.getOperand(0);
3791 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3792 SDValue RHSShiftArg = RHSShift.getOperand(0);
3793 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3795 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3796 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3797 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3798 RHSShiftAmt.getOpcode() == ISD::Constant) {
3799 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3800 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3801 if ((LShVal + RShVal) != OpSizeInBits)
3804 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3805 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
3807 // If there is an AND of either shifted operand, apply it to the result.
3808 if (LHSMask.getNode() || RHSMask.getNode()) {
3809 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3811 if (LHSMask.getNode()) {
3812 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3813 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3815 if (RHSMask.getNode()) {
3816 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3817 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3820 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3823 return Rot.getNode();
3826 // If there is a mask here, and we have a variable shift, we can't be sure
3827 // that we're masking out the right stuff.
3828 if (LHSMask.getNode() || RHSMask.getNode())
3831 // If the shift amount is sign/zext/any-extended just peel it off.
3832 SDValue LExtOp0 = LHSShiftAmt;
3833 SDValue RExtOp0 = RHSShiftAmt;
3834 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3835 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3836 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3837 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3838 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3839 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3840 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3841 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3842 LExtOp0 = LHSShiftAmt.getOperand(0);
3843 RExtOp0 = RHSShiftAmt.getOperand(0);
3846 SDNode *TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt,
3847 LExtOp0, RExtOp0, ISD::ROTL, ISD::ROTR, DL);
3851 SDNode *TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt,
3852 RExtOp0, LExtOp0, ISD::ROTR, ISD::ROTL, DL);
3859 SDValue DAGCombiner::visitXOR(SDNode *N) {
3860 SDValue N0 = N->getOperand(0);
3861 SDValue N1 = N->getOperand(1);
3862 EVT VT = N0.getValueType();
3865 if (VT.isVector()) {
3866 if (SDValue FoldedVOp = SimplifyVBinOp(N))
3869 // fold (xor x, 0) -> x, vector edition
3870 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3872 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3876 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3877 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3878 return DAG.getConstant(0, VT);
3879 // fold (xor x, undef) -> undef
3880 if (N0.getOpcode() == ISD::UNDEF)
3882 if (N1.getOpcode() == ISD::UNDEF)
3884 // fold (xor c1, c2) -> c1^c2
3885 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3886 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3888 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3889 // canonicalize constant to RHS
3891 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
3892 // fold (xor x, 0) -> x
3893 if (N1C && N1C->isNullValue())
3896 if (SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1))
3899 // fold !(x cc y) -> (x !cc y)
3900 SDValue LHS, RHS, CC;
3901 if (TLI.isConstTrueVal(N1.getNode()) && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3902 bool isInt = LHS.getValueType().isInteger();
3903 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3906 if (!LegalOperations ||
3907 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) {
3908 switch (N0.getOpcode()) {
3910 llvm_unreachable("Unhandled SetCC Equivalent!");
3912 return DAG.getSetCC(SDLoc(N), VT, LHS, RHS, NotCC);
3913 case ISD::SELECT_CC:
3914 return DAG.getSelectCC(SDLoc(N), LHS, RHS, N0.getOperand(2),
3915 N0.getOperand(3), NotCC);
3920 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3921 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3922 N0.getNode()->hasOneUse() &&
3923 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3924 SDValue V = N0.getOperand(0);
3925 V = DAG.getNode(ISD::XOR, SDLoc(N0), V.getValueType(), V,
3926 DAG.getConstant(1, V.getValueType()));
3927 AddToWorklist(V.getNode());
3928 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V);
3931 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3932 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3933 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3934 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3935 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3936 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3937 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3938 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3939 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
3940 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3943 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3944 if (N1C && N1C->isAllOnesValue() &&
3945 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3946 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3947 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3948 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3949 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3950 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3951 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
3952 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3955 // fold (xor (and x, y), y) -> (and (not x), y)
3956 if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3957 N0->getOperand(1) == N1) {
3958 SDValue X = N0->getOperand(0);
3959 SDValue NotX = DAG.getNOT(SDLoc(X), X, VT);
3960 AddToWorklist(NotX.getNode());
3961 return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1);
3963 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3964 if (N1C && N0.getOpcode() == ISD::XOR) {
3965 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3966 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3968 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(1),
3969 DAG.getConstant(N1C->getAPIntValue() ^
3970 N00C->getAPIntValue(), VT));
3972 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(0),
3973 DAG.getConstant(N1C->getAPIntValue() ^
3974 N01C->getAPIntValue(), VT));
3976 // fold (xor x, x) -> 0
3978 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
3980 // fold (xor (shl 1, x), -1) -> (rotl ~1, x)
3981 // Here is a concrete example of this equivalence:
3983 // i16 shl == 1 << 14 == 16384 == 0b0100000000000000
3984 // i16 xor == ~(1 << 14) == 49151 == 0b1011111111111111
3988 // i16 ~1 == 0b1111111111111110
3989 // i16 rol(~1, 14) == 0b1011111111111111
3991 // Some additional tips to help conceptualize this transform:
3992 // - Try to see the operation as placing a single zero in a value of all ones.
3993 // - There exists no value for x which would allow the result to contain zero.
3994 // - Values of x larger than the bitwidth are undefined and do not require a
3995 // consistent result.
3996 // - Pushing the zero left requires shifting one bits in from the right.
3997 // A rotate left of ~1 is a nice way of achieving the desired result.
3998 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3999 if (auto *N1C = dyn_cast<ConstantSDNode>(N1.getNode()))
4000 if (N0.getOpcode() == ISD::SHL)
4001 if (auto *ShlLHS = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
4002 if (N1C->isAllOnesValue() && ShlLHS->isOne())
4003 return DAG.getNode(ISD::ROTL, SDLoc(N), VT, DAG.getConstant(~1, VT),
4006 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
4007 if (N0.getOpcode() == N1.getOpcode()) {
4008 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
4009 if (Tmp.getNode()) return Tmp;
4012 // Simplify the expression using non-local knowledge.
4013 if (!VT.isVector() &&
4014 SimplifyDemandedBits(SDValue(N, 0)))
4015 return SDValue(N, 0);
4020 /// Handle transforms common to the three shifts, when the shift amount is a
4022 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, ConstantSDNode *Amt) {
4023 // We can't and shouldn't fold opaque constants.
4024 if (Amt->isOpaque())
4027 SDNode *LHS = N->getOperand(0).getNode();
4028 if (!LHS->hasOneUse()) return SDValue();
4030 // We want to pull some binops through shifts, so that we have (and (shift))
4031 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
4032 // thing happens with address calculations, so it's important to canonicalize
4034 bool HighBitSet = false; // Can we transform this if the high bit is set?
4036 switch (LHS->getOpcode()) {
4037 default: return SDValue();
4040 HighBitSet = false; // We can only transform sra if the high bit is clear.
4043 HighBitSet = true; // We can only transform sra if the high bit is set.
4046 if (N->getOpcode() != ISD::SHL)
4047 return SDValue(); // only shl(add) not sr[al](add).
4048 HighBitSet = false; // We can only transform sra if the high bit is clear.
4052 // We require the RHS of the binop to be a constant and not opaque as well.
4053 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
4054 if (!BinOpCst || BinOpCst->isOpaque()) return SDValue();
4056 // FIXME: disable this unless the input to the binop is a shift by a constant.
4057 // If it is not a shift, it pessimizes some common cases like:
4059 // void foo(int *X, int i) { X[i & 1235] = 1; }
4060 // int bar(int *X, int i) { return X[i & 255]; }
4061 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
4062 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
4063 BinOpLHSVal->getOpcode() != ISD::SRA &&
4064 BinOpLHSVal->getOpcode() != ISD::SRL) ||
4065 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
4068 EVT VT = N->getValueType(0);
4070 // If this is a signed shift right, and the high bit is modified by the
4071 // logical operation, do not perform the transformation. The highBitSet
4072 // boolean indicates the value of the high bit of the constant which would
4073 // cause it to be modified for this operation.
4074 if (N->getOpcode() == ISD::SRA) {
4075 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
4076 if (BinOpRHSSignSet != HighBitSet)
4080 if (!TLI.isDesirableToCommuteWithShift(LHS))
4083 // Fold the constants, shifting the binop RHS by the shift amount.
4084 SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)),
4086 LHS->getOperand(1), N->getOperand(1));
4087 assert(isa<ConstantSDNode>(NewRHS) && "Folding was not successful!");
4089 // Create the new shift.
4090 SDValue NewShift = DAG.getNode(N->getOpcode(),
4091 SDLoc(LHS->getOperand(0)),
4092 VT, LHS->getOperand(0), N->getOperand(1));
4094 // Create the new binop.
4095 return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS);
4098 SDValue DAGCombiner::distributeTruncateThroughAnd(SDNode *N) {
4099 assert(N->getOpcode() == ISD::TRUNCATE);
4100 assert(N->getOperand(0).getOpcode() == ISD::AND);
4102 // (truncate:TruncVT (and N00, N01C)) -> (and (truncate:TruncVT N00), TruncC)
4103 if (N->hasOneUse() && N->getOperand(0).hasOneUse()) {
4104 SDValue N01 = N->getOperand(0).getOperand(1);
4106 if (ConstantSDNode *N01C = isConstOrConstSplat(N01)) {
4107 EVT TruncVT = N->getValueType(0);
4108 SDValue N00 = N->getOperand(0).getOperand(0);
4109 APInt TruncC = N01C->getAPIntValue();
4110 TruncC = TruncC.trunc(TruncVT.getScalarSizeInBits());
4112 return DAG.getNode(ISD::AND, SDLoc(N), TruncVT,
4113 DAG.getNode(ISD::TRUNCATE, SDLoc(N), TruncVT, N00),
4114 DAG.getConstant(TruncC, TruncVT));
4121 SDValue DAGCombiner::visitRotate(SDNode *N) {
4122 // fold (rot* x, (trunc (and y, c))) -> (rot* x, (and (trunc y), (trunc c))).
4123 if (N->getOperand(1).getOpcode() == ISD::TRUNCATE &&
4124 N->getOperand(1).getOperand(0).getOpcode() == ISD::AND) {
4125 SDValue NewOp1 = distributeTruncateThroughAnd(N->getOperand(1).getNode());
4126 if (NewOp1.getNode())
4127 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
4128 N->getOperand(0), NewOp1);
4133 SDValue DAGCombiner::visitSHL(SDNode *N) {
4134 SDValue N0 = N->getOperand(0);
4135 SDValue N1 = N->getOperand(1);
4136 EVT VT = N0.getValueType();
4137 unsigned OpSizeInBits = VT.getScalarSizeInBits();
4140 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4141 if (VT.isVector()) {
4142 if (SDValue FoldedVOp = SimplifyVBinOp(N))
4145 BuildVectorSDNode *N1CV = dyn_cast<BuildVectorSDNode>(N1);
4146 // If setcc produces all-one true value then:
4147 // (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<<N1CV)
4148 if (N1CV && N1CV->isConstant()) {
4149 if (N0.getOpcode() == ISD::AND) {
4150 SDValue N00 = N0->getOperand(0);
4151 SDValue N01 = N0->getOperand(1);
4152 BuildVectorSDNode *N01CV = dyn_cast<BuildVectorSDNode>(N01);
4154 if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC &&
4155 TLI.getBooleanContents(N00.getOperand(0).getValueType()) ==
4156 TargetLowering::ZeroOrNegativeOneBooleanContent) {
4157 if (SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, VT, N01CV, N1CV))
4158 return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C);
4161 N1C = isConstOrConstSplat(N1);
4166 // fold (shl c1, c2) -> c1<<c2
4167 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4169 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
4170 // fold (shl 0, x) -> 0
4171 if (N0C && N0C->isNullValue())
4173 // fold (shl x, c >= size(x)) -> undef
4174 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4175 return DAG.getUNDEF(VT);
4176 // fold (shl x, 0) -> x
4177 if (N1C && N1C->isNullValue())
4179 // fold (shl undef, x) -> 0
4180 if (N0.getOpcode() == ISD::UNDEF)
4181 return DAG.getConstant(0, VT);
4182 // if (shl x, c) is known to be zero, return 0
4183 if (DAG.MaskedValueIsZero(SDValue(N, 0),
4184 APInt::getAllOnesValue(OpSizeInBits)))
4185 return DAG.getConstant(0, VT);
4186 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
4187 if (N1.getOpcode() == ISD::TRUNCATE &&
4188 N1.getOperand(0).getOpcode() == ISD::AND) {
4189 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4190 if (NewOp1.getNode())
4191 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, NewOp1);
4194 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4195 return SDValue(N, 0);
4197 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
4198 if (N1C && N0.getOpcode() == ISD::SHL) {
4199 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4200 uint64_t c1 = N0C1->getZExtValue();
4201 uint64_t c2 = N1C->getZExtValue();
4202 if (c1 + c2 >= OpSizeInBits)
4203 return DAG.getConstant(0, VT);
4204 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4205 DAG.getConstant(c1 + c2, N1.getValueType()));
4209 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
4210 // For this to be valid, the second form must not preserve any of the bits
4211 // that are shifted out by the inner shift in the first form. This means
4212 // the outer shift size must be >= the number of bits added by the ext.
4213 // As a corollary, we don't care what kind of ext it is.
4214 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
4215 N0.getOpcode() == ISD::ANY_EXTEND ||
4216 N0.getOpcode() == ISD::SIGN_EXTEND) &&
4217 N0.getOperand(0).getOpcode() == ISD::SHL) {
4218 SDValue N0Op0 = N0.getOperand(0);
4219 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4220 uint64_t c1 = N0Op0C1->getZExtValue();
4221 uint64_t c2 = N1C->getZExtValue();
4222 EVT InnerShiftVT = N0Op0.getValueType();
4223 uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits();
4224 if (c2 >= OpSizeInBits - InnerShiftSize) {
4225 if (c1 + c2 >= OpSizeInBits)
4226 return DAG.getConstant(0, VT);
4227 return DAG.getNode(ISD::SHL, SDLoc(N0), VT,
4228 DAG.getNode(N0.getOpcode(), SDLoc(N0), VT,
4229 N0Op0->getOperand(0)),
4230 DAG.getConstant(c1 + c2, N1.getValueType()));
4235 // fold (shl (zext (srl x, C)), C) -> (zext (shl (srl x, C), C))
4236 // Only fold this if the inner zext has no other uses to avoid increasing
4237 // the total number of instructions.
4238 if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() &&
4239 N0.getOperand(0).getOpcode() == ISD::SRL) {
4240 SDValue N0Op0 = N0.getOperand(0);
4241 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4242 uint64_t c1 = N0Op0C1->getZExtValue();
4243 if (c1 < VT.getScalarSizeInBits()) {
4244 uint64_t c2 = N1C->getZExtValue();
4246 SDValue NewOp0 = N0.getOperand(0);
4247 EVT CountVT = NewOp0.getOperand(1).getValueType();
4248 SDValue NewSHL = DAG.getNode(ISD::SHL, SDLoc(N), NewOp0.getValueType(),
4249 NewOp0, DAG.getConstant(c2, CountVT));
4250 AddToWorklist(NewSHL.getNode());
4251 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
4257 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
4258 // (and (srl x, (sub c1, c2), MASK)
4259 // Only fold this if the inner shift has no other uses -- if it does, folding
4260 // this will increase the total number of instructions.
4261 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
4262 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4263 uint64_t c1 = N0C1->getZExtValue();
4264 if (c1 < OpSizeInBits) {
4265 uint64_t c2 = N1C->getZExtValue();
4266 APInt Mask = APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - c1);
4269 Mask = Mask.shl(c2 - c1);
4270 Shift = DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4271 DAG.getConstant(c2 - c1, N1.getValueType()));
4273 Mask = Mask.lshr(c1 - c2);
4274 Shift = DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4275 DAG.getConstant(c1 - c2, N1.getValueType()));
4277 return DAG.getNode(ISD::AND, SDLoc(N0), VT, Shift,
4278 DAG.getConstant(Mask, VT));
4282 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
4283 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
4284 unsigned BitSize = VT.getScalarSizeInBits();
4285 SDValue HiBitsMask =
4286 DAG.getConstant(APInt::getHighBitsSet(BitSize,
4287 BitSize - N1C->getZExtValue()), VT);
4288 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4292 // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
4293 // Variant of version done on multiply, except mul by a power of 2 is turned
4296 if (N1C && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
4297 (isa<ConstantSDNode>(N0.getOperand(1)) ||
4298 isConstantSplatVector(N0.getOperand(1).getNode(), Val))) {
4299 SDValue Shl0 = DAG.getNode(ISD::SHL, SDLoc(N0), VT, N0.getOperand(0), N1);
4300 SDValue Shl1 = DAG.getNode(ISD::SHL, SDLoc(N1), VT, N0.getOperand(1), N1);
4301 return DAG.getNode(ISD::ADD, SDLoc(N), VT, Shl0, Shl1);
4305 SDValue NewSHL = visitShiftByConstant(N, N1C);
4306 if (NewSHL.getNode())
4313 SDValue DAGCombiner::visitSRA(SDNode *N) {
4314 SDValue N0 = N->getOperand(0);
4315 SDValue N1 = N->getOperand(1);
4316 EVT VT = N0.getValueType();
4317 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4320 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4321 if (VT.isVector()) {
4322 if (SDValue FoldedVOp = SimplifyVBinOp(N))
4325 N1C = isConstOrConstSplat(N1);
4328 // fold (sra c1, c2) -> (sra c1, c2)
4329 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4331 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
4332 // fold (sra 0, x) -> 0
4333 if (N0C && N0C->isNullValue())
4335 // fold (sra -1, x) -> -1
4336 if (N0C && N0C->isAllOnesValue())
4338 // fold (sra x, (setge c, size(x))) -> undef
4339 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4340 return DAG.getUNDEF(VT);
4341 // fold (sra x, 0) -> x
4342 if (N1C && N1C->isNullValue())
4344 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
4346 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
4347 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
4348 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
4350 ExtVT = EVT::getVectorVT(*DAG.getContext(),
4351 ExtVT, VT.getVectorNumElements());
4352 if ((!LegalOperations ||
4353 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
4354 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
4355 N0.getOperand(0), DAG.getValueType(ExtVT));
4358 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
4359 if (N1C && N0.getOpcode() == ISD::SRA) {
4360 if (ConstantSDNode *C1 = isConstOrConstSplat(N0.getOperand(1))) {
4361 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
4362 if (Sum >= OpSizeInBits)
4363 Sum = OpSizeInBits - 1;
4364 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0.getOperand(0),
4365 DAG.getConstant(Sum, N1.getValueType()));
4369 // fold (sra (shl X, m), (sub result_size, n))
4370 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
4371 // result_size - n != m.
4372 // If truncate is free for the target sext(shl) is likely to result in better
4374 if (N0.getOpcode() == ISD::SHL && N1C) {
4375 // Get the two constanst of the shifts, CN0 = m, CN = n.
4376 const ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1));
4378 LLVMContext &Ctx = *DAG.getContext();
4379 // Determine what the truncate's result bitsize and type would be.
4380 EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - N1C->getZExtValue());
4383 TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorNumElements());
4385 // Determine the residual right-shift amount.
4386 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
4388 // If the shift is not a no-op (in which case this should be just a sign
4389 // extend already), the truncated to type is legal, sign_extend is legal
4390 // on that type, and the truncate to that type is both legal and free,
4391 // perform the transform.
4392 if ((ShiftAmt > 0) &&
4393 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
4394 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
4395 TLI.isTruncateFree(VT, TruncVT)) {
4397 SDValue Amt = DAG.getConstant(ShiftAmt,
4398 getShiftAmountTy(N0.getOperand(0).getValueType()));
4399 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0), VT,
4400 N0.getOperand(0), Amt);
4401 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), TruncVT,
4403 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N),
4404 N->getValueType(0), Trunc);
4409 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
4410 if (N1.getOpcode() == ISD::TRUNCATE &&
4411 N1.getOperand(0).getOpcode() == ISD::AND) {
4412 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4413 if (NewOp1.getNode())
4414 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, NewOp1);
4417 // fold (sra (trunc (srl x, c1)), c2) -> (trunc (sra x, c1 + c2))
4418 // if c1 is equal to the number of bits the trunc removes
4419 if (N0.getOpcode() == ISD::TRUNCATE &&
4420 (N0.getOperand(0).getOpcode() == ISD::SRL ||
4421 N0.getOperand(0).getOpcode() == ISD::SRA) &&
4422 N0.getOperand(0).hasOneUse() &&
4423 N0.getOperand(0).getOperand(1).hasOneUse() &&
4425 SDValue N0Op0 = N0.getOperand(0);
4426 if (ConstantSDNode *LargeShift = isConstOrConstSplat(N0Op0.getOperand(1))) {
4427 unsigned LargeShiftVal = LargeShift->getZExtValue();
4428 EVT LargeVT = N0Op0.getValueType();
4430 if (LargeVT.getScalarSizeInBits() - OpSizeInBits == LargeShiftVal) {
4432 DAG.getConstant(LargeShiftVal + N1C->getZExtValue(),
4433 getShiftAmountTy(N0Op0.getOperand(0).getValueType()));
4434 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), LargeVT,
4435 N0Op0.getOperand(0), Amt);
4436 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, SRA);
4441 // Simplify, based on bits shifted out of the LHS.
4442 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4443 return SDValue(N, 0);
4446 // If the sign bit is known to be zero, switch this to a SRL.
4447 if (DAG.SignBitIsZero(N0))
4448 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1);
4451 SDValue NewSRA = visitShiftByConstant(N, N1C);
4452 if (NewSRA.getNode())
4459 SDValue DAGCombiner::visitSRL(SDNode *N) {
4460 SDValue N0 = N->getOperand(0);
4461 SDValue N1 = N->getOperand(1);
4462 EVT VT = N0.getValueType();
4463 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4466 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4467 if (VT.isVector()) {
4468 if (SDValue FoldedVOp = SimplifyVBinOp(N))
4471 N1C = isConstOrConstSplat(N1);
4474 // fold (srl c1, c2) -> c1 >>u c2
4475 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4477 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
4478 // fold (srl 0, x) -> 0
4479 if (N0C && N0C->isNullValue())
4481 // fold (srl x, c >= size(x)) -> undef
4482 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4483 return DAG.getUNDEF(VT);
4484 // fold (srl x, 0) -> x
4485 if (N1C && N1C->isNullValue())
4487 // if (srl x, c) is known to be zero, return 0
4488 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
4489 APInt::getAllOnesValue(OpSizeInBits)))
4490 return DAG.getConstant(0, VT);
4492 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
4493 if (N1C && N0.getOpcode() == ISD::SRL) {
4494 if (ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1))) {
4495 uint64_t c1 = N01C->getZExtValue();
4496 uint64_t c2 = N1C->getZExtValue();
4497 if (c1 + c2 >= OpSizeInBits)
4498 return DAG.getConstant(0, VT);
4499 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4500 DAG.getConstant(c1 + c2, N1.getValueType()));
4504 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
4505 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
4506 N0.getOperand(0).getOpcode() == ISD::SRL &&
4507 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
4509 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
4510 uint64_t c2 = N1C->getZExtValue();
4511 EVT InnerShiftVT = N0.getOperand(0).getValueType();
4512 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
4513 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
4514 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
4515 if (c1 + OpSizeInBits == InnerShiftSize) {
4516 if (c1 + c2 >= InnerShiftSize)
4517 return DAG.getConstant(0, VT);
4518 return DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT,
4519 DAG.getNode(ISD::SRL, SDLoc(N0), InnerShiftVT,
4520 N0.getOperand(0)->getOperand(0),
4521 DAG.getConstant(c1 + c2, ShiftCountVT)));
4525 // fold (srl (shl x, c), c) -> (and x, cst2)
4526 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1) {
4527 unsigned BitSize = N0.getScalarValueSizeInBits();
4528 if (BitSize <= 64) {
4529 uint64_t ShAmt = N1C->getZExtValue() + 64 - BitSize;
4530 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4531 DAG.getConstant(~0ULL >> ShAmt, VT));
4535 // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask)
4536 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
4537 // Shifting in all undef bits?
4538 EVT SmallVT = N0.getOperand(0).getValueType();
4539 unsigned BitSize = SmallVT.getScalarSizeInBits();
4540 if (N1C->getZExtValue() >= BitSize)
4541 return DAG.getUNDEF(VT);
4543 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
4544 uint64_t ShiftAmt = N1C->getZExtValue();
4545 SDValue SmallShift = DAG.getNode(ISD::SRL, SDLoc(N0), SmallVT,
4547 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
4548 AddToWorklist(SmallShift.getNode());
4549 APInt Mask = APInt::getAllOnesValue(OpSizeInBits).lshr(ShiftAmt);
4550 return DAG.getNode(ISD::AND, SDLoc(N), VT,
4551 DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, SmallShift),
4552 DAG.getConstant(Mask, VT));
4556 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
4557 // bit, which is unmodified by sra.
4558 if (N1C && N1C->getZExtValue() + 1 == OpSizeInBits) {
4559 if (N0.getOpcode() == ISD::SRA)
4560 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1);
4563 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
4564 if (N1C && N0.getOpcode() == ISD::CTLZ &&
4565 N1C->getAPIntValue() == Log2_32(OpSizeInBits)) {
4566 APInt KnownZero, KnownOne;
4567 DAG.computeKnownBits(N0.getOperand(0), KnownZero, KnownOne);
4569 // If any of the input bits are KnownOne, then the input couldn't be all
4570 // zeros, thus the result of the srl will always be zero.
4571 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
4573 // If all of the bits input the to ctlz node are known to be zero, then
4574 // the result of the ctlz is "32" and the result of the shift is one.
4575 APInt UnknownBits = ~KnownZero;
4576 if (UnknownBits == 0) return DAG.getConstant(1, VT);
4578 // Otherwise, check to see if there is exactly one bit input to the ctlz.
4579 if ((UnknownBits & (UnknownBits - 1)) == 0) {
4580 // Okay, we know that only that the single bit specified by UnknownBits
4581 // could be set on input to the CTLZ node. If this bit is set, the SRL
4582 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
4583 // to an SRL/XOR pair, which is likely to simplify more.
4584 unsigned ShAmt = UnknownBits.countTrailingZeros();
4585 SDValue Op = N0.getOperand(0);
4588 Op = DAG.getNode(ISD::SRL, SDLoc(N0), VT, Op,
4589 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
4590 AddToWorklist(Op.getNode());
4593 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
4594 Op, DAG.getConstant(1, VT));
4598 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
4599 if (N1.getOpcode() == ISD::TRUNCATE &&
4600 N1.getOperand(0).getOpcode() == ISD::AND) {
4601 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4602 if (NewOp1.getNode())
4603 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, NewOp1);
4606 // fold operands of srl based on knowledge that the low bits are not
4608 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4609 return SDValue(N, 0);
4612 SDValue NewSRL = visitShiftByConstant(N, N1C);
4613 if (NewSRL.getNode())
4617 // Attempt to convert a srl of a load into a narrower zero-extending load.
4618 SDValue NarrowLoad = ReduceLoadWidth(N);
4619 if (NarrowLoad.getNode())
4622 // Here is a common situation. We want to optimize:
4625 // %b = and i32 %a, 2
4626 // %c = srl i32 %b, 1
4627 // brcond i32 %c ...
4633 // %c = setcc eq %b, 0
4636 // However when after the source operand of SRL is optimized into AND, the SRL
4637 // itself may not be optimized further. Look for it and add the BRCOND into
4639 if (N->hasOneUse()) {
4640 SDNode *Use = *N->use_begin();
4641 if (Use->getOpcode() == ISD::BRCOND)
4643 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
4644 // Also look pass the truncate.
4645 Use = *Use->use_begin();
4646 if (Use->getOpcode() == ISD::BRCOND)
4654 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
4655 SDValue N0 = N->getOperand(0);
4656 EVT VT = N->getValueType(0);
4658 // fold (ctlz c1) -> c2
4659 if (isa<ConstantSDNode>(N0))
4660 return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0);
4664 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4665 SDValue N0 = N->getOperand(0);
4666 EVT VT = N->getValueType(0);
4668 // fold (ctlz_zero_undef c1) -> c2
4669 if (isa<ConstantSDNode>(N0))
4670 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4674 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4675 SDValue N0 = N->getOperand(0);
4676 EVT VT = N->getValueType(0);
4678 // fold (cttz c1) -> c2
4679 if (isa<ConstantSDNode>(N0))
4680 return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0);
4684 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4685 SDValue N0 = N->getOperand(0);
4686 EVT VT = N->getValueType(0);
4688 // fold (cttz_zero_undef c1) -> c2
4689 if (isa<ConstantSDNode>(N0))
4690 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4694 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4695 SDValue N0 = N->getOperand(0);
4696 EVT VT = N->getValueType(0);
4698 // fold (ctpop c1) -> c2
4699 if (isa<ConstantSDNode>(N0))
4700 return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0);
4705 /// \brief Generate Min/Max node
4706 static SDValue combineMinNumMaxNum(SDLoc DL, EVT VT, SDValue LHS, SDValue RHS,
4707 SDValue True, SDValue False,
4708 ISD::CondCode CC, const TargetLowering &TLI,
4709 SelectionDAG &DAG) {
4710 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
4720 unsigned Opcode = (LHS == True) ? ISD::FMINNUM : ISD::FMAXNUM;
4721 if (TLI.isOperationLegal(Opcode, VT))
4722 return DAG.getNode(Opcode, DL, VT, LHS, RHS);
4731 unsigned Opcode = (LHS == True) ? ISD::FMAXNUM : ISD::FMINNUM;
4732 if (TLI.isOperationLegal(Opcode, VT))
4733 return DAG.getNode(Opcode, DL, VT, LHS, RHS);
4741 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4742 SDValue N0 = N->getOperand(0);
4743 SDValue N1 = N->getOperand(1);
4744 SDValue N2 = N->getOperand(2);
4745 EVT VT = N->getValueType(0);
4746 EVT VT0 = N0.getValueType();
4748 // fold (select C, X, X) -> X
4751 // fold (select true, X, Y) -> X
4752 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4753 if (N0C && !N0C->isNullValue())
4755 // fold (select false, X, Y) -> Y
4756 if (N0C && N0C->isNullValue())
4758 // fold (select C, 1, X) -> (or C, X)
4759 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4760 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4761 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4762 // fold (select C, 0, 1) -> (xor C, 1)
4763 // We can't do this reliably if integer based booleans have different contents
4764 // to floating point based booleans. This is because we can't tell whether we
4765 // have an integer-based boolean or a floating-point-based boolean unless we
4766 // can find the SETCC that produced it and inspect its operands. This is
4767 // fairly easy if C is the SETCC node, but it can potentially be
4768 // undiscoverable (or not reasonably discoverable). For example, it could be
4769 // in another basic block or it could require searching a complicated
4771 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
4772 if (VT.isInteger() &&
4773 (VT0 == MVT::i1 || (VT0.isInteger() &&
4774 TLI.getBooleanContents(false, false) ==
4775 TLI.getBooleanContents(false, true) &&
4776 TLI.getBooleanContents(false, false) ==
4777 TargetLowering::ZeroOrOneBooleanContent)) &&
4778 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4781 return DAG.getNode(ISD::XOR, SDLoc(N), VT0,
4782 N0, DAG.getConstant(1, VT0));
4783 XORNode = DAG.getNode(ISD::XOR, SDLoc(N0), VT0,
4784 N0, DAG.getConstant(1, VT0));
4785 AddToWorklist(XORNode.getNode());
4787 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, XORNode);
4788 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, XORNode);
4790 // fold (select C, 0, X) -> (and (not C), X)
4791 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4792 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4793 AddToWorklist(NOTNode.getNode());
4794 return DAG.getNode(ISD::AND, SDLoc(N), VT, NOTNode, N2);
4796 // fold (select C, X, 1) -> (or (not C), X)
4797 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4798 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4799 AddToWorklist(NOTNode.getNode());
4800 return DAG.getNode(ISD::OR, SDLoc(N), VT, NOTNode, N1);
4802 // fold (select C, X, 0) -> (and C, X)
4803 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4804 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4805 // fold (select X, X, Y) -> (or X, Y)
4806 // fold (select X, 1, Y) -> (or X, Y)
4807 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4808 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4809 // fold (select X, Y, X) -> (and X, Y)
4810 // fold (select X, Y, 0) -> (and X, Y)
4811 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4812 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4814 // If we can fold this based on the true/false value, do so.
4815 if (SimplifySelectOps(N, N1, N2))
4816 return SDValue(N, 0); // Don't revisit N.
4818 // fold selects based on a setcc into other things, such as min/max/abs
4819 if (N0.getOpcode() == ISD::SETCC) {
4820 // select x, y (fcmp lt x, y) -> fminnum x, y
4821 // select x, y (fcmp gt x, y) -> fmaxnum x, y
4823 // This is OK if we don't care about what happens if either operand is a
4827 // FIXME: Instead of testing for UnsafeFPMath, this should be checking for
4828 // no signed zeros as well as no nans.
4829 const TargetOptions &Options = DAG.getTarget().Options;
4830 if (Options.UnsafeFPMath &&
4831 VT.isFloatingPoint() && N0.hasOneUse() &&
4832 DAG.isKnownNeverNaN(N1) && DAG.isKnownNeverNaN(N2)) {
4833 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
4836 combineMinNumMaxNum(SDLoc(N), VT, N0.getOperand(0), N0.getOperand(1),
4837 N1, N2, CC, TLI, DAG);
4842 if ((!LegalOperations &&
4843 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT)) ||
4844 TLI.isOperationLegal(ISD::SELECT_CC, VT))
4845 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT,
4846 N0.getOperand(0), N0.getOperand(1),
4847 N1, N2, N0.getOperand(2));
4848 return SimplifySelect(SDLoc(N), N0, N1, N2);
4851 if (VT0 == MVT::i1) {
4852 if (TLI.shouldNormalizeToSelectSequence(*DAG.getContext(), VT)) {
4853 // select (and Cond0, Cond1), X, Y
4854 // -> select Cond0, (select Cond1, X, Y), Y
4855 if (N0->getOpcode() == ISD::AND && N0->hasOneUse()) {
4856 SDValue Cond0 = N0->getOperand(0);
4857 SDValue Cond1 = N0->getOperand(1);
4858 SDValue InnerSelect = DAG.getNode(ISD::SELECT, SDLoc(N),
4859 N1.getValueType(), Cond1, N1, N2);
4860 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Cond0,
4863 // select (or Cond0, Cond1), X, Y -> select Cond0, X, (select Cond1, X, Y)
4864 if (N0->getOpcode() == ISD::OR && N0->hasOneUse()) {
4865 SDValue Cond0 = N0->getOperand(0);
4866 SDValue Cond1 = N0->getOperand(1);
4867 SDValue InnerSelect = DAG.getNode(ISD::SELECT, SDLoc(N),
4868 N1.getValueType(), Cond1, N1, N2);
4869 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Cond0, N1,
4874 // select Cond0, (select Cond1, X, Y), Y -> select (and Cond0, Cond1), X, Y
4875 if (N1->getOpcode() == ISD::SELECT) {
4876 SDValue N1_0 = N1->getOperand(0);
4877 SDValue N1_1 = N1->getOperand(1);
4878 SDValue N1_2 = N1->getOperand(2);
4880 // Create the actual and node if we can generate good code for it.
4881 if (!TLI.shouldNormalizeToSelectSequence(*DAG.getContext(), VT)) {
4882 SDValue And = DAG.getNode(ISD::AND, SDLoc(N), N0.getValueType(),
4884 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), And,
4887 // Otherwise see if we can optimize the "and" to a better pattern.
4888 if (SDValue Combined = visitANDLike(N0, N1_0, N))
4889 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Combined,
4893 // select Cond0, X, (select Cond1, X, Y) -> select (or Cond0, Cond1), X, Y
4894 if (N2->getOpcode() == ISD::SELECT) {
4895 SDValue N2_0 = N2->getOperand(0);
4896 SDValue N2_1 = N2->getOperand(1);
4897 SDValue N2_2 = N2->getOperand(2);
4899 // Create the actual or node if we can generate good code for it.
4900 if (!TLI.shouldNormalizeToSelectSequence(*DAG.getContext(), VT)) {
4901 SDValue Or = DAG.getNode(ISD::OR, SDLoc(N), N0.getValueType(),
4903 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Or,
4906 // Otherwise see if we can optimize to a better pattern.
4907 if (SDValue Combined = visitORLike(N0, N2_0, N))
4908 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Combined,
4918 std::pair<SDValue, SDValue> SplitVSETCC(const SDNode *N, SelectionDAG &DAG) {
4921 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
4923 // Split the inputs.
4924 SDValue Lo, Hi, LL, LH, RL, RH;
4925 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
4926 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
4928 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));
4929 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));
4931 return std::make_pair(Lo, Hi);
4934 // This function assumes all the vselect's arguments are CONCAT_VECTOR
4935 // nodes and that the condition is a BV of ConstantSDNodes (or undefs).
4936 static SDValue ConvertSelectToConcatVector(SDNode *N, SelectionDAG &DAG) {
4938 SDValue Cond = N->getOperand(0);
4939 SDValue LHS = N->getOperand(1);
4940 SDValue RHS = N->getOperand(2);
4941 EVT VT = N->getValueType(0);
4942 int NumElems = VT.getVectorNumElements();
4943 assert(LHS.getOpcode() == ISD::CONCAT_VECTORS &&
4944 RHS.getOpcode() == ISD::CONCAT_VECTORS &&
4945 Cond.getOpcode() == ISD::BUILD_VECTOR);
4947 // CONCAT_VECTOR can take an arbitrary number of arguments. We only care about
4948 // binary ones here.
4949 if (LHS->getNumOperands() != 2 || RHS->getNumOperands() != 2)
4952 // We're sure we have an even number of elements due to the
4953 // concat_vectors we have as arguments to vselect.
4954 // Skip BV elements until we find one that's not an UNDEF
4955 // After we find an UNDEF element, keep looping until we get to half the
4956 // length of the BV and see if all the non-undef nodes are the same.
4957 ConstantSDNode *BottomHalf = nullptr;
4958 for (int i = 0; i < NumElems / 2; ++i) {
4959 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4962 if (BottomHalf == nullptr)
4963 BottomHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4964 else if (Cond->getOperand(i).getNode() != BottomHalf)
4968 // Do the same for the second half of the BuildVector
4969 ConstantSDNode *TopHalf = nullptr;
4970 for (int i = NumElems / 2; i < NumElems; ++i) {
4971 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4974 if (TopHalf == nullptr)
4975 TopHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4976 else if (Cond->getOperand(i).getNode() != TopHalf)
4980 assert(TopHalf && BottomHalf &&
4981 "One half of the selector was all UNDEFs and the other was all the "
4982 "same value. This should have been addressed before this function.");
4984 ISD::CONCAT_VECTORS, dl, VT,
4985 BottomHalf->isNullValue() ? RHS->getOperand(0) : LHS->getOperand(0),
4986 TopHalf->isNullValue() ? RHS->getOperand(1) : LHS->getOperand(1));
4989 SDValue DAGCombiner::visitMSTORE(SDNode *N) {
4991 if (Level >= AfterLegalizeTypes)
4994 MaskedStoreSDNode *MST = dyn_cast<MaskedStoreSDNode>(N);
4995 SDValue Mask = MST->getMask();
4996 SDValue Data = MST->getValue();
4999 // If the MSTORE data type requires splitting and the mask is provided by a
5000 // SETCC, then split both nodes and its operands before legalization. This
5001 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5002 // and enables future optimizations (e.g. min/max pattern matching on X86).
5003 if (Mask.getOpcode() == ISD::SETCC) {
5005 // Check if any splitting is required.
5006 if (TLI.getTypeAction(*DAG.getContext(), Data.getValueType()) !=
5007 TargetLowering::TypeSplitVector)
5010 SDValue MaskLo, MaskHi, Lo, Hi;
5011 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
5014 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MST->getValueType(0));
5016 SDValue Chain = MST->getChain();
5017 SDValue Ptr = MST->getBasePtr();
5019 EVT MemoryVT = MST->getMemoryVT();
5020 unsigned Alignment = MST->getOriginalAlignment();
5022 // if Alignment is equal to the vector size,
5023 // take the half of it for the second part
5024 unsigned SecondHalfAlignment =
5025 (Alignment == Data->getValueType(0).getSizeInBits()/8) ?
5026 Alignment/2 : Alignment;
5028 EVT LoMemVT, HiMemVT;
5029 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
5031 SDValue DataLo, DataHi;
5032 std::tie(DataLo, DataHi) = DAG.SplitVector(Data, DL);
5034 MachineMemOperand *MMO = DAG.getMachineFunction().
5035 getMachineMemOperand(MST->getPointerInfo(),
5036 MachineMemOperand::MOStore, LoMemVT.getStoreSize(),
5037 Alignment, MST->getAAInfo(), MST->getRanges());
5039 Lo = DAG.getMaskedStore(Chain, DL, DataLo, Ptr, MaskLo, LoMemVT, MMO,
5040 MST->isTruncatingStore());
5042 unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
5043 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5044 DAG.getConstant(IncrementSize, Ptr.getValueType()));
5046 MMO = DAG.getMachineFunction().
5047 getMachineMemOperand(MST->getPointerInfo(),
5048 MachineMemOperand::MOStore, HiMemVT.getStoreSize(),
5049 SecondHalfAlignment, MST->getAAInfo(),
5052 Hi = DAG.getMaskedStore(Chain, DL, DataHi, Ptr, MaskHi, HiMemVT, MMO,
5053 MST->isTruncatingStore());
5055 AddToWorklist(Lo.getNode());
5056 AddToWorklist(Hi.getNode());
5058 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);
5063 SDValue DAGCombiner::visitMLOAD(SDNode *N) {
5065 if (Level >= AfterLegalizeTypes)
5068 MaskedLoadSDNode *MLD = dyn_cast<MaskedLoadSDNode>(N);
5069 SDValue Mask = MLD->getMask();
5072 // If the MLOAD result requires splitting and the mask is provided by a
5073 // SETCC, then split both nodes and its operands before legalization. This
5074 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5075 // and enables future optimizations (e.g. min/max pattern matching on X86).
5077 if (Mask.getOpcode() == ISD::SETCC) {
5078 EVT VT = N->getValueType(0);
5080 // Check if any splitting is required.
5081 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
5082 TargetLowering::TypeSplitVector)
5085 SDValue MaskLo, MaskHi, Lo, Hi;
5086 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
5088 SDValue Src0 = MLD->getSrc0();
5089 SDValue Src0Lo, Src0Hi;
5090 std::tie(Src0Lo, Src0Hi) = DAG.SplitVector(Src0, DL);
5093 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MLD->getValueType(0));
5095 SDValue Chain = MLD->getChain();
5096 SDValue Ptr = MLD->getBasePtr();
5097 EVT MemoryVT = MLD->getMemoryVT();
5098 unsigned Alignment = MLD->getOriginalAlignment();
5100 // if Alignment is equal to the vector size,
5101 // take the half of it for the second part
5102 unsigned SecondHalfAlignment =
5103 (Alignment == MLD->getValueType(0).getSizeInBits()/8) ?
5104 Alignment/2 : Alignment;
5106 EVT LoMemVT, HiMemVT;
5107 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
5109 MachineMemOperand *MMO = DAG.getMachineFunction().
5110 getMachineMemOperand(MLD->getPointerInfo(),
5111 MachineMemOperand::MOLoad, LoMemVT.getStoreSize(),
5112 Alignment, MLD->getAAInfo(), MLD->getRanges());
5114 Lo = DAG.getMaskedLoad(LoVT, DL, Chain, Ptr, MaskLo, Src0Lo, LoMemVT, MMO,
5117 unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
5118 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5119 DAG.getConstant(IncrementSize, Ptr.getValueType()));
5121 MMO = DAG.getMachineFunction().
5122 getMachineMemOperand(MLD->getPointerInfo(),
5123 MachineMemOperand::MOLoad, HiMemVT.getStoreSize(),
5124 SecondHalfAlignment, MLD->getAAInfo(), MLD->getRanges());
5126 Hi = DAG.getMaskedLoad(HiVT, DL, Chain, Ptr, MaskHi, Src0Hi, HiMemVT, MMO,
5129 AddToWorklist(Lo.getNode());
5130 AddToWorklist(Hi.getNode());
5132 // Build a factor node to remember that this load is independent of the
5134 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo.getValue(1),
5137 // Legalized the chain result - switch anything that used the old chain to
5139 DAG.ReplaceAllUsesOfValueWith(SDValue(MLD, 1), Chain);
5141 SDValue LoadRes = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
5143 SDValue RetOps[] = { LoadRes, Chain };
5144 return DAG.getMergeValues(RetOps, DL);
5149 SDValue DAGCombiner::visitVSELECT(SDNode *N) {
5150 SDValue N0 = N->getOperand(0);
5151 SDValue N1 = N->getOperand(1);
5152 SDValue N2 = N->getOperand(2);
5155 // Canonicalize integer abs.
5156 // vselect (setg[te] X, 0), X, -X ->
5157 // vselect (setgt X, -1), X, -X ->
5158 // vselect (setl[te] X, 0), -X, X ->
5159 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
5160 if (N0.getOpcode() == ISD::SETCC) {
5161 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
5162 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5164 bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
5166 if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
5167 (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) &&
5168 N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1))
5169 isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode());
5170 else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) &&
5171 N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1))
5172 isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode());
5175 EVT VT = LHS.getValueType();
5176 SDValue Shift = DAG.getNode(
5177 ISD::SRA, DL, VT, LHS,
5178 DAG.getConstant(VT.getScalarType().getSizeInBits() - 1, VT));
5179 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift);
5180 AddToWorklist(Shift.getNode());
5181 AddToWorklist(Add.getNode());
5182 return DAG.getNode(ISD::XOR, DL, VT, Add, Shift);
5186 // If the VSELECT result requires splitting and the mask is provided by a
5187 // SETCC, then split both nodes and its operands before legalization. This
5188 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5189 // and enables future optimizations (e.g. min/max pattern matching on X86).
5190 if (N0.getOpcode() == ISD::SETCC) {
5191 EVT VT = N->getValueType(0);
5193 // Check if any splitting is required.
5194 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
5195 TargetLowering::TypeSplitVector)
5198 SDValue Lo, Hi, CCLo, CCHi, LL, LH, RL, RH;
5199 std::tie(CCLo, CCHi) = SplitVSETCC(N0.getNode(), DAG);
5200 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 1);
5201 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 2);
5203 Lo = DAG.getNode(N->getOpcode(), DL, LL.getValueType(), CCLo, LL, RL);
5204 Hi = DAG.getNode(N->getOpcode(), DL, LH.getValueType(), CCHi, LH, RH);
5206 // Add the new VSELECT nodes to the work list in case they need to be split
5208 AddToWorklist(Lo.getNode());
5209 AddToWorklist(Hi.getNode());
5211 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
5214 // Fold (vselect (build_vector all_ones), N1, N2) -> N1
5215 if (ISD::isBuildVectorAllOnes(N0.getNode()))
5217 // Fold (vselect (build_vector all_zeros), N1, N2) -> N2
5218 if (ISD::isBuildVectorAllZeros(N0.getNode()))
5221 // The ConvertSelectToConcatVector function is assuming both the above
5222 // checks for (vselect (build_vector all{ones,zeros) ...) have been made
5224 if (N1.getOpcode() == ISD::CONCAT_VECTORS &&
5225 N2.getOpcode() == ISD::CONCAT_VECTORS &&
5226 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
5227 SDValue CV = ConvertSelectToConcatVector(N, DAG);
5235 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
5236 SDValue N0 = N->getOperand(0);
5237 SDValue N1 = N->getOperand(1);
5238 SDValue N2 = N->getOperand(2);
5239 SDValue N3 = N->getOperand(3);
5240 SDValue N4 = N->getOperand(4);
5241 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
5243 // fold select_cc lhs, rhs, x, x, cc -> x
5247 // Determine if the condition we're dealing with is constant
5248 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
5249 N0, N1, CC, SDLoc(N), false);
5250 if (SCC.getNode()) {
5251 AddToWorklist(SCC.getNode());
5253 if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) {
5254 if (!SCCC->isNullValue())
5255 return N2; // cond always true -> true val
5257 return N3; // cond always false -> false val
5258 } else if (SCC->getOpcode() == ISD::UNDEF) {
5259 // When the condition is UNDEF, just return the first operand. This is
5260 // coherent the DAG creation, no setcc node is created in this case
5262 } else if (SCC.getOpcode() == ISD::SETCC) {
5263 // Fold to a simpler select_cc
5264 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(),
5265 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
5270 // If we can fold this based on the true/false value, do so.
5271 if (SimplifySelectOps(N, N2, N3))
5272 return SDValue(N, 0); // Don't revisit N.
5274 // fold select_cc into other things, such as min/max/abs
5275 return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC);
5278 SDValue DAGCombiner::visitSETCC(SDNode *N) {
5279 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
5280 cast<CondCodeSDNode>(N->getOperand(2))->get(),
5284 // tryToFoldExtendOfConstant - Try to fold a sext/zext/aext
5285 // dag node into a ConstantSDNode or a build_vector of constants.
5286 // This function is called by the DAGCombiner when visiting sext/zext/aext
5287 // dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND).
5288 // Vector extends are not folded if operations are legal; this is to
5289 // avoid introducing illegal build_vector dag nodes.
5290 static SDNode *tryToFoldExtendOfConstant(SDNode *N, const TargetLowering &TLI,
5291 SelectionDAG &DAG, bool LegalTypes,
5292 bool LegalOperations) {
5293 unsigned Opcode = N->getOpcode();
5294 SDValue N0 = N->getOperand(0);
5295 EVT VT = N->getValueType(0);
5297 assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND ||
5298 Opcode == ISD::ANY_EXTEND) && "Expected EXTEND dag node in input!");
5300 // fold (sext c1) -> c1
5301 // fold (zext c1) -> c1
5302 // fold (aext c1) -> c1
5303 if (isa<ConstantSDNode>(N0))
5304 return DAG.getNode(Opcode, SDLoc(N), VT, N0).getNode();
5306 // fold (sext (build_vector AllConstants) -> (build_vector AllConstants)
5307 // fold (zext (build_vector AllConstants) -> (build_vector AllConstants)
5308 // fold (aext (build_vector AllConstants) -> (build_vector AllConstants)
5309 EVT SVT = VT.getScalarType();
5310 if (!(VT.isVector() &&
5311 (!LegalTypes || (!LegalOperations && TLI.isTypeLegal(SVT))) &&
5312 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())))
5315 // We can fold this node into a build_vector.
5316 unsigned VTBits = SVT.getSizeInBits();
5317 unsigned EVTBits = N0->getValueType(0).getScalarType().getSizeInBits();
5318 unsigned ShAmt = VTBits - EVTBits;
5319 SmallVector<SDValue, 8> Elts;
5320 unsigned NumElts = N0->getNumOperands();
5323 for (unsigned i=0; i != NumElts; ++i) {
5324 SDValue Op = N0->getOperand(i);
5325 if (Op->getOpcode() == ISD::UNDEF) {
5326 Elts.push_back(DAG.getUNDEF(SVT));
5330 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
5331 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
5332 if (Opcode == ISD::SIGN_EXTEND)
5333 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
5336 Elts.push_back(DAG.getConstant(C.shl(ShAmt).lshr(ShAmt).getZExtValue(),
5340 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Elts).getNode();
5343 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
5344 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
5345 // transformation. Returns true if extension are possible and the above
5346 // mentioned transformation is profitable.
5347 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
5349 SmallVectorImpl<SDNode *> &ExtendNodes,
5350 const TargetLowering &TLI) {
5351 bool HasCopyToRegUses = false;
5352 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
5353 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
5354 UE = N0.getNode()->use_end();
5359 if (UI.getUse().getResNo() != N0.getResNo())
5361 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
5362 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
5363 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
5364 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
5365 // Sign bits will be lost after a zext.
5368 for (unsigned i = 0; i != 2; ++i) {
5369 SDValue UseOp = User->getOperand(i);
5372 if (!isa<ConstantSDNode>(UseOp))
5377 ExtendNodes.push_back(User);
5380 // If truncates aren't free and there are users we can't
5381 // extend, it isn't worthwhile.
5384 // Remember if this value is live-out.
5385 if (User->getOpcode() == ISD::CopyToReg)
5386 HasCopyToRegUses = true;
5389 if (HasCopyToRegUses) {
5390 bool BothLiveOut = false;
5391 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
5393 SDUse &Use = UI.getUse();
5394 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
5400 // Both unextended and extended values are live out. There had better be
5401 // a good reason for the transformation.
5402 return ExtendNodes.size();
5407 void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
5408 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
5409 ISD::NodeType ExtType) {
5410 // Extend SetCC uses if necessary.
5411 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
5412 SDNode *SetCC = SetCCs[i];
5413 SmallVector<SDValue, 4> Ops;
5415 for (unsigned j = 0; j != 2; ++j) {
5416 SDValue SOp = SetCC->getOperand(j);
5418 Ops.push_back(ExtLoad);
5420 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
5423 Ops.push_back(SetCC->getOperand(2));
5424 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), Ops));
5428 // FIXME: Bring more similar combines here, common to sext/zext (maybe aext?).
5429 SDValue DAGCombiner::CombineExtLoad(SDNode *N) {
5430 SDValue N0 = N->getOperand(0);
5431 EVT DstVT = N->getValueType(0);
5432 EVT SrcVT = N0.getValueType();
5434 assert((N->getOpcode() == ISD::SIGN_EXTEND ||
5435 N->getOpcode() == ISD::ZERO_EXTEND) &&
5436 "Unexpected node type (not an extend)!");
5438 // fold (sext (load x)) to multiple smaller sextloads; same for zext.
5439 // For example, on a target with legal v4i32, but illegal v8i32, turn:
5440 // (v8i32 (sext (v8i16 (load x))))
5442 // (v8i32 (concat_vectors (v4i32 (sextload x)),
5443 // (v4i32 (sextload (x + 16)))))
5444 // Where uses of the original load, i.e.:
5446 // are replaced with:
5448 // (v8i32 (concat_vectors (v4i32 (sextload x)),
5449 // (v4i32 (sextload (x + 16)))))))
5451 // This combine is only applicable to illegal, but splittable, vectors.
5452 // All legal types, and illegal non-vector types, are handled elsewhere.
5453 // This combine is controlled by TargetLowering::isVectorLoadExtDesirable.
5455 if (N0->getOpcode() != ISD::LOAD)
5458 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5460 if (!ISD::isNON_EXTLoad(LN0) || !ISD::isUNINDEXEDLoad(LN0) ||
5461 !N0.hasOneUse() || LN0->isVolatile() || !DstVT.isVector() ||
5462 !DstVT.isPow2VectorType() || !TLI.isVectorLoadExtDesirable(SDValue(N, 0)))
5465 SmallVector<SDNode *, 4> SetCCs;
5466 if (!ExtendUsesToFormExtLoad(N, N0, N->getOpcode(), SetCCs, TLI))
5469 ISD::LoadExtType ExtType =
5470 N->getOpcode() == ISD::SIGN_EXTEND ? ISD::SEXTLOAD : ISD::ZEXTLOAD;
5472 // Try to split the vector types to get down to legal types.
5473 EVT SplitSrcVT = SrcVT;
5474 EVT SplitDstVT = DstVT;
5475 while (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT) &&
5476 SplitSrcVT.getVectorNumElements() > 1) {
5477 SplitDstVT = DAG.GetSplitDestVTs(SplitDstVT).first;
5478 SplitSrcVT = DAG.GetSplitDestVTs(SplitSrcVT).first;
5481 if (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT))
5485 const unsigned NumSplits =
5486 DstVT.getVectorNumElements() / SplitDstVT.getVectorNumElements();
5487 const unsigned Stride = SplitSrcVT.getStoreSize();
5488 SmallVector<SDValue, 4> Loads;
5489 SmallVector<SDValue, 4> Chains;
5491 SDValue BasePtr = LN0->getBasePtr();
5492 for (unsigned Idx = 0; Idx < NumSplits; Idx++) {
5493 const unsigned Offset = Idx * Stride;
5494 const unsigned Align = MinAlign(LN0->getAlignment(), Offset);
5496 SDValue SplitLoad = DAG.getExtLoad(
5497 ExtType, DL, SplitDstVT, LN0->getChain(), BasePtr,
5498 LN0->getPointerInfo().getWithOffset(Offset), SplitSrcVT,
5499 LN0->isVolatile(), LN0->isNonTemporal(), LN0->isInvariant(),
5500 Align, LN0->getAAInfo());
5502 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
5503 DAG.getConstant(Stride, BasePtr.getValueType()));
5505 Loads.push_back(SplitLoad.getValue(0));
5506 Chains.push_back(SplitLoad.getValue(1));
5509 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
5510 SDValue NewValue = DAG.getNode(ISD::CONCAT_VECTORS, DL, DstVT, Loads);
5512 CombineTo(N, NewValue);
5514 // Replace uses of the original load (before extension)
5515 // with a truncate of the concatenated sextloaded vectors.
5517 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(), NewValue);
5518 CombineTo(N0.getNode(), Trunc, NewChain);
5519 ExtendSetCCUses(SetCCs, Trunc, NewValue, DL,
5520 (ISD::NodeType)N->getOpcode());
5521 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5524 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
5525 SDValue N0 = N->getOperand(0);
5526 EVT VT = N->getValueType(0);
5528 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5530 return SDValue(Res, 0);
5532 // fold (sext (sext x)) -> (sext x)
5533 // fold (sext (aext x)) -> (sext x)
5534 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5535 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT,
5538 if (N0.getOpcode() == ISD::TRUNCATE) {
5539 // fold (sext (truncate (load x))) -> (sext (smaller load x))
5540 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
5541 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5542 if (NarrowLoad.getNode()) {
5543 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5544 if (NarrowLoad.getNode() != N0.getNode()) {
5545 CombineTo(N0.getNode(), NarrowLoad);
5546 // CombineTo deleted the truncate, if needed, but not what's under it.
5549 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5552 // See if the value being truncated is already sign extended. If so, just
5553 // eliminate the trunc/sext pair.
5554 SDValue Op = N0.getOperand(0);
5555 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
5556 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
5557 unsigned DestBits = VT.getScalarType().getSizeInBits();
5558 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
5560 if (OpBits == DestBits) {
5561 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
5562 // bits, it is already ready.
5563 if (NumSignBits > DestBits-MidBits)
5565 } else if (OpBits < DestBits) {
5566 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
5567 // bits, just sext from i32.
5568 if (NumSignBits > OpBits-MidBits)
5569 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, Op);
5571 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
5572 // bits, just truncate to i32.
5573 if (NumSignBits > OpBits-MidBits)
5574 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5577 // fold (sext (truncate x)) -> (sextinreg x).
5578 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
5579 N0.getValueType())) {
5580 if (OpBits < DestBits)
5581 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op);
5582 else if (OpBits > DestBits)
5583 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op);
5584 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, Op,
5585 DAG.getValueType(N0.getValueType()));
5589 // fold (sext (load x)) -> (sext (truncate (sextload x)))
5590 // Only generate vector extloads when 1) they're legal, and 2) they are
5591 // deemed desirable by the target.
5592 if (ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5593 ((!LegalOperations && !VT.isVector() &&
5594 !cast<LoadSDNode>(N0)->isVolatile()) ||
5595 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, N0.getValueType()))) {
5596 bool DoXform = true;
5597 SmallVector<SDNode*, 4> SetCCs;
5598 if (!N0.hasOneUse())
5599 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
5601 DoXform &= TLI.isVectorLoadExtDesirable(SDValue(N, 0));
5603 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5604 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5606 LN0->getBasePtr(), N0.getValueType(),
5607 LN0->getMemOperand());
5608 CombineTo(N, ExtLoad);
5609 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5610 N0.getValueType(), ExtLoad);
5611 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5612 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5614 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5618 // fold (sext (load x)) to multiple smaller sextloads.
5619 // Only on illegal but splittable vectors.
5620 if (SDValue ExtLoad = CombineExtLoad(N))
5623 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
5624 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
5625 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5626 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5627 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5628 EVT MemVT = LN0->getMemoryVT();
5629 if ((!LegalOperations && !LN0->isVolatile()) ||
5630 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, MemVT)) {
5631 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5633 LN0->getBasePtr(), MemVT,
5634 LN0->getMemOperand());
5635 CombineTo(N, ExtLoad);
5636 CombineTo(N0.getNode(),
5637 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5638 N0.getValueType(), ExtLoad),
5639 ExtLoad.getValue(1));
5640 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5644 // fold (sext (and/or/xor (load x), cst)) ->
5645 // (and/or/xor (sextload x), (sext cst))
5646 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5647 N0.getOpcode() == ISD::XOR) &&
5648 isa<LoadSDNode>(N0.getOperand(0)) &&
5649 N0.getOperand(1).getOpcode() == ISD::Constant &&
5650 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, N0.getValueType()) &&
5651 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5652 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5653 if (LN0->getExtensionType() != ISD::ZEXTLOAD && LN0->isUnindexed()) {
5654 bool DoXform = true;
5655 SmallVector<SDNode*, 4> SetCCs;
5656 if (!N0.hasOneUse())
5657 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
5660 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN0), VT,
5661 LN0->getChain(), LN0->getBasePtr(),
5663 LN0->getMemOperand());
5664 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5665 Mask = Mask.sext(VT.getSizeInBits());
5666 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5667 ExtLoad, DAG.getConstant(Mask, VT));
5668 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5669 SDLoc(N0.getOperand(0)),
5670 N0.getOperand(0).getValueType(), ExtLoad);
5672 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5673 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5675 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5680 if (N0.getOpcode() == ISD::SETCC) {
5681 EVT N0VT = N0.getOperand(0).getValueType();
5682 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
5683 // Only do this before legalize for now.
5684 if (VT.isVector() && !LegalOperations &&
5685 TLI.getBooleanContents(N0VT) ==
5686 TargetLowering::ZeroOrNegativeOneBooleanContent) {
5687 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
5688 // of the same size as the compared operands. Only optimize sext(setcc())
5689 // if this is the case.
5690 EVT SVT = getSetCCResultType(N0VT);
5692 // We know that the # elements of the results is the same as the
5693 // # elements of the compare (and the # elements of the compare result
5694 // for that matter). Check to see that they are the same size. If so,
5695 // we know that the element size of the sext'd result matches the
5696 // element size of the compare operands.
5697 if (VT.getSizeInBits() == SVT.getSizeInBits())
5698 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5700 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5702 // If the desired elements are smaller or larger than the source
5703 // elements we can use a matching integer vector type and then
5704 // truncate/sign extend
5705 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5706 if (SVT == MatchingVectorType) {
5707 SDValue VsetCC = DAG.getSetCC(SDLoc(N), MatchingVectorType,
5708 N0.getOperand(0), N0.getOperand(1),
5709 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5710 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
5714 // sext(setcc x, y, cc) -> (select (setcc x, y, cc), -1, 0)
5715 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
5717 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
5719 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5720 NegOne, DAG.getConstant(0, VT),
5721 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5722 if (SCC.getNode()) return SCC;
5724 if (!VT.isVector()) {
5725 EVT SetCCVT = getSetCCResultType(N0.getOperand(0).getValueType());
5726 if (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, SetCCVT)) {
5728 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5729 SDValue SetCC = DAG.getSetCC(DL, SetCCVT,
5730 N0.getOperand(0), N0.getOperand(1), CC);
5731 return DAG.getSelect(DL, VT, SetCC,
5732 NegOne, DAG.getConstant(0, VT));
5737 // fold (sext x) -> (zext x) if the sign bit is known zero.
5738 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
5739 DAG.SignBitIsZero(N0))
5740 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0);
5745 // isTruncateOf - If N is a truncate of some other value, return true, record
5746 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
5747 // This function computes KnownZero to avoid a duplicated call to
5748 // computeKnownBits in the caller.
5749 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
5752 if (N->getOpcode() == ISD::TRUNCATE) {
5753 Op = N->getOperand(0);
5754 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5758 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
5759 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
5762 SDValue Op0 = N->getOperand(0);
5763 SDValue Op1 = N->getOperand(1);
5764 assert(Op0.getValueType() == Op1.getValueType());
5766 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
5767 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
5768 if (COp0 && COp0->isNullValue())
5770 else if (COp1 && COp1->isNullValue())
5775 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5777 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
5783 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
5784 SDValue N0 = N->getOperand(0);
5785 EVT VT = N->getValueType(0);
5787 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5789 return SDValue(Res, 0);
5791 // fold (zext (zext x)) -> (zext x)
5792 // fold (zext (aext x)) -> (zext x)
5793 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5794 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT,
5797 // fold (zext (truncate x)) -> (zext x) or
5798 // (zext (truncate x)) -> (truncate x)
5799 // This is valid when the truncated bits of x are already zero.
5800 // FIXME: We should extend this to work for vectors too.
5803 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
5804 APInt TruncatedBits =
5805 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
5806 APInt(Op.getValueSizeInBits(), 0) :
5807 APInt::getBitsSet(Op.getValueSizeInBits(),
5808 N0.getValueSizeInBits(),
5809 std::min(Op.getValueSizeInBits(),
5810 VT.getSizeInBits()));
5811 if (TruncatedBits == (KnownZero & TruncatedBits)) {
5812 if (VT.bitsGT(Op.getValueType()))
5813 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, Op);
5814 if (VT.bitsLT(Op.getValueType()))
5815 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5821 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5822 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
5823 if (N0.getOpcode() == ISD::TRUNCATE) {
5824 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5825 if (NarrowLoad.getNode()) {
5826 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5827 if (NarrowLoad.getNode() != N0.getNode()) {
5828 CombineTo(N0.getNode(), NarrowLoad);
5829 // CombineTo deleted the truncate, if needed, but not what's under it.
5832 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5836 // fold (zext (truncate x)) -> (and x, mask)
5837 if (N0.getOpcode() == ISD::TRUNCATE &&
5838 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
5840 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5841 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
5842 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5843 if (NarrowLoad.getNode()) {
5844 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5845 if (NarrowLoad.getNode() != N0.getNode()) {
5846 CombineTo(N0.getNode(), NarrowLoad);
5847 // CombineTo deleted the truncate, if needed, but not what's under it.
5850 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5853 SDValue Op = N0.getOperand(0);
5854 if (Op.getValueType().bitsLT(VT)) {
5855 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, Op);
5856 AddToWorklist(Op.getNode());
5857 } else if (Op.getValueType().bitsGT(VT)) {
5858 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5859 AddToWorklist(Op.getNode());
5861 return DAG.getZeroExtendInReg(Op, SDLoc(N),
5862 N0.getValueType().getScalarType());
5865 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
5866 // if either of the casts is not free.
5867 if (N0.getOpcode() == ISD::AND &&
5868 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5869 N0.getOperand(1).getOpcode() == ISD::Constant &&
5870 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5871 N0.getValueType()) ||
5872 !TLI.isZExtFree(N0.getValueType(), VT))) {
5873 SDValue X = N0.getOperand(0).getOperand(0);
5874 if (X.getValueType().bitsLT(VT)) {
5875 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(X), VT, X);
5876 } else if (X.getValueType().bitsGT(VT)) {
5877 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
5879 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5880 Mask = Mask.zext(VT.getSizeInBits());
5881 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5882 X, DAG.getConstant(Mask, VT));
5885 // fold (zext (load x)) -> (zext (truncate (zextload x)))
5886 // Only generate vector extloads when 1) they're legal, and 2) they are
5887 // deemed desirable by the target.
5888 if (ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5889 ((!LegalOperations && !VT.isVector() &&
5890 !cast<LoadSDNode>(N0)->isVolatile()) ||
5891 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, N0.getValueType()))) {
5892 bool DoXform = true;
5893 SmallVector<SDNode*, 4> SetCCs;
5894 if (!N0.hasOneUse())
5895 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
5897 DoXform &= TLI.isVectorLoadExtDesirable(SDValue(N, 0));
5899 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5900 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5902 LN0->getBasePtr(), N0.getValueType(),
5903 LN0->getMemOperand());
5904 CombineTo(N, ExtLoad);
5905 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5906 N0.getValueType(), ExtLoad);
5907 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5909 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5911 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5915 // fold (zext (load x)) to multiple smaller zextloads.
5916 // Only on illegal but splittable vectors.
5917 if (SDValue ExtLoad = CombineExtLoad(N))
5920 // fold (zext (and/or/xor (load x), cst)) ->
5921 // (and/or/xor (zextload x), (zext cst))
5922 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5923 N0.getOpcode() == ISD::XOR) &&
5924 isa<LoadSDNode>(N0.getOperand(0)) &&
5925 N0.getOperand(1).getOpcode() == ISD::Constant &&
5926 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, N0.getValueType()) &&
5927 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5928 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5929 if (LN0->getExtensionType() != ISD::SEXTLOAD && LN0->isUnindexed()) {
5930 bool DoXform = true;
5931 SmallVector<SDNode*, 4> SetCCs;
5932 if (!N0.hasOneUse())
5933 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
5936 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), VT,
5937 LN0->getChain(), LN0->getBasePtr(),
5939 LN0->getMemOperand());
5940 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5941 Mask = Mask.zext(VT.getSizeInBits());
5942 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5943 ExtLoad, DAG.getConstant(Mask, VT));
5944 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5945 SDLoc(N0.getOperand(0)),
5946 N0.getOperand(0).getValueType(), ExtLoad);
5948 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5949 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5951 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5956 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
5957 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
5958 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5959 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5960 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5961 EVT MemVT = LN0->getMemoryVT();
5962 if ((!LegalOperations && !LN0->isVolatile()) ||
5963 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT)) {
5964 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5966 LN0->getBasePtr(), MemVT,
5967 LN0->getMemOperand());
5968 CombineTo(N, ExtLoad);
5969 CombineTo(N0.getNode(),
5970 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(),
5972 ExtLoad.getValue(1));
5973 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5977 if (N0.getOpcode() == ISD::SETCC) {
5978 if (!LegalOperations && VT.isVector() &&
5979 N0.getValueType().getVectorElementType() == MVT::i1) {
5980 EVT N0VT = N0.getOperand(0).getValueType();
5981 if (getSetCCResultType(N0VT) == N0.getValueType())
5984 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
5985 // Only do this before legalize for now.
5986 EVT EltVT = VT.getVectorElementType();
5987 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
5988 DAG.getConstant(1, EltVT));
5989 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5990 // We know that the # elements of the results is the same as the
5991 // # elements of the compare (and the # elements of the compare result
5992 // for that matter). Check to see that they are the same size. If so,
5993 // we know that the element size of the sext'd result matches the
5994 // element size of the compare operands.
5995 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5996 DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5998 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
5999 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT,
6002 // If the desired elements are smaller or larger than the source
6003 // elements we can use a matching integer vector type and then
6004 // truncate/sign extend
6005 EVT MatchingElementType =
6006 EVT::getIntegerVT(*DAG.getContext(),
6007 N0VT.getScalarType().getSizeInBits());
6008 EVT MatchingVectorType =
6009 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
6010 N0VT.getVectorNumElements());
6012 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
6014 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6015 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6016 DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT),
6017 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, OneOps));
6020 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
6022 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
6023 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
6024 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
6025 if (SCC.getNode()) return SCC;
6028 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
6029 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
6030 isa<ConstantSDNode>(N0.getOperand(1)) &&
6031 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
6033 SDValue ShAmt = N0.getOperand(1);
6034 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
6035 if (N0.getOpcode() == ISD::SHL) {
6036 SDValue InnerZExt = N0.getOperand(0);
6037 // If the original shl may be shifting out bits, do not perform this
6039 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
6040 InnerZExt.getOperand(0).getValueType().getSizeInBits();
6041 if (ShAmtVal > KnownZeroBits)
6047 // Ensure that the shift amount is wide enough for the shifted value.
6048 if (VT.getSizeInBits() >= 256)
6049 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
6051 return DAG.getNode(N0.getOpcode(), DL, VT,
6052 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
6059 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
6060 SDValue N0 = N->getOperand(0);
6061 EVT VT = N->getValueType(0);
6063 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
6065 return SDValue(Res, 0);
6067 // fold (aext (aext x)) -> (aext x)
6068 // fold (aext (zext x)) -> (zext x)
6069 // fold (aext (sext x)) -> (sext x)
6070 if (N0.getOpcode() == ISD::ANY_EXTEND ||
6071 N0.getOpcode() == ISD::ZERO_EXTEND ||
6072 N0.getOpcode() == ISD::SIGN_EXTEND)
6073 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0));
6075 // fold (aext (truncate (load x))) -> (aext (smaller load x))
6076 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
6077 if (N0.getOpcode() == ISD::TRUNCATE) {
6078 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
6079 if (NarrowLoad.getNode()) {
6080 SDNode* oye = N0.getNode()->getOperand(0).getNode();
6081 if (NarrowLoad.getNode() != N0.getNode()) {
6082 CombineTo(N0.getNode(), NarrowLoad);
6083 // CombineTo deleted the truncate, if needed, but not what's under it.
6086 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6090 // fold (aext (truncate x))
6091 if (N0.getOpcode() == ISD::TRUNCATE) {
6092 SDValue TruncOp = N0.getOperand(0);
6093 if (TruncOp.getValueType() == VT)
6094 return TruncOp; // x iff x size == zext size.
6095 if (TruncOp.getValueType().bitsGT(VT))
6096 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, TruncOp);
6097 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, TruncOp);
6100 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
6101 // if the trunc is not free.
6102 if (N0.getOpcode() == ISD::AND &&
6103 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
6104 N0.getOperand(1).getOpcode() == ISD::Constant &&
6105 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
6106 N0.getValueType())) {
6107 SDValue X = N0.getOperand(0).getOperand(0);
6108 if (X.getValueType().bitsLT(VT)) {
6109 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, X);
6110 } else if (X.getValueType().bitsGT(VT)) {
6111 X = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, X);
6113 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
6114 Mask = Mask.zext(VT.getSizeInBits());
6115 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6116 X, DAG.getConstant(Mask, VT));
6119 // fold (aext (load x)) -> (aext (truncate (extload x)))
6120 // None of the supported targets knows how to perform load and any_ext
6121 // on vectors in one instruction. We only perform this transformation on
6123 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
6124 ISD::isUNINDEXEDLoad(N0.getNode()) &&
6125 TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) {
6126 bool DoXform = true;
6127 SmallVector<SDNode*, 4> SetCCs;
6128 if (!N0.hasOneUse())
6129 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
6131 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6132 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
6134 LN0->getBasePtr(), N0.getValueType(),
6135 LN0->getMemOperand());
6136 CombineTo(N, ExtLoad);
6137 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
6138 N0.getValueType(), ExtLoad);
6139 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
6140 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
6142 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6146 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
6147 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
6148 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
6149 if (N0.getOpcode() == ISD::LOAD &&
6150 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
6152 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6153 ISD::LoadExtType ExtType = LN0->getExtensionType();
6154 EVT MemVT = LN0->getMemoryVT();
6155 if (!LegalOperations || TLI.isLoadExtLegal(ExtType, VT, MemVT)) {
6156 SDValue ExtLoad = DAG.getExtLoad(ExtType, SDLoc(N),
6157 VT, LN0->getChain(), LN0->getBasePtr(),
6158 MemVT, LN0->getMemOperand());
6159 CombineTo(N, ExtLoad);
6160 CombineTo(N0.getNode(),
6161 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
6162 N0.getValueType(), ExtLoad),
6163 ExtLoad.getValue(1));
6164 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6168 if (N0.getOpcode() == ISD::SETCC) {
6170 // aext(setcc) -> vsetcc
6171 // aext(setcc) -> truncate(vsetcc)
6172 // aext(setcc) -> aext(vsetcc)
6173 // Only do this before legalize for now.
6174 if (VT.isVector() && !LegalOperations) {
6175 EVT N0VT = N0.getOperand(0).getValueType();
6176 // We know that the # elements of the results is the same as the
6177 // # elements of the compare (and the # elements of the compare result
6178 // for that matter). Check to see that they are the same size. If so,
6179 // we know that the element size of the sext'd result matches the
6180 // element size of the compare operands.
6181 if (VT.getSizeInBits() == N0VT.getSizeInBits())
6182 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
6184 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6185 // If the desired elements are smaller or larger than the source
6186 // elements we can use a matching integer vector type and then
6187 // truncate/any extend
6189 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
6191 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
6193 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6194 return DAG.getAnyExtOrTrunc(VsetCC, SDLoc(N), VT);
6198 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
6200 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
6201 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
6202 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
6210 /// See if the specified operand can be simplified with the knowledge that only
6211 /// the bits specified by Mask are used. If so, return the simpler operand,
6212 /// otherwise return a null SDValue.
6213 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
6214 switch (V.getOpcode()) {
6216 case ISD::Constant: {
6217 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
6218 assert(CV && "Const value should be ConstSDNode.");
6219 const APInt &CVal = CV->getAPIntValue();
6220 APInt NewVal = CVal & Mask;
6222 return DAG.getConstant(NewVal, V.getValueType());
6227 // If the LHS or RHS don't contribute bits to the or, drop them.
6228 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
6229 return V.getOperand(1);
6230 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
6231 return V.getOperand(0);
6234 // Only look at single-use SRLs.
6235 if (!V.getNode()->hasOneUse())
6237 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
6238 // See if we can recursively simplify the LHS.
6239 unsigned Amt = RHSC->getZExtValue();
6241 // Watch out for shift count overflow though.
6242 if (Amt >= Mask.getBitWidth()) break;
6243 APInt NewMask = Mask << Amt;
6244 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
6245 if (SimplifyLHS.getNode())
6246 return DAG.getNode(ISD::SRL, SDLoc(V), V.getValueType(),
6247 SimplifyLHS, V.getOperand(1));
6253 /// If the result of a wider load is shifted to right of N bits and then
6254 /// truncated to a narrower type and where N is a multiple of number of bits of
6255 /// the narrower type, transform it to a narrower load from address + N / num of
6256 /// bits of new type. If the result is to be extended, also fold the extension
6257 /// to form a extending load.
6258 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
6259 unsigned Opc = N->getOpcode();
6261 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
6262 SDValue N0 = N->getOperand(0);
6263 EVT VT = N->getValueType(0);
6266 // This transformation isn't valid for vector loads.
6270 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
6272 if (Opc == ISD::SIGN_EXTEND_INREG) {
6273 ExtType = ISD::SEXTLOAD;
6274 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
6275 } else if (Opc == ISD::SRL) {
6276 // Another special-case: SRL is basically zero-extending a narrower value.
6277 ExtType = ISD::ZEXTLOAD;
6279 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
6280 if (!N01) return SDValue();
6281 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
6282 VT.getSizeInBits() - N01->getZExtValue());
6284 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, VT, ExtVT))
6287 unsigned EVTBits = ExtVT.getSizeInBits();
6289 // Do not generate loads of non-round integer types since these can
6290 // be expensive (and would be wrong if the type is not byte sized).
6291 if (!ExtVT.isRound())
6295 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
6296 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
6297 ShAmt = N01->getZExtValue();
6298 // Is the shift amount a multiple of size of VT?
6299 if ((ShAmt & (EVTBits-1)) == 0) {
6300 N0 = N0.getOperand(0);
6301 // Is the load width a multiple of size of VT?
6302 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
6306 // At this point, we must have a load or else we can't do the transform.
6307 if (!isa<LoadSDNode>(N0)) return SDValue();
6309 // Because a SRL must be assumed to *need* to zero-extend the high bits
6310 // (as opposed to anyext the high bits), we can't combine the zextload
6311 // lowering of SRL and an sextload.
6312 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
6315 // If the shift amount is larger than the input type then we're not
6316 // accessing any of the loaded bytes. If the load was a zextload/extload
6317 // then the result of the shift+trunc is zero/undef (handled elsewhere).
6318 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
6323 // If the load is shifted left (and the result isn't shifted back right),
6324 // we can fold the truncate through the shift.
6325 unsigned ShLeftAmt = 0;
6326 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
6327 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
6328 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
6329 ShLeftAmt = N01->getZExtValue();
6330 N0 = N0.getOperand(0);
6334 // If we haven't found a load, we can't narrow it. Don't transform one with
6335 // multiple uses, this would require adding a new load.
6336 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse())
6339 // Don't change the width of a volatile load.
6340 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6341 if (LN0->isVolatile())
6344 // Verify that we are actually reducing a load width here.
6345 if (LN0->getMemoryVT().getSizeInBits() < EVTBits)
6348 // For the transform to be legal, the load must produce only two values
6349 // (the value loaded and the chain). Don't transform a pre-increment
6350 // load, for example, which produces an extra value. Otherwise the
6351 // transformation is not equivalent, and the downstream logic to replace
6352 // uses gets things wrong.
6353 if (LN0->getNumValues() > 2)
6356 // If the load that we're shrinking is an extload and we're not just
6357 // discarding the extension we can't simply shrink the load. Bail.
6358 // TODO: It would be possible to merge the extensions in some cases.
6359 if (LN0->getExtensionType() != ISD::NON_EXTLOAD &&
6360 LN0->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits() + ShAmt)
6363 if (!TLI.shouldReduceLoadWidth(LN0, ExtType, ExtVT))
6366 EVT PtrType = N0.getOperand(1).getValueType();
6368 if (PtrType == MVT::Untyped || PtrType.isExtended())
6369 // It's not possible to generate a constant of extended or untyped type.
6372 // For big endian targets, we need to adjust the offset to the pointer to
6373 // load the correct bytes.
6374 if (TLI.isBigEndian()) {
6375 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
6376 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
6377 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
6380 uint64_t PtrOff = ShAmt / 8;
6381 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
6382 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0),
6383 PtrType, LN0->getBasePtr(),
6384 DAG.getConstant(PtrOff, PtrType));
6385 AddToWorklist(NewPtr.getNode());
6388 if (ExtType == ISD::NON_EXTLOAD)
6389 Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr,
6390 LN0->getPointerInfo().getWithOffset(PtrOff),
6391 LN0->isVolatile(), LN0->isNonTemporal(),
6392 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
6394 Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(),NewPtr,
6395 LN0->getPointerInfo().getWithOffset(PtrOff),
6396 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
6397 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
6399 // Replace the old load's chain with the new load's chain.
6400 WorklistRemover DeadNodes(*this);
6401 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
6403 // Shift the result left, if we've swallowed a left shift.
6404 SDValue Result = Load;
6405 if (ShLeftAmt != 0) {
6406 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
6407 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
6409 // If the shift amount is as large as the result size (but, presumably,
6410 // no larger than the source) then the useful bits of the result are
6411 // zero; we can't simply return the shortened shift, because the result
6412 // of that operation is undefined.
6413 if (ShLeftAmt >= VT.getSizeInBits())
6414 Result = DAG.getConstant(0, VT);
6416 Result = DAG.getNode(ISD::SHL, SDLoc(N0), VT,
6417 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
6420 // Return the new loaded value.
6424 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
6425 SDValue N0 = N->getOperand(0);
6426 SDValue N1 = N->getOperand(1);
6427 EVT VT = N->getValueType(0);
6428 EVT EVT = cast<VTSDNode>(N1)->getVT();
6429 unsigned VTBits = VT.getScalarType().getSizeInBits();
6430 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
6432 // fold (sext_in_reg c1) -> c1
6433 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
6434 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1);
6436 // If the input is already sign extended, just drop the extension.
6437 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
6440 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
6441 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
6442 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT()))
6443 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
6444 N0.getOperand(0), N1);
6446 // fold (sext_in_reg (sext x)) -> (sext x)
6447 // fold (sext_in_reg (aext x)) -> (sext x)
6448 // if x is small enough.
6449 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
6450 SDValue N00 = N0.getOperand(0);
6451 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
6452 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
6453 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
6456 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
6457 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
6458 return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT);
6460 // fold operands of sext_in_reg based on knowledge that the top bits are not
6462 if (SimplifyDemandedBits(SDValue(N, 0)))
6463 return SDValue(N, 0);
6465 // fold (sext_in_reg (load x)) -> (smaller sextload x)
6466 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
6467 SDValue NarrowLoad = ReduceLoadWidth(N);
6468 if (NarrowLoad.getNode())
6471 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
6472 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
6473 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
6474 if (N0.getOpcode() == ISD::SRL) {
6475 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
6476 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
6477 // We can turn this into an SRA iff the input to the SRL is already sign
6479 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
6480 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
6481 return DAG.getNode(ISD::SRA, SDLoc(N), VT,
6482 N0.getOperand(0), N0.getOperand(1));
6486 // fold (sext_inreg (extload x)) -> (sextload x)
6487 if (ISD::isEXTLoad(N0.getNode()) &&
6488 ISD::isUNINDEXEDLoad(N0.getNode()) &&
6489 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
6490 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6491 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) {
6492 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6493 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
6495 LN0->getBasePtr(), EVT,
6496 LN0->getMemOperand());
6497 CombineTo(N, ExtLoad);
6498 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
6499 AddToWorklist(ExtLoad.getNode());
6500 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6502 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
6503 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
6505 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
6506 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6507 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) {
6508 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6509 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
6511 LN0->getBasePtr(), EVT,
6512 LN0->getMemOperand());
6513 CombineTo(N, ExtLoad);
6514 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
6515 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6518 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
6519 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
6520 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
6521 N0.getOperand(1), false);
6522 if (BSwap.getNode())
6523 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
6527 // Fold a sext_inreg of a build_vector of ConstantSDNodes or undefs
6528 // into a build_vector.
6529 if (ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
6530 SmallVector<SDValue, 8> Elts;
6531 unsigned NumElts = N0->getNumOperands();
6532 unsigned ShAmt = VTBits - EVTBits;
6534 for (unsigned i = 0; i != NumElts; ++i) {
6535 SDValue Op = N0->getOperand(i);
6536 if (Op->getOpcode() == ISD::UNDEF) {
6541 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
6542 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
6543 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
6544 Op.getValueType()));
6547 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Elts);
6553 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
6554 SDValue N0 = N->getOperand(0);
6555 EVT VT = N->getValueType(0);
6556 bool isLE = TLI.isLittleEndian();
6559 if (N0.getValueType() == N->getValueType(0))
6561 // fold (truncate c1) -> c1
6562 if (isConstantIntBuildVectorOrConstantInt(N0))
6563 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0);
6564 // fold (truncate (truncate x)) -> (truncate x)
6565 if (N0.getOpcode() == ISD::TRUNCATE)
6566 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6567 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
6568 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
6569 N0.getOpcode() == ISD::SIGN_EXTEND ||
6570 N0.getOpcode() == ISD::ANY_EXTEND) {
6571 if (N0.getOperand(0).getValueType().bitsLT(VT))
6572 // if the source is smaller than the dest, we still need an extend
6573 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
6575 if (N0.getOperand(0).getValueType().bitsGT(VT))
6576 // if the source is larger than the dest, than we just need the truncate
6577 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6578 // if the source and dest are the same type, we can drop both the extend
6579 // and the truncate.
6580 return N0.getOperand(0);
6583 // Fold extract-and-trunc into a narrow extract. For example:
6584 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
6585 // i32 y = TRUNCATE(i64 x)
6587 // v16i8 b = BITCAST (v2i64 val)
6588 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
6590 // Note: We only run this optimization after type legalization (which often
6591 // creates this pattern) and before operation legalization after which
6592 // we need to be more careful about the vector instructions that we generate.
6593 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6594 LegalTypes && !LegalOperations && N0->hasOneUse() && VT != MVT::i1) {
6596 EVT VecTy = N0.getOperand(0).getValueType();
6597 EVT ExTy = N0.getValueType();
6598 EVT TrTy = N->getValueType(0);
6600 unsigned NumElem = VecTy.getVectorNumElements();
6601 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
6603 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
6604 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
6606 SDValue EltNo = N0->getOperand(1);
6607 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
6608 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
6609 EVT IndexTy = TLI.getVectorIdxTy();
6610 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
6612 SDValue V = DAG.getNode(ISD::BITCAST, SDLoc(N),
6613 NVT, N0.getOperand(0));
6615 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
6617 DAG.getConstant(Index, IndexTy));
6621 // trunc (select c, a, b) -> select c, (trunc a), (trunc b)
6622 if (N0.getOpcode() == ISD::SELECT) {
6623 EVT SrcVT = N0.getValueType();
6624 if ((!LegalOperations || TLI.isOperationLegal(ISD::SELECT, SrcVT)) &&
6625 TLI.isTruncateFree(SrcVT, VT)) {
6627 SDValue Cond = N0.getOperand(0);
6628 SDValue TruncOp0 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(1));
6629 SDValue TruncOp1 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(2));
6630 return DAG.getNode(ISD::SELECT, SDLoc(N), VT, Cond, TruncOp0, TruncOp1);
6634 // Fold a series of buildvector, bitcast, and truncate if possible.
6636 // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to
6637 // (2xi32 (buildvector x, y)).
6638 if (Level == AfterLegalizeVectorOps && VT.isVector() &&
6639 N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
6640 N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
6641 N0.getOperand(0).hasOneUse()) {
6643 SDValue BuildVect = N0.getOperand(0);
6644 EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType();
6645 EVT TruncVecEltTy = VT.getVectorElementType();
6647 // Check that the element types match.
6648 if (BuildVectEltTy == TruncVecEltTy) {
6649 // Now we only need to compute the offset of the truncated elements.
6650 unsigned BuildVecNumElts = BuildVect.getNumOperands();
6651 unsigned TruncVecNumElts = VT.getVectorNumElements();
6652 unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts;
6654 assert((BuildVecNumElts % TruncVecNumElts) == 0 &&
6655 "Invalid number of elements");
6657 SmallVector<SDValue, 8> Opnds;
6658 for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset)
6659 Opnds.push_back(BuildVect.getOperand(i));
6661 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
6665 // See if we can simplify the input to this truncate through knowledge that
6666 // only the low bits are being used.
6667 // For example "trunc (or (shl x, 8), y)" // -> trunc y
6668 // Currently we only perform this optimization on scalars because vectors
6669 // may have different active low bits.
6670 if (!VT.isVector()) {
6672 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
6673 VT.getSizeInBits()));
6674 if (Shorter.getNode())
6675 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter);
6677 // fold (truncate (load x)) -> (smaller load x)
6678 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
6679 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
6680 SDValue Reduced = ReduceLoadWidth(N);
6681 if (Reduced.getNode())
6683 // Handle the case where the load remains an extending load even
6684 // after truncation.
6685 if (N0.hasOneUse() && ISD::isUNINDEXEDLoad(N0.getNode())) {
6686 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6687 if (!LN0->isVolatile() &&
6688 LN0->getMemoryVT().getStoreSizeInBits() < VT.getSizeInBits()) {
6689 SDValue NewLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(LN0),
6690 VT, LN0->getChain(), LN0->getBasePtr(),
6692 LN0->getMemOperand());
6693 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLoad.getValue(1));
6698 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
6699 // where ... are all 'undef'.
6700 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
6701 SmallVector<EVT, 8> VTs;
6704 unsigned NumDefs = 0;
6706 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
6707 SDValue X = N0.getOperand(i);
6708 if (X.getOpcode() != ISD::UNDEF) {
6713 // Stop if more than one members are non-undef.
6716 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
6717 VT.getVectorElementType(),
6718 X.getValueType().getVectorNumElements()));
6722 return DAG.getUNDEF(VT);
6725 assert(V.getNode() && "The single defined operand is empty!");
6726 SmallVector<SDValue, 8> Opnds;
6727 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
6729 Opnds.push_back(DAG.getUNDEF(VTs[i]));
6732 SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V);
6733 AddToWorklist(NV.getNode());
6734 Opnds.push_back(NV);
6736 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Opnds);
6740 // Simplify the operands using demanded-bits information.
6741 if (!VT.isVector() &&
6742 SimplifyDemandedBits(SDValue(N, 0)))
6743 return SDValue(N, 0);
6748 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
6749 SDValue Elt = N->getOperand(i);
6750 if (Elt.getOpcode() != ISD::MERGE_VALUES)
6751 return Elt.getNode();
6752 return Elt.getOperand(Elt.getResNo()).getNode();
6755 /// build_pair (load, load) -> load
6756 /// if load locations are consecutive.
6757 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
6758 assert(N->getOpcode() == ISD::BUILD_PAIR);
6760 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
6761 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
6762 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
6763 LD1->getAddressSpace() != LD2->getAddressSpace())
6765 EVT LD1VT = LD1->getValueType(0);
6767 if (ISD::isNON_EXTLoad(LD2) &&
6769 // If both are volatile this would reduce the number of volatile loads.
6770 // If one is volatile it might be ok, but play conservative and bail out.
6771 !LD1->isVolatile() &&
6772 !LD2->isVolatile() &&
6773 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
6774 unsigned Align = LD1->getAlignment();
6775 unsigned NewAlign = TLI.getDataLayout()->
6776 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6778 if (NewAlign <= Align &&
6779 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
6780 return DAG.getLoad(VT, SDLoc(N), LD1->getChain(),
6781 LD1->getBasePtr(), LD1->getPointerInfo(),
6782 false, false, false, Align);
6788 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
6789 SDValue N0 = N->getOperand(0);
6790 EVT VT = N->getValueType(0);
6792 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
6793 // Only do this before legalize, since afterward the target may be depending
6794 // on the bitconvert.
6795 // First check to see if this is all constant.
6797 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
6799 bool isSimple = cast<BuildVectorSDNode>(N0)->isConstant();
6801 EVT DestEltVT = N->getValueType(0).getVectorElementType();
6802 assert(!DestEltVT.isVector() &&
6803 "Element type of vector ValueType must not be vector!");
6805 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
6808 // If the input is a constant, let getNode fold it.
6809 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
6810 // If we can't allow illegal operations, we need to check that this is just
6811 // a fp -> int or int -> conversion and that the resulting operation will
6813 if (!LegalOperations ||
6814 (isa<ConstantSDNode>(N0) && VT.isFloatingPoint() && !VT.isVector() &&
6815 TLI.isOperationLegal(ISD::ConstantFP, VT)) ||
6816 (isa<ConstantFPSDNode>(N0) && VT.isInteger() && !VT.isVector() &&
6817 TLI.isOperationLegal(ISD::Constant, VT)))
6818 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, N0);
6821 // (conv (conv x, t1), t2) -> (conv x, t2)
6822 if (N0.getOpcode() == ISD::BITCAST)
6823 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT,
6826 // fold (conv (load x)) -> (load (conv*)x)
6827 // If the resultant load doesn't need a higher alignment than the original!
6828 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6829 // Do not change the width of a volatile load.
6830 !cast<LoadSDNode>(N0)->isVolatile() &&
6831 // Do not remove the cast if the types differ in endian layout.
6832 TLI.hasBigEndianPartOrdering(N0.getValueType()) ==
6833 TLI.hasBigEndianPartOrdering(VT) &&
6834 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)) &&
6835 TLI.isLoadBitCastBeneficial(N0.getValueType(), VT)) {
6836 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6837 unsigned Align = TLI.getDataLayout()->
6838 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6839 unsigned OrigAlign = LN0->getAlignment();
6841 if (Align <= OrigAlign) {
6842 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(),
6843 LN0->getBasePtr(), LN0->getPointerInfo(),
6844 LN0->isVolatile(), LN0->isNonTemporal(),
6845 LN0->isInvariant(), OrigAlign,
6847 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
6852 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
6853 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
6854 // This often reduces constant pool loads.
6855 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) ||
6856 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) &&
6857 N0.getNode()->hasOneUse() && VT.isInteger() &&
6858 !VT.isVector() && !N0.getValueType().isVector()) {
6859 SDValue NewConv = DAG.getNode(ISD::BITCAST, SDLoc(N0), VT,
6861 AddToWorklist(NewConv.getNode());
6863 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6864 if (N0.getOpcode() == ISD::FNEG)
6865 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
6866 NewConv, DAG.getConstant(SignBit, VT));
6867 assert(N0.getOpcode() == ISD::FABS);
6868 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6869 NewConv, DAG.getConstant(~SignBit, VT));
6872 // fold (bitconvert (fcopysign cst, x)) ->
6873 // (or (and (bitconvert x), sign), (and cst, (not sign)))
6874 // Note that we don't handle (copysign x, cst) because this can always be
6875 // folded to an fneg or fabs.
6876 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
6877 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
6878 VT.isInteger() && !VT.isVector()) {
6879 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
6880 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
6881 if (isTypeLegal(IntXVT)) {
6882 SDValue X = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6883 IntXVT, N0.getOperand(1));
6884 AddToWorklist(X.getNode());
6886 // If X has a different width than the result/lhs, sext it or truncate it.
6887 unsigned VTWidth = VT.getSizeInBits();
6888 if (OrigXWidth < VTWidth) {
6889 X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X);
6890 AddToWorklist(X.getNode());
6891 } else if (OrigXWidth > VTWidth) {
6892 // To get the sign bit in the right place, we have to shift it right
6893 // before truncating.
6894 X = DAG.getNode(ISD::SRL, SDLoc(X),
6895 X.getValueType(), X,
6896 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
6897 AddToWorklist(X.getNode());
6898 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
6899 AddToWorklist(X.getNode());
6902 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6903 X = DAG.getNode(ISD::AND, SDLoc(X), VT,
6904 X, DAG.getConstant(SignBit, VT));
6905 AddToWorklist(X.getNode());
6907 SDValue Cst = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6908 VT, N0.getOperand(0));
6909 Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT,
6910 Cst, DAG.getConstant(~SignBit, VT));
6911 AddToWorklist(Cst.getNode());
6913 return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst);
6917 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
6918 if (N0.getOpcode() == ISD::BUILD_PAIR) {
6919 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
6920 if (CombineLD.getNode())
6927 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
6928 EVT VT = N->getValueType(0);
6929 return CombineConsecutiveLoads(N, VT);
6932 /// We know that BV is a build_vector node with Constant, ConstantFP or Undef
6933 /// operands. DstEltVT indicates the destination element value type.
6934 SDValue DAGCombiner::
6935 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
6936 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
6938 // If this is already the right type, we're done.
6939 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
6941 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
6942 unsigned DstBitSize = DstEltVT.getSizeInBits();
6944 // If this is a conversion of N elements of one type to N elements of another
6945 // type, convert each element. This handles FP<->INT cases.
6946 if (SrcBitSize == DstBitSize) {
6947 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6948 BV->getValueType(0).getVectorNumElements());
6950 // Due to the FP element handling below calling this routine recursively,
6951 // we can end up with a scalar-to-vector node here.
6952 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
6953 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6954 DAG.getNode(ISD::BITCAST, SDLoc(BV),
6955 DstEltVT, BV->getOperand(0)));
6957 SmallVector<SDValue, 8> Ops;
6958 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6959 SDValue Op = BV->getOperand(i);
6960 // If the vector element type is not legal, the BUILD_VECTOR operands
6961 // are promoted and implicitly truncated. Make that explicit here.
6962 if (Op.getValueType() != SrcEltVT)
6963 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op);
6964 Ops.push_back(DAG.getNode(ISD::BITCAST, SDLoc(BV),
6966 AddToWorklist(Ops.back().getNode());
6968 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6971 // Otherwise, we're growing or shrinking the elements. To avoid having to
6972 // handle annoying details of growing/shrinking FP values, we convert them to
6974 if (SrcEltVT.isFloatingPoint()) {
6975 // Convert the input float vector to a int vector where the elements are the
6977 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
6978 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
6982 // Now we know the input is an integer vector. If the output is a FP type,
6983 // convert to integer first, then to FP of the right size.
6984 if (DstEltVT.isFloatingPoint()) {
6985 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
6986 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
6988 // Next, convert to FP elements of the same size.
6989 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
6992 // Okay, we know the src/dst types are both integers of differing types.
6993 // Handling growing first.
6994 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
6995 if (SrcBitSize < DstBitSize) {
6996 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
6998 SmallVector<SDValue, 8> Ops;
6999 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
7000 i += NumInputsPerOutput) {
7001 bool isLE = TLI.isLittleEndian();
7002 APInt NewBits = APInt(DstBitSize, 0);
7003 bool EltIsUndef = true;
7004 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
7005 // Shift the previously computed bits over.
7006 NewBits <<= SrcBitSize;
7007 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
7008 if (Op.getOpcode() == ISD::UNDEF) continue;
7011 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
7012 zextOrTrunc(SrcBitSize).zext(DstBitSize);
7016 Ops.push_back(DAG.getUNDEF(DstEltVT));
7018 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
7021 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
7022 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
7025 // Finally, this must be the case where we are shrinking elements: each input
7026 // turns into multiple outputs.
7027 bool isS2V = ISD::isScalarToVector(BV);
7028 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
7029 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
7030 NumOutputsPerInput*BV->getNumOperands());
7031 SmallVector<SDValue, 8> Ops;
7033 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
7034 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
7035 Ops.append(NumOutputsPerInput, DAG.getUNDEF(DstEltVT));
7039 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
7040 getAPIntValue().zextOrTrunc(SrcBitSize);
7042 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
7043 APInt ThisVal = OpVal.trunc(DstBitSize);
7044 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
7045 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
7046 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
7047 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
7049 OpVal = OpVal.lshr(DstBitSize);
7052 // For big endian targets, swap the order of the pieces of each element.
7053 if (TLI.isBigEndian())
7054 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
7057 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
7060 // Attempt different variants of (fadd (fmul a, b), c) -> fma or fmad
7061 static SDValue performFaddFmulCombines(unsigned FusedOpcode,
7064 const TargetLowering &TLI,
7065 SelectionDAG &DAG) {
7066 SDValue N0 = N->getOperand(0);
7067 SDValue N1 = N->getOperand(1);
7068 EVT VT = N->getValueType(0);
7070 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
7071 if (N0.getOpcode() == ISD::FMUL &&
7072 (Aggressive || N0->hasOneUse())) {
7073 return DAG.getNode(FusedOpcode, SDLoc(N), VT,
7074 N0.getOperand(0), N0.getOperand(1), N1);
7077 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
7078 // Note: Commutes FADD operands.
7079 if (N1.getOpcode() == ISD::FMUL &&
7080 (Aggressive || N1->hasOneUse())) {
7081 return DAG.getNode(FusedOpcode, SDLoc(N), VT,
7082 N1.getOperand(0), N1.getOperand(1), N0);
7085 // More folding opportunities when target permits.
7087 // fold (fadd (fma x, y, (fmul u, v)), z) -> (fma x, y (fma u, v, z))
7088 if (N0.getOpcode() == ISD::FMA &&
7089 N0.getOperand(2).getOpcode() == ISD::FMUL) {
7090 return DAG.getNode(FusedOpcode, SDLoc(N), VT,
7091 N0.getOperand(0), N0.getOperand(1),
7092 DAG.getNode(FusedOpcode, SDLoc(N), VT,
7093 N0.getOperand(2).getOperand(0),
7094 N0.getOperand(2).getOperand(1),
7098 // fold (fadd x, (fma y, z, (fmul u, v)) -> (fma y, z (fma u, v, x))
7099 if (N1->getOpcode() == ISD::FMA &&
7100 N1.getOperand(2).getOpcode() == ISD::FMUL) {
7101 return DAG.getNode(FusedOpcode, SDLoc(N), VT,
7102 N1.getOperand(0), N1.getOperand(1),
7103 DAG.getNode(FusedOpcode, SDLoc(N), VT,
7104 N1.getOperand(2).getOperand(0),
7105 N1.getOperand(2).getOperand(1),
7113 static SDValue performFsubFmulCombines(unsigned FusedOpcode,
7116 const TargetLowering &TLI,
7117 SelectionDAG &DAG) {
7118 SDValue N0 = N->getOperand(0);
7119 SDValue N1 = N->getOperand(1);
7120 EVT VT = N->getValueType(0);
7124 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
7125 if (N0.getOpcode() == ISD::FMUL &&
7126 (Aggressive || N0->hasOneUse())) {
7127 return DAG.getNode(FusedOpcode, SL, VT,
7128 N0.getOperand(0), N0.getOperand(1),
7129 DAG.getNode(ISD::FNEG, SL, VT, N1));
7132 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
7133 // Note: Commutes FSUB operands.
7134 if (N1.getOpcode() == ISD::FMUL &&
7135 (Aggressive || N1->hasOneUse()))
7136 return DAG.getNode(FusedOpcode, SL, VT,
7137 DAG.getNode(ISD::FNEG, SL, VT,
7139 N1.getOperand(1), N0);
7141 // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
7142 if (N0.getOpcode() == ISD::FNEG &&
7143 N0.getOperand(0).getOpcode() == ISD::FMUL &&
7144 (Aggressive || (N0->hasOneUse() && N0.getOperand(0).hasOneUse()))) {
7145 SDValue N00 = N0.getOperand(0).getOperand(0);
7146 SDValue N01 = N0.getOperand(0).getOperand(1);
7147 return DAG.getNode(FusedOpcode, SL, VT,
7148 DAG.getNode(ISD::FNEG, SL, VT, N00), N01,
7149 DAG.getNode(ISD::FNEG, SL, VT, N1));
7152 // More folding opportunities when target permits.
7154 // fold (fsub (fma x, y, (fmul u, v)), z)
7155 // -> (fma x, y (fma u, v, (fneg z)))
7156 if (N0.getOpcode() == FusedOpcode &&
7157 N0.getOperand(2).getOpcode() == ISD::FMUL) {
7158 return DAG.getNode(FusedOpcode, SDLoc(N), VT,
7159 N0.getOperand(0), N0.getOperand(1),
7160 DAG.getNode(FusedOpcode, SDLoc(N), VT,
7161 N0.getOperand(2).getOperand(0),
7162 N0.getOperand(2).getOperand(1),
7163 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7167 // fold (fsub x, (fma y, z, (fmul u, v)))
7168 // -> (fma (fneg y), z, (fma (fneg u), v, x))
7169 if (N1.getOpcode() == FusedOpcode &&
7170 N1.getOperand(2).getOpcode() == ISD::FMUL) {
7171 SDValue N20 = N1.getOperand(2).getOperand(0);
7172 SDValue N21 = N1.getOperand(2).getOperand(1);
7173 return DAG.getNode(FusedOpcode, SDLoc(N), VT,
7174 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7177 DAG.getNode(FusedOpcode, SDLoc(N), VT,
7178 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7187 SDValue DAGCombiner::visitFADD(SDNode *N) {
7188 SDValue N0 = N->getOperand(0);
7189 SDValue N1 = N->getOperand(1);
7190 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7191 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7192 EVT VT = N->getValueType(0);
7193 const TargetOptions &Options = DAG.getTarget().Options;
7197 if (SDValue FoldedVOp = SimplifyVBinOp(N))
7200 // fold (fadd c1, c2) -> c1 + c2
7202 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N1);
7204 // canonicalize constant to RHS
7205 if (N0CFP && !N1CFP)
7206 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N0);
7208 // fold (fadd A, (fneg B)) -> (fsub A, B)
7209 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
7210 isNegatibleForFree(N1, LegalOperations, TLI, &Options) == 2)
7211 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0,
7212 GetNegatedExpression(N1, DAG, LegalOperations));
7214 // fold (fadd (fneg A), B) -> (fsub B, A)
7215 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
7216 isNegatibleForFree(N0, LegalOperations, TLI, &Options) == 2)
7217 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N1,
7218 GetNegatedExpression(N0, DAG, LegalOperations));
7220 // If 'unsafe math' is enabled, fold lots of things.
7221 if (Options.UnsafeFPMath) {
7222 // No FP constant should be created after legalization as Instruction
7223 // Selection pass has a hard time dealing with FP constants.
7224 bool AllowNewConst = (Level < AfterLegalizeDAG);
7226 // fold (fadd A, 0) -> A
7227 if (N1CFP && N1CFP->getValueAPF().isZero())
7230 // fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
7231 if (N1CFP && N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
7232 isa<ConstantFPSDNode>(N0.getOperand(1)))
7233 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0.getOperand(0),
7234 DAG.getNode(ISD::FADD, SDLoc(N), VT,
7235 N0.getOperand(1), N1));
7237 // If allowed, fold (fadd (fneg x), x) -> 0.0
7238 if (AllowNewConst && N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1)
7239 return DAG.getConstantFP(0.0, VT);
7241 // If allowed, fold (fadd x, (fneg x)) -> 0.0
7242 if (AllowNewConst && N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0)
7243 return DAG.getConstantFP(0.0, VT);
7245 // We can fold chains of FADD's of the same value into multiplications.
7246 // This transform is not safe in general because we are reducing the number
7247 // of rounding steps.
7248 if (TLI.isOperationLegalOrCustom(ISD::FMUL, VT) && !N0CFP && !N1CFP) {
7249 if (N0.getOpcode() == ISD::FMUL) {
7250 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
7251 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
7253 // (fadd (fmul x, c), x) -> (fmul x, c+1)
7254 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
7255 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
7257 DAG.getConstantFP(1.0, VT));
7258 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, NewCFP);
7261 // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2)
7262 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
7263 N1.getOperand(0) == N1.getOperand(1) &&
7264 N0.getOperand(0) == N1.getOperand(0)) {
7265 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
7267 DAG.getConstantFP(2.0, VT));
7268 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7269 N0.getOperand(0), NewCFP);
7273 if (N1.getOpcode() == ISD::FMUL) {
7274 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
7275 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
7277 // (fadd x, (fmul x, c)) -> (fmul x, c+1)
7278 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
7279 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
7281 DAG.getConstantFP(1.0, VT));
7282 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, NewCFP);
7285 // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2)
7286 if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD &&
7287 N0.getOperand(0) == N0.getOperand(1) &&
7288 N1.getOperand(0) == N0.getOperand(0)) {
7289 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
7291 DAG.getConstantFP(2.0, VT));
7292 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1.getOperand(0), NewCFP);
7296 if (N0.getOpcode() == ISD::FADD && AllowNewConst) {
7297 ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
7298 // (fadd (fadd x, x), x) -> (fmul x, 3.0)
7299 if (!CFP && N0.getOperand(0) == N0.getOperand(1) &&
7300 (N0.getOperand(0) == N1))
7301 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7302 N1, DAG.getConstantFP(3.0, VT));
7305 if (N1.getOpcode() == ISD::FADD && AllowNewConst) {
7306 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
7307 // (fadd x, (fadd x, x)) -> (fmul x, 3.0)
7308 if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) &&
7309 N1.getOperand(0) == N0)
7310 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7311 N0, DAG.getConstantFP(3.0, VT));
7314 // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0)
7315 if (AllowNewConst &&
7316 N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
7317 N0.getOperand(0) == N0.getOperand(1) &&
7318 N1.getOperand(0) == N1.getOperand(1) &&
7319 N0.getOperand(0) == N1.getOperand(0))
7320 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7321 N0.getOperand(0), DAG.getConstantFP(4.0, VT));
7323 } // enable-unsafe-fp-math
7325 if (LegalOperations && TLI.isOperationLegal(ISD::FMAD, VT)) {
7326 // Assume if there is an fmad instruction that it should be aggressively
7328 if (SDValue Fused = performFaddFmulCombines(ISD::FMAD, true, N, TLI, DAG))
7332 // FADD -> FMA combines:
7333 if ((Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath) &&
7334 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
7335 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
7337 if (!TLI.isOperationLegal(ISD::FMAD, VT)) {
7338 // Don't form FMA if we are preferring FMAD.
7340 = performFaddFmulCombines(ISD::FMA,
7341 TLI.enableAggressiveFMAFusion(VT),
7347 // When FP_EXTEND nodes are free on the target, and there is an opportunity
7348 // to combine into FMA, arrange such nodes accordingly.
7349 if (TLI.isFPExtFree(VT)) {
7351 // fold (fadd (fpext (fmul x, y)), z) -> (fma (fpext x), (fpext y), z)
7352 if (N0.getOpcode() == ISD::FP_EXTEND) {
7353 SDValue N00 = N0.getOperand(0);
7354 if (N00.getOpcode() == ISD::FMUL)
7355 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7356 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7358 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7359 N00.getOperand(1)), N1);
7362 // fold (fadd x, (fpext (fmul y, z)), z) -> (fma (fpext y), (fpext z), x)
7363 // Note: Commutes FADD operands.
7364 if (N1.getOpcode() == ISD::FP_EXTEND) {
7365 SDValue N10 = N1.getOperand(0);
7366 if (N10.getOpcode() == ISD::FMUL)
7367 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7368 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7370 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7371 N10.getOperand(1)), N0);
7379 SDValue DAGCombiner::visitFSUB(SDNode *N) {
7380 SDValue N0 = N->getOperand(0);
7381 SDValue N1 = N->getOperand(1);
7382 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
7383 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
7384 EVT VT = N->getValueType(0);
7386 const TargetOptions &Options = DAG.getTarget().Options;
7390 if (SDValue FoldedVOp = SimplifyVBinOp(N))
7393 // fold (fsub c1, c2) -> c1-c2
7395 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0, N1);
7397 // fold (fsub A, (fneg B)) -> (fadd A, B)
7398 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
7399 return DAG.getNode(ISD::FADD, dl, VT, N0,
7400 GetNegatedExpression(N1, DAG, LegalOperations));
7402 // If 'unsafe math' is enabled, fold lots of things.
7403 if (Options.UnsafeFPMath) {
7405 if (N1CFP && N1CFP->getValueAPF().isZero())
7408 // (fsub 0, B) -> -B
7409 if (N0CFP && N0CFP->getValueAPF().isZero()) {
7410 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
7411 return GetNegatedExpression(N1, DAG, LegalOperations);
7412 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7413 return DAG.getNode(ISD::FNEG, dl, VT, N1);
7416 // (fsub x, x) -> 0.0
7418 return DAG.getConstantFP(0.0f, VT);
7420 // (fsub x, (fadd x, y)) -> (fneg y)
7421 // (fsub x, (fadd y, x)) -> (fneg y)
7422 if (N1.getOpcode() == ISD::FADD) {
7423 SDValue N10 = N1->getOperand(0);
7424 SDValue N11 = N1->getOperand(1);
7426 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI, &Options))
7427 return GetNegatedExpression(N11, DAG, LegalOperations);
7429 if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI, &Options))
7430 return GetNegatedExpression(N10, DAG, LegalOperations);
7434 if (LegalOperations && TLI.isOperationLegal(ISD::FMAD, VT)) {
7435 // Assume if there is an fmad instruction that it should be aggressively
7437 if (SDValue Fused = performFsubFmulCombines(ISD::FMAD, true, N, TLI, DAG))
7441 // FSUB -> FMA combines:
7442 if ((Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath) &&
7443 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
7444 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
7446 if (!TLI.isOperationLegal(ISD::FMAD, VT)) {
7447 // Don't form FMA if we are preferring FMAD.
7450 = performFsubFmulCombines(ISD::FMA,
7451 TLI.enableAggressiveFMAFusion(VT),
7457 // When FP_EXTEND nodes are free on the target, and there is an opportunity
7458 // to combine into FMA, arrange such nodes accordingly.
7459 if (TLI.isFPExtFree(VT)) {
7460 // fold (fsub (fpext (fmul x, y)), z)
7461 // -> (fma (fpext x), (fpext y), (fneg z))
7462 if (N0.getOpcode() == ISD::FP_EXTEND) {
7463 SDValue N00 = N0.getOperand(0);
7464 if (N00.getOpcode() == ISD::FMUL)
7465 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7466 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7468 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7470 DAG.getNode(ISD::FNEG, SDLoc(N), VT, N1));
7473 // fold (fsub x, (fpext (fmul y, z)))
7474 // -> (fma (fneg (fpext y)), (fpext z), x)
7475 // Note: Commutes FSUB operands.
7476 if (N1.getOpcode() == ISD::FP_EXTEND) {
7477 SDValue N10 = N1.getOperand(0);
7478 if (N10.getOpcode() == ISD::FMUL)
7479 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7480 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7481 DAG.getNode(ISD::FP_EXTEND, SDLoc(N),
7482 VT, N10.getOperand(0))),
7483 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7488 // fold (fsub (fpext (fneg (fmul, x, y))), z)
7489 // -> (fma (fneg (fpext x)), (fpext y), (fneg z))
7490 if (N0.getOpcode() == ISD::FP_EXTEND) {
7491 SDValue N00 = N0.getOperand(0);
7492 if (N00.getOpcode() == ISD::FNEG) {
7493 SDValue N000 = N00.getOperand(0);
7494 if (N000.getOpcode() == ISD::FMUL) {
7495 return DAG.getNode(ISD::FMA, dl, VT,
7496 DAG.getNode(ISD::FNEG, dl, VT,
7497 DAG.getNode(ISD::FP_EXTEND, SDLoc(N),
7498 VT, N000.getOperand(0))),
7499 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7500 N000.getOperand(1)),
7501 DAG.getNode(ISD::FNEG, dl, VT, N1));
7506 // fold (fsub (fneg (fpext (fmul, x, y))), z)
7507 // -> (fma (fneg (fpext x)), (fpext y), (fneg z))
7508 if (N0.getOpcode() == ISD::FNEG) {
7509 SDValue N00 = N0.getOperand(0);
7510 if (N00.getOpcode() == ISD::FP_EXTEND) {
7511 SDValue N000 = N00.getOperand(0);
7512 if (N000.getOpcode() == ISD::FMUL) {
7513 return DAG.getNode(ISD::FMA, dl, VT,
7514 DAG.getNode(ISD::FNEG, dl, VT,
7515 DAG.getNode(ISD::FP_EXTEND, SDLoc(N),
7516 VT, N000.getOperand(0))),
7517 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7518 N000.getOperand(1)),
7519 DAG.getNode(ISD::FNEG, dl, VT, N1));
7529 SDValue DAGCombiner::visitFMUL(SDNode *N) {
7530 SDValue N0 = N->getOperand(0);
7531 SDValue N1 = N->getOperand(1);
7532 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
7533 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
7534 EVT VT = N->getValueType(0);
7535 const TargetOptions &Options = DAG.getTarget().Options;
7538 if (VT.isVector()) {
7539 // This just handles C1 * C2 for vectors. Other vector folds are below.
7540 if (SDValue FoldedVOp = SimplifyVBinOp(N))
7543 // Canonicalize vector constant to RHS.
7544 if (N0.getOpcode() == ISD::BUILD_VECTOR &&
7545 N1.getOpcode() != ISD::BUILD_VECTOR)
7546 if (auto *BV0 = dyn_cast<BuildVectorSDNode>(N0))
7547 if (BV0->isConstant())
7548 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, N1, N0);
7551 // fold (fmul c1, c2) -> c1*c2
7553 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, N1);
7555 // canonicalize constant to RHS
7556 if (N0CFP && !N1CFP)
7557 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, N0);
7559 // fold (fmul A, 1.0) -> A
7560 if (N1CFP && N1CFP->isExactlyValue(1.0))
7563 if (Options.UnsafeFPMath) {
7564 // fold (fmul A, 0) -> 0
7565 if (N1CFP && N1CFP->getValueAPF().isZero())
7568 // fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
7569 if (N0.getOpcode() == ISD::FMUL) {
7570 // Fold scalars or any vector constants (not just splats).
7571 // This fold is done in general by InstCombine, but extra fmul insts
7572 // may have been generated during lowering.
7573 SDValue N00 = N0.getOperand(0);
7574 SDValue N01 = N0.getOperand(1);
7575 auto *BV1 = dyn_cast<BuildVectorSDNode>(N1);
7576 auto *BV00 = dyn_cast<BuildVectorSDNode>(N00);
7577 auto *BV01 = dyn_cast<BuildVectorSDNode>(N01);
7579 // Check 1: Make sure that the first operand of the inner multiply is NOT
7580 // a constant. Otherwise, we may induce infinite looping.
7581 if (!(isConstOrConstSplatFP(N00) || (BV00 && BV00->isConstant()))) {
7582 // Check 2: Make sure that the second operand of the inner multiply and
7583 // the second operand of the outer multiply are constants.
7584 if ((N1CFP && isConstOrConstSplatFP(N01)) ||
7585 (BV1 && BV01 && BV1->isConstant() && BV01->isConstant())) {
7587 SDValue MulConsts = DAG.getNode(ISD::FMUL, SL, VT, N01, N1);
7588 return DAG.getNode(ISD::FMUL, SL, VT, N00, MulConsts);
7593 // fold (fmul (fadd x, x), c) -> (fmul x, (fmul 2.0, c))
7594 // Undo the fmul 2.0, x -> fadd x, x transformation, since if it occurs
7595 // during an early run of DAGCombiner can prevent folding with fmuls
7596 // inserted during lowering.
7597 if (N0.getOpcode() == ISD::FADD && N0.getOperand(0) == N0.getOperand(1)) {
7599 const SDValue Two = DAG.getConstantFP(2.0, VT);
7600 SDValue MulConsts = DAG.getNode(ISD::FMUL, SL, VT, Two, N1);
7601 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0.getOperand(0), MulConsts);
7605 // fold (fmul X, 2.0) -> (fadd X, X)
7606 if (N1CFP && N1CFP->isExactlyValue(+2.0))
7607 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N0);
7609 // fold (fmul X, -1.0) -> (fneg X)
7610 if (N1CFP && N1CFP->isExactlyValue(-1.0))
7611 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7612 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
7614 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
7615 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
7616 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
7617 // Both can be negated for free, check to see if at least one is cheaper
7619 if (LHSNeg == 2 || RHSNeg == 2)
7620 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7621 GetNegatedExpression(N0, DAG, LegalOperations),
7622 GetNegatedExpression(N1, DAG, LegalOperations));
7629 SDValue DAGCombiner::visitFMA(SDNode *N) {
7630 SDValue N0 = N->getOperand(0);
7631 SDValue N1 = N->getOperand(1);
7632 SDValue N2 = N->getOperand(2);
7633 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7634 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7635 EVT VT = N->getValueType(0);
7637 const TargetOptions &Options = DAG.getTarget().Options;
7639 // Constant fold FMA.
7640 if (isa<ConstantFPSDNode>(N0) &&
7641 isa<ConstantFPSDNode>(N1) &&
7642 isa<ConstantFPSDNode>(N2)) {
7643 return DAG.getNode(ISD::FMA, dl, VT, N0, N1, N2);
7646 if (Options.UnsafeFPMath) {
7647 if (N0CFP && N0CFP->isZero())
7649 if (N1CFP && N1CFP->isZero())
7652 if (N0CFP && N0CFP->isExactlyValue(1.0))
7653 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2);
7654 if (N1CFP && N1CFP->isExactlyValue(1.0))
7655 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2);
7657 // Canonicalize (fma c, x, y) -> (fma x, c, y)
7658 if (N0CFP && !N1CFP)
7659 return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2);
7661 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
7662 if (Options.UnsafeFPMath && N1CFP &&
7663 N2.getOpcode() == ISD::FMUL &&
7664 N0 == N2.getOperand(0) &&
7665 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
7666 return DAG.getNode(ISD::FMUL, dl, VT, N0,
7667 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
7671 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
7672 if (Options.UnsafeFPMath &&
7673 N0.getOpcode() == ISD::FMUL && N1CFP &&
7674 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
7675 return DAG.getNode(ISD::FMA, dl, VT,
7677 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
7681 // (fma x, 1, y) -> (fadd x, y)
7682 // (fma x, -1, y) -> (fadd (fneg x), y)
7684 if (N1CFP->isExactlyValue(1.0))
7685 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
7687 if (N1CFP->isExactlyValue(-1.0) &&
7688 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
7689 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
7690 AddToWorklist(RHSNeg.getNode());
7691 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
7695 // (fma x, c, x) -> (fmul x, (c+1))
7696 if (Options.UnsafeFPMath && N1CFP && N0 == N2)
7697 return DAG.getNode(ISD::FMUL, dl, VT, N0,
7698 DAG.getNode(ISD::FADD, dl, VT,
7699 N1, DAG.getConstantFP(1.0, VT)));
7701 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
7702 if (Options.UnsafeFPMath && N1CFP &&
7703 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0)
7704 return DAG.getNode(ISD::FMUL, dl, VT, N0,
7705 DAG.getNode(ISD::FADD, dl, VT,
7706 N1, DAG.getConstantFP(-1.0, VT)));
7712 SDValue DAGCombiner::visitFDIV(SDNode *N) {
7713 SDValue N0 = N->getOperand(0);
7714 SDValue N1 = N->getOperand(1);
7715 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7716 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7717 EVT VT = N->getValueType(0);
7719 const TargetOptions &Options = DAG.getTarget().Options;
7723 if (SDValue FoldedVOp = SimplifyVBinOp(N))
7726 // fold (fdiv c1, c2) -> c1/c2
7728 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1);
7730 if (Options.UnsafeFPMath) {
7731 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
7733 // Compute the reciprocal 1.0 / c2.
7734 APFloat N1APF = N1CFP->getValueAPF();
7735 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
7736 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
7737 // Only do the transform if the reciprocal is a legal fp immediate that
7738 // isn't too nasty (eg NaN, denormal, ...).
7739 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
7740 (!LegalOperations ||
7741 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
7742 // backend)... we should handle this gracefully after Legalize.
7743 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
7744 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
7745 TLI.isFPImmLegal(Recip, VT)))
7746 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0,
7747 DAG.getConstantFP(Recip, VT));
7750 // If this FDIV is part of a reciprocal square root, it may be folded
7751 // into a target-specific square root estimate instruction.
7752 if (N1.getOpcode() == ISD::FSQRT) {
7753 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0))) {
7754 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7756 } else if (N1.getOpcode() == ISD::FP_EXTEND &&
7757 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
7758 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0).getOperand(0))) {
7759 RV = DAG.getNode(ISD::FP_EXTEND, SDLoc(N1), VT, RV);
7760 AddToWorklist(RV.getNode());
7761 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7763 } else if (N1.getOpcode() == ISD::FP_ROUND &&
7764 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
7765 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0).getOperand(0))) {
7766 RV = DAG.getNode(ISD::FP_ROUND, SDLoc(N1), VT, RV, N1.getOperand(1));
7767 AddToWorklist(RV.getNode());
7768 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7770 } else if (N1.getOpcode() == ISD::FMUL) {
7771 // Look through an FMUL. Even though this won't remove the FDIV directly,
7772 // it's still worthwhile to get rid of the FSQRT if possible.
7775 if (N1.getOperand(0).getOpcode() == ISD::FSQRT) {
7776 SqrtOp = N1.getOperand(0);
7777 OtherOp = N1.getOperand(1);
7778 } else if (N1.getOperand(1).getOpcode() == ISD::FSQRT) {
7779 SqrtOp = N1.getOperand(1);
7780 OtherOp = N1.getOperand(0);
7782 if (SqrtOp.getNode()) {
7783 // We found a FSQRT, so try to make this fold:
7784 // x / (y * sqrt(z)) -> x * (rsqrt(z) / y)
7785 if (SDValue RV = BuildRsqrtEstimate(SqrtOp.getOperand(0))) {
7786 RV = DAG.getNode(ISD::FDIV, SDLoc(N1), VT, RV, OtherOp);
7787 AddToWorklist(RV.getNode());
7788 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7793 // Fold into a reciprocal estimate and multiply instead of a real divide.
7794 if (SDValue RV = BuildReciprocalEstimate(N1)) {
7795 AddToWorklist(RV.getNode());
7796 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7800 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
7801 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
7802 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
7803 // Both can be negated for free, check to see if at least one is cheaper
7805 if (LHSNeg == 2 || RHSNeg == 2)
7806 return DAG.getNode(ISD::FDIV, SDLoc(N), VT,
7807 GetNegatedExpression(N0, DAG, LegalOperations),
7808 GetNegatedExpression(N1, DAG, LegalOperations));
7812 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
7814 // E.g., (a / D; b / D;) -> (recip = 1.0 / D; a * recip; b * recip)
7815 // Notice that this is not always beneficial. One reason is different target
7816 // may have different costs for FDIV and FMUL, so sometimes the cost of two
7817 // FDIVs may be lower than the cost of one FDIV and two FMULs. Another reason
7818 // is the critical path is increased from "one FDIV" to "one FDIV + one FMUL".
7819 if (Options.UnsafeFPMath) {
7820 // Skip if current node is a reciprocal.
7821 if (N0CFP && N0CFP->isExactlyValue(1.0))
7824 SmallVector<SDNode *, 4> Users;
7825 // Find all FDIV users of the same divisor.
7826 for (SDNode::use_iterator UI = N1.getNode()->use_begin(),
7827 UE = N1.getNode()->use_end();
7829 SDNode *User = UI.getUse().getUser();
7830 if (User->getOpcode() == ISD::FDIV && User->getOperand(1) == N1)
7831 Users.push_back(User);
7834 if (TLI.combineRepeatedFPDivisors(Users.size())) {
7835 SDValue FPOne = DAG.getConstantFP(1.0, VT); // floating point 1.0
7836 SDValue Reciprocal = DAG.getNode(ISD::FDIV, SDLoc(N), VT, FPOne, N1);
7838 // Dividend / Divisor -> Dividend * Reciprocal
7839 for (auto I = Users.begin(), E = Users.end(); I != E; ++I) {
7840 if ((*I)->getOperand(0) != FPOne) {
7841 SDValue NewNode = DAG.getNode(ISD::FMUL, SDLoc(*I), VT,
7842 (*I)->getOperand(0), Reciprocal);
7843 DAG.ReplaceAllUsesWith(*I, NewNode.getNode());
7853 SDValue DAGCombiner::visitFREM(SDNode *N) {
7854 SDValue N0 = N->getOperand(0);
7855 SDValue N1 = N->getOperand(1);
7856 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7857 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7858 EVT VT = N->getValueType(0);
7860 // fold (frem c1, c2) -> fmod(c1,c2)
7862 return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1);
7867 SDValue DAGCombiner::visitFSQRT(SDNode *N) {
7868 if (DAG.getTarget().Options.UnsafeFPMath &&
7869 !TLI.isFsqrtCheap()) {
7870 // Compute this as X * (1/sqrt(X)) = X * (X ** -0.5)
7871 if (SDValue RV = BuildRsqrtEstimate(N->getOperand(0))) {
7872 EVT VT = RV.getValueType();
7873 RV = DAG.getNode(ISD::FMUL, SDLoc(N), VT, N->getOperand(0), RV);
7874 AddToWorklist(RV.getNode());
7876 // Unfortunately, RV is now NaN if the input was exactly 0.
7877 // Select out this case and force the answer to 0.
7878 SDValue Zero = DAG.getConstantFP(0.0, VT);
7880 DAG.getSetCC(SDLoc(N), TLI.getSetCCResultType(*DAG.getContext(), VT),
7881 N->getOperand(0), Zero, ISD::SETEQ);
7882 AddToWorklist(ZeroCmp.getNode());
7883 AddToWorklist(RV.getNode());
7885 RV = DAG.getNode(VT.isVector() ? ISD::VSELECT : ISD::SELECT,
7886 SDLoc(N), VT, ZeroCmp, Zero, RV);
7893 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
7894 SDValue N0 = N->getOperand(0);
7895 SDValue N1 = N->getOperand(1);
7896 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7897 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7898 EVT VT = N->getValueType(0);
7900 if (N0CFP && N1CFP) // Constant fold
7901 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1);
7904 const APFloat& V = N1CFP->getValueAPF();
7905 // copysign(x, c1) -> fabs(x) iff ispos(c1)
7906 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
7907 if (!V.isNegative()) {
7908 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
7909 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7911 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7912 return DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7913 DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0));
7917 // copysign(fabs(x), y) -> copysign(x, y)
7918 // copysign(fneg(x), y) -> copysign(x, y)
7919 // copysign(copysign(x,z), y) -> copysign(x, y)
7920 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
7921 N0.getOpcode() == ISD::FCOPYSIGN)
7922 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7923 N0.getOperand(0), N1);
7925 // copysign(x, abs(y)) -> abs(x)
7926 if (N1.getOpcode() == ISD::FABS)
7927 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7929 // copysign(x, copysign(y,z)) -> copysign(x, z)
7930 if (N1.getOpcode() == ISD::FCOPYSIGN)
7931 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7932 N0, N1.getOperand(1));
7934 // copysign(x, fp_extend(y)) -> copysign(x, y)
7935 // copysign(x, fp_round(y)) -> copysign(x, y)
7936 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
7937 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7938 N0, N1.getOperand(0));
7943 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
7944 SDValue N0 = N->getOperand(0);
7945 EVT VT = N->getValueType(0);
7946 EVT OpVT = N0.getValueType();
7948 // fold (sint_to_fp c1) -> c1fp
7949 if (isConstantIntBuildVectorOrConstantInt(N0) &&
7950 // ...but only if the target supports immediate floating-point values
7951 (!LegalOperations ||
7952 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
7953 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
7955 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
7956 // but UINT_TO_FP is legal on this target, try to convert.
7957 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
7958 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
7959 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
7960 if (DAG.SignBitIsZero(N0))
7961 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
7964 // The next optimizations are desirable only if SELECT_CC can be lowered.
7965 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
7966 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7967 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
7969 (!LegalOperations ||
7970 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7972 { N0.getOperand(0), N0.getOperand(1),
7973 DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT),
7975 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7978 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
7979 // (select_cc x, y, 1.0, 0.0,, cc)
7980 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
7981 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
7982 (!LegalOperations ||
7983 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7985 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
7986 DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT),
7987 N0.getOperand(0).getOperand(2) };
7988 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7995 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
7996 SDValue N0 = N->getOperand(0);
7997 EVT VT = N->getValueType(0);
7998 EVT OpVT = N0.getValueType();
8000 // fold (uint_to_fp c1) -> c1fp
8001 if (isConstantIntBuildVectorOrConstantInt(N0) &&
8002 // ...but only if the target supports immediate floating-point values
8003 (!LegalOperations ||
8004 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
8005 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
8007 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
8008 // but SINT_TO_FP is legal on this target, try to convert.
8009 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
8010 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
8011 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
8012 if (DAG.SignBitIsZero(N0))
8013 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
8016 // The next optimizations are desirable only if SELECT_CC can be lowered.
8017 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
8018 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
8020 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
8021 (!LegalOperations ||
8022 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
8024 { N0.getOperand(0), N0.getOperand(1),
8025 DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT),
8027 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
8034 // Fold (fp_to_{s/u}int ({s/u}int_to_fpx)) -> zext x, sext x, trunc x, or x
8035 static SDValue FoldIntToFPToInt(SDNode *N, SelectionDAG &DAG) {
8036 SDValue N0 = N->getOperand(0);
8037 EVT VT = N->getValueType(0);
8039 if (N0.getOpcode() != ISD::UINT_TO_FP && N0.getOpcode() != ISD::SINT_TO_FP)
8042 SDValue Src = N0.getOperand(0);
8043 EVT SrcVT = Src.getValueType();
8044 bool IsInputSigned = N0.getOpcode() == ISD::SINT_TO_FP;
8045 bool IsOutputSigned = N->getOpcode() == ISD::FP_TO_SINT;
8047 // We can safely assume the conversion won't overflow the output range,
8048 // because (for example) (uint8_t)18293.f is undefined behavior.
8050 // Since we can assume the conversion won't overflow, our decision as to
8051 // whether the input will fit in the float should depend on the minimum
8052 // of the input range and output range.
8054 // This means this is also safe for a signed input and unsigned output, since
8055 // a negative input would lead to undefined behavior.
8056 unsigned InputSize = (int)SrcVT.getScalarSizeInBits() - IsInputSigned;
8057 unsigned OutputSize = (int)VT.getScalarSizeInBits() - IsOutputSigned;
8058 unsigned ActualSize = std::min(InputSize, OutputSize);
8059 const fltSemantics &sem = DAG.EVTToAPFloatSemantics(N0.getValueType());
8061 // We can only fold away the float conversion if the input range can be
8062 // represented exactly in the float range.
8063 if (APFloat::semanticsPrecision(sem) >= ActualSize) {
8064 if (VT.getScalarSizeInBits() > SrcVT.getScalarSizeInBits()) {
8065 unsigned ExtOp = IsInputSigned && IsOutputSigned ? ISD::SIGN_EXTEND
8067 return DAG.getNode(ExtOp, SDLoc(N), VT, Src);
8069 if (VT.getScalarSizeInBits() < SrcVT.getScalarSizeInBits())
8070 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Src);
8073 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Src);
8078 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
8079 SDValue N0 = N->getOperand(0);
8080 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8081 EVT VT = N->getValueType(0);
8083 // fold (fp_to_sint c1fp) -> c1
8085 return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0);
8087 return FoldIntToFPToInt(N, DAG);
8090 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
8091 SDValue N0 = N->getOperand(0);
8092 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8093 EVT VT = N->getValueType(0);
8095 // fold (fp_to_uint c1fp) -> c1
8097 return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0);
8099 return FoldIntToFPToInt(N, DAG);
8102 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
8103 SDValue N0 = N->getOperand(0);
8104 SDValue N1 = N->getOperand(1);
8105 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8106 EVT VT = N->getValueType(0);
8108 // fold (fp_round c1fp) -> c1fp
8110 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1);
8112 // fold (fp_round (fp_extend x)) -> x
8113 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
8114 return N0.getOperand(0);
8116 // fold (fp_round (fp_round x)) -> (fp_round x)
8117 if (N0.getOpcode() == ISD::FP_ROUND) {
8118 const bool NIsTrunc = N->getConstantOperandVal(1) == 1;
8119 const bool N0IsTrunc = N0.getNode()->getConstantOperandVal(1) == 1;
8120 // If the first fp_round isn't a value preserving truncation, it might
8121 // introduce a tie in the second fp_round, that wouldn't occur in the
8122 // single-step fp_round we want to fold to.
8123 // In other words, double rounding isn't the same as rounding.
8124 // Also, this is a value preserving truncation iff both fp_round's are.
8125 if (DAG.getTarget().Options.UnsafeFPMath || N0IsTrunc)
8126 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0.getOperand(0),
8127 DAG.getIntPtrConstant(NIsTrunc && N0IsTrunc));
8130 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
8131 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
8132 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT,
8133 N0.getOperand(0), N1);
8134 AddToWorklist(Tmp.getNode());
8135 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
8136 Tmp, N0.getOperand(1));
8142 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
8143 SDValue N0 = N->getOperand(0);
8144 EVT VT = N->getValueType(0);
8145 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
8146 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8148 // fold (fp_round_inreg c1fp) -> c1fp
8149 if (N0CFP && isTypeLegal(EVT)) {
8150 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
8151 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, Round);
8157 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
8158 SDValue N0 = N->getOperand(0);
8159 EVT VT = N->getValueType(0);
8161 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
8162 if (N->hasOneUse() &&
8163 N->use_begin()->getOpcode() == ISD::FP_ROUND)
8166 // fold (fp_extend c1fp) -> c1fp
8167 if (isConstantFPBuildVectorOrConstantFP(N0))
8168 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0);
8170 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
8172 if (N0.getOpcode() == ISD::FP_ROUND
8173 && N0.getNode()->getConstantOperandVal(1) == 1) {
8174 SDValue In = N0.getOperand(0);
8175 if (In.getValueType() == VT) return In;
8176 if (VT.bitsLT(In.getValueType()))
8177 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT,
8178 In, N0.getOperand(1));
8179 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In);
8182 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
8183 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
8184 TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) {
8185 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
8186 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
8188 LN0->getBasePtr(), N0.getValueType(),
8189 LN0->getMemOperand());
8190 CombineTo(N, ExtLoad);
8191 CombineTo(N0.getNode(),
8192 DAG.getNode(ISD::FP_ROUND, SDLoc(N0),
8193 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
8194 ExtLoad.getValue(1));
8195 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8201 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
8202 SDValue N0 = N->getOperand(0);
8203 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8204 EVT VT = N->getValueType(0);
8206 // fold (fceil c1) -> fceil(c1)
8208 return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0);
8213 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
8214 SDValue N0 = N->getOperand(0);
8215 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8216 EVT VT = N->getValueType(0);
8218 // fold (ftrunc c1) -> ftrunc(c1)
8220 return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0);
8225 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
8226 SDValue N0 = N->getOperand(0);
8227 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8228 EVT VT = N->getValueType(0);
8230 // fold (ffloor c1) -> ffloor(c1)
8232 return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0);
8237 // FIXME: FNEG and FABS have a lot in common; refactor.
8238 SDValue DAGCombiner::visitFNEG(SDNode *N) {
8239 SDValue N0 = N->getOperand(0);
8240 EVT VT = N->getValueType(0);
8242 // Constant fold FNEG.
8243 if (isConstantFPBuildVectorOrConstantFP(N0))
8244 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
8246 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
8247 &DAG.getTarget().Options))
8248 return GetNegatedExpression(N0, DAG, LegalOperations);
8250 // Transform fneg(bitconvert(x)) -> bitconvert(x ^ sign) to avoid loading
8251 // constant pool values.
8252 if (!TLI.isFNegFree(VT) &&
8253 N0.getOpcode() == ISD::BITCAST &&
8254 N0.getNode()->hasOneUse()) {
8255 SDValue Int = N0.getOperand(0);
8256 EVT IntVT = Int.getValueType();
8257 if (IntVT.isInteger() && !IntVT.isVector()) {
8259 if (N0.getValueType().isVector()) {
8260 // For a vector, get a mask such as 0x80... per scalar element
8262 SignMask = APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
8263 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
8265 // For a scalar, just generate 0x80...
8266 SignMask = APInt::getSignBit(IntVT.getSizeInBits());
8268 Int = DAG.getNode(ISD::XOR, SDLoc(N0), IntVT, Int,
8269 DAG.getConstant(SignMask, IntVT));
8270 AddToWorklist(Int.getNode());
8271 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Int);
8275 // (fneg (fmul c, x)) -> (fmul -c, x)
8276 if (N0.getOpcode() == ISD::FMUL) {
8277 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
8279 APFloat CVal = CFP1->getValueAPF();
8281 if (Level >= AfterLegalizeDAG &&
8282 (TLI.isFPImmLegal(CVal, N->getValueType(0)) ||
8283 TLI.isOperationLegal(ISD::ConstantFP, N->getValueType(0))))
8285 ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
8286 DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0.getOperand(1)));
8293 SDValue DAGCombiner::visitFMINNUM(SDNode *N) {
8294 SDValue N0 = N->getOperand(0);
8295 SDValue N1 = N->getOperand(1);
8296 const ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8297 const ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8299 if (N0CFP && N1CFP) {
8300 const APFloat &C0 = N0CFP->getValueAPF();
8301 const APFloat &C1 = N1CFP->getValueAPF();
8302 return DAG.getConstantFP(minnum(C0, C1), N->getValueType(0));
8306 EVT VT = N->getValueType(0);
8307 // Canonicalize to constant on RHS.
8308 return DAG.getNode(ISD::FMINNUM, SDLoc(N), VT, N1, N0);
8314 SDValue DAGCombiner::visitFMAXNUM(SDNode *N) {
8315 SDValue N0 = N->getOperand(0);
8316 SDValue N1 = N->getOperand(1);
8317 const ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8318 const ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8320 if (N0CFP && N1CFP) {
8321 const APFloat &C0 = N0CFP->getValueAPF();
8322 const APFloat &C1 = N1CFP->getValueAPF();
8323 return DAG.getConstantFP(maxnum(C0, C1), N->getValueType(0));
8327 EVT VT = N->getValueType(0);
8328 // Canonicalize to constant on RHS.
8329 return DAG.getNode(ISD::FMAXNUM, SDLoc(N), VT, N1, N0);
8335 SDValue DAGCombiner::visitFABS(SDNode *N) {
8336 SDValue N0 = N->getOperand(0);
8337 EVT VT = N->getValueType(0);
8339 // fold (fabs c1) -> fabs(c1)
8340 if (isConstantFPBuildVectorOrConstantFP(N0))
8341 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
8343 // fold (fabs (fabs x)) -> (fabs x)
8344 if (N0.getOpcode() == ISD::FABS)
8345 return N->getOperand(0);
8347 // fold (fabs (fneg x)) -> (fabs x)
8348 // fold (fabs (fcopysign x, y)) -> (fabs x)
8349 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
8350 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0));
8352 // Transform fabs(bitconvert(x)) -> bitconvert(x & ~sign) to avoid loading
8353 // constant pool values.
8354 if (!TLI.isFAbsFree(VT) &&
8355 N0.getOpcode() == ISD::BITCAST &&
8356 N0.getNode()->hasOneUse()) {
8357 SDValue Int = N0.getOperand(0);
8358 EVT IntVT = Int.getValueType();
8359 if (IntVT.isInteger() && !IntVT.isVector()) {
8361 if (N0.getValueType().isVector()) {
8362 // For a vector, get a mask such as 0x7f... per scalar element
8364 SignMask = ~APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
8365 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
8367 // For a scalar, just generate 0x7f...
8368 SignMask = ~APInt::getSignBit(IntVT.getSizeInBits());
8370 Int = DAG.getNode(ISD::AND, SDLoc(N0), IntVT, Int,
8371 DAG.getConstant(SignMask, IntVT));
8372 AddToWorklist(Int.getNode());
8373 return DAG.getNode(ISD::BITCAST, SDLoc(N), N->getValueType(0), Int);
8380 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
8381 SDValue Chain = N->getOperand(0);
8382 SDValue N1 = N->getOperand(1);
8383 SDValue N2 = N->getOperand(2);
8385 // If N is a constant we could fold this into a fallthrough or unconditional
8386 // branch. However that doesn't happen very often in normal code, because
8387 // Instcombine/SimplifyCFG should have handled the available opportunities.
8388 // If we did this folding here, it would be necessary to update the
8389 // MachineBasicBlock CFG, which is awkward.
8391 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
8393 if (N1.getOpcode() == ISD::SETCC &&
8394 TLI.isOperationLegalOrCustom(ISD::BR_CC,
8395 N1.getOperand(0).getValueType())) {
8396 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
8397 Chain, N1.getOperand(2),
8398 N1.getOperand(0), N1.getOperand(1), N2);
8401 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
8402 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
8403 (N1.getOperand(0).hasOneUse() &&
8404 N1.getOperand(0).getOpcode() == ISD::SRL))) {
8405 SDNode *Trunc = nullptr;
8406 if (N1.getOpcode() == ISD::TRUNCATE) {
8407 // Look pass the truncate.
8408 Trunc = N1.getNode();
8409 N1 = N1.getOperand(0);
8412 // Match this pattern so that we can generate simpler code:
8415 // %b = and i32 %a, 2
8416 // %c = srl i32 %b, 1
8417 // brcond i32 %c ...
8422 // %b = and i32 %a, 2
8423 // %c = setcc eq %b, 0
8426 // This applies only when the AND constant value has one bit set and the
8427 // SRL constant is equal to the log2 of the AND constant. The back-end is
8428 // smart enough to convert the result into a TEST/JMP sequence.
8429 SDValue Op0 = N1.getOperand(0);
8430 SDValue Op1 = N1.getOperand(1);
8432 if (Op0.getOpcode() == ISD::AND &&
8433 Op1.getOpcode() == ISD::Constant) {
8434 SDValue AndOp1 = Op0.getOperand(1);
8436 if (AndOp1.getOpcode() == ISD::Constant) {
8437 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
8439 if (AndConst.isPowerOf2() &&
8440 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
8442 DAG.getSetCC(SDLoc(N),
8443 getSetCCResultType(Op0.getValueType()),
8444 Op0, DAG.getConstant(0, Op0.getValueType()),
8447 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, SDLoc(N),
8448 MVT::Other, Chain, SetCC, N2);
8449 // Don't add the new BRCond into the worklist or else SimplifySelectCC
8450 // will convert it back to (X & C1) >> C2.
8451 CombineTo(N, NewBRCond, false);
8452 // Truncate is dead.
8454 deleteAndRecombine(Trunc);
8455 // Replace the uses of SRL with SETCC
8456 WorklistRemover DeadNodes(*this);
8457 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
8458 deleteAndRecombine(N1.getNode());
8459 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8465 // Restore N1 if the above transformation doesn't match.
8466 N1 = N->getOperand(1);
8469 // Transform br(xor(x, y)) -> br(x != y)
8470 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
8471 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
8472 SDNode *TheXor = N1.getNode();
8473 SDValue Op0 = TheXor->getOperand(0);
8474 SDValue Op1 = TheXor->getOperand(1);
8475 if (Op0.getOpcode() == Op1.getOpcode()) {
8476 // Avoid missing important xor optimizations.
8477 SDValue Tmp = visitXOR(TheXor);
8478 if (Tmp.getNode()) {
8479 if (Tmp.getNode() != TheXor) {
8480 DEBUG(dbgs() << "\nReplacing.8 ";
8482 dbgs() << "\nWith: ";
8483 Tmp.getNode()->dump(&DAG);
8485 WorklistRemover DeadNodes(*this);
8486 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
8487 deleteAndRecombine(TheXor);
8488 return DAG.getNode(ISD::BRCOND, SDLoc(N),
8489 MVT::Other, Chain, Tmp, N2);
8492 // visitXOR has changed XOR's operands or replaced the XOR completely,
8494 return SDValue(N, 0);
8498 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
8500 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
8501 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
8502 Op0.getOpcode() == ISD::XOR) {
8503 TheXor = Op0.getNode();
8507 EVT SetCCVT = N1.getValueType();
8509 SetCCVT = getSetCCResultType(SetCCVT);
8510 SDValue SetCC = DAG.getSetCC(SDLoc(TheXor),
8513 Equal ? ISD::SETEQ : ISD::SETNE);
8514 // Replace the uses of XOR with SETCC
8515 WorklistRemover DeadNodes(*this);
8516 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
8517 deleteAndRecombine(N1.getNode());
8518 return DAG.getNode(ISD::BRCOND, SDLoc(N),
8519 MVT::Other, Chain, SetCC, N2);
8526 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
8528 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
8529 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
8530 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
8532 // If N is a constant we could fold this into a fallthrough or unconditional
8533 // branch. However that doesn't happen very often in normal code, because
8534 // Instcombine/SimplifyCFG should have handled the available opportunities.
8535 // If we did this folding here, it would be necessary to update the
8536 // MachineBasicBlock CFG, which is awkward.
8538 // Use SimplifySetCC to simplify SETCC's.
8539 SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()),
8540 CondLHS, CondRHS, CC->get(), SDLoc(N),
8542 if (Simp.getNode()) AddToWorklist(Simp.getNode());
8544 // fold to a simpler setcc
8545 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
8546 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
8547 N->getOperand(0), Simp.getOperand(2),
8548 Simp.getOperand(0), Simp.getOperand(1),
8554 /// Return true if 'Use' is a load or a store that uses N as its base pointer
8555 /// and that N may be folded in the load / store addressing mode.
8556 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
8558 const TargetLowering &TLI) {
8560 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
8561 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
8563 VT = Use->getValueType(0);
8564 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
8565 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
8567 VT = ST->getValue().getValueType();
8571 TargetLowering::AddrMode AM;
8572 if (N->getOpcode() == ISD::ADD) {
8573 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
8576 AM.BaseOffs = Offset->getSExtValue();
8580 } else if (N->getOpcode() == ISD::SUB) {
8581 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
8584 AM.BaseOffs = -Offset->getSExtValue();
8591 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
8594 /// Try turning a load/store into a pre-indexed load/store when the base
8595 /// pointer is an add or subtract and it has other uses besides the load/store.
8596 /// After the transformation, the new indexed load/store has effectively folded
8597 /// the add/subtract in and all of its other uses are redirected to the
8599 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
8600 if (Level < AfterLegalizeDAG)
8606 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8607 if (LD->isIndexed())
8609 VT = LD->getMemoryVT();
8610 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
8611 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
8613 Ptr = LD->getBasePtr();
8614 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8615 if (ST->isIndexed())
8617 VT = ST->getMemoryVT();
8618 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
8619 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
8621 Ptr = ST->getBasePtr();
8627 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
8628 // out. There is no reason to make this a preinc/predec.
8629 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
8630 Ptr.getNode()->hasOneUse())
8633 // Ask the target to do addressing mode selection.
8636 ISD::MemIndexedMode AM = ISD::UNINDEXED;
8637 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
8640 // Backends without true r+i pre-indexed forms may need to pass a
8641 // constant base with a variable offset so that constant coercion
8642 // will work with the patterns in canonical form.
8643 bool Swapped = false;
8644 if (isa<ConstantSDNode>(BasePtr)) {
8645 std::swap(BasePtr, Offset);
8649 // Don't create a indexed load / store with zero offset.
8650 if (isa<ConstantSDNode>(Offset) &&
8651 cast<ConstantSDNode>(Offset)->isNullValue())
8654 // Try turning it into a pre-indexed load / store except when:
8655 // 1) The new base ptr is a frame index.
8656 // 2) If N is a store and the new base ptr is either the same as or is a
8657 // predecessor of the value being stored.
8658 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
8659 // that would create a cycle.
8660 // 4) All uses are load / store ops that use it as old base ptr.
8662 // Check #1. Preinc'ing a frame index would require copying the stack pointer
8663 // (plus the implicit offset) to a register to preinc anyway.
8664 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
8669 SDValue Val = cast<StoreSDNode>(N)->getValue();
8670 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
8674 // If the offset is a constant, there may be other adds of constants that
8675 // can be folded with this one. We should do this to avoid having to keep
8676 // a copy of the original base pointer.
8677 SmallVector<SDNode *, 16> OtherUses;
8678 if (isa<ConstantSDNode>(Offset))
8679 for (SDNode *Use : BasePtr.getNode()->uses()) {
8680 if (Use == Ptr.getNode())
8683 if (Use->isPredecessorOf(N))
8686 if (Use->getOpcode() != ISD::ADD && Use->getOpcode() != ISD::SUB) {
8691 SDValue Op0 = Use->getOperand(0), Op1 = Use->getOperand(1);
8692 if (Op1.getNode() == BasePtr.getNode())
8693 std::swap(Op0, Op1);
8694 assert(Op0.getNode() == BasePtr.getNode() &&
8695 "Use of ADD/SUB but not an operand");
8697 if (!isa<ConstantSDNode>(Op1)) {
8702 // FIXME: In some cases, we can be smarter about this.
8703 if (Op1.getValueType() != Offset.getValueType()) {
8708 OtherUses.push_back(Use);
8712 std::swap(BasePtr, Offset);
8714 // Now check for #3 and #4.
8715 bool RealUse = false;
8717 // Caches for hasPredecessorHelper
8718 SmallPtrSet<const SDNode *, 32> Visited;
8719 SmallVector<const SDNode *, 16> Worklist;
8721 for (SDNode *Use : Ptr.getNode()->uses()) {
8724 if (N->hasPredecessorHelper(Use, Visited, Worklist))
8727 // If Ptr may be folded in addressing mode of other use, then it's
8728 // not profitable to do this transformation.
8729 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
8738 Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
8739 BasePtr, Offset, AM);
8741 Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
8742 BasePtr, Offset, AM);
8745 DEBUG(dbgs() << "\nReplacing.4 ";
8747 dbgs() << "\nWith: ";
8748 Result.getNode()->dump(&DAG);
8750 WorklistRemover DeadNodes(*this);
8752 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
8753 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
8755 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
8758 // Finally, since the node is now dead, remove it from the graph.
8759 deleteAndRecombine(N);
8762 std::swap(BasePtr, Offset);
8764 // Replace other uses of BasePtr that can be updated to use Ptr
8765 for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) {
8766 unsigned OffsetIdx = 1;
8767 if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode())
8769 assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() ==
8770 BasePtr.getNode() && "Expected BasePtr operand");
8772 // We need to replace ptr0 in the following expression:
8773 // x0 * offset0 + y0 * ptr0 = t0
8775 // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store)
8777 // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the
8778 // indexed load/store and the expresion that needs to be re-written.
8780 // Therefore, we have:
8781 // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1
8783 ConstantSDNode *CN =
8784 cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx));
8786 APInt Offset0 = CN->getAPIntValue();
8787 APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue();
8789 X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1;
8790 Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1;
8791 X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1;
8792 Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1;
8794 unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD;
8796 APInt CNV = Offset0;
8797 if (X0 < 0) CNV = -CNV;
8798 if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1;
8799 else CNV = CNV - Offset1;
8801 // We can now generate the new expression.
8802 SDValue NewOp1 = DAG.getConstant(CNV, CN->getValueType(0));
8803 SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0);
8805 SDValue NewUse = DAG.getNode(Opcode,
8806 SDLoc(OtherUses[i]),
8807 OtherUses[i]->getValueType(0), NewOp1, NewOp2);
8808 DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse);
8809 deleteAndRecombine(OtherUses[i]);
8812 // Replace the uses of Ptr with uses of the updated base value.
8813 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
8814 deleteAndRecombine(Ptr.getNode());
8819 /// Try to combine a load/store with a add/sub of the base pointer node into a
8820 /// post-indexed load/store. The transformation folded the add/subtract into the
8821 /// new indexed load/store effectively and all of its uses are redirected to the
8823 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
8824 if (Level < AfterLegalizeDAG)
8830 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8831 if (LD->isIndexed())
8833 VT = LD->getMemoryVT();
8834 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
8835 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
8837 Ptr = LD->getBasePtr();
8838 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8839 if (ST->isIndexed())
8841 VT = ST->getMemoryVT();
8842 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
8843 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
8845 Ptr = ST->getBasePtr();
8851 if (Ptr.getNode()->hasOneUse())
8854 for (SDNode *Op : Ptr.getNode()->uses()) {
8856 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
8861 ISD::MemIndexedMode AM = ISD::UNINDEXED;
8862 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
8863 // Don't create a indexed load / store with zero offset.
8864 if (isa<ConstantSDNode>(Offset) &&
8865 cast<ConstantSDNode>(Offset)->isNullValue())
8868 // Try turning it into a post-indexed load / store except when
8869 // 1) All uses are load / store ops that use it as base ptr (and
8870 // it may be folded as addressing mmode).
8871 // 2) Op must be independent of N, i.e. Op is neither a predecessor
8872 // nor a successor of N. Otherwise, if Op is folded that would
8875 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
8879 bool TryNext = false;
8880 for (SDNode *Use : BasePtr.getNode()->uses()) {
8881 if (Use == Ptr.getNode())
8884 // If all the uses are load / store addresses, then don't do the
8886 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
8887 bool RealUse = false;
8888 for (SDNode *UseUse : Use->uses()) {
8889 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
8904 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
8905 SDValue Result = isLoad
8906 ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
8907 BasePtr, Offset, AM)
8908 : DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
8909 BasePtr, Offset, AM);
8912 DEBUG(dbgs() << "\nReplacing.5 ";
8914 dbgs() << "\nWith: ";
8915 Result.getNode()->dump(&DAG);
8917 WorklistRemover DeadNodes(*this);
8919 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
8920 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
8922 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
8925 // Finally, since the node is now dead, remove it from the graph.
8926 deleteAndRecombine(N);
8928 // Replace the uses of Use with uses of the updated base value.
8929 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
8930 Result.getValue(isLoad ? 1 : 0));
8931 deleteAndRecombine(Op);
8940 /// \brief Return the base-pointer arithmetic from an indexed \p LD.
8941 SDValue DAGCombiner::SplitIndexingFromLoad(LoadSDNode *LD) {
8942 ISD::MemIndexedMode AM = LD->getAddressingMode();
8943 assert(AM != ISD::UNINDEXED);
8944 SDValue BP = LD->getOperand(1);
8945 SDValue Inc = LD->getOperand(2);
8947 // Some backends use TargetConstants for load offsets, but don't expect
8948 // TargetConstants in general ADD nodes. We can convert these constants into
8949 // regular Constants (if the constant is not opaque).
8950 assert((Inc.getOpcode() != ISD::TargetConstant ||
8951 !cast<ConstantSDNode>(Inc)->isOpaque()) &&
8952 "Cannot split out indexing using opaque target constants");
8953 if (Inc.getOpcode() == ISD::TargetConstant) {
8954 ConstantSDNode *ConstInc = cast<ConstantSDNode>(Inc);
8955 Inc = DAG.getConstant(*ConstInc->getConstantIntValue(),
8956 ConstInc->getValueType(0));
8960 (AM == ISD::PRE_INC || AM == ISD::POST_INC ? ISD::ADD : ISD::SUB);
8961 return DAG.getNode(Opc, SDLoc(LD), BP.getSimpleValueType(), BP, Inc);
8964 SDValue DAGCombiner::visitLOAD(SDNode *N) {
8965 LoadSDNode *LD = cast<LoadSDNode>(N);
8966 SDValue Chain = LD->getChain();
8967 SDValue Ptr = LD->getBasePtr();
8969 // If load is not volatile and there are no uses of the loaded value (and
8970 // the updated indexed value in case of indexed loads), change uses of the
8971 // chain value into uses of the chain input (i.e. delete the dead load).
8972 if (!LD->isVolatile()) {
8973 if (N->getValueType(1) == MVT::Other) {
8975 if (!N->hasAnyUseOfValue(0)) {
8976 // It's not safe to use the two value CombineTo variant here. e.g.
8977 // v1, chain2 = load chain1, loc
8978 // v2, chain3 = load chain2, loc
8980 // Now we replace use of chain2 with chain1. This makes the second load
8981 // isomorphic to the one we are deleting, and thus makes this load live.
8982 DEBUG(dbgs() << "\nReplacing.6 ";
8984 dbgs() << "\nWith chain: ";
8985 Chain.getNode()->dump(&DAG);
8987 WorklistRemover DeadNodes(*this);
8988 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
8991 deleteAndRecombine(N);
8993 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8997 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
8999 // If this load has an opaque TargetConstant offset, then we cannot split
9000 // the indexing into an add/sub directly (that TargetConstant may not be
9001 // valid for a different type of node, and we cannot convert an opaque
9002 // target constant into a regular constant).
9003 bool HasOTCInc = LD->getOperand(2).getOpcode() == ISD::TargetConstant &&
9004 cast<ConstantSDNode>(LD->getOperand(2))->isOpaque();
9006 if (!N->hasAnyUseOfValue(0) &&
9007 ((MaySplitLoadIndex && !HasOTCInc) || !N->hasAnyUseOfValue(1))) {
9008 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
9010 if (N->hasAnyUseOfValue(1) && MaySplitLoadIndex && !HasOTCInc) {
9011 Index = SplitIndexingFromLoad(LD);
9012 // Try to fold the base pointer arithmetic into subsequent loads and
9014 AddUsersToWorklist(N);
9016 Index = DAG.getUNDEF(N->getValueType(1));
9017 DEBUG(dbgs() << "\nReplacing.7 ";
9019 dbgs() << "\nWith: ";
9020 Undef.getNode()->dump(&DAG);
9021 dbgs() << " and 2 other values\n");
9022 WorklistRemover DeadNodes(*this);
9023 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
9024 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Index);
9025 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
9026 deleteAndRecombine(N);
9027 return SDValue(N, 0); // Return N so it doesn't get rechecked!
9032 // If this load is directly stored, replace the load value with the stored
9034 // TODO: Handle store large -> read small portion.
9035 // TODO: Handle TRUNCSTORE/LOADEXT
9036 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
9037 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
9038 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
9039 if (PrevST->getBasePtr() == Ptr &&
9040 PrevST->getValue().getValueType() == N->getValueType(0))
9041 return CombineTo(N, Chain.getOperand(1), Chain);
9045 // Try to infer better alignment information than the load already has.
9046 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
9047 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
9048 if (Align > LD->getMemOperand()->getBaseAlignment()) {
9050 DAG.getExtLoad(LD->getExtensionType(), SDLoc(N),
9051 LD->getValueType(0),
9052 Chain, Ptr, LD->getPointerInfo(),
9054 LD->isVolatile(), LD->isNonTemporal(),
9055 LD->isInvariant(), Align, LD->getAAInfo());
9056 if (NewLoad.getNode() != N)
9057 return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true);
9062 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
9063 : DAG.getSubtarget().useAA();
9065 if (CombinerAAOnlyFunc.getNumOccurrences() &&
9066 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
9069 if (UseAA && LD->isUnindexed()) {
9070 // Walk up chain skipping non-aliasing memory nodes.
9071 SDValue BetterChain = FindBetterChain(N, Chain);
9073 // If there is a better chain.
9074 if (Chain != BetterChain) {
9077 // Replace the chain to void dependency.
9078 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
9079 ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD),
9080 BetterChain, Ptr, LD->getMemOperand());
9082 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD),
9083 LD->getValueType(0),
9084 BetterChain, Ptr, LD->getMemoryVT(),
9085 LD->getMemOperand());
9088 // Create token factor to keep old chain connected.
9089 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
9090 MVT::Other, Chain, ReplLoad.getValue(1));
9092 // Make sure the new and old chains are cleaned up.
9093 AddToWorklist(Token.getNode());
9095 // Replace uses with load result and token factor. Don't add users
9097 return CombineTo(N, ReplLoad.getValue(0), Token, false);
9101 // Try transforming N to an indexed load.
9102 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
9103 return SDValue(N, 0);
9105 // Try to slice up N to more direct loads if the slices are mapped to
9106 // different register banks or pairing can take place.
9108 return SDValue(N, 0);
9114 /// \brief Helper structure used to slice a load in smaller loads.
9115 /// Basically a slice is obtained from the following sequence:
9116 /// Origin = load Ty1, Base
9117 /// Shift = srl Ty1 Origin, CstTy Amount
9118 /// Inst = trunc Shift to Ty2
9120 /// Then, it will be rewriten into:
9121 /// Slice = load SliceTy, Base + SliceOffset
9122 /// [Inst = zext Slice to Ty2], only if SliceTy <> Ty2
9124 /// SliceTy is deduced from the number of bits that are actually used to
9126 struct LoadedSlice {
9127 /// \brief Helper structure used to compute the cost of a slice.
9129 /// Are we optimizing for code size.
9134 unsigned CrossRegisterBanksCopies;
9138 Cost(bool ForCodeSize = false)
9139 : ForCodeSize(ForCodeSize), Loads(0), Truncates(0),
9140 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {}
9142 /// \brief Get the cost of one isolated slice.
9143 Cost(const LoadedSlice &LS, bool ForCodeSize = false)
9144 : ForCodeSize(ForCodeSize), Loads(1), Truncates(0),
9145 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {
9146 EVT TruncType = LS.Inst->getValueType(0);
9147 EVT LoadedType = LS.getLoadedType();
9148 if (TruncType != LoadedType &&
9149 !LS.DAG->getTargetLoweringInfo().isZExtFree(LoadedType, TruncType))
9153 /// \brief Account for slicing gain in the current cost.
9154 /// Slicing provide a few gains like removing a shift or a
9155 /// truncate. This method allows to grow the cost of the original
9156 /// load with the gain from this slice.
9157 void addSliceGain(const LoadedSlice &LS) {
9158 // Each slice saves a truncate.
9159 const TargetLowering &TLI = LS.DAG->getTargetLoweringInfo();
9160 if (!TLI.isTruncateFree(LS.Inst->getValueType(0),
9161 LS.Inst->getOperand(0).getValueType()))
9163 // If there is a shift amount, this slice gets rid of it.
9166 // If this slice can merge a cross register bank copy, account for it.
9167 if (LS.canMergeExpensiveCrossRegisterBankCopy())
9168 ++CrossRegisterBanksCopies;
9171 Cost &operator+=(const Cost &RHS) {
9173 Truncates += RHS.Truncates;
9174 CrossRegisterBanksCopies += RHS.CrossRegisterBanksCopies;
9180 bool operator==(const Cost &RHS) const {
9181 return Loads == RHS.Loads && Truncates == RHS.Truncates &&
9182 CrossRegisterBanksCopies == RHS.CrossRegisterBanksCopies &&
9183 ZExts == RHS.ZExts && Shift == RHS.Shift;
9186 bool operator!=(const Cost &RHS) const { return !(*this == RHS); }
9188 bool operator<(const Cost &RHS) const {
9189 // Assume cross register banks copies are as expensive as loads.
9190 // FIXME: Do we want some more target hooks?
9191 unsigned ExpensiveOpsLHS = Loads + CrossRegisterBanksCopies;
9192 unsigned ExpensiveOpsRHS = RHS.Loads + RHS.CrossRegisterBanksCopies;
9193 // Unless we are optimizing for code size, consider the
9194 // expensive operation first.
9195 if (!ForCodeSize && ExpensiveOpsLHS != ExpensiveOpsRHS)
9196 return ExpensiveOpsLHS < ExpensiveOpsRHS;
9197 return (Truncates + ZExts + Shift + ExpensiveOpsLHS) <
9198 (RHS.Truncates + RHS.ZExts + RHS.Shift + ExpensiveOpsRHS);
9201 bool operator>(const Cost &RHS) const { return RHS < *this; }
9203 bool operator<=(const Cost &RHS) const { return !(RHS < *this); }
9205 bool operator>=(const Cost &RHS) const { return !(*this < RHS); }
9207 // The last instruction that represent the slice. This should be a
9208 // truncate instruction.
9210 // The original load instruction.
9212 // The right shift amount in bits from the original load.
9214 // The DAG from which Origin came from.
9215 // This is used to get some contextual information about legal types, etc.
9218 LoadedSlice(SDNode *Inst = nullptr, LoadSDNode *Origin = nullptr,
9219 unsigned Shift = 0, SelectionDAG *DAG = nullptr)
9220 : Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {}
9222 /// \brief Get the bits used in a chunk of bits \p BitWidth large.
9223 /// \return Result is \p BitWidth and has used bits set to 1 and
9224 /// not used bits set to 0.
9225 APInt getUsedBits() const {
9226 // Reproduce the trunc(lshr) sequence:
9227 // - Start from the truncated value.
9228 // - Zero extend to the desired bit width.
9230 assert(Origin && "No original load to compare against.");
9231 unsigned BitWidth = Origin->getValueSizeInBits(0);
9232 assert(Inst && "This slice is not bound to an instruction");
9233 assert(Inst->getValueSizeInBits(0) <= BitWidth &&
9234 "Extracted slice is bigger than the whole type!");
9235 APInt UsedBits(Inst->getValueSizeInBits(0), 0);
9236 UsedBits.setAllBits();
9237 UsedBits = UsedBits.zext(BitWidth);
9242 /// \brief Get the size of the slice to be loaded in bytes.
9243 unsigned getLoadedSize() const {
9244 unsigned SliceSize = getUsedBits().countPopulation();
9245 assert(!(SliceSize & 0x7) && "Size is not a multiple of a byte.");
9246 return SliceSize / 8;
9249 /// \brief Get the type that will be loaded for this slice.
9250 /// Note: This may not be the final type for the slice.
9251 EVT getLoadedType() const {
9252 assert(DAG && "Missing context");
9253 LLVMContext &Ctxt = *DAG->getContext();
9254 return EVT::getIntegerVT(Ctxt, getLoadedSize() * 8);
9257 /// \brief Get the alignment of the load used for this slice.
9258 unsigned getAlignment() const {
9259 unsigned Alignment = Origin->getAlignment();
9260 unsigned Offset = getOffsetFromBase();
9262 Alignment = MinAlign(Alignment, Alignment + Offset);
9266 /// \brief Check if this slice can be rewritten with legal operations.
9267 bool isLegal() const {
9268 // An invalid slice is not legal.
9269 if (!Origin || !Inst || !DAG)
9272 // Offsets are for indexed load only, we do not handle that.
9273 if (Origin->getOffset().getOpcode() != ISD::UNDEF)
9276 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
9278 // Check that the type is legal.
9279 EVT SliceType = getLoadedType();
9280 if (!TLI.isTypeLegal(SliceType))
9283 // Check that the load is legal for this type.
9284 if (!TLI.isOperationLegal(ISD::LOAD, SliceType))
9287 // Check that the offset can be computed.
9288 // 1. Check its type.
9289 EVT PtrType = Origin->getBasePtr().getValueType();
9290 if (PtrType == MVT::Untyped || PtrType.isExtended())
9293 // 2. Check that it fits in the immediate.
9294 if (!TLI.isLegalAddImmediate(getOffsetFromBase()))
9297 // 3. Check that the computation is legal.
9298 if (!TLI.isOperationLegal(ISD::ADD, PtrType))
9301 // Check that the zext is legal if it needs one.
9302 EVT TruncateType = Inst->getValueType(0);
9303 if (TruncateType != SliceType &&
9304 !TLI.isOperationLegal(ISD::ZERO_EXTEND, TruncateType))
9310 /// \brief Get the offset in bytes of this slice in the original chunk of
9312 /// \pre DAG != nullptr.
9313 uint64_t getOffsetFromBase() const {
9314 assert(DAG && "Missing context.");
9316 DAG->getTargetLoweringInfo().getDataLayout()->isBigEndian();
9317 assert(!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported.");
9318 uint64_t Offset = Shift / 8;
9319 unsigned TySizeInBytes = Origin->getValueSizeInBits(0) / 8;
9320 assert(!(Origin->getValueSizeInBits(0) & 0x7) &&
9321 "The size of the original loaded type is not a multiple of a"
9323 // If Offset is bigger than TySizeInBytes, it means we are loading all
9324 // zeros. This should have been optimized before in the process.
9325 assert(TySizeInBytes > Offset &&
9326 "Invalid shift amount for given loaded size");
9328 Offset = TySizeInBytes - Offset - getLoadedSize();
9332 /// \brief Generate the sequence of instructions to load the slice
9333 /// represented by this object and redirect the uses of this slice to
9334 /// this new sequence of instructions.
9335 /// \pre this->Inst && this->Origin are valid Instructions and this
9336 /// object passed the legal check: LoadedSlice::isLegal returned true.
9337 /// \return The last instruction of the sequence used to load the slice.
9338 SDValue loadSlice() const {
9339 assert(Inst && Origin && "Unable to replace a non-existing slice.");
9340 const SDValue &OldBaseAddr = Origin->getBasePtr();
9341 SDValue BaseAddr = OldBaseAddr;
9342 // Get the offset in that chunk of bytes w.r.t. the endianess.
9343 int64_t Offset = static_cast<int64_t>(getOffsetFromBase());
9344 assert(Offset >= 0 && "Offset too big to fit in int64_t!");
9346 // BaseAddr = BaseAddr + Offset.
9347 EVT ArithType = BaseAddr.getValueType();
9348 BaseAddr = DAG->getNode(ISD::ADD, SDLoc(Origin), ArithType, BaseAddr,
9349 DAG->getConstant(Offset, ArithType));
9352 // Create the type of the loaded slice according to its size.
9353 EVT SliceType = getLoadedType();
9355 // Create the load for the slice.
9356 SDValue LastInst = DAG->getLoad(
9357 SliceType, SDLoc(Origin), Origin->getChain(), BaseAddr,
9358 Origin->getPointerInfo().getWithOffset(Offset), Origin->isVolatile(),
9359 Origin->isNonTemporal(), Origin->isInvariant(), getAlignment());
9360 // If the final type is not the same as the loaded type, this means that
9361 // we have to pad with zero. Create a zero extend for that.
9362 EVT FinalType = Inst->getValueType(0);
9363 if (SliceType != FinalType)
9365 DAG->getNode(ISD::ZERO_EXTEND, SDLoc(LastInst), FinalType, LastInst);
9369 /// \brief Check if this slice can be merged with an expensive cross register
9370 /// bank copy. E.g.,
9372 /// f = bitcast i32 i to float
9373 bool canMergeExpensiveCrossRegisterBankCopy() const {
9374 if (!Inst || !Inst->hasOneUse())
9376 SDNode *Use = *Inst->use_begin();
9377 if (Use->getOpcode() != ISD::BITCAST)
9379 assert(DAG && "Missing context");
9380 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
9381 EVT ResVT = Use->getValueType(0);
9382 const TargetRegisterClass *ResRC = TLI.getRegClassFor(ResVT.getSimpleVT());
9383 const TargetRegisterClass *ArgRC =
9384 TLI.getRegClassFor(Use->getOperand(0).getValueType().getSimpleVT());
9385 if (ArgRC == ResRC || !TLI.isOperationLegal(ISD::LOAD, ResVT))
9388 // At this point, we know that we perform a cross-register-bank copy.
9389 // Check if it is expensive.
9390 const TargetRegisterInfo *TRI = DAG->getSubtarget().getRegisterInfo();
9391 // Assume bitcasts are cheap, unless both register classes do not
9392 // explicitly share a common sub class.
9393 if (!TRI || TRI->getCommonSubClass(ArgRC, ResRC))
9396 // Check if it will be merged with the load.
9397 // 1. Check the alignment constraint.
9398 unsigned RequiredAlignment = TLI.getDataLayout()->getABITypeAlignment(
9399 ResVT.getTypeForEVT(*DAG->getContext()));
9401 if (RequiredAlignment > getAlignment())
9404 // 2. Check that the load is a legal operation for that type.
9405 if (!TLI.isOperationLegal(ISD::LOAD, ResVT))
9408 // 3. Check that we do not have a zext in the way.
9409 if (Inst->getValueType(0) != getLoadedType())
9417 /// \brief Check that all bits set in \p UsedBits form a dense region, i.e.,
9418 /// \p UsedBits looks like 0..0 1..1 0..0.
9419 static bool areUsedBitsDense(const APInt &UsedBits) {
9420 // If all the bits are one, this is dense!
9421 if (UsedBits.isAllOnesValue())
9424 // Get rid of the unused bits on the right.
9425 APInt NarrowedUsedBits = UsedBits.lshr(UsedBits.countTrailingZeros());
9426 // Get rid of the unused bits on the left.
9427 if (NarrowedUsedBits.countLeadingZeros())
9428 NarrowedUsedBits = NarrowedUsedBits.trunc(NarrowedUsedBits.getActiveBits());
9429 // Check that the chunk of bits is completely used.
9430 return NarrowedUsedBits.isAllOnesValue();
9433 /// \brief Check whether or not \p First and \p Second are next to each other
9434 /// in memory. This means that there is no hole between the bits loaded
9435 /// by \p First and the bits loaded by \p Second.
9436 static bool areSlicesNextToEachOther(const LoadedSlice &First,
9437 const LoadedSlice &Second) {
9438 assert(First.Origin == Second.Origin && First.Origin &&
9439 "Unable to match different memory origins.");
9440 APInt UsedBits = First.getUsedBits();
9441 assert((UsedBits & Second.getUsedBits()) == 0 &&
9442 "Slices are not supposed to overlap.");
9443 UsedBits |= Second.getUsedBits();
9444 return areUsedBitsDense(UsedBits);
9447 /// \brief Adjust the \p GlobalLSCost according to the target
9448 /// paring capabilities and the layout of the slices.
9449 /// \pre \p GlobalLSCost should account for at least as many loads as
9450 /// there is in the slices in \p LoadedSlices.
9451 static void adjustCostForPairing(SmallVectorImpl<LoadedSlice> &LoadedSlices,
9452 LoadedSlice::Cost &GlobalLSCost) {
9453 unsigned NumberOfSlices = LoadedSlices.size();
9454 // If there is less than 2 elements, no pairing is possible.
9455 if (NumberOfSlices < 2)
9458 // Sort the slices so that elements that are likely to be next to each
9459 // other in memory are next to each other in the list.
9460 std::sort(LoadedSlices.begin(), LoadedSlices.end(),
9461 [](const LoadedSlice &LHS, const LoadedSlice &RHS) {
9462 assert(LHS.Origin == RHS.Origin && "Different bases not implemented.");
9463 return LHS.getOffsetFromBase() < RHS.getOffsetFromBase();
9465 const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo();
9466 // First (resp. Second) is the first (resp. Second) potentially candidate
9467 // to be placed in a paired load.
9468 const LoadedSlice *First = nullptr;
9469 const LoadedSlice *Second = nullptr;
9470 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice,
9471 // Set the beginning of the pair.
9474 Second = &LoadedSlices[CurrSlice];
9476 // If First is NULL, it means we start a new pair.
9477 // Get to the next slice.
9481 EVT LoadedType = First->getLoadedType();
9483 // If the types of the slices are different, we cannot pair them.
9484 if (LoadedType != Second->getLoadedType())
9487 // Check if the target supplies paired loads for this type.
9488 unsigned RequiredAlignment = 0;
9489 if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) {
9490 // move to the next pair, this type is hopeless.
9494 // Check if we meet the alignment requirement.
9495 if (RequiredAlignment > First->getAlignment())
9498 // Check that both loads are next to each other in memory.
9499 if (!areSlicesNextToEachOther(*First, *Second))
9502 assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!");
9503 --GlobalLSCost.Loads;
9504 // Move to the next pair.
9509 /// \brief Check the profitability of all involved LoadedSlice.
9510 /// Currently, it is considered profitable if there is exactly two
9511 /// involved slices (1) which are (2) next to each other in memory, and
9512 /// whose cost (\see LoadedSlice::Cost) is smaller than the original load (3).
9514 /// Note: The order of the elements in \p LoadedSlices may be modified, but not
9515 /// the elements themselves.
9517 /// FIXME: When the cost model will be mature enough, we can relax
9518 /// constraints (1) and (2).
9519 static bool isSlicingProfitable(SmallVectorImpl<LoadedSlice> &LoadedSlices,
9520 const APInt &UsedBits, bool ForCodeSize) {
9521 unsigned NumberOfSlices = LoadedSlices.size();
9522 if (StressLoadSlicing)
9523 return NumberOfSlices > 1;
9526 if (NumberOfSlices != 2)
9530 if (!areUsedBitsDense(UsedBits))
9534 LoadedSlice::Cost OrigCost(ForCodeSize), GlobalSlicingCost(ForCodeSize);
9535 // The original code has one big load.
9537 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice) {
9538 const LoadedSlice &LS = LoadedSlices[CurrSlice];
9539 // Accumulate the cost of all the slices.
9540 LoadedSlice::Cost SliceCost(LS, ForCodeSize);
9541 GlobalSlicingCost += SliceCost;
9543 // Account as cost in the original configuration the gain obtained
9544 // with the current slices.
9545 OrigCost.addSliceGain(LS);
9548 // If the target supports paired load, adjust the cost accordingly.
9549 adjustCostForPairing(LoadedSlices, GlobalSlicingCost);
9550 return OrigCost > GlobalSlicingCost;
9553 /// \brief If the given load, \p LI, is used only by trunc or trunc(lshr)
9554 /// operations, split it in the various pieces being extracted.
9556 /// This sort of thing is introduced by SROA.
9557 /// This slicing takes care not to insert overlapping loads.
9558 /// \pre LI is a simple load (i.e., not an atomic or volatile load).
9559 bool DAGCombiner::SliceUpLoad(SDNode *N) {
9560 if (Level < AfterLegalizeDAG)
9563 LoadSDNode *LD = cast<LoadSDNode>(N);
9564 if (LD->isVolatile() || !ISD::isNormalLoad(LD) ||
9565 !LD->getValueType(0).isInteger())
9568 // Keep track of already used bits to detect overlapping values.
9569 // In that case, we will just abort the transformation.
9570 APInt UsedBits(LD->getValueSizeInBits(0), 0);
9572 SmallVector<LoadedSlice, 4> LoadedSlices;
9574 // Check if this load is used as several smaller chunks of bits.
9575 // Basically, look for uses in trunc or trunc(lshr) and record a new chain
9576 // of computation for each trunc.
9577 for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end();
9578 UI != UIEnd; ++UI) {
9579 // Skip the uses of the chain.
9580 if (UI.getUse().getResNo() != 0)
9586 // Check if this is a trunc(lshr).
9587 if (User->getOpcode() == ISD::SRL && User->hasOneUse() &&
9588 isa<ConstantSDNode>(User->getOperand(1))) {
9589 Shift = cast<ConstantSDNode>(User->getOperand(1))->getZExtValue();
9590 User = *User->use_begin();
9593 // At this point, User is a Truncate, iff we encountered, trunc or
9595 if (User->getOpcode() != ISD::TRUNCATE)
9598 // The width of the type must be a power of 2 and greater than 8-bits.
9599 // Otherwise the load cannot be represented in LLVM IR.
9600 // Moreover, if we shifted with a non-8-bits multiple, the slice
9601 // will be across several bytes. We do not support that.
9602 unsigned Width = User->getValueSizeInBits(0);
9603 if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7))
9606 // Build the slice for this chain of computations.
9607 LoadedSlice LS(User, LD, Shift, &DAG);
9608 APInt CurrentUsedBits = LS.getUsedBits();
9610 // Check if this slice overlaps with another.
9611 if ((CurrentUsedBits & UsedBits) != 0)
9613 // Update the bits used globally.
9614 UsedBits |= CurrentUsedBits;
9616 // Check if the new slice would be legal.
9620 // Record the slice.
9621 LoadedSlices.push_back(LS);
9624 // Abort slicing if it does not seem to be profitable.
9625 if (!isSlicingProfitable(LoadedSlices, UsedBits, ForCodeSize))
9630 // Rewrite each chain to use an independent load.
9631 // By construction, each chain can be represented by a unique load.
9633 // Prepare the argument for the new token factor for all the slices.
9634 SmallVector<SDValue, 8> ArgChains;
9635 for (SmallVectorImpl<LoadedSlice>::const_iterator
9636 LSIt = LoadedSlices.begin(),
9637 LSItEnd = LoadedSlices.end();
9638 LSIt != LSItEnd; ++LSIt) {
9639 SDValue SliceInst = LSIt->loadSlice();
9640 CombineTo(LSIt->Inst, SliceInst, true);
9641 if (SliceInst.getNode()->getOpcode() != ISD::LOAD)
9642 SliceInst = SliceInst.getOperand(0);
9643 assert(SliceInst->getOpcode() == ISD::LOAD &&
9644 "It takes more than a zext to get to the loaded slice!!");
9645 ArgChains.push_back(SliceInst.getValue(1));
9648 SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other,
9650 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
9654 /// Check to see if V is (and load (ptr), imm), where the load is having
9655 /// specific bytes cleared out. If so, return the byte size being masked out
9656 /// and the shift amount.
9657 static std::pair<unsigned, unsigned>
9658 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
9659 std::pair<unsigned, unsigned> Result(0, 0);
9661 // Check for the structure we're looking for.
9662 if (V->getOpcode() != ISD::AND ||
9663 !isa<ConstantSDNode>(V->getOperand(1)) ||
9664 !ISD::isNormalLoad(V->getOperand(0).getNode()))
9667 // Check the chain and pointer.
9668 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
9669 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
9671 // The store should be chained directly to the load or be an operand of a
9673 if (LD == Chain.getNode())
9675 else if (Chain->getOpcode() != ISD::TokenFactor)
9676 return Result; // Fail.
9679 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
9680 if (Chain->getOperand(i).getNode() == LD) {
9684 if (!isOk) return Result;
9687 // This only handles simple types.
9688 if (V.getValueType() != MVT::i16 &&
9689 V.getValueType() != MVT::i32 &&
9690 V.getValueType() != MVT::i64)
9693 // Check the constant mask. Invert it so that the bits being masked out are
9694 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
9695 // follow the sign bit for uniformity.
9696 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
9697 unsigned NotMaskLZ = countLeadingZeros(NotMask);
9698 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
9699 unsigned NotMaskTZ = countTrailingZeros(NotMask);
9700 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
9701 if (NotMaskLZ == 64) return Result; // All zero mask.
9703 // See if we have a continuous run of bits. If so, we have 0*1+0*
9704 if (countTrailingOnes(NotMask >> NotMaskTZ) + NotMaskTZ + NotMaskLZ != 64)
9707 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
9708 if (V.getValueType() != MVT::i64 && NotMaskLZ)
9709 NotMaskLZ -= 64-V.getValueSizeInBits();
9711 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
9712 switch (MaskedBytes) {
9716 default: return Result; // All one mask, or 5-byte mask.
9719 // Verify that the first bit starts at a multiple of mask so that the access
9720 // is aligned the same as the access width.
9721 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
9723 Result.first = MaskedBytes;
9724 Result.second = NotMaskTZ/8;
9729 /// Check to see if IVal is something that provides a value as specified by
9730 /// MaskInfo. If so, replace the specified store with a narrower store of
9733 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
9734 SDValue IVal, StoreSDNode *St,
9736 unsigned NumBytes = MaskInfo.first;
9737 unsigned ByteShift = MaskInfo.second;
9738 SelectionDAG &DAG = DC->getDAG();
9740 // Check to see if IVal is all zeros in the part being masked in by the 'or'
9741 // that uses this. If not, this is not a replacement.
9742 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
9743 ByteShift*8, (ByteShift+NumBytes)*8);
9744 if (!DAG.MaskedValueIsZero(IVal, Mask)) return nullptr;
9746 // Check that it is legal on the target to do this. It is legal if the new
9747 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
9749 MVT VT = MVT::getIntegerVT(NumBytes*8);
9750 if (!DC->isTypeLegal(VT))
9753 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
9754 // shifted by ByteShift and truncated down to NumBytes.
9756 IVal = DAG.getNode(ISD::SRL, SDLoc(IVal), IVal.getValueType(), IVal,
9757 DAG.getConstant(ByteShift*8,
9758 DC->getShiftAmountTy(IVal.getValueType())));
9760 // Figure out the offset for the store and the alignment of the access.
9762 unsigned NewAlign = St->getAlignment();
9764 if (DAG.getTargetLoweringInfo().isLittleEndian())
9765 StOffset = ByteShift;
9767 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
9769 SDValue Ptr = St->getBasePtr();
9771 Ptr = DAG.getNode(ISD::ADD, SDLoc(IVal), Ptr.getValueType(),
9772 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
9773 NewAlign = MinAlign(NewAlign, StOffset);
9776 // Truncate down to the new size.
9777 IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal);
9780 return DAG.getStore(St->getChain(), SDLoc(St), IVal, Ptr,
9781 St->getPointerInfo().getWithOffset(StOffset),
9782 false, false, NewAlign).getNode();
9786 /// Look for sequence of load / op / store where op is one of 'or', 'xor', and
9787 /// 'and' of immediates. If 'op' is only touching some of the loaded bits, try
9788 /// narrowing the load and store if it would end up being a win for performance
9790 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
9791 StoreSDNode *ST = cast<StoreSDNode>(N);
9792 if (ST->isVolatile())
9795 SDValue Chain = ST->getChain();
9796 SDValue Value = ST->getValue();
9797 SDValue Ptr = ST->getBasePtr();
9798 EVT VT = Value.getValueType();
9800 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
9803 unsigned Opc = Value.getOpcode();
9805 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
9806 // is a byte mask indicating a consecutive number of bytes, check to see if
9807 // Y is known to provide just those bytes. If so, we try to replace the
9808 // load + replace + store sequence with a single (narrower) store, which makes
9810 if (Opc == ISD::OR) {
9811 std::pair<unsigned, unsigned> MaskedLoad;
9812 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
9813 if (MaskedLoad.first)
9814 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
9815 Value.getOperand(1), ST,this))
9816 return SDValue(NewST, 0);
9818 // Or is commutative, so try swapping X and Y.
9819 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
9820 if (MaskedLoad.first)
9821 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
9822 Value.getOperand(0), ST,this))
9823 return SDValue(NewST, 0);
9826 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
9827 Value.getOperand(1).getOpcode() != ISD::Constant)
9830 SDValue N0 = Value.getOperand(0);
9831 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
9832 Chain == SDValue(N0.getNode(), 1)) {
9833 LoadSDNode *LD = cast<LoadSDNode>(N0);
9834 if (LD->getBasePtr() != Ptr ||
9835 LD->getPointerInfo().getAddrSpace() !=
9836 ST->getPointerInfo().getAddrSpace())
9839 // Find the type to narrow it the load / op / store to.
9840 SDValue N1 = Value.getOperand(1);
9841 unsigned BitWidth = N1.getValueSizeInBits();
9842 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
9843 if (Opc == ISD::AND)
9844 Imm ^= APInt::getAllOnesValue(BitWidth);
9845 if (Imm == 0 || Imm.isAllOnesValue())
9847 unsigned ShAmt = Imm.countTrailingZeros();
9848 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
9849 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
9850 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
9851 // The narrowing should be profitable, the load/store operation should be
9852 // legal (or custom) and the store size should be equal to the NewVT width.
9853 while (NewBW < BitWidth &&
9854 (NewVT.getStoreSizeInBits() != NewBW ||
9855 !TLI.isOperationLegalOrCustom(Opc, NewVT) ||
9856 !TLI.isNarrowingProfitable(VT, NewVT))) {
9857 NewBW = NextPowerOf2(NewBW);
9858 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
9860 if (NewBW >= BitWidth)
9863 // If the lsb changed does not start at the type bitwidth boundary,
9864 // start at the previous one.
9866 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
9867 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt,
9868 std::min(BitWidth, ShAmt + NewBW));
9869 if ((Imm & Mask) == Imm) {
9870 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
9871 if (Opc == ISD::AND)
9872 NewImm ^= APInt::getAllOnesValue(NewBW);
9873 uint64_t PtrOff = ShAmt / 8;
9874 // For big endian targets, we need to adjust the offset to the pointer to
9875 // load the correct bytes.
9876 if (TLI.isBigEndian())
9877 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
9879 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
9880 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
9881 if (NewAlign < TLI.getDataLayout()->getABITypeAlignment(NewVTTy))
9884 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD),
9885 Ptr.getValueType(), Ptr,
9886 DAG.getConstant(PtrOff, Ptr.getValueType()));
9887 SDValue NewLD = DAG.getLoad(NewVT, SDLoc(N0),
9888 LD->getChain(), NewPtr,
9889 LD->getPointerInfo().getWithOffset(PtrOff),
9890 LD->isVolatile(), LD->isNonTemporal(),
9891 LD->isInvariant(), NewAlign,
9893 SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD,
9894 DAG.getConstant(NewImm, NewVT));
9895 SDValue NewST = DAG.getStore(Chain, SDLoc(N),
9897 ST->getPointerInfo().getWithOffset(PtrOff),
9898 false, false, NewAlign);
9900 AddToWorklist(NewPtr.getNode());
9901 AddToWorklist(NewLD.getNode());
9902 AddToWorklist(NewVal.getNode());
9903 WorklistRemover DeadNodes(*this);
9904 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
9913 /// For a given floating point load / store pair, if the load value isn't used
9914 /// by any other operations, then consider transforming the pair to integer
9915 /// load / store operations if the target deems the transformation profitable.
9916 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
9917 StoreSDNode *ST = cast<StoreSDNode>(N);
9918 SDValue Chain = ST->getChain();
9919 SDValue Value = ST->getValue();
9920 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
9921 Value.hasOneUse() &&
9922 Chain == SDValue(Value.getNode(), 1)) {
9923 LoadSDNode *LD = cast<LoadSDNode>(Value);
9924 EVT VT = LD->getMemoryVT();
9925 if (!VT.isFloatingPoint() ||
9926 VT != ST->getMemoryVT() ||
9927 LD->isNonTemporal() ||
9928 ST->isNonTemporal() ||
9929 LD->getPointerInfo().getAddrSpace() != 0 ||
9930 ST->getPointerInfo().getAddrSpace() != 0)
9933 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
9934 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
9935 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
9936 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
9937 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
9940 unsigned LDAlign = LD->getAlignment();
9941 unsigned STAlign = ST->getAlignment();
9942 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
9943 unsigned ABIAlign = TLI.getDataLayout()->getABITypeAlignment(IntVTTy);
9944 if (LDAlign < ABIAlign || STAlign < ABIAlign)
9947 SDValue NewLD = DAG.getLoad(IntVT, SDLoc(Value),
9948 LD->getChain(), LD->getBasePtr(),
9949 LD->getPointerInfo(),
9950 false, false, false, LDAlign);
9952 SDValue NewST = DAG.getStore(NewLD.getValue(1), SDLoc(N),
9953 NewLD, ST->getBasePtr(),
9954 ST->getPointerInfo(),
9955 false, false, STAlign);
9957 AddToWorklist(NewLD.getNode());
9958 AddToWorklist(NewST.getNode());
9959 WorklistRemover DeadNodes(*this);
9960 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
9969 /// Helper struct to parse and store a memory address as base + index + offset.
9970 /// We ignore sign extensions when it is safe to do so.
9971 /// The following two expressions are not equivalent. To differentiate we need
9972 /// to store whether there was a sign extension involved in the index
9974 /// (load (i64 add (i64 copyfromreg %c)
9975 /// (i64 signextend (add (i8 load %index)
9979 /// (load (i64 add (i64 copyfromreg %c)
9980 /// (i64 signextend (i32 add (i32 signextend (i8 load %index))
9982 struct BaseIndexOffset {
9986 bool IsIndexSignExt;
9988 BaseIndexOffset() : Offset(0), IsIndexSignExt(false) {}
9990 BaseIndexOffset(SDValue Base, SDValue Index, int64_t Offset,
9991 bool IsIndexSignExt) :
9992 Base(Base), Index(Index), Offset(Offset), IsIndexSignExt(IsIndexSignExt) {}
9994 bool equalBaseIndex(const BaseIndexOffset &Other) {
9995 return Other.Base == Base && Other.Index == Index &&
9996 Other.IsIndexSignExt == IsIndexSignExt;
9999 /// Parses tree in Ptr for base, index, offset addresses.
10000 static BaseIndexOffset match(SDValue Ptr) {
10001 bool IsIndexSignExt = false;
10003 // We only can pattern match BASE + INDEX + OFFSET. If Ptr is not an ADD
10004 // instruction, then it could be just the BASE or everything else we don't
10005 // know how to handle. Just use Ptr as BASE and give up.
10006 if (Ptr->getOpcode() != ISD::ADD)
10007 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
10009 // We know that we have at least an ADD instruction. Try to pattern match
10010 // the simple case of BASE + OFFSET.
10011 if (isa<ConstantSDNode>(Ptr->getOperand(1))) {
10012 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
10013 return BaseIndexOffset(Ptr->getOperand(0), SDValue(), Offset,
10017 // Inside a loop the current BASE pointer is calculated using an ADD and a
10018 // MUL instruction. In this case Ptr is the actual BASE pointer.
10019 // (i64 add (i64 %array_ptr)
10020 // (i64 mul (i64 %induction_var)
10021 // (i64 %element_size)))
10022 if (Ptr->getOperand(1)->getOpcode() == ISD::MUL)
10023 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
10025 // Look at Base + Index + Offset cases.
10026 SDValue Base = Ptr->getOperand(0);
10027 SDValue IndexOffset = Ptr->getOperand(1);
10029 // Skip signextends.
10030 if (IndexOffset->getOpcode() == ISD::SIGN_EXTEND) {
10031 IndexOffset = IndexOffset->getOperand(0);
10032 IsIndexSignExt = true;
10035 // Either the case of Base + Index (no offset) or something else.
10036 if (IndexOffset->getOpcode() != ISD::ADD)
10037 return BaseIndexOffset(Base, IndexOffset, 0, IsIndexSignExt);
10039 // Now we have the case of Base + Index + offset.
10040 SDValue Index = IndexOffset->getOperand(0);
10041 SDValue Offset = IndexOffset->getOperand(1);
10043 if (!isa<ConstantSDNode>(Offset))
10044 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
10046 // Ignore signextends.
10047 if (Index->getOpcode() == ISD::SIGN_EXTEND) {
10048 Index = Index->getOperand(0);
10049 IsIndexSignExt = true;
10050 } else IsIndexSignExt = false;
10052 int64_t Off = cast<ConstantSDNode>(Offset)->getSExtValue();
10053 return BaseIndexOffset(Base, Index, Off, IsIndexSignExt);
10058 bool DAGCombiner::MergeStoresOfConstantsOrVecElts(
10059 SmallVectorImpl<MemOpLink> &StoreNodes, EVT MemVT,
10060 unsigned NumElem, bool IsConstantSrc, bool UseVector) {
10061 // Make sure we have something to merge.
10065 int64_t ElementSizeBytes = MemVT.getSizeInBits() / 8;
10066 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
10067 unsigned EarliestNodeUsed = 0;
10069 for (unsigned i=0; i < NumElem; ++i) {
10070 // Find a chain for the new wide-store operand. Notice that some
10071 // of the store nodes that we found may not be selected for inclusion
10072 // in the wide store. The chain we use needs to be the chain of the
10073 // earliest store node which is *used* and replaced by the wide store.
10074 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
10075 EarliestNodeUsed = i;
10078 // The earliest Node in the DAG.
10079 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
10080 SDLoc DL(StoreNodes[0].MemNode);
10084 // Find a legal type for the vector store.
10085 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
10086 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
10087 if (IsConstantSrc) {
10088 // A vector store with a constant source implies that the constant is
10089 // zero; we only handle merging stores of constant zeros because the zero
10090 // can be materialized without a load.
10091 // It may be beneficial to loosen this restriction to allow non-zero
10093 StoredVal = DAG.getConstant(0, Ty);
10095 SmallVector<SDValue, 8> Ops;
10096 for (unsigned i = 0; i < NumElem ; ++i) {
10097 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10098 SDValue Val = St->getValue();
10099 // All of the operands of a BUILD_VECTOR must have the same type.
10100 if (Val.getValueType() != MemVT)
10102 Ops.push_back(Val);
10105 // Build the extracted vector elements back into a vector.
10106 StoredVal = DAG.getNode(ISD::BUILD_VECTOR, DL, Ty, Ops);
10109 // We should always use a vector store when merging extracted vector
10110 // elements, so this path implies a store of constants.
10111 assert(IsConstantSrc && "Merged vector elements should use vector store");
10113 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
10114 APInt StoreInt(StoreBW, 0);
10116 // Construct a single integer constant which is made of the smaller
10117 // constant inputs.
10118 bool IsLE = TLI.isLittleEndian();
10119 for (unsigned i = 0; i < NumElem ; ++i) {
10120 unsigned Idx = IsLE ? (NumElem - 1 - i) : i;
10121 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
10122 SDValue Val = St->getValue();
10123 StoreInt <<= ElementSizeBytes*8;
10124 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
10125 StoreInt |= C->getAPIntValue().zext(StoreBW);
10126 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
10127 StoreInt |= C->getValueAPF().bitcastToAPInt().zext(StoreBW);
10129 llvm_unreachable("Invalid constant element type");
10133 // Create the new Load and Store operations.
10134 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
10135 StoredVal = DAG.getConstant(StoreInt, StoreTy);
10138 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), DL, StoredVal,
10139 FirstInChain->getBasePtr(),
10140 FirstInChain->getPointerInfo(),
10142 FirstInChain->getAlignment());
10144 // Replace the first store with the new store
10145 CombineTo(EarliestOp, NewStore);
10146 // Erase all other stores.
10147 for (unsigned i = 0; i < NumElem ; ++i) {
10148 if (StoreNodes[i].MemNode == EarliestOp)
10150 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10151 // ReplaceAllUsesWith will replace all uses that existed when it was
10152 // called, but graph optimizations may cause new ones to appear. For
10153 // example, the case in pr14333 looks like
10155 // St's chain -> St -> another store -> X
10157 // And the only difference from St to the other store is the chain.
10158 // When we change it's chain to be St's chain they become identical,
10159 // get CSEed and the net result is that X is now a use of St.
10160 // Since we know that St is redundant, just iterate.
10161 while (!St->use_empty())
10162 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain());
10163 deleteAndRecombine(St);
10169 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
10170 if (OptLevel == CodeGenOpt::None)
10173 EVT MemVT = St->getMemoryVT();
10174 int64_t ElementSizeBytes = MemVT.getSizeInBits()/8;
10175 bool NoVectors = DAG.getMachineFunction().getFunction()->hasFnAttribute(
10176 Attribute::NoImplicitFloat);
10178 // Don't merge vectors into wider inputs.
10179 if (MemVT.isVector() || !MemVT.isSimple())
10182 // Perform an early exit check. Do not bother looking at stored values that
10183 // are not constants, loads, or extracted vector elements.
10184 SDValue StoredVal = St->getValue();
10185 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
10186 bool IsConstantSrc = isa<ConstantSDNode>(StoredVal) ||
10187 isa<ConstantFPSDNode>(StoredVal);
10188 bool IsExtractVecEltSrc = (StoredVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT);
10190 if (!IsConstantSrc && !IsLoadSrc && !IsExtractVecEltSrc)
10193 // Only look at ends of store sequences.
10194 SDValue Chain = SDValue(St, 0);
10195 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
10198 // This holds the base pointer, index, and the offset in bytes from the base
10200 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr());
10202 // We must have a base and an offset.
10203 if (!BasePtr.Base.getNode())
10206 // Do not handle stores to undef base pointers.
10207 if (BasePtr.Base.getOpcode() == ISD::UNDEF)
10210 // Save the LoadSDNodes that we find in the chain.
10211 // We need to make sure that these nodes do not interfere with
10212 // any of the store nodes.
10213 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
10215 // Save the StoreSDNodes that we find in the chain.
10216 SmallVector<MemOpLink, 8> StoreNodes;
10218 // Walk up the chain and look for nodes with offsets from the same
10219 // base pointer. Stop when reaching an instruction with a different kind
10220 // or instruction which has a different base pointer.
10222 StoreSDNode *Index = St;
10224 // If the chain has more than one use, then we can't reorder the mem ops.
10225 if (Index != St && !SDValue(Index, 0)->hasOneUse())
10228 // Find the base pointer and offset for this memory node.
10229 BaseIndexOffset Ptr = BaseIndexOffset::match(Index->getBasePtr());
10231 // Check that the base pointer is the same as the original one.
10232 if (!Ptr.equalBaseIndex(BasePtr))
10235 // Check that the alignment is the same.
10236 if (Index->getAlignment() != St->getAlignment())
10239 // The memory operands must not be volatile.
10240 if (Index->isVolatile() || Index->isIndexed())
10244 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
10245 if (St->isTruncatingStore())
10248 // The stored memory type must be the same.
10249 if (Index->getMemoryVT() != MemVT)
10252 // We do not allow unaligned stores because we want to prevent overriding
10254 if (Index->getAlignment()*8 != MemVT.getSizeInBits())
10257 // We found a potential memory operand to merge.
10258 StoreNodes.push_back(MemOpLink(Index, Ptr.Offset, Seq++));
10260 // Find the next memory operand in the chain. If the next operand in the
10261 // chain is a store then move up and continue the scan with the next
10262 // memory operand. If the next operand is a load save it and use alias
10263 // information to check if it interferes with anything.
10264 SDNode *NextInChain = Index->getChain().getNode();
10266 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
10267 // We found a store node. Use it for the next iteration.
10270 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
10271 if (Ldn->isVolatile()) {
10276 // Save the load node for later. Continue the scan.
10277 AliasLoadNodes.push_back(Ldn);
10278 NextInChain = Ldn->getChain().getNode();
10287 // Check if there is anything to merge.
10288 if (StoreNodes.size() < 2)
10291 // Sort the memory operands according to their distance from the base pointer.
10292 std::sort(StoreNodes.begin(), StoreNodes.end(),
10293 [](MemOpLink LHS, MemOpLink RHS) {
10294 return LHS.OffsetFromBase < RHS.OffsetFromBase ||
10295 (LHS.OffsetFromBase == RHS.OffsetFromBase &&
10296 LHS.SequenceNum > RHS.SequenceNum);
10299 // Scan the memory operations on the chain and find the first non-consecutive
10300 // store memory address.
10301 unsigned LastConsecutiveStore = 0;
10302 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
10303 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) {
10305 // Check that the addresses are consecutive starting from the second
10306 // element in the list of stores.
10308 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
10309 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
10313 bool Alias = false;
10314 // Check if this store interferes with any of the loads that we found.
10315 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld)
10316 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) {
10320 // We found a load that alias with this store. Stop the sequence.
10324 // Mark this node as useful.
10325 LastConsecutiveStore = i;
10328 // The node with the lowest store address.
10329 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
10331 // Store the constants into memory as one consecutive store.
10332 if (IsConstantSrc) {
10333 unsigned LastLegalType = 0;
10334 unsigned LastLegalVectorType = 0;
10335 bool NonZero = false;
10336 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
10337 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10338 SDValue StoredVal = St->getValue();
10340 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
10341 NonZero |= !C->isNullValue();
10342 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) {
10343 NonZero |= !C->getConstantFPValue()->isNullValue();
10349 // Find a legal type for the constant store.
10350 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
10351 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
10352 if (TLI.isTypeLegal(StoreTy))
10353 LastLegalType = i+1;
10354 // Or check whether a truncstore is legal.
10355 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
10356 TargetLowering::TypePromoteInteger) {
10357 EVT LegalizedStoredValueTy =
10358 TLI.getTypeToTransformTo(*DAG.getContext(), StoredVal.getValueType());
10359 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy))
10360 LastLegalType = i+1;
10363 // Find a legal type for the vector store.
10364 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
10365 if (TLI.isTypeLegal(Ty))
10366 LastLegalVectorType = i + 1;
10369 // We only use vectors if the constant is known to be zero and the
10370 // function is not marked with the noimplicitfloat attribute.
10371 if (NonZero || NoVectors)
10372 LastLegalVectorType = 0;
10374 // Check if we found a legal integer type to store.
10375 if (LastLegalType == 0 && LastLegalVectorType == 0)
10378 bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors;
10379 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType;
10381 return MergeStoresOfConstantsOrVecElts(StoreNodes, MemVT, NumElem,
10385 // When extracting multiple vector elements, try to store them
10386 // in one vector store rather than a sequence of scalar stores.
10387 if (IsExtractVecEltSrc) {
10388 unsigned NumElem = 0;
10389 for (unsigned i = 0; i < LastConsecutiveStore + 1; ++i) {
10390 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10391 SDValue StoredVal = St->getValue();
10392 // This restriction could be loosened.
10393 // Bail out if any stored values are not elements extracted from a vector.
10394 // It should be possible to handle mixed sources, but load sources need
10395 // more careful handling (see the block of code below that handles
10396 // consecutive loads).
10397 if (StoredVal.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
10400 // Find a legal type for the vector store.
10401 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
10402 if (TLI.isTypeLegal(Ty))
10406 return MergeStoresOfConstantsOrVecElts(StoreNodes, MemVT, NumElem,
10410 // Below we handle the case of multiple consecutive stores that
10411 // come from multiple consecutive loads. We merge them into a single
10412 // wide load and a single wide store.
10414 // Look for load nodes which are used by the stored values.
10415 SmallVector<MemOpLink, 8> LoadNodes;
10417 // Find acceptable loads. Loads need to have the same chain (token factor),
10418 // must not be zext, volatile, indexed, and they must be consecutive.
10419 BaseIndexOffset LdBasePtr;
10420 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
10421 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10422 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
10425 // Loads must only have one use.
10426 if (!Ld->hasNUsesOfValue(1, 0))
10429 // Check that the alignment is the same as the stores.
10430 if (Ld->getAlignment() != St->getAlignment())
10433 // The memory operands must not be volatile.
10434 if (Ld->isVolatile() || Ld->isIndexed())
10437 // We do not accept ext loads.
10438 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
10441 // The stored memory type must be the same.
10442 if (Ld->getMemoryVT() != MemVT)
10445 BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld->getBasePtr());
10446 // If this is not the first ptr that we check.
10447 if (LdBasePtr.Base.getNode()) {
10448 // The base ptr must be the same.
10449 if (!LdPtr.equalBaseIndex(LdBasePtr))
10452 // Check that all other base pointers are the same as this one.
10456 // We found a potential memory operand to merge.
10457 LoadNodes.push_back(MemOpLink(Ld, LdPtr.Offset, 0));
10460 if (LoadNodes.size() < 2)
10463 // If we have load/store pair instructions and we only have two values,
10465 unsigned RequiredAlignment;
10466 if (LoadNodes.size() == 2 && TLI.hasPairedLoad(MemVT, RequiredAlignment) &&
10467 St->getAlignment() >= RequiredAlignment)
10470 // Scan the memory operations on the chain and find the first non-consecutive
10471 // load memory address. These variables hold the index in the store node
10473 unsigned LastConsecutiveLoad = 0;
10474 // This variable refers to the size and not index in the array.
10475 unsigned LastLegalVectorType = 0;
10476 unsigned LastLegalIntegerType = 0;
10477 StartAddress = LoadNodes[0].OffsetFromBase;
10478 SDValue FirstChain = LoadNodes[0].MemNode->getChain();
10479 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
10480 // All loads much share the same chain.
10481 if (LoadNodes[i].MemNode->getChain() != FirstChain)
10484 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
10485 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
10487 LastConsecutiveLoad = i;
10489 // Find a legal type for the vector store.
10490 EVT StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
10491 if (TLI.isTypeLegal(StoreTy))
10492 LastLegalVectorType = i + 1;
10494 // Find a legal type for the integer store.
10495 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
10496 StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
10497 if (TLI.isTypeLegal(StoreTy))
10498 LastLegalIntegerType = i + 1;
10499 // Or check whether a truncstore and extload is legal.
10500 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
10501 TargetLowering::TypePromoteInteger) {
10502 EVT LegalizedStoredValueTy =
10503 TLI.getTypeToTransformTo(*DAG.getContext(), StoreTy);
10504 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
10505 TLI.isLoadExtLegal(ISD::ZEXTLOAD, LegalizedStoredValueTy, StoreTy) &&
10506 TLI.isLoadExtLegal(ISD::SEXTLOAD, LegalizedStoredValueTy, StoreTy) &&
10507 TLI.isLoadExtLegal(ISD::EXTLOAD, LegalizedStoredValueTy, StoreTy))
10508 LastLegalIntegerType = i+1;
10512 // Only use vector types if the vector type is larger than the integer type.
10513 // If they are the same, use integers.
10514 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors;
10515 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType);
10517 // We add +1 here because the LastXXX variables refer to location while
10518 // the NumElem refers to array/index size.
10519 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1;
10520 NumElem = std::min(LastLegalType, NumElem);
10525 // The earliest Node in the DAG.
10526 unsigned EarliestNodeUsed = 0;
10527 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
10528 for (unsigned i=1; i<NumElem; ++i) {
10529 // Find a chain for the new wide-store operand. Notice that some
10530 // of the store nodes that we found may not be selected for inclusion
10531 // in the wide store. The chain we use needs to be the chain of the
10532 // earliest store node which is *used* and replaced by the wide store.
10533 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
10534 EarliestNodeUsed = i;
10537 // Find if it is better to use vectors or integers to load and store
10541 JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
10543 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
10544 JointMemOpVT = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
10547 SDLoc LoadDL(LoadNodes[0].MemNode);
10548 SDLoc StoreDL(StoreNodes[0].MemNode);
10550 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
10551 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL,
10552 FirstLoad->getChain(),
10553 FirstLoad->getBasePtr(),
10554 FirstLoad->getPointerInfo(),
10555 false, false, false,
10556 FirstLoad->getAlignment());
10558 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), StoreDL, NewLoad,
10559 FirstInChain->getBasePtr(),
10560 FirstInChain->getPointerInfo(), false, false,
10561 FirstInChain->getAlignment());
10563 // Replace one of the loads with the new load.
10564 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode);
10565 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
10566 SDValue(NewLoad.getNode(), 1));
10568 // Remove the rest of the load chains.
10569 for (unsigned i = 1; i < NumElem ; ++i) {
10570 // Replace all chain users of the old load nodes with the chain of the new
10572 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
10573 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain());
10576 // Replace the first store with the new store.
10577 CombineTo(EarliestOp, NewStore);
10578 // Erase all other stores.
10579 for (unsigned i = 0; i < NumElem ; ++i) {
10580 // Remove all Store nodes.
10581 if (StoreNodes[i].MemNode == EarliestOp)
10583 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10584 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain());
10585 deleteAndRecombine(St);
10591 SDValue DAGCombiner::visitSTORE(SDNode *N) {
10592 StoreSDNode *ST = cast<StoreSDNode>(N);
10593 SDValue Chain = ST->getChain();
10594 SDValue Value = ST->getValue();
10595 SDValue Ptr = ST->getBasePtr();
10597 // If this is a store of a bit convert, store the input value if the
10598 // resultant store does not need a higher alignment than the original.
10599 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
10600 ST->isUnindexed()) {
10601 unsigned OrigAlign = ST->getAlignment();
10602 EVT SVT = Value.getOperand(0).getValueType();
10603 unsigned Align = TLI.getDataLayout()->
10604 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
10605 if (Align <= OrigAlign &&
10606 ((!LegalOperations && !ST->isVolatile()) ||
10607 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
10608 return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0),
10609 Ptr, ST->getPointerInfo(), ST->isVolatile(),
10610 ST->isNonTemporal(), OrigAlign,
10614 // Turn 'store undef, Ptr' -> nothing.
10615 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
10618 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
10619 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
10620 // NOTE: If the original store is volatile, this transform must not increase
10621 // the number of stores. For example, on x86-32 an f64 can be stored in one
10622 // processor operation but an i64 (which is not legal) requires two. So the
10623 // transform should not be done in this case.
10624 if (Value.getOpcode() != ISD::TargetConstantFP) {
10626 switch (CFP->getSimpleValueType(0).SimpleTy) {
10627 default: llvm_unreachable("Unknown FP type");
10628 case MVT::f16: // We don't do this for these yet.
10634 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
10635 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
10636 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
10637 bitcastToAPInt().getZExtValue(), MVT::i32);
10638 return DAG.getStore(Chain, SDLoc(N), Tmp,
10639 Ptr, ST->getMemOperand());
10643 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
10644 !ST->isVolatile()) ||
10645 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
10646 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
10647 getZExtValue(), MVT::i64);
10648 return DAG.getStore(Chain, SDLoc(N), Tmp,
10649 Ptr, ST->getMemOperand());
10652 if (!ST->isVolatile() &&
10653 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
10654 // Many FP stores are not made apparent until after legalize, e.g. for
10655 // argument passing. Since this is so common, custom legalize the
10656 // 64-bit integer store into two 32-bit stores.
10657 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
10658 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
10659 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
10660 if (TLI.isBigEndian()) std::swap(Lo, Hi);
10662 unsigned Alignment = ST->getAlignment();
10663 bool isVolatile = ST->isVolatile();
10664 bool isNonTemporal = ST->isNonTemporal();
10665 AAMDNodes AAInfo = ST->getAAInfo();
10667 SDValue St0 = DAG.getStore(Chain, SDLoc(ST), Lo,
10668 Ptr, ST->getPointerInfo(),
10669 isVolatile, isNonTemporal,
10670 ST->getAlignment(), AAInfo);
10671 Ptr = DAG.getNode(ISD::ADD, SDLoc(N), Ptr.getValueType(), Ptr,
10672 DAG.getConstant(4, Ptr.getValueType()));
10673 Alignment = MinAlign(Alignment, 4U);
10674 SDValue St1 = DAG.getStore(Chain, SDLoc(ST), Hi,
10675 Ptr, ST->getPointerInfo().getWithOffset(4),
10676 isVolatile, isNonTemporal,
10677 Alignment, AAInfo);
10678 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other,
10687 // Try to infer better alignment information than the store already has.
10688 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
10689 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
10690 if (Align > ST->getAlignment()) {
10692 DAG.getTruncStore(Chain, SDLoc(N), Value,
10693 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
10694 ST->isVolatile(), ST->isNonTemporal(), Align,
10696 if (NewStore.getNode() != N)
10697 return CombineTo(ST, NewStore, true);
10702 // Try transforming a pair floating point load / store ops to integer
10703 // load / store ops.
10704 SDValue NewST = TransformFPLoadStorePair(N);
10705 if (NewST.getNode())
10708 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
10709 : DAG.getSubtarget().useAA();
10711 if (CombinerAAOnlyFunc.getNumOccurrences() &&
10712 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
10715 if (UseAA && ST->isUnindexed()) {
10716 // Walk up chain skipping non-aliasing memory nodes.
10717 SDValue BetterChain = FindBetterChain(N, Chain);
10719 // If there is a better chain.
10720 if (Chain != BetterChain) {
10723 // Replace the chain to avoid dependency.
10724 if (ST->isTruncatingStore()) {
10725 ReplStore = DAG.getTruncStore(BetterChain, SDLoc(N), Value, Ptr,
10726 ST->getMemoryVT(), ST->getMemOperand());
10728 ReplStore = DAG.getStore(BetterChain, SDLoc(N), Value, Ptr,
10729 ST->getMemOperand());
10732 // Create token to keep both nodes around.
10733 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
10734 MVT::Other, Chain, ReplStore);
10736 // Make sure the new and old chains are cleaned up.
10737 AddToWorklist(Token.getNode());
10739 // Don't add users to work list.
10740 return CombineTo(N, Token, false);
10744 // Try transforming N to an indexed store.
10745 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
10746 return SDValue(N, 0);
10748 // FIXME: is there such a thing as a truncating indexed store?
10749 if (ST->isTruncatingStore() && ST->isUnindexed() &&
10750 Value.getValueType().isInteger()) {
10751 // See if we can simplify the input to this truncstore with knowledge that
10752 // only the low bits are being used. For example:
10753 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
10755 GetDemandedBits(Value,
10756 APInt::getLowBitsSet(
10757 Value.getValueType().getScalarType().getSizeInBits(),
10758 ST->getMemoryVT().getScalarType().getSizeInBits()));
10759 AddToWorklist(Value.getNode());
10760 if (Shorter.getNode())
10761 return DAG.getTruncStore(Chain, SDLoc(N), Shorter,
10762 Ptr, ST->getMemoryVT(), ST->getMemOperand());
10764 // Otherwise, see if we can simplify the operation with
10765 // SimplifyDemandedBits, which only works if the value has a single use.
10766 if (SimplifyDemandedBits(Value,
10767 APInt::getLowBitsSet(
10768 Value.getValueType().getScalarType().getSizeInBits(),
10769 ST->getMemoryVT().getScalarType().getSizeInBits())))
10770 return SDValue(N, 0);
10773 // If this is a load followed by a store to the same location, then the store
10775 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
10776 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
10777 ST->isUnindexed() && !ST->isVolatile() &&
10778 // There can't be any side effects between the load and store, such as
10779 // a call or store.
10780 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
10781 // The store is dead, remove it.
10786 // If this is a store followed by a store with the same value to the same
10787 // location, then the store is dead/noop.
10788 if (StoreSDNode *ST1 = dyn_cast<StoreSDNode>(Chain)) {
10789 if (ST1->getBasePtr() == Ptr && ST->getMemoryVT() == ST1->getMemoryVT() &&
10790 ST1->getValue() == Value && ST->isUnindexed() && !ST->isVolatile() &&
10791 ST1->isUnindexed() && !ST1->isVolatile()) {
10792 // The store is dead, remove it.
10797 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
10798 // truncating store. We can do this even if this is already a truncstore.
10799 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
10800 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
10801 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
10802 ST->getMemoryVT())) {
10803 return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0),
10804 Ptr, ST->getMemoryVT(), ST->getMemOperand());
10807 // Only perform this optimization before the types are legal, because we
10808 // don't want to perform this optimization on every DAGCombine invocation.
10810 bool EverChanged = false;
10813 // There can be multiple store sequences on the same chain.
10814 // Keep trying to merge store sequences until we are unable to do so
10815 // or until we merge the last store on the chain.
10816 bool Changed = MergeConsecutiveStores(ST);
10817 EverChanged |= Changed;
10818 if (!Changed) break;
10819 } while (ST->getOpcode() != ISD::DELETED_NODE);
10822 return SDValue(N, 0);
10825 return ReduceLoadOpStoreWidth(N);
10828 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
10829 SDValue InVec = N->getOperand(0);
10830 SDValue InVal = N->getOperand(1);
10831 SDValue EltNo = N->getOperand(2);
10834 // If the inserted element is an UNDEF, just use the input vector.
10835 if (InVal.getOpcode() == ISD::UNDEF)
10838 EVT VT = InVec.getValueType();
10840 // If we can't generate a legal BUILD_VECTOR, exit
10841 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
10844 // Check that we know which element is being inserted
10845 if (!isa<ConstantSDNode>(EltNo))
10847 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
10849 // Canonicalize insert_vector_elt dag nodes.
10851 // (insert_vector_elt (insert_vector_elt A, Idx0), Idx1)
10852 // -> (insert_vector_elt (insert_vector_elt A, Idx1), Idx0)
10854 // Do this only if the child insert_vector node has one use; also
10855 // do this only if indices are both constants and Idx1 < Idx0.
10856 if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT && InVec.hasOneUse()
10857 && isa<ConstantSDNode>(InVec.getOperand(2))) {
10858 unsigned OtherElt =
10859 cast<ConstantSDNode>(InVec.getOperand(2))->getZExtValue();
10860 if (Elt < OtherElt) {
10862 SDValue NewOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(N), VT,
10863 InVec.getOperand(0), InVal, EltNo);
10864 AddToWorklist(NewOp.getNode());
10865 return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(InVec.getNode()),
10866 VT, NewOp, InVec.getOperand(1), InVec.getOperand(2));
10870 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
10871 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
10872 // vector elements.
10873 SmallVector<SDValue, 8> Ops;
10874 // Do not combine these two vectors if the output vector will not replace
10875 // the input vector.
10876 if (InVec.getOpcode() == ISD::BUILD_VECTOR && InVec.hasOneUse()) {
10877 Ops.append(InVec.getNode()->op_begin(),
10878 InVec.getNode()->op_end());
10879 } else if (InVec.getOpcode() == ISD::UNDEF) {
10880 unsigned NElts = VT.getVectorNumElements();
10881 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
10886 // Insert the element
10887 if (Elt < Ops.size()) {
10888 // All the operands of BUILD_VECTOR must have the same type;
10889 // we enforce that here.
10890 EVT OpVT = Ops[0].getValueType();
10891 if (InVal.getValueType() != OpVT)
10892 InVal = OpVT.bitsGT(InVal.getValueType()) ?
10893 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
10894 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
10898 // Return the new vector
10899 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
10902 SDValue DAGCombiner::ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
10903 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad) {
10904 EVT ResultVT = EVE->getValueType(0);
10905 EVT VecEltVT = InVecVT.getVectorElementType();
10906 unsigned Align = OriginalLoad->getAlignment();
10907 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
10908 VecEltVT.getTypeForEVT(*DAG.getContext()));
10910 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VecEltVT))
10915 SDValue NewPtr = OriginalLoad->getBasePtr();
10917 EVT PtrType = NewPtr.getValueType();
10918 MachinePointerInfo MPI;
10919 if (auto *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo)) {
10920 int Elt = ConstEltNo->getZExtValue();
10921 unsigned PtrOff = VecEltVT.getSizeInBits() * Elt / 8;
10922 if (TLI.isBigEndian())
10923 PtrOff = InVecVT.getSizeInBits() / 8 - PtrOff;
10924 Offset = DAG.getConstant(PtrOff, PtrType);
10925 MPI = OriginalLoad->getPointerInfo().getWithOffset(PtrOff);
10927 Offset = DAG.getNode(
10928 ISD::MUL, SDLoc(EVE), EltNo.getValueType(), EltNo,
10929 DAG.getConstant(VecEltVT.getStoreSize(), EltNo.getValueType()));
10930 if (TLI.isBigEndian())
10931 Offset = DAG.getNode(
10932 ISD::SUB, SDLoc(EVE), EltNo.getValueType(),
10933 DAG.getConstant(InVecVT.getStoreSize(), EltNo.getValueType()), Offset);
10934 MPI = OriginalLoad->getPointerInfo();
10936 NewPtr = DAG.getNode(ISD::ADD, SDLoc(EVE), PtrType, NewPtr, Offset);
10938 // The replacement we need to do here is a little tricky: we need to
10939 // replace an extractelement of a load with a load.
10940 // Use ReplaceAllUsesOfValuesWith to do the replacement.
10941 // Note that this replacement assumes that the extractvalue is the only
10942 // use of the load; that's okay because we don't want to perform this
10943 // transformation in other cases anyway.
10946 if (ResultVT.bitsGT(VecEltVT)) {
10947 // If the result type of vextract is wider than the load, then issue an
10948 // extending load instead.
10949 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, ResultVT,
10953 Load = DAG.getExtLoad(
10954 ExtType, SDLoc(EVE), ResultVT, OriginalLoad->getChain(), NewPtr, MPI,
10955 VecEltVT, OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
10956 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
10957 Chain = Load.getValue(1);
10959 Load = DAG.getLoad(
10960 VecEltVT, SDLoc(EVE), OriginalLoad->getChain(), NewPtr, MPI,
10961 OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
10962 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
10963 Chain = Load.getValue(1);
10964 if (ResultVT.bitsLT(VecEltVT))
10965 Load = DAG.getNode(ISD::TRUNCATE, SDLoc(EVE), ResultVT, Load);
10967 Load = DAG.getNode(ISD::BITCAST, SDLoc(EVE), ResultVT, Load);
10969 WorklistRemover DeadNodes(*this);
10970 SDValue From[] = { SDValue(EVE, 0), SDValue(OriginalLoad, 1) };
10971 SDValue To[] = { Load, Chain };
10972 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
10973 // Since we're explicitly calling ReplaceAllUses, add the new node to the
10974 // worklist explicitly as well.
10975 AddToWorklist(Load.getNode());
10976 AddUsersToWorklist(Load.getNode()); // Add users too
10977 // Make sure to revisit this node to clean it up; it will usually be dead.
10978 AddToWorklist(EVE);
10980 return SDValue(EVE, 0);
10983 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
10984 // (vextract (scalar_to_vector val, 0) -> val
10985 SDValue InVec = N->getOperand(0);
10986 EVT VT = InVec.getValueType();
10987 EVT NVT = N->getValueType(0);
10989 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
10990 // Check if the result type doesn't match the inserted element type. A
10991 // SCALAR_TO_VECTOR may truncate the inserted element and the
10992 // EXTRACT_VECTOR_ELT may widen the extracted vector.
10993 SDValue InOp = InVec.getOperand(0);
10994 if (InOp.getValueType() != NVT) {
10995 assert(InOp.getValueType().isInteger() && NVT.isInteger());
10996 return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT);
11001 SDValue EltNo = N->getOperand(1);
11002 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
11004 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
11005 // We only perform this optimization before the op legalization phase because
11006 // we may introduce new vector instructions which are not backed by TD
11007 // patterns. For example on AVX, extracting elements from a wide vector
11008 // without using extract_subvector. However, if we can find an underlying
11009 // scalar value, then we can always use that.
11010 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
11012 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
11013 int NumElem = VT.getVectorNumElements();
11014 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
11015 // Find the new index to extract from.
11016 int OrigElt = SVOp->getMaskElt(Elt);
11018 // Extracting an undef index is undef.
11020 return DAG.getUNDEF(NVT);
11022 // Select the right vector half to extract from.
11024 if (OrigElt < NumElem) {
11025 SVInVec = InVec->getOperand(0);
11027 SVInVec = InVec->getOperand(1);
11028 OrigElt -= NumElem;
11031 if (SVInVec.getOpcode() == ISD::BUILD_VECTOR) {
11032 SDValue InOp = SVInVec.getOperand(OrigElt);
11033 if (InOp.getValueType() != NVT) {
11034 assert(InOp.getValueType().isInteger() && NVT.isInteger());
11035 InOp = DAG.getSExtOrTrunc(InOp, SDLoc(SVInVec), NVT);
11041 // FIXME: We should handle recursing on other vector shuffles and
11042 // scalar_to_vector here as well.
11044 if (!LegalOperations) {
11045 EVT IndexTy = TLI.getVectorIdxTy();
11046 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT,
11047 SVInVec, DAG.getConstant(OrigElt, IndexTy));
11051 bool BCNumEltsChanged = false;
11052 EVT ExtVT = VT.getVectorElementType();
11055 // If the result of load has to be truncated, then it's not necessarily
11057 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
11060 if (InVec.getOpcode() == ISD::BITCAST) {
11061 // Don't duplicate a load with other uses.
11062 if (!InVec.hasOneUse())
11065 EVT BCVT = InVec.getOperand(0).getValueType();
11066 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
11068 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
11069 BCNumEltsChanged = true;
11070 InVec = InVec.getOperand(0);
11071 ExtVT = BCVT.getVectorElementType();
11074 // (vextract (vN[if]M load $addr), i) -> ([if]M load $addr + i * size)
11075 if (!LegalOperations && !ConstEltNo && InVec.hasOneUse() &&
11076 ISD::isNormalLoad(InVec.getNode()) &&
11077 !N->getOperand(1)->hasPredecessor(InVec.getNode())) {
11078 SDValue Index = N->getOperand(1);
11079 if (LoadSDNode *OrigLoad = dyn_cast<LoadSDNode>(InVec))
11080 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, Index,
11084 // Perform only after legalization to ensure build_vector / vector_shuffle
11085 // optimizations have already been done.
11086 if (!LegalOperations) return SDValue();
11088 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
11089 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
11090 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
11093 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
11095 LoadSDNode *LN0 = nullptr;
11096 const ShuffleVectorSDNode *SVN = nullptr;
11097 if (ISD::isNormalLoad(InVec.getNode())) {
11098 LN0 = cast<LoadSDNode>(InVec);
11099 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
11100 InVec.getOperand(0).getValueType() == ExtVT &&
11101 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
11102 // Don't duplicate a load with other uses.
11103 if (!InVec.hasOneUse())
11106 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
11107 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
11108 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
11110 // (load $addr+1*size)
11112 // Don't duplicate a load with other uses.
11113 if (!InVec.hasOneUse())
11116 // If the bit convert changed the number of elements, it is unsafe
11117 // to examine the mask.
11118 if (BCNumEltsChanged)
11121 // Select the input vector, guarding against out of range extract vector.
11122 unsigned NumElems = VT.getVectorNumElements();
11123 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
11124 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
11126 if (InVec.getOpcode() == ISD::BITCAST) {
11127 // Don't duplicate a load with other uses.
11128 if (!InVec.hasOneUse())
11131 InVec = InVec.getOperand(0);
11133 if (ISD::isNormalLoad(InVec.getNode())) {
11134 LN0 = cast<LoadSDNode>(InVec);
11135 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
11136 EltNo = DAG.getConstant(Elt, EltNo.getValueType());
11140 // Make sure we found a non-volatile load and the extractelement is
11142 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
11145 // If Idx was -1 above, Elt is going to be -1, so just return undef.
11147 return DAG.getUNDEF(LVT);
11149 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, EltNo, LN0);
11155 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
11156 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
11157 // We perform this optimization post type-legalization because
11158 // the type-legalizer often scalarizes integer-promoted vectors.
11159 // Performing this optimization before may create bit-casts which
11160 // will be type-legalized to complex code sequences.
11161 // We perform this optimization only before the operation legalizer because we
11162 // may introduce illegal operations.
11163 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
11166 unsigned NumInScalars = N->getNumOperands();
11168 EVT VT = N->getValueType(0);
11170 // Check to see if this is a BUILD_VECTOR of a bunch of values
11171 // which come from any_extend or zero_extend nodes. If so, we can create
11172 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
11173 // optimizations. We do not handle sign-extend because we can't fill the sign
11175 EVT SourceType = MVT::Other;
11176 bool AllAnyExt = true;
11178 for (unsigned i = 0; i != NumInScalars; ++i) {
11179 SDValue In = N->getOperand(i);
11180 // Ignore undef inputs.
11181 if (In.getOpcode() == ISD::UNDEF) continue;
11183 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
11184 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
11186 // Abort if the element is not an extension.
11187 if (!ZeroExt && !AnyExt) {
11188 SourceType = MVT::Other;
11192 // The input is a ZeroExt or AnyExt. Check the original type.
11193 EVT InTy = In.getOperand(0).getValueType();
11195 // Check that all of the widened source types are the same.
11196 if (SourceType == MVT::Other)
11199 else if (InTy != SourceType) {
11200 // Multiple income types. Abort.
11201 SourceType = MVT::Other;
11205 // Check if all of the extends are ANY_EXTENDs.
11206 AllAnyExt &= AnyExt;
11209 // In order to have valid types, all of the inputs must be extended from the
11210 // same source type and all of the inputs must be any or zero extend.
11211 // Scalar sizes must be a power of two.
11212 EVT OutScalarTy = VT.getScalarType();
11213 bool ValidTypes = SourceType != MVT::Other &&
11214 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
11215 isPowerOf2_32(SourceType.getSizeInBits());
11217 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
11218 // turn into a single shuffle instruction.
11222 bool isLE = TLI.isLittleEndian();
11223 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
11224 assert(ElemRatio > 1 && "Invalid element size ratio");
11225 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
11226 DAG.getConstant(0, SourceType);
11228 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
11229 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
11231 // Populate the new build_vector
11232 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
11233 SDValue Cast = N->getOperand(i);
11234 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
11235 Cast.getOpcode() == ISD::ZERO_EXTEND ||
11236 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
11238 if (Cast.getOpcode() == ISD::UNDEF)
11239 In = DAG.getUNDEF(SourceType);
11241 In = Cast->getOperand(0);
11242 unsigned Index = isLE ? (i * ElemRatio) :
11243 (i * ElemRatio + (ElemRatio - 1));
11245 assert(Index < Ops.size() && "Invalid index");
11249 // The type of the new BUILD_VECTOR node.
11250 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
11251 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
11252 "Invalid vector size");
11253 // Check if the new vector type is legal.
11254 if (!isTypeLegal(VecVT)) return SDValue();
11256 // Make the new BUILD_VECTOR.
11257 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
11259 // The new BUILD_VECTOR node has the potential to be further optimized.
11260 AddToWorklist(BV.getNode());
11261 // Bitcast to the desired type.
11262 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
11265 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
11266 EVT VT = N->getValueType(0);
11268 unsigned NumInScalars = N->getNumOperands();
11271 EVT SrcVT = MVT::Other;
11272 unsigned Opcode = ISD::DELETED_NODE;
11273 unsigned NumDefs = 0;
11275 for (unsigned i = 0; i != NumInScalars; ++i) {
11276 SDValue In = N->getOperand(i);
11277 unsigned Opc = In.getOpcode();
11279 if (Opc == ISD::UNDEF)
11282 // If all scalar values are floats and converted from integers.
11283 if (Opcode == ISD::DELETED_NODE &&
11284 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
11291 EVT InVT = In.getOperand(0).getValueType();
11293 // If all scalar values are typed differently, bail out. It's chosen to
11294 // simplify BUILD_VECTOR of integer types.
11295 if (SrcVT == MVT::Other)
11302 // If the vector has just one element defined, it's not worth to fold it into
11303 // a vectorized one.
11307 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
11308 && "Should only handle conversion from integer to float.");
11309 assert(SrcVT != MVT::Other && "Cannot determine source type!");
11311 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
11313 if (!TLI.isOperationLegalOrCustom(Opcode, NVT))
11316 // Just because the floating-point vector type is legal does not necessarily
11317 // mean that the corresponding integer vector type is.
11318 if (!isTypeLegal(NVT))
11321 SmallVector<SDValue, 8> Opnds;
11322 for (unsigned i = 0; i != NumInScalars; ++i) {
11323 SDValue In = N->getOperand(i);
11325 if (In.getOpcode() == ISD::UNDEF)
11326 Opnds.push_back(DAG.getUNDEF(SrcVT));
11328 Opnds.push_back(In.getOperand(0));
11330 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Opnds);
11331 AddToWorklist(BV.getNode());
11333 return DAG.getNode(Opcode, dl, VT, BV);
11336 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
11337 unsigned NumInScalars = N->getNumOperands();
11339 EVT VT = N->getValueType(0);
11341 // A vector built entirely of undefs is undef.
11342 if (ISD::allOperandsUndef(N))
11343 return DAG.getUNDEF(VT);
11345 if (SDValue V = reduceBuildVecExtToExtBuildVec(N))
11348 if (SDValue V = reduceBuildVecConvertToConvertBuildVec(N))
11351 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
11352 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
11353 // at most two distinct vectors, turn this into a shuffle node.
11355 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
11356 if (!isTypeLegal(VT))
11359 // May only combine to shuffle after legalize if shuffle is legal.
11360 if (LegalOperations && !TLI.isOperationLegal(ISD::VECTOR_SHUFFLE, VT))
11363 SDValue VecIn1, VecIn2;
11364 bool UsesZeroVector = false;
11365 for (unsigned i = 0; i != NumInScalars; ++i) {
11366 SDValue Op = N->getOperand(i);
11367 // Ignore undef inputs.
11368 if (Op.getOpcode() == ISD::UNDEF) continue;
11370 // See if we can combine this build_vector into a blend with a zero vector.
11371 if (!VecIn2.getNode() && ((Op.getOpcode() == ISD::Constant &&
11372 cast<ConstantSDNode>(Op.getNode())->isNullValue()) ||
11373 (Op.getOpcode() == ISD::ConstantFP &&
11374 cast<ConstantFPSDNode>(Op.getNode())->getValueAPF().isZero()))) {
11375 UsesZeroVector = true;
11379 // If this input is something other than a EXTRACT_VECTOR_ELT with a
11380 // constant index, bail out.
11381 if (Op.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
11382 !isa<ConstantSDNode>(Op.getOperand(1))) {
11383 VecIn1 = VecIn2 = SDValue(nullptr, 0);
11387 // We allow up to two distinct input vectors.
11388 SDValue ExtractedFromVec = Op.getOperand(0);
11389 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
11392 if (!VecIn1.getNode()) {
11393 VecIn1 = ExtractedFromVec;
11394 } else if (!VecIn2.getNode() && !UsesZeroVector) {
11395 VecIn2 = ExtractedFromVec;
11397 // Too many inputs.
11398 VecIn1 = VecIn2 = SDValue(nullptr, 0);
11403 // If everything is good, we can make a shuffle operation.
11404 if (VecIn1.getNode()) {
11405 unsigned InNumElements = VecIn1.getValueType().getVectorNumElements();
11406 SmallVector<int, 8> Mask;
11407 for (unsigned i = 0; i != NumInScalars; ++i) {
11408 unsigned Opcode = N->getOperand(i).getOpcode();
11409 if (Opcode == ISD::UNDEF) {
11410 Mask.push_back(-1);
11414 // Operands can also be zero.
11415 if (Opcode != ISD::EXTRACT_VECTOR_ELT) {
11416 assert(UsesZeroVector &&
11417 (Opcode == ISD::Constant || Opcode == ISD::ConstantFP) &&
11418 "Unexpected node found!");
11419 Mask.push_back(NumInScalars+i);
11423 // If extracting from the first vector, just use the index directly.
11424 SDValue Extract = N->getOperand(i);
11425 SDValue ExtVal = Extract.getOperand(1);
11426 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
11427 if (Extract.getOperand(0) == VecIn1) {
11428 Mask.push_back(ExtIndex);
11432 // Otherwise, use InIdx + InputVecSize
11433 Mask.push_back(InNumElements + ExtIndex);
11436 // Avoid introducing illegal shuffles with zero.
11437 if (UsesZeroVector && !TLI.isVectorClearMaskLegal(Mask, VT))
11440 // We can't generate a shuffle node with mismatched input and output types.
11441 // Attempt to transform a single input vector to the correct type.
11442 if ((VT != VecIn1.getValueType())) {
11443 // If the input vector type has a different base type to the output
11444 // vector type, bail out.
11445 EVT VTElemType = VT.getVectorElementType();
11446 if ((VecIn1.getValueType().getVectorElementType() != VTElemType) ||
11447 (VecIn2.getNode() &&
11448 (VecIn2.getValueType().getVectorElementType() != VTElemType)))
11451 // If the input vector is too small, widen it.
11452 // We only support widening of vectors which are half the size of the
11453 // output registers. For example XMM->YMM widening on X86 with AVX.
11454 EVT VecInT = VecIn1.getValueType();
11455 if (VecInT.getSizeInBits() * 2 == VT.getSizeInBits()) {
11456 // If we only have one small input, widen it by adding undef values.
11457 if (!VecIn2.getNode())
11458 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, VecIn1,
11459 DAG.getUNDEF(VecIn1.getValueType()));
11460 else if (VecIn1.getValueType() == VecIn2.getValueType()) {
11461 // If we have two small inputs of the same type, try to concat them.
11462 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, VecIn1, VecIn2);
11463 VecIn2 = SDValue(nullptr, 0);
11466 } else if (VecInT.getSizeInBits() == VT.getSizeInBits() * 2) {
11467 // If the input vector is too large, try to split it.
11468 // We don't support having two input vectors that are too large.
11469 // If the zero vector was used, we can not split the vector,
11470 // since we'd need 3 inputs.
11471 if (UsesZeroVector || VecIn2.getNode())
11474 if (!TLI.isExtractSubvectorCheap(VT, VT.getVectorNumElements()))
11477 // Try to replace VecIn1 with two extract_subvectors
11478 // No need to update the masks, they should still be correct.
11479 VecIn2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, VecIn1,
11480 DAG.getConstant(VT.getVectorNumElements(), TLI.getVectorIdxTy()));
11481 VecIn1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, VecIn1,
11482 DAG.getConstant(0, TLI.getVectorIdxTy()));
11487 if (UsesZeroVector)
11488 VecIn2 = VT.isInteger() ? DAG.getConstant(0, VT) :
11489 DAG.getConstantFP(0.0, VT);
11491 // If VecIn2 is unused then change it to undef.
11492 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
11494 // Check that we were able to transform all incoming values to the same
11496 if (VecIn2.getValueType() != VecIn1.getValueType() ||
11497 VecIn1.getValueType() != VT)
11500 // Return the new VECTOR_SHUFFLE node.
11504 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]);
11510 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
11511 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
11512 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
11513 // inputs come from at most two distinct vectors, turn this into a shuffle
11516 // If we only have one input vector, we don't need to do any concatenation.
11517 if (N->getNumOperands() == 1)
11518 return N->getOperand(0);
11520 // Check if all of the operands are undefs.
11521 EVT VT = N->getValueType(0);
11522 if (ISD::allOperandsUndef(N))
11523 return DAG.getUNDEF(VT);
11525 // Optimize concat_vectors where one of the vectors is undef.
11526 if (N->getNumOperands() == 2 &&
11527 N->getOperand(1)->getOpcode() == ISD::UNDEF) {
11528 SDValue In = N->getOperand(0);
11529 assert(In.getValueType().isVector() && "Must concat vectors");
11531 // Transform: concat_vectors(scalar, undef) -> scalar_to_vector(sclr).
11532 if (In->getOpcode() == ISD::BITCAST &&
11533 !In->getOperand(0)->getValueType(0).isVector()) {
11534 SDValue Scalar = In->getOperand(0);
11535 EVT SclTy = Scalar->getValueType(0);
11537 if (!SclTy.isFloatingPoint() && !SclTy.isInteger())
11540 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SclTy,
11541 VT.getSizeInBits() / SclTy.getSizeInBits());
11542 if (!TLI.isTypeLegal(NVT) || !TLI.isTypeLegal(Scalar.getValueType()))
11545 SDLoc dl = SDLoc(N);
11546 SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NVT, Scalar);
11547 return DAG.getNode(ISD::BITCAST, dl, VT, Res);
11551 // Fold any combination of BUILD_VECTOR or UNDEF nodes into one BUILD_VECTOR.
11552 // We have already tested above for an UNDEF only concatenation.
11553 // fold (concat_vectors (BUILD_VECTOR A, B, ...), (BUILD_VECTOR C, D, ...))
11554 // -> (BUILD_VECTOR A, B, ..., C, D, ...)
11555 auto IsBuildVectorOrUndef = [](const SDValue &Op) {
11556 return ISD::UNDEF == Op.getOpcode() || ISD::BUILD_VECTOR == Op.getOpcode();
11558 bool AllBuildVectorsOrUndefs =
11559 std::all_of(N->op_begin(), N->op_end(), IsBuildVectorOrUndef);
11560 if (AllBuildVectorsOrUndefs) {
11561 SmallVector<SDValue, 8> Opnds;
11562 EVT SVT = VT.getScalarType();
11565 if (!SVT.isFloatingPoint()) {
11566 // If BUILD_VECTOR are from built from integer, they may have different
11567 // operand types. Get the smallest type and truncate all operands to it.
11568 bool FoundMinVT = false;
11569 for (const SDValue &Op : N->ops())
11570 if (ISD::BUILD_VECTOR == Op.getOpcode()) {
11571 EVT OpSVT = Op.getOperand(0)->getValueType(0);
11572 MinVT = (!FoundMinVT || OpSVT.bitsLE(MinVT)) ? OpSVT : MinVT;
11575 assert(FoundMinVT && "Concat vector type mismatch");
11578 for (const SDValue &Op : N->ops()) {
11579 EVT OpVT = Op.getValueType();
11580 unsigned NumElts = OpVT.getVectorNumElements();
11582 if (ISD::UNDEF == Op.getOpcode())
11583 Opnds.append(NumElts, DAG.getUNDEF(MinVT));
11585 if (ISD::BUILD_VECTOR == Op.getOpcode()) {
11586 if (SVT.isFloatingPoint()) {
11587 assert(SVT == OpVT.getScalarType() && "Concat vector type mismatch");
11588 Opnds.append(Op->op_begin(), Op->op_begin() + NumElts);
11590 for (unsigned i = 0; i != NumElts; ++i)
11592 DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinVT, Op.getOperand(i)));
11597 assert(VT.getVectorNumElements() == Opnds.size() &&
11598 "Concat vector type mismatch");
11599 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
11602 // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
11603 // nodes often generate nop CONCAT_VECTOR nodes.
11604 // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that
11605 // place the incoming vectors at the exact same location.
11606 SDValue SingleSource = SDValue();
11607 unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements();
11609 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
11610 SDValue Op = N->getOperand(i);
11612 if (Op.getOpcode() == ISD::UNDEF)
11615 // Check if this is the identity extract:
11616 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
11619 // Find the single incoming vector for the extract_subvector.
11620 if (SingleSource.getNode()) {
11621 if (Op.getOperand(0) != SingleSource)
11624 SingleSource = Op.getOperand(0);
11626 // Check the source type is the same as the type of the result.
11627 // If not, this concat may extend the vector, so we can not
11628 // optimize it away.
11629 if (SingleSource.getValueType() != N->getValueType(0))
11633 unsigned IdentityIndex = i * PartNumElem;
11634 ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1));
11635 // The extract index must be constant.
11639 // Check that we are reading from the identity index.
11640 if (CS->getZExtValue() != IdentityIndex)
11644 if (SingleSource.getNode())
11645 return SingleSource;
11650 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
11651 EVT NVT = N->getValueType(0);
11652 SDValue V = N->getOperand(0);
11654 if (V->getOpcode() == ISD::CONCAT_VECTORS) {
11656 // (extract_subvec (concat V1, V2, ...), i)
11659 // Only operand 0 is checked as 'concat' assumes all inputs of the same
11661 if (V->getOperand(0).getValueType() != NVT)
11663 unsigned Idx = dyn_cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
11664 unsigned NumElems = NVT.getVectorNumElements();
11665 assert((Idx % NumElems) == 0 &&
11666 "IDX in concat is not a multiple of the result vector length.");
11667 return V->getOperand(Idx / NumElems);
11671 if (V->getOpcode() == ISD::BITCAST)
11672 V = V.getOperand(0);
11674 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
11676 // Handle only simple case where vector being inserted and vector
11677 // being extracted are of same type, and are half size of larger vectors.
11678 EVT BigVT = V->getOperand(0).getValueType();
11679 EVT SmallVT = V->getOperand(1).getValueType();
11680 if (!NVT.bitsEq(SmallVT) || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
11683 // Only handle cases where both indexes are constants with the same type.
11684 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
11685 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
11687 if (InsIdx && ExtIdx &&
11688 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
11689 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
11691 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
11693 // indices are equal or bit offsets are equal => V1
11694 // otherwise => (extract_subvec V1, ExtIdx)
11695 if (InsIdx->getZExtValue() * SmallVT.getScalarType().getSizeInBits() ==
11696 ExtIdx->getZExtValue() * NVT.getScalarType().getSizeInBits())
11697 return DAG.getNode(ISD::BITCAST, dl, NVT, V->getOperand(1));
11698 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT,
11699 DAG.getNode(ISD::BITCAST, dl,
11700 N->getOperand(0).getValueType(),
11701 V->getOperand(0)), N->getOperand(1));
11708 static SDValue simplifyShuffleOperandRecursively(SmallBitVector &UsedElements,
11709 SDValue V, SelectionDAG &DAG) {
11711 EVT VT = V.getValueType();
11713 switch (V.getOpcode()) {
11717 case ISD::CONCAT_VECTORS: {
11718 EVT OpVT = V->getOperand(0).getValueType();
11719 int OpSize = OpVT.getVectorNumElements();
11720 SmallBitVector OpUsedElements(OpSize, false);
11721 bool FoundSimplification = false;
11722 SmallVector<SDValue, 4> NewOps;
11723 NewOps.reserve(V->getNumOperands());
11724 for (int i = 0, NumOps = V->getNumOperands(); i < NumOps; ++i) {
11725 SDValue Op = V->getOperand(i);
11726 bool OpUsed = false;
11727 for (int j = 0; j < OpSize; ++j)
11728 if (UsedElements[i * OpSize + j]) {
11729 OpUsedElements[j] = true;
11733 OpUsed ? simplifyShuffleOperandRecursively(OpUsedElements, Op, DAG)
11734 : DAG.getUNDEF(OpVT));
11735 FoundSimplification |= Op == NewOps.back();
11736 OpUsedElements.reset();
11738 if (FoundSimplification)
11739 V = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, NewOps);
11743 case ISD::INSERT_SUBVECTOR: {
11744 SDValue BaseV = V->getOperand(0);
11745 SDValue SubV = V->getOperand(1);
11746 auto *IdxN = dyn_cast<ConstantSDNode>(V->getOperand(2));
11750 int SubSize = SubV.getValueType().getVectorNumElements();
11751 int Idx = IdxN->getZExtValue();
11752 bool SubVectorUsed = false;
11753 SmallBitVector SubUsedElements(SubSize, false);
11754 for (int i = 0; i < SubSize; ++i)
11755 if (UsedElements[i + Idx]) {
11756 SubVectorUsed = true;
11757 SubUsedElements[i] = true;
11758 UsedElements[i + Idx] = false;
11761 // Now recurse on both the base and sub vectors.
11762 SDValue SimplifiedSubV =
11764 ? simplifyShuffleOperandRecursively(SubUsedElements, SubV, DAG)
11765 : DAG.getUNDEF(SubV.getValueType());
11766 SDValue SimplifiedBaseV = simplifyShuffleOperandRecursively(UsedElements, BaseV, DAG);
11767 if (SimplifiedSubV != SubV || SimplifiedBaseV != BaseV)
11768 V = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT,
11769 SimplifiedBaseV, SimplifiedSubV, V->getOperand(2));
11775 static SDValue simplifyShuffleOperands(ShuffleVectorSDNode *SVN, SDValue N0,
11776 SDValue N1, SelectionDAG &DAG) {
11777 EVT VT = SVN->getValueType(0);
11778 int NumElts = VT.getVectorNumElements();
11779 SmallBitVector N0UsedElements(NumElts, false), N1UsedElements(NumElts, false);
11780 for (int M : SVN->getMask())
11781 if (M >= 0 && M < NumElts)
11782 N0UsedElements[M] = true;
11783 else if (M >= NumElts)
11784 N1UsedElements[M - NumElts] = true;
11786 SDValue S0 = simplifyShuffleOperandRecursively(N0UsedElements, N0, DAG);
11787 SDValue S1 = simplifyShuffleOperandRecursively(N1UsedElements, N1, DAG);
11788 if (S0 == N0 && S1 == N1)
11791 return DAG.getVectorShuffle(VT, SDLoc(SVN), S0, S1, SVN->getMask());
11794 // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat,
11795 // or turn a shuffle of a single concat into simpler shuffle then concat.
11796 static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
11797 EVT VT = N->getValueType(0);
11798 unsigned NumElts = VT.getVectorNumElements();
11800 SDValue N0 = N->getOperand(0);
11801 SDValue N1 = N->getOperand(1);
11802 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
11804 SmallVector<SDValue, 4> Ops;
11805 EVT ConcatVT = N0.getOperand(0).getValueType();
11806 unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements();
11807 unsigned NumConcats = NumElts / NumElemsPerConcat;
11809 // Special case: shuffle(concat(A,B)) can be more efficiently represented
11810 // as concat(shuffle(A,B),UNDEF) if the shuffle doesn't set any of the high
11811 // half vector elements.
11812 if (NumElemsPerConcat * 2 == NumElts && N1.getOpcode() == ISD::UNDEF &&
11813 std::all_of(SVN->getMask().begin() + NumElemsPerConcat,
11814 SVN->getMask().end(), [](int i) { return i == -1; })) {
11815 N0 = DAG.getVectorShuffle(ConcatVT, SDLoc(N), N0.getOperand(0), N0.getOperand(1),
11816 ArrayRef<int>(SVN->getMask().begin(), NumElemsPerConcat));
11817 N1 = DAG.getUNDEF(ConcatVT);
11818 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, N0, N1);
11821 // Look at every vector that's inserted. We're looking for exact
11822 // subvector-sized copies from a concatenated vector
11823 for (unsigned I = 0; I != NumConcats; ++I) {
11824 // Make sure we're dealing with a copy.
11825 unsigned Begin = I * NumElemsPerConcat;
11826 bool AllUndef = true, NoUndef = true;
11827 for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) {
11828 if (SVN->getMaskElt(J) >= 0)
11835 if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0)
11838 for (unsigned J = 1; J != NumElemsPerConcat; ++J)
11839 if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J))
11842 unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat;
11843 if (FirstElt < N0.getNumOperands())
11844 Ops.push_back(N0.getOperand(FirstElt));
11846 Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands()));
11848 } else if (AllUndef) {
11849 Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType()));
11850 } else { // Mixed with general masks and undefs, can't do optimization.
11855 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops);
11858 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
11859 EVT VT = N->getValueType(0);
11860 unsigned NumElts = VT.getVectorNumElements();
11862 SDValue N0 = N->getOperand(0);
11863 SDValue N1 = N->getOperand(1);
11865 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
11867 // Canonicalize shuffle undef, undef -> undef
11868 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
11869 return DAG.getUNDEF(VT);
11871 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
11873 // Canonicalize shuffle v, v -> v, undef
11875 SmallVector<int, 8> NewMask;
11876 for (unsigned i = 0; i != NumElts; ++i) {
11877 int Idx = SVN->getMaskElt(i);
11878 if (Idx >= (int)NumElts) Idx -= NumElts;
11879 NewMask.push_back(Idx);
11881 return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT),
11885 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
11886 if (N0.getOpcode() == ISD::UNDEF) {
11887 SmallVector<int, 8> NewMask;
11888 for (unsigned i = 0; i != NumElts; ++i) {
11889 int Idx = SVN->getMaskElt(i);
11891 if (Idx >= (int)NumElts)
11894 Idx = -1; // remove reference to lhs
11896 NewMask.push_back(Idx);
11898 return DAG.getVectorShuffle(VT, SDLoc(N), N1, DAG.getUNDEF(VT),
11902 // Remove references to rhs if it is undef
11903 if (N1.getOpcode() == ISD::UNDEF) {
11904 bool Changed = false;
11905 SmallVector<int, 8> NewMask;
11906 for (unsigned i = 0; i != NumElts; ++i) {
11907 int Idx = SVN->getMaskElt(i);
11908 if (Idx >= (int)NumElts) {
11912 NewMask.push_back(Idx);
11915 return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, &NewMask[0]);
11918 // If it is a splat, check if the argument vector is another splat or a
11920 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
11921 SDNode *V = N0.getNode();
11923 // If this is a bit convert that changes the element type of the vector but
11924 // not the number of vector elements, look through it. Be careful not to
11925 // look though conversions that change things like v4f32 to v2f64.
11926 if (V->getOpcode() == ISD::BITCAST) {
11927 SDValue ConvInput = V->getOperand(0);
11928 if (ConvInput.getValueType().isVector() &&
11929 ConvInput.getValueType().getVectorNumElements() == NumElts)
11930 V = ConvInput.getNode();
11933 if (V->getOpcode() == ISD::BUILD_VECTOR) {
11934 assert(V->getNumOperands() == NumElts &&
11935 "BUILD_VECTOR has wrong number of operands");
11937 bool AllSame = true;
11938 for (unsigned i = 0; i != NumElts; ++i) {
11939 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
11940 Base = V->getOperand(i);
11944 // Splat of <u, u, u, u>, return <u, u, u, u>
11945 if (!Base.getNode())
11947 for (unsigned i = 0; i != NumElts; ++i) {
11948 if (V->getOperand(i) != Base) {
11953 // Splat of <x, x, x, x>, return <x, x, x, x>
11957 // Canonicalize any other splat as a build_vector.
11958 const SDValue &Splatted = V->getOperand(SVN->getSplatIndex());
11959 SmallVector<SDValue, 8> Ops(NumElts, Splatted);
11960 SDValue NewBV = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
11961 V->getValueType(0), Ops);
11963 // We may have jumped through bitcasts, so the type of the
11964 // BUILD_VECTOR may not match the type of the shuffle.
11965 if (V->getValueType(0) != VT)
11966 NewBV = DAG.getNode(ISD::BITCAST, SDLoc(N), VT, NewBV);
11971 // There are various patterns used to build up a vector from smaller vectors,
11972 // subvectors, or elements. Scan chains of these and replace unused insertions
11973 // or components with undef.
11974 if (SDValue S = simplifyShuffleOperands(SVN, N0, N1, DAG))
11977 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
11978 Level < AfterLegalizeVectorOps &&
11979 (N1.getOpcode() == ISD::UNDEF ||
11980 (N1.getOpcode() == ISD::CONCAT_VECTORS &&
11981 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) {
11982 SDValue V = partitionShuffleOfConcats(N, DAG);
11988 // If this shuffle only has a single input that is a bitcasted shuffle,
11989 // attempt to merge the 2 shuffles and suitably bitcast the inputs/output
11990 // back to their original types.
11991 if (N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
11992 N1.getOpcode() == ISD::UNDEF && Level < AfterLegalizeVectorOps &&
11993 TLI.isTypeLegal(VT)) {
11995 // Peek through the bitcast only if there is one user.
11997 while (BC0.getOpcode() == ISD::BITCAST) {
11998 if (!BC0.hasOneUse())
12000 BC0 = BC0.getOperand(0);
12003 auto ScaleShuffleMask = [](ArrayRef<int> Mask, int Scale) {
12005 return SmallVector<int, 8>(Mask.begin(), Mask.end());
12007 SmallVector<int, 8> NewMask;
12009 for (int s = 0; s != Scale; ++s)
12010 NewMask.push_back(M < 0 ? -1 : Scale * M + s);
12014 if (BC0.getOpcode() == ISD::VECTOR_SHUFFLE && BC0.hasOneUse()) {
12015 EVT SVT = VT.getScalarType();
12016 EVT InnerVT = BC0->getValueType(0);
12017 EVT InnerSVT = InnerVT.getScalarType();
12019 // Determine which shuffle works with the smaller scalar type.
12020 EVT ScaleVT = SVT.bitsLT(InnerSVT) ? VT : InnerVT;
12021 EVT ScaleSVT = ScaleVT.getScalarType();
12023 if (TLI.isTypeLegal(ScaleVT) &&
12024 0 == (InnerSVT.getSizeInBits() % ScaleSVT.getSizeInBits()) &&
12025 0 == (SVT.getSizeInBits() % ScaleSVT.getSizeInBits())) {
12027 int InnerScale = InnerSVT.getSizeInBits() / ScaleSVT.getSizeInBits();
12028 int OuterScale = SVT.getSizeInBits() / ScaleSVT.getSizeInBits();
12030 // Scale the shuffle masks to the smaller scalar type.
12031 ShuffleVectorSDNode *InnerSVN = cast<ShuffleVectorSDNode>(BC0);
12032 SmallVector<int, 8> InnerMask =
12033 ScaleShuffleMask(InnerSVN->getMask(), InnerScale);
12034 SmallVector<int, 8> OuterMask =
12035 ScaleShuffleMask(SVN->getMask(), OuterScale);
12037 // Merge the shuffle masks.
12038 SmallVector<int, 8> NewMask;
12039 for (int M : OuterMask)
12040 NewMask.push_back(M < 0 ? -1 : InnerMask[M]);
12042 // Test for shuffle mask legality over both commutations.
12043 SDValue SV0 = BC0->getOperand(0);
12044 SDValue SV1 = BC0->getOperand(1);
12045 bool LegalMask = TLI.isShuffleMaskLegal(NewMask, ScaleVT);
12047 std::swap(SV0, SV1);
12048 ShuffleVectorSDNode::commuteMask(NewMask);
12049 LegalMask = TLI.isShuffleMaskLegal(NewMask, ScaleVT);
12053 SV0 = DAG.getNode(ISD::BITCAST, SDLoc(N), ScaleVT, SV0);
12054 SV1 = DAG.getNode(ISD::BITCAST, SDLoc(N), ScaleVT, SV1);
12055 return DAG.getNode(
12056 ISD::BITCAST, SDLoc(N), VT,
12057 DAG.getVectorShuffle(ScaleVT, SDLoc(N), SV0, SV1, NewMask));
12063 // Canonicalize shuffles according to rules:
12064 // shuffle(A, shuffle(A, B)) -> shuffle(shuffle(A,B), A)
12065 // shuffle(B, shuffle(A, B)) -> shuffle(shuffle(A,B), B)
12066 // shuffle(B, shuffle(A, Undef)) -> shuffle(shuffle(A, Undef), B)
12067 if (N1.getOpcode() == ISD::VECTOR_SHUFFLE &&
12068 N0.getOpcode() != ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
12069 TLI.isTypeLegal(VT)) {
12070 // The incoming shuffle must be of the same type as the result of the
12071 // current shuffle.
12072 assert(N1->getOperand(0).getValueType() == VT &&
12073 "Shuffle types don't match");
12075 SDValue SV0 = N1->getOperand(0);
12076 SDValue SV1 = N1->getOperand(1);
12077 bool HasSameOp0 = N0 == SV0;
12078 bool IsSV1Undef = SV1.getOpcode() == ISD::UNDEF;
12079 if (HasSameOp0 || IsSV1Undef || N0 == SV1)
12080 // Commute the operands of this shuffle so that next rule
12082 return DAG.getCommutedVectorShuffle(*SVN);
12085 // Try to fold according to rules:
12086 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2)
12087 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2)
12088 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2)
12089 // Don't try to fold shuffles with illegal type.
12090 // Only fold if this shuffle is the only user of the other shuffle.
12091 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && N->isOnlyUserOf(N0.getNode()) &&
12092 Level < AfterLegalizeDAG && TLI.isTypeLegal(VT)) {
12093 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
12095 // The incoming shuffle must be of the same type as the result of the
12096 // current shuffle.
12097 assert(OtherSV->getOperand(0).getValueType() == VT &&
12098 "Shuffle types don't match");
12101 SmallVector<int, 4> Mask;
12102 // Compute the combined shuffle mask for a shuffle with SV0 as the first
12103 // operand, and SV1 as the second operand.
12104 for (unsigned i = 0; i != NumElts; ++i) {
12105 int Idx = SVN->getMaskElt(i);
12107 // Propagate Undef.
12108 Mask.push_back(Idx);
12112 SDValue CurrentVec;
12113 if (Idx < (int)NumElts) {
12114 // This shuffle index refers to the inner shuffle N0. Lookup the inner
12115 // shuffle mask to identify which vector is actually referenced.
12116 Idx = OtherSV->getMaskElt(Idx);
12118 // Propagate Undef.
12119 Mask.push_back(Idx);
12123 CurrentVec = (Idx < (int) NumElts) ? OtherSV->getOperand(0)
12124 : OtherSV->getOperand(1);
12126 // This shuffle index references an element within N1.
12130 // Simple case where 'CurrentVec' is UNDEF.
12131 if (CurrentVec.getOpcode() == ISD::UNDEF) {
12132 Mask.push_back(-1);
12136 // Canonicalize the shuffle index. We don't know yet if CurrentVec
12137 // will be the first or second operand of the combined shuffle.
12138 Idx = Idx % NumElts;
12139 if (!SV0.getNode() || SV0 == CurrentVec) {
12140 // Ok. CurrentVec is the left hand side.
12141 // Update the mask accordingly.
12143 Mask.push_back(Idx);
12147 // Bail out if we cannot convert the shuffle pair into a single shuffle.
12148 if (SV1.getNode() && SV1 != CurrentVec)
12151 // Ok. CurrentVec is the right hand side.
12152 // Update the mask accordingly.
12154 Mask.push_back(Idx + NumElts);
12157 // Check if all indices in Mask are Undef. In case, propagate Undef.
12158 bool isUndefMask = true;
12159 for (unsigned i = 0; i != NumElts && isUndefMask; ++i)
12160 isUndefMask &= Mask[i] < 0;
12163 return DAG.getUNDEF(VT);
12165 if (!SV0.getNode())
12166 SV0 = DAG.getUNDEF(VT);
12167 if (!SV1.getNode())
12168 SV1 = DAG.getUNDEF(VT);
12170 // Avoid introducing shuffles with illegal mask.
12171 if (!TLI.isShuffleMaskLegal(Mask, VT)) {
12172 ShuffleVectorSDNode::commuteMask(Mask);
12174 if (!TLI.isShuffleMaskLegal(Mask, VT))
12177 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, A, M2)
12178 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, A, M2)
12179 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, B, M2)
12180 std::swap(SV0, SV1);
12183 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2)
12184 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2)
12185 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2)
12186 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, &Mask[0]);
12192 SDValue DAGCombiner::visitSCALAR_TO_VECTOR(SDNode *N) {
12193 SDValue InVal = N->getOperand(0);
12194 EVT VT = N->getValueType(0);
12196 // Replace a SCALAR_TO_VECTOR(EXTRACT_VECTOR_ELT(V,C0)) pattern
12197 // with a VECTOR_SHUFFLE.
12198 if (InVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
12199 SDValue InVec = InVal->getOperand(0);
12200 SDValue EltNo = InVal->getOperand(1);
12202 // FIXME: We could support implicit truncation if the shuffle can be
12203 // scaled to a smaller vector scalar type.
12204 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(EltNo);
12205 if (C0 && VT == InVec.getValueType() &&
12206 VT.getScalarType() == InVal.getValueType()) {
12207 SmallVector<int, 8> NewMask(VT.getVectorNumElements(), -1);
12208 int Elt = C0->getZExtValue();
12211 if (TLI.isShuffleMaskLegal(NewMask, VT))
12212 return DAG.getVectorShuffle(VT, SDLoc(N), InVec, DAG.getUNDEF(VT),
12220 SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
12221 SDValue N0 = N->getOperand(0);
12222 SDValue N2 = N->getOperand(2);
12224 // If the input vector is a concatenation, and the insert replaces
12225 // one of the halves, we can optimize into a single concat_vectors.
12226 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
12227 N0->getNumOperands() == 2 && N2.getOpcode() == ISD::Constant) {
12228 APInt InsIdx = cast<ConstantSDNode>(N2)->getAPIntValue();
12229 EVT VT = N->getValueType(0);
12231 // Lower half: fold (insert_subvector (concat_vectors X, Y), Z) ->
12232 // (concat_vectors Z, Y)
12234 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
12235 N->getOperand(1), N0.getOperand(1));
12237 // Upper half: fold (insert_subvector (concat_vectors X, Y), Z) ->
12238 // (concat_vectors X, Z)
12239 if (InsIdx == VT.getVectorNumElements()/2)
12240 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
12241 N0.getOperand(0), N->getOperand(1));
12247 /// Returns a vector_shuffle if it able to transform an AND to a vector_shuffle
12248 /// with the destination vector and a zero vector.
12249 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
12250 /// vector_shuffle V, Zero, <0, 4, 2, 4>
12251 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
12252 EVT VT = N->getValueType(0);
12253 SDValue LHS = N->getOperand(0);
12254 SDValue RHS = N->getOperand(1);
12257 // Make sure we're not running after operation legalization where it
12258 // may have custom lowered the vector shuffles.
12259 if (LegalOperations)
12262 if (N->getOpcode() != ISD::AND)
12265 if (RHS.getOpcode() == ISD::BITCAST)
12266 RHS = RHS.getOperand(0);
12268 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
12269 SmallVector<int, 8> Indices;
12270 unsigned NumElts = RHS.getNumOperands();
12272 for (unsigned i = 0; i != NumElts; ++i) {
12273 SDValue Elt = RHS.getOperand(i);
12274 if (!isa<ConstantSDNode>(Elt))
12277 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
12278 Indices.push_back(i);
12279 else if (cast<ConstantSDNode>(Elt)->isNullValue())
12280 Indices.push_back(NumElts+i);
12285 // Let's see if the target supports this vector_shuffle.
12286 EVT RVT = RHS.getValueType();
12287 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
12290 // Return the new VECTOR_SHUFFLE node.
12291 EVT EltVT = RVT.getVectorElementType();
12292 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
12293 DAG.getConstant(0, EltVT));
12294 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), RVT, ZeroOps);
12295 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
12296 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
12297 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
12303 /// Visit a binary vector operation, like ADD.
12304 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
12305 assert(N->getValueType(0).isVector() &&
12306 "SimplifyVBinOp only works on vectors!");
12308 SDValue LHS = N->getOperand(0);
12309 SDValue RHS = N->getOperand(1);
12311 if (SDValue Shuffle = XformToShuffleWithZero(N))
12314 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
12316 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
12317 RHS.getOpcode() == ISD::BUILD_VECTOR) {
12318 // Check if both vectors are constants. If not bail out.
12319 if (!(cast<BuildVectorSDNode>(LHS)->isConstant() &&
12320 cast<BuildVectorSDNode>(RHS)->isConstant()))
12323 SmallVector<SDValue, 8> Ops;
12324 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
12325 SDValue LHSOp = LHS.getOperand(i);
12326 SDValue RHSOp = RHS.getOperand(i);
12328 // Can't fold divide by zero.
12329 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
12330 N->getOpcode() == ISD::FDIV) {
12331 if ((RHSOp.getOpcode() == ISD::Constant &&
12332 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
12333 (RHSOp.getOpcode() == ISD::ConstantFP &&
12334 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
12338 EVT VT = LHSOp.getValueType();
12339 EVT RVT = RHSOp.getValueType();
12341 // Integer BUILD_VECTOR operands may have types larger than the element
12342 // size (e.g., when the element type is not legal). Prior to type
12343 // legalization, the types may not match between the two BUILD_VECTORS.
12344 // Truncate one of the operands to make them match.
12345 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
12346 RHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, RHSOp);
12348 LHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), RVT, LHSOp);
12352 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(LHS), VT,
12354 if (FoldOp.getOpcode() != ISD::UNDEF &&
12355 FoldOp.getOpcode() != ISD::Constant &&
12356 FoldOp.getOpcode() != ISD::ConstantFP)
12358 Ops.push_back(FoldOp);
12359 AddToWorklist(FoldOp.getNode());
12362 if (Ops.size() == LHS.getNumOperands())
12363 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), LHS.getValueType(), Ops);
12366 // Type legalization might introduce new shuffles in the DAG.
12367 // Fold (VBinOp (shuffle (A, Undef, Mask)), (shuffle (B, Undef, Mask)))
12368 // -> (shuffle (VBinOp (A, B)), Undef, Mask).
12369 if (LegalTypes && isa<ShuffleVectorSDNode>(LHS) &&
12370 isa<ShuffleVectorSDNode>(RHS) && LHS.hasOneUse() && RHS.hasOneUse() &&
12371 LHS.getOperand(1).getOpcode() == ISD::UNDEF &&
12372 RHS.getOperand(1).getOpcode() == ISD::UNDEF) {
12373 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(LHS);
12374 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(RHS);
12376 if (SVN0->getMask().equals(SVN1->getMask())) {
12377 EVT VT = N->getValueType(0);
12378 SDValue UndefVector = LHS.getOperand(1);
12379 SDValue NewBinOp = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
12380 LHS.getOperand(0), RHS.getOperand(0));
12381 AddUsersToWorklist(N);
12382 return DAG.getVectorShuffle(VT, SDLoc(N), NewBinOp, UndefVector,
12383 &SVN0->getMask()[0]);
12390 SDValue DAGCombiner::SimplifySelect(SDLoc DL, SDValue N0,
12391 SDValue N1, SDValue N2){
12392 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
12394 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
12395 cast<CondCodeSDNode>(N0.getOperand(2))->get());
12397 // If we got a simplified select_cc node back from SimplifySelectCC, then
12398 // break it down into a new SETCC node, and a new SELECT node, and then return
12399 // the SELECT node, since we were called with a SELECT node.
12400 if (SCC.getNode()) {
12401 // Check to see if we got a select_cc back (to turn into setcc/select).
12402 // Otherwise, just return whatever node we got back, like fabs.
12403 if (SCC.getOpcode() == ISD::SELECT_CC) {
12404 SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0),
12406 SCC.getOperand(0), SCC.getOperand(1),
12407 SCC.getOperand(4));
12408 AddToWorklist(SETCC.getNode());
12409 return DAG.getSelect(SDLoc(SCC), SCC.getValueType(), SETCC,
12410 SCC.getOperand(2), SCC.getOperand(3));
12418 /// Given a SELECT or a SELECT_CC node, where LHS and RHS are the two values
12419 /// being selected between, see if we can simplify the select. Callers of this
12420 /// should assume that TheSelect is deleted if this returns true. As such, they
12421 /// should return the appropriate thing (e.g. the node) back to the top-level of
12422 /// the DAG combiner loop to avoid it being looked at.
12423 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
12426 // Cannot simplify select with vector condition
12427 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
12429 // If this is a select from two identical things, try to pull the operation
12430 // through the select.
12431 if (LHS.getOpcode() != RHS.getOpcode() ||
12432 !LHS.hasOneUse() || !RHS.hasOneUse())
12435 // If this is a load and the token chain is identical, replace the select
12436 // of two loads with a load through a select of the address to load from.
12437 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
12438 // constants have been dropped into the constant pool.
12439 if (LHS.getOpcode() == ISD::LOAD) {
12440 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
12441 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
12443 // Token chains must be identical.
12444 if (LHS.getOperand(0) != RHS.getOperand(0) ||
12445 // Do not let this transformation reduce the number of volatile loads.
12446 LLD->isVolatile() || RLD->isVolatile() ||
12447 // If this is an EXTLOAD, the VT's must match.
12448 LLD->getMemoryVT() != RLD->getMemoryVT() ||
12449 // If this is an EXTLOAD, the kind of extension must match.
12450 (LLD->getExtensionType() != RLD->getExtensionType() &&
12451 // The only exception is if one of the extensions is anyext.
12452 LLD->getExtensionType() != ISD::EXTLOAD &&
12453 RLD->getExtensionType() != ISD::EXTLOAD) ||
12454 // FIXME: this discards src value information. This is
12455 // over-conservative. It would be beneficial to be able to remember
12456 // both potential memory locations. Since we are discarding
12457 // src value info, don't do the transformation if the memory
12458 // locations are not in the default address space.
12459 LLD->getPointerInfo().getAddrSpace() != 0 ||
12460 RLD->getPointerInfo().getAddrSpace() != 0 ||
12461 !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(),
12462 LLD->getBasePtr().getValueType()))
12465 // Check that the select condition doesn't reach either load. If so,
12466 // folding this will induce a cycle into the DAG. If not, this is safe to
12467 // xform, so create a select of the addresses.
12469 if (TheSelect->getOpcode() == ISD::SELECT) {
12470 SDNode *CondNode = TheSelect->getOperand(0).getNode();
12471 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
12472 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
12474 // The loads must not depend on one another.
12475 if (LLD->isPredecessorOf(RLD) ||
12476 RLD->isPredecessorOf(LLD))
12478 Addr = DAG.getSelect(SDLoc(TheSelect),
12479 LLD->getBasePtr().getValueType(),
12480 TheSelect->getOperand(0), LLD->getBasePtr(),
12481 RLD->getBasePtr());
12482 } else { // Otherwise SELECT_CC
12483 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
12484 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
12486 if ((LLD->hasAnyUseOfValue(1) &&
12487 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
12488 (RLD->hasAnyUseOfValue(1) &&
12489 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
12492 Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect),
12493 LLD->getBasePtr().getValueType(),
12494 TheSelect->getOperand(0),
12495 TheSelect->getOperand(1),
12496 LLD->getBasePtr(), RLD->getBasePtr(),
12497 TheSelect->getOperand(4));
12501 // It is safe to replace the two loads if they have different alignments,
12502 // but the new load must be the minimum (most restrictive) alignment of the
12504 bool isInvariant = LLD->isInvariant() & RLD->isInvariant();
12505 unsigned Alignment = std::min(LLD->getAlignment(), RLD->getAlignment());
12506 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
12507 Load = DAG.getLoad(TheSelect->getValueType(0),
12509 // FIXME: Discards pointer and AA info.
12510 LLD->getChain(), Addr, MachinePointerInfo(),
12511 LLD->isVolatile(), LLD->isNonTemporal(),
12512 isInvariant, Alignment);
12514 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
12515 RLD->getExtensionType() : LLD->getExtensionType(),
12517 TheSelect->getValueType(0),
12518 // FIXME: Discards pointer and AA info.
12519 LLD->getChain(), Addr, MachinePointerInfo(),
12520 LLD->getMemoryVT(), LLD->isVolatile(),
12521 LLD->isNonTemporal(), isInvariant, Alignment);
12524 // Users of the select now use the result of the load.
12525 CombineTo(TheSelect, Load);
12527 // Users of the old loads now use the new load's chain. We know the
12528 // old-load value is dead now.
12529 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
12530 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
12537 /// Simplify an expression of the form (N0 cond N1) ? N2 : N3
12538 /// where 'cond' is the comparison specified by CC.
12539 SDValue DAGCombiner::SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1,
12540 SDValue N2, SDValue N3,
12541 ISD::CondCode CC, bool NotExtCompare) {
12542 // (x ? y : y) -> y.
12543 if (N2 == N3) return N2;
12545 EVT VT = N2.getValueType();
12546 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
12547 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
12548 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
12550 // Determine if the condition we're dealing with is constant
12551 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
12552 N0, N1, CC, DL, false);
12553 if (SCC.getNode()) AddToWorklist(SCC.getNode());
12554 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
12556 // fold select_cc true, x, y -> x
12557 if (SCCC && !SCCC->isNullValue())
12559 // fold select_cc false, x, y -> y
12560 if (SCCC && SCCC->isNullValue())
12563 // Check to see if we can simplify the select into an fabs node
12564 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
12565 // Allow either -0.0 or 0.0
12566 if (CFP->getValueAPF().isZero()) {
12567 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
12568 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
12569 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
12570 N2 == N3.getOperand(0))
12571 return DAG.getNode(ISD::FABS, DL, VT, N0);
12573 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
12574 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
12575 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
12576 N2.getOperand(0) == N3)
12577 return DAG.getNode(ISD::FABS, DL, VT, N3);
12581 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
12582 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
12583 // in it. This is a win when the constant is not otherwise available because
12584 // it replaces two constant pool loads with one. We only do this if the FP
12585 // type is known to be legal, because if it isn't, then we are before legalize
12586 // types an we want the other legalization to happen first (e.g. to avoid
12587 // messing with soft float) and if the ConstantFP is not legal, because if
12588 // it is legal, we may not need to store the FP constant in a constant pool.
12589 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
12590 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
12591 if (TLI.isTypeLegal(N2.getValueType()) &&
12592 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
12593 TargetLowering::Legal &&
12594 !TLI.isFPImmLegal(TV->getValueAPF(), TV->getValueType(0)) &&
12595 !TLI.isFPImmLegal(FV->getValueAPF(), FV->getValueType(0))) &&
12596 // If both constants have multiple uses, then we won't need to do an
12597 // extra load, they are likely around in registers for other users.
12598 (TV->hasOneUse() || FV->hasOneUse())) {
12599 Constant *Elts[] = {
12600 const_cast<ConstantFP*>(FV->getConstantFPValue()),
12601 const_cast<ConstantFP*>(TV->getConstantFPValue())
12603 Type *FPTy = Elts[0]->getType();
12604 const DataLayout &TD = *TLI.getDataLayout();
12606 // Create a ConstantArray of the two constants.
12607 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
12608 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
12609 TD.getPrefTypeAlignment(FPTy));
12610 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
12612 // Get the offsets to the 0 and 1 element of the array so that we can
12613 // select between them.
12614 SDValue Zero = DAG.getIntPtrConstant(0);
12615 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
12616 SDValue One = DAG.getIntPtrConstant(EltSize);
12618 SDValue Cond = DAG.getSetCC(DL,
12619 getSetCCResultType(N0.getValueType()),
12621 AddToWorklist(Cond.getNode());
12622 SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(),
12624 AddToWorklist(CstOffset.getNode());
12625 CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx,
12627 AddToWorklist(CPIdx.getNode());
12628 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
12629 MachinePointerInfo::getConstantPool(), false,
12630 false, false, Alignment);
12635 // Check to see if we can perform the "gzip trick", transforming
12636 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
12637 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
12638 (N1C->isNullValue() || // (a < 0) ? b : 0
12639 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
12640 EVT XType = N0.getValueType();
12641 EVT AType = N2.getValueType();
12642 if (XType.bitsGE(AType)) {
12643 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
12644 // single-bit constant.
12645 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
12646 unsigned ShCtV = N2C->getAPIntValue().logBase2();
12647 ShCtV = XType.getSizeInBits()-ShCtV-1;
12648 SDValue ShCt = DAG.getConstant(ShCtV,
12649 getShiftAmountTy(N0.getValueType()));
12650 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0),
12652 AddToWorklist(Shift.getNode());
12654 if (XType.bitsGT(AType)) {
12655 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
12656 AddToWorklist(Shift.getNode());
12659 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
12662 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0),
12664 DAG.getConstant(XType.getSizeInBits()-1,
12665 getShiftAmountTy(N0.getValueType())));
12666 AddToWorklist(Shift.getNode());
12668 if (XType.bitsGT(AType)) {
12669 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
12670 AddToWorklist(Shift.getNode());
12673 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
12677 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
12678 // where y is has a single bit set.
12679 // A plaintext description would be, we can turn the SELECT_CC into an AND
12680 // when the condition can be materialized as an all-ones register. Any
12681 // single bit-test can be materialized as an all-ones register with
12682 // shift-left and shift-right-arith.
12683 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
12684 N0->getValueType(0) == VT &&
12685 N1C && N1C->isNullValue() &&
12686 N2C && N2C->isNullValue()) {
12687 SDValue AndLHS = N0->getOperand(0);
12688 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
12689 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
12690 // Shift the tested bit over the sign bit.
12691 APInt AndMask = ConstAndRHS->getAPIntValue();
12693 DAG.getConstant(AndMask.countLeadingZeros(),
12694 getShiftAmountTy(AndLHS.getValueType()));
12695 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt);
12697 // Now arithmetic right shift it all the way over, so the result is either
12698 // all-ones, or zero.
12700 DAG.getConstant(AndMask.getBitWidth()-1,
12701 getShiftAmountTy(Shl.getValueType()));
12702 SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt);
12704 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
12708 // fold select C, 16, 0 -> shl C, 4
12709 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
12710 TLI.getBooleanContents(N0.getValueType()) ==
12711 TargetLowering::ZeroOrOneBooleanContent) {
12713 // If the caller doesn't want us to simplify this into a zext of a compare,
12715 if (NotExtCompare && N2C->getAPIntValue() == 1)
12718 // Get a SetCC of the condition
12719 // NOTE: Don't create a SETCC if it's not legal on this target.
12720 if (!LegalOperations ||
12721 TLI.isOperationLegal(ISD::SETCC,
12722 LegalTypes ? getSetCCResultType(N0.getValueType()) : MVT::i1)) {
12724 // cast from setcc result type to select result type
12726 SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()),
12728 if (N2.getValueType().bitsLT(SCC.getValueType()))
12729 Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2),
12730 N2.getValueType());
12732 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
12733 N2.getValueType(), SCC);
12735 SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC);
12736 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
12737 N2.getValueType(), SCC);
12740 AddToWorklist(SCC.getNode());
12741 AddToWorklist(Temp.getNode());
12743 if (N2C->getAPIntValue() == 1)
12746 // shl setcc result by log2 n2c
12747 return DAG.getNode(
12748 ISD::SHL, DL, N2.getValueType(), Temp,
12749 DAG.getConstant(N2C->getAPIntValue().logBase2(),
12750 getShiftAmountTy(Temp.getValueType())));
12754 // Check to see if this is the equivalent of setcc
12755 // FIXME: Turn all of these into setcc if setcc if setcc is legal
12756 // otherwise, go ahead with the folds.
12757 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
12758 EVT XType = N0.getValueType();
12759 if (!LegalOperations ||
12760 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(XType))) {
12761 SDValue Res = DAG.getSetCC(DL, getSetCCResultType(XType), N0, N1, CC);
12762 if (Res.getValueType() != VT)
12763 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
12767 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
12768 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
12769 (!LegalOperations ||
12770 TLI.isOperationLegal(ISD::CTLZ, XType))) {
12771 SDValue Ctlz = DAG.getNode(ISD::CTLZ, SDLoc(N0), XType, N0);
12772 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
12773 DAG.getConstant(Log2_32(XType.getSizeInBits()),
12774 getShiftAmountTy(Ctlz.getValueType())));
12776 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
12777 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
12778 SDValue NegN0 = DAG.getNode(ISD::SUB, SDLoc(N0),
12779 XType, DAG.getConstant(0, XType), N0);
12780 SDValue NotN0 = DAG.getNOT(SDLoc(N0), N0, XType);
12781 return DAG.getNode(ISD::SRL, DL, XType,
12782 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
12783 DAG.getConstant(XType.getSizeInBits()-1,
12784 getShiftAmountTy(XType)));
12786 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
12787 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
12788 SDValue Sign = DAG.getNode(ISD::SRL, SDLoc(N0), XType, N0,
12789 DAG.getConstant(XType.getSizeInBits()-1,
12790 getShiftAmountTy(N0.getValueType())));
12791 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
12795 // Check to see if this is an integer abs.
12796 // select_cc setg[te] X, 0, X, -X ->
12797 // select_cc setgt X, -1, X, -X ->
12798 // select_cc setl[te] X, 0, -X, X ->
12799 // select_cc setlt X, 1, -X, X ->
12800 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
12802 ConstantSDNode *SubC = nullptr;
12803 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
12804 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
12805 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
12806 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
12807 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
12808 (N1C->isOne() && CC == ISD::SETLT)) &&
12809 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
12810 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
12812 EVT XType = N0.getValueType();
12813 if (SubC && SubC->isNullValue() && XType.isInteger()) {
12814 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0), XType,
12816 DAG.getConstant(XType.getSizeInBits()-1,
12817 getShiftAmountTy(N0.getValueType())));
12818 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0),
12820 AddToWorklist(Shift.getNode());
12821 AddToWorklist(Add.getNode());
12822 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
12829 /// This is a stub for TargetLowering::SimplifySetCC.
12830 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
12831 SDValue N1, ISD::CondCode Cond,
12832 SDLoc DL, bool foldBooleans) {
12833 TargetLowering::DAGCombinerInfo
12834 DagCombineInfo(DAG, Level, false, this);
12835 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
12838 /// Given an ISD::SDIV node expressing a divide by constant, return
12839 /// a DAG expression to select that will generate the same value by multiplying
12840 /// by a magic number.
12841 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
12842 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
12843 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
12847 // Avoid division by zero.
12848 if (!C->getAPIntValue())
12851 std::vector<SDNode*> Built;
12853 TLI.BuildSDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
12855 for (SDNode *N : Built)
12860 /// Given an ISD::SDIV node expressing a divide by constant power of 2, return a
12861 /// DAG expression that will generate the same value by right shifting.
12862 SDValue DAGCombiner::BuildSDIVPow2(SDNode *N) {
12863 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
12867 // Avoid division by zero.
12868 if (!C->getAPIntValue())
12871 std::vector<SDNode *> Built;
12872 SDValue S = TLI.BuildSDIVPow2(N, C->getAPIntValue(), DAG, &Built);
12874 for (SDNode *N : Built)
12879 /// Given an ISD::UDIV node expressing a divide by constant, return a DAG
12880 /// expression that will generate the same value by multiplying by a magic
12882 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
12883 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
12884 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
12888 // Avoid division by zero.
12889 if (!C->getAPIntValue())
12892 std::vector<SDNode*> Built;
12894 TLI.BuildUDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
12896 for (SDNode *N : Built)
12901 SDValue DAGCombiner::BuildReciprocalEstimate(SDValue Op) {
12902 if (Level >= AfterLegalizeDAG)
12905 // Expose the DAG combiner to the target combiner implementations.
12906 TargetLowering::DAGCombinerInfo DCI(DAG, Level, false, this);
12908 unsigned Iterations = 0;
12909 if (SDValue Est = TLI.getRecipEstimate(Op, DCI, Iterations)) {
12911 // Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
12912 // For the reciprocal, we need to find the zero of the function:
12913 // F(X) = A X - 1 [which has a zero at X = 1/A]
12915 // X_{i+1} = X_i (2 - A X_i) = X_i + X_i (1 - A X_i) [this second form
12916 // does not require additional intermediate precision]
12917 EVT VT = Op.getValueType();
12919 SDValue FPOne = DAG.getConstantFP(1.0, VT);
12921 AddToWorklist(Est.getNode());
12923 // Newton iterations: Est = Est + Est (1 - Arg * Est)
12924 for (unsigned i = 0; i < Iterations; ++i) {
12925 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Op, Est);
12926 AddToWorklist(NewEst.getNode());
12928 NewEst = DAG.getNode(ISD::FSUB, DL, VT, FPOne, NewEst);
12929 AddToWorklist(NewEst.getNode());
12931 NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst);
12932 AddToWorklist(NewEst.getNode());
12934 Est = DAG.getNode(ISD::FADD, DL, VT, Est, NewEst);
12935 AddToWorklist(Est.getNode());
12944 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
12945 /// For the reciprocal sqrt, we need to find the zero of the function:
12946 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
12948 /// X_{i+1} = X_i (1.5 - A X_i^2 / 2)
12949 /// As a result, we precompute A/2 prior to the iteration loop.
12950 SDValue DAGCombiner::BuildRsqrtNROneConst(SDValue Arg, SDValue Est,
12951 unsigned Iterations) {
12952 EVT VT = Arg.getValueType();
12954 SDValue ThreeHalves = DAG.getConstantFP(1.5, VT);
12956 // We now need 0.5 * Arg which we can write as (1.5 * Arg - Arg) so that
12957 // this entire sequence requires only one FP constant.
12958 SDValue HalfArg = DAG.getNode(ISD::FMUL, DL, VT, ThreeHalves, Arg);
12959 AddToWorklist(HalfArg.getNode());
12961 HalfArg = DAG.getNode(ISD::FSUB, DL, VT, HalfArg, Arg);
12962 AddToWorklist(HalfArg.getNode());
12964 // Newton iterations: Est = Est * (1.5 - HalfArg * Est * Est)
12965 for (unsigned i = 0; i < Iterations; ++i) {
12966 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, Est);
12967 AddToWorklist(NewEst.getNode());
12969 NewEst = DAG.getNode(ISD::FMUL, DL, VT, HalfArg, NewEst);
12970 AddToWorklist(NewEst.getNode());
12972 NewEst = DAG.getNode(ISD::FSUB, DL, VT, ThreeHalves, NewEst);
12973 AddToWorklist(NewEst.getNode());
12975 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst);
12976 AddToWorklist(Est.getNode());
12981 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
12982 /// For the reciprocal sqrt, we need to find the zero of the function:
12983 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
12985 /// X_{i+1} = (-0.5 * X_i) * (A * X_i * X_i + (-3.0))
12986 SDValue DAGCombiner::BuildRsqrtNRTwoConst(SDValue Arg, SDValue Est,
12987 unsigned Iterations) {
12988 EVT VT = Arg.getValueType();
12990 SDValue MinusThree = DAG.getConstantFP(-3.0, VT);
12991 SDValue MinusHalf = DAG.getConstantFP(-0.5, VT);
12993 // Newton iterations: Est = -0.5 * Est * (-3.0 + Arg * Est * Est)
12994 for (unsigned i = 0; i < Iterations; ++i) {
12995 SDValue HalfEst = DAG.getNode(ISD::FMUL, DL, VT, Est, MinusHalf);
12996 AddToWorklist(HalfEst.getNode());
12998 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Est);
12999 AddToWorklist(Est.getNode());
13001 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Arg);
13002 AddToWorklist(Est.getNode());
13004 Est = DAG.getNode(ISD::FADD, DL, VT, Est, MinusThree);
13005 AddToWorklist(Est.getNode());
13007 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, HalfEst);
13008 AddToWorklist(Est.getNode());
13013 SDValue DAGCombiner::BuildRsqrtEstimate(SDValue Op) {
13014 if (Level >= AfterLegalizeDAG)
13017 // Expose the DAG combiner to the target combiner implementations.
13018 TargetLowering::DAGCombinerInfo DCI(DAG, Level, false, this);
13019 unsigned Iterations = 0;
13020 bool UseOneConstNR = false;
13021 if (SDValue Est = TLI.getRsqrtEstimate(Op, DCI, Iterations, UseOneConstNR)) {
13022 AddToWorklist(Est.getNode());
13024 Est = UseOneConstNR ?
13025 BuildRsqrtNROneConst(Op, Est, Iterations) :
13026 BuildRsqrtNRTwoConst(Op, Est, Iterations);
13034 /// Return true if base is a frame index, which is known not to alias with
13035 /// anything but itself. Provides base object and offset as results.
13036 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
13037 const GlobalValue *&GV, const void *&CV) {
13038 // Assume it is a primitive operation.
13039 Base = Ptr; Offset = 0; GV = nullptr; CV = nullptr;
13041 // If it's an adding a simple constant then integrate the offset.
13042 if (Base.getOpcode() == ISD::ADD) {
13043 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
13044 Base = Base.getOperand(0);
13045 Offset += C->getZExtValue();
13049 // Return the underlying GlobalValue, and update the Offset. Return false
13050 // for GlobalAddressSDNode since the same GlobalAddress may be represented
13051 // by multiple nodes with different offsets.
13052 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
13053 GV = G->getGlobal();
13054 Offset += G->getOffset();
13058 // Return the underlying Constant value, and update the Offset. Return false
13059 // for ConstantSDNodes since the same constant pool entry may be represented
13060 // by multiple nodes with different offsets.
13061 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
13062 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
13063 : (const void *)C->getConstVal();
13064 Offset += C->getOffset();
13067 // If it's any of the following then it can't alias with anything but itself.
13068 return isa<FrameIndexSDNode>(Base);
13071 /// Return true if there is any possibility that the two addresses overlap.
13072 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const {
13073 // If they are the same then they must be aliases.
13074 if (Op0->getBasePtr() == Op1->getBasePtr()) return true;
13076 // If they are both volatile then they cannot be reordered.
13077 if (Op0->isVolatile() && Op1->isVolatile()) return true;
13079 // Gather base node and offset information.
13080 SDValue Base1, Base2;
13081 int64_t Offset1, Offset2;
13082 const GlobalValue *GV1, *GV2;
13083 const void *CV1, *CV2;
13084 bool isFrameIndex1 = FindBaseOffset(Op0->getBasePtr(),
13085 Base1, Offset1, GV1, CV1);
13086 bool isFrameIndex2 = FindBaseOffset(Op1->getBasePtr(),
13087 Base2, Offset2, GV2, CV2);
13089 // If they have a same base address then check to see if they overlap.
13090 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
13091 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
13092 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
13094 // It is possible for different frame indices to alias each other, mostly
13095 // when tail call optimization reuses return address slots for arguments.
13096 // To catch this case, look up the actual index of frame indices to compute
13097 // the real alias relationship.
13098 if (isFrameIndex1 && isFrameIndex2) {
13099 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
13100 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
13101 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
13102 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
13103 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
13106 // Otherwise, if we know what the bases are, and they aren't identical, then
13107 // we know they cannot alias.
13108 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
13111 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
13112 // compared to the size and offset of the access, we may be able to prove they
13113 // do not alias. This check is conservative for now to catch cases created by
13114 // splitting vector types.
13115 if ((Op0->getOriginalAlignment() == Op1->getOriginalAlignment()) &&
13116 (Op0->getSrcValueOffset() != Op1->getSrcValueOffset()) &&
13117 (Op0->getMemoryVT().getSizeInBits() >> 3 ==
13118 Op1->getMemoryVT().getSizeInBits() >> 3) &&
13119 (Op0->getOriginalAlignment() > Op0->getMemoryVT().getSizeInBits()) >> 3) {
13120 int64_t OffAlign1 = Op0->getSrcValueOffset() % Op0->getOriginalAlignment();
13121 int64_t OffAlign2 = Op1->getSrcValueOffset() % Op1->getOriginalAlignment();
13123 // There is no overlap between these relatively aligned accesses of similar
13124 // size, return no alias.
13125 if ((OffAlign1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign2 ||
13126 (OffAlign2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign1)
13130 bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0
13132 : DAG.getSubtarget().useAA();
13134 if (CombinerAAOnlyFunc.getNumOccurrences() &&
13135 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
13139 Op0->getMemOperand()->getValue() && Op1->getMemOperand()->getValue()) {
13140 // Use alias analysis information.
13141 int64_t MinOffset = std::min(Op0->getSrcValueOffset(),
13142 Op1->getSrcValueOffset());
13143 int64_t Overlap1 = (Op0->getMemoryVT().getSizeInBits() >> 3) +
13144 Op0->getSrcValueOffset() - MinOffset;
13145 int64_t Overlap2 = (Op1->getMemoryVT().getSizeInBits() >> 3) +
13146 Op1->getSrcValueOffset() - MinOffset;
13147 AliasAnalysis::AliasResult AAResult =
13148 AA.alias(AliasAnalysis::Location(Op0->getMemOperand()->getValue(),
13150 UseTBAA ? Op0->getAAInfo() : AAMDNodes()),
13151 AliasAnalysis::Location(Op1->getMemOperand()->getValue(),
13153 UseTBAA ? Op1->getAAInfo() : AAMDNodes()));
13154 if (AAResult == AliasAnalysis::NoAlias)
13158 // Otherwise we have to assume they alias.
13162 /// Walk up chain skipping non-aliasing memory nodes,
13163 /// looking for aliasing nodes and adding them to the Aliases vector.
13164 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
13165 SmallVectorImpl<SDValue> &Aliases) {
13166 SmallVector<SDValue, 8> Chains; // List of chains to visit.
13167 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
13169 // Get alias information for node.
13170 bool IsLoad = isa<LoadSDNode>(N) && !cast<LSBaseSDNode>(N)->isVolatile();
13173 Chains.push_back(OriginalChain);
13174 unsigned Depth = 0;
13176 // Look at each chain and determine if it is an alias. If so, add it to the
13177 // aliases list. If not, then continue up the chain looking for the next
13179 while (!Chains.empty()) {
13180 SDValue Chain = Chains.back();
13183 // For TokenFactor nodes, look at each operand and only continue up the
13184 // chain until we find two aliases. If we've seen two aliases, assume we'll
13185 // find more and revert to original chain since the xform is unlikely to be
13188 // FIXME: The depth check could be made to return the last non-aliasing
13189 // chain we found before we hit a tokenfactor rather than the original
13191 if (Depth > 6 || Aliases.size() == 2) {
13193 Aliases.push_back(OriginalChain);
13197 // Don't bother if we've been before.
13198 if (!Visited.insert(Chain.getNode()).second)
13201 switch (Chain.getOpcode()) {
13202 case ISD::EntryToken:
13203 // Entry token is ideal chain operand, but handled in FindBetterChain.
13208 // Get alias information for Chain.
13209 bool IsOpLoad = isa<LoadSDNode>(Chain.getNode()) &&
13210 !cast<LSBaseSDNode>(Chain.getNode())->isVolatile();
13212 // If chain is alias then stop here.
13213 if (!(IsLoad && IsOpLoad) &&
13214 isAlias(cast<LSBaseSDNode>(N), cast<LSBaseSDNode>(Chain.getNode()))) {
13215 Aliases.push_back(Chain);
13217 // Look further up the chain.
13218 Chains.push_back(Chain.getOperand(0));
13224 case ISD::TokenFactor:
13225 // We have to check each of the operands of the token factor for "small"
13226 // token factors, so we queue them up. Adding the operands to the queue
13227 // (stack) in reverse order maintains the original order and increases the
13228 // likelihood that getNode will find a matching token factor (CSE.)
13229 if (Chain.getNumOperands() > 16) {
13230 Aliases.push_back(Chain);
13233 for (unsigned n = Chain.getNumOperands(); n;)
13234 Chains.push_back(Chain.getOperand(--n));
13239 // For all other instructions we will just have to take what we can get.
13240 Aliases.push_back(Chain);
13245 // We need to be careful here to also search for aliases through the
13246 // value operand of a store, etc. Consider the following situation:
13248 // L1 = load Token1, %52
13249 // S1 = store Token1, L1, %51
13250 // L2 = load Token1, %52+8
13251 // S2 = store Token1, L2, %51+8
13252 // Token2 = Token(S1, S2)
13253 // L3 = load Token2, %53
13254 // S3 = store Token2, L3, %52
13255 // L4 = load Token2, %53+8
13256 // S4 = store Token2, L4, %52+8
13257 // If we search for aliases of S3 (which loads address %52), and we look
13258 // only through the chain, then we'll miss the trivial dependence on L1
13259 // (which also loads from %52). We then might change all loads and
13260 // stores to use Token1 as their chain operand, which could result in
13261 // copying %53 into %52 before copying %52 into %51 (which should
13264 // The problem is, however, that searching for such data dependencies
13265 // can become expensive, and the cost is not directly related to the
13266 // chain depth. Instead, we'll rule out such configurations here by
13267 // insisting that we've visited all chain users (except for users
13268 // of the original chain, which is not necessary). When doing this,
13269 // we need to look through nodes we don't care about (otherwise, things
13270 // like register copies will interfere with trivial cases).
13272 SmallVector<const SDNode *, 16> Worklist;
13273 for (const SDNode *N : Visited)
13274 if (N != OriginalChain.getNode())
13275 Worklist.push_back(N);
13277 while (!Worklist.empty()) {
13278 const SDNode *M = Worklist.pop_back_val();
13280 // We have already visited M, and want to make sure we've visited any uses
13281 // of M that we care about. For uses that we've not visisted, and don't
13282 // care about, queue them to the worklist.
13284 for (SDNode::use_iterator UI = M->use_begin(),
13285 UIE = M->use_end(); UI != UIE; ++UI)
13286 if (UI.getUse().getValueType() == MVT::Other &&
13287 Visited.insert(*UI).second) {
13288 if (isa<MemIntrinsicSDNode>(*UI) || isa<MemSDNode>(*UI)) {
13289 // We've not visited this use, and we care about it (it could have an
13290 // ordering dependency with the original node).
13292 Aliases.push_back(OriginalChain);
13296 // We've not visited this use, but we don't care about it. Mark it as
13297 // visited and enqueue it to the worklist.
13298 Worklist.push_back(*UI);
13303 /// Walk up chain skipping non-aliasing memory nodes, looking for a better chain
13304 /// (aliasing node.)
13305 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
13306 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
13308 // Accumulate all the aliases to this node.
13309 GatherAllAliases(N, OldChain, Aliases);
13311 // If no operands then chain to entry token.
13312 if (Aliases.size() == 0)
13313 return DAG.getEntryNode();
13315 // If a single operand then chain to it. We don't need to revisit it.
13316 if (Aliases.size() == 1)
13319 // Construct a custom tailored token factor.
13320 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Aliases);
13323 /// This is the entry point for the file.
13324 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
13325 CodeGenOpt::Level OptLevel) {
13326 /// This is the main entry point to this class.
13327 DAGCombiner(*this, AA, OptLevel).Run(Level);