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 SimplifyVUnaryOp(SDNode *N);
255 SDValue visitSHL(SDNode *N);
256 SDValue visitSRA(SDNode *N);
257 SDValue visitSRL(SDNode *N);
258 SDValue visitRotate(SDNode *N);
259 SDValue visitCTLZ(SDNode *N);
260 SDValue visitCTLZ_ZERO_UNDEF(SDNode *N);
261 SDValue visitCTTZ(SDNode *N);
262 SDValue visitCTTZ_ZERO_UNDEF(SDNode *N);
263 SDValue visitCTPOP(SDNode *N);
264 SDValue visitSELECT(SDNode *N);
265 SDValue visitVSELECT(SDNode *N);
266 SDValue visitSELECT_CC(SDNode *N);
267 SDValue visitSETCC(SDNode *N);
268 SDValue visitSIGN_EXTEND(SDNode *N);
269 SDValue visitZERO_EXTEND(SDNode *N);
270 SDValue visitANY_EXTEND(SDNode *N);
271 SDValue visitSIGN_EXTEND_INREG(SDNode *N);
272 SDValue visitTRUNCATE(SDNode *N);
273 SDValue visitBITCAST(SDNode *N);
274 SDValue visitBUILD_PAIR(SDNode *N);
275 SDValue visitFADD(SDNode *N);
276 SDValue visitFSUB(SDNode *N);
277 SDValue visitFMUL(SDNode *N);
278 SDValue visitFMA(SDNode *N);
279 SDValue visitFDIV(SDNode *N);
280 SDValue visitFREM(SDNode *N);
281 SDValue visitFSQRT(SDNode *N);
282 SDValue visitFCOPYSIGN(SDNode *N);
283 SDValue visitSINT_TO_FP(SDNode *N);
284 SDValue visitUINT_TO_FP(SDNode *N);
285 SDValue visitFP_TO_SINT(SDNode *N);
286 SDValue visitFP_TO_UINT(SDNode *N);
287 SDValue visitFP_ROUND(SDNode *N);
288 SDValue visitFP_ROUND_INREG(SDNode *N);
289 SDValue visitFP_EXTEND(SDNode *N);
290 SDValue visitFNEG(SDNode *N);
291 SDValue visitFABS(SDNode *N);
292 SDValue visitFCEIL(SDNode *N);
293 SDValue visitFTRUNC(SDNode *N);
294 SDValue visitFFLOOR(SDNode *N);
295 SDValue visitFMINNUM(SDNode *N);
296 SDValue visitFMAXNUM(SDNode *N);
297 SDValue visitBRCOND(SDNode *N);
298 SDValue visitBR_CC(SDNode *N);
299 SDValue visitLOAD(SDNode *N);
300 SDValue visitSTORE(SDNode *N);
301 SDValue visitINSERT_VECTOR_ELT(SDNode *N);
302 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
303 SDValue visitBUILD_VECTOR(SDNode *N);
304 SDValue visitCONCAT_VECTORS(SDNode *N);
305 SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
306 SDValue visitVECTOR_SHUFFLE(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 BuildVector or constant.
709 static SDNode *isConstantBuildVectorOrConstantInt(SDValue N) {
710 if (isa<ConstantSDNode>(N))
712 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N);
713 if (BV && BV->isConstant())
718 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
720 static ConstantSDNode *isConstOrConstSplat(SDValue N) {
721 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N))
724 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
725 BitVector UndefElements;
726 ConstantSDNode *CN = BV->getConstantSplatNode(&UndefElements);
728 // BuildVectors can truncate their operands. Ignore that case here.
729 // FIXME: We blindly ignore splats which include undef which is overly
731 if (CN && UndefElements.none() &&
732 CN->getValueType(0) == N.getValueType().getScalarType())
739 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
741 static ConstantFPSDNode *isConstOrConstSplatFP(SDValue N) {
742 if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N))
745 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
746 BitVector UndefElements;
747 ConstantFPSDNode *CN = BV->getConstantFPSplatNode(&UndefElements);
749 if (CN && UndefElements.none())
756 SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL,
757 SDValue N0, SDValue N1) {
758 EVT VT = N0.getValueType();
759 if (N0.getOpcode() == Opc) {
760 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0.getOperand(1))) {
761 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1)) {
762 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
763 if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, L, R))
764 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
767 if (N0.hasOneUse()) {
768 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one
770 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1);
771 if (!OpNode.getNode())
773 AddToWorklist(OpNode.getNode());
774 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
779 if (N1.getOpcode() == Opc) {
780 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1.getOperand(1))) {
781 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0)) {
782 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
783 if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, R, L))
784 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
787 if (N1.hasOneUse()) {
788 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one
790 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N1.getOperand(0), N0);
791 if (!OpNode.getNode())
793 AddToWorklist(OpNode.getNode());
794 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
802 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
804 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
806 DEBUG(dbgs() << "\nReplacing.1 ";
808 dbgs() << "\nWith: ";
809 To[0].getNode()->dump(&DAG);
810 dbgs() << " and " << NumTo-1 << " other values\n");
811 for (unsigned i = 0, e = NumTo; i != e; ++i)
812 assert((!To[i].getNode() ||
813 N->getValueType(i) == To[i].getValueType()) &&
814 "Cannot combine value to value of different type!");
816 WorklistRemover DeadNodes(*this);
817 DAG.ReplaceAllUsesWith(N, To);
819 // Push the new nodes and any users onto the worklist
820 for (unsigned i = 0, e = NumTo; i != e; ++i) {
821 if (To[i].getNode()) {
822 AddToWorklist(To[i].getNode());
823 AddUsersToWorklist(To[i].getNode());
828 // Finally, if the node is now dead, remove it from the graph. The node
829 // may not be dead if the replacement process recursively simplified to
830 // something else needing this node.
832 deleteAndRecombine(N);
833 return SDValue(N, 0);
837 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
838 // Replace all uses. If any nodes become isomorphic to other nodes and
839 // are deleted, make sure to remove them from our worklist.
840 WorklistRemover DeadNodes(*this);
841 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New);
843 // Push the new node and any (possibly new) users onto the worklist.
844 AddToWorklist(TLO.New.getNode());
845 AddUsersToWorklist(TLO.New.getNode());
847 // Finally, if the node is now dead, remove it from the graph. The node
848 // may not be dead if the replacement process recursively simplified to
849 // something else needing this node.
850 if (TLO.Old.getNode()->use_empty())
851 deleteAndRecombine(TLO.Old.getNode());
854 /// Check the specified integer node value to see if it can be simplified or if
855 /// things it uses can be simplified by bit propagation. If so, return true.
856 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
857 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
858 APInt KnownZero, KnownOne;
859 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
863 AddToWorklist(Op.getNode());
865 // Replace the old value with the new one.
867 DEBUG(dbgs() << "\nReplacing.2 ";
868 TLO.Old.getNode()->dump(&DAG);
869 dbgs() << "\nWith: ";
870 TLO.New.getNode()->dump(&DAG);
873 CommitTargetLoweringOpt(TLO);
877 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
879 EVT VT = Load->getValueType(0);
880 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
882 DEBUG(dbgs() << "\nReplacing.9 ";
884 dbgs() << "\nWith: ";
885 Trunc.getNode()->dump(&DAG);
887 WorklistRemover DeadNodes(*this);
888 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc);
889 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1));
890 deleteAndRecombine(Load);
891 AddToWorklist(Trunc.getNode());
894 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
897 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
898 EVT MemVT = LD->getMemoryVT();
899 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
900 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, PVT, MemVT) ? ISD::ZEXTLOAD
902 : LD->getExtensionType();
904 return DAG.getExtLoad(ExtType, dl, PVT,
905 LD->getChain(), LD->getBasePtr(),
906 MemVT, LD->getMemOperand());
909 unsigned Opc = Op.getOpcode();
912 case ISD::AssertSext:
913 return DAG.getNode(ISD::AssertSext, dl, PVT,
914 SExtPromoteOperand(Op.getOperand(0), PVT),
916 case ISD::AssertZext:
917 return DAG.getNode(ISD::AssertZext, dl, PVT,
918 ZExtPromoteOperand(Op.getOperand(0), PVT),
920 case ISD::Constant: {
922 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
923 return DAG.getNode(ExtOpc, dl, PVT, Op);
927 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
929 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
932 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
933 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
935 EVT OldVT = Op.getValueType();
937 bool Replace = false;
938 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
939 if (!NewOp.getNode())
941 AddToWorklist(NewOp.getNode());
944 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
945 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
946 DAG.getValueType(OldVT));
949 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
950 EVT OldVT = Op.getValueType();
952 bool Replace = false;
953 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
954 if (!NewOp.getNode())
956 AddToWorklist(NewOp.getNode());
959 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
960 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
963 /// Promote the specified integer binary operation if the target indicates it is
964 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
965 /// i32 since i16 instructions are longer.
966 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
967 if (!LegalOperations)
970 EVT VT = Op.getValueType();
971 if (VT.isVector() || !VT.isInteger())
974 // If operation type is 'undesirable', e.g. i16 on x86, consider
976 unsigned Opc = Op.getOpcode();
977 if (TLI.isTypeDesirableForOp(Opc, VT))
981 // Consult target whether it is a good idea to promote this operation and
982 // what's the right type to promote it to.
983 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
984 assert(PVT != VT && "Don't know what type to promote to!");
986 bool Replace0 = false;
987 SDValue N0 = Op.getOperand(0);
988 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
992 bool Replace1 = false;
993 SDValue N1 = Op.getOperand(1);
998 NN1 = PromoteOperand(N1, PVT, Replace1);
1003 AddToWorklist(NN0.getNode());
1005 AddToWorklist(NN1.getNode());
1008 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
1010 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
1012 DEBUG(dbgs() << "\nPromoting ";
1013 Op.getNode()->dump(&DAG));
1015 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1016 DAG.getNode(Opc, dl, PVT, NN0, NN1));
1021 /// Promote the specified integer shift operation if the target indicates it is
1022 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
1023 /// i32 since i16 instructions are longer.
1024 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
1025 if (!LegalOperations)
1028 EVT VT = Op.getValueType();
1029 if (VT.isVector() || !VT.isInteger())
1032 // If operation type is 'undesirable', e.g. i16 on x86, consider
1034 unsigned Opc = Op.getOpcode();
1035 if (TLI.isTypeDesirableForOp(Opc, VT))
1039 // Consult target whether it is a good idea to promote this operation and
1040 // what's the right type to promote it to.
1041 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1042 assert(PVT != VT && "Don't know what type to promote to!");
1044 bool Replace = false;
1045 SDValue N0 = Op.getOperand(0);
1046 if (Opc == ISD::SRA)
1047 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
1048 else if (Opc == ISD::SRL)
1049 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
1051 N0 = PromoteOperand(N0, PVT, Replace);
1055 AddToWorklist(N0.getNode());
1057 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
1059 DEBUG(dbgs() << "\nPromoting ";
1060 Op.getNode()->dump(&DAG));
1062 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1063 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
1068 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
1069 if (!LegalOperations)
1072 EVT VT = Op.getValueType();
1073 if (VT.isVector() || !VT.isInteger())
1076 // If operation type is 'undesirable', e.g. i16 on x86, consider
1078 unsigned Opc = Op.getOpcode();
1079 if (TLI.isTypeDesirableForOp(Opc, VT))
1083 // Consult target whether it is a good idea to promote this operation and
1084 // what's the right type to promote it to.
1085 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1086 assert(PVT != VT && "Don't know what type to promote to!");
1087 // fold (aext (aext x)) -> (aext x)
1088 // fold (aext (zext x)) -> (zext x)
1089 // fold (aext (sext x)) -> (sext x)
1090 DEBUG(dbgs() << "\nPromoting ";
1091 Op.getNode()->dump(&DAG));
1092 return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0));
1097 bool DAGCombiner::PromoteLoad(SDValue Op) {
1098 if (!LegalOperations)
1101 EVT VT = Op.getValueType();
1102 if (VT.isVector() || !VT.isInteger())
1105 // If operation type is 'undesirable', e.g. i16 on x86, consider
1107 unsigned Opc = Op.getOpcode();
1108 if (TLI.isTypeDesirableForOp(Opc, VT))
1112 // Consult target whether it is a good idea to promote this operation and
1113 // what's the right type to promote it to.
1114 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1115 assert(PVT != VT && "Don't know what type to promote to!");
1118 SDNode *N = Op.getNode();
1119 LoadSDNode *LD = cast<LoadSDNode>(N);
1120 EVT MemVT = LD->getMemoryVT();
1121 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
1122 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, PVT, MemVT) ? ISD::ZEXTLOAD
1124 : LD->getExtensionType();
1125 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
1126 LD->getChain(), LD->getBasePtr(),
1127 MemVT, LD->getMemOperand());
1128 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
1130 DEBUG(dbgs() << "\nPromoting ";
1133 Result.getNode()->dump(&DAG);
1135 WorklistRemover DeadNodes(*this);
1136 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
1137 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1));
1138 deleteAndRecombine(N);
1139 AddToWorklist(Result.getNode());
1145 /// \brief Recursively delete a node which has no uses and any operands for
1146 /// which it is the only use.
1148 /// Note that this both deletes the nodes and removes them from the worklist.
1149 /// It also adds any nodes who have had a user deleted to the worklist as they
1150 /// may now have only one use and subject to other combines.
1151 bool DAGCombiner::recursivelyDeleteUnusedNodes(SDNode *N) {
1152 if (!N->use_empty())
1155 SmallSetVector<SDNode *, 16> Nodes;
1158 N = Nodes.pop_back_val();
1162 if (N->use_empty()) {
1163 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1164 Nodes.insert(N->getOperand(i).getNode());
1166 removeFromWorklist(N);
1171 } while (!Nodes.empty());
1175 //===----------------------------------------------------------------------===//
1176 // Main DAG Combiner implementation
1177 //===----------------------------------------------------------------------===//
1179 void DAGCombiner::Run(CombineLevel AtLevel) {
1180 // set the instance variables, so that the various visit routines may use it.
1182 LegalOperations = Level >= AfterLegalizeVectorOps;
1183 LegalTypes = Level >= AfterLegalizeTypes;
1185 // Early exit if this basic block is in an optnone function.
1186 if (DAG.getMachineFunction().getFunction()->hasFnAttribute(
1187 Attribute::OptimizeNone))
1190 // Add all the dag nodes to the worklist.
1191 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
1192 E = DAG.allnodes_end(); I != E; ++I)
1195 // Create a dummy node (which is not added to allnodes), that adds a reference
1196 // to the root node, preventing it from being deleted, and tracking any
1197 // changes of the root.
1198 HandleSDNode Dummy(DAG.getRoot());
1200 // while the worklist isn't empty, find a node and
1201 // try and combine it.
1202 while (!WorklistMap.empty()) {
1204 // The Worklist holds the SDNodes in order, but it may contain null entries.
1206 N = Worklist.pop_back_val();
1209 bool GoodWorklistEntry = WorklistMap.erase(N);
1210 (void)GoodWorklistEntry;
1211 assert(GoodWorklistEntry &&
1212 "Found a worklist entry without a corresponding map entry!");
1214 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1215 // N is deleted from the DAG, since they too may now be dead or may have a
1216 // reduced number of uses, allowing other xforms.
1217 if (recursivelyDeleteUnusedNodes(N))
1220 WorklistRemover DeadNodes(*this);
1222 // If this combine is running after legalizing the DAG, re-legalize any
1223 // nodes pulled off the worklist.
1224 if (Level == AfterLegalizeDAG) {
1225 SmallSetVector<SDNode *, 16> UpdatedNodes;
1226 bool NIsValid = DAG.LegalizeOp(N, UpdatedNodes);
1228 for (SDNode *LN : UpdatedNodes) {
1230 AddUsersToWorklist(LN);
1236 DEBUG(dbgs() << "\nCombining: "; N->dump(&DAG));
1238 // Add any operands of the new node which have not yet been combined to the
1239 // worklist as well. Because the worklist uniques things already, this
1240 // won't repeatedly process the same operand.
1241 CombinedNodes.insert(N);
1242 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1243 if (!CombinedNodes.count(N->getOperand(i).getNode()))
1244 AddToWorklist(N->getOperand(i).getNode());
1246 SDValue RV = combine(N);
1253 // If we get back the same node we passed in, rather than a new node or
1254 // zero, we know that the node must have defined multiple values and
1255 // CombineTo was used. Since CombineTo takes care of the worklist
1256 // mechanics for us, we have no work to do in this case.
1257 if (RV.getNode() == N)
1260 assert(N->getOpcode() != ISD::DELETED_NODE &&
1261 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1262 "Node was deleted but visit returned new node!");
1264 DEBUG(dbgs() << " ... into: ";
1265 RV.getNode()->dump(&DAG));
1267 // Transfer debug value.
1268 DAG.TransferDbgValues(SDValue(N, 0), RV);
1269 if (N->getNumValues() == RV.getNode()->getNumValues())
1270 DAG.ReplaceAllUsesWith(N, RV.getNode());
1272 assert(N->getValueType(0) == RV.getValueType() &&
1273 N->getNumValues() == 1 && "Type mismatch");
1275 DAG.ReplaceAllUsesWith(N, &OpV);
1278 // Push the new node and any users onto the worklist
1279 AddToWorklist(RV.getNode());
1280 AddUsersToWorklist(RV.getNode());
1282 // Finally, if the node is now dead, remove it from the graph. The node
1283 // may not be dead if the replacement process recursively simplified to
1284 // something else needing this node. This will also take care of adding any
1285 // operands which have lost a user to the worklist.
1286 recursivelyDeleteUnusedNodes(N);
1289 // If the root changed (e.g. it was a dead load, update the root).
1290 DAG.setRoot(Dummy.getValue());
1291 DAG.RemoveDeadNodes();
1294 SDValue DAGCombiner::visit(SDNode *N) {
1295 switch (N->getOpcode()) {
1297 case ISD::TokenFactor: return visitTokenFactor(N);
1298 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1299 case ISD::ADD: return visitADD(N);
1300 case ISD::SUB: return visitSUB(N);
1301 case ISD::ADDC: return visitADDC(N);
1302 case ISD::SUBC: return visitSUBC(N);
1303 case ISD::ADDE: return visitADDE(N);
1304 case ISD::SUBE: return visitSUBE(N);
1305 case ISD::MUL: return visitMUL(N);
1306 case ISD::SDIV: return visitSDIV(N);
1307 case ISD::UDIV: return visitUDIV(N);
1308 case ISD::SREM: return visitSREM(N);
1309 case ISD::UREM: return visitUREM(N);
1310 case ISD::MULHU: return visitMULHU(N);
1311 case ISD::MULHS: return visitMULHS(N);
1312 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1313 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1314 case ISD::SMULO: return visitSMULO(N);
1315 case ISD::UMULO: return visitUMULO(N);
1316 case ISD::SDIVREM: return visitSDIVREM(N);
1317 case ISD::UDIVREM: return visitUDIVREM(N);
1318 case ISD::AND: return visitAND(N);
1319 case ISD::OR: return visitOR(N);
1320 case ISD::XOR: return visitXOR(N);
1321 case ISD::SHL: return visitSHL(N);
1322 case ISD::SRA: return visitSRA(N);
1323 case ISD::SRL: return visitSRL(N);
1325 case ISD::ROTL: return visitRotate(N);
1326 case ISD::CTLZ: return visitCTLZ(N);
1327 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1328 case ISD::CTTZ: return visitCTTZ(N);
1329 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1330 case ISD::CTPOP: return visitCTPOP(N);
1331 case ISD::SELECT: return visitSELECT(N);
1332 case ISD::VSELECT: return visitVSELECT(N);
1333 case ISD::SELECT_CC: return visitSELECT_CC(N);
1334 case ISD::SETCC: return visitSETCC(N);
1335 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1336 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1337 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1338 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1339 case ISD::TRUNCATE: return visitTRUNCATE(N);
1340 case ISD::BITCAST: return visitBITCAST(N);
1341 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1342 case ISD::FADD: return visitFADD(N);
1343 case ISD::FSUB: return visitFSUB(N);
1344 case ISD::FMUL: return visitFMUL(N);
1345 case ISD::FMA: return visitFMA(N);
1346 case ISD::FDIV: return visitFDIV(N);
1347 case ISD::FREM: return visitFREM(N);
1348 case ISD::FSQRT: return visitFSQRT(N);
1349 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1350 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1351 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1352 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1353 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1354 case ISD::FP_ROUND: return visitFP_ROUND(N);
1355 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1356 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1357 case ISD::FNEG: return visitFNEG(N);
1358 case ISD::FABS: return visitFABS(N);
1359 case ISD::FFLOOR: return visitFFLOOR(N);
1360 case ISD::FMINNUM: return visitFMINNUM(N);
1361 case ISD::FMAXNUM: return visitFMAXNUM(N);
1362 case ISD::FCEIL: return visitFCEIL(N);
1363 case ISD::FTRUNC: return visitFTRUNC(N);
1364 case ISD::BRCOND: return visitBRCOND(N);
1365 case ISD::BR_CC: return visitBR_CC(N);
1366 case ISD::LOAD: return visitLOAD(N);
1367 case ISD::STORE: return visitSTORE(N);
1368 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1369 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1370 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1371 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1372 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1373 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1374 case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N);
1375 case ISD::MLOAD: return visitMLOAD(N);
1376 case ISD::MSTORE: return visitMSTORE(N);
1381 SDValue DAGCombiner::combine(SDNode *N) {
1382 SDValue RV = visit(N);
1384 // If nothing happened, try a target-specific DAG combine.
1385 if (!RV.getNode()) {
1386 assert(N->getOpcode() != ISD::DELETED_NODE &&
1387 "Node was deleted but visit returned NULL!");
1389 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1390 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1392 // Expose the DAG combiner to the target combiner impls.
1393 TargetLowering::DAGCombinerInfo
1394 DagCombineInfo(DAG, Level, false, this);
1396 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1400 // If nothing happened still, try promoting the operation.
1401 if (!RV.getNode()) {
1402 switch (N->getOpcode()) {
1410 RV = PromoteIntBinOp(SDValue(N, 0));
1415 RV = PromoteIntShiftOp(SDValue(N, 0));
1417 case ISD::SIGN_EXTEND:
1418 case ISD::ZERO_EXTEND:
1419 case ISD::ANY_EXTEND:
1420 RV = PromoteExtend(SDValue(N, 0));
1423 if (PromoteLoad(SDValue(N, 0)))
1429 // If N is a commutative binary node, try commuting it to enable more
1431 if (!RV.getNode() && SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1432 N->getNumValues() == 1) {
1433 SDValue N0 = N->getOperand(0);
1434 SDValue N1 = N->getOperand(1);
1436 // Constant operands are canonicalized to RHS.
1437 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1438 SDValue Ops[] = {N1, N0};
1440 if (const BinaryWithFlagsSDNode *BinNode =
1441 dyn_cast<BinaryWithFlagsSDNode>(N)) {
1442 CSENode = DAG.getNodeIfExists(
1443 N->getOpcode(), N->getVTList(), Ops, BinNode->hasNoUnsignedWrap(),
1444 BinNode->hasNoSignedWrap(), BinNode->isExact());
1446 CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops);
1449 return SDValue(CSENode, 0);
1456 /// Given a node, return its input chain if it has one, otherwise return a null
1458 static SDValue getInputChainForNode(SDNode *N) {
1459 if (unsigned NumOps = N->getNumOperands()) {
1460 if (N->getOperand(0).getValueType() == MVT::Other)
1461 return N->getOperand(0);
1462 if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1463 return N->getOperand(NumOps-1);
1464 for (unsigned i = 1; i < NumOps-1; ++i)
1465 if (N->getOperand(i).getValueType() == MVT::Other)
1466 return N->getOperand(i);
1471 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1472 // If N has two operands, where one has an input chain equal to the other,
1473 // the 'other' chain is redundant.
1474 if (N->getNumOperands() == 2) {
1475 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1476 return N->getOperand(0);
1477 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1478 return N->getOperand(1);
1481 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1482 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1483 SmallPtrSet<SDNode*, 16> SeenOps;
1484 bool Changed = false; // If we should replace this token factor.
1486 // Start out with this token factor.
1489 // Iterate through token factors. The TFs grows when new token factors are
1491 for (unsigned i = 0; i < TFs.size(); ++i) {
1492 SDNode *TF = TFs[i];
1494 // Check each of the operands.
1495 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1496 SDValue Op = TF->getOperand(i);
1498 switch (Op.getOpcode()) {
1499 case ISD::EntryToken:
1500 // Entry tokens don't need to be added to the list. They are
1505 case ISD::TokenFactor:
1506 if (Op.hasOneUse() &&
1507 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1508 // Queue up for processing.
1509 TFs.push_back(Op.getNode());
1510 // Clean up in case the token factor is removed.
1511 AddToWorklist(Op.getNode());
1518 // Only add if it isn't already in the list.
1519 if (SeenOps.insert(Op.getNode()).second)
1530 // If we've changed things around then replace token factor.
1533 // The entry token is the only possible outcome.
1534 Result = DAG.getEntryNode();
1536 // New and improved token factor.
1537 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Ops);
1540 // Add users to worklist if AA is enabled, since it may introduce
1541 // a lot of new chained token factors while removing memory deps.
1542 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
1543 : DAG.getSubtarget().useAA();
1544 return CombineTo(N, Result, UseAA /*add to worklist*/);
1550 /// MERGE_VALUES can always be eliminated.
1551 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1552 WorklistRemover DeadNodes(*this);
1553 // Replacing results may cause a different MERGE_VALUES to suddenly
1554 // be CSE'd with N, and carry its uses with it. Iterate until no
1555 // uses remain, to ensure that the node can be safely deleted.
1556 // First add the users of this node to the work list so that they
1557 // can be tried again once they have new operands.
1558 AddUsersToWorklist(N);
1560 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1561 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1562 } while (!N->use_empty());
1563 deleteAndRecombine(N);
1564 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1567 SDValue DAGCombiner::visitADD(SDNode *N) {
1568 SDValue N0 = N->getOperand(0);
1569 SDValue N1 = N->getOperand(1);
1570 EVT VT = N0.getValueType();
1573 if (VT.isVector()) {
1574 SDValue FoldedVOp = SimplifyVBinOp(N);
1575 if (FoldedVOp.getNode()) return FoldedVOp;
1577 // fold (add x, 0) -> x, vector edition
1578 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1580 if (ISD::isBuildVectorAllZeros(N0.getNode()))
1584 // fold (add x, undef) -> undef
1585 if (N0.getOpcode() == ISD::UNDEF)
1587 if (N1.getOpcode() == ISD::UNDEF)
1589 // fold (add c1, c2) -> c1+c2
1590 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1591 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1593 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1594 // canonicalize constant to RHS
1596 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, N0);
1597 // fold (add x, 0) -> x
1598 if (N1C && N1C->isNullValue())
1600 // fold (add Sym, c) -> Sym+c
1601 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1602 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1603 GA->getOpcode() == ISD::GlobalAddress)
1604 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1606 (uint64_t)N1C->getSExtValue());
1607 // fold ((c1-A)+c2) -> (c1+c2)-A
1608 if (N1C && N0.getOpcode() == ISD::SUB)
1609 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1610 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1611 DAG.getConstant(N1C->getAPIntValue()+
1612 N0C->getAPIntValue(), VT),
1615 SDValue RADD = ReassociateOps(ISD::ADD, SDLoc(N), N0, N1);
1618 // fold ((0-A) + B) -> B-A
1619 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1620 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1621 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, N0.getOperand(1));
1622 // fold (A + (0-B)) -> A-B
1623 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1624 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1625 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1.getOperand(1));
1626 // fold (A+(B-A)) -> B
1627 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1628 return N1.getOperand(0);
1629 // fold ((B-A)+A) -> B
1630 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1631 return N0.getOperand(0);
1632 // fold (A+(B-(A+C))) to (B-C)
1633 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1634 N0 == N1.getOperand(1).getOperand(0))
1635 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1636 N1.getOperand(1).getOperand(1));
1637 // fold (A+(B-(C+A))) to (B-C)
1638 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1639 N0 == N1.getOperand(1).getOperand(1))
1640 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1641 N1.getOperand(1).getOperand(0));
1642 // fold (A+((B-A)+or-C)) to (B+or-C)
1643 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1644 N1.getOperand(0).getOpcode() == ISD::SUB &&
1645 N0 == N1.getOperand(0).getOperand(1))
1646 return DAG.getNode(N1.getOpcode(), SDLoc(N), VT,
1647 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1649 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1650 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1651 SDValue N00 = N0.getOperand(0);
1652 SDValue N01 = N0.getOperand(1);
1653 SDValue N10 = N1.getOperand(0);
1654 SDValue N11 = N1.getOperand(1);
1656 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1657 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1658 DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10),
1659 DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11));
1662 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1663 return SDValue(N, 0);
1665 // fold (a+b) -> (a|b) iff a and b share no bits.
1666 if (VT.isInteger() && !VT.isVector()) {
1667 APInt LHSZero, LHSOne;
1668 APInt RHSZero, RHSOne;
1669 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1671 if (LHSZero.getBoolValue()) {
1672 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1674 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1675 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1676 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero){
1677 if (!LegalOperations || TLI.isOperationLegal(ISD::OR, VT))
1678 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1);
1683 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1684 if (N1.getOpcode() == ISD::SHL &&
1685 N1.getOperand(0).getOpcode() == ISD::SUB)
1686 if (ConstantSDNode *C =
1687 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1688 if (C->getAPIntValue() == 0)
1689 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0,
1690 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1691 N1.getOperand(0).getOperand(1),
1693 if (N0.getOpcode() == ISD::SHL &&
1694 N0.getOperand(0).getOpcode() == ISD::SUB)
1695 if (ConstantSDNode *C =
1696 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1697 if (C->getAPIntValue() == 0)
1698 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1,
1699 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1700 N0.getOperand(0).getOperand(1),
1703 if (N1.getOpcode() == ISD::AND) {
1704 SDValue AndOp0 = N1.getOperand(0);
1705 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1706 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1707 unsigned DestBits = VT.getScalarType().getSizeInBits();
1709 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1710 // and similar xforms where the inner op is either ~0 or 0.
1711 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1713 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1717 // add (sext i1), X -> sub X, (zext i1)
1718 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1719 N0.getOperand(0).getValueType() == MVT::i1 &&
1720 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1722 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1723 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1726 // add X, (sextinreg Y i1) -> sub X, (and Y 1)
1727 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
1728 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
1729 if (TN->getVT() == MVT::i1) {
1731 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
1732 DAG.getConstant(1, VT));
1733 return DAG.getNode(ISD::SUB, DL, VT, N0, ZExt);
1740 SDValue DAGCombiner::visitADDC(SDNode *N) {
1741 SDValue N0 = N->getOperand(0);
1742 SDValue N1 = N->getOperand(1);
1743 EVT VT = N0.getValueType();
1745 // If the flag result is dead, turn this into an ADD.
1746 if (!N->hasAnyUseOfValue(1))
1747 return CombineTo(N, DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N1),
1748 DAG.getNode(ISD::CARRY_FALSE,
1749 SDLoc(N), MVT::Glue));
1751 // canonicalize constant to RHS.
1752 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1753 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1755 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N1, N0);
1757 // fold (addc x, 0) -> x + no carry out
1758 if (N1C && N1C->isNullValue())
1759 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1760 SDLoc(N), MVT::Glue));
1762 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1763 APInt LHSZero, LHSOne;
1764 APInt RHSZero, RHSOne;
1765 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1767 if (LHSZero.getBoolValue()) {
1768 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1770 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1771 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1772 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1773 return CombineTo(N, DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1),
1774 DAG.getNode(ISD::CARRY_FALSE,
1775 SDLoc(N), MVT::Glue));
1781 SDValue DAGCombiner::visitADDE(SDNode *N) {
1782 SDValue N0 = N->getOperand(0);
1783 SDValue N1 = N->getOperand(1);
1784 SDValue CarryIn = N->getOperand(2);
1786 // canonicalize constant to RHS
1787 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1788 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1790 return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(),
1793 // fold (adde x, y, false) -> (addc x, y)
1794 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1795 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1);
1800 // Since it may not be valid to emit a fold to zero for vector initializers
1801 // check if we can before folding.
1802 static SDValue tryFoldToZero(SDLoc DL, const TargetLowering &TLI, EVT VT,
1804 bool LegalOperations, bool LegalTypes) {
1806 return DAG.getConstant(0, VT);
1807 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
1808 return DAG.getConstant(0, VT);
1812 SDValue DAGCombiner::visitSUB(SDNode *N) {
1813 SDValue N0 = N->getOperand(0);
1814 SDValue N1 = N->getOperand(1);
1815 EVT VT = N0.getValueType();
1818 if (VT.isVector()) {
1819 SDValue FoldedVOp = SimplifyVBinOp(N);
1820 if (FoldedVOp.getNode()) return FoldedVOp;
1822 // fold (sub x, 0) -> x, vector edition
1823 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1827 // fold (sub x, x) -> 0
1828 // FIXME: Refactor this and xor and other similar operations together.
1830 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
1831 // fold (sub c1, c2) -> c1-c2
1832 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1833 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1835 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1836 // fold (sub x, c) -> (add x, -c)
1838 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0,
1839 DAG.getConstant(-N1C->getAPIntValue(), VT));
1840 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1841 if (N0C && N0C->isAllOnesValue())
1842 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
1843 // fold A-(A-B) -> B
1844 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1845 return N1.getOperand(1);
1846 // fold (A+B)-A -> B
1847 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1848 return N0.getOperand(1);
1849 // fold (A+B)-B -> A
1850 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1851 return N0.getOperand(0);
1852 // fold C2-(A+C1) -> (C2-C1)-A
1853 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? nullptr :
1854 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1855 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1856 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(),
1858 return DAG.getNode(ISD::SUB, SDLoc(N), VT, NewC,
1861 // fold ((A+(B+or-C))-B) -> A+or-C
1862 if (N0.getOpcode() == ISD::ADD &&
1863 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1864 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1865 N0.getOperand(1).getOperand(0) == N1)
1866 return DAG.getNode(N0.getOperand(1).getOpcode(), SDLoc(N), VT,
1867 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1868 // fold ((A+(C+B))-B) -> A+C
1869 if (N0.getOpcode() == ISD::ADD &&
1870 N0.getOperand(1).getOpcode() == ISD::ADD &&
1871 N0.getOperand(1).getOperand(1) == N1)
1872 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1873 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1874 // fold ((A-(B-C))-C) -> A-B
1875 if (N0.getOpcode() == ISD::SUB &&
1876 N0.getOperand(1).getOpcode() == ISD::SUB &&
1877 N0.getOperand(1).getOperand(1) == N1)
1878 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1879 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1881 // If either operand of a sub is undef, the result is undef
1882 if (N0.getOpcode() == ISD::UNDEF)
1884 if (N1.getOpcode() == ISD::UNDEF)
1887 // If the relocation model supports it, consider symbol offsets.
1888 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1889 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1890 // fold (sub Sym, c) -> Sym-c
1891 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1892 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1894 (uint64_t)N1C->getSExtValue());
1895 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1896 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1897 if (GA->getGlobal() == GB->getGlobal())
1898 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1902 // sub X, (sextinreg Y i1) -> add X, (and Y 1)
1903 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
1904 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
1905 if (TN->getVT() == MVT::i1) {
1907 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
1908 DAG.getConstant(1, VT));
1909 return DAG.getNode(ISD::ADD, DL, VT, N0, ZExt);
1916 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1917 SDValue N0 = N->getOperand(0);
1918 SDValue N1 = N->getOperand(1);
1919 EVT VT = N0.getValueType();
1921 // If the flag result is dead, turn this into an SUB.
1922 if (!N->hasAnyUseOfValue(1))
1923 return CombineTo(N, DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1),
1924 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1927 // fold (subc x, x) -> 0 + no borrow
1929 return CombineTo(N, DAG.getConstant(0, VT),
1930 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1933 // fold (subc x, 0) -> x + no borrow
1934 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1935 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1936 if (N1C && N1C->isNullValue())
1937 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1940 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1941 if (N0C && N0C->isAllOnesValue())
1942 return CombineTo(N, DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0),
1943 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1949 SDValue DAGCombiner::visitSUBE(SDNode *N) {
1950 SDValue N0 = N->getOperand(0);
1951 SDValue N1 = N->getOperand(1);
1952 SDValue CarryIn = N->getOperand(2);
1954 // fold (sube x, y, false) -> (subc x, y)
1955 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1956 return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1);
1961 SDValue DAGCombiner::visitMUL(SDNode *N) {
1962 SDValue N0 = N->getOperand(0);
1963 SDValue N1 = N->getOperand(1);
1964 EVT VT = N0.getValueType();
1966 // fold (mul x, undef) -> 0
1967 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1968 return DAG.getConstant(0, VT);
1970 bool N0IsConst = false;
1971 bool N1IsConst = false;
1972 APInt ConstValue0, ConstValue1;
1974 if (VT.isVector()) {
1975 SDValue FoldedVOp = SimplifyVBinOp(N);
1976 if (FoldedVOp.getNode()) return FoldedVOp;
1978 N0IsConst = isConstantSplatVector(N0.getNode(), ConstValue0);
1979 N1IsConst = isConstantSplatVector(N1.getNode(), ConstValue1);
1981 N0IsConst = dyn_cast<ConstantSDNode>(N0) != nullptr;
1982 ConstValue0 = N0IsConst ? (dyn_cast<ConstantSDNode>(N0))->getAPIntValue()
1984 N1IsConst = dyn_cast<ConstantSDNode>(N1) != nullptr;
1985 ConstValue1 = N1IsConst ? (dyn_cast<ConstantSDNode>(N1))->getAPIntValue()
1989 // fold (mul c1, c2) -> c1*c2
1990 if (N0IsConst && N1IsConst)
1991 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0.getNode(), N1.getNode());
1993 // canonicalize constant to RHS
1994 if (N0IsConst && !N1IsConst)
1995 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0);
1996 // fold (mul x, 0) -> 0
1997 if (N1IsConst && ConstValue1 == 0)
1999 // We require a splat of the entire scalar bit width for non-contiguous
2002 ConstValue1.getBitWidth() == VT.getScalarType().getSizeInBits();
2003 // fold (mul x, 1) -> x
2004 if (N1IsConst && ConstValue1 == 1 && IsFullSplat)
2006 // fold (mul x, -1) -> 0-x
2007 if (N1IsConst && ConstValue1.isAllOnesValue())
2008 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2009 DAG.getConstant(0, VT), N0);
2010 // fold (mul x, (1 << c)) -> x << c
2011 if (N1IsConst && ConstValue1.isPowerOf2() && IsFullSplat)
2012 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
2013 DAG.getConstant(ConstValue1.logBase2(),
2014 getShiftAmountTy(N0.getValueType())));
2015 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
2016 if (N1IsConst && (-ConstValue1).isPowerOf2() && IsFullSplat) {
2017 unsigned Log2Val = (-ConstValue1).logBase2();
2018 // FIXME: If the input is something that is easily negated (e.g. a
2019 // single-use add), we should put the negate there.
2020 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2021 DAG.getConstant(0, VT),
2022 DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
2023 DAG.getConstant(Log2Val,
2024 getShiftAmountTy(N0.getValueType()))));
2028 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
2029 if (N1IsConst && N0.getOpcode() == ISD::SHL &&
2030 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2031 isa<ConstantSDNode>(N0.getOperand(1)))) {
2032 SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT,
2033 N1, N0.getOperand(1));
2034 AddToWorklist(C3.getNode());
2035 return DAG.getNode(ISD::MUL, SDLoc(N), VT,
2036 N0.getOperand(0), C3);
2039 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
2042 SDValue Sh(nullptr,0), Y(nullptr,0);
2043 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
2044 if (N0.getOpcode() == ISD::SHL &&
2045 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2046 isa<ConstantSDNode>(N0.getOperand(1))) &&
2047 N0.getNode()->hasOneUse()) {
2049 } else if (N1.getOpcode() == ISD::SHL &&
2050 isa<ConstantSDNode>(N1.getOperand(1)) &&
2051 N1.getNode()->hasOneUse()) {
2056 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2057 Sh.getOperand(0), Y);
2058 return DAG.getNode(ISD::SHL, SDLoc(N), VT,
2059 Mul, Sh.getOperand(1));
2063 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
2064 if (N1IsConst && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
2065 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2066 isa<ConstantSDNode>(N0.getOperand(1))))
2067 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
2068 DAG.getNode(ISD::MUL, SDLoc(N0), VT,
2069 N0.getOperand(0), N1),
2070 DAG.getNode(ISD::MUL, SDLoc(N1), VT,
2071 N0.getOperand(1), N1));
2074 SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1);
2081 SDValue DAGCombiner::visitSDIV(SDNode *N) {
2082 SDValue N0 = N->getOperand(0);
2083 SDValue N1 = N->getOperand(1);
2084 EVT VT = N->getValueType(0);
2087 if (VT.isVector()) {
2088 SDValue FoldedVOp = SimplifyVBinOp(N);
2089 if (FoldedVOp.getNode()) return FoldedVOp;
2092 // fold (sdiv c1, c2) -> c1/c2
2093 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2094 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2095 if (N0C && N1C && !N1C->isNullValue())
2096 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
2097 // fold (sdiv X, 1) -> X
2098 if (N1C && N1C->getAPIntValue() == 1LL)
2100 // fold (sdiv X, -1) -> 0-X
2101 if (N1C && N1C->isAllOnesValue())
2102 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2103 DAG.getConstant(0, VT), N0);
2104 // If we know the sign bits of both operands are zero, strength reduce to a
2105 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
2106 if (!VT.isVector()) {
2107 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2108 return DAG.getNode(ISD::UDIV, SDLoc(N), N1.getValueType(),
2112 // fold (sdiv X, pow2) -> simple ops after legalize
2113 if (N1C && !N1C->isNullValue() && (N1C->getAPIntValue().isPowerOf2() ||
2114 (-N1C->getAPIntValue()).isPowerOf2())) {
2115 // If dividing by powers of two is cheap, then don't perform the following
2117 if (TLI.isPow2SDivCheap())
2120 // Target-specific implementation of sdiv x, pow2.
2121 SDValue Res = BuildSDIVPow2(N);
2125 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
2127 // Splat the sign bit into the register
2129 DAG.getNode(ISD::SRA, SDLoc(N), VT, N0,
2130 DAG.getConstant(VT.getScalarSizeInBits() - 1,
2131 getShiftAmountTy(N0.getValueType())));
2132 AddToWorklist(SGN.getNode());
2134 // Add (N0 < 0) ? abs2 - 1 : 0;
2136 DAG.getNode(ISD::SRL, SDLoc(N), VT, SGN,
2137 DAG.getConstant(VT.getScalarSizeInBits() - lg2,
2138 getShiftAmountTy(SGN.getValueType())));
2139 SDValue ADD = DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, SRL);
2140 AddToWorklist(SRL.getNode());
2141 AddToWorklist(ADD.getNode()); // Divide by pow2
2142 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), VT, ADD,
2143 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType())));
2145 // If we're dividing by a positive value, we're done. Otherwise, we must
2146 // negate the result.
2147 if (N1C->getAPIntValue().isNonNegative())
2150 AddToWorklist(SRA.getNode());
2151 return DAG.getNode(ISD::SUB, SDLoc(N), VT, DAG.getConstant(0, VT), SRA);
2154 // if integer divide is expensive and we satisfy the requirements, emit an
2155 // alternate sequence.
2156 if (N1C && !TLI.isIntDivCheap()) {
2157 SDValue Op = BuildSDIV(N);
2158 if (Op.getNode()) return Op;
2162 if (N0.getOpcode() == ISD::UNDEF)
2163 return DAG.getConstant(0, VT);
2164 // X / undef -> undef
2165 if (N1.getOpcode() == ISD::UNDEF)
2171 SDValue DAGCombiner::visitUDIV(SDNode *N) {
2172 SDValue N0 = N->getOperand(0);
2173 SDValue N1 = N->getOperand(1);
2174 EVT VT = N->getValueType(0);
2177 if (VT.isVector()) {
2178 SDValue FoldedVOp = SimplifyVBinOp(N);
2179 if (FoldedVOp.getNode()) return FoldedVOp;
2182 // fold (udiv c1, c2) -> c1/c2
2183 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2184 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2185 if (N0C && N1C && !N1C->isNullValue())
2186 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
2187 // fold (udiv x, (1 << c)) -> x >>u c
2188 if (N1C && N1C->getAPIntValue().isPowerOf2())
2189 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0,
2190 DAG.getConstant(N1C->getAPIntValue().logBase2(),
2191 getShiftAmountTy(N0.getValueType())));
2192 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
2193 if (N1.getOpcode() == ISD::SHL) {
2194 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2195 if (SHC->getAPIntValue().isPowerOf2()) {
2196 EVT ADDVT = N1.getOperand(1).getValueType();
2197 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N), ADDVT,
2199 DAG.getConstant(SHC->getAPIntValue()
2202 AddToWorklist(Add.getNode());
2203 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, Add);
2207 // fold (udiv x, c) -> alternate
2208 if (N1C && !TLI.isIntDivCheap()) {
2209 SDValue Op = BuildUDIV(N);
2210 if (Op.getNode()) return Op;
2214 if (N0.getOpcode() == ISD::UNDEF)
2215 return DAG.getConstant(0, VT);
2216 // X / undef -> undef
2217 if (N1.getOpcode() == ISD::UNDEF)
2223 SDValue DAGCombiner::visitSREM(SDNode *N) {
2224 SDValue N0 = N->getOperand(0);
2225 SDValue N1 = N->getOperand(1);
2226 EVT VT = N->getValueType(0);
2228 // fold (srem c1, c2) -> c1%c2
2229 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2230 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2231 if (N0C && N1C && !N1C->isNullValue())
2232 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
2233 // If we know the sign bits of both operands are zero, strength reduce to a
2234 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
2235 if (!VT.isVector()) {
2236 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2237 return DAG.getNode(ISD::UREM, SDLoc(N), VT, N0, N1);
2240 // If X/C can be simplified by the division-by-constant logic, lower
2241 // X%C to the equivalent of X-X/C*C.
2242 if (N1C && !N1C->isNullValue()) {
2243 SDValue Div = DAG.getNode(ISD::SDIV, SDLoc(N), VT, N0, N1);
2244 AddToWorklist(Div.getNode());
2245 SDValue OptimizedDiv = combine(Div.getNode());
2246 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2247 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2249 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2250 AddToWorklist(Mul.getNode());
2256 if (N0.getOpcode() == ISD::UNDEF)
2257 return DAG.getConstant(0, VT);
2258 // X % undef -> undef
2259 if (N1.getOpcode() == ISD::UNDEF)
2265 SDValue DAGCombiner::visitUREM(SDNode *N) {
2266 SDValue N0 = N->getOperand(0);
2267 SDValue N1 = N->getOperand(1);
2268 EVT VT = N->getValueType(0);
2270 // fold (urem c1, c2) -> c1%c2
2271 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2272 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2273 if (N0C && N1C && !N1C->isNullValue())
2274 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
2275 // fold (urem x, pow2) -> (and x, pow2-1)
2276 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
2277 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0,
2278 DAG.getConstant(N1C->getAPIntValue()-1,VT));
2279 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2280 if (N1.getOpcode() == ISD::SHL) {
2281 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2282 if (SHC->getAPIntValue().isPowerOf2()) {
2284 DAG.getNode(ISD::ADD, SDLoc(N), VT, N1,
2285 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
2287 AddToWorklist(Add.getNode());
2288 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, Add);
2293 // If X/C can be simplified by the division-by-constant logic, lower
2294 // X%C to the equivalent of X-X/C*C.
2295 if (N1C && !N1C->isNullValue()) {
2296 SDValue Div = DAG.getNode(ISD::UDIV, SDLoc(N), VT, N0, N1);
2297 AddToWorklist(Div.getNode());
2298 SDValue OptimizedDiv = combine(Div.getNode());
2299 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2300 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2302 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2303 AddToWorklist(Mul.getNode());
2309 if (N0.getOpcode() == ISD::UNDEF)
2310 return DAG.getConstant(0, VT);
2311 // X % undef -> undef
2312 if (N1.getOpcode() == ISD::UNDEF)
2318 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2319 SDValue N0 = N->getOperand(0);
2320 SDValue N1 = N->getOperand(1);
2321 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2322 EVT VT = N->getValueType(0);
2325 // fold (mulhs x, 0) -> 0
2326 if (N1C && N1C->isNullValue())
2328 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2329 if (N1C && N1C->getAPIntValue() == 1)
2330 return DAG.getNode(ISD::SRA, SDLoc(N), N0.getValueType(), N0,
2331 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2332 getShiftAmountTy(N0.getValueType())));
2333 // fold (mulhs x, undef) -> 0
2334 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2335 return DAG.getConstant(0, VT);
2337 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2339 if (VT.isSimple() && !VT.isVector()) {
2340 MVT Simple = VT.getSimpleVT();
2341 unsigned SimpleSize = Simple.getSizeInBits();
2342 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2343 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2344 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2345 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2346 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2347 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2348 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2349 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2356 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2357 SDValue N0 = N->getOperand(0);
2358 SDValue N1 = N->getOperand(1);
2359 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2360 EVT VT = N->getValueType(0);
2363 // fold (mulhu x, 0) -> 0
2364 if (N1C && N1C->isNullValue())
2366 // fold (mulhu x, 1) -> 0
2367 if (N1C && N1C->getAPIntValue() == 1)
2368 return DAG.getConstant(0, N0.getValueType());
2369 // fold (mulhu x, undef) -> 0
2370 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2371 return DAG.getConstant(0, VT);
2373 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2375 if (VT.isSimple() && !VT.isVector()) {
2376 MVT Simple = VT.getSimpleVT();
2377 unsigned SimpleSize = Simple.getSizeInBits();
2378 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2379 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2380 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2381 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2382 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2383 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2384 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2385 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2392 /// Perform optimizations common to nodes that compute two values. LoOp and HiOp
2393 /// give the opcodes for the two computations that are being performed. Return
2394 /// true if a simplification was made.
2395 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2397 // If the high half is not needed, just compute the low half.
2398 bool HiExists = N->hasAnyUseOfValue(1);
2400 (!LegalOperations ||
2401 TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) {
2402 SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2403 return CombineTo(N, Res, Res);
2406 // If the low half is not needed, just compute the high half.
2407 bool LoExists = N->hasAnyUseOfValue(0);
2409 (!LegalOperations ||
2410 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2411 SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2412 return CombineTo(N, Res, Res);
2415 // If both halves are used, return as it is.
2416 if (LoExists && HiExists)
2419 // If the two computed results can be simplified separately, separate them.
2421 SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2422 AddToWorklist(Lo.getNode());
2423 SDValue LoOpt = combine(Lo.getNode());
2424 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2425 (!LegalOperations ||
2426 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2427 return CombineTo(N, LoOpt, LoOpt);
2431 SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2432 AddToWorklist(Hi.getNode());
2433 SDValue HiOpt = combine(Hi.getNode());
2434 if (HiOpt.getNode() && HiOpt != Hi &&
2435 (!LegalOperations ||
2436 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2437 return CombineTo(N, HiOpt, HiOpt);
2443 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2444 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2445 if (Res.getNode()) return Res;
2447 EVT VT = N->getValueType(0);
2450 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2452 if (VT.isSimple() && !VT.isVector()) {
2453 MVT Simple = VT.getSimpleVT();
2454 unsigned SimpleSize = Simple.getSizeInBits();
2455 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2456 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2457 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2458 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2459 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2460 // Compute the high part as N1.
2461 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2462 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2463 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2464 // Compute the low part as N0.
2465 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2466 return CombineTo(N, Lo, Hi);
2473 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2474 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2475 if (Res.getNode()) return Res;
2477 EVT VT = N->getValueType(0);
2480 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2482 if (VT.isSimple() && !VT.isVector()) {
2483 MVT Simple = VT.getSimpleVT();
2484 unsigned SimpleSize = Simple.getSizeInBits();
2485 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2486 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2487 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2488 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2489 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2490 // Compute the high part as N1.
2491 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2492 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2493 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2494 // Compute the low part as N0.
2495 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2496 return CombineTo(N, Lo, Hi);
2503 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2504 // (smulo x, 2) -> (saddo x, x)
2505 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2506 if (C2->getAPIntValue() == 2)
2507 return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(),
2508 N->getOperand(0), N->getOperand(0));
2513 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2514 // (umulo x, 2) -> (uaddo x, x)
2515 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2516 if (C2->getAPIntValue() == 2)
2517 return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(),
2518 N->getOperand(0), N->getOperand(0));
2523 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2524 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2525 if (Res.getNode()) return Res;
2530 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2531 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2532 if (Res.getNode()) return Res;
2537 /// If this is a binary operator with two operands of the same opcode, try to
2539 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2540 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2541 EVT VT = N0.getValueType();
2542 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2544 // Bail early if none of these transforms apply.
2545 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2547 // For each of OP in AND/OR/XOR:
2548 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2549 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2550 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2551 // fold (OP (bswap x), (bswap y)) -> (bswap (OP x, y))
2552 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2554 // do not sink logical op inside of a vector extend, since it may combine
2556 EVT Op0VT = N0.getOperand(0).getValueType();
2557 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2558 N0.getOpcode() == ISD::SIGN_EXTEND ||
2559 N0.getOpcode() == ISD::BSWAP ||
2560 // Avoid infinite looping with PromoteIntBinOp.
2561 (N0.getOpcode() == ISD::ANY_EXTEND &&
2562 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2563 (N0.getOpcode() == ISD::TRUNCATE &&
2564 (!TLI.isZExtFree(VT, Op0VT) ||
2565 !TLI.isTruncateFree(Op0VT, VT)) &&
2566 TLI.isTypeLegal(Op0VT))) &&
2568 Op0VT == N1.getOperand(0).getValueType() &&
2569 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2570 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2571 N0.getOperand(0).getValueType(),
2572 N0.getOperand(0), N1.getOperand(0));
2573 AddToWorklist(ORNode.getNode());
2574 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode);
2577 // For each of OP in SHL/SRL/SRA/AND...
2578 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2579 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2580 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2581 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2582 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2583 N0.getOperand(1) == N1.getOperand(1)) {
2584 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2585 N0.getOperand(0).getValueType(),
2586 N0.getOperand(0), N1.getOperand(0));
2587 AddToWorklist(ORNode.getNode());
2588 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
2589 ORNode, N0.getOperand(1));
2592 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2593 // Only perform this optimization after type legalization and before
2594 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2595 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2596 // we don't want to undo this promotion.
2597 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2599 if ((N0.getOpcode() == ISD::BITCAST ||
2600 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) &&
2601 Level == AfterLegalizeTypes) {
2602 SDValue In0 = N0.getOperand(0);
2603 SDValue In1 = N1.getOperand(0);
2604 EVT In0Ty = In0.getValueType();
2605 EVT In1Ty = In1.getValueType();
2607 // If both incoming values are integers, and the original types are the
2609 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2610 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1);
2611 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op);
2612 AddToWorklist(Op.getNode());
2617 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2618 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2619 // If both shuffles use the same mask, and both shuffle within a single
2620 // vector, then it is worthwhile to move the swizzle after the operation.
2621 // The type-legalizer generates this pattern when loading illegal
2622 // vector types from memory. In many cases this allows additional shuffle
2624 // There are other cases where moving the shuffle after the xor/and/or
2625 // is profitable even if shuffles don't perform a swizzle.
2626 // If both shuffles use the same mask, and both shuffles have the same first
2627 // or second operand, then it might still be profitable to move the shuffle
2628 // after the xor/and/or operation.
2629 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) {
2630 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2631 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2633 assert(N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() &&
2634 "Inputs to shuffles are not the same type");
2636 // Check that both shuffles use the same mask. The masks are known to be of
2637 // the same length because the result vector type is the same.
2638 // Check also that shuffles have only one use to avoid introducing extra
2640 if (SVN0->hasOneUse() && SVN1->hasOneUse() &&
2641 SVN0->getMask().equals(SVN1->getMask())) {
2642 SDValue ShOp = N0->getOperand(1);
2644 // Don't try to fold this node if it requires introducing a
2645 // build vector of all zeros that might be illegal at this stage.
2646 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2648 ShOp = DAG.getConstant(0, VT);
2653 // (AND (shuf (A, C), shuf (B, C)) -> shuf (AND (A, B), C)
2654 // (OR (shuf (A, C), shuf (B, C)) -> shuf (OR (A, B), C)
2655 // (XOR (shuf (A, C), shuf (B, C)) -> shuf (XOR (A, B), V_0)
2656 if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) {
2657 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2658 N0->getOperand(0), N1->getOperand(0));
2659 AddToWorklist(NewNode.getNode());
2660 return DAG.getVectorShuffle(VT, SDLoc(N), NewNode, ShOp,
2661 &SVN0->getMask()[0]);
2664 // Don't try to fold this node if it requires introducing a
2665 // build vector of all zeros that might be illegal at this stage.
2666 ShOp = N0->getOperand(0);
2667 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2669 ShOp = DAG.getConstant(0, VT);
2674 // (AND (shuf (C, A), shuf (C, B)) -> shuf (C, AND (A, B))
2675 // (OR (shuf (C, A), shuf (C, B)) -> shuf (C, OR (A, B))
2676 // (XOR (shuf (C, A), shuf (C, B)) -> shuf (V_0, XOR (A, B))
2677 if (N0->getOperand(0) == N1->getOperand(0) && ShOp.getNode()) {
2678 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2679 N0->getOperand(1), N1->getOperand(1));
2680 AddToWorklist(NewNode.getNode());
2681 return DAG.getVectorShuffle(VT, SDLoc(N), ShOp, NewNode,
2682 &SVN0->getMask()[0]);
2690 /// This contains all DAGCombine rules which reduce two values combined by
2691 /// an And operation to a single value. This makes them reusable in the context
2692 /// of visitSELECT(). Rules involving constants are not included as
2693 /// visitSELECT() already handles those cases.
2694 SDValue DAGCombiner::visitANDLike(SDValue N0, SDValue N1,
2695 SDNode *LocReference) {
2696 EVT VT = N1.getValueType();
2698 // fold (and x, undef) -> 0
2699 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2700 return DAG.getConstant(0, VT);
2701 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2702 SDValue LL, LR, RL, RR, CC0, CC1;
2703 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2704 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2705 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2707 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2708 LL.getValueType().isInteger()) {
2709 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2710 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
2711 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2712 LR.getValueType(), LL, RL);
2713 AddToWorklist(ORNode.getNode());
2714 return DAG.getSetCC(SDLoc(LocReference), VT, ORNode, LR, Op1);
2716 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2717 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2718 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(N0),
2719 LR.getValueType(), LL, RL);
2720 AddToWorklist(ANDNode.getNode());
2721 return DAG.getSetCC(SDLoc(LocReference), VT, ANDNode, LR, Op1);
2723 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2724 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2725 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2726 LR.getValueType(), LL, RL);
2727 AddToWorklist(ORNode.getNode());
2728 return DAG.getSetCC(SDLoc(LocReference), VT, ORNode, LR, Op1);
2731 // Simplify (and (setne X, 0), (setne X, -1)) -> (setuge (add X, 1), 2)
2732 if (LL == RL && isa<ConstantSDNode>(LR) && isa<ConstantSDNode>(RR) &&
2733 Op0 == Op1 && LL.getValueType().isInteger() &&
2734 Op0 == ISD::SETNE && ((cast<ConstantSDNode>(LR)->isNullValue() &&
2735 cast<ConstantSDNode>(RR)->isAllOnesValue()) ||
2736 (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
2737 cast<ConstantSDNode>(RR)->isNullValue()))) {
2738 SDValue ADDNode = DAG.getNode(ISD::ADD, SDLoc(N0), LL.getValueType(),
2739 LL, DAG.getConstant(1, LL.getValueType()));
2740 AddToWorklist(ADDNode.getNode());
2741 return DAG.getSetCC(SDLoc(LocReference), VT, ADDNode,
2742 DAG.getConstant(2, LL.getValueType()), ISD::SETUGE);
2744 // canonicalize equivalent to ll == rl
2745 if (LL == RR && LR == RL) {
2746 Op1 = ISD::getSetCCSwappedOperands(Op1);
2749 if (LL == RL && LR == RR) {
2750 bool isInteger = LL.getValueType().isInteger();
2751 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2752 if (Result != ISD::SETCC_INVALID &&
2753 (!LegalOperations ||
2754 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
2755 TLI.isOperationLegal(ISD::SETCC,
2756 getSetCCResultType(N0.getSimpleValueType())))))
2757 return DAG.getSetCC(SDLoc(LocReference), N0.getValueType(),
2762 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
2763 VT.getSizeInBits() <= 64) {
2764 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2765 APInt ADDC = ADDI->getAPIntValue();
2766 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2767 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
2768 // immediate for an add, but it is legal if its top c2 bits are set,
2769 // transform the ADD so the immediate doesn't need to be materialized
2771 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
2772 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
2773 SRLI->getZExtValue());
2774 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
2776 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2778 DAG.getNode(ISD::ADD, SDLoc(N0), VT,
2779 N0.getOperand(0), DAG.getConstant(ADDC, VT));
2780 CombineTo(N0.getNode(), NewAdd);
2781 // Return N so it doesn't get rechecked!
2782 return SDValue(LocReference, 0);
2793 SDValue DAGCombiner::visitAND(SDNode *N) {
2794 SDValue N0 = N->getOperand(0);
2795 SDValue N1 = N->getOperand(1);
2796 EVT VT = N1.getValueType();
2799 if (VT.isVector()) {
2800 SDValue FoldedVOp = SimplifyVBinOp(N);
2801 if (FoldedVOp.getNode()) return FoldedVOp;
2803 // fold (and x, 0) -> 0, vector edition
2804 if (ISD::isBuildVectorAllZeros(N0.getNode()))
2805 // do not return N0, because undef node may exist in N0
2806 return DAG.getConstant(
2807 APInt::getNullValue(
2808 N0.getValueType().getScalarType().getSizeInBits()),
2810 if (ISD::isBuildVectorAllZeros(N1.getNode()))
2811 // do not return N1, because undef node may exist in N1
2812 return DAG.getConstant(
2813 APInt::getNullValue(
2814 N1.getValueType().getScalarType().getSizeInBits()),
2817 // fold (and x, -1) -> x, vector edition
2818 if (ISD::isBuildVectorAllOnes(N0.getNode()))
2820 if (ISD::isBuildVectorAllOnes(N1.getNode()))
2824 // fold (and c1, c2) -> c1&c2
2825 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2826 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2828 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2829 // canonicalize constant to RHS
2831 return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0);
2832 // fold (and x, -1) -> x
2833 if (N1C && N1C->isAllOnesValue())
2835 // if (and x, c) is known to be zero, return 0
2836 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2837 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2838 APInt::getAllOnesValue(BitWidth)))
2839 return DAG.getConstant(0, VT);
2841 SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1);
2844 // fold (and (or x, C), D) -> D if (C & D) == D
2845 if (N1C && N0.getOpcode() == ISD::OR)
2846 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2847 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2849 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2850 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2851 SDValue N0Op0 = N0.getOperand(0);
2852 APInt Mask = ~N1C->getAPIntValue();
2853 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2854 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2855 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
2856 N0.getValueType(), N0Op0);
2858 // Replace uses of the AND with uses of the Zero extend node.
2861 // We actually want to replace all uses of the any_extend with the
2862 // zero_extend, to avoid duplicating things. This will later cause this
2863 // AND to be folded.
2864 CombineTo(N0.getNode(), Zext);
2865 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2868 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2869 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2870 // already be zero by virtue of the width of the base type of the load.
2872 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2874 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2875 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2876 N0.getOpcode() == ISD::LOAD) {
2877 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2878 N0 : N0.getOperand(0) );
2880 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2881 // This can be a pure constant or a vector splat, in which case we treat the
2882 // vector as a scalar and use the splat value.
2883 APInt Constant = APInt::getNullValue(1);
2884 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2885 Constant = C->getAPIntValue();
2886 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2887 APInt SplatValue, SplatUndef;
2888 unsigned SplatBitSize;
2890 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2891 SplatBitSize, HasAnyUndefs);
2893 // Undef bits can contribute to a possible optimisation if set, so
2895 SplatValue |= SplatUndef;
2897 // The splat value may be something like "0x00FFFFFF", which means 0 for
2898 // the first vector value and FF for the rest, repeating. We need a mask
2899 // that will apply equally to all members of the vector, so AND all the
2900 // lanes of the constant together.
2901 EVT VT = Vector->getValueType(0);
2902 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2904 // If the splat value has been compressed to a bitlength lower
2905 // than the size of the vector lane, we need to re-expand it to
2907 if (BitWidth > SplatBitSize)
2908 for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
2909 SplatBitSize < BitWidth;
2910 SplatBitSize = SplatBitSize * 2)
2911 SplatValue |= SplatValue.shl(SplatBitSize);
2913 Constant = APInt::getAllOnesValue(BitWidth);
2914 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
2915 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
2919 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
2920 // actually legal and isn't going to get expanded, else this is a false
2922 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
2923 Load->getValueType(0),
2924 Load->getMemoryVT());
2926 // Resize the constant to the same size as the original memory access before
2927 // extension. If it is still the AllOnesValue then this AND is completely
2930 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
2933 switch (Load->getExtensionType()) {
2934 default: B = false; break;
2935 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
2937 case ISD::NON_EXTLOAD: B = true; break;
2940 if (B && Constant.isAllOnesValue()) {
2941 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
2942 // preserve semantics once we get rid of the AND.
2943 SDValue NewLoad(Load, 0);
2944 if (Load->getExtensionType() == ISD::EXTLOAD) {
2945 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
2946 Load->getValueType(0), SDLoc(Load),
2947 Load->getChain(), Load->getBasePtr(),
2948 Load->getOffset(), Load->getMemoryVT(),
2949 Load->getMemOperand());
2950 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
2951 if (Load->getNumValues() == 3) {
2952 // PRE/POST_INC loads have 3 values.
2953 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
2954 NewLoad.getValue(2) };
2955 CombineTo(Load, To, 3, true);
2957 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
2961 // Fold the AND away, taking care not to fold to the old load node if we
2963 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
2965 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2969 // fold (and (load x), 255) -> (zextload x, i8)
2970 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2971 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2972 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2973 (N0.getOpcode() == ISD::ANY_EXTEND &&
2974 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2975 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2976 LoadSDNode *LN0 = HasAnyExt
2977 ? cast<LoadSDNode>(N0.getOperand(0))
2978 : cast<LoadSDNode>(N0);
2979 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2980 LN0->isUnindexed() && N0.hasOneUse() && SDValue(LN0, 0).hasOneUse()) {
2981 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2982 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2983 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2984 EVT LoadedVT = LN0->getMemoryVT();
2985 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2987 if (ExtVT == LoadedVT &&
2988 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy,
2992 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
2993 LN0->getChain(), LN0->getBasePtr(), ExtVT,
2994 LN0->getMemOperand());
2996 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
2997 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3000 // Do not change the width of a volatile load.
3001 // Do not generate loads of non-round integer types since these can
3002 // be expensive (and would be wrong if the type is not byte sized).
3003 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
3004 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy,
3006 EVT PtrType = LN0->getOperand(1).getValueType();
3008 unsigned Alignment = LN0->getAlignment();
3009 SDValue NewPtr = LN0->getBasePtr();
3011 // For big endian targets, we need to add an offset to the pointer
3012 // to load the correct bytes. For little endian systems, we merely
3013 // need to read fewer bytes from the same pointer.
3014 if (TLI.isBigEndian()) {
3015 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
3016 unsigned EVTStoreBytes = ExtVT.getStoreSize();
3017 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
3018 NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0), PtrType,
3019 NewPtr, DAG.getConstant(PtrOff, PtrType));
3020 Alignment = MinAlign(Alignment, PtrOff);
3023 AddToWorklist(NewPtr.getNode());
3026 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
3027 LN0->getChain(), NewPtr,
3028 LN0->getPointerInfo(),
3029 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
3030 LN0->isInvariant(), Alignment, LN0->getAAInfo());
3032 CombineTo(LN0, Load, Load.getValue(1));
3033 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3039 if (SDValue Combined = visitANDLike(N0, N1, N))
3042 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
3043 if (N0.getOpcode() == N1.getOpcode()) {
3044 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3045 if (Tmp.getNode()) return Tmp;
3048 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
3049 // fold (and (sra)) -> (and (srl)) when possible.
3050 if (!VT.isVector() &&
3051 SimplifyDemandedBits(SDValue(N, 0)))
3052 return SDValue(N, 0);
3054 // fold (zext_inreg (extload x)) -> (zextload x)
3055 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
3056 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3057 EVT MemVT = LN0->getMemoryVT();
3058 // If we zero all the possible extended bits, then we can turn this into
3059 // a zextload if we are running before legalize or the operation is legal.
3060 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
3061 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
3062 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
3063 ((!LegalOperations && !LN0->isVolatile()) ||
3064 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) {
3065 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
3066 LN0->getChain(), LN0->getBasePtr(),
3067 MemVT, LN0->getMemOperand());
3069 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
3070 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3073 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
3074 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
3076 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3077 EVT MemVT = LN0->getMemoryVT();
3078 // If we zero all the possible extended bits, then we can turn this into
3079 // a zextload if we are running before legalize or the operation is legal.
3080 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
3081 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
3082 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
3083 ((!LegalOperations && !LN0->isVolatile()) ||
3084 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) {
3085 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
3086 LN0->getChain(), LN0->getBasePtr(),
3087 MemVT, LN0->getMemOperand());
3089 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
3090 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3093 // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const)
3094 if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) {
3095 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
3096 N0.getOperand(1), false);
3097 if (BSwap.getNode())
3104 /// Match (a >> 8) | (a << 8) as (bswap a) >> 16.
3105 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
3106 bool DemandHighBits) {
3107 if (!LegalOperations)
3110 EVT VT = N->getValueType(0);
3111 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
3113 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3116 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
3117 bool LookPassAnd0 = false;
3118 bool LookPassAnd1 = false;
3119 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
3121 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
3123 if (N0.getOpcode() == ISD::AND) {
3124 if (!N0.getNode()->hasOneUse())
3126 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3127 if (!N01C || N01C->getZExtValue() != 0xFF00)
3129 N0 = N0.getOperand(0);
3130 LookPassAnd0 = true;
3133 if (N1.getOpcode() == ISD::AND) {
3134 if (!N1.getNode()->hasOneUse())
3136 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3137 if (!N11C || N11C->getZExtValue() != 0xFF)
3139 N1 = N1.getOperand(0);
3140 LookPassAnd1 = true;
3143 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
3145 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
3147 if (!N0.getNode()->hasOneUse() ||
3148 !N1.getNode()->hasOneUse())
3151 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3152 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3155 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
3158 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
3159 SDValue N00 = N0->getOperand(0);
3160 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
3161 if (!N00.getNode()->hasOneUse())
3163 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
3164 if (!N001C || N001C->getZExtValue() != 0xFF)
3166 N00 = N00.getOperand(0);
3167 LookPassAnd0 = true;
3170 SDValue N10 = N1->getOperand(0);
3171 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
3172 if (!N10.getNode()->hasOneUse())
3174 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
3175 if (!N101C || N101C->getZExtValue() != 0xFF00)
3177 N10 = N10.getOperand(0);
3178 LookPassAnd1 = true;
3184 // Make sure everything beyond the low halfword gets set to zero since the SRL
3185 // 16 will clear the top bits.
3186 unsigned OpSizeInBits = VT.getSizeInBits();
3187 if (DemandHighBits && OpSizeInBits > 16) {
3188 // If the left-shift isn't masked out then the only way this is a bswap is
3189 // if all bits beyond the low 8 are 0. In that case the entire pattern
3190 // reduces to a left shift anyway: leave it for other parts of the combiner.
3194 // However, if the right shift isn't masked out then it might be because
3195 // it's not needed. See if we can spot that too.
3196 if (!LookPassAnd1 &&
3197 !DAG.MaskedValueIsZero(
3198 N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16)))
3202 SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00);
3203 if (OpSizeInBits > 16)
3204 Res = DAG.getNode(ISD::SRL, SDLoc(N), VT, Res,
3205 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
3209 /// Return true if the specified node is an element that makes up a 32-bit
3210 /// packed halfword byteswap.
3211 /// ((x & 0x000000ff) << 8) |
3212 /// ((x & 0x0000ff00) >> 8) |
3213 /// ((x & 0x00ff0000) << 8) |
3214 /// ((x & 0xff000000) >> 8)
3215 static bool isBSwapHWordElement(SDValue N, MutableArrayRef<SDNode *> Parts) {
3216 if (!N.getNode()->hasOneUse())
3219 unsigned Opc = N.getOpcode();
3220 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
3223 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3228 switch (N1C->getZExtValue()) {
3231 case 0xFF: Num = 0; break;
3232 case 0xFF00: Num = 1; break;
3233 case 0xFF0000: Num = 2; break;
3234 case 0xFF000000: Num = 3; break;
3237 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
3238 SDValue N0 = N.getOperand(0);
3239 if (Opc == ISD::AND) {
3240 if (Num == 0 || Num == 2) {
3242 // (x >> 8) & 0xff0000
3243 if (N0.getOpcode() != ISD::SRL)
3245 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3246 if (!C || C->getZExtValue() != 8)
3249 // (x << 8) & 0xff00
3250 // (x << 8) & 0xff000000
3251 if (N0.getOpcode() != ISD::SHL)
3253 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3254 if (!C || C->getZExtValue() != 8)
3257 } else if (Opc == ISD::SHL) {
3259 // (x & 0xff0000) << 8
3260 if (Num != 0 && Num != 2)
3262 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3263 if (!C || C->getZExtValue() != 8)
3265 } else { // Opc == ISD::SRL
3266 // (x & 0xff00) >> 8
3267 // (x & 0xff000000) >> 8
3268 if (Num != 1 && Num != 3)
3270 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3271 if (!C || C->getZExtValue() != 8)
3278 Parts[Num] = N0.getOperand(0).getNode();
3282 /// Match a 32-bit packed halfword bswap. That is
3283 /// ((x & 0x000000ff) << 8) |
3284 /// ((x & 0x0000ff00) >> 8) |
3285 /// ((x & 0x00ff0000) << 8) |
3286 /// ((x & 0xff000000) >> 8)
3287 /// => (rotl (bswap x), 16)
3288 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
3289 if (!LegalOperations)
3292 EVT VT = N->getValueType(0);
3295 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3299 // (or (or (and), (and)), (or (and), (and)))
3300 // (or (or (or (and), (and)), (and)), (and))
3301 if (N0.getOpcode() != ISD::OR)
3303 SDValue N00 = N0.getOperand(0);
3304 SDValue N01 = N0.getOperand(1);
3305 SDNode *Parts[4] = {};
3307 if (N1.getOpcode() == ISD::OR &&
3308 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) {
3309 // (or (or (and), (and)), (or (and), (and)))
3310 SDValue N000 = N00.getOperand(0);
3311 if (!isBSwapHWordElement(N000, Parts))
3314 SDValue N001 = N00.getOperand(1);
3315 if (!isBSwapHWordElement(N001, Parts))
3317 SDValue N010 = N01.getOperand(0);
3318 if (!isBSwapHWordElement(N010, Parts))
3320 SDValue N011 = N01.getOperand(1);
3321 if (!isBSwapHWordElement(N011, Parts))
3324 // (or (or (or (and), (and)), (and)), (and))
3325 if (!isBSwapHWordElement(N1, Parts))
3327 if (!isBSwapHWordElement(N01, Parts))
3329 if (N00.getOpcode() != ISD::OR)
3331 SDValue N000 = N00.getOperand(0);
3332 if (!isBSwapHWordElement(N000, Parts))
3334 SDValue N001 = N00.getOperand(1);
3335 if (!isBSwapHWordElement(N001, Parts))
3339 // Make sure the parts are all coming from the same node.
3340 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
3343 SDValue BSwap = DAG.getNode(ISD::BSWAP, SDLoc(N), VT,
3344 SDValue(Parts[0],0));
3346 // Result of the bswap should be rotated by 16. If it's not legal, then
3347 // do (x << 16) | (x >> 16).
3348 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
3349 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3350 return DAG.getNode(ISD::ROTL, SDLoc(N), VT, BSwap, ShAmt);
3351 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3352 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, BSwap, ShAmt);
3353 return DAG.getNode(ISD::OR, SDLoc(N), VT,
3354 DAG.getNode(ISD::SHL, SDLoc(N), VT, BSwap, ShAmt),
3355 DAG.getNode(ISD::SRL, SDLoc(N), VT, BSwap, ShAmt));
3358 /// This contains all DAGCombine rules which reduce two values combined by
3359 /// an Or operation to a single value \see visitANDLike().
3360 SDValue DAGCombiner::visitORLike(SDValue N0, SDValue N1, SDNode *LocReference) {
3361 EVT VT = N1.getValueType();
3362 // fold (or x, undef) -> -1
3363 if (!LegalOperations &&
3364 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3365 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3366 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
3368 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3369 SDValue LL, LR, RL, RR, CC0, CC1;
3370 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3371 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3372 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3374 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3375 LL.getValueType().isInteger()) {
3376 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3377 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3378 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3379 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3380 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(LR),
3381 LR.getValueType(), LL, RL);
3382 AddToWorklist(ORNode.getNode());
3383 return DAG.getSetCC(SDLoc(LocReference), VT, ORNode, LR, Op1);
3385 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3386 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3387 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3388 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3389 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(LR),
3390 LR.getValueType(), LL, RL);
3391 AddToWorklist(ANDNode.getNode());
3392 return DAG.getSetCC(SDLoc(LocReference), VT, ANDNode, LR, Op1);
3395 // canonicalize equivalent to ll == rl
3396 if (LL == RR && LR == RL) {
3397 Op1 = ISD::getSetCCSwappedOperands(Op1);
3400 if (LL == RL && LR == RR) {
3401 bool isInteger = LL.getValueType().isInteger();
3402 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3403 if (Result != ISD::SETCC_INVALID &&
3404 (!LegalOperations ||
3405 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
3406 TLI.isOperationLegal(ISD::SETCC,
3407 getSetCCResultType(N0.getValueType())))))
3408 return DAG.getSetCC(SDLoc(LocReference), N0.getValueType(),
3413 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3414 if (N0.getOpcode() == ISD::AND &&
3415 N1.getOpcode() == ISD::AND &&
3416 N0.getOperand(1).getOpcode() == ISD::Constant &&
3417 N1.getOperand(1).getOpcode() == ISD::Constant &&
3418 // Don't increase # computations.
3419 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3420 // We can only do this xform if we know that bits from X that are set in C2
3421 // but not in C1 are already zero. Likewise for Y.
3422 const APInt &LHSMask =
3423 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3424 const APInt &RHSMask =
3425 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3427 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3428 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3429 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3430 N0.getOperand(0), N1.getOperand(0));
3431 return DAG.getNode(ISD::AND, SDLoc(LocReference), VT, X,
3432 DAG.getConstant(LHSMask | RHSMask, VT));
3436 // (or (and X, M), (and X, N)) -> (and X, (or M, N))
3437 if (N0.getOpcode() == ISD::AND &&
3438 N1.getOpcode() == ISD::AND &&
3439 N0.getOperand(0) == N1.getOperand(0) &&
3440 // Don't increase # computations.
3441 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3442 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3443 N0.getOperand(1), N1.getOperand(1));
3444 return DAG.getNode(ISD::AND, SDLoc(LocReference), VT, N0.getOperand(0), X);
3450 SDValue DAGCombiner::visitOR(SDNode *N) {
3451 SDValue N0 = N->getOperand(0);
3452 SDValue N1 = N->getOperand(1);
3453 EVT VT = N1.getValueType();
3456 if (VT.isVector()) {
3457 SDValue FoldedVOp = SimplifyVBinOp(N);
3458 if (FoldedVOp.getNode()) return FoldedVOp;
3460 // fold (or x, 0) -> x, vector edition
3461 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3463 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3466 // fold (or x, -1) -> -1, vector edition
3467 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3468 // do not return N0, because undef node may exist in N0
3469 return DAG.getConstant(
3470 APInt::getAllOnesValue(
3471 N0.getValueType().getScalarType().getSizeInBits()),
3473 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3474 // do not return N1, because undef node may exist in N1
3475 return DAG.getConstant(
3476 APInt::getAllOnesValue(
3477 N1.getValueType().getScalarType().getSizeInBits()),
3480 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask1)
3481 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf B, A, Mask2)
3482 // Do this only if the resulting shuffle is legal.
3483 if (isa<ShuffleVectorSDNode>(N0) &&
3484 isa<ShuffleVectorSDNode>(N1) &&
3485 // Avoid folding a node with illegal type.
3486 TLI.isTypeLegal(VT) &&
3487 N0->getOperand(1) == N1->getOperand(1) &&
3488 ISD::isBuildVectorAllZeros(N0.getOperand(1).getNode())) {
3489 bool CanFold = true;
3490 unsigned NumElts = VT.getVectorNumElements();
3491 const ShuffleVectorSDNode *SV0 = cast<ShuffleVectorSDNode>(N0);
3492 const ShuffleVectorSDNode *SV1 = cast<ShuffleVectorSDNode>(N1);
3493 // We construct two shuffle masks:
3494 // - Mask1 is a shuffle mask for a shuffle with N0 as the first operand
3495 // and N1 as the second operand.
3496 // - Mask2 is a shuffle mask for a shuffle with N1 as the first operand
3497 // and N0 as the second operand.
3498 // We do this because OR is commutable and therefore there might be
3499 // two ways to fold this node into a shuffle.
3500 SmallVector<int,4> Mask1;
3501 SmallVector<int,4> Mask2;
3503 for (unsigned i = 0; i != NumElts && CanFold; ++i) {
3504 int M0 = SV0->getMaskElt(i);
3505 int M1 = SV1->getMaskElt(i);
3507 // Both shuffle indexes are undef. Propagate Undef.
3508 if (M0 < 0 && M1 < 0) {
3509 Mask1.push_back(M0);
3510 Mask2.push_back(M0);
3514 if (M0 < 0 || M1 < 0 ||
3515 (M0 < (int)NumElts && M1 < (int)NumElts) ||
3516 (M0 >= (int)NumElts && M1 >= (int)NumElts)) {
3521 Mask1.push_back(M0 < (int)NumElts ? M0 : M1 + NumElts);
3522 Mask2.push_back(M1 < (int)NumElts ? M1 : M0 + NumElts);
3526 // Fold this sequence only if the resulting shuffle is 'legal'.
3527 if (TLI.isShuffleMaskLegal(Mask1, VT))
3528 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0),
3529 N1->getOperand(0), &Mask1[0]);
3530 if (TLI.isShuffleMaskLegal(Mask2, VT))
3531 return DAG.getVectorShuffle(VT, SDLoc(N), N1->getOperand(0),
3532 N0->getOperand(0), &Mask2[0]);
3537 // fold (or c1, c2) -> c1|c2
3538 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3539 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3541 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
3542 // canonicalize constant to RHS
3544 return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0);
3545 // fold (or x, 0) -> x
3546 if (N1C && N1C->isNullValue())
3548 // fold (or x, -1) -> -1
3549 if (N1C && N1C->isAllOnesValue())
3551 // fold (or x, c) -> c iff (x & ~c) == 0
3552 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3555 if (SDValue Combined = visitORLike(N0, N1, N))
3558 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3559 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
3560 if (BSwap.getNode())
3562 BSwap = MatchBSwapHWordLow(N, N0, N1);
3563 if (BSwap.getNode())
3567 SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1);
3570 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3571 // iff (c1 & c2) == 0.
3572 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3573 isa<ConstantSDNode>(N0.getOperand(1))) {
3574 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3575 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) {
3576 if (SDValue COR = DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1))
3578 ISD::AND, SDLoc(N), VT,
3579 DAG.getNode(ISD::OR, SDLoc(N0), VT, N0.getOperand(0), N1), COR);
3583 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3584 if (N0.getOpcode() == N1.getOpcode()) {
3585 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3586 if (Tmp.getNode()) return Tmp;
3589 // See if this is some rotate idiom.
3590 if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N)))
3591 return SDValue(Rot, 0);
3593 // Simplify the operands using demanded-bits information.
3594 if (!VT.isVector() &&
3595 SimplifyDemandedBits(SDValue(N, 0)))
3596 return SDValue(N, 0);
3601 /// Match "(X shl/srl V1) & V2" where V2 may not be present.
3602 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3603 if (Op.getOpcode() == ISD::AND) {
3604 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3605 Mask = Op.getOperand(1);
3606 Op = Op.getOperand(0);
3612 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3620 // Return true if we can prove that, whenever Neg and Pos are both in the
3621 // range [0, OpSize), Neg == (Pos == 0 ? 0 : OpSize - Pos). This means that
3622 // for two opposing shifts shift1 and shift2 and a value X with OpBits bits:
3624 // (or (shift1 X, Neg), (shift2 X, Pos))
3626 // reduces to a rotate in direction shift2 by Pos or (equivalently) a rotate
3627 // in direction shift1 by Neg. The range [0, OpSize) means that we only need
3628 // to consider shift amounts with defined behavior.
3629 static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned OpSize) {
3630 // If OpSize is a power of 2 then:
3632 // (a) (Pos == 0 ? 0 : OpSize - Pos) == (OpSize - Pos) & (OpSize - 1)
3633 // (b) Neg == Neg & (OpSize - 1) whenever Neg is in [0, OpSize).
3635 // So if OpSize is a power of 2 and Neg is (and Neg', OpSize-1), we check
3636 // for the stronger condition:
3638 // Neg & (OpSize - 1) == (OpSize - Pos) & (OpSize - 1) [A]
3640 // for all Neg and Pos. Since Neg & (OpSize - 1) == Neg' & (OpSize - 1)
3641 // we can just replace Neg with Neg' for the rest of the function.
3643 // In other cases we check for the even stronger condition:
3645 // Neg == OpSize - Pos [B]
3647 // for all Neg and Pos. Note that the (or ...) then invokes undefined
3648 // behavior if Pos == 0 (and consequently Neg == OpSize).
3650 // We could actually use [A] whenever OpSize is a power of 2, but the
3651 // only extra cases that it would match are those uninteresting ones
3652 // where Neg and Pos are never in range at the same time. E.g. for
3653 // OpSize == 32, using [A] would allow a Neg of the form (sub 64, Pos)
3654 // as well as (sub 32, Pos), but:
3656 // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos))
3658 // always invokes undefined behavior for 32-bit X.
3660 // Below, Mask == OpSize - 1 when using [A] and is all-ones otherwise.
3661 unsigned MaskLoBits = 0;
3662 if (Neg.getOpcode() == ISD::AND &&
3663 isPowerOf2_64(OpSize) &&
3664 Neg.getOperand(1).getOpcode() == ISD::Constant &&
3665 cast<ConstantSDNode>(Neg.getOperand(1))->getAPIntValue() == OpSize - 1) {
3666 Neg = Neg.getOperand(0);
3667 MaskLoBits = Log2_64(OpSize);
3670 // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1.
3671 if (Neg.getOpcode() != ISD::SUB)
3673 ConstantSDNode *NegC = dyn_cast<ConstantSDNode>(Neg.getOperand(0));
3676 SDValue NegOp1 = Neg.getOperand(1);
3678 // On the RHS of [A], if Pos is Pos' & (OpSize - 1), just replace Pos with
3679 // Pos'. The truncation is redundant for the purpose of the equality.
3681 Pos.getOpcode() == ISD::AND &&
3682 Pos.getOperand(1).getOpcode() == ISD::Constant &&
3683 cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() == OpSize - 1)
3684 Pos = Pos.getOperand(0);
3686 // The condition we need is now:
3688 // (NegC - NegOp1) & Mask == (OpSize - Pos) & Mask
3690 // If NegOp1 == Pos then we need:
3692 // OpSize & Mask == NegC & Mask
3694 // (because "x & Mask" is a truncation and distributes through subtraction).
3697 Width = NegC->getAPIntValue();
3698 // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC.
3699 // Then the condition we want to prove becomes:
3701 // (NegC - NegOp1) & Mask == (OpSize - (NegOp1 + PosC)) & Mask
3703 // which, again because "x & Mask" is a truncation, becomes:
3705 // NegC & Mask == (OpSize - PosC) & Mask
3706 // OpSize & Mask == (NegC + PosC) & Mask
3707 else if (Pos.getOpcode() == ISD::ADD &&
3708 Pos.getOperand(0) == NegOp1 &&
3709 Pos.getOperand(1).getOpcode() == ISD::Constant)
3710 Width = (cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() +
3711 NegC->getAPIntValue());
3715 // Now we just need to check that OpSize & Mask == Width & Mask.
3717 // Opsize & Mask is 0 since Mask is Opsize - 1.
3718 return Width.getLoBits(MaskLoBits) == 0;
3719 return Width == OpSize;
3722 // A subroutine of MatchRotate used once we have found an OR of two opposite
3723 // shifts of Shifted. If Neg == <operand size> - Pos then the OR reduces
3724 // to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the
3725 // former being preferred if supported. InnerPos and InnerNeg are Pos and
3726 // Neg with outer conversions stripped away.
3727 SDNode *DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos,
3728 SDValue Neg, SDValue InnerPos,
3729 SDValue InnerNeg, unsigned PosOpcode,
3730 unsigned NegOpcode, SDLoc DL) {
3731 // fold (or (shl x, (*ext y)),
3732 // (srl x, (*ext (sub 32, y)))) ->
3733 // (rotl x, y) or (rotr x, (sub 32, y))
3735 // fold (or (shl x, (*ext (sub 32, y))),
3736 // (srl x, (*ext y))) ->
3737 // (rotr x, y) or (rotl x, (sub 32, y))
3738 EVT VT = Shifted.getValueType();
3739 if (matchRotateSub(InnerPos, InnerNeg, VT.getSizeInBits())) {
3740 bool HasPos = TLI.isOperationLegalOrCustom(PosOpcode, VT);
3741 return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted,
3742 HasPos ? Pos : Neg).getNode();
3748 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3749 // idioms for rotate, and if the target supports rotation instructions, generate
3751 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL) {
3752 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3753 EVT VT = LHS.getValueType();
3754 if (!TLI.isTypeLegal(VT)) return nullptr;
3756 // The target must have at least one rotate flavor.
3757 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3758 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3759 if (!HasROTL && !HasROTR) return nullptr;
3761 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3762 SDValue LHSShift; // The shift.
3763 SDValue LHSMask; // AND value if any.
3764 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3765 return nullptr; // Not part of a rotate.
3767 SDValue RHSShift; // The shift.
3768 SDValue RHSMask; // AND value if any.
3769 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3770 return nullptr; // Not part of a rotate.
3772 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3773 return nullptr; // Not shifting the same value.
3775 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3776 return nullptr; // Shifts must disagree.
3778 // Canonicalize shl to left side in a shl/srl pair.
3779 if (RHSShift.getOpcode() == ISD::SHL) {
3780 std::swap(LHS, RHS);
3781 std::swap(LHSShift, RHSShift);
3782 std::swap(LHSMask , RHSMask );
3785 unsigned OpSizeInBits = VT.getSizeInBits();
3786 SDValue LHSShiftArg = LHSShift.getOperand(0);
3787 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3788 SDValue RHSShiftArg = RHSShift.getOperand(0);
3789 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3791 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3792 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3793 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3794 RHSShiftAmt.getOpcode() == ISD::Constant) {
3795 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3796 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3797 if ((LShVal + RShVal) != OpSizeInBits)
3800 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3801 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
3803 // If there is an AND of either shifted operand, apply it to the result.
3804 if (LHSMask.getNode() || RHSMask.getNode()) {
3805 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3807 if (LHSMask.getNode()) {
3808 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3809 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3811 if (RHSMask.getNode()) {
3812 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3813 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3816 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3819 return Rot.getNode();
3822 // If there is a mask here, and we have a variable shift, we can't be sure
3823 // that we're masking out the right stuff.
3824 if (LHSMask.getNode() || RHSMask.getNode())
3827 // If the shift amount is sign/zext/any-extended just peel it off.
3828 SDValue LExtOp0 = LHSShiftAmt;
3829 SDValue RExtOp0 = RHSShiftAmt;
3830 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3831 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3832 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3833 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3834 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3835 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3836 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3837 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3838 LExtOp0 = LHSShiftAmt.getOperand(0);
3839 RExtOp0 = RHSShiftAmt.getOperand(0);
3842 SDNode *TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt,
3843 LExtOp0, RExtOp0, ISD::ROTL, ISD::ROTR, DL);
3847 SDNode *TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt,
3848 RExtOp0, LExtOp0, ISD::ROTR, ISD::ROTL, DL);
3855 SDValue DAGCombiner::visitXOR(SDNode *N) {
3856 SDValue N0 = N->getOperand(0);
3857 SDValue N1 = N->getOperand(1);
3858 EVT VT = N0.getValueType();
3861 if (VT.isVector()) {
3862 SDValue FoldedVOp = SimplifyVBinOp(N);
3863 if (FoldedVOp.getNode()) return FoldedVOp;
3865 // fold (xor x, 0) -> x, vector edition
3866 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3868 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3872 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3873 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3874 return DAG.getConstant(0, VT);
3875 // fold (xor x, undef) -> undef
3876 if (N0.getOpcode() == ISD::UNDEF)
3878 if (N1.getOpcode() == ISD::UNDEF)
3880 // fold (xor c1, c2) -> c1^c2
3881 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3882 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3884 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3885 // canonicalize constant to RHS
3887 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
3888 // fold (xor x, 0) -> x
3889 if (N1C && N1C->isNullValue())
3892 SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1);
3896 // fold !(x cc y) -> (x !cc y)
3897 SDValue LHS, RHS, CC;
3898 if (TLI.isConstTrueVal(N1.getNode()) && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3899 bool isInt = LHS.getValueType().isInteger();
3900 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3903 if (!LegalOperations ||
3904 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) {
3905 switch (N0.getOpcode()) {
3907 llvm_unreachable("Unhandled SetCC Equivalent!");
3909 return DAG.getSetCC(SDLoc(N), VT, LHS, RHS, NotCC);
3910 case ISD::SELECT_CC:
3911 return DAG.getSelectCC(SDLoc(N), LHS, RHS, N0.getOperand(2),
3912 N0.getOperand(3), NotCC);
3917 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3918 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3919 N0.getNode()->hasOneUse() &&
3920 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3921 SDValue V = N0.getOperand(0);
3922 V = DAG.getNode(ISD::XOR, SDLoc(N0), V.getValueType(), V,
3923 DAG.getConstant(1, V.getValueType()));
3924 AddToWorklist(V.getNode());
3925 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V);
3928 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3929 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3930 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3931 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3932 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3933 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3934 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3935 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3936 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
3937 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3940 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3941 if (N1C && N1C->isAllOnesValue() &&
3942 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3943 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3944 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3945 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3946 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3947 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3948 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
3949 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3952 // fold (xor (and x, y), y) -> (and (not x), y)
3953 if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3954 N0->getOperand(1) == N1) {
3955 SDValue X = N0->getOperand(0);
3956 SDValue NotX = DAG.getNOT(SDLoc(X), X, VT);
3957 AddToWorklist(NotX.getNode());
3958 return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1);
3960 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3961 if (N1C && N0.getOpcode() == ISD::XOR) {
3962 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3963 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3965 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(1),
3966 DAG.getConstant(N1C->getAPIntValue() ^
3967 N00C->getAPIntValue(), VT));
3969 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(0),
3970 DAG.getConstant(N1C->getAPIntValue() ^
3971 N01C->getAPIntValue(), VT));
3973 // fold (xor x, x) -> 0
3975 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
3977 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
3978 if (N0.getOpcode() == N1.getOpcode()) {
3979 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3980 if (Tmp.getNode()) return Tmp;
3983 // Simplify the expression using non-local knowledge.
3984 if (!VT.isVector() &&
3985 SimplifyDemandedBits(SDValue(N, 0)))
3986 return SDValue(N, 0);
3991 /// Handle transforms common to the three shifts, when the shift amount is a
3993 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, ConstantSDNode *Amt) {
3994 // We can't and shouldn't fold opaque constants.
3995 if (Amt->isOpaque())
3998 SDNode *LHS = N->getOperand(0).getNode();
3999 if (!LHS->hasOneUse()) return SDValue();
4001 // We want to pull some binops through shifts, so that we have (and (shift))
4002 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
4003 // thing happens with address calculations, so it's important to canonicalize
4005 bool HighBitSet = false; // Can we transform this if the high bit is set?
4007 switch (LHS->getOpcode()) {
4008 default: return SDValue();
4011 HighBitSet = false; // We can only transform sra if the high bit is clear.
4014 HighBitSet = true; // We can only transform sra if the high bit is set.
4017 if (N->getOpcode() != ISD::SHL)
4018 return SDValue(); // only shl(add) not sr[al](add).
4019 HighBitSet = false; // We can only transform sra if the high bit is clear.
4023 // We require the RHS of the binop to be a constant and not opaque as well.
4024 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
4025 if (!BinOpCst || BinOpCst->isOpaque()) return SDValue();
4027 // FIXME: disable this unless the input to the binop is a shift by a constant.
4028 // If it is not a shift, it pessimizes some common cases like:
4030 // void foo(int *X, int i) { X[i & 1235] = 1; }
4031 // int bar(int *X, int i) { return X[i & 255]; }
4032 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
4033 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
4034 BinOpLHSVal->getOpcode() != ISD::SRA &&
4035 BinOpLHSVal->getOpcode() != ISD::SRL) ||
4036 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
4039 EVT VT = N->getValueType(0);
4041 // If this is a signed shift right, and the high bit is modified by the
4042 // logical operation, do not perform the transformation. The highBitSet
4043 // boolean indicates the value of the high bit of the constant which would
4044 // cause it to be modified for this operation.
4045 if (N->getOpcode() == ISD::SRA) {
4046 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
4047 if (BinOpRHSSignSet != HighBitSet)
4051 if (!TLI.isDesirableToCommuteWithShift(LHS))
4054 // Fold the constants, shifting the binop RHS by the shift amount.
4055 SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)),
4057 LHS->getOperand(1), N->getOperand(1));
4058 assert(isa<ConstantSDNode>(NewRHS) && "Folding was not successful!");
4060 // Create the new shift.
4061 SDValue NewShift = DAG.getNode(N->getOpcode(),
4062 SDLoc(LHS->getOperand(0)),
4063 VT, LHS->getOperand(0), N->getOperand(1));
4065 // Create the new binop.
4066 return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS);
4069 SDValue DAGCombiner::distributeTruncateThroughAnd(SDNode *N) {
4070 assert(N->getOpcode() == ISD::TRUNCATE);
4071 assert(N->getOperand(0).getOpcode() == ISD::AND);
4073 // (truncate:TruncVT (and N00, N01C)) -> (and (truncate:TruncVT N00), TruncC)
4074 if (N->hasOneUse() && N->getOperand(0).hasOneUse()) {
4075 SDValue N01 = N->getOperand(0).getOperand(1);
4077 if (ConstantSDNode *N01C = isConstOrConstSplat(N01)) {
4078 EVT TruncVT = N->getValueType(0);
4079 SDValue N00 = N->getOperand(0).getOperand(0);
4080 APInt TruncC = N01C->getAPIntValue();
4081 TruncC = TruncC.trunc(TruncVT.getScalarSizeInBits());
4083 return DAG.getNode(ISD::AND, SDLoc(N), TruncVT,
4084 DAG.getNode(ISD::TRUNCATE, SDLoc(N), TruncVT, N00),
4085 DAG.getConstant(TruncC, TruncVT));
4092 SDValue DAGCombiner::visitRotate(SDNode *N) {
4093 // fold (rot* x, (trunc (and y, c))) -> (rot* x, (and (trunc y), (trunc c))).
4094 if (N->getOperand(1).getOpcode() == ISD::TRUNCATE &&
4095 N->getOperand(1).getOperand(0).getOpcode() == ISD::AND) {
4096 SDValue NewOp1 = distributeTruncateThroughAnd(N->getOperand(1).getNode());
4097 if (NewOp1.getNode())
4098 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
4099 N->getOperand(0), NewOp1);
4104 SDValue DAGCombiner::visitSHL(SDNode *N) {
4105 SDValue N0 = N->getOperand(0);
4106 SDValue N1 = N->getOperand(1);
4107 EVT VT = N0.getValueType();
4108 unsigned OpSizeInBits = VT.getScalarSizeInBits();
4111 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4112 if (VT.isVector()) {
4113 SDValue FoldedVOp = SimplifyVBinOp(N);
4114 if (FoldedVOp.getNode()) return FoldedVOp;
4116 BuildVectorSDNode *N1CV = dyn_cast<BuildVectorSDNode>(N1);
4117 // If setcc produces all-one true value then:
4118 // (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<<N1CV)
4119 if (N1CV && N1CV->isConstant()) {
4120 if (N0.getOpcode() == ISD::AND) {
4121 SDValue N00 = N0->getOperand(0);
4122 SDValue N01 = N0->getOperand(1);
4123 BuildVectorSDNode *N01CV = dyn_cast<BuildVectorSDNode>(N01);
4125 if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC &&
4126 TLI.getBooleanContents(N00.getOperand(0).getValueType()) ==
4127 TargetLowering::ZeroOrNegativeOneBooleanContent) {
4128 if (SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, VT, N01CV, N1CV))
4129 return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C);
4132 N1C = isConstOrConstSplat(N1);
4137 // fold (shl c1, c2) -> c1<<c2
4138 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4140 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
4141 // fold (shl 0, x) -> 0
4142 if (N0C && N0C->isNullValue())
4144 // fold (shl x, c >= size(x)) -> undef
4145 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4146 return DAG.getUNDEF(VT);
4147 // fold (shl x, 0) -> x
4148 if (N1C && N1C->isNullValue())
4150 // fold (shl undef, x) -> 0
4151 if (N0.getOpcode() == ISD::UNDEF)
4152 return DAG.getConstant(0, VT);
4153 // if (shl x, c) is known to be zero, return 0
4154 if (DAG.MaskedValueIsZero(SDValue(N, 0),
4155 APInt::getAllOnesValue(OpSizeInBits)))
4156 return DAG.getConstant(0, VT);
4157 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
4158 if (N1.getOpcode() == ISD::TRUNCATE &&
4159 N1.getOperand(0).getOpcode() == ISD::AND) {
4160 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4161 if (NewOp1.getNode())
4162 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, NewOp1);
4165 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4166 return SDValue(N, 0);
4168 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
4169 if (N1C && N0.getOpcode() == ISD::SHL) {
4170 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4171 uint64_t c1 = N0C1->getZExtValue();
4172 uint64_t c2 = N1C->getZExtValue();
4173 if (c1 + c2 >= OpSizeInBits)
4174 return DAG.getConstant(0, VT);
4175 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4176 DAG.getConstant(c1 + c2, N1.getValueType()));
4180 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
4181 // For this to be valid, the second form must not preserve any of the bits
4182 // that are shifted out by the inner shift in the first form. This means
4183 // the outer shift size must be >= the number of bits added by the ext.
4184 // As a corollary, we don't care what kind of ext it is.
4185 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
4186 N0.getOpcode() == ISD::ANY_EXTEND ||
4187 N0.getOpcode() == ISD::SIGN_EXTEND) &&
4188 N0.getOperand(0).getOpcode() == ISD::SHL) {
4189 SDValue N0Op0 = N0.getOperand(0);
4190 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4191 uint64_t c1 = N0Op0C1->getZExtValue();
4192 uint64_t c2 = N1C->getZExtValue();
4193 EVT InnerShiftVT = N0Op0.getValueType();
4194 uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits();
4195 if (c2 >= OpSizeInBits - InnerShiftSize) {
4196 if (c1 + c2 >= OpSizeInBits)
4197 return DAG.getConstant(0, VT);
4198 return DAG.getNode(ISD::SHL, SDLoc(N0), VT,
4199 DAG.getNode(N0.getOpcode(), SDLoc(N0), VT,
4200 N0Op0->getOperand(0)),
4201 DAG.getConstant(c1 + c2, N1.getValueType()));
4206 // fold (shl (zext (srl x, C)), C) -> (zext (shl (srl x, C), C))
4207 // Only fold this if the inner zext has no other uses to avoid increasing
4208 // the total number of instructions.
4209 if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() &&
4210 N0.getOperand(0).getOpcode() == ISD::SRL) {
4211 SDValue N0Op0 = N0.getOperand(0);
4212 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4213 uint64_t c1 = N0Op0C1->getZExtValue();
4214 if (c1 < VT.getScalarSizeInBits()) {
4215 uint64_t c2 = N1C->getZExtValue();
4217 SDValue NewOp0 = N0.getOperand(0);
4218 EVT CountVT = NewOp0.getOperand(1).getValueType();
4219 SDValue NewSHL = DAG.getNode(ISD::SHL, SDLoc(N), NewOp0.getValueType(),
4220 NewOp0, DAG.getConstant(c2, CountVT));
4221 AddToWorklist(NewSHL.getNode());
4222 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
4228 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
4229 // (and (srl x, (sub c1, c2), MASK)
4230 // Only fold this if the inner shift has no other uses -- if it does, folding
4231 // this will increase the total number of instructions.
4232 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
4233 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4234 uint64_t c1 = N0C1->getZExtValue();
4235 if (c1 < OpSizeInBits) {
4236 uint64_t c2 = N1C->getZExtValue();
4237 APInt Mask = APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - c1);
4240 Mask = Mask.shl(c2 - c1);
4241 Shift = DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4242 DAG.getConstant(c2 - c1, N1.getValueType()));
4244 Mask = Mask.lshr(c1 - c2);
4245 Shift = DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4246 DAG.getConstant(c1 - c2, N1.getValueType()));
4248 return DAG.getNode(ISD::AND, SDLoc(N0), VT, Shift,
4249 DAG.getConstant(Mask, VT));
4253 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
4254 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
4255 unsigned BitSize = VT.getScalarSizeInBits();
4256 SDValue HiBitsMask =
4257 DAG.getConstant(APInt::getHighBitsSet(BitSize,
4258 BitSize - N1C->getZExtValue()), VT);
4259 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4263 // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
4264 // Variant of version done on multiply, except mul by a power of 2 is turned
4267 if (N1C && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
4268 (isa<ConstantSDNode>(N0.getOperand(1)) ||
4269 isConstantSplatVector(N0.getOperand(1).getNode(), Val))) {
4270 SDValue Shl0 = DAG.getNode(ISD::SHL, SDLoc(N0), VT, N0.getOperand(0), N1);
4271 SDValue Shl1 = DAG.getNode(ISD::SHL, SDLoc(N1), VT, N0.getOperand(1), N1);
4272 return DAG.getNode(ISD::ADD, SDLoc(N), VT, Shl0, Shl1);
4276 SDValue NewSHL = visitShiftByConstant(N, N1C);
4277 if (NewSHL.getNode())
4284 SDValue DAGCombiner::visitSRA(SDNode *N) {
4285 SDValue N0 = N->getOperand(0);
4286 SDValue N1 = N->getOperand(1);
4287 EVT VT = N0.getValueType();
4288 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4291 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4292 if (VT.isVector()) {
4293 SDValue FoldedVOp = SimplifyVBinOp(N);
4294 if (FoldedVOp.getNode()) return FoldedVOp;
4296 N1C = isConstOrConstSplat(N1);
4299 // fold (sra c1, c2) -> (sra c1, c2)
4300 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4302 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
4303 // fold (sra 0, x) -> 0
4304 if (N0C && N0C->isNullValue())
4306 // fold (sra -1, x) -> -1
4307 if (N0C && N0C->isAllOnesValue())
4309 // fold (sra x, (setge c, size(x))) -> undef
4310 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4311 return DAG.getUNDEF(VT);
4312 // fold (sra x, 0) -> x
4313 if (N1C && N1C->isNullValue())
4315 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
4317 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
4318 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
4319 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
4321 ExtVT = EVT::getVectorVT(*DAG.getContext(),
4322 ExtVT, VT.getVectorNumElements());
4323 if ((!LegalOperations ||
4324 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
4325 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
4326 N0.getOperand(0), DAG.getValueType(ExtVT));
4329 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
4330 if (N1C && N0.getOpcode() == ISD::SRA) {
4331 if (ConstantSDNode *C1 = isConstOrConstSplat(N0.getOperand(1))) {
4332 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
4333 if (Sum >= OpSizeInBits)
4334 Sum = OpSizeInBits - 1;
4335 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0.getOperand(0),
4336 DAG.getConstant(Sum, N1.getValueType()));
4340 // fold (sra (shl X, m), (sub result_size, n))
4341 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
4342 // result_size - n != m.
4343 // If truncate is free for the target sext(shl) is likely to result in better
4345 if (N0.getOpcode() == ISD::SHL && N1C) {
4346 // Get the two constanst of the shifts, CN0 = m, CN = n.
4347 const ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1));
4349 LLVMContext &Ctx = *DAG.getContext();
4350 // Determine what the truncate's result bitsize and type would be.
4351 EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - N1C->getZExtValue());
4354 TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorNumElements());
4356 // Determine the residual right-shift amount.
4357 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
4359 // If the shift is not a no-op (in which case this should be just a sign
4360 // extend already), the truncated to type is legal, sign_extend is legal
4361 // on that type, and the truncate to that type is both legal and free,
4362 // perform the transform.
4363 if ((ShiftAmt > 0) &&
4364 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
4365 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
4366 TLI.isTruncateFree(VT, TruncVT)) {
4368 SDValue Amt = DAG.getConstant(ShiftAmt,
4369 getShiftAmountTy(N0.getOperand(0).getValueType()));
4370 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0), VT,
4371 N0.getOperand(0), Amt);
4372 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), TruncVT,
4374 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N),
4375 N->getValueType(0), Trunc);
4380 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
4381 if (N1.getOpcode() == ISD::TRUNCATE &&
4382 N1.getOperand(0).getOpcode() == ISD::AND) {
4383 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4384 if (NewOp1.getNode())
4385 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, NewOp1);
4388 // fold (sra (trunc (srl x, c1)), c2) -> (trunc (sra x, c1 + c2))
4389 // if c1 is equal to the number of bits the trunc removes
4390 if (N0.getOpcode() == ISD::TRUNCATE &&
4391 (N0.getOperand(0).getOpcode() == ISD::SRL ||
4392 N0.getOperand(0).getOpcode() == ISD::SRA) &&
4393 N0.getOperand(0).hasOneUse() &&
4394 N0.getOperand(0).getOperand(1).hasOneUse() &&
4396 SDValue N0Op0 = N0.getOperand(0);
4397 if (ConstantSDNode *LargeShift = isConstOrConstSplat(N0Op0.getOperand(1))) {
4398 unsigned LargeShiftVal = LargeShift->getZExtValue();
4399 EVT LargeVT = N0Op0.getValueType();
4401 if (LargeVT.getScalarSizeInBits() - OpSizeInBits == LargeShiftVal) {
4403 DAG.getConstant(LargeShiftVal + N1C->getZExtValue(),
4404 getShiftAmountTy(N0Op0.getOperand(0).getValueType()));
4405 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), LargeVT,
4406 N0Op0.getOperand(0), Amt);
4407 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, SRA);
4412 // Simplify, based on bits shifted out of the LHS.
4413 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4414 return SDValue(N, 0);
4417 // If the sign bit is known to be zero, switch this to a SRL.
4418 if (DAG.SignBitIsZero(N0))
4419 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1);
4422 SDValue NewSRA = visitShiftByConstant(N, N1C);
4423 if (NewSRA.getNode())
4430 SDValue DAGCombiner::visitSRL(SDNode *N) {
4431 SDValue N0 = N->getOperand(0);
4432 SDValue N1 = N->getOperand(1);
4433 EVT VT = N0.getValueType();
4434 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4437 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4438 if (VT.isVector()) {
4439 SDValue FoldedVOp = SimplifyVBinOp(N);
4440 if (FoldedVOp.getNode()) return FoldedVOp;
4442 N1C = isConstOrConstSplat(N1);
4445 // fold (srl c1, c2) -> c1 >>u c2
4446 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4448 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
4449 // fold (srl 0, x) -> 0
4450 if (N0C && N0C->isNullValue())
4452 // fold (srl x, c >= size(x)) -> undef
4453 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4454 return DAG.getUNDEF(VT);
4455 // fold (srl x, 0) -> x
4456 if (N1C && N1C->isNullValue())
4458 // if (srl x, c) is known to be zero, return 0
4459 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
4460 APInt::getAllOnesValue(OpSizeInBits)))
4461 return DAG.getConstant(0, VT);
4463 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
4464 if (N1C && N0.getOpcode() == ISD::SRL) {
4465 if (ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1))) {
4466 uint64_t c1 = N01C->getZExtValue();
4467 uint64_t c2 = N1C->getZExtValue();
4468 if (c1 + c2 >= OpSizeInBits)
4469 return DAG.getConstant(0, VT);
4470 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4471 DAG.getConstant(c1 + c2, N1.getValueType()));
4475 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
4476 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
4477 N0.getOperand(0).getOpcode() == ISD::SRL &&
4478 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
4480 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
4481 uint64_t c2 = N1C->getZExtValue();
4482 EVT InnerShiftVT = N0.getOperand(0).getValueType();
4483 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
4484 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
4485 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
4486 if (c1 + OpSizeInBits == InnerShiftSize) {
4487 if (c1 + c2 >= InnerShiftSize)
4488 return DAG.getConstant(0, VT);
4489 return DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT,
4490 DAG.getNode(ISD::SRL, SDLoc(N0), InnerShiftVT,
4491 N0.getOperand(0)->getOperand(0),
4492 DAG.getConstant(c1 + c2, ShiftCountVT)));
4496 // fold (srl (shl x, c), c) -> (and x, cst2)
4497 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1) {
4498 unsigned BitSize = N0.getScalarValueSizeInBits();
4499 if (BitSize <= 64) {
4500 uint64_t ShAmt = N1C->getZExtValue() + 64 - BitSize;
4501 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4502 DAG.getConstant(~0ULL >> ShAmt, VT));
4506 // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask)
4507 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
4508 // Shifting in all undef bits?
4509 EVT SmallVT = N0.getOperand(0).getValueType();
4510 unsigned BitSize = SmallVT.getScalarSizeInBits();
4511 if (N1C->getZExtValue() >= BitSize)
4512 return DAG.getUNDEF(VT);
4514 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
4515 uint64_t ShiftAmt = N1C->getZExtValue();
4516 SDValue SmallShift = DAG.getNode(ISD::SRL, SDLoc(N0), SmallVT,
4518 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
4519 AddToWorklist(SmallShift.getNode());
4520 APInt Mask = APInt::getAllOnesValue(OpSizeInBits).lshr(ShiftAmt);
4521 return DAG.getNode(ISD::AND, SDLoc(N), VT,
4522 DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, SmallShift),
4523 DAG.getConstant(Mask, VT));
4527 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
4528 // bit, which is unmodified by sra.
4529 if (N1C && N1C->getZExtValue() + 1 == OpSizeInBits) {
4530 if (N0.getOpcode() == ISD::SRA)
4531 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1);
4534 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
4535 if (N1C && N0.getOpcode() == ISD::CTLZ &&
4536 N1C->getAPIntValue() == Log2_32(OpSizeInBits)) {
4537 APInt KnownZero, KnownOne;
4538 DAG.computeKnownBits(N0.getOperand(0), KnownZero, KnownOne);
4540 // If any of the input bits are KnownOne, then the input couldn't be all
4541 // zeros, thus the result of the srl will always be zero.
4542 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
4544 // If all of the bits input the to ctlz node are known to be zero, then
4545 // the result of the ctlz is "32" and the result of the shift is one.
4546 APInt UnknownBits = ~KnownZero;
4547 if (UnknownBits == 0) return DAG.getConstant(1, VT);
4549 // Otherwise, check to see if there is exactly one bit input to the ctlz.
4550 if ((UnknownBits & (UnknownBits - 1)) == 0) {
4551 // Okay, we know that only that the single bit specified by UnknownBits
4552 // could be set on input to the CTLZ node. If this bit is set, the SRL
4553 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
4554 // to an SRL/XOR pair, which is likely to simplify more.
4555 unsigned ShAmt = UnknownBits.countTrailingZeros();
4556 SDValue Op = N0.getOperand(0);
4559 Op = DAG.getNode(ISD::SRL, SDLoc(N0), VT, Op,
4560 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
4561 AddToWorklist(Op.getNode());
4564 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
4565 Op, DAG.getConstant(1, VT));
4569 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
4570 if (N1.getOpcode() == ISD::TRUNCATE &&
4571 N1.getOperand(0).getOpcode() == ISD::AND) {
4572 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4573 if (NewOp1.getNode())
4574 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, NewOp1);
4577 // fold operands of srl based on knowledge that the low bits are not
4579 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4580 return SDValue(N, 0);
4583 SDValue NewSRL = visitShiftByConstant(N, N1C);
4584 if (NewSRL.getNode())
4588 // Attempt to convert a srl of a load into a narrower zero-extending load.
4589 SDValue NarrowLoad = ReduceLoadWidth(N);
4590 if (NarrowLoad.getNode())
4593 // Here is a common situation. We want to optimize:
4596 // %b = and i32 %a, 2
4597 // %c = srl i32 %b, 1
4598 // brcond i32 %c ...
4604 // %c = setcc eq %b, 0
4607 // However when after the source operand of SRL is optimized into AND, the SRL
4608 // itself may not be optimized further. Look for it and add the BRCOND into
4610 if (N->hasOneUse()) {
4611 SDNode *Use = *N->use_begin();
4612 if (Use->getOpcode() == ISD::BRCOND)
4614 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
4615 // Also look pass the truncate.
4616 Use = *Use->use_begin();
4617 if (Use->getOpcode() == ISD::BRCOND)
4625 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
4626 SDValue N0 = N->getOperand(0);
4627 EVT VT = N->getValueType(0);
4629 // fold (ctlz c1) -> c2
4630 if (isa<ConstantSDNode>(N0))
4631 return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0);
4635 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4636 SDValue N0 = N->getOperand(0);
4637 EVT VT = N->getValueType(0);
4639 // fold (ctlz_zero_undef c1) -> c2
4640 if (isa<ConstantSDNode>(N0))
4641 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4645 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4646 SDValue N0 = N->getOperand(0);
4647 EVT VT = N->getValueType(0);
4649 // fold (cttz c1) -> c2
4650 if (isa<ConstantSDNode>(N0))
4651 return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0);
4655 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4656 SDValue N0 = N->getOperand(0);
4657 EVT VT = N->getValueType(0);
4659 // fold (cttz_zero_undef c1) -> c2
4660 if (isa<ConstantSDNode>(N0))
4661 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4665 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4666 SDValue N0 = N->getOperand(0);
4667 EVT VT = N->getValueType(0);
4669 // fold (ctpop c1) -> c2
4670 if (isa<ConstantSDNode>(N0))
4671 return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0);
4676 /// \brief Generate Min/Max node
4677 static SDValue combineMinNumMaxNum(SDLoc DL, EVT VT, SDValue LHS, SDValue RHS,
4678 SDValue True, SDValue False,
4679 ISD::CondCode CC, const TargetLowering &TLI,
4680 SelectionDAG &DAG) {
4681 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
4691 unsigned Opcode = (LHS == True) ? ISD::FMINNUM : ISD::FMAXNUM;
4692 if (TLI.isOperationLegal(Opcode, VT))
4693 return DAG.getNode(Opcode, DL, VT, LHS, RHS);
4702 unsigned Opcode = (LHS == True) ? ISD::FMAXNUM : ISD::FMINNUM;
4703 if (TLI.isOperationLegal(Opcode, VT))
4704 return DAG.getNode(Opcode, DL, VT, LHS, RHS);
4712 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4713 SDValue N0 = N->getOperand(0);
4714 SDValue N1 = N->getOperand(1);
4715 SDValue N2 = N->getOperand(2);
4716 EVT VT = N->getValueType(0);
4717 EVT VT0 = N0.getValueType();
4719 // fold (select C, X, X) -> X
4722 // fold (select true, X, Y) -> X
4723 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4724 if (N0C && !N0C->isNullValue())
4726 // fold (select false, X, Y) -> Y
4727 if (N0C && N0C->isNullValue())
4729 // fold (select C, 1, X) -> (or C, X)
4730 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4731 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4732 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4733 // fold (select C, 0, 1) -> (xor C, 1)
4734 // We can't do this reliably if integer based booleans have different contents
4735 // to floating point based booleans. This is because we can't tell whether we
4736 // have an integer-based boolean or a floating-point-based boolean unless we
4737 // can find the SETCC that produced it and inspect its operands. This is
4738 // fairly easy if C is the SETCC node, but it can potentially be
4739 // undiscoverable (or not reasonably discoverable). For example, it could be
4740 // in another basic block or it could require searching a complicated
4742 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
4743 if (VT.isInteger() &&
4744 (VT0 == MVT::i1 || (VT0.isInteger() &&
4745 TLI.getBooleanContents(false, false) ==
4746 TLI.getBooleanContents(false, true) &&
4747 TLI.getBooleanContents(false, false) ==
4748 TargetLowering::ZeroOrOneBooleanContent)) &&
4749 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4752 return DAG.getNode(ISD::XOR, SDLoc(N), VT0,
4753 N0, DAG.getConstant(1, VT0));
4754 XORNode = DAG.getNode(ISD::XOR, SDLoc(N0), VT0,
4755 N0, DAG.getConstant(1, VT0));
4756 AddToWorklist(XORNode.getNode());
4758 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, XORNode);
4759 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, XORNode);
4761 // fold (select C, 0, X) -> (and (not C), X)
4762 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4763 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4764 AddToWorklist(NOTNode.getNode());
4765 return DAG.getNode(ISD::AND, SDLoc(N), VT, NOTNode, N2);
4767 // fold (select C, X, 1) -> (or (not C), X)
4768 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4769 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4770 AddToWorklist(NOTNode.getNode());
4771 return DAG.getNode(ISD::OR, SDLoc(N), VT, NOTNode, N1);
4773 // fold (select C, X, 0) -> (and C, X)
4774 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4775 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4776 // fold (select X, X, Y) -> (or X, Y)
4777 // fold (select X, 1, Y) -> (or X, Y)
4778 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4779 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4780 // fold (select X, Y, X) -> (and X, Y)
4781 // fold (select X, Y, 0) -> (and X, Y)
4782 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4783 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4785 // If we can fold this based on the true/false value, do so.
4786 if (SimplifySelectOps(N, N1, N2))
4787 return SDValue(N, 0); // Don't revisit N.
4789 // fold selects based on a setcc into other things, such as min/max/abs
4790 if (N0.getOpcode() == ISD::SETCC) {
4791 // select x, y (fcmp lt x, y) -> fminnum x, y
4792 // select x, y (fcmp gt x, y) -> fmaxnum x, y
4794 // This is OK if we don't care about what happens if either operand is a
4798 // FIXME: Instead of testing for UnsafeFPMath, this should be checking for
4799 // no signed zeros as well as no nans.
4800 const TargetOptions &Options = DAG.getTarget().Options;
4801 if (Options.UnsafeFPMath &&
4802 VT.isFloatingPoint() && N0.hasOneUse() &&
4803 DAG.isKnownNeverNaN(N1) && DAG.isKnownNeverNaN(N2)) {
4804 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
4807 combineMinNumMaxNum(SDLoc(N), VT, N0.getOperand(0), N0.getOperand(1),
4808 N1, N2, CC, TLI, DAG);
4813 if ((!LegalOperations &&
4814 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT)) ||
4815 TLI.isOperationLegal(ISD::SELECT_CC, VT))
4816 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT,
4817 N0.getOperand(0), N0.getOperand(1),
4818 N1, N2, N0.getOperand(2));
4819 return SimplifySelect(SDLoc(N), N0, N1, N2);
4826 std::pair<SDValue, SDValue> SplitVSETCC(const SDNode *N, SelectionDAG &DAG) {
4829 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
4831 // Split the inputs.
4832 SDValue Lo, Hi, LL, LH, RL, RH;
4833 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
4834 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
4836 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));
4837 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));
4839 return std::make_pair(Lo, Hi);
4842 // This function assumes all the vselect's arguments are CONCAT_VECTOR
4843 // nodes and that the condition is a BV of ConstantSDNodes (or undefs).
4844 static SDValue ConvertSelectToConcatVector(SDNode *N, SelectionDAG &DAG) {
4846 SDValue Cond = N->getOperand(0);
4847 SDValue LHS = N->getOperand(1);
4848 SDValue RHS = N->getOperand(2);
4849 EVT VT = N->getValueType(0);
4850 int NumElems = VT.getVectorNumElements();
4851 assert(LHS.getOpcode() == ISD::CONCAT_VECTORS &&
4852 RHS.getOpcode() == ISD::CONCAT_VECTORS &&
4853 Cond.getOpcode() == ISD::BUILD_VECTOR);
4855 // CONCAT_VECTOR can take an arbitrary number of arguments. We only care about
4856 // binary ones here.
4857 if (LHS->getNumOperands() != 2 || RHS->getNumOperands() != 2)
4860 // We're sure we have an even number of elements due to the
4861 // concat_vectors we have as arguments to vselect.
4862 // Skip BV elements until we find one that's not an UNDEF
4863 // After we find an UNDEF element, keep looping until we get to half the
4864 // length of the BV and see if all the non-undef nodes are the same.
4865 ConstantSDNode *BottomHalf = nullptr;
4866 for (int i = 0; i < NumElems / 2; ++i) {
4867 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4870 if (BottomHalf == nullptr)
4871 BottomHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4872 else if (Cond->getOperand(i).getNode() != BottomHalf)
4876 // Do the same for the second half of the BuildVector
4877 ConstantSDNode *TopHalf = nullptr;
4878 for (int i = NumElems / 2; i < NumElems; ++i) {
4879 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4882 if (TopHalf == nullptr)
4883 TopHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4884 else if (Cond->getOperand(i).getNode() != TopHalf)
4888 assert(TopHalf && BottomHalf &&
4889 "One half of the selector was all UNDEFs and the other was all the "
4890 "same value. This should have been addressed before this function.");
4892 ISD::CONCAT_VECTORS, dl, VT,
4893 BottomHalf->isNullValue() ? RHS->getOperand(0) : LHS->getOperand(0),
4894 TopHalf->isNullValue() ? RHS->getOperand(1) : LHS->getOperand(1));
4897 SDValue DAGCombiner::visitMSTORE(SDNode *N) {
4899 if (Level >= AfterLegalizeTypes)
4902 MaskedStoreSDNode *MST = dyn_cast<MaskedStoreSDNode>(N);
4903 SDValue Mask = MST->getMask();
4904 SDValue Data = MST->getValue();
4907 // If the MSTORE data type requires splitting and the mask is provided by a
4908 // SETCC, then split both nodes and its operands before legalization. This
4909 // prevents the type legalizer from unrolling SETCC into scalar comparisons
4910 // and enables future optimizations (e.g. min/max pattern matching on X86).
4911 if (Mask.getOpcode() == ISD::SETCC) {
4913 // Check if any splitting is required.
4914 if (TLI.getTypeAction(*DAG.getContext(), Data.getValueType()) !=
4915 TargetLowering::TypeSplitVector)
4918 SDValue MaskLo, MaskHi, Lo, Hi;
4919 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
4922 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MST->getValueType(0));
4924 SDValue Chain = MST->getChain();
4925 SDValue Ptr = MST->getBasePtr();
4927 EVT MemoryVT = MST->getMemoryVT();
4928 unsigned Alignment = MST->getOriginalAlignment();
4930 // if Alignment is equal to the vector size,
4931 // take the half of it for the second part
4932 unsigned SecondHalfAlignment =
4933 (Alignment == Data->getValueType(0).getSizeInBits()/8) ?
4934 Alignment/2 : Alignment;
4936 EVT LoMemVT, HiMemVT;
4937 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
4939 SDValue DataLo, DataHi;
4940 std::tie(DataLo, DataHi) = DAG.SplitVector(Data, DL);
4942 MachineMemOperand *MMO = DAG.getMachineFunction().
4943 getMachineMemOperand(MST->getPointerInfo(),
4944 MachineMemOperand::MOStore, LoMemVT.getStoreSize(),
4945 Alignment, MST->getAAInfo(), MST->getRanges());
4947 Lo = DAG.getMaskedStore(Chain, DL, DataLo, Ptr, MaskLo, LoMemVT, MMO,
4948 MST->isTruncatingStore());
4950 unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
4951 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
4952 DAG.getConstant(IncrementSize, Ptr.getValueType()));
4954 MMO = DAG.getMachineFunction().
4955 getMachineMemOperand(MST->getPointerInfo(),
4956 MachineMemOperand::MOStore, HiMemVT.getStoreSize(),
4957 SecondHalfAlignment, MST->getAAInfo(),
4960 Hi = DAG.getMaskedStore(Chain, DL, DataHi, Ptr, MaskHi, HiMemVT, MMO,
4961 MST->isTruncatingStore());
4963 AddToWorklist(Lo.getNode());
4964 AddToWorklist(Hi.getNode());
4966 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);
4971 SDValue DAGCombiner::visitMLOAD(SDNode *N) {
4973 if (Level >= AfterLegalizeTypes)
4976 MaskedLoadSDNode *MLD = dyn_cast<MaskedLoadSDNode>(N);
4977 SDValue Mask = MLD->getMask();
4980 // If the MLOAD result requires splitting and the mask is provided by a
4981 // SETCC, then split both nodes and its operands before legalization. This
4982 // prevents the type legalizer from unrolling SETCC into scalar comparisons
4983 // and enables future optimizations (e.g. min/max pattern matching on X86).
4985 if (Mask.getOpcode() == ISD::SETCC) {
4986 EVT VT = N->getValueType(0);
4988 // Check if any splitting is required.
4989 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
4990 TargetLowering::TypeSplitVector)
4993 SDValue MaskLo, MaskHi, Lo, Hi;
4994 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
4996 SDValue Src0 = MLD->getSrc0();
4997 SDValue Src0Lo, Src0Hi;
4998 std::tie(Src0Lo, Src0Hi) = DAG.SplitVector(Src0, DL);
5001 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MLD->getValueType(0));
5003 SDValue Chain = MLD->getChain();
5004 SDValue Ptr = MLD->getBasePtr();
5005 EVT MemoryVT = MLD->getMemoryVT();
5006 unsigned Alignment = MLD->getOriginalAlignment();
5008 // if Alignment is equal to the vector size,
5009 // take the half of it for the second part
5010 unsigned SecondHalfAlignment =
5011 (Alignment == MLD->getValueType(0).getSizeInBits()/8) ?
5012 Alignment/2 : Alignment;
5014 EVT LoMemVT, HiMemVT;
5015 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
5017 MachineMemOperand *MMO = DAG.getMachineFunction().
5018 getMachineMemOperand(MLD->getPointerInfo(),
5019 MachineMemOperand::MOLoad, LoMemVT.getStoreSize(),
5020 Alignment, MLD->getAAInfo(), MLD->getRanges());
5022 Lo = DAG.getMaskedLoad(LoVT, DL, Chain, Ptr, MaskLo, Src0Lo, LoMemVT, MMO,
5025 unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
5026 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5027 DAG.getConstant(IncrementSize, Ptr.getValueType()));
5029 MMO = DAG.getMachineFunction().
5030 getMachineMemOperand(MLD->getPointerInfo(),
5031 MachineMemOperand::MOLoad, HiMemVT.getStoreSize(),
5032 SecondHalfAlignment, MLD->getAAInfo(), MLD->getRanges());
5034 Hi = DAG.getMaskedLoad(HiVT, DL, Chain, Ptr, MaskHi, Src0Hi, HiMemVT, MMO,
5037 AddToWorklist(Lo.getNode());
5038 AddToWorklist(Hi.getNode());
5040 // Build a factor node to remember that this load is independent of the
5042 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo.getValue(1),
5045 // Legalized the chain result - switch anything that used the old chain to
5047 DAG.ReplaceAllUsesOfValueWith(SDValue(MLD, 1), Chain);
5049 SDValue LoadRes = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
5051 SDValue RetOps[] = { LoadRes, Chain };
5052 return DAG.getMergeValues(RetOps, DL);
5057 SDValue DAGCombiner::visitVSELECT(SDNode *N) {
5058 SDValue N0 = N->getOperand(0);
5059 SDValue N1 = N->getOperand(1);
5060 SDValue N2 = N->getOperand(2);
5063 // Canonicalize integer abs.
5064 // vselect (setg[te] X, 0), X, -X ->
5065 // vselect (setgt X, -1), X, -X ->
5066 // vselect (setl[te] X, 0), -X, X ->
5067 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
5068 if (N0.getOpcode() == ISD::SETCC) {
5069 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
5070 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5072 bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
5074 if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
5075 (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) &&
5076 N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1))
5077 isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode());
5078 else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) &&
5079 N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1))
5080 isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode());
5083 EVT VT = LHS.getValueType();
5084 SDValue Shift = DAG.getNode(
5085 ISD::SRA, DL, VT, LHS,
5086 DAG.getConstant(VT.getScalarType().getSizeInBits() - 1, VT));
5087 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift);
5088 AddToWorklist(Shift.getNode());
5089 AddToWorklist(Add.getNode());
5090 return DAG.getNode(ISD::XOR, DL, VT, Add, Shift);
5094 // If the VSELECT result requires splitting and the mask is provided by a
5095 // SETCC, then split both nodes and its operands before legalization. This
5096 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5097 // and enables future optimizations (e.g. min/max pattern matching on X86).
5098 if (N0.getOpcode() == ISD::SETCC) {
5099 EVT VT = N->getValueType(0);
5101 // Check if any splitting is required.
5102 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
5103 TargetLowering::TypeSplitVector)
5106 SDValue Lo, Hi, CCLo, CCHi, LL, LH, RL, RH;
5107 std::tie(CCLo, CCHi) = SplitVSETCC(N0.getNode(), DAG);
5108 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 1);
5109 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 2);
5111 Lo = DAG.getNode(N->getOpcode(), DL, LL.getValueType(), CCLo, LL, RL);
5112 Hi = DAG.getNode(N->getOpcode(), DL, LH.getValueType(), CCHi, LH, RH);
5114 // Add the new VSELECT nodes to the work list in case they need to be split
5116 AddToWorklist(Lo.getNode());
5117 AddToWorklist(Hi.getNode());
5119 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
5122 // Fold (vselect (build_vector all_ones), N1, N2) -> N1
5123 if (ISD::isBuildVectorAllOnes(N0.getNode()))
5125 // Fold (vselect (build_vector all_zeros), N1, N2) -> N2
5126 if (ISD::isBuildVectorAllZeros(N0.getNode()))
5129 // The ConvertSelectToConcatVector function is assuming both the above
5130 // checks for (vselect (build_vector all{ones,zeros) ...) have been made
5132 if (N1.getOpcode() == ISD::CONCAT_VECTORS &&
5133 N2.getOpcode() == ISD::CONCAT_VECTORS &&
5134 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
5135 SDValue CV = ConvertSelectToConcatVector(N, DAG);
5143 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
5144 SDValue N0 = N->getOperand(0);
5145 SDValue N1 = N->getOperand(1);
5146 SDValue N2 = N->getOperand(2);
5147 SDValue N3 = N->getOperand(3);
5148 SDValue N4 = N->getOperand(4);
5149 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
5151 // fold select_cc lhs, rhs, x, x, cc -> x
5155 // Determine if the condition we're dealing with is constant
5156 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
5157 N0, N1, CC, SDLoc(N), false);
5158 if (SCC.getNode()) {
5159 AddToWorklist(SCC.getNode());
5161 if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) {
5162 if (!SCCC->isNullValue())
5163 return N2; // cond always true -> true val
5165 return N3; // cond always false -> false val
5166 } else if (SCC->getOpcode() == ISD::UNDEF) {
5167 // When the condition is UNDEF, just return the first operand. This is
5168 // coherent the DAG creation, no setcc node is created in this case
5170 } else if (SCC.getOpcode() == ISD::SETCC) {
5171 // Fold to a simpler select_cc
5172 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(),
5173 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
5178 // If we can fold this based on the true/false value, do so.
5179 if (SimplifySelectOps(N, N2, N3))
5180 return SDValue(N, 0); // Don't revisit N.
5182 // fold select_cc into other things, such as min/max/abs
5183 return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC);
5186 SDValue DAGCombiner::visitSETCC(SDNode *N) {
5187 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
5188 cast<CondCodeSDNode>(N->getOperand(2))->get(),
5192 // tryToFoldExtendOfConstant - Try to fold a sext/zext/aext
5193 // dag node into a ConstantSDNode or a build_vector of constants.
5194 // This function is called by the DAGCombiner when visiting sext/zext/aext
5195 // dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND).
5196 // Vector extends are not folded if operations are legal; this is to
5197 // avoid introducing illegal build_vector dag nodes.
5198 static SDNode *tryToFoldExtendOfConstant(SDNode *N, const TargetLowering &TLI,
5199 SelectionDAG &DAG, bool LegalTypes,
5200 bool LegalOperations) {
5201 unsigned Opcode = N->getOpcode();
5202 SDValue N0 = N->getOperand(0);
5203 EVT VT = N->getValueType(0);
5205 assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND ||
5206 Opcode == ISD::ANY_EXTEND) && "Expected EXTEND dag node in input!");
5208 // fold (sext c1) -> c1
5209 // fold (zext c1) -> c1
5210 // fold (aext c1) -> c1
5211 if (isa<ConstantSDNode>(N0))
5212 return DAG.getNode(Opcode, SDLoc(N), VT, N0).getNode();
5214 // fold (sext (build_vector AllConstants) -> (build_vector AllConstants)
5215 // fold (zext (build_vector AllConstants) -> (build_vector AllConstants)
5216 // fold (aext (build_vector AllConstants) -> (build_vector AllConstants)
5217 EVT SVT = VT.getScalarType();
5218 if (!(VT.isVector() &&
5219 (!LegalTypes || (!LegalOperations && TLI.isTypeLegal(SVT))) &&
5220 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())))
5223 // We can fold this node into a build_vector.
5224 unsigned VTBits = SVT.getSizeInBits();
5225 unsigned EVTBits = N0->getValueType(0).getScalarType().getSizeInBits();
5226 unsigned ShAmt = VTBits - EVTBits;
5227 SmallVector<SDValue, 8> Elts;
5228 unsigned NumElts = N0->getNumOperands();
5231 for (unsigned i=0; i != NumElts; ++i) {
5232 SDValue Op = N0->getOperand(i);
5233 if (Op->getOpcode() == ISD::UNDEF) {
5234 Elts.push_back(DAG.getUNDEF(SVT));
5238 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
5239 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
5240 if (Opcode == ISD::SIGN_EXTEND)
5241 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
5244 Elts.push_back(DAG.getConstant(C.shl(ShAmt).lshr(ShAmt).getZExtValue(),
5248 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Elts).getNode();
5251 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
5252 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
5253 // transformation. Returns true if extension are possible and the above
5254 // mentioned transformation is profitable.
5255 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
5257 SmallVectorImpl<SDNode *> &ExtendNodes,
5258 const TargetLowering &TLI) {
5259 bool HasCopyToRegUses = false;
5260 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
5261 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
5262 UE = N0.getNode()->use_end();
5267 if (UI.getUse().getResNo() != N0.getResNo())
5269 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
5270 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
5271 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
5272 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
5273 // Sign bits will be lost after a zext.
5276 for (unsigned i = 0; i != 2; ++i) {
5277 SDValue UseOp = User->getOperand(i);
5280 if (!isa<ConstantSDNode>(UseOp))
5285 ExtendNodes.push_back(User);
5288 // If truncates aren't free and there are users we can't
5289 // extend, it isn't worthwhile.
5292 // Remember if this value is live-out.
5293 if (User->getOpcode() == ISD::CopyToReg)
5294 HasCopyToRegUses = true;
5297 if (HasCopyToRegUses) {
5298 bool BothLiveOut = false;
5299 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
5301 SDUse &Use = UI.getUse();
5302 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
5308 // Both unextended and extended values are live out. There had better be
5309 // a good reason for the transformation.
5310 return ExtendNodes.size();
5315 void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
5316 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
5317 ISD::NodeType ExtType) {
5318 // Extend SetCC uses if necessary.
5319 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
5320 SDNode *SetCC = SetCCs[i];
5321 SmallVector<SDValue, 4> Ops;
5323 for (unsigned j = 0; j != 2; ++j) {
5324 SDValue SOp = SetCC->getOperand(j);
5326 Ops.push_back(ExtLoad);
5328 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
5331 Ops.push_back(SetCC->getOperand(2));
5332 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), Ops));
5336 // FIXME: Bring more similar combines here, common to sext/zext (maybe aext?).
5337 SDValue DAGCombiner::CombineExtLoad(SDNode *N) {
5338 SDValue N0 = N->getOperand(0);
5339 EVT DstVT = N->getValueType(0);
5340 EVT SrcVT = N0.getValueType();
5342 assert((N->getOpcode() == ISD::SIGN_EXTEND ||
5343 N->getOpcode() == ISD::ZERO_EXTEND) &&
5344 "Unexpected node type (not an extend)!");
5346 // fold (sext (load x)) to multiple smaller sextloads; same for zext.
5347 // For example, on a target with legal v4i32, but illegal v8i32, turn:
5348 // (v8i32 (sext (v8i16 (load x))))
5350 // (v8i32 (concat_vectors (v4i32 (sextload x)),
5351 // (v4i32 (sextload (x + 16)))))
5352 // Where uses of the original load, i.e.:
5354 // are replaced with:
5356 // (v8i32 (concat_vectors (v4i32 (sextload x)),
5357 // (v4i32 (sextload (x + 16)))))))
5359 // This combine is only applicable to illegal, but splittable, vectors.
5360 // All legal types, and illegal non-vector types, are handled elsewhere.
5361 // This combine is controlled by TargetLowering::isVectorLoadExtDesirable.
5363 if (N0->getOpcode() != ISD::LOAD)
5366 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5368 if (!ISD::isNON_EXTLoad(LN0) || !ISD::isUNINDEXEDLoad(LN0) ||
5369 !N0.hasOneUse() || LN0->isVolatile() || !DstVT.isVector() ||
5370 !DstVT.isPow2VectorType() || !TLI.isVectorLoadExtDesirable(SDValue(N, 0)))
5373 SmallVector<SDNode *, 4> SetCCs;
5374 if (!ExtendUsesToFormExtLoad(N, N0, N->getOpcode(), SetCCs, TLI))
5377 ISD::LoadExtType ExtType =
5378 N->getOpcode() == ISD::SIGN_EXTEND ? ISD::SEXTLOAD : ISD::ZEXTLOAD;
5380 // Try to split the vector types to get down to legal types.
5381 EVT SplitSrcVT = SrcVT;
5382 EVT SplitDstVT = DstVT;
5383 while (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT) &&
5384 SplitSrcVT.getVectorNumElements() > 1) {
5385 SplitDstVT = DAG.GetSplitDestVTs(SplitDstVT).first;
5386 SplitSrcVT = DAG.GetSplitDestVTs(SplitSrcVT).first;
5389 if (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT))
5393 const unsigned NumSplits =
5394 DstVT.getVectorNumElements() / SplitDstVT.getVectorNumElements();
5395 const unsigned Stride = SplitSrcVT.getStoreSize();
5396 SmallVector<SDValue, 4> Loads;
5397 SmallVector<SDValue, 4> Chains;
5399 SDValue BasePtr = LN0->getBasePtr();
5400 for (unsigned Idx = 0; Idx < NumSplits; Idx++) {
5401 const unsigned Offset = Idx * Stride;
5402 const unsigned Align = MinAlign(LN0->getAlignment(), Offset);
5404 SDValue SplitLoad = DAG.getExtLoad(
5405 ExtType, DL, SplitDstVT, LN0->getChain(), BasePtr,
5406 LN0->getPointerInfo().getWithOffset(Offset), SplitSrcVT,
5407 LN0->isVolatile(), LN0->isNonTemporal(), LN0->isInvariant(),
5408 Align, LN0->getAAInfo());
5410 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
5411 DAG.getConstant(Stride, BasePtr.getValueType()));
5413 Loads.push_back(SplitLoad.getValue(0));
5414 Chains.push_back(SplitLoad.getValue(1));
5417 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
5418 SDValue NewValue = DAG.getNode(ISD::CONCAT_VECTORS, DL, DstVT, Loads);
5420 CombineTo(N, NewValue);
5422 // Replace uses of the original load (before extension)
5423 // with a truncate of the concatenated sextloaded vectors.
5425 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(), NewValue);
5426 CombineTo(N0.getNode(), Trunc, NewChain);
5427 ExtendSetCCUses(SetCCs, Trunc, NewValue, DL,
5428 (ISD::NodeType)N->getOpcode());
5429 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5432 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
5433 SDValue N0 = N->getOperand(0);
5434 EVT VT = N->getValueType(0);
5436 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5438 return SDValue(Res, 0);
5440 // fold (sext (sext x)) -> (sext x)
5441 // fold (sext (aext x)) -> (sext x)
5442 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5443 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT,
5446 if (N0.getOpcode() == ISD::TRUNCATE) {
5447 // fold (sext (truncate (load x))) -> (sext (smaller load x))
5448 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
5449 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5450 if (NarrowLoad.getNode()) {
5451 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5452 if (NarrowLoad.getNode() != N0.getNode()) {
5453 CombineTo(N0.getNode(), NarrowLoad);
5454 // CombineTo deleted the truncate, if needed, but not what's under it.
5457 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5460 // See if the value being truncated is already sign extended. If so, just
5461 // eliminate the trunc/sext pair.
5462 SDValue Op = N0.getOperand(0);
5463 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
5464 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
5465 unsigned DestBits = VT.getScalarType().getSizeInBits();
5466 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
5468 if (OpBits == DestBits) {
5469 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
5470 // bits, it is already ready.
5471 if (NumSignBits > DestBits-MidBits)
5473 } else if (OpBits < DestBits) {
5474 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
5475 // bits, just sext from i32.
5476 if (NumSignBits > OpBits-MidBits)
5477 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, Op);
5479 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
5480 // bits, just truncate to i32.
5481 if (NumSignBits > OpBits-MidBits)
5482 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5485 // fold (sext (truncate x)) -> (sextinreg x).
5486 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
5487 N0.getValueType())) {
5488 if (OpBits < DestBits)
5489 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op);
5490 else if (OpBits > DestBits)
5491 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op);
5492 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, Op,
5493 DAG.getValueType(N0.getValueType()));
5497 // fold (sext (load x)) -> (sext (truncate (sextload x)))
5498 // Only generate vector extloads when 1) they're legal, and 2) they are
5499 // deemed desirable by the target.
5500 if (ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5501 ((!LegalOperations && !VT.isVector() &&
5502 !cast<LoadSDNode>(N0)->isVolatile()) ||
5503 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, N0.getValueType()))) {
5504 bool DoXform = true;
5505 SmallVector<SDNode*, 4> SetCCs;
5506 if (!N0.hasOneUse())
5507 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
5509 DoXform &= TLI.isVectorLoadExtDesirable(SDValue(N, 0));
5511 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5512 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5514 LN0->getBasePtr(), N0.getValueType(),
5515 LN0->getMemOperand());
5516 CombineTo(N, ExtLoad);
5517 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5518 N0.getValueType(), ExtLoad);
5519 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5520 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5522 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5526 // fold (sext (load x)) to multiple smaller sextloads.
5527 // Only on illegal but splittable vectors.
5528 if (SDValue ExtLoad = CombineExtLoad(N))
5531 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
5532 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
5533 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5534 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5535 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5536 EVT MemVT = LN0->getMemoryVT();
5537 if ((!LegalOperations && !LN0->isVolatile()) ||
5538 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, MemVT)) {
5539 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5541 LN0->getBasePtr(), MemVT,
5542 LN0->getMemOperand());
5543 CombineTo(N, ExtLoad);
5544 CombineTo(N0.getNode(),
5545 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5546 N0.getValueType(), ExtLoad),
5547 ExtLoad.getValue(1));
5548 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5552 // fold (sext (and/or/xor (load x), cst)) ->
5553 // (and/or/xor (sextload x), (sext cst))
5554 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5555 N0.getOpcode() == ISD::XOR) &&
5556 isa<LoadSDNode>(N0.getOperand(0)) &&
5557 N0.getOperand(1).getOpcode() == ISD::Constant &&
5558 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, N0.getValueType()) &&
5559 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5560 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5561 if (LN0->getExtensionType() != ISD::ZEXTLOAD && LN0->isUnindexed()) {
5562 bool DoXform = true;
5563 SmallVector<SDNode*, 4> SetCCs;
5564 if (!N0.hasOneUse())
5565 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
5568 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN0), VT,
5569 LN0->getChain(), LN0->getBasePtr(),
5571 LN0->getMemOperand());
5572 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5573 Mask = Mask.sext(VT.getSizeInBits());
5574 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5575 ExtLoad, DAG.getConstant(Mask, VT));
5576 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5577 SDLoc(N0.getOperand(0)),
5578 N0.getOperand(0).getValueType(), ExtLoad);
5580 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5581 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5583 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5588 if (N0.getOpcode() == ISD::SETCC) {
5589 EVT N0VT = N0.getOperand(0).getValueType();
5590 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
5591 // Only do this before legalize for now.
5592 if (VT.isVector() && !LegalOperations &&
5593 TLI.getBooleanContents(N0VT) ==
5594 TargetLowering::ZeroOrNegativeOneBooleanContent) {
5595 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
5596 // of the same size as the compared operands. Only optimize sext(setcc())
5597 // if this is the case.
5598 EVT SVT = getSetCCResultType(N0VT);
5600 // We know that the # elements of the results is the same as the
5601 // # elements of the compare (and the # elements of the compare result
5602 // for that matter). Check to see that they are the same size. If so,
5603 // we know that the element size of the sext'd result matches the
5604 // element size of the compare operands.
5605 if (VT.getSizeInBits() == SVT.getSizeInBits())
5606 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5608 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5610 // If the desired elements are smaller or larger than the source
5611 // elements we can use a matching integer vector type and then
5612 // truncate/sign extend
5613 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5614 if (SVT == MatchingVectorType) {
5615 SDValue VsetCC = DAG.getSetCC(SDLoc(N), MatchingVectorType,
5616 N0.getOperand(0), N0.getOperand(1),
5617 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5618 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
5622 // sext(setcc x, y, cc) -> (select (setcc x, y, cc), -1, 0)
5623 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
5625 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
5627 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5628 NegOne, DAG.getConstant(0, VT),
5629 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5630 if (SCC.getNode()) return SCC;
5632 if (!VT.isVector()) {
5633 EVT SetCCVT = getSetCCResultType(N0.getOperand(0).getValueType());
5634 if (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, SetCCVT)) {
5636 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5637 SDValue SetCC = DAG.getSetCC(DL, SetCCVT,
5638 N0.getOperand(0), N0.getOperand(1), CC);
5639 return DAG.getSelect(DL, VT, SetCC,
5640 NegOne, DAG.getConstant(0, VT));
5645 // fold (sext x) -> (zext x) if the sign bit is known zero.
5646 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
5647 DAG.SignBitIsZero(N0))
5648 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0);
5653 // isTruncateOf - If N is a truncate of some other value, return true, record
5654 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
5655 // This function computes KnownZero to avoid a duplicated call to
5656 // computeKnownBits in the caller.
5657 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
5660 if (N->getOpcode() == ISD::TRUNCATE) {
5661 Op = N->getOperand(0);
5662 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5666 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
5667 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
5670 SDValue Op0 = N->getOperand(0);
5671 SDValue Op1 = N->getOperand(1);
5672 assert(Op0.getValueType() == Op1.getValueType());
5674 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
5675 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
5676 if (COp0 && COp0->isNullValue())
5678 else if (COp1 && COp1->isNullValue())
5683 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5685 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
5691 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
5692 SDValue N0 = N->getOperand(0);
5693 EVT VT = N->getValueType(0);
5695 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5697 return SDValue(Res, 0);
5699 // fold (zext (zext x)) -> (zext x)
5700 // fold (zext (aext x)) -> (zext x)
5701 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5702 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT,
5705 // fold (zext (truncate x)) -> (zext x) or
5706 // (zext (truncate x)) -> (truncate x)
5707 // This is valid when the truncated bits of x are already zero.
5708 // FIXME: We should extend this to work for vectors too.
5711 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
5712 APInt TruncatedBits =
5713 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
5714 APInt(Op.getValueSizeInBits(), 0) :
5715 APInt::getBitsSet(Op.getValueSizeInBits(),
5716 N0.getValueSizeInBits(),
5717 std::min(Op.getValueSizeInBits(),
5718 VT.getSizeInBits()));
5719 if (TruncatedBits == (KnownZero & TruncatedBits)) {
5720 if (VT.bitsGT(Op.getValueType()))
5721 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, Op);
5722 if (VT.bitsLT(Op.getValueType()))
5723 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5729 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5730 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
5731 if (N0.getOpcode() == ISD::TRUNCATE) {
5732 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5733 if (NarrowLoad.getNode()) {
5734 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5735 if (NarrowLoad.getNode() != N0.getNode()) {
5736 CombineTo(N0.getNode(), NarrowLoad);
5737 // CombineTo deleted the truncate, if needed, but not what's under it.
5740 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5744 // fold (zext (truncate x)) -> (and x, mask)
5745 if (N0.getOpcode() == ISD::TRUNCATE &&
5746 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
5748 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5749 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
5750 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5751 if (NarrowLoad.getNode()) {
5752 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5753 if (NarrowLoad.getNode() != N0.getNode()) {
5754 CombineTo(N0.getNode(), NarrowLoad);
5755 // CombineTo deleted the truncate, if needed, but not what's under it.
5758 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5761 SDValue Op = N0.getOperand(0);
5762 if (Op.getValueType().bitsLT(VT)) {
5763 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, Op);
5764 AddToWorklist(Op.getNode());
5765 } else if (Op.getValueType().bitsGT(VT)) {
5766 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5767 AddToWorklist(Op.getNode());
5769 return DAG.getZeroExtendInReg(Op, SDLoc(N),
5770 N0.getValueType().getScalarType());
5773 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
5774 // if either of the casts is not free.
5775 if (N0.getOpcode() == ISD::AND &&
5776 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5777 N0.getOperand(1).getOpcode() == ISD::Constant &&
5778 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5779 N0.getValueType()) ||
5780 !TLI.isZExtFree(N0.getValueType(), VT))) {
5781 SDValue X = N0.getOperand(0).getOperand(0);
5782 if (X.getValueType().bitsLT(VT)) {
5783 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(X), VT, X);
5784 } else if (X.getValueType().bitsGT(VT)) {
5785 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
5787 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5788 Mask = Mask.zext(VT.getSizeInBits());
5789 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5790 X, DAG.getConstant(Mask, VT));
5793 // fold (zext (load x)) -> (zext (truncate (zextload x)))
5794 // Only generate vector extloads when 1) they're legal, and 2) they are
5795 // deemed desirable by the target.
5796 if (ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5797 ((!LegalOperations && !VT.isVector() &&
5798 !cast<LoadSDNode>(N0)->isVolatile()) ||
5799 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, N0.getValueType()))) {
5800 bool DoXform = true;
5801 SmallVector<SDNode*, 4> SetCCs;
5802 if (!N0.hasOneUse())
5803 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
5805 DoXform &= TLI.isVectorLoadExtDesirable(SDValue(N, 0));
5807 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5808 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5810 LN0->getBasePtr(), N0.getValueType(),
5811 LN0->getMemOperand());
5812 CombineTo(N, ExtLoad);
5813 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5814 N0.getValueType(), ExtLoad);
5815 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5817 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5819 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5823 // fold (zext (load x)) to multiple smaller zextloads.
5824 // Only on illegal but splittable vectors.
5825 if (SDValue ExtLoad = CombineExtLoad(N))
5828 // fold (zext (and/or/xor (load x), cst)) ->
5829 // (and/or/xor (zextload x), (zext cst))
5830 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5831 N0.getOpcode() == ISD::XOR) &&
5832 isa<LoadSDNode>(N0.getOperand(0)) &&
5833 N0.getOperand(1).getOpcode() == ISD::Constant &&
5834 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, N0.getValueType()) &&
5835 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5836 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5837 if (LN0->getExtensionType() != ISD::SEXTLOAD && LN0->isUnindexed()) {
5838 bool DoXform = true;
5839 SmallVector<SDNode*, 4> SetCCs;
5840 if (!N0.hasOneUse())
5841 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
5844 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), VT,
5845 LN0->getChain(), LN0->getBasePtr(),
5847 LN0->getMemOperand());
5848 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5849 Mask = Mask.zext(VT.getSizeInBits());
5850 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5851 ExtLoad, DAG.getConstant(Mask, VT));
5852 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5853 SDLoc(N0.getOperand(0)),
5854 N0.getOperand(0).getValueType(), ExtLoad);
5856 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5857 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5859 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5864 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
5865 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
5866 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5867 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5868 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5869 EVT MemVT = LN0->getMemoryVT();
5870 if ((!LegalOperations && !LN0->isVolatile()) ||
5871 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT)) {
5872 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5874 LN0->getBasePtr(), MemVT,
5875 LN0->getMemOperand());
5876 CombineTo(N, ExtLoad);
5877 CombineTo(N0.getNode(),
5878 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(),
5880 ExtLoad.getValue(1));
5881 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5885 if (N0.getOpcode() == ISD::SETCC) {
5886 if (!LegalOperations && VT.isVector() &&
5887 N0.getValueType().getVectorElementType() == MVT::i1) {
5888 EVT N0VT = N0.getOperand(0).getValueType();
5889 if (getSetCCResultType(N0VT) == N0.getValueType())
5892 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
5893 // Only do this before legalize for now.
5894 EVT EltVT = VT.getVectorElementType();
5895 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
5896 DAG.getConstant(1, EltVT));
5897 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5898 // We know that the # elements of the results is the same as the
5899 // # elements of the compare (and the # elements of the compare result
5900 // for that matter). Check to see that they are the same size. If so,
5901 // we know that the element size of the sext'd result matches the
5902 // element size of the compare operands.
5903 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5904 DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5906 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
5907 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT,
5910 // If the desired elements are smaller or larger than the source
5911 // elements we can use a matching integer vector type and then
5912 // truncate/sign extend
5913 EVT MatchingElementType =
5914 EVT::getIntegerVT(*DAG.getContext(),
5915 N0VT.getScalarType().getSizeInBits());
5916 EVT MatchingVectorType =
5917 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
5918 N0VT.getVectorNumElements());
5920 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5922 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5923 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5924 DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT),
5925 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, OneOps));
5928 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5930 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5931 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5932 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5933 if (SCC.getNode()) return SCC;
5936 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
5937 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
5938 isa<ConstantSDNode>(N0.getOperand(1)) &&
5939 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
5941 SDValue ShAmt = N0.getOperand(1);
5942 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
5943 if (N0.getOpcode() == ISD::SHL) {
5944 SDValue InnerZExt = N0.getOperand(0);
5945 // If the original shl may be shifting out bits, do not perform this
5947 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
5948 InnerZExt.getOperand(0).getValueType().getSizeInBits();
5949 if (ShAmtVal > KnownZeroBits)
5955 // Ensure that the shift amount is wide enough for the shifted value.
5956 if (VT.getSizeInBits() >= 256)
5957 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
5959 return DAG.getNode(N0.getOpcode(), DL, VT,
5960 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
5967 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
5968 SDValue N0 = N->getOperand(0);
5969 EVT VT = N->getValueType(0);
5971 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5973 return SDValue(Res, 0);
5975 // fold (aext (aext x)) -> (aext x)
5976 // fold (aext (zext x)) -> (zext x)
5977 // fold (aext (sext x)) -> (sext x)
5978 if (N0.getOpcode() == ISD::ANY_EXTEND ||
5979 N0.getOpcode() == ISD::ZERO_EXTEND ||
5980 N0.getOpcode() == ISD::SIGN_EXTEND)
5981 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0));
5983 // fold (aext (truncate (load x))) -> (aext (smaller load x))
5984 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
5985 if (N0.getOpcode() == ISD::TRUNCATE) {
5986 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5987 if (NarrowLoad.getNode()) {
5988 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5989 if (NarrowLoad.getNode() != N0.getNode()) {
5990 CombineTo(N0.getNode(), NarrowLoad);
5991 // CombineTo deleted the truncate, if needed, but not what's under it.
5994 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5998 // fold (aext (truncate x))
5999 if (N0.getOpcode() == ISD::TRUNCATE) {
6000 SDValue TruncOp = N0.getOperand(0);
6001 if (TruncOp.getValueType() == VT)
6002 return TruncOp; // x iff x size == zext size.
6003 if (TruncOp.getValueType().bitsGT(VT))
6004 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, TruncOp);
6005 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, TruncOp);
6008 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
6009 // if the trunc is not free.
6010 if (N0.getOpcode() == ISD::AND &&
6011 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
6012 N0.getOperand(1).getOpcode() == ISD::Constant &&
6013 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
6014 N0.getValueType())) {
6015 SDValue X = N0.getOperand(0).getOperand(0);
6016 if (X.getValueType().bitsLT(VT)) {
6017 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, X);
6018 } else if (X.getValueType().bitsGT(VT)) {
6019 X = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, X);
6021 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
6022 Mask = Mask.zext(VT.getSizeInBits());
6023 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6024 X, DAG.getConstant(Mask, VT));
6027 // fold (aext (load x)) -> (aext (truncate (extload x)))
6028 // None of the supported targets knows how to perform load and any_ext
6029 // on vectors in one instruction. We only perform this transformation on
6031 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
6032 ISD::isUNINDEXEDLoad(N0.getNode()) &&
6033 TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) {
6034 bool DoXform = true;
6035 SmallVector<SDNode*, 4> SetCCs;
6036 if (!N0.hasOneUse())
6037 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
6039 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6040 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
6042 LN0->getBasePtr(), N0.getValueType(),
6043 LN0->getMemOperand());
6044 CombineTo(N, ExtLoad);
6045 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
6046 N0.getValueType(), ExtLoad);
6047 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
6048 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
6050 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6054 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
6055 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
6056 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
6057 if (N0.getOpcode() == ISD::LOAD &&
6058 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
6060 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6061 ISD::LoadExtType ExtType = LN0->getExtensionType();
6062 EVT MemVT = LN0->getMemoryVT();
6063 if (!LegalOperations || TLI.isLoadExtLegal(ExtType, VT, MemVT)) {
6064 SDValue ExtLoad = DAG.getExtLoad(ExtType, SDLoc(N),
6065 VT, LN0->getChain(), LN0->getBasePtr(),
6066 MemVT, LN0->getMemOperand());
6067 CombineTo(N, ExtLoad);
6068 CombineTo(N0.getNode(),
6069 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
6070 N0.getValueType(), ExtLoad),
6071 ExtLoad.getValue(1));
6072 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6076 if (N0.getOpcode() == ISD::SETCC) {
6078 // aext(setcc) -> vsetcc
6079 // aext(setcc) -> truncate(vsetcc)
6080 // aext(setcc) -> aext(vsetcc)
6081 // Only do this before legalize for now.
6082 if (VT.isVector() && !LegalOperations) {
6083 EVT N0VT = N0.getOperand(0).getValueType();
6084 // We know that the # elements of the results is the same as the
6085 // # elements of the compare (and the # elements of the compare result
6086 // for that matter). Check to see that they are the same size. If so,
6087 // we know that the element size of the sext'd result matches the
6088 // element size of the compare operands.
6089 if (VT.getSizeInBits() == N0VT.getSizeInBits())
6090 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
6092 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6093 // If the desired elements are smaller or larger than the source
6094 // elements we can use a matching integer vector type and then
6095 // truncate/any extend
6097 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
6099 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
6101 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6102 return DAG.getAnyExtOrTrunc(VsetCC, SDLoc(N), VT);
6106 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
6108 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
6109 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
6110 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
6118 /// See if the specified operand can be simplified with the knowledge that only
6119 /// the bits specified by Mask are used. If so, return the simpler operand,
6120 /// otherwise return a null SDValue.
6121 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
6122 switch (V.getOpcode()) {
6124 case ISD::Constant: {
6125 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
6126 assert(CV && "Const value should be ConstSDNode.");
6127 const APInt &CVal = CV->getAPIntValue();
6128 APInt NewVal = CVal & Mask;
6130 return DAG.getConstant(NewVal, V.getValueType());
6135 // If the LHS or RHS don't contribute bits to the or, drop them.
6136 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
6137 return V.getOperand(1);
6138 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
6139 return V.getOperand(0);
6142 // Only look at single-use SRLs.
6143 if (!V.getNode()->hasOneUse())
6145 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
6146 // See if we can recursively simplify the LHS.
6147 unsigned Amt = RHSC->getZExtValue();
6149 // Watch out for shift count overflow though.
6150 if (Amt >= Mask.getBitWidth()) break;
6151 APInt NewMask = Mask << Amt;
6152 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
6153 if (SimplifyLHS.getNode())
6154 return DAG.getNode(ISD::SRL, SDLoc(V), V.getValueType(),
6155 SimplifyLHS, V.getOperand(1));
6161 /// If the result of a wider load is shifted to right of N bits and then
6162 /// truncated to a narrower type and where N is a multiple of number of bits of
6163 /// the narrower type, transform it to a narrower load from address + N / num of
6164 /// bits of new type. If the result is to be extended, also fold the extension
6165 /// to form a extending load.
6166 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
6167 unsigned Opc = N->getOpcode();
6169 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
6170 SDValue N0 = N->getOperand(0);
6171 EVT VT = N->getValueType(0);
6174 // This transformation isn't valid for vector loads.
6178 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
6180 if (Opc == ISD::SIGN_EXTEND_INREG) {
6181 ExtType = ISD::SEXTLOAD;
6182 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
6183 } else if (Opc == ISD::SRL) {
6184 // Another special-case: SRL is basically zero-extending a narrower value.
6185 ExtType = ISD::ZEXTLOAD;
6187 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
6188 if (!N01) return SDValue();
6189 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
6190 VT.getSizeInBits() - N01->getZExtValue());
6192 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, VT, ExtVT))
6195 unsigned EVTBits = ExtVT.getSizeInBits();
6197 // Do not generate loads of non-round integer types since these can
6198 // be expensive (and would be wrong if the type is not byte sized).
6199 if (!ExtVT.isRound())
6203 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
6204 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
6205 ShAmt = N01->getZExtValue();
6206 // Is the shift amount a multiple of size of VT?
6207 if ((ShAmt & (EVTBits-1)) == 0) {
6208 N0 = N0.getOperand(0);
6209 // Is the load width a multiple of size of VT?
6210 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
6214 // At this point, we must have a load or else we can't do the transform.
6215 if (!isa<LoadSDNode>(N0)) return SDValue();
6217 // Because a SRL must be assumed to *need* to zero-extend the high bits
6218 // (as opposed to anyext the high bits), we can't combine the zextload
6219 // lowering of SRL and an sextload.
6220 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
6223 // If the shift amount is larger than the input type then we're not
6224 // accessing any of the loaded bytes. If the load was a zextload/extload
6225 // then the result of the shift+trunc is zero/undef (handled elsewhere).
6226 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
6231 // If the load is shifted left (and the result isn't shifted back right),
6232 // we can fold the truncate through the shift.
6233 unsigned ShLeftAmt = 0;
6234 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
6235 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
6236 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
6237 ShLeftAmt = N01->getZExtValue();
6238 N0 = N0.getOperand(0);
6242 // If we haven't found a load, we can't narrow it. Don't transform one with
6243 // multiple uses, this would require adding a new load.
6244 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse())
6247 // Don't change the width of a volatile load.
6248 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6249 if (LN0->isVolatile())
6252 // Verify that we are actually reducing a load width here.
6253 if (LN0->getMemoryVT().getSizeInBits() < EVTBits)
6256 // For the transform to be legal, the load must produce only two values
6257 // (the value loaded and the chain). Don't transform a pre-increment
6258 // load, for example, which produces an extra value. Otherwise the
6259 // transformation is not equivalent, and the downstream logic to replace
6260 // uses gets things wrong.
6261 if (LN0->getNumValues() > 2)
6264 // If the load that we're shrinking is an extload and we're not just
6265 // discarding the extension we can't simply shrink the load. Bail.
6266 // TODO: It would be possible to merge the extensions in some cases.
6267 if (LN0->getExtensionType() != ISD::NON_EXTLOAD &&
6268 LN0->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits() + ShAmt)
6271 if (!TLI.shouldReduceLoadWidth(LN0, ExtType, ExtVT))
6274 EVT PtrType = N0.getOperand(1).getValueType();
6276 if (PtrType == MVT::Untyped || PtrType.isExtended())
6277 // It's not possible to generate a constant of extended or untyped type.
6280 // For big endian targets, we need to adjust the offset to the pointer to
6281 // load the correct bytes.
6282 if (TLI.isBigEndian()) {
6283 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
6284 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
6285 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
6288 uint64_t PtrOff = ShAmt / 8;
6289 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
6290 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0),
6291 PtrType, LN0->getBasePtr(),
6292 DAG.getConstant(PtrOff, PtrType));
6293 AddToWorklist(NewPtr.getNode());
6296 if (ExtType == ISD::NON_EXTLOAD)
6297 Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr,
6298 LN0->getPointerInfo().getWithOffset(PtrOff),
6299 LN0->isVolatile(), LN0->isNonTemporal(),
6300 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
6302 Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(),NewPtr,
6303 LN0->getPointerInfo().getWithOffset(PtrOff),
6304 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
6305 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
6307 // Replace the old load's chain with the new load's chain.
6308 WorklistRemover DeadNodes(*this);
6309 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
6311 // Shift the result left, if we've swallowed a left shift.
6312 SDValue Result = Load;
6313 if (ShLeftAmt != 0) {
6314 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
6315 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
6317 // If the shift amount is as large as the result size (but, presumably,
6318 // no larger than the source) then the useful bits of the result are
6319 // zero; we can't simply return the shortened shift, because the result
6320 // of that operation is undefined.
6321 if (ShLeftAmt >= VT.getSizeInBits())
6322 Result = DAG.getConstant(0, VT);
6324 Result = DAG.getNode(ISD::SHL, SDLoc(N0), VT,
6325 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
6328 // Return the new loaded value.
6332 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
6333 SDValue N0 = N->getOperand(0);
6334 SDValue N1 = N->getOperand(1);
6335 EVT VT = N->getValueType(0);
6336 EVT EVT = cast<VTSDNode>(N1)->getVT();
6337 unsigned VTBits = VT.getScalarType().getSizeInBits();
6338 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
6340 // fold (sext_in_reg c1) -> c1
6341 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
6342 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1);
6344 // If the input is already sign extended, just drop the extension.
6345 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
6348 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
6349 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
6350 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT()))
6351 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
6352 N0.getOperand(0), N1);
6354 // fold (sext_in_reg (sext x)) -> (sext x)
6355 // fold (sext_in_reg (aext x)) -> (sext x)
6356 // if x is small enough.
6357 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
6358 SDValue N00 = N0.getOperand(0);
6359 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
6360 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
6361 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
6364 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
6365 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
6366 return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT);
6368 // fold operands of sext_in_reg based on knowledge that the top bits are not
6370 if (SimplifyDemandedBits(SDValue(N, 0)))
6371 return SDValue(N, 0);
6373 // fold (sext_in_reg (load x)) -> (smaller sextload x)
6374 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
6375 SDValue NarrowLoad = ReduceLoadWidth(N);
6376 if (NarrowLoad.getNode())
6379 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
6380 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
6381 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
6382 if (N0.getOpcode() == ISD::SRL) {
6383 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
6384 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
6385 // We can turn this into an SRA iff the input to the SRL is already sign
6387 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
6388 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
6389 return DAG.getNode(ISD::SRA, SDLoc(N), VT,
6390 N0.getOperand(0), N0.getOperand(1));
6394 // fold (sext_inreg (extload x)) -> (sextload x)
6395 if (ISD::isEXTLoad(N0.getNode()) &&
6396 ISD::isUNINDEXEDLoad(N0.getNode()) &&
6397 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
6398 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6399 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) {
6400 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6401 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
6403 LN0->getBasePtr(), EVT,
6404 LN0->getMemOperand());
6405 CombineTo(N, ExtLoad);
6406 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
6407 AddToWorklist(ExtLoad.getNode());
6408 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6410 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
6411 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
6413 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
6414 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6415 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) {
6416 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6417 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
6419 LN0->getBasePtr(), EVT,
6420 LN0->getMemOperand());
6421 CombineTo(N, ExtLoad);
6422 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
6423 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6426 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
6427 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
6428 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
6429 N0.getOperand(1), false);
6430 if (BSwap.getNode())
6431 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
6435 // Fold a sext_inreg of a build_vector of ConstantSDNodes or undefs
6436 // into a build_vector.
6437 if (ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
6438 SmallVector<SDValue, 8> Elts;
6439 unsigned NumElts = N0->getNumOperands();
6440 unsigned ShAmt = VTBits - EVTBits;
6442 for (unsigned i = 0; i != NumElts; ++i) {
6443 SDValue Op = N0->getOperand(i);
6444 if (Op->getOpcode() == ISD::UNDEF) {
6449 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
6450 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
6451 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
6452 Op.getValueType()));
6455 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Elts);
6461 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
6462 SDValue N0 = N->getOperand(0);
6463 EVT VT = N->getValueType(0);
6464 bool isLE = TLI.isLittleEndian();
6467 if (N0.getValueType() == N->getValueType(0))
6469 // fold (truncate c1) -> c1
6470 if (isa<ConstantSDNode>(N0))
6471 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0);
6472 // fold (truncate (truncate x)) -> (truncate x)
6473 if (N0.getOpcode() == ISD::TRUNCATE)
6474 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6475 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
6476 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
6477 N0.getOpcode() == ISD::SIGN_EXTEND ||
6478 N0.getOpcode() == ISD::ANY_EXTEND) {
6479 if (N0.getOperand(0).getValueType().bitsLT(VT))
6480 // if the source is smaller than the dest, we still need an extend
6481 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
6483 if (N0.getOperand(0).getValueType().bitsGT(VT))
6484 // if the source is larger than the dest, than we just need the truncate
6485 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6486 // if the source and dest are the same type, we can drop both the extend
6487 // and the truncate.
6488 return N0.getOperand(0);
6491 // Fold extract-and-trunc into a narrow extract. For example:
6492 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
6493 // i32 y = TRUNCATE(i64 x)
6495 // v16i8 b = BITCAST (v2i64 val)
6496 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
6498 // Note: We only run this optimization after type legalization (which often
6499 // creates this pattern) and before operation legalization after which
6500 // we need to be more careful about the vector instructions that we generate.
6501 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6502 LegalTypes && !LegalOperations && N0->hasOneUse() && VT != MVT::i1) {
6504 EVT VecTy = N0.getOperand(0).getValueType();
6505 EVT ExTy = N0.getValueType();
6506 EVT TrTy = N->getValueType(0);
6508 unsigned NumElem = VecTy.getVectorNumElements();
6509 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
6511 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
6512 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
6514 SDValue EltNo = N0->getOperand(1);
6515 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
6516 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
6517 EVT IndexTy = TLI.getVectorIdxTy();
6518 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
6520 SDValue V = DAG.getNode(ISD::BITCAST, SDLoc(N),
6521 NVT, N0.getOperand(0));
6523 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
6525 DAG.getConstant(Index, IndexTy));
6529 // trunc (select c, a, b) -> select c, (trunc a), (trunc b)
6530 if (N0.getOpcode() == ISD::SELECT) {
6531 EVT SrcVT = N0.getValueType();
6532 if ((!LegalOperations || TLI.isOperationLegal(ISD::SELECT, SrcVT)) &&
6533 TLI.isTruncateFree(SrcVT, VT)) {
6535 SDValue Cond = N0.getOperand(0);
6536 SDValue TruncOp0 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(1));
6537 SDValue TruncOp1 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(2));
6538 return DAG.getNode(ISD::SELECT, SDLoc(N), VT, Cond, TruncOp0, TruncOp1);
6542 // Fold a series of buildvector, bitcast, and truncate if possible.
6544 // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to
6545 // (2xi32 (buildvector x, y)).
6546 if (Level == AfterLegalizeVectorOps && VT.isVector() &&
6547 N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
6548 N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
6549 N0.getOperand(0).hasOneUse()) {
6551 SDValue BuildVect = N0.getOperand(0);
6552 EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType();
6553 EVT TruncVecEltTy = VT.getVectorElementType();
6555 // Check that the element types match.
6556 if (BuildVectEltTy == TruncVecEltTy) {
6557 // Now we only need to compute the offset of the truncated elements.
6558 unsigned BuildVecNumElts = BuildVect.getNumOperands();
6559 unsigned TruncVecNumElts = VT.getVectorNumElements();
6560 unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts;
6562 assert((BuildVecNumElts % TruncVecNumElts) == 0 &&
6563 "Invalid number of elements");
6565 SmallVector<SDValue, 8> Opnds;
6566 for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset)
6567 Opnds.push_back(BuildVect.getOperand(i));
6569 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
6573 // See if we can simplify the input to this truncate through knowledge that
6574 // only the low bits are being used.
6575 // For example "trunc (or (shl x, 8), y)" // -> trunc y
6576 // Currently we only perform this optimization on scalars because vectors
6577 // may have different active low bits.
6578 if (!VT.isVector()) {
6580 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
6581 VT.getSizeInBits()));
6582 if (Shorter.getNode())
6583 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter);
6585 // fold (truncate (load x)) -> (smaller load x)
6586 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
6587 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
6588 SDValue Reduced = ReduceLoadWidth(N);
6589 if (Reduced.getNode())
6591 // Handle the case where the load remains an extending load even
6592 // after truncation.
6593 if (N0.hasOneUse() && ISD::isUNINDEXEDLoad(N0.getNode())) {
6594 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6595 if (!LN0->isVolatile() &&
6596 LN0->getMemoryVT().getStoreSizeInBits() < VT.getSizeInBits()) {
6597 SDValue NewLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(LN0),
6598 VT, LN0->getChain(), LN0->getBasePtr(),
6600 LN0->getMemOperand());
6601 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLoad.getValue(1));
6606 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
6607 // where ... are all 'undef'.
6608 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
6609 SmallVector<EVT, 8> VTs;
6612 unsigned NumDefs = 0;
6614 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
6615 SDValue X = N0.getOperand(i);
6616 if (X.getOpcode() != ISD::UNDEF) {
6621 // Stop if more than one members are non-undef.
6624 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
6625 VT.getVectorElementType(),
6626 X.getValueType().getVectorNumElements()));
6630 return DAG.getUNDEF(VT);
6633 assert(V.getNode() && "The single defined operand is empty!");
6634 SmallVector<SDValue, 8> Opnds;
6635 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
6637 Opnds.push_back(DAG.getUNDEF(VTs[i]));
6640 SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V);
6641 AddToWorklist(NV.getNode());
6642 Opnds.push_back(NV);
6644 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Opnds);
6648 // Simplify the operands using demanded-bits information.
6649 if (!VT.isVector() &&
6650 SimplifyDemandedBits(SDValue(N, 0)))
6651 return SDValue(N, 0);
6656 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
6657 SDValue Elt = N->getOperand(i);
6658 if (Elt.getOpcode() != ISD::MERGE_VALUES)
6659 return Elt.getNode();
6660 return Elt.getOperand(Elt.getResNo()).getNode();
6663 /// build_pair (load, load) -> load
6664 /// if load locations are consecutive.
6665 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
6666 assert(N->getOpcode() == ISD::BUILD_PAIR);
6668 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
6669 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
6670 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
6671 LD1->getAddressSpace() != LD2->getAddressSpace())
6673 EVT LD1VT = LD1->getValueType(0);
6675 if (ISD::isNON_EXTLoad(LD2) &&
6677 // If both are volatile this would reduce the number of volatile loads.
6678 // If one is volatile it might be ok, but play conservative and bail out.
6679 !LD1->isVolatile() &&
6680 !LD2->isVolatile() &&
6681 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
6682 unsigned Align = LD1->getAlignment();
6683 unsigned NewAlign = TLI.getDataLayout()->
6684 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6686 if (NewAlign <= Align &&
6687 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
6688 return DAG.getLoad(VT, SDLoc(N), LD1->getChain(),
6689 LD1->getBasePtr(), LD1->getPointerInfo(),
6690 false, false, false, Align);
6696 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
6697 SDValue N0 = N->getOperand(0);
6698 EVT VT = N->getValueType(0);
6700 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
6701 // Only do this before legalize, since afterward the target may be depending
6702 // on the bitconvert.
6703 // First check to see if this is all constant.
6705 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
6707 bool isSimple = cast<BuildVectorSDNode>(N0)->isConstant();
6709 EVT DestEltVT = N->getValueType(0).getVectorElementType();
6710 assert(!DestEltVT.isVector() &&
6711 "Element type of vector ValueType must not be vector!");
6713 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
6716 // If the input is a constant, let getNode fold it.
6717 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
6718 // If we can't allow illegal operations, we need to check that this is just
6719 // a fp -> int or int -> conversion and that the resulting operation will
6721 if (!LegalOperations ||
6722 (isa<ConstantSDNode>(N0) && VT.isFloatingPoint() && !VT.isVector() &&
6723 TLI.isOperationLegal(ISD::ConstantFP, VT)) ||
6724 (isa<ConstantFPSDNode>(N0) && VT.isInteger() && !VT.isVector() &&
6725 TLI.isOperationLegal(ISD::Constant, VT)))
6726 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, N0);
6729 // (conv (conv x, t1), t2) -> (conv x, t2)
6730 if (N0.getOpcode() == ISD::BITCAST)
6731 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT,
6734 // fold (conv (load x)) -> (load (conv*)x)
6735 // If the resultant load doesn't need a higher alignment than the original!
6736 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6737 // Do not change the width of a volatile load.
6738 !cast<LoadSDNode>(N0)->isVolatile() &&
6739 // Do not remove the cast if the types differ in endian layout.
6740 TLI.hasBigEndianPartOrdering(N0.getValueType()) ==
6741 TLI.hasBigEndianPartOrdering(VT) &&
6742 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)) &&
6743 TLI.isLoadBitCastBeneficial(N0.getValueType(), VT)) {
6744 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6745 unsigned Align = TLI.getDataLayout()->
6746 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6747 unsigned OrigAlign = LN0->getAlignment();
6749 if (Align <= OrigAlign) {
6750 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(),
6751 LN0->getBasePtr(), LN0->getPointerInfo(),
6752 LN0->isVolatile(), LN0->isNonTemporal(),
6753 LN0->isInvariant(), OrigAlign,
6755 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
6760 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
6761 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
6762 // This often reduces constant pool loads.
6763 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) ||
6764 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) &&
6765 N0.getNode()->hasOneUse() && VT.isInteger() &&
6766 !VT.isVector() && !N0.getValueType().isVector()) {
6767 SDValue NewConv = DAG.getNode(ISD::BITCAST, SDLoc(N0), VT,
6769 AddToWorklist(NewConv.getNode());
6771 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6772 if (N0.getOpcode() == ISD::FNEG)
6773 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
6774 NewConv, DAG.getConstant(SignBit, VT));
6775 assert(N0.getOpcode() == ISD::FABS);
6776 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6777 NewConv, DAG.getConstant(~SignBit, VT));
6780 // fold (bitconvert (fcopysign cst, x)) ->
6781 // (or (and (bitconvert x), sign), (and cst, (not sign)))
6782 // Note that we don't handle (copysign x, cst) because this can always be
6783 // folded to an fneg or fabs.
6784 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
6785 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
6786 VT.isInteger() && !VT.isVector()) {
6787 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
6788 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
6789 if (isTypeLegal(IntXVT)) {
6790 SDValue X = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6791 IntXVT, N0.getOperand(1));
6792 AddToWorklist(X.getNode());
6794 // If X has a different width than the result/lhs, sext it or truncate it.
6795 unsigned VTWidth = VT.getSizeInBits();
6796 if (OrigXWidth < VTWidth) {
6797 X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X);
6798 AddToWorklist(X.getNode());
6799 } else if (OrigXWidth > VTWidth) {
6800 // To get the sign bit in the right place, we have to shift it right
6801 // before truncating.
6802 X = DAG.getNode(ISD::SRL, SDLoc(X),
6803 X.getValueType(), X,
6804 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
6805 AddToWorklist(X.getNode());
6806 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
6807 AddToWorklist(X.getNode());
6810 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6811 X = DAG.getNode(ISD::AND, SDLoc(X), VT,
6812 X, DAG.getConstant(SignBit, VT));
6813 AddToWorklist(X.getNode());
6815 SDValue Cst = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6816 VT, N0.getOperand(0));
6817 Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT,
6818 Cst, DAG.getConstant(~SignBit, VT));
6819 AddToWorklist(Cst.getNode());
6821 return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst);
6825 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
6826 if (N0.getOpcode() == ISD::BUILD_PAIR) {
6827 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
6828 if (CombineLD.getNode())
6835 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
6836 EVT VT = N->getValueType(0);
6837 return CombineConsecutiveLoads(N, VT);
6840 /// We know that BV is a build_vector node with Constant, ConstantFP or Undef
6841 /// operands. DstEltVT indicates the destination element value type.
6842 SDValue DAGCombiner::
6843 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
6844 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
6846 // If this is already the right type, we're done.
6847 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
6849 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
6850 unsigned DstBitSize = DstEltVT.getSizeInBits();
6852 // If this is a conversion of N elements of one type to N elements of another
6853 // type, convert each element. This handles FP<->INT cases.
6854 if (SrcBitSize == DstBitSize) {
6855 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6856 BV->getValueType(0).getVectorNumElements());
6858 // Due to the FP element handling below calling this routine recursively,
6859 // we can end up with a scalar-to-vector node here.
6860 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
6861 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6862 DAG.getNode(ISD::BITCAST, SDLoc(BV),
6863 DstEltVT, BV->getOperand(0)));
6865 SmallVector<SDValue, 8> Ops;
6866 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6867 SDValue Op = BV->getOperand(i);
6868 // If the vector element type is not legal, the BUILD_VECTOR operands
6869 // are promoted and implicitly truncated. Make that explicit here.
6870 if (Op.getValueType() != SrcEltVT)
6871 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op);
6872 Ops.push_back(DAG.getNode(ISD::BITCAST, SDLoc(BV),
6874 AddToWorklist(Ops.back().getNode());
6876 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6879 // Otherwise, we're growing or shrinking the elements. To avoid having to
6880 // handle annoying details of growing/shrinking FP values, we convert them to
6882 if (SrcEltVT.isFloatingPoint()) {
6883 // Convert the input float vector to a int vector where the elements are the
6885 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
6886 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
6890 // Now we know the input is an integer vector. If the output is a FP type,
6891 // convert to integer first, then to FP of the right size.
6892 if (DstEltVT.isFloatingPoint()) {
6893 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
6894 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
6896 // Next, convert to FP elements of the same size.
6897 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
6900 // Okay, we know the src/dst types are both integers of differing types.
6901 // Handling growing first.
6902 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
6903 if (SrcBitSize < DstBitSize) {
6904 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
6906 SmallVector<SDValue, 8> Ops;
6907 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
6908 i += NumInputsPerOutput) {
6909 bool isLE = TLI.isLittleEndian();
6910 APInt NewBits = APInt(DstBitSize, 0);
6911 bool EltIsUndef = true;
6912 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
6913 // Shift the previously computed bits over.
6914 NewBits <<= SrcBitSize;
6915 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
6916 if (Op.getOpcode() == ISD::UNDEF) continue;
6919 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
6920 zextOrTrunc(SrcBitSize).zext(DstBitSize);
6924 Ops.push_back(DAG.getUNDEF(DstEltVT));
6926 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
6929 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
6930 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6933 // Finally, this must be the case where we are shrinking elements: each input
6934 // turns into multiple outputs.
6935 bool isS2V = ISD::isScalarToVector(BV);
6936 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
6937 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6938 NumOutputsPerInput*BV->getNumOperands());
6939 SmallVector<SDValue, 8> Ops;
6941 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6942 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
6943 Ops.append(NumOutputsPerInput, DAG.getUNDEF(DstEltVT));
6947 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
6948 getAPIntValue().zextOrTrunc(SrcBitSize);
6950 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
6951 APInt ThisVal = OpVal.trunc(DstBitSize);
6952 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
6953 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
6954 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
6955 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6957 OpVal = OpVal.lshr(DstBitSize);
6960 // For big endian targets, swap the order of the pieces of each element.
6961 if (TLI.isBigEndian())
6962 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
6965 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6968 // Attempt different variants of (fadd (fmul a, b), c) -> fma or fmad
6969 static SDValue performFaddFmulCombines(unsigned FusedOpcode,
6972 const TargetLowering &TLI,
6973 SelectionDAG &DAG) {
6974 SDValue N0 = N->getOperand(0);
6975 SDValue N1 = N->getOperand(1);
6976 EVT VT = N->getValueType(0);
6978 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
6979 if (N0.getOpcode() == ISD::FMUL &&
6980 (Aggressive || N0->hasOneUse())) {
6981 return DAG.getNode(FusedOpcode, SDLoc(N), VT,
6982 N0.getOperand(0), N0.getOperand(1), N1);
6985 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
6986 // Note: Commutes FADD operands.
6987 if (N1.getOpcode() == ISD::FMUL &&
6988 (Aggressive || N1->hasOneUse())) {
6989 return DAG.getNode(FusedOpcode, SDLoc(N), VT,
6990 N1.getOperand(0), N1.getOperand(1), N0);
6993 // More folding opportunities when target permits.
6995 // fold (fadd (fma x, y, (fmul u, v)), z) -> (fma x, y (fma u, v, z))
6996 if (N0.getOpcode() == ISD::FMA &&
6997 N0.getOperand(2).getOpcode() == ISD::FMUL) {
6998 return DAG.getNode(FusedOpcode, SDLoc(N), VT,
6999 N0.getOperand(0), N0.getOperand(1),
7000 DAG.getNode(FusedOpcode, SDLoc(N), VT,
7001 N0.getOperand(2).getOperand(0),
7002 N0.getOperand(2).getOperand(1),
7006 // fold (fadd x, (fma y, z, (fmul u, v)) -> (fma y, z (fma u, v, x))
7007 if (N1->getOpcode() == ISD::FMA &&
7008 N1.getOperand(2).getOpcode() == ISD::FMUL) {
7009 return DAG.getNode(FusedOpcode, SDLoc(N), VT,
7010 N1.getOperand(0), N1.getOperand(1),
7011 DAG.getNode(FusedOpcode, SDLoc(N), VT,
7012 N1.getOperand(2).getOperand(0),
7013 N1.getOperand(2).getOperand(1),
7021 static SDValue performFsubFmulCombines(unsigned FusedOpcode,
7024 const TargetLowering &TLI,
7025 SelectionDAG &DAG) {
7026 SDValue N0 = N->getOperand(0);
7027 SDValue N1 = N->getOperand(1);
7028 EVT VT = N->getValueType(0);
7032 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
7033 if (N0.getOpcode() == ISD::FMUL &&
7034 (Aggressive || N0->hasOneUse())) {
7035 return DAG.getNode(FusedOpcode, SL, VT,
7036 N0.getOperand(0), N0.getOperand(1),
7037 DAG.getNode(ISD::FNEG, SL, VT, N1));
7040 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
7041 // Note: Commutes FSUB operands.
7042 if (N1.getOpcode() == ISD::FMUL &&
7043 (Aggressive || N1->hasOneUse()))
7044 return DAG.getNode(FusedOpcode, SL, VT,
7045 DAG.getNode(ISD::FNEG, SL, VT,
7047 N1.getOperand(1), N0);
7049 // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
7050 if (N0.getOpcode() == ISD::FNEG &&
7051 N0.getOperand(0).getOpcode() == ISD::FMUL &&
7052 (Aggressive || (N0->hasOneUse() && N0.getOperand(0).hasOneUse()))) {
7053 SDValue N00 = N0.getOperand(0).getOperand(0);
7054 SDValue N01 = N0.getOperand(0).getOperand(1);
7055 return DAG.getNode(FusedOpcode, SL, VT,
7056 DAG.getNode(ISD::FNEG, SL, VT, N00), N01,
7057 DAG.getNode(ISD::FNEG, SL, VT, N1));
7060 // More folding opportunities when target permits.
7062 // fold (fsub (fma x, y, (fmul u, v)), z)
7063 // -> (fma x, y (fma u, v, (fneg z)))
7064 if (N0.getOpcode() == FusedOpcode &&
7065 N0.getOperand(2).getOpcode() == ISD::FMUL) {
7066 return DAG.getNode(FusedOpcode, SDLoc(N), VT,
7067 N0.getOperand(0), N0.getOperand(1),
7068 DAG.getNode(FusedOpcode, SDLoc(N), VT,
7069 N0.getOperand(2).getOperand(0),
7070 N0.getOperand(2).getOperand(1),
7071 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7075 // fold (fsub x, (fma y, z, (fmul u, v)))
7076 // -> (fma (fneg y), z, (fma (fneg u), v, x))
7077 if (N1.getOpcode() == FusedOpcode &&
7078 N1.getOperand(2).getOpcode() == ISD::FMUL) {
7079 SDValue N20 = N1.getOperand(2).getOperand(0);
7080 SDValue N21 = N1.getOperand(2).getOperand(1);
7081 return DAG.getNode(FusedOpcode, SDLoc(N), VT,
7082 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7085 DAG.getNode(FusedOpcode, SDLoc(N), VT,
7086 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7095 SDValue DAGCombiner::visitFADD(SDNode *N) {
7096 SDValue N0 = N->getOperand(0);
7097 SDValue N1 = N->getOperand(1);
7098 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7099 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7100 EVT VT = N->getValueType(0);
7101 const TargetOptions &Options = DAG.getTarget().Options;
7104 if (VT.isVector()) {
7105 SDValue FoldedVOp = SimplifyVBinOp(N);
7106 if (FoldedVOp.getNode()) return FoldedVOp;
7109 // fold (fadd c1, c2) -> c1 + c2
7111 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N1);
7113 // canonicalize constant to RHS
7114 if (N0CFP && !N1CFP)
7115 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N0);
7117 // fold (fadd A, (fneg B)) -> (fsub A, B)
7118 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
7119 isNegatibleForFree(N1, LegalOperations, TLI, &Options) == 2)
7120 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0,
7121 GetNegatedExpression(N1, DAG, LegalOperations));
7123 // fold (fadd (fneg A), B) -> (fsub B, A)
7124 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
7125 isNegatibleForFree(N0, LegalOperations, TLI, &Options) == 2)
7126 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N1,
7127 GetNegatedExpression(N0, DAG, LegalOperations));
7129 // If 'unsafe math' is enabled, fold lots of things.
7130 if (Options.UnsafeFPMath) {
7131 // No FP constant should be created after legalization as Instruction
7132 // Selection pass has a hard time dealing with FP constants.
7133 bool AllowNewConst = (Level < AfterLegalizeDAG);
7135 // fold (fadd A, 0) -> A
7136 if (N1CFP && N1CFP->getValueAPF().isZero())
7139 // fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
7140 if (N1CFP && N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
7141 isa<ConstantFPSDNode>(N0.getOperand(1)))
7142 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0.getOperand(0),
7143 DAG.getNode(ISD::FADD, SDLoc(N), VT,
7144 N0.getOperand(1), N1));
7146 // If allowed, fold (fadd (fneg x), x) -> 0.0
7147 if (AllowNewConst && N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1)
7148 return DAG.getConstantFP(0.0, VT);
7150 // If allowed, fold (fadd x, (fneg x)) -> 0.0
7151 if (AllowNewConst && N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0)
7152 return DAG.getConstantFP(0.0, VT);
7154 // We can fold chains of FADD's of the same value into multiplications.
7155 // This transform is not safe in general because we are reducing the number
7156 // of rounding steps.
7157 if (TLI.isOperationLegalOrCustom(ISD::FMUL, VT) && !N0CFP && !N1CFP) {
7158 if (N0.getOpcode() == ISD::FMUL) {
7159 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
7160 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
7162 // (fadd (fmul x, c), x) -> (fmul x, c+1)
7163 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
7164 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
7166 DAG.getConstantFP(1.0, VT));
7167 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, NewCFP);
7170 // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2)
7171 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
7172 N1.getOperand(0) == N1.getOperand(1) &&
7173 N0.getOperand(0) == N1.getOperand(0)) {
7174 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
7176 DAG.getConstantFP(2.0, VT));
7177 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7178 N0.getOperand(0), NewCFP);
7182 if (N1.getOpcode() == ISD::FMUL) {
7183 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
7184 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
7186 // (fadd x, (fmul x, c)) -> (fmul x, c+1)
7187 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
7188 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
7190 DAG.getConstantFP(1.0, VT));
7191 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, NewCFP);
7194 // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2)
7195 if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD &&
7196 N0.getOperand(0) == N0.getOperand(1) &&
7197 N1.getOperand(0) == N0.getOperand(0)) {
7198 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
7200 DAG.getConstantFP(2.0, VT));
7201 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1.getOperand(0), NewCFP);
7205 if (N0.getOpcode() == ISD::FADD && AllowNewConst) {
7206 ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
7207 // (fadd (fadd x, x), x) -> (fmul x, 3.0)
7208 if (!CFP && N0.getOperand(0) == N0.getOperand(1) &&
7209 (N0.getOperand(0) == N1))
7210 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7211 N1, DAG.getConstantFP(3.0, VT));
7214 if (N1.getOpcode() == ISD::FADD && AllowNewConst) {
7215 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
7216 // (fadd x, (fadd x, x)) -> (fmul x, 3.0)
7217 if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) &&
7218 N1.getOperand(0) == N0)
7219 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7220 N0, DAG.getConstantFP(3.0, VT));
7223 // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0)
7224 if (AllowNewConst &&
7225 N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
7226 N0.getOperand(0) == N0.getOperand(1) &&
7227 N1.getOperand(0) == N1.getOperand(1) &&
7228 N0.getOperand(0) == N1.getOperand(0))
7229 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7230 N0.getOperand(0), DAG.getConstantFP(4.0, VT));
7232 } // enable-unsafe-fp-math
7234 if (LegalOperations && TLI.isOperationLegal(ISD::FMAD, VT)) {
7235 // Assume if there is an fmad instruction that it should be aggressively
7237 if (SDValue Fused = performFaddFmulCombines(ISD::FMAD, true, N, TLI, DAG))
7241 // FADD -> FMA combines:
7242 if ((Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath) &&
7243 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
7244 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
7246 if (!TLI.isOperationLegal(ISD::FMAD, VT)) {
7247 // Don't form FMA if we are preferring FMAD.
7249 = performFaddFmulCombines(ISD::FMA,
7250 TLI.enableAggressiveFMAFusion(VT),
7256 // When FP_EXTEND nodes are free on the target, and there is an opportunity
7257 // to combine into FMA, arrange such nodes accordingly.
7258 if (TLI.isFPExtFree(VT)) {
7260 // fold (fadd (fpext (fmul x, y)), z) -> (fma (fpext x), (fpext y), z)
7261 if (N0.getOpcode() == ISD::FP_EXTEND) {
7262 SDValue N00 = N0.getOperand(0);
7263 if (N00.getOpcode() == ISD::FMUL)
7264 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7265 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7267 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7268 N00.getOperand(1)), N1);
7271 // fold (fadd x, (fpext (fmul y, z)), z) -> (fma (fpext y), (fpext z), x)
7272 // Note: Commutes FADD operands.
7273 if (N1.getOpcode() == ISD::FP_EXTEND) {
7274 SDValue N10 = N1.getOperand(0);
7275 if (N10.getOpcode() == ISD::FMUL)
7276 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7277 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7279 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7280 N10.getOperand(1)), N0);
7288 SDValue DAGCombiner::visitFSUB(SDNode *N) {
7289 SDValue N0 = N->getOperand(0);
7290 SDValue N1 = N->getOperand(1);
7291 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
7292 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
7293 EVT VT = N->getValueType(0);
7295 const TargetOptions &Options = DAG.getTarget().Options;
7298 if (VT.isVector()) {
7299 SDValue FoldedVOp = SimplifyVBinOp(N);
7300 if (FoldedVOp.getNode()) return FoldedVOp;
7303 // fold (fsub c1, c2) -> c1-c2
7305 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0, N1);
7307 // fold (fsub A, (fneg B)) -> (fadd A, B)
7308 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
7309 return DAG.getNode(ISD::FADD, dl, VT, N0,
7310 GetNegatedExpression(N1, DAG, LegalOperations));
7312 // If 'unsafe math' is enabled, fold lots of things.
7313 if (Options.UnsafeFPMath) {
7315 if (N1CFP && N1CFP->getValueAPF().isZero())
7318 // (fsub 0, B) -> -B
7319 if (N0CFP && N0CFP->getValueAPF().isZero()) {
7320 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
7321 return GetNegatedExpression(N1, DAG, LegalOperations);
7322 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7323 return DAG.getNode(ISD::FNEG, dl, VT, N1);
7326 // (fsub x, x) -> 0.0
7328 return DAG.getConstantFP(0.0f, VT);
7330 // (fsub x, (fadd x, y)) -> (fneg y)
7331 // (fsub x, (fadd y, x)) -> (fneg y)
7332 if (N1.getOpcode() == ISD::FADD) {
7333 SDValue N10 = N1->getOperand(0);
7334 SDValue N11 = N1->getOperand(1);
7336 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI, &Options))
7337 return GetNegatedExpression(N11, DAG, LegalOperations);
7339 if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI, &Options))
7340 return GetNegatedExpression(N10, DAG, LegalOperations);
7344 if (LegalOperations && TLI.isOperationLegal(ISD::FMAD, VT)) {
7345 // Assume if there is an fmad instruction that it should be aggressively
7347 if (SDValue Fused = performFsubFmulCombines(ISD::FMAD, true, N, TLI, DAG))
7351 // FSUB -> FMA combines:
7352 if ((Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath) &&
7353 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
7354 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
7356 if (!TLI.isOperationLegal(ISD::FMAD, VT)) {
7357 // Don't form FMA if we are preferring FMAD.
7360 = performFsubFmulCombines(ISD::FMA,
7361 TLI.enableAggressiveFMAFusion(VT),
7367 // When FP_EXTEND nodes are free on the target, and there is an opportunity
7368 // to combine into FMA, arrange such nodes accordingly.
7369 if (TLI.isFPExtFree(VT)) {
7370 // fold (fsub (fpext (fmul x, y)), z)
7371 // -> (fma (fpext x), (fpext y), (fneg z))
7372 if (N0.getOpcode() == ISD::FP_EXTEND) {
7373 SDValue N00 = N0.getOperand(0);
7374 if (N00.getOpcode() == ISD::FMUL)
7375 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7376 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7378 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7380 DAG.getNode(ISD::FNEG, SDLoc(N), VT, N1));
7383 // fold (fsub x, (fpext (fmul y, z)))
7384 // -> (fma (fneg (fpext y)), (fpext z), x)
7385 // Note: Commutes FSUB operands.
7386 if (N1.getOpcode() == ISD::FP_EXTEND) {
7387 SDValue N10 = N1.getOperand(0);
7388 if (N10.getOpcode() == ISD::FMUL)
7389 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7390 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7391 DAG.getNode(ISD::FP_EXTEND, SDLoc(N),
7392 VT, N10.getOperand(0))),
7393 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7398 // fold (fsub (fpext (fneg (fmul, x, y))), z)
7399 // -> (fma (fneg (fpext x)), (fpext y), (fneg z))
7400 if (N0.getOpcode() == ISD::FP_EXTEND) {
7401 SDValue N00 = N0.getOperand(0);
7402 if (N00.getOpcode() == ISD::FNEG) {
7403 SDValue N000 = N00.getOperand(0);
7404 if (N000.getOpcode() == ISD::FMUL) {
7405 return DAG.getNode(ISD::FMA, dl, VT,
7406 DAG.getNode(ISD::FNEG, dl, VT,
7407 DAG.getNode(ISD::FP_EXTEND, SDLoc(N),
7408 VT, N000.getOperand(0))),
7409 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7410 N000.getOperand(1)),
7411 DAG.getNode(ISD::FNEG, dl, VT, N1));
7416 // fold (fsub (fneg (fpext (fmul, x, y))), z)
7417 // -> (fma (fneg (fpext x)), (fpext y), (fneg z))
7418 if (N0.getOpcode() == ISD::FNEG) {
7419 SDValue N00 = N0.getOperand(0);
7420 if (N00.getOpcode() == ISD::FP_EXTEND) {
7421 SDValue N000 = N00.getOperand(0);
7422 if (N000.getOpcode() == ISD::FMUL) {
7423 return DAG.getNode(ISD::FMA, dl, VT,
7424 DAG.getNode(ISD::FNEG, dl, VT,
7425 DAG.getNode(ISD::FP_EXTEND, SDLoc(N),
7426 VT, N000.getOperand(0))),
7427 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7428 N000.getOperand(1)),
7429 DAG.getNode(ISD::FNEG, dl, VT, N1));
7439 SDValue DAGCombiner::visitFMUL(SDNode *N) {
7440 SDValue N0 = N->getOperand(0);
7441 SDValue N1 = N->getOperand(1);
7442 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
7443 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
7444 EVT VT = N->getValueType(0);
7445 const TargetOptions &Options = DAG.getTarget().Options;
7448 if (VT.isVector()) {
7449 // This just handles C1 * C2 for vectors. Other vector folds are below.
7450 SDValue FoldedVOp = SimplifyVBinOp(N);
7451 if (FoldedVOp.getNode())
7453 // Canonicalize vector constant to RHS.
7454 if (N0.getOpcode() == ISD::BUILD_VECTOR &&
7455 N1.getOpcode() != ISD::BUILD_VECTOR)
7456 if (auto *BV0 = dyn_cast<BuildVectorSDNode>(N0))
7457 if (BV0->isConstant())
7458 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, N1, N0);
7461 // fold (fmul c1, c2) -> c1*c2
7463 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, N1);
7465 // canonicalize constant to RHS
7466 if (N0CFP && !N1CFP)
7467 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, N0);
7469 // fold (fmul A, 1.0) -> A
7470 if (N1CFP && N1CFP->isExactlyValue(1.0))
7473 if (Options.UnsafeFPMath) {
7474 // fold (fmul A, 0) -> 0
7475 if (N1CFP && N1CFP->getValueAPF().isZero())
7478 // fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
7479 if (N0.getOpcode() == ISD::FMUL) {
7480 // Fold scalars or any vector constants (not just splats).
7481 // This fold is done in general by InstCombine, but extra fmul insts
7482 // may have been generated during lowering.
7483 SDValue N00 = N0.getOperand(0);
7484 SDValue N01 = N0.getOperand(1);
7485 auto *BV1 = dyn_cast<BuildVectorSDNode>(N1);
7486 auto *BV00 = dyn_cast<BuildVectorSDNode>(N00);
7487 auto *BV01 = dyn_cast<BuildVectorSDNode>(N01);
7489 // Check 1: Make sure that the first operand of the inner multiply is NOT
7490 // a constant. Otherwise, we may induce infinite looping.
7491 if (!(isConstOrConstSplatFP(N00) || (BV00 && BV00->isConstant()))) {
7492 // Check 2: Make sure that the second operand of the inner multiply and
7493 // the second operand of the outer multiply are constants.
7494 if ((N1CFP && isConstOrConstSplatFP(N01)) ||
7495 (BV1 && BV01 && BV1->isConstant() && BV01->isConstant())) {
7497 SDValue MulConsts = DAG.getNode(ISD::FMUL, SL, VT, N01, N1);
7498 return DAG.getNode(ISD::FMUL, SL, VT, N00, MulConsts);
7503 // fold (fmul (fadd x, x), c) -> (fmul x, (fmul 2.0, c))
7504 // Undo the fmul 2.0, x -> fadd x, x transformation, since if it occurs
7505 // during an early run of DAGCombiner can prevent folding with fmuls
7506 // inserted during lowering.
7507 if (N0.getOpcode() == ISD::FADD && N0.getOperand(0) == N0.getOperand(1)) {
7509 const SDValue Two = DAG.getConstantFP(2.0, VT);
7510 SDValue MulConsts = DAG.getNode(ISD::FMUL, SL, VT, Two, N1);
7511 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0.getOperand(0), MulConsts);
7515 // fold (fmul X, 2.0) -> (fadd X, X)
7516 if (N1CFP && N1CFP->isExactlyValue(+2.0))
7517 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N0);
7519 // fold (fmul X, -1.0) -> (fneg X)
7520 if (N1CFP && N1CFP->isExactlyValue(-1.0))
7521 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7522 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
7524 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
7525 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
7526 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
7527 // Both can be negated for free, check to see if at least one is cheaper
7529 if (LHSNeg == 2 || RHSNeg == 2)
7530 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7531 GetNegatedExpression(N0, DAG, LegalOperations),
7532 GetNegatedExpression(N1, DAG, LegalOperations));
7539 SDValue DAGCombiner::visitFMA(SDNode *N) {
7540 SDValue N0 = N->getOperand(0);
7541 SDValue N1 = N->getOperand(1);
7542 SDValue N2 = N->getOperand(2);
7543 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7544 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7545 EVT VT = N->getValueType(0);
7547 const TargetOptions &Options = DAG.getTarget().Options;
7549 // Constant fold FMA.
7550 if (isa<ConstantFPSDNode>(N0) &&
7551 isa<ConstantFPSDNode>(N1) &&
7552 isa<ConstantFPSDNode>(N2)) {
7553 return DAG.getNode(ISD::FMA, dl, VT, N0, N1, N2);
7556 if (Options.UnsafeFPMath) {
7557 if (N0CFP && N0CFP->isZero())
7559 if (N1CFP && N1CFP->isZero())
7562 if (N0CFP && N0CFP->isExactlyValue(1.0))
7563 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2);
7564 if (N1CFP && N1CFP->isExactlyValue(1.0))
7565 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2);
7567 // Canonicalize (fma c, x, y) -> (fma x, c, y)
7568 if (N0CFP && !N1CFP)
7569 return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2);
7571 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
7572 if (Options.UnsafeFPMath && N1CFP &&
7573 N2.getOpcode() == ISD::FMUL &&
7574 N0 == N2.getOperand(0) &&
7575 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
7576 return DAG.getNode(ISD::FMUL, dl, VT, N0,
7577 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
7581 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
7582 if (Options.UnsafeFPMath &&
7583 N0.getOpcode() == ISD::FMUL && N1CFP &&
7584 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
7585 return DAG.getNode(ISD::FMA, dl, VT,
7587 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
7591 // (fma x, 1, y) -> (fadd x, y)
7592 // (fma x, -1, y) -> (fadd (fneg x), y)
7594 if (N1CFP->isExactlyValue(1.0))
7595 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
7597 if (N1CFP->isExactlyValue(-1.0) &&
7598 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
7599 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
7600 AddToWorklist(RHSNeg.getNode());
7601 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
7605 // (fma x, c, x) -> (fmul x, (c+1))
7606 if (Options.UnsafeFPMath && N1CFP && N0 == N2)
7607 return DAG.getNode(ISD::FMUL, dl, VT, N0,
7608 DAG.getNode(ISD::FADD, dl, VT,
7609 N1, DAG.getConstantFP(1.0, VT)));
7611 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
7612 if (Options.UnsafeFPMath && N1CFP &&
7613 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0)
7614 return DAG.getNode(ISD::FMUL, dl, VT, N0,
7615 DAG.getNode(ISD::FADD, dl, VT,
7616 N1, DAG.getConstantFP(-1.0, VT)));
7622 SDValue DAGCombiner::visitFDIV(SDNode *N) {
7623 SDValue N0 = N->getOperand(0);
7624 SDValue N1 = N->getOperand(1);
7625 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7626 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7627 EVT VT = N->getValueType(0);
7629 const TargetOptions &Options = DAG.getTarget().Options;
7632 if (VT.isVector()) {
7633 SDValue FoldedVOp = SimplifyVBinOp(N);
7634 if (FoldedVOp.getNode()) return FoldedVOp;
7637 // fold (fdiv c1, c2) -> c1/c2
7639 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1);
7641 if (Options.UnsafeFPMath) {
7642 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
7644 // Compute the reciprocal 1.0 / c2.
7645 APFloat N1APF = N1CFP->getValueAPF();
7646 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
7647 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
7648 // Only do the transform if the reciprocal is a legal fp immediate that
7649 // isn't too nasty (eg NaN, denormal, ...).
7650 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
7651 (!LegalOperations ||
7652 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
7653 // backend)... we should handle this gracefully after Legalize.
7654 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
7655 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
7656 TLI.isFPImmLegal(Recip, VT)))
7657 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0,
7658 DAG.getConstantFP(Recip, VT));
7661 // If this FDIV is part of a reciprocal square root, it may be folded
7662 // into a target-specific square root estimate instruction.
7663 if (N1.getOpcode() == ISD::FSQRT) {
7664 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0))) {
7665 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7667 } else if (N1.getOpcode() == ISD::FP_EXTEND &&
7668 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
7669 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0).getOperand(0))) {
7670 RV = DAG.getNode(ISD::FP_EXTEND, SDLoc(N1), VT, RV);
7671 AddToWorklist(RV.getNode());
7672 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7674 } else if (N1.getOpcode() == ISD::FP_ROUND &&
7675 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
7676 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0).getOperand(0))) {
7677 RV = DAG.getNode(ISD::FP_ROUND, SDLoc(N1), VT, RV, N1.getOperand(1));
7678 AddToWorklist(RV.getNode());
7679 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7681 } else if (N1.getOpcode() == ISD::FMUL) {
7682 // Look through an FMUL. Even though this won't remove the FDIV directly,
7683 // it's still worthwhile to get rid of the FSQRT if possible.
7686 if (N1.getOperand(0).getOpcode() == ISD::FSQRT) {
7687 SqrtOp = N1.getOperand(0);
7688 OtherOp = N1.getOperand(1);
7689 } else if (N1.getOperand(1).getOpcode() == ISD::FSQRT) {
7690 SqrtOp = N1.getOperand(1);
7691 OtherOp = N1.getOperand(0);
7693 if (SqrtOp.getNode()) {
7694 // We found a FSQRT, so try to make this fold:
7695 // x / (y * sqrt(z)) -> x * (rsqrt(z) / y)
7696 if (SDValue RV = BuildRsqrtEstimate(SqrtOp.getOperand(0))) {
7697 RV = DAG.getNode(ISD::FDIV, SDLoc(N1), VT, RV, OtherOp);
7698 AddToWorklist(RV.getNode());
7699 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7704 // Fold into a reciprocal estimate and multiply instead of a real divide.
7705 if (SDValue RV = BuildReciprocalEstimate(N1)) {
7706 AddToWorklist(RV.getNode());
7707 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7711 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
7712 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
7713 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
7714 // Both can be negated for free, check to see if at least one is cheaper
7716 if (LHSNeg == 2 || RHSNeg == 2)
7717 return DAG.getNode(ISD::FDIV, SDLoc(N), VT,
7718 GetNegatedExpression(N0, DAG, LegalOperations),
7719 GetNegatedExpression(N1, DAG, LegalOperations));
7723 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
7725 // E.g., (a / D; b / D;) -> (recip = 1.0 / D; a * recip; b * recip)
7726 // Notice that this is not always beneficial. One reason is different target
7727 // may have different costs for FDIV and FMUL, so sometimes the cost of two
7728 // FDIVs may be lower than the cost of one FDIV and two FMULs. Another reason
7729 // is the critical path is increased from "one FDIV" to "one FDIV + one FMUL".
7730 if (Options.UnsafeFPMath) {
7731 // Skip if current node is a reciprocal.
7732 if (N0CFP && N0CFP->isExactlyValue(1.0))
7735 SmallVector<SDNode *, 4> Users;
7736 // Find all FDIV users of the same divisor.
7737 for (SDNode::use_iterator UI = N1.getNode()->use_begin(),
7738 UE = N1.getNode()->use_end();
7740 SDNode *User = UI.getUse().getUser();
7741 if (User->getOpcode() == ISD::FDIV && User->getOperand(1) == N1)
7742 Users.push_back(User);
7745 if (TLI.combineRepeatedFPDivisors(Users.size())) {
7746 SDValue FPOne = DAG.getConstantFP(1.0, VT); // floating point 1.0
7747 SDValue Reciprocal = DAG.getNode(ISD::FDIV, SDLoc(N), VT, FPOne, N1);
7749 // Dividend / Divisor -> Dividend * Reciprocal
7750 for (auto I = Users.begin(), E = Users.end(); I != E; ++I) {
7751 if ((*I)->getOperand(0) != FPOne) {
7752 SDValue NewNode = DAG.getNode(ISD::FMUL, SDLoc(*I), VT,
7753 (*I)->getOperand(0), Reciprocal);
7754 DAG.ReplaceAllUsesWith(*I, NewNode.getNode());
7764 SDValue DAGCombiner::visitFREM(SDNode *N) {
7765 SDValue N0 = N->getOperand(0);
7766 SDValue N1 = N->getOperand(1);
7767 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7768 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7769 EVT VT = N->getValueType(0);
7771 // fold (frem c1, c2) -> fmod(c1,c2)
7773 return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1);
7778 SDValue DAGCombiner::visitFSQRT(SDNode *N) {
7779 if (DAG.getTarget().Options.UnsafeFPMath &&
7780 !TLI.isFsqrtCheap()) {
7781 // Compute this as X * (1/sqrt(X)) = X * (X ** -0.5)
7782 if (SDValue RV = BuildRsqrtEstimate(N->getOperand(0))) {
7783 EVT VT = RV.getValueType();
7784 RV = DAG.getNode(ISD::FMUL, SDLoc(N), VT, N->getOperand(0), RV);
7785 AddToWorklist(RV.getNode());
7787 // Unfortunately, RV is now NaN if the input was exactly 0.
7788 // Select out this case and force the answer to 0.
7789 SDValue Zero = DAG.getConstantFP(0.0, VT);
7791 DAG.getSetCC(SDLoc(N), TLI.getSetCCResultType(*DAG.getContext(), VT),
7792 N->getOperand(0), Zero, ISD::SETEQ);
7793 AddToWorklist(ZeroCmp.getNode());
7794 AddToWorklist(RV.getNode());
7796 RV = DAG.getNode(VT.isVector() ? ISD::VSELECT : ISD::SELECT,
7797 SDLoc(N), VT, ZeroCmp, Zero, RV);
7804 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
7805 SDValue N0 = N->getOperand(0);
7806 SDValue N1 = N->getOperand(1);
7807 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7808 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7809 EVT VT = N->getValueType(0);
7811 if (N0CFP && N1CFP) // Constant fold
7812 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1);
7815 const APFloat& V = N1CFP->getValueAPF();
7816 // copysign(x, c1) -> fabs(x) iff ispos(c1)
7817 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
7818 if (!V.isNegative()) {
7819 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
7820 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7822 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7823 return DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7824 DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0));
7828 // copysign(fabs(x), y) -> copysign(x, y)
7829 // copysign(fneg(x), y) -> copysign(x, y)
7830 // copysign(copysign(x,z), y) -> copysign(x, y)
7831 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
7832 N0.getOpcode() == ISD::FCOPYSIGN)
7833 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7834 N0.getOperand(0), N1);
7836 // copysign(x, abs(y)) -> abs(x)
7837 if (N1.getOpcode() == ISD::FABS)
7838 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7840 // copysign(x, copysign(y,z)) -> copysign(x, z)
7841 if (N1.getOpcode() == ISD::FCOPYSIGN)
7842 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7843 N0, N1.getOperand(1));
7845 // copysign(x, fp_extend(y)) -> copysign(x, y)
7846 // copysign(x, fp_round(y)) -> copysign(x, y)
7847 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
7848 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7849 N0, N1.getOperand(0));
7854 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
7855 SDValue N0 = N->getOperand(0);
7856 EVT VT = N->getValueType(0);
7857 EVT OpVT = N0.getValueType();
7859 // fold (sint_to_fp c1) -> c1fp
7860 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
7862 // ...but only if the target supports immediate floating-point values
7863 (!LegalOperations ||
7864 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
7865 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
7867 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
7868 // but UINT_TO_FP is legal on this target, try to convert.
7869 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
7870 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
7871 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
7872 if (DAG.SignBitIsZero(N0))
7873 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
7876 // The next optimizations are desirable only if SELECT_CC can be lowered.
7877 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
7878 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7879 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
7881 (!LegalOperations ||
7882 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7884 { N0.getOperand(0), N0.getOperand(1),
7885 DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT),
7887 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7890 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
7891 // (select_cc x, y, 1.0, 0.0,, cc)
7892 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
7893 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
7894 (!LegalOperations ||
7895 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7897 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
7898 DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT),
7899 N0.getOperand(0).getOperand(2) };
7900 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7907 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
7908 SDValue N0 = N->getOperand(0);
7909 EVT VT = N->getValueType(0);
7910 EVT OpVT = N0.getValueType();
7912 // fold (uint_to_fp c1) -> c1fp
7913 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
7915 // ...but only if the target supports immediate floating-point values
7916 (!LegalOperations ||
7917 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
7918 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
7920 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
7921 // but SINT_TO_FP is legal on this target, try to convert.
7922 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
7923 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
7924 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
7925 if (DAG.SignBitIsZero(N0))
7926 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
7929 // The next optimizations are desirable only if SELECT_CC can be lowered.
7930 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
7931 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7933 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
7934 (!LegalOperations ||
7935 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7937 { N0.getOperand(0), N0.getOperand(1),
7938 DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT),
7940 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7947 // Fold (fp_to_{s/u}int ({s/u}int_to_fpx)) -> zext x, sext x, trunc x, or x
7948 static SDValue FoldIntToFPToInt(SDNode *N, SelectionDAG &DAG) {
7949 SDValue N0 = N->getOperand(0);
7950 EVT VT = N->getValueType(0);
7952 if (N0.getOpcode() != ISD::UINT_TO_FP && N0.getOpcode() != ISD::SINT_TO_FP)
7955 SDValue Src = N0.getOperand(0);
7956 EVT SrcVT = Src.getValueType();
7957 bool IsInputSigned = N0.getOpcode() == ISD::SINT_TO_FP;
7958 bool IsOutputSigned = N->getOpcode() == ISD::FP_TO_SINT;
7960 // We can safely assume the conversion won't overflow the output range,
7961 // because (for example) (uint8_t)18293.f is undefined behavior.
7963 // Since we can assume the conversion won't overflow, our decision as to
7964 // whether the input will fit in the float should depend on the minimum
7965 // of the input range and output range.
7967 // This means this is also safe for a signed input and unsigned output, since
7968 // a negative input would lead to undefined behavior.
7969 unsigned InputSize = (int)SrcVT.getScalarSizeInBits() - IsInputSigned;
7970 unsigned OutputSize = (int)VT.getScalarSizeInBits() - IsOutputSigned;
7971 unsigned ActualSize = std::min(InputSize, OutputSize);
7972 const fltSemantics &sem = DAG.EVTToAPFloatSemantics(N0.getValueType());
7974 // We can only fold away the float conversion if the input range can be
7975 // represented exactly in the float range.
7976 if (APFloat::semanticsPrecision(sem) >= ActualSize) {
7977 if (VT.getScalarSizeInBits() > SrcVT.getScalarSizeInBits()) {
7978 unsigned ExtOp = IsInputSigned && IsOutputSigned ? ISD::SIGN_EXTEND
7980 return DAG.getNode(ExtOp, SDLoc(N), VT, Src);
7982 if (VT.getScalarSizeInBits() < SrcVT.getScalarSizeInBits())
7983 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Src);
7986 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Src);
7991 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
7992 SDValue N0 = N->getOperand(0);
7993 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7994 EVT VT = N->getValueType(0);
7996 // fold (fp_to_sint c1fp) -> c1
7998 return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0);
8000 return FoldIntToFPToInt(N, DAG);
8003 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
8004 SDValue N0 = N->getOperand(0);
8005 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8006 EVT VT = N->getValueType(0);
8008 // fold (fp_to_uint c1fp) -> c1
8010 return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0);
8012 return FoldIntToFPToInt(N, DAG);
8015 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
8016 SDValue N0 = N->getOperand(0);
8017 SDValue N1 = N->getOperand(1);
8018 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8019 EVT VT = N->getValueType(0);
8021 // fold (fp_round c1fp) -> c1fp
8023 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1);
8025 // fold (fp_round (fp_extend x)) -> x
8026 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
8027 return N0.getOperand(0);
8029 // fold (fp_round (fp_round x)) -> (fp_round x)
8030 if (N0.getOpcode() == ISD::FP_ROUND) {
8031 const bool NIsTrunc = N->getConstantOperandVal(1) == 1;
8032 const bool N0IsTrunc = N0.getNode()->getConstantOperandVal(1) == 1;
8033 // If the first fp_round isn't a value preserving truncation, it might
8034 // introduce a tie in the second fp_round, that wouldn't occur in the
8035 // single-step fp_round we want to fold to.
8036 // In other words, double rounding isn't the same as rounding.
8037 // Also, this is a value preserving truncation iff both fp_round's are.
8038 if (DAG.getTarget().Options.UnsafeFPMath || N0IsTrunc)
8039 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0.getOperand(0),
8040 DAG.getIntPtrConstant(NIsTrunc && N0IsTrunc));
8043 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
8044 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
8045 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT,
8046 N0.getOperand(0), N1);
8047 AddToWorklist(Tmp.getNode());
8048 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
8049 Tmp, N0.getOperand(1));
8055 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
8056 SDValue N0 = N->getOperand(0);
8057 EVT VT = N->getValueType(0);
8058 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
8059 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8061 // fold (fp_round_inreg c1fp) -> c1fp
8062 if (N0CFP && isTypeLegal(EVT)) {
8063 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
8064 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, Round);
8070 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
8071 SDValue N0 = N->getOperand(0);
8072 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8073 EVT VT = N->getValueType(0);
8075 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
8076 if (N->hasOneUse() &&
8077 N->use_begin()->getOpcode() == ISD::FP_ROUND)
8080 // fold (fp_extend c1fp) -> c1fp
8082 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0);
8084 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
8086 if (N0.getOpcode() == ISD::FP_ROUND
8087 && N0.getNode()->getConstantOperandVal(1) == 1) {
8088 SDValue In = N0.getOperand(0);
8089 if (In.getValueType() == VT) return In;
8090 if (VT.bitsLT(In.getValueType()))
8091 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT,
8092 In, N0.getOperand(1));
8093 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In);
8096 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
8097 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
8098 TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) {
8099 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
8100 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
8102 LN0->getBasePtr(), N0.getValueType(),
8103 LN0->getMemOperand());
8104 CombineTo(N, ExtLoad);
8105 CombineTo(N0.getNode(),
8106 DAG.getNode(ISD::FP_ROUND, SDLoc(N0),
8107 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
8108 ExtLoad.getValue(1));
8109 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8115 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
8116 SDValue N0 = N->getOperand(0);
8117 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8118 EVT VT = N->getValueType(0);
8120 // fold (fceil c1) -> fceil(c1)
8122 return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0);
8127 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
8128 SDValue N0 = N->getOperand(0);
8129 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8130 EVT VT = N->getValueType(0);
8132 // fold (ftrunc c1) -> ftrunc(c1)
8134 return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0);
8139 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
8140 SDValue N0 = N->getOperand(0);
8141 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8142 EVT VT = N->getValueType(0);
8144 // fold (ffloor c1) -> ffloor(c1)
8146 return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0);
8151 // FIXME: FNEG and FABS have a lot in common; refactor.
8152 SDValue DAGCombiner::visitFNEG(SDNode *N) {
8153 SDValue N0 = N->getOperand(0);
8154 EVT VT = N->getValueType(0);
8156 if (VT.isVector()) {
8157 SDValue FoldedVOp = SimplifyVUnaryOp(N);
8158 if (FoldedVOp.getNode()) return FoldedVOp;
8161 // Constant fold FNEG.
8162 if (isa<ConstantFPSDNode>(N0))
8163 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N->getOperand(0));
8165 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
8166 &DAG.getTarget().Options))
8167 return GetNegatedExpression(N0, DAG, LegalOperations);
8169 // Transform fneg(bitconvert(x)) -> bitconvert(x ^ sign) to avoid loading
8170 // constant pool values.
8171 if (!TLI.isFNegFree(VT) &&
8172 N0.getOpcode() == ISD::BITCAST &&
8173 N0.getNode()->hasOneUse()) {
8174 SDValue Int = N0.getOperand(0);
8175 EVT IntVT = Int.getValueType();
8176 if (IntVT.isInteger() && !IntVT.isVector()) {
8178 if (N0.getValueType().isVector()) {
8179 // For a vector, get a mask such as 0x80... per scalar element
8181 SignMask = APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
8182 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
8184 // For a scalar, just generate 0x80...
8185 SignMask = APInt::getSignBit(IntVT.getSizeInBits());
8187 Int = DAG.getNode(ISD::XOR, SDLoc(N0), IntVT, Int,
8188 DAG.getConstant(SignMask, IntVT));
8189 AddToWorklist(Int.getNode());
8190 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Int);
8194 // (fneg (fmul c, x)) -> (fmul -c, x)
8195 if (N0.getOpcode() == ISD::FMUL) {
8196 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
8198 APFloat CVal = CFP1->getValueAPF();
8200 if (Level >= AfterLegalizeDAG &&
8201 (TLI.isFPImmLegal(CVal, N->getValueType(0)) ||
8202 TLI.isOperationLegal(ISD::ConstantFP, N->getValueType(0))))
8204 ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
8205 DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0.getOperand(1)));
8212 SDValue DAGCombiner::visitFMINNUM(SDNode *N) {
8213 SDValue N0 = N->getOperand(0);
8214 SDValue N1 = N->getOperand(1);
8215 const ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8216 const ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8218 if (N0CFP && N1CFP) {
8219 const APFloat &C0 = N0CFP->getValueAPF();
8220 const APFloat &C1 = N1CFP->getValueAPF();
8221 return DAG.getConstantFP(minnum(C0, C1), N->getValueType(0));
8225 EVT VT = N->getValueType(0);
8226 // Canonicalize to constant on RHS.
8227 return DAG.getNode(ISD::FMINNUM, SDLoc(N), VT, N1, N0);
8233 SDValue DAGCombiner::visitFMAXNUM(SDNode *N) {
8234 SDValue N0 = N->getOperand(0);
8235 SDValue N1 = N->getOperand(1);
8236 const ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8237 const ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8239 if (N0CFP && N1CFP) {
8240 const APFloat &C0 = N0CFP->getValueAPF();
8241 const APFloat &C1 = N1CFP->getValueAPF();
8242 return DAG.getConstantFP(maxnum(C0, C1), N->getValueType(0));
8246 EVT VT = N->getValueType(0);
8247 // Canonicalize to constant on RHS.
8248 return DAG.getNode(ISD::FMAXNUM, SDLoc(N), VT, N1, N0);
8254 SDValue DAGCombiner::visitFABS(SDNode *N) {
8255 SDValue N0 = N->getOperand(0);
8256 EVT VT = N->getValueType(0);
8258 if (VT.isVector()) {
8259 SDValue FoldedVOp = SimplifyVUnaryOp(N);
8260 if (FoldedVOp.getNode()) return FoldedVOp;
8263 // fold (fabs c1) -> fabs(c1)
8264 if (isa<ConstantFPSDNode>(N0))
8265 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
8267 // fold (fabs (fabs x)) -> (fabs x)
8268 if (N0.getOpcode() == ISD::FABS)
8269 return N->getOperand(0);
8271 // fold (fabs (fneg x)) -> (fabs x)
8272 // fold (fabs (fcopysign x, y)) -> (fabs x)
8273 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
8274 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0));
8276 // Transform fabs(bitconvert(x)) -> bitconvert(x & ~sign) to avoid loading
8277 // constant pool values.
8278 if (!TLI.isFAbsFree(VT) &&
8279 N0.getOpcode() == ISD::BITCAST &&
8280 N0.getNode()->hasOneUse()) {
8281 SDValue Int = N0.getOperand(0);
8282 EVT IntVT = Int.getValueType();
8283 if (IntVT.isInteger() && !IntVT.isVector()) {
8285 if (N0.getValueType().isVector()) {
8286 // For a vector, get a mask such as 0x7f... per scalar element
8288 SignMask = ~APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
8289 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
8291 // For a scalar, just generate 0x7f...
8292 SignMask = ~APInt::getSignBit(IntVT.getSizeInBits());
8294 Int = DAG.getNode(ISD::AND, SDLoc(N0), IntVT, Int,
8295 DAG.getConstant(SignMask, IntVT));
8296 AddToWorklist(Int.getNode());
8297 return DAG.getNode(ISD::BITCAST, SDLoc(N), N->getValueType(0), Int);
8304 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
8305 SDValue Chain = N->getOperand(0);
8306 SDValue N1 = N->getOperand(1);
8307 SDValue N2 = N->getOperand(2);
8309 // If N is a constant we could fold this into a fallthrough or unconditional
8310 // branch. However that doesn't happen very often in normal code, because
8311 // Instcombine/SimplifyCFG should have handled the available opportunities.
8312 // If we did this folding here, it would be necessary to update the
8313 // MachineBasicBlock CFG, which is awkward.
8315 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
8317 if (N1.getOpcode() == ISD::SETCC &&
8318 TLI.isOperationLegalOrCustom(ISD::BR_CC,
8319 N1.getOperand(0).getValueType())) {
8320 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
8321 Chain, N1.getOperand(2),
8322 N1.getOperand(0), N1.getOperand(1), N2);
8325 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
8326 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
8327 (N1.getOperand(0).hasOneUse() &&
8328 N1.getOperand(0).getOpcode() == ISD::SRL))) {
8329 SDNode *Trunc = nullptr;
8330 if (N1.getOpcode() == ISD::TRUNCATE) {
8331 // Look pass the truncate.
8332 Trunc = N1.getNode();
8333 N1 = N1.getOperand(0);
8336 // Match this pattern so that we can generate simpler code:
8339 // %b = and i32 %a, 2
8340 // %c = srl i32 %b, 1
8341 // brcond i32 %c ...
8346 // %b = and i32 %a, 2
8347 // %c = setcc eq %b, 0
8350 // This applies only when the AND constant value has one bit set and the
8351 // SRL constant is equal to the log2 of the AND constant. The back-end is
8352 // smart enough to convert the result into a TEST/JMP sequence.
8353 SDValue Op0 = N1.getOperand(0);
8354 SDValue Op1 = N1.getOperand(1);
8356 if (Op0.getOpcode() == ISD::AND &&
8357 Op1.getOpcode() == ISD::Constant) {
8358 SDValue AndOp1 = Op0.getOperand(1);
8360 if (AndOp1.getOpcode() == ISD::Constant) {
8361 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
8363 if (AndConst.isPowerOf2() &&
8364 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
8366 DAG.getSetCC(SDLoc(N),
8367 getSetCCResultType(Op0.getValueType()),
8368 Op0, DAG.getConstant(0, Op0.getValueType()),
8371 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, SDLoc(N),
8372 MVT::Other, Chain, SetCC, N2);
8373 // Don't add the new BRCond into the worklist or else SimplifySelectCC
8374 // will convert it back to (X & C1) >> C2.
8375 CombineTo(N, NewBRCond, false);
8376 // Truncate is dead.
8378 deleteAndRecombine(Trunc);
8379 // Replace the uses of SRL with SETCC
8380 WorklistRemover DeadNodes(*this);
8381 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
8382 deleteAndRecombine(N1.getNode());
8383 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8389 // Restore N1 if the above transformation doesn't match.
8390 N1 = N->getOperand(1);
8393 // Transform br(xor(x, y)) -> br(x != y)
8394 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
8395 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
8396 SDNode *TheXor = N1.getNode();
8397 SDValue Op0 = TheXor->getOperand(0);
8398 SDValue Op1 = TheXor->getOperand(1);
8399 if (Op0.getOpcode() == Op1.getOpcode()) {
8400 // Avoid missing important xor optimizations.
8401 SDValue Tmp = visitXOR(TheXor);
8402 if (Tmp.getNode()) {
8403 if (Tmp.getNode() != TheXor) {
8404 DEBUG(dbgs() << "\nReplacing.8 ";
8406 dbgs() << "\nWith: ";
8407 Tmp.getNode()->dump(&DAG);
8409 WorklistRemover DeadNodes(*this);
8410 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
8411 deleteAndRecombine(TheXor);
8412 return DAG.getNode(ISD::BRCOND, SDLoc(N),
8413 MVT::Other, Chain, Tmp, N2);
8416 // visitXOR has changed XOR's operands or replaced the XOR completely,
8418 return SDValue(N, 0);
8422 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
8424 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
8425 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
8426 Op0.getOpcode() == ISD::XOR) {
8427 TheXor = Op0.getNode();
8431 EVT SetCCVT = N1.getValueType();
8433 SetCCVT = getSetCCResultType(SetCCVT);
8434 SDValue SetCC = DAG.getSetCC(SDLoc(TheXor),
8437 Equal ? ISD::SETEQ : ISD::SETNE);
8438 // Replace the uses of XOR with SETCC
8439 WorklistRemover DeadNodes(*this);
8440 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
8441 deleteAndRecombine(N1.getNode());
8442 return DAG.getNode(ISD::BRCOND, SDLoc(N),
8443 MVT::Other, Chain, SetCC, N2);
8450 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
8452 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
8453 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
8454 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
8456 // If N is a constant we could fold this into a fallthrough or unconditional
8457 // branch. However that doesn't happen very often in normal code, because
8458 // Instcombine/SimplifyCFG should have handled the available opportunities.
8459 // If we did this folding here, it would be necessary to update the
8460 // MachineBasicBlock CFG, which is awkward.
8462 // Use SimplifySetCC to simplify SETCC's.
8463 SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()),
8464 CondLHS, CondRHS, CC->get(), SDLoc(N),
8466 if (Simp.getNode()) AddToWorklist(Simp.getNode());
8468 // fold to a simpler setcc
8469 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
8470 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
8471 N->getOperand(0), Simp.getOperand(2),
8472 Simp.getOperand(0), Simp.getOperand(1),
8478 /// Return true if 'Use' is a load or a store that uses N as its base pointer
8479 /// and that N may be folded in the load / store addressing mode.
8480 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
8482 const TargetLowering &TLI) {
8484 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
8485 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
8487 VT = Use->getValueType(0);
8488 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
8489 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
8491 VT = ST->getValue().getValueType();
8495 TargetLowering::AddrMode AM;
8496 if (N->getOpcode() == ISD::ADD) {
8497 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
8500 AM.BaseOffs = Offset->getSExtValue();
8504 } else if (N->getOpcode() == ISD::SUB) {
8505 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
8508 AM.BaseOffs = -Offset->getSExtValue();
8515 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
8518 /// Try turning a load/store into a pre-indexed load/store when the base
8519 /// pointer is an add or subtract and it has other uses besides the load/store.
8520 /// After the transformation, the new indexed load/store has effectively folded
8521 /// the add/subtract in and all of its other uses are redirected to the
8523 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
8524 if (Level < AfterLegalizeDAG)
8530 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8531 if (LD->isIndexed())
8533 VT = LD->getMemoryVT();
8534 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
8535 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
8537 Ptr = LD->getBasePtr();
8538 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8539 if (ST->isIndexed())
8541 VT = ST->getMemoryVT();
8542 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
8543 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
8545 Ptr = ST->getBasePtr();
8551 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
8552 // out. There is no reason to make this a preinc/predec.
8553 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
8554 Ptr.getNode()->hasOneUse())
8557 // Ask the target to do addressing mode selection.
8560 ISD::MemIndexedMode AM = ISD::UNINDEXED;
8561 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
8564 // Backends without true r+i pre-indexed forms may need to pass a
8565 // constant base with a variable offset so that constant coercion
8566 // will work with the patterns in canonical form.
8567 bool Swapped = false;
8568 if (isa<ConstantSDNode>(BasePtr)) {
8569 std::swap(BasePtr, Offset);
8573 // Don't create a indexed load / store with zero offset.
8574 if (isa<ConstantSDNode>(Offset) &&
8575 cast<ConstantSDNode>(Offset)->isNullValue())
8578 // Try turning it into a pre-indexed load / store except when:
8579 // 1) The new base ptr is a frame index.
8580 // 2) If N is a store and the new base ptr is either the same as or is a
8581 // predecessor of the value being stored.
8582 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
8583 // that would create a cycle.
8584 // 4) All uses are load / store ops that use it as old base ptr.
8586 // Check #1. Preinc'ing a frame index would require copying the stack pointer
8587 // (plus the implicit offset) to a register to preinc anyway.
8588 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
8593 SDValue Val = cast<StoreSDNode>(N)->getValue();
8594 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
8598 // If the offset is a constant, there may be other adds of constants that
8599 // can be folded with this one. We should do this to avoid having to keep
8600 // a copy of the original base pointer.
8601 SmallVector<SDNode *, 16> OtherUses;
8602 if (isa<ConstantSDNode>(Offset))
8603 for (SDNode *Use : BasePtr.getNode()->uses()) {
8604 if (Use == Ptr.getNode())
8607 if (Use->isPredecessorOf(N))
8610 if (Use->getOpcode() != ISD::ADD && Use->getOpcode() != ISD::SUB) {
8615 SDValue Op0 = Use->getOperand(0), Op1 = Use->getOperand(1);
8616 if (Op1.getNode() == BasePtr.getNode())
8617 std::swap(Op0, Op1);
8618 assert(Op0.getNode() == BasePtr.getNode() &&
8619 "Use of ADD/SUB but not an operand");
8621 if (!isa<ConstantSDNode>(Op1)) {
8626 // FIXME: In some cases, we can be smarter about this.
8627 if (Op1.getValueType() != Offset.getValueType()) {
8632 OtherUses.push_back(Use);
8636 std::swap(BasePtr, Offset);
8638 // Now check for #3 and #4.
8639 bool RealUse = false;
8641 // Caches for hasPredecessorHelper
8642 SmallPtrSet<const SDNode *, 32> Visited;
8643 SmallVector<const SDNode *, 16> Worklist;
8645 for (SDNode *Use : Ptr.getNode()->uses()) {
8648 if (N->hasPredecessorHelper(Use, Visited, Worklist))
8651 // If Ptr may be folded in addressing mode of other use, then it's
8652 // not profitable to do this transformation.
8653 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
8662 Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
8663 BasePtr, Offset, AM);
8665 Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
8666 BasePtr, Offset, AM);
8669 DEBUG(dbgs() << "\nReplacing.4 ";
8671 dbgs() << "\nWith: ";
8672 Result.getNode()->dump(&DAG);
8674 WorklistRemover DeadNodes(*this);
8676 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
8677 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
8679 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
8682 // Finally, since the node is now dead, remove it from the graph.
8683 deleteAndRecombine(N);
8686 std::swap(BasePtr, Offset);
8688 // Replace other uses of BasePtr that can be updated to use Ptr
8689 for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) {
8690 unsigned OffsetIdx = 1;
8691 if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode())
8693 assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() ==
8694 BasePtr.getNode() && "Expected BasePtr operand");
8696 // We need to replace ptr0 in the following expression:
8697 // x0 * offset0 + y0 * ptr0 = t0
8699 // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store)
8701 // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the
8702 // indexed load/store and the expresion that needs to be re-written.
8704 // Therefore, we have:
8705 // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1
8707 ConstantSDNode *CN =
8708 cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx));
8710 APInt Offset0 = CN->getAPIntValue();
8711 APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue();
8713 X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1;
8714 Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1;
8715 X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1;
8716 Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1;
8718 unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD;
8720 APInt CNV = Offset0;
8721 if (X0 < 0) CNV = -CNV;
8722 if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1;
8723 else CNV = CNV - Offset1;
8725 // We can now generate the new expression.
8726 SDValue NewOp1 = DAG.getConstant(CNV, CN->getValueType(0));
8727 SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0);
8729 SDValue NewUse = DAG.getNode(Opcode,
8730 SDLoc(OtherUses[i]),
8731 OtherUses[i]->getValueType(0), NewOp1, NewOp2);
8732 DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse);
8733 deleteAndRecombine(OtherUses[i]);
8736 // Replace the uses of Ptr with uses of the updated base value.
8737 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
8738 deleteAndRecombine(Ptr.getNode());
8743 /// Try to combine a load/store with a add/sub of the base pointer node into a
8744 /// post-indexed load/store. The transformation folded the add/subtract into the
8745 /// new indexed load/store effectively and all of its uses are redirected to the
8747 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
8748 if (Level < AfterLegalizeDAG)
8754 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8755 if (LD->isIndexed())
8757 VT = LD->getMemoryVT();
8758 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
8759 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
8761 Ptr = LD->getBasePtr();
8762 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8763 if (ST->isIndexed())
8765 VT = ST->getMemoryVT();
8766 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
8767 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
8769 Ptr = ST->getBasePtr();
8775 if (Ptr.getNode()->hasOneUse())
8778 for (SDNode *Op : Ptr.getNode()->uses()) {
8780 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
8785 ISD::MemIndexedMode AM = ISD::UNINDEXED;
8786 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
8787 // Don't create a indexed load / store with zero offset.
8788 if (isa<ConstantSDNode>(Offset) &&
8789 cast<ConstantSDNode>(Offset)->isNullValue())
8792 // Try turning it into a post-indexed load / store except when
8793 // 1) All uses are load / store ops that use it as base ptr (and
8794 // it may be folded as addressing mmode).
8795 // 2) Op must be independent of N, i.e. Op is neither a predecessor
8796 // nor a successor of N. Otherwise, if Op is folded that would
8799 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
8803 bool TryNext = false;
8804 for (SDNode *Use : BasePtr.getNode()->uses()) {
8805 if (Use == Ptr.getNode())
8808 // If all the uses are load / store addresses, then don't do the
8810 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
8811 bool RealUse = false;
8812 for (SDNode *UseUse : Use->uses()) {
8813 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
8828 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
8829 SDValue Result = isLoad
8830 ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
8831 BasePtr, Offset, AM)
8832 : DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
8833 BasePtr, Offset, AM);
8836 DEBUG(dbgs() << "\nReplacing.5 ";
8838 dbgs() << "\nWith: ";
8839 Result.getNode()->dump(&DAG);
8841 WorklistRemover DeadNodes(*this);
8843 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
8844 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
8846 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
8849 // Finally, since the node is now dead, remove it from the graph.
8850 deleteAndRecombine(N);
8852 // Replace the uses of Use with uses of the updated base value.
8853 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
8854 Result.getValue(isLoad ? 1 : 0));
8855 deleteAndRecombine(Op);
8864 /// \brief Return the base-pointer arithmetic from an indexed \p LD.
8865 SDValue DAGCombiner::SplitIndexingFromLoad(LoadSDNode *LD) {
8866 ISD::MemIndexedMode AM = LD->getAddressingMode();
8867 assert(AM != ISD::UNINDEXED);
8868 SDValue BP = LD->getOperand(1);
8869 SDValue Inc = LD->getOperand(2);
8871 // Some backends use TargetConstants for load offsets, but don't expect
8872 // TargetConstants in general ADD nodes. We can convert these constants into
8873 // regular Constants (if the constant is not opaque).
8874 assert((Inc.getOpcode() != ISD::TargetConstant ||
8875 !cast<ConstantSDNode>(Inc)->isOpaque()) &&
8876 "Cannot split out indexing using opaque target constants");
8877 if (Inc.getOpcode() == ISD::TargetConstant) {
8878 ConstantSDNode *ConstInc = cast<ConstantSDNode>(Inc);
8879 Inc = DAG.getConstant(*ConstInc->getConstantIntValue(),
8880 ConstInc->getValueType(0));
8884 (AM == ISD::PRE_INC || AM == ISD::POST_INC ? ISD::ADD : ISD::SUB);
8885 return DAG.getNode(Opc, SDLoc(LD), BP.getSimpleValueType(), BP, Inc);
8888 SDValue DAGCombiner::visitLOAD(SDNode *N) {
8889 LoadSDNode *LD = cast<LoadSDNode>(N);
8890 SDValue Chain = LD->getChain();
8891 SDValue Ptr = LD->getBasePtr();
8893 // If load is not volatile and there are no uses of the loaded value (and
8894 // the updated indexed value in case of indexed loads), change uses of the
8895 // chain value into uses of the chain input (i.e. delete the dead load).
8896 if (!LD->isVolatile()) {
8897 if (N->getValueType(1) == MVT::Other) {
8899 if (!N->hasAnyUseOfValue(0)) {
8900 // It's not safe to use the two value CombineTo variant here. e.g.
8901 // v1, chain2 = load chain1, loc
8902 // v2, chain3 = load chain2, loc
8904 // Now we replace use of chain2 with chain1. This makes the second load
8905 // isomorphic to the one we are deleting, and thus makes this load live.
8906 DEBUG(dbgs() << "\nReplacing.6 ";
8908 dbgs() << "\nWith chain: ";
8909 Chain.getNode()->dump(&DAG);
8911 WorklistRemover DeadNodes(*this);
8912 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
8915 deleteAndRecombine(N);
8917 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8921 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
8923 // If this load has an opaque TargetConstant offset, then we cannot split
8924 // the indexing into an add/sub directly (that TargetConstant may not be
8925 // valid for a different type of node, and we cannot convert an opaque
8926 // target constant into a regular constant).
8927 bool HasOTCInc = LD->getOperand(2).getOpcode() == ISD::TargetConstant &&
8928 cast<ConstantSDNode>(LD->getOperand(2))->isOpaque();
8930 if (!N->hasAnyUseOfValue(0) &&
8931 ((MaySplitLoadIndex && !HasOTCInc) || !N->hasAnyUseOfValue(1))) {
8932 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
8934 if (N->hasAnyUseOfValue(1) && MaySplitLoadIndex && !HasOTCInc) {
8935 Index = SplitIndexingFromLoad(LD);
8936 // Try to fold the base pointer arithmetic into subsequent loads and
8938 AddUsersToWorklist(N);
8940 Index = DAG.getUNDEF(N->getValueType(1));
8941 DEBUG(dbgs() << "\nReplacing.7 ";
8943 dbgs() << "\nWith: ";
8944 Undef.getNode()->dump(&DAG);
8945 dbgs() << " and 2 other values\n");
8946 WorklistRemover DeadNodes(*this);
8947 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
8948 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Index);
8949 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
8950 deleteAndRecombine(N);
8951 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8956 // If this load is directly stored, replace the load value with the stored
8958 // TODO: Handle store large -> read small portion.
8959 // TODO: Handle TRUNCSTORE/LOADEXT
8960 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
8961 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
8962 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
8963 if (PrevST->getBasePtr() == Ptr &&
8964 PrevST->getValue().getValueType() == N->getValueType(0))
8965 return CombineTo(N, Chain.getOperand(1), Chain);
8969 // Try to infer better alignment information than the load already has.
8970 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
8971 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
8972 if (Align > LD->getMemOperand()->getBaseAlignment()) {
8974 DAG.getExtLoad(LD->getExtensionType(), SDLoc(N),
8975 LD->getValueType(0),
8976 Chain, Ptr, LD->getPointerInfo(),
8978 LD->isVolatile(), LD->isNonTemporal(),
8979 LD->isInvariant(), Align, LD->getAAInfo());
8980 return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true);
8985 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
8986 : DAG.getSubtarget().useAA();
8988 if (CombinerAAOnlyFunc.getNumOccurrences() &&
8989 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
8992 if (UseAA && LD->isUnindexed()) {
8993 // Walk up chain skipping non-aliasing memory nodes.
8994 SDValue BetterChain = FindBetterChain(N, Chain);
8996 // If there is a better chain.
8997 if (Chain != BetterChain) {
9000 // Replace the chain to void dependency.
9001 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
9002 ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD),
9003 BetterChain, Ptr, LD->getMemOperand());
9005 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD),
9006 LD->getValueType(0),
9007 BetterChain, Ptr, LD->getMemoryVT(),
9008 LD->getMemOperand());
9011 // Create token factor to keep old chain connected.
9012 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
9013 MVT::Other, Chain, ReplLoad.getValue(1));
9015 // Make sure the new and old chains are cleaned up.
9016 AddToWorklist(Token.getNode());
9018 // Replace uses with load result and token factor. Don't add users
9020 return CombineTo(N, ReplLoad.getValue(0), Token, false);
9024 // Try transforming N to an indexed load.
9025 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
9026 return SDValue(N, 0);
9028 // Try to slice up N to more direct loads if the slices are mapped to
9029 // different register banks or pairing can take place.
9031 return SDValue(N, 0);
9037 /// \brief Helper structure used to slice a load in smaller loads.
9038 /// Basically a slice is obtained from the following sequence:
9039 /// Origin = load Ty1, Base
9040 /// Shift = srl Ty1 Origin, CstTy Amount
9041 /// Inst = trunc Shift to Ty2
9043 /// Then, it will be rewriten into:
9044 /// Slice = load SliceTy, Base + SliceOffset
9045 /// [Inst = zext Slice to Ty2], only if SliceTy <> Ty2
9047 /// SliceTy is deduced from the number of bits that are actually used to
9049 struct LoadedSlice {
9050 /// \brief Helper structure used to compute the cost of a slice.
9052 /// Are we optimizing for code size.
9057 unsigned CrossRegisterBanksCopies;
9061 Cost(bool ForCodeSize = false)
9062 : ForCodeSize(ForCodeSize), Loads(0), Truncates(0),
9063 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {}
9065 /// \brief Get the cost of one isolated slice.
9066 Cost(const LoadedSlice &LS, bool ForCodeSize = false)
9067 : ForCodeSize(ForCodeSize), Loads(1), Truncates(0),
9068 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {
9069 EVT TruncType = LS.Inst->getValueType(0);
9070 EVT LoadedType = LS.getLoadedType();
9071 if (TruncType != LoadedType &&
9072 !LS.DAG->getTargetLoweringInfo().isZExtFree(LoadedType, TruncType))
9076 /// \brief Account for slicing gain in the current cost.
9077 /// Slicing provide a few gains like removing a shift or a
9078 /// truncate. This method allows to grow the cost of the original
9079 /// load with the gain from this slice.
9080 void addSliceGain(const LoadedSlice &LS) {
9081 // Each slice saves a truncate.
9082 const TargetLowering &TLI = LS.DAG->getTargetLoweringInfo();
9083 if (!TLI.isTruncateFree(LS.Inst->getValueType(0),
9084 LS.Inst->getOperand(0).getValueType()))
9086 // If there is a shift amount, this slice gets rid of it.
9089 // If this slice can merge a cross register bank copy, account for it.
9090 if (LS.canMergeExpensiveCrossRegisterBankCopy())
9091 ++CrossRegisterBanksCopies;
9094 Cost &operator+=(const Cost &RHS) {
9096 Truncates += RHS.Truncates;
9097 CrossRegisterBanksCopies += RHS.CrossRegisterBanksCopies;
9103 bool operator==(const Cost &RHS) const {
9104 return Loads == RHS.Loads && Truncates == RHS.Truncates &&
9105 CrossRegisterBanksCopies == RHS.CrossRegisterBanksCopies &&
9106 ZExts == RHS.ZExts && Shift == RHS.Shift;
9109 bool operator!=(const Cost &RHS) const { return !(*this == RHS); }
9111 bool operator<(const Cost &RHS) const {
9112 // Assume cross register banks copies are as expensive as loads.
9113 // FIXME: Do we want some more target hooks?
9114 unsigned ExpensiveOpsLHS = Loads + CrossRegisterBanksCopies;
9115 unsigned ExpensiveOpsRHS = RHS.Loads + RHS.CrossRegisterBanksCopies;
9116 // Unless we are optimizing for code size, consider the
9117 // expensive operation first.
9118 if (!ForCodeSize && ExpensiveOpsLHS != ExpensiveOpsRHS)
9119 return ExpensiveOpsLHS < ExpensiveOpsRHS;
9120 return (Truncates + ZExts + Shift + ExpensiveOpsLHS) <
9121 (RHS.Truncates + RHS.ZExts + RHS.Shift + ExpensiveOpsRHS);
9124 bool operator>(const Cost &RHS) const { return RHS < *this; }
9126 bool operator<=(const Cost &RHS) const { return !(RHS < *this); }
9128 bool operator>=(const Cost &RHS) const { return !(*this < RHS); }
9130 // The last instruction that represent the slice. This should be a
9131 // truncate instruction.
9133 // The original load instruction.
9135 // The right shift amount in bits from the original load.
9137 // The DAG from which Origin came from.
9138 // This is used to get some contextual information about legal types, etc.
9141 LoadedSlice(SDNode *Inst = nullptr, LoadSDNode *Origin = nullptr,
9142 unsigned Shift = 0, SelectionDAG *DAG = nullptr)
9143 : Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {}
9145 /// \brief Get the bits used in a chunk of bits \p BitWidth large.
9146 /// \return Result is \p BitWidth and has used bits set to 1 and
9147 /// not used bits set to 0.
9148 APInt getUsedBits() const {
9149 // Reproduce the trunc(lshr) sequence:
9150 // - Start from the truncated value.
9151 // - Zero extend to the desired bit width.
9153 assert(Origin && "No original load to compare against.");
9154 unsigned BitWidth = Origin->getValueSizeInBits(0);
9155 assert(Inst && "This slice is not bound to an instruction");
9156 assert(Inst->getValueSizeInBits(0) <= BitWidth &&
9157 "Extracted slice is bigger than the whole type!");
9158 APInt UsedBits(Inst->getValueSizeInBits(0), 0);
9159 UsedBits.setAllBits();
9160 UsedBits = UsedBits.zext(BitWidth);
9165 /// \brief Get the size of the slice to be loaded in bytes.
9166 unsigned getLoadedSize() const {
9167 unsigned SliceSize = getUsedBits().countPopulation();
9168 assert(!(SliceSize & 0x7) && "Size is not a multiple of a byte.");
9169 return SliceSize / 8;
9172 /// \brief Get the type that will be loaded for this slice.
9173 /// Note: This may not be the final type for the slice.
9174 EVT getLoadedType() const {
9175 assert(DAG && "Missing context");
9176 LLVMContext &Ctxt = *DAG->getContext();
9177 return EVT::getIntegerVT(Ctxt, getLoadedSize() * 8);
9180 /// \brief Get the alignment of the load used for this slice.
9181 unsigned getAlignment() const {
9182 unsigned Alignment = Origin->getAlignment();
9183 unsigned Offset = getOffsetFromBase();
9185 Alignment = MinAlign(Alignment, Alignment + Offset);
9189 /// \brief Check if this slice can be rewritten with legal operations.
9190 bool isLegal() const {
9191 // An invalid slice is not legal.
9192 if (!Origin || !Inst || !DAG)
9195 // Offsets are for indexed load only, we do not handle that.
9196 if (Origin->getOffset().getOpcode() != ISD::UNDEF)
9199 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
9201 // Check that the type is legal.
9202 EVT SliceType = getLoadedType();
9203 if (!TLI.isTypeLegal(SliceType))
9206 // Check that the load is legal for this type.
9207 if (!TLI.isOperationLegal(ISD::LOAD, SliceType))
9210 // Check that the offset can be computed.
9211 // 1. Check its type.
9212 EVT PtrType = Origin->getBasePtr().getValueType();
9213 if (PtrType == MVT::Untyped || PtrType.isExtended())
9216 // 2. Check that it fits in the immediate.
9217 if (!TLI.isLegalAddImmediate(getOffsetFromBase()))
9220 // 3. Check that the computation is legal.
9221 if (!TLI.isOperationLegal(ISD::ADD, PtrType))
9224 // Check that the zext is legal if it needs one.
9225 EVT TruncateType = Inst->getValueType(0);
9226 if (TruncateType != SliceType &&
9227 !TLI.isOperationLegal(ISD::ZERO_EXTEND, TruncateType))
9233 /// \brief Get the offset in bytes of this slice in the original chunk of
9235 /// \pre DAG != nullptr.
9236 uint64_t getOffsetFromBase() const {
9237 assert(DAG && "Missing context.");
9239 DAG->getTargetLoweringInfo().getDataLayout()->isBigEndian();
9240 assert(!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported.");
9241 uint64_t Offset = Shift / 8;
9242 unsigned TySizeInBytes = Origin->getValueSizeInBits(0) / 8;
9243 assert(!(Origin->getValueSizeInBits(0) & 0x7) &&
9244 "The size of the original loaded type is not a multiple of a"
9246 // If Offset is bigger than TySizeInBytes, it means we are loading all
9247 // zeros. This should have been optimized before in the process.
9248 assert(TySizeInBytes > Offset &&
9249 "Invalid shift amount for given loaded size");
9251 Offset = TySizeInBytes - Offset - getLoadedSize();
9255 /// \brief Generate the sequence of instructions to load the slice
9256 /// represented by this object and redirect the uses of this slice to
9257 /// this new sequence of instructions.
9258 /// \pre this->Inst && this->Origin are valid Instructions and this
9259 /// object passed the legal check: LoadedSlice::isLegal returned true.
9260 /// \return The last instruction of the sequence used to load the slice.
9261 SDValue loadSlice() const {
9262 assert(Inst && Origin && "Unable to replace a non-existing slice.");
9263 const SDValue &OldBaseAddr = Origin->getBasePtr();
9264 SDValue BaseAddr = OldBaseAddr;
9265 // Get the offset in that chunk of bytes w.r.t. the endianess.
9266 int64_t Offset = static_cast<int64_t>(getOffsetFromBase());
9267 assert(Offset >= 0 && "Offset too big to fit in int64_t!");
9269 // BaseAddr = BaseAddr + Offset.
9270 EVT ArithType = BaseAddr.getValueType();
9271 BaseAddr = DAG->getNode(ISD::ADD, SDLoc(Origin), ArithType, BaseAddr,
9272 DAG->getConstant(Offset, ArithType));
9275 // Create the type of the loaded slice according to its size.
9276 EVT SliceType = getLoadedType();
9278 // Create the load for the slice.
9279 SDValue LastInst = DAG->getLoad(
9280 SliceType, SDLoc(Origin), Origin->getChain(), BaseAddr,
9281 Origin->getPointerInfo().getWithOffset(Offset), Origin->isVolatile(),
9282 Origin->isNonTemporal(), Origin->isInvariant(), getAlignment());
9283 // If the final type is not the same as the loaded type, this means that
9284 // we have to pad with zero. Create a zero extend for that.
9285 EVT FinalType = Inst->getValueType(0);
9286 if (SliceType != FinalType)
9288 DAG->getNode(ISD::ZERO_EXTEND, SDLoc(LastInst), FinalType, LastInst);
9292 /// \brief Check if this slice can be merged with an expensive cross register
9293 /// bank copy. E.g.,
9295 /// f = bitcast i32 i to float
9296 bool canMergeExpensiveCrossRegisterBankCopy() const {
9297 if (!Inst || !Inst->hasOneUse())
9299 SDNode *Use = *Inst->use_begin();
9300 if (Use->getOpcode() != ISD::BITCAST)
9302 assert(DAG && "Missing context");
9303 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
9304 EVT ResVT = Use->getValueType(0);
9305 const TargetRegisterClass *ResRC = TLI.getRegClassFor(ResVT.getSimpleVT());
9306 const TargetRegisterClass *ArgRC =
9307 TLI.getRegClassFor(Use->getOperand(0).getValueType().getSimpleVT());
9308 if (ArgRC == ResRC || !TLI.isOperationLegal(ISD::LOAD, ResVT))
9311 // At this point, we know that we perform a cross-register-bank copy.
9312 // Check if it is expensive.
9313 const TargetRegisterInfo *TRI = DAG->getSubtarget().getRegisterInfo();
9314 // Assume bitcasts are cheap, unless both register classes do not
9315 // explicitly share a common sub class.
9316 if (!TRI || TRI->getCommonSubClass(ArgRC, ResRC))
9319 // Check if it will be merged with the load.
9320 // 1. Check the alignment constraint.
9321 unsigned RequiredAlignment = TLI.getDataLayout()->getABITypeAlignment(
9322 ResVT.getTypeForEVT(*DAG->getContext()));
9324 if (RequiredAlignment > getAlignment())
9327 // 2. Check that the load is a legal operation for that type.
9328 if (!TLI.isOperationLegal(ISD::LOAD, ResVT))
9331 // 3. Check that we do not have a zext in the way.
9332 if (Inst->getValueType(0) != getLoadedType())
9340 /// \brief Check that all bits set in \p UsedBits form a dense region, i.e.,
9341 /// \p UsedBits looks like 0..0 1..1 0..0.
9342 static bool areUsedBitsDense(const APInt &UsedBits) {
9343 // If all the bits are one, this is dense!
9344 if (UsedBits.isAllOnesValue())
9347 // Get rid of the unused bits on the right.
9348 APInt NarrowedUsedBits = UsedBits.lshr(UsedBits.countTrailingZeros());
9349 // Get rid of the unused bits on the left.
9350 if (NarrowedUsedBits.countLeadingZeros())
9351 NarrowedUsedBits = NarrowedUsedBits.trunc(NarrowedUsedBits.getActiveBits());
9352 // Check that the chunk of bits is completely used.
9353 return NarrowedUsedBits.isAllOnesValue();
9356 /// \brief Check whether or not \p First and \p Second are next to each other
9357 /// in memory. This means that there is no hole between the bits loaded
9358 /// by \p First and the bits loaded by \p Second.
9359 static bool areSlicesNextToEachOther(const LoadedSlice &First,
9360 const LoadedSlice &Second) {
9361 assert(First.Origin == Second.Origin && First.Origin &&
9362 "Unable to match different memory origins.");
9363 APInt UsedBits = First.getUsedBits();
9364 assert((UsedBits & Second.getUsedBits()) == 0 &&
9365 "Slices are not supposed to overlap.");
9366 UsedBits |= Second.getUsedBits();
9367 return areUsedBitsDense(UsedBits);
9370 /// \brief Adjust the \p GlobalLSCost according to the target
9371 /// paring capabilities and the layout of the slices.
9372 /// \pre \p GlobalLSCost should account for at least as many loads as
9373 /// there is in the slices in \p LoadedSlices.
9374 static void adjustCostForPairing(SmallVectorImpl<LoadedSlice> &LoadedSlices,
9375 LoadedSlice::Cost &GlobalLSCost) {
9376 unsigned NumberOfSlices = LoadedSlices.size();
9377 // If there is less than 2 elements, no pairing is possible.
9378 if (NumberOfSlices < 2)
9381 // Sort the slices so that elements that are likely to be next to each
9382 // other in memory are next to each other in the list.
9383 std::sort(LoadedSlices.begin(), LoadedSlices.end(),
9384 [](const LoadedSlice &LHS, const LoadedSlice &RHS) {
9385 assert(LHS.Origin == RHS.Origin && "Different bases not implemented.");
9386 return LHS.getOffsetFromBase() < RHS.getOffsetFromBase();
9388 const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo();
9389 // First (resp. Second) is the first (resp. Second) potentially candidate
9390 // to be placed in a paired load.
9391 const LoadedSlice *First = nullptr;
9392 const LoadedSlice *Second = nullptr;
9393 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice,
9394 // Set the beginning of the pair.
9397 Second = &LoadedSlices[CurrSlice];
9399 // If First is NULL, it means we start a new pair.
9400 // Get to the next slice.
9404 EVT LoadedType = First->getLoadedType();
9406 // If the types of the slices are different, we cannot pair them.
9407 if (LoadedType != Second->getLoadedType())
9410 // Check if the target supplies paired loads for this type.
9411 unsigned RequiredAlignment = 0;
9412 if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) {
9413 // move to the next pair, this type is hopeless.
9417 // Check if we meet the alignment requirement.
9418 if (RequiredAlignment > First->getAlignment())
9421 // Check that both loads are next to each other in memory.
9422 if (!areSlicesNextToEachOther(*First, *Second))
9425 assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!");
9426 --GlobalLSCost.Loads;
9427 // Move to the next pair.
9432 /// \brief Check the profitability of all involved LoadedSlice.
9433 /// Currently, it is considered profitable if there is exactly two
9434 /// involved slices (1) which are (2) next to each other in memory, and
9435 /// whose cost (\see LoadedSlice::Cost) is smaller than the original load (3).
9437 /// Note: The order of the elements in \p LoadedSlices may be modified, but not
9438 /// the elements themselves.
9440 /// FIXME: When the cost model will be mature enough, we can relax
9441 /// constraints (1) and (2).
9442 static bool isSlicingProfitable(SmallVectorImpl<LoadedSlice> &LoadedSlices,
9443 const APInt &UsedBits, bool ForCodeSize) {
9444 unsigned NumberOfSlices = LoadedSlices.size();
9445 if (StressLoadSlicing)
9446 return NumberOfSlices > 1;
9449 if (NumberOfSlices != 2)
9453 if (!areUsedBitsDense(UsedBits))
9457 LoadedSlice::Cost OrigCost(ForCodeSize), GlobalSlicingCost(ForCodeSize);
9458 // The original code has one big load.
9460 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice) {
9461 const LoadedSlice &LS = LoadedSlices[CurrSlice];
9462 // Accumulate the cost of all the slices.
9463 LoadedSlice::Cost SliceCost(LS, ForCodeSize);
9464 GlobalSlicingCost += SliceCost;
9466 // Account as cost in the original configuration the gain obtained
9467 // with the current slices.
9468 OrigCost.addSliceGain(LS);
9471 // If the target supports paired load, adjust the cost accordingly.
9472 adjustCostForPairing(LoadedSlices, GlobalSlicingCost);
9473 return OrigCost > GlobalSlicingCost;
9476 /// \brief If the given load, \p LI, is used only by trunc or trunc(lshr)
9477 /// operations, split it in the various pieces being extracted.
9479 /// This sort of thing is introduced by SROA.
9480 /// This slicing takes care not to insert overlapping loads.
9481 /// \pre LI is a simple load (i.e., not an atomic or volatile load).
9482 bool DAGCombiner::SliceUpLoad(SDNode *N) {
9483 if (Level < AfterLegalizeDAG)
9486 LoadSDNode *LD = cast<LoadSDNode>(N);
9487 if (LD->isVolatile() || !ISD::isNormalLoad(LD) ||
9488 !LD->getValueType(0).isInteger())
9491 // Keep track of already used bits to detect overlapping values.
9492 // In that case, we will just abort the transformation.
9493 APInt UsedBits(LD->getValueSizeInBits(0), 0);
9495 SmallVector<LoadedSlice, 4> LoadedSlices;
9497 // Check if this load is used as several smaller chunks of bits.
9498 // Basically, look for uses in trunc or trunc(lshr) and record a new chain
9499 // of computation for each trunc.
9500 for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end();
9501 UI != UIEnd; ++UI) {
9502 // Skip the uses of the chain.
9503 if (UI.getUse().getResNo() != 0)
9509 // Check if this is a trunc(lshr).
9510 if (User->getOpcode() == ISD::SRL && User->hasOneUse() &&
9511 isa<ConstantSDNode>(User->getOperand(1))) {
9512 Shift = cast<ConstantSDNode>(User->getOperand(1))->getZExtValue();
9513 User = *User->use_begin();
9516 // At this point, User is a Truncate, iff we encountered, trunc or
9518 if (User->getOpcode() != ISD::TRUNCATE)
9521 // The width of the type must be a power of 2 and greater than 8-bits.
9522 // Otherwise the load cannot be represented in LLVM IR.
9523 // Moreover, if we shifted with a non-8-bits multiple, the slice
9524 // will be across several bytes. We do not support that.
9525 unsigned Width = User->getValueSizeInBits(0);
9526 if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7))
9529 // Build the slice for this chain of computations.
9530 LoadedSlice LS(User, LD, Shift, &DAG);
9531 APInt CurrentUsedBits = LS.getUsedBits();
9533 // Check if this slice overlaps with another.
9534 if ((CurrentUsedBits & UsedBits) != 0)
9536 // Update the bits used globally.
9537 UsedBits |= CurrentUsedBits;
9539 // Check if the new slice would be legal.
9543 // Record the slice.
9544 LoadedSlices.push_back(LS);
9547 // Abort slicing if it does not seem to be profitable.
9548 if (!isSlicingProfitable(LoadedSlices, UsedBits, ForCodeSize))
9553 // Rewrite each chain to use an independent load.
9554 // By construction, each chain can be represented by a unique load.
9556 // Prepare the argument for the new token factor for all the slices.
9557 SmallVector<SDValue, 8> ArgChains;
9558 for (SmallVectorImpl<LoadedSlice>::const_iterator
9559 LSIt = LoadedSlices.begin(),
9560 LSItEnd = LoadedSlices.end();
9561 LSIt != LSItEnd; ++LSIt) {
9562 SDValue SliceInst = LSIt->loadSlice();
9563 CombineTo(LSIt->Inst, SliceInst, true);
9564 if (SliceInst.getNode()->getOpcode() != ISD::LOAD)
9565 SliceInst = SliceInst.getOperand(0);
9566 assert(SliceInst->getOpcode() == ISD::LOAD &&
9567 "It takes more than a zext to get to the loaded slice!!");
9568 ArgChains.push_back(SliceInst.getValue(1));
9571 SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other,
9573 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
9577 /// Check to see if V is (and load (ptr), imm), where the load is having
9578 /// specific bytes cleared out. If so, return the byte size being masked out
9579 /// and the shift amount.
9580 static std::pair<unsigned, unsigned>
9581 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
9582 std::pair<unsigned, unsigned> Result(0, 0);
9584 // Check for the structure we're looking for.
9585 if (V->getOpcode() != ISD::AND ||
9586 !isa<ConstantSDNode>(V->getOperand(1)) ||
9587 !ISD::isNormalLoad(V->getOperand(0).getNode()))
9590 // Check the chain and pointer.
9591 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
9592 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
9594 // The store should be chained directly to the load or be an operand of a
9596 if (LD == Chain.getNode())
9598 else if (Chain->getOpcode() != ISD::TokenFactor)
9599 return Result; // Fail.
9602 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
9603 if (Chain->getOperand(i).getNode() == LD) {
9607 if (!isOk) return Result;
9610 // This only handles simple types.
9611 if (V.getValueType() != MVT::i16 &&
9612 V.getValueType() != MVT::i32 &&
9613 V.getValueType() != MVT::i64)
9616 // Check the constant mask. Invert it so that the bits being masked out are
9617 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
9618 // follow the sign bit for uniformity.
9619 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
9620 unsigned NotMaskLZ = countLeadingZeros(NotMask);
9621 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
9622 unsigned NotMaskTZ = countTrailingZeros(NotMask);
9623 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
9624 if (NotMaskLZ == 64) return Result; // All zero mask.
9626 // See if we have a continuous run of bits. If so, we have 0*1+0*
9627 if (countTrailingOnes(NotMask >> NotMaskTZ) + NotMaskTZ + NotMaskLZ != 64)
9630 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
9631 if (V.getValueType() != MVT::i64 && NotMaskLZ)
9632 NotMaskLZ -= 64-V.getValueSizeInBits();
9634 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
9635 switch (MaskedBytes) {
9639 default: return Result; // All one mask, or 5-byte mask.
9642 // Verify that the first bit starts at a multiple of mask so that the access
9643 // is aligned the same as the access width.
9644 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
9646 Result.first = MaskedBytes;
9647 Result.second = NotMaskTZ/8;
9652 /// Check to see if IVal is something that provides a value as specified by
9653 /// MaskInfo. If so, replace the specified store with a narrower store of
9656 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
9657 SDValue IVal, StoreSDNode *St,
9659 unsigned NumBytes = MaskInfo.first;
9660 unsigned ByteShift = MaskInfo.second;
9661 SelectionDAG &DAG = DC->getDAG();
9663 // Check to see if IVal is all zeros in the part being masked in by the 'or'
9664 // that uses this. If not, this is not a replacement.
9665 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
9666 ByteShift*8, (ByteShift+NumBytes)*8);
9667 if (!DAG.MaskedValueIsZero(IVal, Mask)) return nullptr;
9669 // Check that it is legal on the target to do this. It is legal if the new
9670 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
9672 MVT VT = MVT::getIntegerVT(NumBytes*8);
9673 if (!DC->isTypeLegal(VT))
9676 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
9677 // shifted by ByteShift and truncated down to NumBytes.
9679 IVal = DAG.getNode(ISD::SRL, SDLoc(IVal), IVal.getValueType(), IVal,
9680 DAG.getConstant(ByteShift*8,
9681 DC->getShiftAmountTy(IVal.getValueType())));
9683 // Figure out the offset for the store and the alignment of the access.
9685 unsigned NewAlign = St->getAlignment();
9687 if (DAG.getTargetLoweringInfo().isLittleEndian())
9688 StOffset = ByteShift;
9690 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
9692 SDValue Ptr = St->getBasePtr();
9694 Ptr = DAG.getNode(ISD::ADD, SDLoc(IVal), Ptr.getValueType(),
9695 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
9696 NewAlign = MinAlign(NewAlign, StOffset);
9699 // Truncate down to the new size.
9700 IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal);
9703 return DAG.getStore(St->getChain(), SDLoc(St), IVal, Ptr,
9704 St->getPointerInfo().getWithOffset(StOffset),
9705 false, false, NewAlign).getNode();
9709 /// Look for sequence of load / op / store where op is one of 'or', 'xor', and
9710 /// 'and' of immediates. If 'op' is only touching some of the loaded bits, try
9711 /// narrowing the load and store if it would end up being a win for performance
9713 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
9714 StoreSDNode *ST = cast<StoreSDNode>(N);
9715 if (ST->isVolatile())
9718 SDValue Chain = ST->getChain();
9719 SDValue Value = ST->getValue();
9720 SDValue Ptr = ST->getBasePtr();
9721 EVT VT = Value.getValueType();
9723 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
9726 unsigned Opc = Value.getOpcode();
9728 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
9729 // is a byte mask indicating a consecutive number of bytes, check to see if
9730 // Y is known to provide just those bytes. If so, we try to replace the
9731 // load + replace + store sequence with a single (narrower) store, which makes
9733 if (Opc == ISD::OR) {
9734 std::pair<unsigned, unsigned> MaskedLoad;
9735 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
9736 if (MaskedLoad.first)
9737 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
9738 Value.getOperand(1), ST,this))
9739 return SDValue(NewST, 0);
9741 // Or is commutative, so try swapping X and Y.
9742 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
9743 if (MaskedLoad.first)
9744 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
9745 Value.getOperand(0), ST,this))
9746 return SDValue(NewST, 0);
9749 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
9750 Value.getOperand(1).getOpcode() != ISD::Constant)
9753 SDValue N0 = Value.getOperand(0);
9754 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
9755 Chain == SDValue(N0.getNode(), 1)) {
9756 LoadSDNode *LD = cast<LoadSDNode>(N0);
9757 if (LD->getBasePtr() != Ptr ||
9758 LD->getPointerInfo().getAddrSpace() !=
9759 ST->getPointerInfo().getAddrSpace())
9762 // Find the type to narrow it the load / op / store to.
9763 SDValue N1 = Value.getOperand(1);
9764 unsigned BitWidth = N1.getValueSizeInBits();
9765 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
9766 if (Opc == ISD::AND)
9767 Imm ^= APInt::getAllOnesValue(BitWidth);
9768 if (Imm == 0 || Imm.isAllOnesValue())
9770 unsigned ShAmt = Imm.countTrailingZeros();
9771 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
9772 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
9773 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
9774 // The narrowing should be profitable, the load/store operation should be
9775 // legal (or custom) and the store size should be equal to the NewVT width.
9776 while (NewBW < BitWidth &&
9777 (NewVT.getStoreSizeInBits() != NewBW ||
9778 !TLI.isOperationLegalOrCustom(Opc, NewVT) ||
9779 !TLI.isNarrowingProfitable(VT, NewVT))) {
9780 NewBW = NextPowerOf2(NewBW);
9781 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
9783 if (NewBW >= BitWidth)
9786 // If the lsb changed does not start at the type bitwidth boundary,
9787 // start at the previous one.
9789 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
9790 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt,
9791 std::min(BitWidth, ShAmt + NewBW));
9792 if ((Imm & Mask) == Imm) {
9793 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
9794 if (Opc == ISD::AND)
9795 NewImm ^= APInt::getAllOnesValue(NewBW);
9796 uint64_t PtrOff = ShAmt / 8;
9797 // For big endian targets, we need to adjust the offset to the pointer to
9798 // load the correct bytes.
9799 if (TLI.isBigEndian())
9800 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
9802 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
9803 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
9804 if (NewAlign < TLI.getDataLayout()->getABITypeAlignment(NewVTTy))
9807 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD),
9808 Ptr.getValueType(), Ptr,
9809 DAG.getConstant(PtrOff, Ptr.getValueType()));
9810 SDValue NewLD = DAG.getLoad(NewVT, SDLoc(N0),
9811 LD->getChain(), NewPtr,
9812 LD->getPointerInfo().getWithOffset(PtrOff),
9813 LD->isVolatile(), LD->isNonTemporal(),
9814 LD->isInvariant(), NewAlign,
9816 SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD,
9817 DAG.getConstant(NewImm, NewVT));
9818 SDValue NewST = DAG.getStore(Chain, SDLoc(N),
9820 ST->getPointerInfo().getWithOffset(PtrOff),
9821 false, false, NewAlign);
9823 AddToWorklist(NewPtr.getNode());
9824 AddToWorklist(NewLD.getNode());
9825 AddToWorklist(NewVal.getNode());
9826 WorklistRemover DeadNodes(*this);
9827 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
9836 /// For a given floating point load / store pair, if the load value isn't used
9837 /// by any other operations, then consider transforming the pair to integer
9838 /// load / store operations if the target deems the transformation profitable.
9839 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
9840 StoreSDNode *ST = cast<StoreSDNode>(N);
9841 SDValue Chain = ST->getChain();
9842 SDValue Value = ST->getValue();
9843 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
9844 Value.hasOneUse() &&
9845 Chain == SDValue(Value.getNode(), 1)) {
9846 LoadSDNode *LD = cast<LoadSDNode>(Value);
9847 EVT VT = LD->getMemoryVT();
9848 if (!VT.isFloatingPoint() ||
9849 VT != ST->getMemoryVT() ||
9850 LD->isNonTemporal() ||
9851 ST->isNonTemporal() ||
9852 LD->getPointerInfo().getAddrSpace() != 0 ||
9853 ST->getPointerInfo().getAddrSpace() != 0)
9856 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
9857 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
9858 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
9859 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
9860 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
9863 unsigned LDAlign = LD->getAlignment();
9864 unsigned STAlign = ST->getAlignment();
9865 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
9866 unsigned ABIAlign = TLI.getDataLayout()->getABITypeAlignment(IntVTTy);
9867 if (LDAlign < ABIAlign || STAlign < ABIAlign)
9870 SDValue NewLD = DAG.getLoad(IntVT, SDLoc(Value),
9871 LD->getChain(), LD->getBasePtr(),
9872 LD->getPointerInfo(),
9873 false, false, false, LDAlign);
9875 SDValue NewST = DAG.getStore(NewLD.getValue(1), SDLoc(N),
9876 NewLD, ST->getBasePtr(),
9877 ST->getPointerInfo(),
9878 false, false, STAlign);
9880 AddToWorklist(NewLD.getNode());
9881 AddToWorklist(NewST.getNode());
9882 WorklistRemover DeadNodes(*this);
9883 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
9891 /// Helper struct to parse and store a memory address as base + index + offset.
9892 /// We ignore sign extensions when it is safe to do so.
9893 /// The following two expressions are not equivalent. To differentiate we need
9894 /// to store whether there was a sign extension involved in the index
9896 /// (load (i64 add (i64 copyfromreg %c)
9897 /// (i64 signextend (add (i8 load %index)
9901 /// (load (i64 add (i64 copyfromreg %c)
9902 /// (i64 signextend (i32 add (i32 signextend (i8 load %index))
9904 struct BaseIndexOffset {
9908 bool IsIndexSignExt;
9910 BaseIndexOffset() : Offset(0), IsIndexSignExt(false) {}
9912 BaseIndexOffset(SDValue Base, SDValue Index, int64_t Offset,
9913 bool IsIndexSignExt) :
9914 Base(Base), Index(Index), Offset(Offset), IsIndexSignExt(IsIndexSignExt) {}
9916 bool equalBaseIndex(const BaseIndexOffset &Other) {
9917 return Other.Base == Base && Other.Index == Index &&
9918 Other.IsIndexSignExt == IsIndexSignExt;
9921 /// Parses tree in Ptr for base, index, offset addresses.
9922 static BaseIndexOffset match(SDValue Ptr) {
9923 bool IsIndexSignExt = false;
9925 // We only can pattern match BASE + INDEX + OFFSET. If Ptr is not an ADD
9926 // instruction, then it could be just the BASE or everything else we don't
9927 // know how to handle. Just use Ptr as BASE and give up.
9928 if (Ptr->getOpcode() != ISD::ADD)
9929 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
9931 // We know that we have at least an ADD instruction. Try to pattern match
9932 // the simple case of BASE + OFFSET.
9933 if (isa<ConstantSDNode>(Ptr->getOperand(1))) {
9934 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
9935 return BaseIndexOffset(Ptr->getOperand(0), SDValue(), Offset,
9939 // Inside a loop the current BASE pointer is calculated using an ADD and a
9940 // MUL instruction. In this case Ptr is the actual BASE pointer.
9941 // (i64 add (i64 %array_ptr)
9942 // (i64 mul (i64 %induction_var)
9943 // (i64 %element_size)))
9944 if (Ptr->getOperand(1)->getOpcode() == ISD::MUL)
9945 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
9947 // Look at Base + Index + Offset cases.
9948 SDValue Base = Ptr->getOperand(0);
9949 SDValue IndexOffset = Ptr->getOperand(1);
9951 // Skip signextends.
9952 if (IndexOffset->getOpcode() == ISD::SIGN_EXTEND) {
9953 IndexOffset = IndexOffset->getOperand(0);
9954 IsIndexSignExt = true;
9957 // Either the case of Base + Index (no offset) or something else.
9958 if (IndexOffset->getOpcode() != ISD::ADD)
9959 return BaseIndexOffset(Base, IndexOffset, 0, IsIndexSignExt);
9961 // Now we have the case of Base + Index + offset.
9962 SDValue Index = IndexOffset->getOperand(0);
9963 SDValue Offset = IndexOffset->getOperand(1);
9965 if (!isa<ConstantSDNode>(Offset))
9966 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
9968 // Ignore signextends.
9969 if (Index->getOpcode() == ISD::SIGN_EXTEND) {
9970 Index = Index->getOperand(0);
9971 IsIndexSignExt = true;
9972 } else IsIndexSignExt = false;
9974 int64_t Off = cast<ConstantSDNode>(Offset)->getSExtValue();
9975 return BaseIndexOffset(Base, Index, Off, IsIndexSignExt);
9979 bool DAGCombiner::MergeStoresOfConstantsOrVecElts(
9980 SmallVectorImpl<MemOpLink> &StoreNodes, EVT MemVT,
9981 unsigned NumElem, bool IsConstantSrc, bool UseVector) {
9982 // Make sure we have something to merge.
9986 int64_t ElementSizeBytes = MemVT.getSizeInBits() / 8;
9987 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
9988 unsigned EarliestNodeUsed = 0;
9990 for (unsigned i=0; i < NumElem; ++i) {
9991 // Find a chain for the new wide-store operand. Notice that some
9992 // of the store nodes that we found may not be selected for inclusion
9993 // in the wide store. The chain we use needs to be the chain of the
9994 // earliest store node which is *used* and replaced by the wide store.
9995 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
9996 EarliestNodeUsed = i;
9999 // The earliest Node in the DAG.
10000 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
10001 SDLoc DL(StoreNodes[0].MemNode);
10005 // Find a legal type for the vector store.
10006 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
10007 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
10008 if (IsConstantSrc) {
10009 // A vector store with a constant source implies that the constant is
10010 // zero; we only handle merging stores of constant zeros because the zero
10011 // can be materialized without a load.
10012 // It may be beneficial to loosen this restriction to allow non-zero
10014 StoredVal = DAG.getConstant(0, Ty);
10016 SmallVector<SDValue, 8> Ops;
10017 for (unsigned i = 0; i < NumElem ; ++i) {
10018 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10019 SDValue Val = St->getValue();
10020 // All of the operands of a BUILD_VECTOR must have the same type.
10021 if (Val.getValueType() != MemVT)
10023 Ops.push_back(Val);
10026 // Build the extracted vector elements back into a vector.
10027 StoredVal = DAG.getNode(ISD::BUILD_VECTOR, DL, Ty, Ops);
10030 // We should always use a vector store when merging extracted vector
10031 // elements, so this path implies a store of constants.
10032 assert(IsConstantSrc && "Merged vector elements should use vector store");
10034 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
10035 APInt StoreInt(StoreBW, 0);
10037 // Construct a single integer constant which is made of the smaller
10038 // constant inputs.
10039 bool IsLE = TLI.isLittleEndian();
10040 for (unsigned i = 0; i < NumElem ; ++i) {
10041 unsigned Idx = IsLE ? (NumElem - 1 - i) : i;
10042 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
10043 SDValue Val = St->getValue();
10044 StoreInt <<= ElementSizeBytes*8;
10045 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
10046 StoreInt |= C->getAPIntValue().zext(StoreBW);
10047 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
10048 StoreInt |= C->getValueAPF().bitcastToAPInt().zext(StoreBW);
10050 llvm_unreachable("Invalid constant element type");
10054 // Create the new Load and Store operations.
10055 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
10056 StoredVal = DAG.getConstant(StoreInt, StoreTy);
10059 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), DL, StoredVal,
10060 FirstInChain->getBasePtr(),
10061 FirstInChain->getPointerInfo(),
10063 FirstInChain->getAlignment());
10065 // Replace the first store with the new store
10066 CombineTo(EarliestOp, NewStore);
10067 // Erase all other stores.
10068 for (unsigned i = 0; i < NumElem ; ++i) {
10069 if (StoreNodes[i].MemNode == EarliestOp)
10071 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10072 // ReplaceAllUsesWith will replace all uses that existed when it was
10073 // called, but graph optimizations may cause new ones to appear. For
10074 // example, the case in pr14333 looks like
10076 // St's chain -> St -> another store -> X
10078 // And the only difference from St to the other store is the chain.
10079 // When we change it's chain to be St's chain they become identical,
10080 // get CSEed and the net result is that X is now a use of St.
10081 // Since we know that St is redundant, just iterate.
10082 while (!St->use_empty())
10083 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain());
10084 deleteAndRecombine(St);
10090 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
10091 if (OptLevel == CodeGenOpt::None)
10094 EVT MemVT = St->getMemoryVT();
10095 int64_t ElementSizeBytes = MemVT.getSizeInBits()/8;
10096 bool NoVectors = DAG.getMachineFunction().getFunction()->hasFnAttribute(
10097 Attribute::NoImplicitFloat);
10099 // Don't merge vectors into wider inputs.
10100 if (MemVT.isVector() || !MemVT.isSimple())
10103 // Perform an early exit check. Do not bother looking at stored values that
10104 // are not constants, loads, or extracted vector elements.
10105 SDValue StoredVal = St->getValue();
10106 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
10107 bool IsConstantSrc = isa<ConstantSDNode>(StoredVal) ||
10108 isa<ConstantFPSDNode>(StoredVal);
10109 bool IsExtractVecEltSrc = (StoredVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT);
10111 if (!IsConstantSrc && !IsLoadSrc && !IsExtractVecEltSrc)
10114 // Only look at ends of store sequences.
10115 SDValue Chain = SDValue(St, 0);
10116 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
10119 // This holds the base pointer, index, and the offset in bytes from the base
10121 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr());
10123 // We must have a base and an offset.
10124 if (!BasePtr.Base.getNode())
10127 // Do not handle stores to undef base pointers.
10128 if (BasePtr.Base.getOpcode() == ISD::UNDEF)
10131 // Save the LoadSDNodes that we find in the chain.
10132 // We need to make sure that these nodes do not interfere with
10133 // any of the store nodes.
10134 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
10136 // Save the StoreSDNodes that we find in the chain.
10137 SmallVector<MemOpLink, 8> StoreNodes;
10139 // Walk up the chain and look for nodes with offsets from the same
10140 // base pointer. Stop when reaching an instruction with a different kind
10141 // or instruction which has a different base pointer.
10143 StoreSDNode *Index = St;
10145 // If the chain has more than one use, then we can't reorder the mem ops.
10146 if (Index != St && !SDValue(Index, 0)->hasOneUse())
10149 // Find the base pointer and offset for this memory node.
10150 BaseIndexOffset Ptr = BaseIndexOffset::match(Index->getBasePtr());
10152 // Check that the base pointer is the same as the original one.
10153 if (!Ptr.equalBaseIndex(BasePtr))
10156 // Check that the alignment is the same.
10157 if (Index->getAlignment() != St->getAlignment())
10160 // The memory operands must not be volatile.
10161 if (Index->isVolatile() || Index->isIndexed())
10165 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
10166 if (St->isTruncatingStore())
10169 // The stored memory type must be the same.
10170 if (Index->getMemoryVT() != MemVT)
10173 // We do not allow unaligned stores because we want to prevent overriding
10175 if (Index->getAlignment()*8 != MemVT.getSizeInBits())
10178 // We found a potential memory operand to merge.
10179 StoreNodes.push_back(MemOpLink(Index, Ptr.Offset, Seq++));
10181 // Find the next memory operand in the chain. If the next operand in the
10182 // chain is a store then move up and continue the scan with the next
10183 // memory operand. If the next operand is a load save it and use alias
10184 // information to check if it interferes with anything.
10185 SDNode *NextInChain = Index->getChain().getNode();
10187 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
10188 // We found a store node. Use it for the next iteration.
10191 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
10192 if (Ldn->isVolatile()) {
10197 // Save the load node for later. Continue the scan.
10198 AliasLoadNodes.push_back(Ldn);
10199 NextInChain = Ldn->getChain().getNode();
10208 // Check if there is anything to merge.
10209 if (StoreNodes.size() < 2)
10212 // Sort the memory operands according to their distance from the base pointer.
10213 std::sort(StoreNodes.begin(), StoreNodes.end(),
10214 [](MemOpLink LHS, MemOpLink RHS) {
10215 return LHS.OffsetFromBase < RHS.OffsetFromBase ||
10216 (LHS.OffsetFromBase == RHS.OffsetFromBase &&
10217 LHS.SequenceNum > RHS.SequenceNum);
10220 // Scan the memory operations on the chain and find the first non-consecutive
10221 // store memory address.
10222 unsigned LastConsecutiveStore = 0;
10223 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
10224 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) {
10226 // Check that the addresses are consecutive starting from the second
10227 // element in the list of stores.
10229 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
10230 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
10234 bool Alias = false;
10235 // Check if this store interferes with any of the loads that we found.
10236 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld)
10237 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) {
10241 // We found a load that alias with this store. Stop the sequence.
10245 // Mark this node as useful.
10246 LastConsecutiveStore = i;
10249 // The node with the lowest store address.
10250 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
10252 // Store the constants into memory as one consecutive store.
10253 if (IsConstantSrc) {
10254 unsigned LastLegalType = 0;
10255 unsigned LastLegalVectorType = 0;
10256 bool NonZero = false;
10257 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
10258 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10259 SDValue StoredVal = St->getValue();
10261 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
10262 NonZero |= !C->isNullValue();
10263 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) {
10264 NonZero |= !C->getConstantFPValue()->isNullValue();
10270 // Find a legal type for the constant store.
10271 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
10272 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
10273 if (TLI.isTypeLegal(StoreTy))
10274 LastLegalType = i+1;
10275 // Or check whether a truncstore is legal.
10276 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
10277 TargetLowering::TypePromoteInteger) {
10278 EVT LegalizedStoredValueTy =
10279 TLI.getTypeToTransformTo(*DAG.getContext(), StoredVal.getValueType());
10280 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy))
10281 LastLegalType = i+1;
10284 // Find a legal type for the vector store.
10285 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
10286 if (TLI.isTypeLegal(Ty))
10287 LastLegalVectorType = i + 1;
10290 // We only use vectors if the constant is known to be zero and the
10291 // function is not marked with the noimplicitfloat attribute.
10292 if (NonZero || NoVectors)
10293 LastLegalVectorType = 0;
10295 // Check if we found a legal integer type to store.
10296 if (LastLegalType == 0 && LastLegalVectorType == 0)
10299 bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors;
10300 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType;
10302 return MergeStoresOfConstantsOrVecElts(StoreNodes, MemVT, NumElem,
10306 // When extracting multiple vector elements, try to store them
10307 // in one vector store rather than a sequence of scalar stores.
10308 if (IsExtractVecEltSrc) {
10309 unsigned NumElem = 0;
10310 for (unsigned i = 0; i < LastConsecutiveStore + 1; ++i) {
10311 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10312 SDValue StoredVal = St->getValue();
10313 // This restriction could be loosened.
10314 // Bail out if any stored values are not elements extracted from a vector.
10315 // It should be possible to handle mixed sources, but load sources need
10316 // more careful handling (see the block of code below that handles
10317 // consecutive loads).
10318 if (StoredVal.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
10321 // Find a legal type for the vector store.
10322 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
10323 if (TLI.isTypeLegal(Ty))
10327 return MergeStoresOfConstantsOrVecElts(StoreNodes, MemVT, NumElem,
10331 // Below we handle the case of multiple consecutive stores that
10332 // come from multiple consecutive loads. We merge them into a single
10333 // wide load and a single wide store.
10335 // Look for load nodes which are used by the stored values.
10336 SmallVector<MemOpLink, 8> LoadNodes;
10338 // Find acceptable loads. Loads need to have the same chain (token factor),
10339 // must not be zext, volatile, indexed, and they must be consecutive.
10340 BaseIndexOffset LdBasePtr;
10341 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
10342 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10343 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
10346 // Loads must only have one use.
10347 if (!Ld->hasNUsesOfValue(1, 0))
10350 // Check that the alignment is the same as the stores.
10351 if (Ld->getAlignment() != St->getAlignment())
10354 // The memory operands must not be volatile.
10355 if (Ld->isVolatile() || Ld->isIndexed())
10358 // We do not accept ext loads.
10359 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
10362 // The stored memory type must be the same.
10363 if (Ld->getMemoryVT() != MemVT)
10366 BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld->getBasePtr());
10367 // If this is not the first ptr that we check.
10368 if (LdBasePtr.Base.getNode()) {
10369 // The base ptr must be the same.
10370 if (!LdPtr.equalBaseIndex(LdBasePtr))
10373 // Check that all other base pointers are the same as this one.
10377 // We found a potential memory operand to merge.
10378 LoadNodes.push_back(MemOpLink(Ld, LdPtr.Offset, 0));
10381 if (LoadNodes.size() < 2)
10384 // If we have load/store pair instructions and we only have two values,
10386 unsigned RequiredAlignment;
10387 if (LoadNodes.size() == 2 && TLI.hasPairedLoad(MemVT, RequiredAlignment) &&
10388 St->getAlignment() >= RequiredAlignment)
10391 // Scan the memory operations on the chain and find the first non-consecutive
10392 // load memory address. These variables hold the index in the store node
10394 unsigned LastConsecutiveLoad = 0;
10395 // This variable refers to the size and not index in the array.
10396 unsigned LastLegalVectorType = 0;
10397 unsigned LastLegalIntegerType = 0;
10398 StartAddress = LoadNodes[0].OffsetFromBase;
10399 SDValue FirstChain = LoadNodes[0].MemNode->getChain();
10400 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
10401 // All loads much share the same chain.
10402 if (LoadNodes[i].MemNode->getChain() != FirstChain)
10405 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
10406 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
10408 LastConsecutiveLoad = i;
10410 // Find a legal type for the vector store.
10411 EVT StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
10412 if (TLI.isTypeLegal(StoreTy))
10413 LastLegalVectorType = i + 1;
10415 // Find a legal type for the integer store.
10416 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
10417 StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
10418 if (TLI.isTypeLegal(StoreTy))
10419 LastLegalIntegerType = i + 1;
10420 // Or check whether a truncstore and extload is legal.
10421 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
10422 TargetLowering::TypePromoteInteger) {
10423 EVT LegalizedStoredValueTy =
10424 TLI.getTypeToTransformTo(*DAG.getContext(), StoreTy);
10425 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
10426 TLI.isLoadExtLegal(ISD::ZEXTLOAD, LegalizedStoredValueTy, StoreTy) &&
10427 TLI.isLoadExtLegal(ISD::SEXTLOAD, LegalizedStoredValueTy, StoreTy) &&
10428 TLI.isLoadExtLegal(ISD::EXTLOAD, LegalizedStoredValueTy, StoreTy))
10429 LastLegalIntegerType = i+1;
10433 // Only use vector types if the vector type is larger than the integer type.
10434 // If they are the same, use integers.
10435 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors;
10436 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType);
10438 // We add +1 here because the LastXXX variables refer to location while
10439 // the NumElem refers to array/index size.
10440 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1;
10441 NumElem = std::min(LastLegalType, NumElem);
10446 // The earliest Node in the DAG.
10447 unsigned EarliestNodeUsed = 0;
10448 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
10449 for (unsigned i=1; i<NumElem; ++i) {
10450 // Find a chain for the new wide-store operand. Notice that some
10451 // of the store nodes that we found may not be selected for inclusion
10452 // in the wide store. The chain we use needs to be the chain of the
10453 // earliest store node which is *used* and replaced by the wide store.
10454 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
10455 EarliestNodeUsed = i;
10458 // Find if it is better to use vectors or integers to load and store
10462 JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
10464 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
10465 JointMemOpVT = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
10468 SDLoc LoadDL(LoadNodes[0].MemNode);
10469 SDLoc StoreDL(StoreNodes[0].MemNode);
10471 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
10472 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL,
10473 FirstLoad->getChain(),
10474 FirstLoad->getBasePtr(),
10475 FirstLoad->getPointerInfo(),
10476 false, false, false,
10477 FirstLoad->getAlignment());
10479 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), StoreDL, NewLoad,
10480 FirstInChain->getBasePtr(),
10481 FirstInChain->getPointerInfo(), false, false,
10482 FirstInChain->getAlignment());
10484 // Replace one of the loads with the new load.
10485 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode);
10486 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
10487 SDValue(NewLoad.getNode(), 1));
10489 // Remove the rest of the load chains.
10490 for (unsigned i = 1; i < NumElem ; ++i) {
10491 // Replace all chain users of the old load nodes with the chain of the new
10493 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
10494 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain());
10497 // Replace the first store with the new store.
10498 CombineTo(EarliestOp, NewStore);
10499 // Erase all other stores.
10500 for (unsigned i = 0; i < NumElem ; ++i) {
10501 // Remove all Store nodes.
10502 if (StoreNodes[i].MemNode == EarliestOp)
10504 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10505 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain());
10506 deleteAndRecombine(St);
10512 SDValue DAGCombiner::visitSTORE(SDNode *N) {
10513 StoreSDNode *ST = cast<StoreSDNode>(N);
10514 SDValue Chain = ST->getChain();
10515 SDValue Value = ST->getValue();
10516 SDValue Ptr = ST->getBasePtr();
10518 // If this is a store of a bit convert, store the input value if the
10519 // resultant store does not need a higher alignment than the original.
10520 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
10521 ST->isUnindexed()) {
10522 unsigned OrigAlign = ST->getAlignment();
10523 EVT SVT = Value.getOperand(0).getValueType();
10524 unsigned Align = TLI.getDataLayout()->
10525 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
10526 if (Align <= OrigAlign &&
10527 ((!LegalOperations && !ST->isVolatile()) ||
10528 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
10529 return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0),
10530 Ptr, ST->getPointerInfo(), ST->isVolatile(),
10531 ST->isNonTemporal(), OrigAlign,
10535 // Turn 'store undef, Ptr' -> nothing.
10536 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
10539 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
10540 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
10541 // NOTE: If the original store is volatile, this transform must not increase
10542 // the number of stores. For example, on x86-32 an f64 can be stored in one
10543 // processor operation but an i64 (which is not legal) requires two. So the
10544 // transform should not be done in this case.
10545 if (Value.getOpcode() != ISD::TargetConstantFP) {
10547 switch (CFP->getSimpleValueType(0).SimpleTy) {
10548 default: llvm_unreachable("Unknown FP type");
10549 case MVT::f16: // We don't do this for these yet.
10555 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
10556 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
10557 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
10558 bitcastToAPInt().getZExtValue(), MVT::i32);
10559 return DAG.getStore(Chain, SDLoc(N), Tmp,
10560 Ptr, ST->getMemOperand());
10564 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
10565 !ST->isVolatile()) ||
10566 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
10567 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
10568 getZExtValue(), MVT::i64);
10569 return DAG.getStore(Chain, SDLoc(N), Tmp,
10570 Ptr, ST->getMemOperand());
10573 if (!ST->isVolatile() &&
10574 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
10575 // Many FP stores are not made apparent until after legalize, e.g. for
10576 // argument passing. Since this is so common, custom legalize the
10577 // 64-bit integer store into two 32-bit stores.
10578 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
10579 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
10580 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
10581 if (TLI.isBigEndian()) std::swap(Lo, Hi);
10583 unsigned Alignment = ST->getAlignment();
10584 bool isVolatile = ST->isVolatile();
10585 bool isNonTemporal = ST->isNonTemporal();
10586 AAMDNodes AAInfo = ST->getAAInfo();
10588 SDValue St0 = DAG.getStore(Chain, SDLoc(ST), Lo,
10589 Ptr, ST->getPointerInfo(),
10590 isVolatile, isNonTemporal,
10591 ST->getAlignment(), AAInfo);
10592 Ptr = DAG.getNode(ISD::ADD, SDLoc(N), Ptr.getValueType(), Ptr,
10593 DAG.getConstant(4, Ptr.getValueType()));
10594 Alignment = MinAlign(Alignment, 4U);
10595 SDValue St1 = DAG.getStore(Chain, SDLoc(ST), Hi,
10596 Ptr, ST->getPointerInfo().getWithOffset(4),
10597 isVolatile, isNonTemporal,
10598 Alignment, AAInfo);
10599 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other,
10608 // Try to infer better alignment information than the store already has.
10609 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
10610 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
10611 if (Align > ST->getAlignment())
10612 return DAG.getTruncStore(Chain, SDLoc(N), Value,
10613 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
10614 ST->isVolatile(), ST->isNonTemporal(), Align,
10619 // Try transforming a pair floating point load / store ops to integer
10620 // load / store ops.
10621 SDValue NewST = TransformFPLoadStorePair(N);
10622 if (NewST.getNode())
10625 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
10626 : DAG.getSubtarget().useAA();
10628 if (CombinerAAOnlyFunc.getNumOccurrences() &&
10629 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
10632 if (UseAA && ST->isUnindexed()) {
10633 // Walk up chain skipping non-aliasing memory nodes.
10634 SDValue BetterChain = FindBetterChain(N, Chain);
10636 // If there is a better chain.
10637 if (Chain != BetterChain) {
10640 // Replace the chain to avoid dependency.
10641 if (ST->isTruncatingStore()) {
10642 ReplStore = DAG.getTruncStore(BetterChain, SDLoc(N), Value, Ptr,
10643 ST->getMemoryVT(), ST->getMemOperand());
10645 ReplStore = DAG.getStore(BetterChain, SDLoc(N), Value, Ptr,
10646 ST->getMemOperand());
10649 // Create token to keep both nodes around.
10650 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
10651 MVT::Other, Chain, ReplStore);
10653 // Make sure the new and old chains are cleaned up.
10654 AddToWorklist(Token.getNode());
10656 // Don't add users to work list.
10657 return CombineTo(N, Token, false);
10661 // Try transforming N to an indexed store.
10662 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
10663 return SDValue(N, 0);
10665 // FIXME: is there such a thing as a truncating indexed store?
10666 if (ST->isTruncatingStore() && ST->isUnindexed() &&
10667 Value.getValueType().isInteger()) {
10668 // See if we can simplify the input to this truncstore with knowledge that
10669 // only the low bits are being used. For example:
10670 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
10672 GetDemandedBits(Value,
10673 APInt::getLowBitsSet(
10674 Value.getValueType().getScalarType().getSizeInBits(),
10675 ST->getMemoryVT().getScalarType().getSizeInBits()));
10676 AddToWorklist(Value.getNode());
10677 if (Shorter.getNode())
10678 return DAG.getTruncStore(Chain, SDLoc(N), Shorter,
10679 Ptr, ST->getMemoryVT(), ST->getMemOperand());
10681 // Otherwise, see if we can simplify the operation with
10682 // SimplifyDemandedBits, which only works if the value has a single use.
10683 if (SimplifyDemandedBits(Value,
10684 APInt::getLowBitsSet(
10685 Value.getValueType().getScalarType().getSizeInBits(),
10686 ST->getMemoryVT().getScalarType().getSizeInBits())))
10687 return SDValue(N, 0);
10690 // If this is a load followed by a store to the same location, then the store
10692 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
10693 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
10694 ST->isUnindexed() && !ST->isVolatile() &&
10695 // There can't be any side effects between the load and store, such as
10696 // a call or store.
10697 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
10698 // The store is dead, remove it.
10703 // If this is a store followed by a store with the same value to the same
10704 // location, then the store is dead/noop.
10705 if (StoreSDNode *ST1 = dyn_cast<StoreSDNode>(Chain)) {
10706 if (ST1->getBasePtr() == Ptr && ST->getMemoryVT() == ST1->getMemoryVT() &&
10707 ST1->getValue() == Value && ST->isUnindexed() && !ST->isVolatile() &&
10708 ST1->isUnindexed() && !ST1->isVolatile()) {
10709 // The store is dead, remove it.
10714 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
10715 // truncating store. We can do this even if this is already a truncstore.
10716 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
10717 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
10718 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
10719 ST->getMemoryVT())) {
10720 return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0),
10721 Ptr, ST->getMemoryVT(), ST->getMemOperand());
10724 // Only perform this optimization before the types are legal, because we
10725 // don't want to perform this optimization on every DAGCombine invocation.
10727 bool EverChanged = false;
10730 // There can be multiple store sequences on the same chain.
10731 // Keep trying to merge store sequences until we are unable to do so
10732 // or until we merge the last store on the chain.
10733 bool Changed = MergeConsecutiveStores(ST);
10734 EverChanged |= Changed;
10735 if (!Changed) break;
10736 } while (ST->getOpcode() != ISD::DELETED_NODE);
10739 return SDValue(N, 0);
10742 return ReduceLoadOpStoreWidth(N);
10745 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
10746 SDValue InVec = N->getOperand(0);
10747 SDValue InVal = N->getOperand(1);
10748 SDValue EltNo = N->getOperand(2);
10751 // If the inserted element is an UNDEF, just use the input vector.
10752 if (InVal.getOpcode() == ISD::UNDEF)
10755 EVT VT = InVec.getValueType();
10757 // If we can't generate a legal BUILD_VECTOR, exit
10758 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
10761 // Check that we know which element is being inserted
10762 if (!isa<ConstantSDNode>(EltNo))
10764 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
10766 // Canonicalize insert_vector_elt dag nodes.
10768 // (insert_vector_elt (insert_vector_elt A, Idx0), Idx1)
10769 // -> (insert_vector_elt (insert_vector_elt A, Idx1), Idx0)
10771 // Do this only if the child insert_vector node has one use; also
10772 // do this only if indices are both constants and Idx1 < Idx0.
10773 if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT && InVec.hasOneUse()
10774 && isa<ConstantSDNode>(InVec.getOperand(2))) {
10775 unsigned OtherElt =
10776 cast<ConstantSDNode>(InVec.getOperand(2))->getZExtValue();
10777 if (Elt < OtherElt) {
10779 SDValue NewOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(N), VT,
10780 InVec.getOperand(0), InVal, EltNo);
10781 AddToWorklist(NewOp.getNode());
10782 return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(InVec.getNode()),
10783 VT, NewOp, InVec.getOperand(1), InVec.getOperand(2));
10787 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
10788 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
10789 // vector elements.
10790 SmallVector<SDValue, 8> Ops;
10791 // Do not combine these two vectors if the output vector will not replace
10792 // the input vector.
10793 if (InVec.getOpcode() == ISD::BUILD_VECTOR && InVec.hasOneUse()) {
10794 Ops.append(InVec.getNode()->op_begin(),
10795 InVec.getNode()->op_end());
10796 } else if (InVec.getOpcode() == ISD::UNDEF) {
10797 unsigned NElts = VT.getVectorNumElements();
10798 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
10803 // Insert the element
10804 if (Elt < Ops.size()) {
10805 // All the operands of BUILD_VECTOR must have the same type;
10806 // we enforce that here.
10807 EVT OpVT = Ops[0].getValueType();
10808 if (InVal.getValueType() != OpVT)
10809 InVal = OpVT.bitsGT(InVal.getValueType()) ?
10810 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
10811 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
10815 // Return the new vector
10816 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
10819 SDValue DAGCombiner::ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
10820 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad) {
10821 EVT ResultVT = EVE->getValueType(0);
10822 EVT VecEltVT = InVecVT.getVectorElementType();
10823 unsigned Align = OriginalLoad->getAlignment();
10824 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
10825 VecEltVT.getTypeForEVT(*DAG.getContext()));
10827 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VecEltVT))
10832 SDValue NewPtr = OriginalLoad->getBasePtr();
10834 EVT PtrType = NewPtr.getValueType();
10835 MachinePointerInfo MPI;
10836 if (auto *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo)) {
10837 int Elt = ConstEltNo->getZExtValue();
10838 unsigned PtrOff = VecEltVT.getSizeInBits() * Elt / 8;
10839 if (TLI.isBigEndian())
10840 PtrOff = InVecVT.getSizeInBits() / 8 - PtrOff;
10841 Offset = DAG.getConstant(PtrOff, PtrType);
10842 MPI = OriginalLoad->getPointerInfo().getWithOffset(PtrOff);
10844 Offset = DAG.getNode(
10845 ISD::MUL, SDLoc(EVE), EltNo.getValueType(), EltNo,
10846 DAG.getConstant(VecEltVT.getStoreSize(), EltNo.getValueType()));
10847 if (TLI.isBigEndian())
10848 Offset = DAG.getNode(
10849 ISD::SUB, SDLoc(EVE), EltNo.getValueType(),
10850 DAG.getConstant(InVecVT.getStoreSize(), EltNo.getValueType()), Offset);
10851 MPI = OriginalLoad->getPointerInfo();
10853 NewPtr = DAG.getNode(ISD::ADD, SDLoc(EVE), PtrType, NewPtr, Offset);
10855 // The replacement we need to do here is a little tricky: we need to
10856 // replace an extractelement of a load with a load.
10857 // Use ReplaceAllUsesOfValuesWith to do the replacement.
10858 // Note that this replacement assumes that the extractvalue is the only
10859 // use of the load; that's okay because we don't want to perform this
10860 // transformation in other cases anyway.
10863 if (ResultVT.bitsGT(VecEltVT)) {
10864 // If the result type of vextract is wider than the load, then issue an
10865 // extending load instead.
10866 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, ResultVT,
10870 Load = DAG.getExtLoad(
10871 ExtType, SDLoc(EVE), ResultVT, OriginalLoad->getChain(), NewPtr, MPI,
10872 VecEltVT, OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
10873 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
10874 Chain = Load.getValue(1);
10876 Load = DAG.getLoad(
10877 VecEltVT, SDLoc(EVE), OriginalLoad->getChain(), NewPtr, MPI,
10878 OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
10879 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
10880 Chain = Load.getValue(1);
10881 if (ResultVT.bitsLT(VecEltVT))
10882 Load = DAG.getNode(ISD::TRUNCATE, SDLoc(EVE), ResultVT, Load);
10884 Load = DAG.getNode(ISD::BITCAST, SDLoc(EVE), ResultVT, Load);
10886 WorklistRemover DeadNodes(*this);
10887 SDValue From[] = { SDValue(EVE, 0), SDValue(OriginalLoad, 1) };
10888 SDValue To[] = { Load, Chain };
10889 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
10890 // Since we're explicitly calling ReplaceAllUses, add the new node to the
10891 // worklist explicitly as well.
10892 AddToWorklist(Load.getNode());
10893 AddUsersToWorklist(Load.getNode()); // Add users too
10894 // Make sure to revisit this node to clean it up; it will usually be dead.
10895 AddToWorklist(EVE);
10897 return SDValue(EVE, 0);
10900 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
10901 // (vextract (scalar_to_vector val, 0) -> val
10902 SDValue InVec = N->getOperand(0);
10903 EVT VT = InVec.getValueType();
10904 EVT NVT = N->getValueType(0);
10906 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
10907 // Check if the result type doesn't match the inserted element type. A
10908 // SCALAR_TO_VECTOR may truncate the inserted element and the
10909 // EXTRACT_VECTOR_ELT may widen the extracted vector.
10910 SDValue InOp = InVec.getOperand(0);
10911 if (InOp.getValueType() != NVT) {
10912 assert(InOp.getValueType().isInteger() && NVT.isInteger());
10913 return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT);
10918 SDValue EltNo = N->getOperand(1);
10919 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
10921 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
10922 // We only perform this optimization before the op legalization phase because
10923 // we may introduce new vector instructions which are not backed by TD
10924 // patterns. For example on AVX, extracting elements from a wide vector
10925 // without using extract_subvector. However, if we can find an underlying
10926 // scalar value, then we can always use that.
10927 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
10929 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
10930 int NumElem = VT.getVectorNumElements();
10931 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
10932 // Find the new index to extract from.
10933 int OrigElt = SVOp->getMaskElt(Elt);
10935 // Extracting an undef index is undef.
10937 return DAG.getUNDEF(NVT);
10939 // Select the right vector half to extract from.
10941 if (OrigElt < NumElem) {
10942 SVInVec = InVec->getOperand(0);
10944 SVInVec = InVec->getOperand(1);
10945 OrigElt -= NumElem;
10948 if (SVInVec.getOpcode() == ISD::BUILD_VECTOR) {
10949 SDValue InOp = SVInVec.getOperand(OrigElt);
10950 if (InOp.getValueType() != NVT) {
10951 assert(InOp.getValueType().isInteger() && NVT.isInteger());
10952 InOp = DAG.getSExtOrTrunc(InOp, SDLoc(SVInVec), NVT);
10958 // FIXME: We should handle recursing on other vector shuffles and
10959 // scalar_to_vector here as well.
10961 if (!LegalOperations) {
10962 EVT IndexTy = TLI.getVectorIdxTy();
10963 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT,
10964 SVInVec, DAG.getConstant(OrigElt, IndexTy));
10968 bool BCNumEltsChanged = false;
10969 EVT ExtVT = VT.getVectorElementType();
10972 // If the result of load has to be truncated, then it's not necessarily
10974 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
10977 if (InVec.getOpcode() == ISD::BITCAST) {
10978 // Don't duplicate a load with other uses.
10979 if (!InVec.hasOneUse())
10982 EVT BCVT = InVec.getOperand(0).getValueType();
10983 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
10985 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
10986 BCNumEltsChanged = true;
10987 InVec = InVec.getOperand(0);
10988 ExtVT = BCVT.getVectorElementType();
10991 // (vextract (vN[if]M load $addr), i) -> ([if]M load $addr + i * size)
10992 if (!LegalOperations && !ConstEltNo && InVec.hasOneUse() &&
10993 ISD::isNormalLoad(InVec.getNode()) &&
10994 !N->getOperand(1)->hasPredecessor(InVec.getNode())) {
10995 SDValue Index = N->getOperand(1);
10996 if (LoadSDNode *OrigLoad = dyn_cast<LoadSDNode>(InVec))
10997 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, Index,
11001 // Perform only after legalization to ensure build_vector / vector_shuffle
11002 // optimizations have already been done.
11003 if (!LegalOperations) return SDValue();
11005 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
11006 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
11007 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
11010 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
11012 LoadSDNode *LN0 = nullptr;
11013 const ShuffleVectorSDNode *SVN = nullptr;
11014 if (ISD::isNormalLoad(InVec.getNode())) {
11015 LN0 = cast<LoadSDNode>(InVec);
11016 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
11017 InVec.getOperand(0).getValueType() == ExtVT &&
11018 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
11019 // Don't duplicate a load with other uses.
11020 if (!InVec.hasOneUse())
11023 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
11024 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
11025 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
11027 // (load $addr+1*size)
11029 // Don't duplicate a load with other uses.
11030 if (!InVec.hasOneUse())
11033 // If the bit convert changed the number of elements, it is unsafe
11034 // to examine the mask.
11035 if (BCNumEltsChanged)
11038 // Select the input vector, guarding against out of range extract vector.
11039 unsigned NumElems = VT.getVectorNumElements();
11040 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
11041 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
11043 if (InVec.getOpcode() == ISD::BITCAST) {
11044 // Don't duplicate a load with other uses.
11045 if (!InVec.hasOneUse())
11048 InVec = InVec.getOperand(0);
11050 if (ISD::isNormalLoad(InVec.getNode())) {
11051 LN0 = cast<LoadSDNode>(InVec);
11052 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
11053 EltNo = DAG.getConstant(Elt, EltNo.getValueType());
11057 // Make sure we found a non-volatile load and the extractelement is
11059 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
11062 // If Idx was -1 above, Elt is going to be -1, so just return undef.
11064 return DAG.getUNDEF(LVT);
11066 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, EltNo, LN0);
11072 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
11073 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
11074 // We perform this optimization post type-legalization because
11075 // the type-legalizer often scalarizes integer-promoted vectors.
11076 // Performing this optimization before may create bit-casts which
11077 // will be type-legalized to complex code sequences.
11078 // We perform this optimization only before the operation legalizer because we
11079 // may introduce illegal operations.
11080 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
11083 unsigned NumInScalars = N->getNumOperands();
11085 EVT VT = N->getValueType(0);
11087 // Check to see if this is a BUILD_VECTOR of a bunch of values
11088 // which come from any_extend or zero_extend nodes. If so, we can create
11089 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
11090 // optimizations. We do not handle sign-extend because we can't fill the sign
11092 EVT SourceType = MVT::Other;
11093 bool AllAnyExt = true;
11095 for (unsigned i = 0; i != NumInScalars; ++i) {
11096 SDValue In = N->getOperand(i);
11097 // Ignore undef inputs.
11098 if (In.getOpcode() == ISD::UNDEF) continue;
11100 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
11101 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
11103 // Abort if the element is not an extension.
11104 if (!ZeroExt && !AnyExt) {
11105 SourceType = MVT::Other;
11109 // The input is a ZeroExt or AnyExt. Check the original type.
11110 EVT InTy = In.getOperand(0).getValueType();
11112 // Check that all of the widened source types are the same.
11113 if (SourceType == MVT::Other)
11116 else if (InTy != SourceType) {
11117 // Multiple income types. Abort.
11118 SourceType = MVT::Other;
11122 // Check if all of the extends are ANY_EXTENDs.
11123 AllAnyExt &= AnyExt;
11126 // In order to have valid types, all of the inputs must be extended from the
11127 // same source type and all of the inputs must be any or zero extend.
11128 // Scalar sizes must be a power of two.
11129 EVT OutScalarTy = VT.getScalarType();
11130 bool ValidTypes = SourceType != MVT::Other &&
11131 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
11132 isPowerOf2_32(SourceType.getSizeInBits());
11134 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
11135 // turn into a single shuffle instruction.
11139 bool isLE = TLI.isLittleEndian();
11140 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
11141 assert(ElemRatio > 1 && "Invalid element size ratio");
11142 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
11143 DAG.getConstant(0, SourceType);
11145 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
11146 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
11148 // Populate the new build_vector
11149 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
11150 SDValue Cast = N->getOperand(i);
11151 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
11152 Cast.getOpcode() == ISD::ZERO_EXTEND ||
11153 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
11155 if (Cast.getOpcode() == ISD::UNDEF)
11156 In = DAG.getUNDEF(SourceType);
11158 In = Cast->getOperand(0);
11159 unsigned Index = isLE ? (i * ElemRatio) :
11160 (i * ElemRatio + (ElemRatio - 1));
11162 assert(Index < Ops.size() && "Invalid index");
11166 // The type of the new BUILD_VECTOR node.
11167 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
11168 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
11169 "Invalid vector size");
11170 // Check if the new vector type is legal.
11171 if (!isTypeLegal(VecVT)) return SDValue();
11173 // Make the new BUILD_VECTOR.
11174 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
11176 // The new BUILD_VECTOR node has the potential to be further optimized.
11177 AddToWorklist(BV.getNode());
11178 // Bitcast to the desired type.
11179 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
11182 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
11183 EVT VT = N->getValueType(0);
11185 unsigned NumInScalars = N->getNumOperands();
11188 EVT SrcVT = MVT::Other;
11189 unsigned Opcode = ISD::DELETED_NODE;
11190 unsigned NumDefs = 0;
11192 for (unsigned i = 0; i != NumInScalars; ++i) {
11193 SDValue In = N->getOperand(i);
11194 unsigned Opc = In.getOpcode();
11196 if (Opc == ISD::UNDEF)
11199 // If all scalar values are floats and converted from integers.
11200 if (Opcode == ISD::DELETED_NODE &&
11201 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
11208 EVT InVT = In.getOperand(0).getValueType();
11210 // If all scalar values are typed differently, bail out. It's chosen to
11211 // simplify BUILD_VECTOR of integer types.
11212 if (SrcVT == MVT::Other)
11219 // If the vector has just one element defined, it's not worth to fold it into
11220 // a vectorized one.
11224 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
11225 && "Should only handle conversion from integer to float.");
11226 assert(SrcVT != MVT::Other && "Cannot determine source type!");
11228 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
11230 if (!TLI.isOperationLegalOrCustom(Opcode, NVT))
11233 // Just because the floating-point vector type is legal does not necessarily
11234 // mean that the corresponding integer vector type is.
11235 if (!isTypeLegal(NVT))
11238 SmallVector<SDValue, 8> Opnds;
11239 for (unsigned i = 0; i != NumInScalars; ++i) {
11240 SDValue In = N->getOperand(i);
11242 if (In.getOpcode() == ISD::UNDEF)
11243 Opnds.push_back(DAG.getUNDEF(SrcVT));
11245 Opnds.push_back(In.getOperand(0));
11247 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Opnds);
11248 AddToWorklist(BV.getNode());
11250 return DAG.getNode(Opcode, dl, VT, BV);
11253 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
11254 unsigned NumInScalars = N->getNumOperands();
11256 EVT VT = N->getValueType(0);
11258 // A vector built entirely of undefs is undef.
11259 if (ISD::allOperandsUndef(N))
11260 return DAG.getUNDEF(VT);
11262 SDValue V = reduceBuildVecExtToExtBuildVec(N);
11266 V = reduceBuildVecConvertToConvertBuildVec(N);
11270 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
11271 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
11272 // at most two distinct vectors, turn this into a shuffle node.
11274 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
11275 if (!isTypeLegal(VT))
11278 // May only combine to shuffle after legalize if shuffle is legal.
11279 if (LegalOperations && !TLI.isOperationLegal(ISD::VECTOR_SHUFFLE, VT))
11282 SDValue VecIn1, VecIn2;
11283 bool UsesZeroVector = false;
11284 for (unsigned i = 0; i != NumInScalars; ++i) {
11285 SDValue Op = N->getOperand(i);
11286 // Ignore undef inputs.
11287 if (Op.getOpcode() == ISD::UNDEF) continue;
11289 // See if we can combine this build_vector into a blend with a zero vector.
11290 if (!VecIn2.getNode() && ((Op.getOpcode() == ISD::Constant &&
11291 cast<ConstantSDNode>(Op.getNode())->isNullValue()) ||
11292 (Op.getOpcode() == ISD::ConstantFP &&
11293 cast<ConstantFPSDNode>(Op.getNode())->getValueAPF().isZero()))) {
11294 UsesZeroVector = true;
11298 // If this input is something other than a EXTRACT_VECTOR_ELT with a
11299 // constant index, bail out.
11300 if (Op.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
11301 !isa<ConstantSDNode>(Op.getOperand(1))) {
11302 VecIn1 = VecIn2 = SDValue(nullptr, 0);
11306 // We allow up to two distinct input vectors.
11307 SDValue ExtractedFromVec = Op.getOperand(0);
11308 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
11311 if (!VecIn1.getNode()) {
11312 VecIn1 = ExtractedFromVec;
11313 } else if (!VecIn2.getNode() && !UsesZeroVector) {
11314 VecIn2 = ExtractedFromVec;
11316 // Too many inputs.
11317 VecIn1 = VecIn2 = SDValue(nullptr, 0);
11322 // If everything is good, we can make a shuffle operation.
11323 if (VecIn1.getNode()) {
11324 unsigned InNumElements = VecIn1.getValueType().getVectorNumElements();
11325 SmallVector<int, 8> Mask;
11326 for (unsigned i = 0; i != NumInScalars; ++i) {
11327 unsigned Opcode = N->getOperand(i).getOpcode();
11328 if (Opcode == ISD::UNDEF) {
11329 Mask.push_back(-1);
11333 // Operands can also be zero.
11334 if (Opcode != ISD::EXTRACT_VECTOR_ELT) {
11335 assert(UsesZeroVector &&
11336 (Opcode == ISD::Constant || Opcode == ISD::ConstantFP) &&
11337 "Unexpected node found!");
11338 Mask.push_back(NumInScalars+i);
11342 // If extracting from the first vector, just use the index directly.
11343 SDValue Extract = N->getOperand(i);
11344 SDValue ExtVal = Extract.getOperand(1);
11345 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
11346 if (Extract.getOperand(0) == VecIn1) {
11347 Mask.push_back(ExtIndex);
11351 // Otherwise, use InIdx + InputVecSize
11352 Mask.push_back(InNumElements + ExtIndex);
11355 // Avoid introducing illegal shuffles with zero.
11356 if (UsesZeroVector && !TLI.isVectorClearMaskLegal(Mask, VT))
11359 // We can't generate a shuffle node with mismatched input and output types.
11360 // Attempt to transform a single input vector to the correct type.
11361 if ((VT != VecIn1.getValueType())) {
11362 // If the input vector type has a different base type to the output
11363 // vector type, bail out.
11364 EVT VTElemType = VT.getVectorElementType();
11365 if ((VecIn1.getValueType().getVectorElementType() != VTElemType) ||
11366 (VecIn2.getNode() &&
11367 (VecIn2.getValueType().getVectorElementType() != VTElemType)))
11370 // If the input vector is too small, widen it.
11371 // We only support widening of vectors which are half the size of the
11372 // output registers. For example XMM->YMM widening on X86 with AVX.
11373 EVT VecInT = VecIn1.getValueType();
11374 if (VecInT.getSizeInBits() * 2 == VT.getSizeInBits()) {
11375 // If we only have one small input, widen it by adding undef values.
11376 if (!VecIn2.getNode())
11377 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, VecIn1,
11378 DAG.getUNDEF(VecIn1.getValueType()));
11379 else if (VecIn1.getValueType() == VecIn2.getValueType()) {
11380 // If we have two small inputs of the same type, try to concat them.
11381 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, VecIn1, VecIn2);
11382 VecIn2 = SDValue(nullptr, 0);
11385 } else if (VecInT.getSizeInBits() == VT.getSizeInBits() * 2) {
11386 // If the input vector is too large, try to split it.
11387 // We don't support having two input vectors that are too large.
11388 // If the zero vector was used, we can not split the vector,
11389 // since we'd need 3 inputs.
11390 if (UsesZeroVector || VecIn2.getNode())
11393 if (!TLI.isExtractSubvectorCheap(VT, VT.getVectorNumElements()))
11396 // Try to replace VecIn1 with two extract_subvectors
11397 // No need to update the masks, they should still be correct.
11398 VecIn2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, VecIn1,
11399 DAG.getConstant(VT.getVectorNumElements(), TLI.getVectorIdxTy()));
11400 VecIn1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, VecIn1,
11401 DAG.getConstant(0, TLI.getVectorIdxTy()));
11406 if (UsesZeroVector)
11407 VecIn2 = VT.isInteger() ? DAG.getConstant(0, VT) :
11408 DAG.getConstantFP(0.0, VT);
11410 // If VecIn2 is unused then change it to undef.
11411 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
11413 // Check that we were able to transform all incoming values to the same
11415 if (VecIn2.getValueType() != VecIn1.getValueType() ||
11416 VecIn1.getValueType() != VT)
11419 // Return the new VECTOR_SHUFFLE node.
11423 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]);
11429 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
11430 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
11431 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
11432 // inputs come from at most two distinct vectors, turn this into a shuffle
11435 // If we only have one input vector, we don't need to do any concatenation.
11436 if (N->getNumOperands() == 1)
11437 return N->getOperand(0);
11439 // Check if all of the operands are undefs.
11440 EVT VT = N->getValueType(0);
11441 if (ISD::allOperandsUndef(N))
11442 return DAG.getUNDEF(VT);
11444 // Optimize concat_vectors where one of the vectors is undef.
11445 if (N->getNumOperands() == 2 &&
11446 N->getOperand(1)->getOpcode() == ISD::UNDEF) {
11447 SDValue In = N->getOperand(0);
11448 assert(In.getValueType().isVector() && "Must concat vectors");
11450 // Transform: concat_vectors(scalar, undef) -> scalar_to_vector(sclr).
11451 if (In->getOpcode() == ISD::BITCAST &&
11452 !In->getOperand(0)->getValueType(0).isVector()) {
11453 SDValue Scalar = In->getOperand(0);
11454 EVT SclTy = Scalar->getValueType(0);
11456 if (!SclTy.isFloatingPoint() && !SclTy.isInteger())
11459 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SclTy,
11460 VT.getSizeInBits() / SclTy.getSizeInBits());
11461 if (!TLI.isTypeLegal(NVT) || !TLI.isTypeLegal(Scalar.getValueType()))
11464 SDLoc dl = SDLoc(N);
11465 SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NVT, Scalar);
11466 return DAG.getNode(ISD::BITCAST, dl, VT, Res);
11470 // Fold any combination of BUILD_VECTOR or UNDEF nodes into one BUILD_VECTOR.
11471 // We have already tested above for an UNDEF only concatenation.
11472 // fold (concat_vectors (BUILD_VECTOR A, B, ...), (BUILD_VECTOR C, D, ...))
11473 // -> (BUILD_VECTOR A, B, ..., C, D, ...)
11474 auto IsBuildVectorOrUndef = [](const SDValue &Op) {
11475 return ISD::UNDEF == Op.getOpcode() || ISD::BUILD_VECTOR == Op.getOpcode();
11477 bool AllBuildVectorsOrUndefs =
11478 std::all_of(N->op_begin(), N->op_end(), IsBuildVectorOrUndef);
11479 if (AllBuildVectorsOrUndefs) {
11480 SmallVector<SDValue, 8> Opnds;
11481 EVT SVT = VT.getScalarType();
11484 if (!SVT.isFloatingPoint()) {
11485 // If BUILD_VECTOR are from built from integer, they may have different
11486 // operand types. Get the smallest type and truncate all operands to it.
11487 bool FoundMinVT = false;
11488 for (const SDValue &Op : N->ops())
11489 if (ISD::BUILD_VECTOR == Op.getOpcode()) {
11490 EVT OpSVT = Op.getOperand(0)->getValueType(0);
11491 MinVT = (!FoundMinVT || OpSVT.bitsLE(MinVT)) ? OpSVT : MinVT;
11494 assert(FoundMinVT && "Concat vector type mismatch");
11497 for (const SDValue &Op : N->ops()) {
11498 EVT OpVT = Op.getValueType();
11499 unsigned NumElts = OpVT.getVectorNumElements();
11501 if (ISD::UNDEF == Op.getOpcode())
11502 Opnds.append(NumElts, DAG.getUNDEF(MinVT));
11504 if (ISD::BUILD_VECTOR == Op.getOpcode()) {
11505 if (SVT.isFloatingPoint()) {
11506 assert(SVT == OpVT.getScalarType() && "Concat vector type mismatch");
11507 Opnds.append(Op->op_begin(), Op->op_begin() + NumElts);
11509 for (unsigned i = 0; i != NumElts; ++i)
11511 DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinVT, Op.getOperand(i)));
11516 assert(VT.getVectorNumElements() == Opnds.size() &&
11517 "Concat vector type mismatch");
11518 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
11521 // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
11522 // nodes often generate nop CONCAT_VECTOR nodes.
11523 // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that
11524 // place the incoming vectors at the exact same location.
11525 SDValue SingleSource = SDValue();
11526 unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements();
11528 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
11529 SDValue Op = N->getOperand(i);
11531 if (Op.getOpcode() == ISD::UNDEF)
11534 // Check if this is the identity extract:
11535 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
11538 // Find the single incoming vector for the extract_subvector.
11539 if (SingleSource.getNode()) {
11540 if (Op.getOperand(0) != SingleSource)
11543 SingleSource = Op.getOperand(0);
11545 // Check the source type is the same as the type of the result.
11546 // If not, this concat may extend the vector, so we can not
11547 // optimize it away.
11548 if (SingleSource.getValueType() != N->getValueType(0))
11552 unsigned IdentityIndex = i * PartNumElem;
11553 ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1));
11554 // The extract index must be constant.
11558 // Check that we are reading from the identity index.
11559 if (CS->getZExtValue() != IdentityIndex)
11563 if (SingleSource.getNode())
11564 return SingleSource;
11569 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
11570 EVT NVT = N->getValueType(0);
11571 SDValue V = N->getOperand(0);
11573 if (V->getOpcode() == ISD::CONCAT_VECTORS) {
11575 // (extract_subvec (concat V1, V2, ...), i)
11578 // Only operand 0 is checked as 'concat' assumes all inputs of the same
11580 if (V->getOperand(0).getValueType() != NVT)
11582 unsigned Idx = dyn_cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
11583 unsigned NumElems = NVT.getVectorNumElements();
11584 assert((Idx % NumElems) == 0 &&
11585 "IDX in concat is not a multiple of the result vector length.");
11586 return V->getOperand(Idx / NumElems);
11590 if (V->getOpcode() == ISD::BITCAST)
11591 V = V.getOperand(0);
11593 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
11595 // Handle only simple case where vector being inserted and vector
11596 // being extracted are of same type, and are half size of larger vectors.
11597 EVT BigVT = V->getOperand(0).getValueType();
11598 EVT SmallVT = V->getOperand(1).getValueType();
11599 if (!NVT.bitsEq(SmallVT) || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
11602 // Only handle cases where both indexes are constants with the same type.
11603 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
11604 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
11606 if (InsIdx && ExtIdx &&
11607 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
11608 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
11610 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
11612 // indices are equal or bit offsets are equal => V1
11613 // otherwise => (extract_subvec V1, ExtIdx)
11614 if (InsIdx->getZExtValue() * SmallVT.getScalarType().getSizeInBits() ==
11615 ExtIdx->getZExtValue() * NVT.getScalarType().getSizeInBits())
11616 return DAG.getNode(ISD::BITCAST, dl, NVT, V->getOperand(1));
11617 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT,
11618 DAG.getNode(ISD::BITCAST, dl,
11619 N->getOperand(0).getValueType(),
11620 V->getOperand(0)), N->getOperand(1));
11627 static SDValue simplifyShuffleOperandRecursively(SmallBitVector &UsedElements,
11628 SDValue V, SelectionDAG &DAG) {
11630 EVT VT = V.getValueType();
11632 switch (V.getOpcode()) {
11636 case ISD::CONCAT_VECTORS: {
11637 EVT OpVT = V->getOperand(0).getValueType();
11638 int OpSize = OpVT.getVectorNumElements();
11639 SmallBitVector OpUsedElements(OpSize, false);
11640 bool FoundSimplification = false;
11641 SmallVector<SDValue, 4> NewOps;
11642 NewOps.reserve(V->getNumOperands());
11643 for (int i = 0, NumOps = V->getNumOperands(); i < NumOps; ++i) {
11644 SDValue Op = V->getOperand(i);
11645 bool OpUsed = false;
11646 for (int j = 0; j < OpSize; ++j)
11647 if (UsedElements[i * OpSize + j]) {
11648 OpUsedElements[j] = true;
11652 OpUsed ? simplifyShuffleOperandRecursively(OpUsedElements, Op, DAG)
11653 : DAG.getUNDEF(OpVT));
11654 FoundSimplification |= Op == NewOps.back();
11655 OpUsedElements.reset();
11657 if (FoundSimplification)
11658 V = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, NewOps);
11662 case ISD::INSERT_SUBVECTOR: {
11663 SDValue BaseV = V->getOperand(0);
11664 SDValue SubV = V->getOperand(1);
11665 auto *IdxN = dyn_cast<ConstantSDNode>(V->getOperand(2));
11669 int SubSize = SubV.getValueType().getVectorNumElements();
11670 int Idx = IdxN->getZExtValue();
11671 bool SubVectorUsed = false;
11672 SmallBitVector SubUsedElements(SubSize, false);
11673 for (int i = 0; i < SubSize; ++i)
11674 if (UsedElements[i + Idx]) {
11675 SubVectorUsed = true;
11676 SubUsedElements[i] = true;
11677 UsedElements[i + Idx] = false;
11680 // Now recurse on both the base and sub vectors.
11681 SDValue SimplifiedSubV =
11683 ? simplifyShuffleOperandRecursively(SubUsedElements, SubV, DAG)
11684 : DAG.getUNDEF(SubV.getValueType());
11685 SDValue SimplifiedBaseV = simplifyShuffleOperandRecursively(UsedElements, BaseV, DAG);
11686 if (SimplifiedSubV != SubV || SimplifiedBaseV != BaseV)
11687 V = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT,
11688 SimplifiedBaseV, SimplifiedSubV, V->getOperand(2));
11694 static SDValue simplifyShuffleOperands(ShuffleVectorSDNode *SVN, SDValue N0,
11695 SDValue N1, SelectionDAG &DAG) {
11696 EVT VT = SVN->getValueType(0);
11697 int NumElts = VT.getVectorNumElements();
11698 SmallBitVector N0UsedElements(NumElts, false), N1UsedElements(NumElts, false);
11699 for (int M : SVN->getMask())
11700 if (M >= 0 && M < NumElts)
11701 N0UsedElements[M] = true;
11702 else if (M >= NumElts)
11703 N1UsedElements[M - NumElts] = true;
11705 SDValue S0 = simplifyShuffleOperandRecursively(N0UsedElements, N0, DAG);
11706 SDValue S1 = simplifyShuffleOperandRecursively(N1UsedElements, N1, DAG);
11707 if (S0 == N0 && S1 == N1)
11710 return DAG.getVectorShuffle(VT, SDLoc(SVN), S0, S1, SVN->getMask());
11713 // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat,
11714 // or turn a shuffle of a single concat into simpler shuffle then concat.
11715 static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
11716 EVT VT = N->getValueType(0);
11717 unsigned NumElts = VT.getVectorNumElements();
11719 SDValue N0 = N->getOperand(0);
11720 SDValue N1 = N->getOperand(1);
11721 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
11723 SmallVector<SDValue, 4> Ops;
11724 EVT ConcatVT = N0.getOperand(0).getValueType();
11725 unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements();
11726 unsigned NumConcats = NumElts / NumElemsPerConcat;
11728 // Special case: shuffle(concat(A,B)) can be more efficiently represented
11729 // as concat(shuffle(A,B),UNDEF) if the shuffle doesn't set any of the high
11730 // half vector elements.
11731 if (NumElemsPerConcat * 2 == NumElts && N1.getOpcode() == ISD::UNDEF &&
11732 std::all_of(SVN->getMask().begin() + NumElemsPerConcat,
11733 SVN->getMask().end(), [](int i) { return i == -1; })) {
11734 N0 = DAG.getVectorShuffle(ConcatVT, SDLoc(N), N0.getOperand(0), N0.getOperand(1),
11735 ArrayRef<int>(SVN->getMask().begin(), NumElemsPerConcat));
11736 N1 = DAG.getUNDEF(ConcatVT);
11737 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, N0, N1);
11740 // Look at every vector that's inserted. We're looking for exact
11741 // subvector-sized copies from a concatenated vector
11742 for (unsigned I = 0; I != NumConcats; ++I) {
11743 // Make sure we're dealing with a copy.
11744 unsigned Begin = I * NumElemsPerConcat;
11745 bool AllUndef = true, NoUndef = true;
11746 for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) {
11747 if (SVN->getMaskElt(J) >= 0)
11754 if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0)
11757 for (unsigned J = 1; J != NumElemsPerConcat; ++J)
11758 if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J))
11761 unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat;
11762 if (FirstElt < N0.getNumOperands())
11763 Ops.push_back(N0.getOperand(FirstElt));
11765 Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands()));
11767 } else if (AllUndef) {
11768 Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType()));
11769 } else { // Mixed with general masks and undefs, can't do optimization.
11774 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops);
11777 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
11778 EVT VT = N->getValueType(0);
11779 unsigned NumElts = VT.getVectorNumElements();
11781 SDValue N0 = N->getOperand(0);
11782 SDValue N1 = N->getOperand(1);
11784 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
11786 // Canonicalize shuffle undef, undef -> undef
11787 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
11788 return DAG.getUNDEF(VT);
11790 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
11792 // Canonicalize shuffle v, v -> v, undef
11794 SmallVector<int, 8> NewMask;
11795 for (unsigned i = 0; i != NumElts; ++i) {
11796 int Idx = SVN->getMaskElt(i);
11797 if (Idx >= (int)NumElts) Idx -= NumElts;
11798 NewMask.push_back(Idx);
11800 return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT),
11804 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
11805 if (N0.getOpcode() == ISD::UNDEF) {
11806 SmallVector<int, 8> NewMask;
11807 for (unsigned i = 0; i != NumElts; ++i) {
11808 int Idx = SVN->getMaskElt(i);
11810 if (Idx >= (int)NumElts)
11813 Idx = -1; // remove reference to lhs
11815 NewMask.push_back(Idx);
11817 return DAG.getVectorShuffle(VT, SDLoc(N), N1, DAG.getUNDEF(VT),
11821 // Remove references to rhs if it is undef
11822 if (N1.getOpcode() == ISD::UNDEF) {
11823 bool Changed = false;
11824 SmallVector<int, 8> NewMask;
11825 for (unsigned i = 0; i != NumElts; ++i) {
11826 int Idx = SVN->getMaskElt(i);
11827 if (Idx >= (int)NumElts) {
11831 NewMask.push_back(Idx);
11834 return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, &NewMask[0]);
11837 // If it is a splat, check if the argument vector is another splat or a
11839 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
11840 SDNode *V = N0.getNode();
11842 // If this is a bit convert that changes the element type of the vector but
11843 // not the number of vector elements, look through it. Be careful not to
11844 // look though conversions that change things like v4f32 to v2f64.
11845 if (V->getOpcode() == ISD::BITCAST) {
11846 SDValue ConvInput = V->getOperand(0);
11847 if (ConvInput.getValueType().isVector() &&
11848 ConvInput.getValueType().getVectorNumElements() == NumElts)
11849 V = ConvInput.getNode();
11852 if (V->getOpcode() == ISD::BUILD_VECTOR) {
11853 assert(V->getNumOperands() == NumElts &&
11854 "BUILD_VECTOR has wrong number of operands");
11856 bool AllSame = true;
11857 for (unsigned i = 0; i != NumElts; ++i) {
11858 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
11859 Base = V->getOperand(i);
11863 // Splat of <u, u, u, u>, return <u, u, u, u>
11864 if (!Base.getNode())
11866 for (unsigned i = 0; i != NumElts; ++i) {
11867 if (V->getOperand(i) != Base) {
11872 // Splat of <x, x, x, x>, return <x, x, x, x>
11876 // Canonicalize any other splat as a build_vector.
11877 const SDValue &Splatted = V->getOperand(SVN->getSplatIndex());
11878 SmallVector<SDValue, 8> Ops(NumElts, Splatted);
11879 SDValue NewBV = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
11880 V->getValueType(0), Ops);
11882 // We may have jumped through bitcasts, so the type of the
11883 // BUILD_VECTOR may not match the type of the shuffle.
11884 if (V->getValueType(0) != VT)
11885 NewBV = DAG.getNode(ISD::BITCAST, SDLoc(N), VT, NewBV);
11890 // There are various patterns used to build up a vector from smaller vectors,
11891 // subvectors, or elements. Scan chains of these and replace unused insertions
11892 // or components with undef.
11893 if (SDValue S = simplifyShuffleOperands(SVN, N0, N1, DAG))
11896 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
11897 Level < AfterLegalizeVectorOps &&
11898 (N1.getOpcode() == ISD::UNDEF ||
11899 (N1.getOpcode() == ISD::CONCAT_VECTORS &&
11900 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) {
11901 SDValue V = partitionShuffleOfConcats(N, DAG);
11907 // If this shuffle only has a single input that is a bitcasted shuffle,
11908 // attempt to merge the 2 shuffles and suitably bitcast the inputs/output
11909 // back to their original types.
11910 if (N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
11911 N1.getOpcode() == ISD::UNDEF && Level < AfterLegalizeVectorOps &&
11912 TLI.isTypeLegal(VT)) {
11914 // Peek through the bitcast only if there is one user.
11916 while (BC0.getOpcode() == ISD::BITCAST) {
11917 if (!BC0.hasOneUse())
11919 BC0 = BC0.getOperand(0);
11922 auto ScaleShuffleMask = [](ArrayRef<int> Mask, int Scale) {
11924 return SmallVector<int, 8>(Mask.begin(), Mask.end());
11926 SmallVector<int, 8> NewMask;
11928 for (int s = 0; s != Scale; ++s)
11929 NewMask.push_back(M < 0 ? -1 : Scale * M + s);
11933 if (BC0.getOpcode() == ISD::VECTOR_SHUFFLE && BC0.hasOneUse()) {
11934 EVT SVT = VT.getScalarType();
11935 EVT InnerVT = BC0->getValueType(0);
11936 EVT InnerSVT = InnerVT.getScalarType();
11938 // Determine which shuffle works with the smaller scalar type.
11939 EVT ScaleVT = SVT.bitsLT(InnerSVT) ? VT : InnerVT;
11940 EVT ScaleSVT = ScaleVT.getScalarType();
11942 if (TLI.isTypeLegal(ScaleVT) &&
11943 0 == (InnerSVT.getSizeInBits() % ScaleSVT.getSizeInBits()) &&
11944 0 == (SVT.getSizeInBits() % ScaleSVT.getSizeInBits())) {
11946 int InnerScale = InnerSVT.getSizeInBits() / ScaleSVT.getSizeInBits();
11947 int OuterScale = SVT.getSizeInBits() / ScaleSVT.getSizeInBits();
11949 // Scale the shuffle masks to the smaller scalar type.
11950 ShuffleVectorSDNode *InnerSVN = cast<ShuffleVectorSDNode>(BC0);
11951 SmallVector<int, 8> InnerMask =
11952 ScaleShuffleMask(InnerSVN->getMask(), InnerScale);
11953 SmallVector<int, 8> OuterMask =
11954 ScaleShuffleMask(SVN->getMask(), OuterScale);
11956 // Merge the shuffle masks.
11957 SmallVector<int, 8> NewMask;
11958 for (int M : OuterMask)
11959 NewMask.push_back(M < 0 ? -1 : InnerMask[M]);
11961 // Test for shuffle mask legality over both commutations.
11962 SDValue SV0 = BC0->getOperand(0);
11963 SDValue SV1 = BC0->getOperand(1);
11964 bool LegalMask = TLI.isShuffleMaskLegal(NewMask, ScaleVT);
11966 for (int i = 0, e = (int)NewMask.size(); i != e; ++i) {
11967 int idx = NewMask[i];
11971 NewMask[i] = idx + e;
11973 NewMask[i] = idx - e;
11975 std::swap(SV0, SV1);
11976 LegalMask = TLI.isShuffleMaskLegal(NewMask, ScaleVT);
11980 SV0 = DAG.getNode(ISD::BITCAST, SDLoc(N), ScaleVT, SV0);
11981 SV1 = DAG.getNode(ISD::BITCAST, SDLoc(N), ScaleVT, SV1);
11982 return DAG.getNode(
11983 ISD::BITCAST, SDLoc(N), VT,
11984 DAG.getVectorShuffle(ScaleVT, SDLoc(N), SV0, SV1, NewMask));
11990 // Canonicalize shuffles according to rules:
11991 // shuffle(A, shuffle(A, B)) -> shuffle(shuffle(A,B), A)
11992 // shuffle(B, shuffle(A, B)) -> shuffle(shuffle(A,B), B)
11993 // shuffle(B, shuffle(A, Undef)) -> shuffle(shuffle(A, Undef), B)
11994 if (N1.getOpcode() == ISD::VECTOR_SHUFFLE &&
11995 N0.getOpcode() != ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
11996 TLI.isTypeLegal(VT)) {
11997 // The incoming shuffle must be of the same type as the result of the
11998 // current shuffle.
11999 assert(N1->getOperand(0).getValueType() == VT &&
12000 "Shuffle types don't match");
12002 SDValue SV0 = N1->getOperand(0);
12003 SDValue SV1 = N1->getOperand(1);
12004 bool HasSameOp0 = N0 == SV0;
12005 bool IsSV1Undef = SV1.getOpcode() == ISD::UNDEF;
12006 if (HasSameOp0 || IsSV1Undef || N0 == SV1)
12007 // Commute the operands of this shuffle so that next rule
12009 return DAG.getCommutedVectorShuffle(*SVN);
12012 // Try to fold according to rules:
12013 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2)
12014 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2)
12015 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2)
12016 // Don't try to fold shuffles with illegal type.
12017 // Only fold if this shuffle is the only user of the other shuffle.
12018 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && N->isOnlyUserOf(N0.getNode()) &&
12019 Level < AfterLegalizeDAG && TLI.isTypeLegal(VT)) {
12020 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
12022 // The incoming shuffle must be of the same type as the result of the
12023 // current shuffle.
12024 assert(OtherSV->getOperand(0).getValueType() == VT &&
12025 "Shuffle types don't match");
12028 SmallVector<int, 4> Mask;
12029 // Compute the combined shuffle mask for a shuffle with SV0 as the first
12030 // operand, and SV1 as the second operand.
12031 for (unsigned i = 0; i != NumElts; ++i) {
12032 int Idx = SVN->getMaskElt(i);
12034 // Propagate Undef.
12035 Mask.push_back(Idx);
12039 SDValue CurrentVec;
12040 if (Idx < (int)NumElts) {
12041 // This shuffle index refers to the inner shuffle N0. Lookup the inner
12042 // shuffle mask to identify which vector is actually referenced.
12043 Idx = OtherSV->getMaskElt(Idx);
12045 // Propagate Undef.
12046 Mask.push_back(Idx);
12050 CurrentVec = (Idx < (int) NumElts) ? OtherSV->getOperand(0)
12051 : OtherSV->getOperand(1);
12053 // This shuffle index references an element within N1.
12057 // Simple case where 'CurrentVec' is UNDEF.
12058 if (CurrentVec.getOpcode() == ISD::UNDEF) {
12059 Mask.push_back(-1);
12063 // Canonicalize the shuffle index. We don't know yet if CurrentVec
12064 // will be the first or second operand of the combined shuffle.
12065 Idx = Idx % NumElts;
12066 if (!SV0.getNode() || SV0 == CurrentVec) {
12067 // Ok. CurrentVec is the left hand side.
12068 // Update the mask accordingly.
12070 Mask.push_back(Idx);
12074 // Bail out if we cannot convert the shuffle pair into a single shuffle.
12075 if (SV1.getNode() && SV1 != CurrentVec)
12078 // Ok. CurrentVec is the right hand side.
12079 // Update the mask accordingly.
12081 Mask.push_back(Idx + NumElts);
12084 // Check if all indices in Mask are Undef. In case, propagate Undef.
12085 bool isUndefMask = true;
12086 for (unsigned i = 0; i != NumElts && isUndefMask; ++i)
12087 isUndefMask &= Mask[i] < 0;
12090 return DAG.getUNDEF(VT);
12092 if (!SV0.getNode())
12093 SV0 = DAG.getUNDEF(VT);
12094 if (!SV1.getNode())
12095 SV1 = DAG.getUNDEF(VT);
12097 // Avoid introducing shuffles with illegal mask.
12098 if (!TLI.isShuffleMaskLegal(Mask, VT)) {
12099 // Compute the commuted shuffle mask and test again.
12100 for (unsigned i = 0; i != NumElts; ++i) {
12104 else if (idx < (int)NumElts)
12105 Mask[i] = idx + NumElts;
12107 Mask[i] = idx - NumElts;
12110 if (!TLI.isShuffleMaskLegal(Mask, VT))
12113 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, A, M2)
12114 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, A, M2)
12115 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, B, M2)
12116 std::swap(SV0, SV1);
12119 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2)
12120 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2)
12121 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2)
12122 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, &Mask[0]);
12128 SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
12129 SDValue N0 = N->getOperand(0);
12130 SDValue N2 = N->getOperand(2);
12132 // If the input vector is a concatenation, and the insert replaces
12133 // one of the halves, we can optimize into a single concat_vectors.
12134 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
12135 N0->getNumOperands() == 2 && N2.getOpcode() == ISD::Constant) {
12136 APInt InsIdx = cast<ConstantSDNode>(N2)->getAPIntValue();
12137 EVT VT = N->getValueType(0);
12139 // Lower half: fold (insert_subvector (concat_vectors X, Y), Z) ->
12140 // (concat_vectors Z, Y)
12142 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
12143 N->getOperand(1), N0.getOperand(1));
12145 // Upper half: fold (insert_subvector (concat_vectors X, Y), Z) ->
12146 // (concat_vectors X, Z)
12147 if (InsIdx == VT.getVectorNumElements()/2)
12148 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
12149 N0.getOperand(0), N->getOperand(1));
12155 /// Returns a vector_shuffle if it able to transform an AND to a vector_shuffle
12156 /// with the destination vector and a zero vector.
12157 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
12158 /// vector_shuffle V, Zero, <0, 4, 2, 4>
12159 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
12160 EVT VT = N->getValueType(0);
12162 SDValue LHS = N->getOperand(0);
12163 SDValue RHS = N->getOperand(1);
12164 if (N->getOpcode() == ISD::AND) {
12165 if (RHS.getOpcode() == ISD::BITCAST)
12166 RHS = RHS.getOperand(0);
12167 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
12168 SmallVector<int, 8> Indices;
12169 unsigned NumElts = RHS.getNumOperands();
12170 for (unsigned i = 0; i != NumElts; ++i) {
12171 SDValue Elt = RHS.getOperand(i);
12172 if (!isa<ConstantSDNode>(Elt))
12175 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
12176 Indices.push_back(i);
12177 else if (cast<ConstantSDNode>(Elt)->isNullValue())
12178 Indices.push_back(NumElts+i);
12183 // Let's see if the target supports this vector_shuffle and make sure
12184 // we're not running after operation legalization where it may have
12185 // custom lowered the vector shuffles.
12186 EVT RVT = RHS.getValueType();
12187 if (LegalOperations || !TLI.isVectorClearMaskLegal(Indices, RVT))
12190 // Return the new VECTOR_SHUFFLE node.
12191 EVT EltVT = RVT.getVectorElementType();
12192 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
12193 DAG.getConstant(0, EltVT));
12194 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), RVT, ZeroOps);
12195 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
12196 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
12197 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
12204 /// Visit a binary vector operation, like ADD.
12205 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
12206 assert(N->getValueType(0).isVector() &&
12207 "SimplifyVBinOp only works on vectors!");
12209 SDValue LHS = N->getOperand(0);
12210 SDValue RHS = N->getOperand(1);
12211 SDValue Shuffle = XformToShuffleWithZero(N);
12212 if (Shuffle.getNode()) return Shuffle;
12214 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
12216 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
12217 RHS.getOpcode() == ISD::BUILD_VECTOR) {
12218 // Check if both vectors are constants. If not bail out.
12219 if (!(cast<BuildVectorSDNode>(LHS)->isConstant() &&
12220 cast<BuildVectorSDNode>(RHS)->isConstant()))
12223 SmallVector<SDValue, 8> Ops;
12224 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
12225 SDValue LHSOp = LHS.getOperand(i);
12226 SDValue RHSOp = RHS.getOperand(i);
12228 // Can't fold divide by zero.
12229 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
12230 N->getOpcode() == ISD::FDIV) {
12231 if ((RHSOp.getOpcode() == ISD::Constant &&
12232 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
12233 (RHSOp.getOpcode() == ISD::ConstantFP &&
12234 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
12238 EVT VT = LHSOp.getValueType();
12239 EVT RVT = RHSOp.getValueType();
12241 // Integer BUILD_VECTOR operands may have types larger than the element
12242 // size (e.g., when the element type is not legal). Prior to type
12243 // legalization, the types may not match between the two BUILD_VECTORS.
12244 // Truncate one of the operands to make them match.
12245 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
12246 RHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, RHSOp);
12248 LHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), RVT, LHSOp);
12252 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(LHS), VT,
12254 if (FoldOp.getOpcode() != ISD::UNDEF &&
12255 FoldOp.getOpcode() != ISD::Constant &&
12256 FoldOp.getOpcode() != ISD::ConstantFP)
12258 Ops.push_back(FoldOp);
12259 AddToWorklist(FoldOp.getNode());
12262 if (Ops.size() == LHS.getNumOperands())
12263 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), LHS.getValueType(), Ops);
12266 // Type legalization might introduce new shuffles in the DAG.
12267 // Fold (VBinOp (shuffle (A, Undef, Mask)), (shuffle (B, Undef, Mask)))
12268 // -> (shuffle (VBinOp (A, B)), Undef, Mask).
12269 if (LegalTypes && isa<ShuffleVectorSDNode>(LHS) &&
12270 isa<ShuffleVectorSDNode>(RHS) && LHS.hasOneUse() && RHS.hasOneUse() &&
12271 LHS.getOperand(1).getOpcode() == ISD::UNDEF &&
12272 RHS.getOperand(1).getOpcode() == ISD::UNDEF) {
12273 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(LHS);
12274 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(RHS);
12276 if (SVN0->getMask().equals(SVN1->getMask())) {
12277 EVT VT = N->getValueType(0);
12278 SDValue UndefVector = LHS.getOperand(1);
12279 SDValue NewBinOp = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
12280 LHS.getOperand(0), RHS.getOperand(0));
12281 AddUsersToWorklist(N);
12282 return DAG.getVectorShuffle(VT, SDLoc(N), NewBinOp, UndefVector,
12283 &SVN0->getMask()[0]);
12290 /// Visit a binary vector operation, like FABS/FNEG.
12291 SDValue DAGCombiner::SimplifyVUnaryOp(SDNode *N) {
12292 assert(N->getValueType(0).isVector() &&
12293 "SimplifyVUnaryOp only works on vectors!");
12295 SDValue N0 = N->getOperand(0);
12297 if (N0.getOpcode() != ISD::BUILD_VECTOR)
12300 // Operand is a BUILD_VECTOR node, see if we can constant fold it.
12301 SmallVector<SDValue, 8> Ops;
12302 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
12303 SDValue Op = N0.getOperand(i);
12304 if (Op.getOpcode() != ISD::UNDEF &&
12305 Op.getOpcode() != ISD::ConstantFP)
12307 EVT EltVT = Op.getValueType();
12308 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(N0), EltVT, Op);
12309 if (FoldOp.getOpcode() != ISD::UNDEF &&
12310 FoldOp.getOpcode() != ISD::ConstantFP)
12312 Ops.push_back(FoldOp);
12313 AddToWorklist(FoldOp.getNode());
12316 if (Ops.size() != N0.getNumOperands())
12319 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), N0.getValueType(), Ops);
12322 SDValue DAGCombiner::SimplifySelect(SDLoc DL, SDValue N0,
12323 SDValue N1, SDValue N2){
12324 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
12326 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
12327 cast<CondCodeSDNode>(N0.getOperand(2))->get());
12329 // If we got a simplified select_cc node back from SimplifySelectCC, then
12330 // break it down into a new SETCC node, and a new SELECT node, and then return
12331 // the SELECT node, since we were called with a SELECT node.
12332 if (SCC.getNode()) {
12333 // Check to see if we got a select_cc back (to turn into setcc/select).
12334 // Otherwise, just return whatever node we got back, like fabs.
12335 if (SCC.getOpcode() == ISD::SELECT_CC) {
12336 SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0),
12338 SCC.getOperand(0), SCC.getOperand(1),
12339 SCC.getOperand(4));
12340 AddToWorklist(SETCC.getNode());
12341 return DAG.getSelect(SDLoc(SCC), SCC.getValueType(), SETCC,
12342 SCC.getOperand(2), SCC.getOperand(3));
12350 /// Given a SELECT or a SELECT_CC node, where LHS and RHS are the two values
12351 /// being selected between, see if we can simplify the select. Callers of this
12352 /// should assume that TheSelect is deleted if this returns true. As such, they
12353 /// should return the appropriate thing (e.g. the node) back to the top-level of
12354 /// the DAG combiner loop to avoid it being looked at.
12355 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
12358 // Cannot simplify select with vector condition
12359 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
12361 // If this is a select from two identical things, try to pull the operation
12362 // through the select.
12363 if (LHS.getOpcode() != RHS.getOpcode() ||
12364 !LHS.hasOneUse() || !RHS.hasOneUse())
12367 // If this is a load and the token chain is identical, replace the select
12368 // of two loads with a load through a select of the address to load from.
12369 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
12370 // constants have been dropped into the constant pool.
12371 if (LHS.getOpcode() == ISD::LOAD) {
12372 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
12373 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
12375 // Token chains must be identical.
12376 if (LHS.getOperand(0) != RHS.getOperand(0) ||
12377 // Do not let this transformation reduce the number of volatile loads.
12378 LLD->isVolatile() || RLD->isVolatile() ||
12379 // If this is an EXTLOAD, the VT's must match.
12380 LLD->getMemoryVT() != RLD->getMemoryVT() ||
12381 // If this is an EXTLOAD, the kind of extension must match.
12382 (LLD->getExtensionType() != RLD->getExtensionType() &&
12383 // The only exception is if one of the extensions is anyext.
12384 LLD->getExtensionType() != ISD::EXTLOAD &&
12385 RLD->getExtensionType() != ISD::EXTLOAD) ||
12386 // FIXME: this discards src value information. This is
12387 // over-conservative. It would be beneficial to be able to remember
12388 // both potential memory locations. Since we are discarding
12389 // src value info, don't do the transformation if the memory
12390 // locations are not in the default address space.
12391 LLD->getPointerInfo().getAddrSpace() != 0 ||
12392 RLD->getPointerInfo().getAddrSpace() != 0 ||
12393 !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(),
12394 LLD->getBasePtr().getValueType()))
12397 // Check that the select condition doesn't reach either load. If so,
12398 // folding this will induce a cycle into the DAG. If not, this is safe to
12399 // xform, so create a select of the addresses.
12401 if (TheSelect->getOpcode() == ISD::SELECT) {
12402 SDNode *CondNode = TheSelect->getOperand(0).getNode();
12403 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
12404 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
12406 // The loads must not depend on one another.
12407 if (LLD->isPredecessorOf(RLD) ||
12408 RLD->isPredecessorOf(LLD))
12410 Addr = DAG.getSelect(SDLoc(TheSelect),
12411 LLD->getBasePtr().getValueType(),
12412 TheSelect->getOperand(0), LLD->getBasePtr(),
12413 RLD->getBasePtr());
12414 } else { // Otherwise SELECT_CC
12415 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
12416 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
12418 if ((LLD->hasAnyUseOfValue(1) &&
12419 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
12420 (RLD->hasAnyUseOfValue(1) &&
12421 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
12424 Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect),
12425 LLD->getBasePtr().getValueType(),
12426 TheSelect->getOperand(0),
12427 TheSelect->getOperand(1),
12428 LLD->getBasePtr(), RLD->getBasePtr(),
12429 TheSelect->getOperand(4));
12433 // It is safe to replace the two loads if they have different alignments,
12434 // but the new load must be the minimum (most restrictive) alignment of the
12436 bool isInvariant = LLD->isInvariant() & RLD->isInvariant();
12437 unsigned Alignment = std::min(LLD->getAlignment(), RLD->getAlignment());
12438 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
12439 Load = DAG.getLoad(TheSelect->getValueType(0),
12441 // FIXME: Discards pointer and AA info.
12442 LLD->getChain(), Addr, MachinePointerInfo(),
12443 LLD->isVolatile(), LLD->isNonTemporal(),
12444 isInvariant, Alignment);
12446 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
12447 RLD->getExtensionType() : LLD->getExtensionType(),
12449 TheSelect->getValueType(0),
12450 // FIXME: Discards pointer and AA info.
12451 LLD->getChain(), Addr, MachinePointerInfo(),
12452 LLD->getMemoryVT(), LLD->isVolatile(),
12453 LLD->isNonTemporal(), isInvariant, Alignment);
12456 // Users of the select now use the result of the load.
12457 CombineTo(TheSelect, Load);
12459 // Users of the old loads now use the new load's chain. We know the
12460 // old-load value is dead now.
12461 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
12462 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
12469 /// Simplify an expression of the form (N0 cond N1) ? N2 : N3
12470 /// where 'cond' is the comparison specified by CC.
12471 SDValue DAGCombiner::SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1,
12472 SDValue N2, SDValue N3,
12473 ISD::CondCode CC, bool NotExtCompare) {
12474 // (x ? y : y) -> y.
12475 if (N2 == N3) return N2;
12477 EVT VT = N2.getValueType();
12478 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
12479 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
12480 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
12482 // Determine if the condition we're dealing with is constant
12483 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
12484 N0, N1, CC, DL, false);
12485 if (SCC.getNode()) AddToWorklist(SCC.getNode());
12486 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
12488 // fold select_cc true, x, y -> x
12489 if (SCCC && !SCCC->isNullValue())
12491 // fold select_cc false, x, y -> y
12492 if (SCCC && SCCC->isNullValue())
12495 // Check to see if we can simplify the select into an fabs node
12496 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
12497 // Allow either -0.0 or 0.0
12498 if (CFP->getValueAPF().isZero()) {
12499 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
12500 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
12501 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
12502 N2 == N3.getOperand(0))
12503 return DAG.getNode(ISD::FABS, DL, VT, N0);
12505 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
12506 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
12507 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
12508 N2.getOperand(0) == N3)
12509 return DAG.getNode(ISD::FABS, DL, VT, N3);
12513 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
12514 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
12515 // in it. This is a win when the constant is not otherwise available because
12516 // it replaces two constant pool loads with one. We only do this if the FP
12517 // type is known to be legal, because if it isn't, then we are before legalize
12518 // types an we want the other legalization to happen first (e.g. to avoid
12519 // messing with soft float) and if the ConstantFP is not legal, because if
12520 // it is legal, we may not need to store the FP constant in a constant pool.
12521 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
12522 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
12523 if (TLI.isTypeLegal(N2.getValueType()) &&
12524 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
12525 TargetLowering::Legal &&
12526 !TLI.isFPImmLegal(TV->getValueAPF(), TV->getValueType(0)) &&
12527 !TLI.isFPImmLegal(FV->getValueAPF(), FV->getValueType(0))) &&
12528 // If both constants have multiple uses, then we won't need to do an
12529 // extra load, they are likely around in registers for other users.
12530 (TV->hasOneUse() || FV->hasOneUse())) {
12531 Constant *Elts[] = {
12532 const_cast<ConstantFP*>(FV->getConstantFPValue()),
12533 const_cast<ConstantFP*>(TV->getConstantFPValue())
12535 Type *FPTy = Elts[0]->getType();
12536 const DataLayout &TD = *TLI.getDataLayout();
12538 // Create a ConstantArray of the two constants.
12539 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
12540 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
12541 TD.getPrefTypeAlignment(FPTy));
12542 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
12544 // Get the offsets to the 0 and 1 element of the array so that we can
12545 // select between them.
12546 SDValue Zero = DAG.getIntPtrConstant(0);
12547 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
12548 SDValue One = DAG.getIntPtrConstant(EltSize);
12550 SDValue Cond = DAG.getSetCC(DL,
12551 getSetCCResultType(N0.getValueType()),
12553 AddToWorklist(Cond.getNode());
12554 SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(),
12556 AddToWorklist(CstOffset.getNode());
12557 CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx,
12559 AddToWorklist(CPIdx.getNode());
12560 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
12561 MachinePointerInfo::getConstantPool(), false,
12562 false, false, Alignment);
12567 // Check to see if we can perform the "gzip trick", transforming
12568 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
12569 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
12570 (N1C->isNullValue() || // (a < 0) ? b : 0
12571 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
12572 EVT XType = N0.getValueType();
12573 EVT AType = N2.getValueType();
12574 if (XType.bitsGE(AType)) {
12575 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
12576 // single-bit constant.
12577 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
12578 unsigned ShCtV = N2C->getAPIntValue().logBase2();
12579 ShCtV = XType.getSizeInBits()-ShCtV-1;
12580 SDValue ShCt = DAG.getConstant(ShCtV,
12581 getShiftAmountTy(N0.getValueType()));
12582 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0),
12584 AddToWorklist(Shift.getNode());
12586 if (XType.bitsGT(AType)) {
12587 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
12588 AddToWorklist(Shift.getNode());
12591 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
12594 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0),
12596 DAG.getConstant(XType.getSizeInBits()-1,
12597 getShiftAmountTy(N0.getValueType())));
12598 AddToWorklist(Shift.getNode());
12600 if (XType.bitsGT(AType)) {
12601 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
12602 AddToWorklist(Shift.getNode());
12605 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
12609 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
12610 // where y is has a single bit set.
12611 // A plaintext description would be, we can turn the SELECT_CC into an AND
12612 // when the condition can be materialized as an all-ones register. Any
12613 // single bit-test can be materialized as an all-ones register with
12614 // shift-left and shift-right-arith.
12615 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
12616 N0->getValueType(0) == VT &&
12617 N1C && N1C->isNullValue() &&
12618 N2C && N2C->isNullValue()) {
12619 SDValue AndLHS = N0->getOperand(0);
12620 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
12621 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
12622 // Shift the tested bit over the sign bit.
12623 APInt AndMask = ConstAndRHS->getAPIntValue();
12625 DAG.getConstant(AndMask.countLeadingZeros(),
12626 getShiftAmountTy(AndLHS.getValueType()));
12627 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt);
12629 // Now arithmetic right shift it all the way over, so the result is either
12630 // all-ones, or zero.
12632 DAG.getConstant(AndMask.getBitWidth()-1,
12633 getShiftAmountTy(Shl.getValueType()));
12634 SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt);
12636 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
12640 // fold select C, 16, 0 -> shl C, 4
12641 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
12642 TLI.getBooleanContents(N0.getValueType()) ==
12643 TargetLowering::ZeroOrOneBooleanContent) {
12645 // If the caller doesn't want us to simplify this into a zext of a compare,
12647 if (NotExtCompare && N2C->getAPIntValue() == 1)
12650 // Get a SetCC of the condition
12651 // NOTE: Don't create a SETCC if it's not legal on this target.
12652 if (!LegalOperations ||
12653 TLI.isOperationLegal(ISD::SETCC,
12654 LegalTypes ? getSetCCResultType(N0.getValueType()) : MVT::i1)) {
12656 // cast from setcc result type to select result type
12658 SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()),
12660 if (N2.getValueType().bitsLT(SCC.getValueType()))
12661 Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2),
12662 N2.getValueType());
12664 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
12665 N2.getValueType(), SCC);
12667 SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC);
12668 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
12669 N2.getValueType(), SCC);
12672 AddToWorklist(SCC.getNode());
12673 AddToWorklist(Temp.getNode());
12675 if (N2C->getAPIntValue() == 1)
12678 // shl setcc result by log2 n2c
12679 return DAG.getNode(
12680 ISD::SHL, DL, N2.getValueType(), Temp,
12681 DAG.getConstant(N2C->getAPIntValue().logBase2(),
12682 getShiftAmountTy(Temp.getValueType())));
12686 // Check to see if this is the equivalent of setcc
12687 // FIXME: Turn all of these into setcc if setcc if setcc is legal
12688 // otherwise, go ahead with the folds.
12689 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
12690 EVT XType = N0.getValueType();
12691 if (!LegalOperations ||
12692 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(XType))) {
12693 SDValue Res = DAG.getSetCC(DL, getSetCCResultType(XType), N0, N1, CC);
12694 if (Res.getValueType() != VT)
12695 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
12699 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
12700 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
12701 (!LegalOperations ||
12702 TLI.isOperationLegal(ISD::CTLZ, XType))) {
12703 SDValue Ctlz = DAG.getNode(ISD::CTLZ, SDLoc(N0), XType, N0);
12704 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
12705 DAG.getConstant(Log2_32(XType.getSizeInBits()),
12706 getShiftAmountTy(Ctlz.getValueType())));
12708 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
12709 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
12710 SDValue NegN0 = DAG.getNode(ISD::SUB, SDLoc(N0),
12711 XType, DAG.getConstant(0, XType), N0);
12712 SDValue NotN0 = DAG.getNOT(SDLoc(N0), N0, XType);
12713 return DAG.getNode(ISD::SRL, DL, XType,
12714 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
12715 DAG.getConstant(XType.getSizeInBits()-1,
12716 getShiftAmountTy(XType)));
12718 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
12719 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
12720 SDValue Sign = DAG.getNode(ISD::SRL, SDLoc(N0), XType, N0,
12721 DAG.getConstant(XType.getSizeInBits()-1,
12722 getShiftAmountTy(N0.getValueType())));
12723 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
12727 // Check to see if this is an integer abs.
12728 // select_cc setg[te] X, 0, X, -X ->
12729 // select_cc setgt X, -1, X, -X ->
12730 // select_cc setl[te] X, 0, -X, X ->
12731 // select_cc setlt X, 1, -X, X ->
12732 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
12734 ConstantSDNode *SubC = nullptr;
12735 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
12736 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
12737 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
12738 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
12739 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
12740 (N1C->isOne() && CC == ISD::SETLT)) &&
12741 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
12742 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
12744 EVT XType = N0.getValueType();
12745 if (SubC && SubC->isNullValue() && XType.isInteger()) {
12746 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0), XType,
12748 DAG.getConstant(XType.getSizeInBits()-1,
12749 getShiftAmountTy(N0.getValueType())));
12750 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0),
12752 AddToWorklist(Shift.getNode());
12753 AddToWorklist(Add.getNode());
12754 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
12761 /// This is a stub for TargetLowering::SimplifySetCC.
12762 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
12763 SDValue N1, ISD::CondCode Cond,
12764 SDLoc DL, bool foldBooleans) {
12765 TargetLowering::DAGCombinerInfo
12766 DagCombineInfo(DAG, Level, false, this);
12767 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
12770 /// Given an ISD::SDIV node expressing a divide by constant, return
12771 /// a DAG expression to select that will generate the same value by multiplying
12772 /// by a magic number.
12773 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
12774 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
12775 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
12779 // Avoid division by zero.
12780 if (!C->getAPIntValue())
12783 std::vector<SDNode*> Built;
12785 TLI.BuildSDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
12787 for (SDNode *N : Built)
12792 /// Given an ISD::SDIV node expressing a divide by constant power of 2, return a
12793 /// DAG expression that will generate the same value by right shifting.
12794 SDValue DAGCombiner::BuildSDIVPow2(SDNode *N) {
12795 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
12799 // Avoid division by zero.
12800 if (!C->getAPIntValue())
12803 std::vector<SDNode *> Built;
12804 SDValue S = TLI.BuildSDIVPow2(N, C->getAPIntValue(), DAG, &Built);
12806 for (SDNode *N : Built)
12811 /// Given an ISD::UDIV node expressing a divide by constant, return a DAG
12812 /// expression that will generate the same value by multiplying by a magic
12814 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
12815 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
12816 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
12820 // Avoid division by zero.
12821 if (!C->getAPIntValue())
12824 std::vector<SDNode*> Built;
12826 TLI.BuildUDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
12828 for (SDNode *N : Built)
12833 SDValue DAGCombiner::BuildReciprocalEstimate(SDValue Op) {
12834 if (Level >= AfterLegalizeDAG)
12837 // Expose the DAG combiner to the target combiner implementations.
12838 TargetLowering::DAGCombinerInfo DCI(DAG, Level, false, this);
12840 unsigned Iterations = 0;
12841 if (SDValue Est = TLI.getRecipEstimate(Op, DCI, Iterations)) {
12843 // Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
12844 // For the reciprocal, we need to find the zero of the function:
12845 // F(X) = A X - 1 [which has a zero at X = 1/A]
12847 // X_{i+1} = X_i (2 - A X_i) = X_i + X_i (1 - A X_i) [this second form
12848 // does not require additional intermediate precision]
12849 EVT VT = Op.getValueType();
12851 SDValue FPOne = DAG.getConstantFP(1.0, VT);
12853 AddToWorklist(Est.getNode());
12855 // Newton iterations: Est = Est + Est (1 - Arg * Est)
12856 for (unsigned i = 0; i < Iterations; ++i) {
12857 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Op, Est);
12858 AddToWorklist(NewEst.getNode());
12860 NewEst = DAG.getNode(ISD::FSUB, DL, VT, FPOne, NewEst);
12861 AddToWorklist(NewEst.getNode());
12863 NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst);
12864 AddToWorklist(NewEst.getNode());
12866 Est = DAG.getNode(ISD::FADD, DL, VT, Est, NewEst);
12867 AddToWorklist(Est.getNode());
12876 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
12877 /// For the reciprocal sqrt, we need to find the zero of the function:
12878 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
12880 /// X_{i+1} = X_i (1.5 - A X_i^2 / 2)
12881 /// As a result, we precompute A/2 prior to the iteration loop.
12882 SDValue DAGCombiner::BuildRsqrtNROneConst(SDValue Arg, SDValue Est,
12883 unsigned Iterations) {
12884 EVT VT = Arg.getValueType();
12886 SDValue ThreeHalves = DAG.getConstantFP(1.5, VT);
12888 // We now need 0.5 * Arg which we can write as (1.5 * Arg - Arg) so that
12889 // this entire sequence requires only one FP constant.
12890 SDValue HalfArg = DAG.getNode(ISD::FMUL, DL, VT, ThreeHalves, Arg);
12891 AddToWorklist(HalfArg.getNode());
12893 HalfArg = DAG.getNode(ISD::FSUB, DL, VT, HalfArg, Arg);
12894 AddToWorklist(HalfArg.getNode());
12896 // Newton iterations: Est = Est * (1.5 - HalfArg * Est * Est)
12897 for (unsigned i = 0; i < Iterations; ++i) {
12898 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, Est);
12899 AddToWorklist(NewEst.getNode());
12901 NewEst = DAG.getNode(ISD::FMUL, DL, VT, HalfArg, NewEst);
12902 AddToWorklist(NewEst.getNode());
12904 NewEst = DAG.getNode(ISD::FSUB, DL, VT, ThreeHalves, NewEst);
12905 AddToWorklist(NewEst.getNode());
12907 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst);
12908 AddToWorklist(Est.getNode());
12913 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
12914 /// For the reciprocal sqrt, we need to find the zero of the function:
12915 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
12917 /// X_{i+1} = (-0.5 * X_i) * (A * X_i * X_i + (-3.0))
12918 SDValue DAGCombiner::BuildRsqrtNRTwoConst(SDValue Arg, SDValue Est,
12919 unsigned Iterations) {
12920 EVT VT = Arg.getValueType();
12922 SDValue MinusThree = DAG.getConstantFP(-3.0, VT);
12923 SDValue MinusHalf = DAG.getConstantFP(-0.5, VT);
12925 // Newton iterations: Est = -0.5 * Est * (-3.0 + Arg * Est * Est)
12926 for (unsigned i = 0; i < Iterations; ++i) {
12927 SDValue HalfEst = DAG.getNode(ISD::FMUL, DL, VT, Est, MinusHalf);
12928 AddToWorklist(HalfEst.getNode());
12930 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Est);
12931 AddToWorklist(Est.getNode());
12933 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Arg);
12934 AddToWorklist(Est.getNode());
12936 Est = DAG.getNode(ISD::FADD, DL, VT, Est, MinusThree);
12937 AddToWorklist(Est.getNode());
12939 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, HalfEst);
12940 AddToWorklist(Est.getNode());
12945 SDValue DAGCombiner::BuildRsqrtEstimate(SDValue Op) {
12946 if (Level >= AfterLegalizeDAG)
12949 // Expose the DAG combiner to the target combiner implementations.
12950 TargetLowering::DAGCombinerInfo DCI(DAG, Level, false, this);
12951 unsigned Iterations = 0;
12952 bool UseOneConstNR = false;
12953 if (SDValue Est = TLI.getRsqrtEstimate(Op, DCI, Iterations, UseOneConstNR)) {
12954 AddToWorklist(Est.getNode());
12956 Est = UseOneConstNR ?
12957 BuildRsqrtNROneConst(Op, Est, Iterations) :
12958 BuildRsqrtNRTwoConst(Op, Est, Iterations);
12966 /// Return true if base is a frame index, which is known not to alias with
12967 /// anything but itself. Provides base object and offset as results.
12968 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
12969 const GlobalValue *&GV, const void *&CV) {
12970 // Assume it is a primitive operation.
12971 Base = Ptr; Offset = 0; GV = nullptr; CV = nullptr;
12973 // If it's an adding a simple constant then integrate the offset.
12974 if (Base.getOpcode() == ISD::ADD) {
12975 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
12976 Base = Base.getOperand(0);
12977 Offset += C->getZExtValue();
12981 // Return the underlying GlobalValue, and update the Offset. Return false
12982 // for GlobalAddressSDNode since the same GlobalAddress may be represented
12983 // by multiple nodes with different offsets.
12984 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
12985 GV = G->getGlobal();
12986 Offset += G->getOffset();
12990 // Return the underlying Constant value, and update the Offset. Return false
12991 // for ConstantSDNodes since the same constant pool entry may be represented
12992 // by multiple nodes with different offsets.
12993 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
12994 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
12995 : (const void *)C->getConstVal();
12996 Offset += C->getOffset();
12999 // If it's any of the following then it can't alias with anything but itself.
13000 return isa<FrameIndexSDNode>(Base);
13003 /// Return true if there is any possibility that the two addresses overlap.
13004 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const {
13005 // If they are the same then they must be aliases.
13006 if (Op0->getBasePtr() == Op1->getBasePtr()) return true;
13008 // If they are both volatile then they cannot be reordered.
13009 if (Op0->isVolatile() && Op1->isVolatile()) return true;
13011 // Gather base node and offset information.
13012 SDValue Base1, Base2;
13013 int64_t Offset1, Offset2;
13014 const GlobalValue *GV1, *GV2;
13015 const void *CV1, *CV2;
13016 bool isFrameIndex1 = FindBaseOffset(Op0->getBasePtr(),
13017 Base1, Offset1, GV1, CV1);
13018 bool isFrameIndex2 = FindBaseOffset(Op1->getBasePtr(),
13019 Base2, Offset2, GV2, CV2);
13021 // If they have a same base address then check to see if they overlap.
13022 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
13023 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
13024 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
13026 // It is possible for different frame indices to alias each other, mostly
13027 // when tail call optimization reuses return address slots for arguments.
13028 // To catch this case, look up the actual index of frame indices to compute
13029 // the real alias relationship.
13030 if (isFrameIndex1 && isFrameIndex2) {
13031 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
13032 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
13033 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
13034 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
13035 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
13038 // Otherwise, if we know what the bases are, and they aren't identical, then
13039 // we know they cannot alias.
13040 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
13043 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
13044 // compared to the size and offset of the access, we may be able to prove they
13045 // do not alias. This check is conservative for now to catch cases created by
13046 // splitting vector types.
13047 if ((Op0->getOriginalAlignment() == Op1->getOriginalAlignment()) &&
13048 (Op0->getSrcValueOffset() != Op1->getSrcValueOffset()) &&
13049 (Op0->getMemoryVT().getSizeInBits() >> 3 ==
13050 Op1->getMemoryVT().getSizeInBits() >> 3) &&
13051 (Op0->getOriginalAlignment() > Op0->getMemoryVT().getSizeInBits()) >> 3) {
13052 int64_t OffAlign1 = Op0->getSrcValueOffset() % Op0->getOriginalAlignment();
13053 int64_t OffAlign2 = Op1->getSrcValueOffset() % Op1->getOriginalAlignment();
13055 // There is no overlap between these relatively aligned accesses of similar
13056 // size, return no alias.
13057 if ((OffAlign1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign2 ||
13058 (OffAlign2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign1)
13062 bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0
13064 : DAG.getSubtarget().useAA();
13066 if (CombinerAAOnlyFunc.getNumOccurrences() &&
13067 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
13071 Op0->getMemOperand()->getValue() && Op1->getMemOperand()->getValue()) {
13072 // Use alias analysis information.
13073 int64_t MinOffset = std::min(Op0->getSrcValueOffset(),
13074 Op1->getSrcValueOffset());
13075 int64_t Overlap1 = (Op0->getMemoryVT().getSizeInBits() >> 3) +
13076 Op0->getSrcValueOffset() - MinOffset;
13077 int64_t Overlap2 = (Op1->getMemoryVT().getSizeInBits() >> 3) +
13078 Op1->getSrcValueOffset() - MinOffset;
13079 AliasAnalysis::AliasResult AAResult =
13080 AA.alias(AliasAnalysis::Location(Op0->getMemOperand()->getValue(),
13082 UseTBAA ? Op0->getAAInfo() : AAMDNodes()),
13083 AliasAnalysis::Location(Op1->getMemOperand()->getValue(),
13085 UseTBAA ? Op1->getAAInfo() : AAMDNodes()));
13086 if (AAResult == AliasAnalysis::NoAlias)
13090 // Otherwise we have to assume they alias.
13094 /// Walk up chain skipping non-aliasing memory nodes,
13095 /// looking for aliasing nodes and adding them to the Aliases vector.
13096 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
13097 SmallVectorImpl<SDValue> &Aliases) {
13098 SmallVector<SDValue, 8> Chains; // List of chains to visit.
13099 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
13101 // Get alias information for node.
13102 bool IsLoad = isa<LoadSDNode>(N) && !cast<LSBaseSDNode>(N)->isVolatile();
13105 Chains.push_back(OriginalChain);
13106 unsigned Depth = 0;
13108 // Look at each chain and determine if it is an alias. If so, add it to the
13109 // aliases list. If not, then continue up the chain looking for the next
13111 while (!Chains.empty()) {
13112 SDValue Chain = Chains.back();
13115 // For TokenFactor nodes, look at each operand and only continue up the
13116 // chain until we find two aliases. If we've seen two aliases, assume we'll
13117 // find more and revert to original chain since the xform is unlikely to be
13120 // FIXME: The depth check could be made to return the last non-aliasing
13121 // chain we found before we hit a tokenfactor rather than the original
13123 if (Depth > 6 || Aliases.size() == 2) {
13125 Aliases.push_back(OriginalChain);
13129 // Don't bother if we've been before.
13130 if (!Visited.insert(Chain.getNode()).second)
13133 switch (Chain.getOpcode()) {
13134 case ISD::EntryToken:
13135 // Entry token is ideal chain operand, but handled in FindBetterChain.
13140 // Get alias information for Chain.
13141 bool IsOpLoad = isa<LoadSDNode>(Chain.getNode()) &&
13142 !cast<LSBaseSDNode>(Chain.getNode())->isVolatile();
13144 // If chain is alias then stop here.
13145 if (!(IsLoad && IsOpLoad) &&
13146 isAlias(cast<LSBaseSDNode>(N), cast<LSBaseSDNode>(Chain.getNode()))) {
13147 Aliases.push_back(Chain);
13149 // Look further up the chain.
13150 Chains.push_back(Chain.getOperand(0));
13156 case ISD::TokenFactor:
13157 // We have to check each of the operands of the token factor for "small"
13158 // token factors, so we queue them up. Adding the operands to the queue
13159 // (stack) in reverse order maintains the original order and increases the
13160 // likelihood that getNode will find a matching token factor (CSE.)
13161 if (Chain.getNumOperands() > 16) {
13162 Aliases.push_back(Chain);
13165 for (unsigned n = Chain.getNumOperands(); n;)
13166 Chains.push_back(Chain.getOperand(--n));
13171 // For all other instructions we will just have to take what we can get.
13172 Aliases.push_back(Chain);
13177 // We need to be careful here to also search for aliases through the
13178 // value operand of a store, etc. Consider the following situation:
13180 // L1 = load Token1, %52
13181 // S1 = store Token1, L1, %51
13182 // L2 = load Token1, %52+8
13183 // S2 = store Token1, L2, %51+8
13184 // Token2 = Token(S1, S2)
13185 // L3 = load Token2, %53
13186 // S3 = store Token2, L3, %52
13187 // L4 = load Token2, %53+8
13188 // S4 = store Token2, L4, %52+8
13189 // If we search for aliases of S3 (which loads address %52), and we look
13190 // only through the chain, then we'll miss the trivial dependence on L1
13191 // (which also loads from %52). We then might change all loads and
13192 // stores to use Token1 as their chain operand, which could result in
13193 // copying %53 into %52 before copying %52 into %51 (which should
13196 // The problem is, however, that searching for such data dependencies
13197 // can become expensive, and the cost is not directly related to the
13198 // chain depth. Instead, we'll rule out such configurations here by
13199 // insisting that we've visited all chain users (except for users
13200 // of the original chain, which is not necessary). When doing this,
13201 // we need to look through nodes we don't care about (otherwise, things
13202 // like register copies will interfere with trivial cases).
13204 SmallVector<const SDNode *, 16> Worklist;
13205 for (const SDNode *N : Visited)
13206 if (N != OriginalChain.getNode())
13207 Worklist.push_back(N);
13209 while (!Worklist.empty()) {
13210 const SDNode *M = Worklist.pop_back_val();
13212 // We have already visited M, and want to make sure we've visited any uses
13213 // of M that we care about. For uses that we've not visisted, and don't
13214 // care about, queue them to the worklist.
13216 for (SDNode::use_iterator UI = M->use_begin(),
13217 UIE = M->use_end(); UI != UIE; ++UI)
13218 if (UI.getUse().getValueType() == MVT::Other &&
13219 Visited.insert(*UI).second) {
13220 if (isa<MemIntrinsicSDNode>(*UI) || isa<MemSDNode>(*UI)) {
13221 // We've not visited this use, and we care about it (it could have an
13222 // ordering dependency with the original node).
13224 Aliases.push_back(OriginalChain);
13228 // We've not visited this use, but we don't care about it. Mark it as
13229 // visited and enqueue it to the worklist.
13230 Worklist.push_back(*UI);
13235 /// Walk up chain skipping non-aliasing memory nodes, looking for a better chain
13236 /// (aliasing node.)
13237 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
13238 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
13240 // Accumulate all the aliases to this node.
13241 GatherAllAliases(N, OldChain, Aliases);
13243 // If no operands then chain to entry token.
13244 if (Aliases.size() == 0)
13245 return DAG.getEntryNode();
13247 // If a single operand then chain to it. We don't need to revisit it.
13248 if (Aliases.size() == 1)
13251 // Construct a custom tailored token factor.
13252 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Aliases);
13255 /// This is the entry point for the file.
13256 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
13257 CodeGenOpt::Level OptLevel) {
13258 /// This is the main entry point to this class.
13259 DAGCombiner(*this, AA, OptLevel).Run(Level);