1 //===-- DAGCombiner.cpp - Implement a DAG node combiner -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass combines dag nodes to form fewer, simpler DAG nodes. It can be run
11 // both before and after the DAG is legalized.
13 // This pass is not a substitute for the LLVM IR instcombine pass. This pass is
14 // primarily intended to handle simplification opportunities that are implicit
15 // in the LLVM IR and exposed by the various codegen lowering phases.
17 //===----------------------------------------------------------------------===//
19 #include "llvm/CodeGen/SelectionDAG.h"
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/ADT/SmallBitVector.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/Analysis/AliasAnalysis.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/DerivedTypes.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/LLVMContext.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Target/TargetLowering.h"
37 #include "llvm/Target/TargetOptions.h"
38 #include "llvm/Target/TargetRegisterInfo.h"
39 #include "llvm/Target/TargetSubtargetInfo.h"
43 #define DEBUG_TYPE "dagcombine"
45 STATISTIC(NodesCombined , "Number of dag nodes combined");
46 STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created");
47 STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created");
48 STATISTIC(OpsNarrowed , "Number of load/op/store narrowed");
49 STATISTIC(LdStFP2Int , "Number of fp load/store pairs transformed to int");
50 STATISTIC(SlicedLoads, "Number of load sliced");
54 CombinerAA("combiner-alias-analysis", cl::Hidden,
55 cl::desc("Enable DAG combiner alias-analysis heuristics"));
58 CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden,
59 cl::desc("Enable DAG combiner's use of IR alias analysis"));
62 UseTBAA("combiner-use-tbaa", cl::Hidden, cl::init(true),
63 cl::desc("Enable DAG combiner's use of TBAA"));
66 static cl::opt<std::string>
67 CombinerAAOnlyFunc("combiner-aa-only-func", cl::Hidden,
68 cl::desc("Only use DAG-combiner alias analysis in this"
72 /// Hidden option to stress test load slicing, i.e., when this option
73 /// is enabled, load slicing bypasses most of its profitability guards.
75 StressLoadSlicing("combiner-stress-load-slicing", cl::Hidden,
76 cl::desc("Bypass the profitability model of load "
81 MaySplitLoadIndex("combiner-split-load-index", cl::Hidden, cl::init(true),
82 cl::desc("DAG combiner may split indexing from loads"));
84 //------------------------------ DAGCombiner ---------------------------------//
88 const TargetLowering &TLI;
90 CodeGenOpt::Level OptLevel;
95 /// \brief Worklist of all of the nodes that need to be simplified.
97 /// This must behave as a stack -- new nodes to process are pushed onto the
98 /// back and when processing we pop off of the back.
100 /// The worklist will not contain duplicates but may contain null entries
101 /// due to nodes being deleted from the underlying DAG.
102 SmallVector<SDNode *, 64> Worklist;
104 /// \brief Mapping from an SDNode to its position on the worklist.
106 /// This is used to find and remove nodes from the worklist (by nulling
107 /// them) when they are deleted from the underlying DAG. It relies on
108 /// stable indices of nodes within the worklist.
109 DenseMap<SDNode *, unsigned> WorklistMap;
111 /// \brief Set of nodes which have been combined (at least once).
113 /// This is used to allow us to reliably add any operands of a DAG node
114 /// which have not yet been combined to the worklist.
115 SmallPtrSet<SDNode *, 64> CombinedNodes;
117 // AA - Used for DAG load/store alias analysis.
120 /// When an instruction is simplified, add all users of the instruction to
121 /// the work lists because they might get more simplified now.
122 void AddUsersToWorklist(SDNode *N) {
123 for (SDNode *Node : N->uses())
127 /// Call the node-specific routine that folds each particular type of node.
128 SDValue visit(SDNode *N);
131 /// Add to the worklist making sure its instance is at the back (next to be
133 void AddToWorklist(SDNode *N) {
134 // Skip handle nodes as they can't usefully be combined and confuse the
135 // zero-use deletion strategy.
136 if (N->getOpcode() == ISD::HANDLENODE)
139 if (WorklistMap.insert(std::make_pair(N, Worklist.size())).second)
140 Worklist.push_back(N);
143 /// Remove all instances of N from the worklist.
144 void removeFromWorklist(SDNode *N) {
145 CombinedNodes.erase(N);
147 auto It = WorklistMap.find(N);
148 if (It == WorklistMap.end())
149 return; // Not in the worklist.
151 // Null out the entry rather than erasing it to avoid a linear operation.
152 Worklist[It->second] = nullptr;
153 WorklistMap.erase(It);
156 void deleteAndRecombine(SDNode *N);
157 bool recursivelyDeleteUnusedNodes(SDNode *N);
159 SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
162 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) {
163 return CombineTo(N, &Res, 1, AddTo);
166 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1,
168 SDValue To[] = { Res0, Res1 };
169 return CombineTo(N, To, 2, AddTo);
172 void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO);
176 /// Check the specified integer node value to see if it can be simplified or
177 /// if things it uses can be simplified by bit propagation.
178 /// If so, return true.
179 bool SimplifyDemandedBits(SDValue Op) {
180 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
181 APInt Demanded = APInt::getAllOnesValue(BitWidth);
182 return SimplifyDemandedBits(Op, Demanded);
185 bool SimplifyDemandedBits(SDValue Op, const APInt &Demanded);
187 bool CombineToPreIndexedLoadStore(SDNode *N);
188 bool CombineToPostIndexedLoadStore(SDNode *N);
189 SDValue SplitIndexingFromLoad(LoadSDNode *LD);
190 bool SliceUpLoad(SDNode *N);
192 /// \brief Replace an ISD::EXTRACT_VECTOR_ELT of a load with a narrowed
195 /// \param EVE ISD::EXTRACT_VECTOR_ELT to be replaced.
196 /// \param InVecVT type of the input vector to EVE with bitcasts resolved.
197 /// \param EltNo index of the vector element to load.
198 /// \param OriginalLoad load that EVE came from to be replaced.
199 /// \returns EVE on success SDValue() on failure.
200 SDValue ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
201 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad);
202 void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad);
203 SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace);
204 SDValue SExtPromoteOperand(SDValue Op, EVT PVT);
205 SDValue ZExtPromoteOperand(SDValue Op, EVT PVT);
206 SDValue PromoteIntBinOp(SDValue Op);
207 SDValue PromoteIntShiftOp(SDValue Op);
208 SDValue PromoteExtend(SDValue Op);
209 bool PromoteLoad(SDValue Op);
211 void ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
212 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
213 ISD::NodeType ExtType);
215 /// Call the node-specific routine that knows how to fold each
216 /// particular type of node. If that doesn't do anything, try the
217 /// target-specific DAG combines.
218 SDValue combine(SDNode *N);
220 // Visitation implementation - Implement dag node combining for different
221 // node types. The semantics are as follows:
223 // SDValue.getNode() == 0 - No change was made
224 // SDValue.getNode() == N - N was replaced, is dead and has been handled.
225 // otherwise - N should be replaced by the returned Operand.
227 SDValue visitTokenFactor(SDNode *N);
228 SDValue visitMERGE_VALUES(SDNode *N);
229 SDValue visitADD(SDNode *N);
230 SDValue visitSUB(SDNode *N);
231 SDValue visitADDC(SDNode *N);
232 SDValue visitSUBC(SDNode *N);
233 SDValue visitADDE(SDNode *N);
234 SDValue visitSUBE(SDNode *N);
235 SDValue visitMUL(SDNode *N);
236 SDValue visitSDIV(SDNode *N);
237 SDValue visitUDIV(SDNode *N);
238 SDValue visitSREM(SDNode *N);
239 SDValue visitUREM(SDNode *N);
240 SDValue visitMULHU(SDNode *N);
241 SDValue visitMULHS(SDNode *N);
242 SDValue visitSMUL_LOHI(SDNode *N);
243 SDValue visitUMUL_LOHI(SDNode *N);
244 SDValue visitSMULO(SDNode *N);
245 SDValue visitUMULO(SDNode *N);
246 SDValue visitSDIVREM(SDNode *N);
247 SDValue visitUDIVREM(SDNode *N);
248 SDValue visitAND(SDNode *N);
249 SDValue visitANDLike(SDValue N0, SDValue N1, SDNode *LocReference);
250 SDValue visitOR(SDNode *N);
251 SDValue visitORLike(SDValue N0, SDValue N1, SDNode *LocReference);
252 SDValue visitXOR(SDNode *N);
253 SDValue SimplifyVBinOp(SDNode *N);
254 SDValue visitSHL(SDNode *N);
255 SDValue visitSRA(SDNode *N);
256 SDValue visitSRL(SDNode *N);
257 SDValue visitRotate(SDNode *N);
258 SDValue visitCTLZ(SDNode *N);
259 SDValue visitCTLZ_ZERO_UNDEF(SDNode *N);
260 SDValue visitCTTZ(SDNode *N);
261 SDValue visitCTTZ_ZERO_UNDEF(SDNode *N);
262 SDValue visitCTPOP(SDNode *N);
263 SDValue visitSELECT(SDNode *N);
264 SDValue visitVSELECT(SDNode *N);
265 SDValue visitSELECT_CC(SDNode *N);
266 SDValue visitSETCC(SDNode *N);
267 SDValue visitSIGN_EXTEND(SDNode *N);
268 SDValue visitZERO_EXTEND(SDNode *N);
269 SDValue visitANY_EXTEND(SDNode *N);
270 SDValue visitSIGN_EXTEND_INREG(SDNode *N);
271 SDValue visitTRUNCATE(SDNode *N);
272 SDValue visitBITCAST(SDNode *N);
273 SDValue visitBUILD_PAIR(SDNode *N);
274 SDValue visitFADD(SDNode *N);
275 SDValue visitFSUB(SDNode *N);
276 SDValue visitFMUL(SDNode *N);
277 SDValue visitFMA(SDNode *N);
278 SDValue visitFDIV(SDNode *N);
279 SDValue visitFREM(SDNode *N);
280 SDValue visitFSQRT(SDNode *N);
281 SDValue visitFCOPYSIGN(SDNode *N);
282 SDValue visitSINT_TO_FP(SDNode *N);
283 SDValue visitUINT_TO_FP(SDNode *N);
284 SDValue visitFP_TO_SINT(SDNode *N);
285 SDValue visitFP_TO_UINT(SDNode *N);
286 SDValue visitFP_ROUND(SDNode *N);
287 SDValue visitFP_ROUND_INREG(SDNode *N);
288 SDValue visitFP_EXTEND(SDNode *N);
289 SDValue visitFNEG(SDNode *N);
290 SDValue visitFABS(SDNode *N);
291 SDValue visitFCEIL(SDNode *N);
292 SDValue visitFTRUNC(SDNode *N);
293 SDValue visitFFLOOR(SDNode *N);
294 SDValue visitFMINNUM(SDNode *N);
295 SDValue visitFMAXNUM(SDNode *N);
296 SDValue visitBRCOND(SDNode *N);
297 SDValue visitBR_CC(SDNode *N);
298 SDValue visitLOAD(SDNode *N);
299 SDValue visitSTORE(SDNode *N);
300 SDValue visitINSERT_VECTOR_ELT(SDNode *N);
301 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
302 SDValue visitBUILD_VECTOR(SDNode *N);
303 SDValue visitCONCAT_VECTORS(SDNode *N);
304 SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
305 SDValue visitVECTOR_SHUFFLE(SDNode *N);
306 SDValue visitSCALAR_TO_VECTOR(SDNode *N);
307 SDValue visitINSERT_SUBVECTOR(SDNode *N);
308 SDValue visitMLOAD(SDNode *N);
309 SDValue visitMSTORE(SDNode *N);
311 SDValue visitFADDForFMACombine(SDNode *N);
312 SDValue visitFSUBForFMACombine(SDNode *N);
314 SDValue XformToShuffleWithZero(SDNode *N);
315 SDValue ReassociateOps(unsigned Opc, SDLoc DL, SDValue LHS, SDValue RHS);
317 SDValue visitShiftByConstant(SDNode *N, ConstantSDNode *Amt);
319 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
320 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
321 SDValue SimplifySelect(SDLoc DL, SDValue N0, SDValue N1, SDValue N2);
322 SDValue SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1, SDValue N2,
323 SDValue N3, ISD::CondCode CC,
324 bool NotExtCompare = false);
325 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
326 SDLoc DL, bool foldBooleans = true);
328 bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
330 bool isOneUseSetCC(SDValue N) const;
332 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
334 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
335 SDValue CombineExtLoad(SDNode *N);
336 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
337 SDValue BuildSDIV(SDNode *N);
338 SDValue BuildSDIVPow2(SDNode *N);
339 SDValue BuildUDIV(SDNode *N);
340 SDValue BuildReciprocalEstimate(SDValue Op);
341 SDValue BuildRsqrtEstimate(SDValue Op);
342 SDValue BuildRsqrtNROneConst(SDValue Op, SDValue Est, unsigned Iterations);
343 SDValue BuildRsqrtNRTwoConst(SDValue Op, SDValue Est, unsigned Iterations);
344 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
345 bool DemandHighBits = true);
346 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
347 SDNode *MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg,
348 SDValue InnerPos, SDValue InnerNeg,
349 unsigned PosOpcode, unsigned NegOpcode,
351 SDNode *MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL);
352 SDValue ReduceLoadWidth(SDNode *N);
353 SDValue ReduceLoadOpStoreWidth(SDNode *N);
354 SDValue TransformFPLoadStorePair(SDNode *N);
355 SDValue reduceBuildVecExtToExtBuildVec(SDNode *N);
356 SDValue reduceBuildVecConvertToConvertBuildVec(SDNode *N);
358 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
360 /// Walk up chain skipping non-aliasing memory nodes,
361 /// looking for aliasing nodes and adding them to the Aliases vector.
362 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
363 SmallVectorImpl<SDValue> &Aliases);
365 /// Return true if there is any possibility that the two addresses overlap.
366 bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const;
368 /// Walk up chain skipping non-aliasing memory nodes, looking for a better
369 /// chain (aliasing node.)
370 SDValue FindBetterChain(SDNode *N, SDValue Chain);
372 /// Holds a pointer to an LSBaseSDNode as well as information on where it
373 /// is located in a sequence of memory operations connected by a chain.
375 MemOpLink (LSBaseSDNode *N, int64_t Offset, unsigned Seq):
376 MemNode(N), OffsetFromBase(Offset), SequenceNum(Seq) { }
377 // Ptr to the mem node.
378 LSBaseSDNode *MemNode;
379 // Offset from the base ptr.
380 int64_t OffsetFromBase;
381 // What is the sequence number of this mem node.
382 // Lowest mem operand in the DAG starts at zero.
383 unsigned SequenceNum;
386 /// This is a helper function for MergeConsecutiveStores. When the source
387 /// elements of the consecutive stores are all constants or all extracted
388 /// vector elements, try to merge them into one larger store.
389 /// \return True if a merged store was created.
390 bool MergeStoresOfConstantsOrVecElts(SmallVectorImpl<MemOpLink> &StoreNodes,
391 EVT MemVT, unsigned NumElem,
392 bool IsConstantSrc, bool UseVector);
394 /// Merge consecutive store operations into a wide store.
395 /// This optimization uses wide integers or vectors when possible.
396 /// \return True if some memory operations were changed.
397 bool MergeConsecutiveStores(StoreSDNode *N);
399 /// \brief Try to transform a truncation where C is a constant:
400 /// (trunc (and X, C)) -> (and (trunc X), (trunc C))
402 /// \p N needs to be a truncation and its first operand an AND. Other
403 /// requirements are checked by the function (e.g. that trunc is
404 /// single-use) and if missed an empty SDValue is returned.
405 SDValue distributeTruncateThroughAnd(SDNode *N);
408 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
409 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
410 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {
411 auto *F = DAG.getMachineFunction().getFunction();
412 ForCodeSize = F->hasFnAttribute(Attribute::OptimizeForSize) ||
413 F->hasFnAttribute(Attribute::MinSize);
416 /// Runs the dag combiner on all nodes in the work list
417 void Run(CombineLevel AtLevel);
419 SelectionDAG &getDAG() const { return DAG; }
421 /// Returns a type large enough to hold any valid shift amount - before type
422 /// legalization these can be huge.
423 EVT getShiftAmountTy(EVT LHSTy) {
424 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
425 if (LHSTy.isVector())
427 return LegalTypes ? TLI.getScalarShiftAmountTy(LHSTy)
428 : TLI.getPointerTy();
431 /// This method returns true if we are running before type legalization or
432 /// if the specified VT is legal.
433 bool isTypeLegal(const EVT &VT) {
434 if (!LegalTypes) return true;
435 return TLI.isTypeLegal(VT);
438 /// Convenience wrapper around TargetLowering::getSetCCResultType
439 EVT getSetCCResultType(EVT VT) const {
440 return TLI.getSetCCResultType(*DAG.getContext(), VT);
447 /// This class is a DAGUpdateListener that removes any deleted
448 /// nodes from the worklist.
449 class WorklistRemover : public SelectionDAG::DAGUpdateListener {
452 explicit WorklistRemover(DAGCombiner &dc)
453 : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {}
455 void NodeDeleted(SDNode *N, SDNode *E) override {
456 DC.removeFromWorklist(N);
461 //===----------------------------------------------------------------------===//
462 // TargetLowering::DAGCombinerInfo implementation
463 //===----------------------------------------------------------------------===//
465 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
466 ((DAGCombiner*)DC)->AddToWorklist(N);
469 void TargetLowering::DAGCombinerInfo::RemoveFromWorklist(SDNode *N) {
470 ((DAGCombiner*)DC)->removeFromWorklist(N);
473 SDValue TargetLowering::DAGCombinerInfo::
474 CombineTo(SDNode *N, ArrayRef<SDValue> To, bool AddTo) {
475 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
478 SDValue TargetLowering::DAGCombinerInfo::
479 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
480 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
484 SDValue TargetLowering::DAGCombinerInfo::
485 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
486 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
489 void TargetLowering::DAGCombinerInfo::
490 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
491 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
494 //===----------------------------------------------------------------------===//
496 //===----------------------------------------------------------------------===//
498 void DAGCombiner::deleteAndRecombine(SDNode *N) {
499 removeFromWorklist(N);
501 // If the operands of this node are only used by the node, they will now be
502 // dead. Make sure to re-visit them and recursively delete dead nodes.
503 for (const SDValue &Op : N->ops())
504 // For an operand generating multiple values, one of the values may
505 // become dead allowing further simplification (e.g. split index
506 // arithmetic from an indexed load).
507 if (Op->hasOneUse() || Op->getNumValues() > 1)
508 AddToWorklist(Op.getNode());
513 /// Return 1 if we can compute the negated form of the specified expression for
514 /// the same cost as the expression itself, or 2 if we can compute the negated
515 /// form more cheaply than the expression itself.
516 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
517 const TargetLowering &TLI,
518 const TargetOptions *Options,
519 unsigned Depth = 0) {
520 // fneg is removable even if it has multiple uses.
521 if (Op.getOpcode() == ISD::FNEG) return 2;
523 // Don't allow anything with multiple uses.
524 if (!Op.hasOneUse()) return 0;
526 // Don't recurse exponentially.
527 if (Depth > 6) return 0;
529 switch (Op.getOpcode()) {
530 default: return false;
531 case ISD::ConstantFP:
532 // Don't invert constant FP values after legalize. The negated constant
533 // isn't necessarily legal.
534 return LegalOperations ? 0 : 1;
536 // FIXME: determine better conditions for this xform.
537 if (!Options->UnsafeFPMath) return 0;
539 // After operation legalization, it might not be legal to create new FSUBs.
540 if (LegalOperations &&
541 !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType()))
544 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
545 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
548 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
549 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
552 // We can't turn -(A-B) into B-A when we honor signed zeros.
553 if (!Options->UnsafeFPMath) return 0;
555 // fold (fneg (fsub A, B)) -> (fsub B, A)
560 if (Options->HonorSignDependentRoundingFPMath()) return 0;
562 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
563 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
567 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
573 return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options,
578 /// If isNegatibleForFree returns true, return the newly negated expression.
579 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
580 bool LegalOperations, unsigned Depth = 0) {
581 const TargetOptions &Options = DAG.getTarget().Options;
582 // fneg is removable even if it has multiple uses.
583 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
585 // Don't allow anything with multiple uses.
586 assert(Op.hasOneUse() && "Unknown reuse!");
588 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
589 switch (Op.getOpcode()) {
590 default: llvm_unreachable("Unknown code");
591 case ISD::ConstantFP: {
592 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
594 return DAG.getConstantFP(V, Op.getValueType());
597 // FIXME: determine better conditions for this xform.
598 assert(Options.UnsafeFPMath);
600 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
601 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
602 DAG.getTargetLoweringInfo(), &Options, Depth+1))
603 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
604 GetNegatedExpression(Op.getOperand(0), DAG,
605 LegalOperations, Depth+1),
607 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
608 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
609 GetNegatedExpression(Op.getOperand(1), DAG,
610 LegalOperations, Depth+1),
613 // We can't turn -(A-B) into B-A when we honor signed zeros.
614 assert(Options.UnsafeFPMath);
616 // fold (fneg (fsub 0, B)) -> B
617 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
618 if (N0CFP->getValueAPF().isZero())
619 return Op.getOperand(1);
621 // fold (fneg (fsub A, B)) -> (fsub B, A)
622 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
623 Op.getOperand(1), Op.getOperand(0));
627 assert(!Options.HonorSignDependentRoundingFPMath());
629 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
630 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
631 DAG.getTargetLoweringInfo(), &Options, Depth+1))
632 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
633 GetNegatedExpression(Op.getOperand(0), DAG,
634 LegalOperations, Depth+1),
637 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
638 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
640 GetNegatedExpression(Op.getOperand(1), DAG,
641 LegalOperations, Depth+1));
645 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
646 GetNegatedExpression(Op.getOperand(0), DAG,
647 LegalOperations, Depth+1));
649 return DAG.getNode(ISD::FP_ROUND, SDLoc(Op), Op.getValueType(),
650 GetNegatedExpression(Op.getOperand(0), DAG,
651 LegalOperations, Depth+1),
656 // Return true if this node is a setcc, or is a select_cc
657 // that selects between the target values used for true and false, making it
658 // equivalent to a setcc. Also, set the incoming LHS, RHS, and CC references to
659 // the appropriate nodes based on the type of node we are checking. This
660 // simplifies life a bit for the callers.
661 bool DAGCombiner::isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
663 if (N.getOpcode() == ISD::SETCC) {
664 LHS = N.getOperand(0);
665 RHS = N.getOperand(1);
666 CC = N.getOperand(2);
670 if (N.getOpcode() != ISD::SELECT_CC ||
671 !TLI.isConstTrueVal(N.getOperand(2).getNode()) ||
672 !TLI.isConstFalseVal(N.getOperand(3).getNode()))
675 if (TLI.getBooleanContents(N.getValueType()) ==
676 TargetLowering::UndefinedBooleanContent)
679 LHS = N.getOperand(0);
680 RHS = N.getOperand(1);
681 CC = N.getOperand(4);
685 /// Return true if this is a SetCC-equivalent operation with only one use.
686 /// If this is true, it allows the users to invert the operation for free when
687 /// it is profitable to do so.
688 bool DAGCombiner::isOneUseSetCC(SDValue N) const {
690 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
695 /// Returns true if N is a BUILD_VECTOR node whose
696 /// elements are all the same constant or undefined.
697 static bool isConstantSplatVector(SDNode *N, APInt& SplatValue) {
698 BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(N);
703 unsigned SplatBitSize;
705 EVT EltVT = N->getValueType(0).getVectorElementType();
706 return (C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
708 EltVT.getSizeInBits() >= SplatBitSize);
711 // \brief Returns the SDNode if it is a constant integer BuildVector
712 // or constant integer.
713 static SDNode *isConstantIntBuildVectorOrConstantInt(SDValue N) {
714 if (isa<ConstantSDNode>(N))
716 if (ISD::isBuildVectorOfConstantSDNodes(N.getNode()))
721 // \brief Returns the SDNode if it is a constant float BuildVector
722 // or constant float.
723 static SDNode *isConstantFPBuildVectorOrConstantFP(SDValue N) {
724 if (isa<ConstantFPSDNode>(N))
726 if (ISD::isBuildVectorOfConstantFPSDNodes(N.getNode()))
731 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
733 static ConstantSDNode *isConstOrConstSplat(SDValue N) {
734 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N))
737 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
738 BitVector UndefElements;
739 ConstantSDNode *CN = BV->getConstantSplatNode(&UndefElements);
741 // BuildVectors can truncate their operands. Ignore that case here.
742 // FIXME: We blindly ignore splats which include undef which is overly
744 if (CN && UndefElements.none() &&
745 CN->getValueType(0) == N.getValueType().getScalarType())
752 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
754 static ConstantFPSDNode *isConstOrConstSplatFP(SDValue N) {
755 if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N))
758 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
759 BitVector UndefElements;
760 ConstantFPSDNode *CN = BV->getConstantFPSplatNode(&UndefElements);
762 if (CN && UndefElements.none())
769 SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL,
770 SDValue N0, SDValue N1) {
771 EVT VT = N0.getValueType();
772 if (N0.getOpcode() == Opc) {
773 if (SDNode *L = isConstantIntBuildVectorOrConstantInt(N0.getOperand(1))) {
774 if (SDNode *R = isConstantIntBuildVectorOrConstantInt(N1)) {
775 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
776 if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, L, R))
777 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
780 if (N0.hasOneUse()) {
781 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one
783 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1);
784 if (!OpNode.getNode())
786 AddToWorklist(OpNode.getNode());
787 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
792 if (N1.getOpcode() == Opc) {
793 if (SDNode *R = isConstantIntBuildVectorOrConstantInt(N1.getOperand(1))) {
794 if (SDNode *L = isConstantIntBuildVectorOrConstantInt(N0)) {
795 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
796 if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, R, L))
797 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
800 if (N1.hasOneUse()) {
801 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one
803 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N1.getOperand(0), N0);
804 if (!OpNode.getNode())
806 AddToWorklist(OpNode.getNode());
807 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
815 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
817 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
819 DEBUG(dbgs() << "\nReplacing.1 ";
821 dbgs() << "\nWith: ";
822 To[0].getNode()->dump(&DAG);
823 dbgs() << " and " << NumTo-1 << " other values\n");
824 for (unsigned i = 0, e = NumTo; i != e; ++i)
825 assert((!To[i].getNode() ||
826 N->getValueType(i) == To[i].getValueType()) &&
827 "Cannot combine value to value of different type!");
829 WorklistRemover DeadNodes(*this);
830 DAG.ReplaceAllUsesWith(N, To);
832 // Push the new nodes and any users onto the worklist
833 for (unsigned i = 0, e = NumTo; i != e; ++i) {
834 if (To[i].getNode()) {
835 AddToWorklist(To[i].getNode());
836 AddUsersToWorklist(To[i].getNode());
841 // Finally, if the node is now dead, remove it from the graph. The node
842 // may not be dead if the replacement process recursively simplified to
843 // something else needing this node.
845 deleteAndRecombine(N);
846 return SDValue(N, 0);
850 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
851 // Replace all uses. If any nodes become isomorphic to other nodes and
852 // are deleted, make sure to remove them from our worklist.
853 WorklistRemover DeadNodes(*this);
854 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New);
856 // Push the new node and any (possibly new) users onto the worklist.
857 AddToWorklist(TLO.New.getNode());
858 AddUsersToWorklist(TLO.New.getNode());
860 // Finally, if the node is now dead, remove it from the graph. The node
861 // may not be dead if the replacement process recursively simplified to
862 // something else needing this node.
863 if (TLO.Old.getNode()->use_empty())
864 deleteAndRecombine(TLO.Old.getNode());
867 /// Check the specified integer node value to see if it can be simplified or if
868 /// things it uses can be simplified by bit propagation. If so, return true.
869 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
870 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
871 APInt KnownZero, KnownOne;
872 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
876 AddToWorklist(Op.getNode());
878 // Replace the old value with the new one.
880 DEBUG(dbgs() << "\nReplacing.2 ";
881 TLO.Old.getNode()->dump(&DAG);
882 dbgs() << "\nWith: ";
883 TLO.New.getNode()->dump(&DAG);
886 CommitTargetLoweringOpt(TLO);
890 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
892 EVT VT = Load->getValueType(0);
893 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
895 DEBUG(dbgs() << "\nReplacing.9 ";
897 dbgs() << "\nWith: ";
898 Trunc.getNode()->dump(&DAG);
900 WorklistRemover DeadNodes(*this);
901 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc);
902 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1));
903 deleteAndRecombine(Load);
904 AddToWorklist(Trunc.getNode());
907 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
910 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
911 EVT MemVT = LD->getMemoryVT();
912 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
913 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, PVT, MemVT) ? ISD::ZEXTLOAD
915 : LD->getExtensionType();
917 return DAG.getExtLoad(ExtType, dl, PVT,
918 LD->getChain(), LD->getBasePtr(),
919 MemVT, LD->getMemOperand());
922 unsigned Opc = Op.getOpcode();
925 case ISD::AssertSext:
926 return DAG.getNode(ISD::AssertSext, dl, PVT,
927 SExtPromoteOperand(Op.getOperand(0), PVT),
929 case ISD::AssertZext:
930 return DAG.getNode(ISD::AssertZext, dl, PVT,
931 ZExtPromoteOperand(Op.getOperand(0), PVT),
933 case ISD::Constant: {
935 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
936 return DAG.getNode(ExtOpc, dl, PVT, Op);
940 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
942 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
945 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
946 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
948 EVT OldVT = Op.getValueType();
950 bool Replace = false;
951 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
952 if (!NewOp.getNode())
954 AddToWorklist(NewOp.getNode());
957 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
958 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
959 DAG.getValueType(OldVT));
962 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
963 EVT OldVT = Op.getValueType();
965 bool Replace = false;
966 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
967 if (!NewOp.getNode())
969 AddToWorklist(NewOp.getNode());
972 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
973 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
976 /// Promote the specified integer binary operation if the target indicates it is
977 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
978 /// i32 since i16 instructions are longer.
979 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
980 if (!LegalOperations)
983 EVT VT = Op.getValueType();
984 if (VT.isVector() || !VT.isInteger())
987 // If operation type is 'undesirable', e.g. i16 on x86, consider
989 unsigned Opc = Op.getOpcode();
990 if (TLI.isTypeDesirableForOp(Opc, VT))
994 // Consult target whether it is a good idea to promote this operation and
995 // what's the right type to promote it to.
996 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
997 assert(PVT != VT && "Don't know what type to promote to!");
999 bool Replace0 = false;
1000 SDValue N0 = Op.getOperand(0);
1001 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
1005 bool Replace1 = false;
1006 SDValue N1 = Op.getOperand(1);
1011 NN1 = PromoteOperand(N1, PVT, Replace1);
1016 AddToWorklist(NN0.getNode());
1018 AddToWorklist(NN1.getNode());
1021 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
1023 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
1025 DEBUG(dbgs() << "\nPromoting ";
1026 Op.getNode()->dump(&DAG));
1028 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1029 DAG.getNode(Opc, dl, PVT, NN0, NN1));
1034 /// Promote the specified integer shift operation if the target indicates it is
1035 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
1036 /// i32 since i16 instructions are longer.
1037 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
1038 if (!LegalOperations)
1041 EVT VT = Op.getValueType();
1042 if (VT.isVector() || !VT.isInteger())
1045 // If operation type is 'undesirable', e.g. i16 on x86, consider
1047 unsigned Opc = Op.getOpcode();
1048 if (TLI.isTypeDesirableForOp(Opc, VT))
1052 // Consult target whether it is a good idea to promote this operation and
1053 // what's the right type to promote it to.
1054 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1055 assert(PVT != VT && "Don't know what type to promote to!");
1057 bool Replace = false;
1058 SDValue N0 = Op.getOperand(0);
1059 if (Opc == ISD::SRA)
1060 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
1061 else if (Opc == ISD::SRL)
1062 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
1064 N0 = PromoteOperand(N0, PVT, Replace);
1068 AddToWorklist(N0.getNode());
1070 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
1072 DEBUG(dbgs() << "\nPromoting ";
1073 Op.getNode()->dump(&DAG));
1075 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1076 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
1081 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
1082 if (!LegalOperations)
1085 EVT VT = Op.getValueType();
1086 if (VT.isVector() || !VT.isInteger())
1089 // If operation type is 'undesirable', e.g. i16 on x86, consider
1091 unsigned Opc = Op.getOpcode();
1092 if (TLI.isTypeDesirableForOp(Opc, VT))
1096 // Consult target whether it is a good idea to promote this operation and
1097 // what's the right type to promote it to.
1098 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1099 assert(PVT != VT && "Don't know what type to promote to!");
1100 // fold (aext (aext x)) -> (aext x)
1101 // fold (aext (zext x)) -> (zext x)
1102 // fold (aext (sext x)) -> (sext x)
1103 DEBUG(dbgs() << "\nPromoting ";
1104 Op.getNode()->dump(&DAG));
1105 return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0));
1110 bool DAGCombiner::PromoteLoad(SDValue Op) {
1111 if (!LegalOperations)
1114 EVT VT = Op.getValueType();
1115 if (VT.isVector() || !VT.isInteger())
1118 // If operation type is 'undesirable', e.g. i16 on x86, consider
1120 unsigned Opc = Op.getOpcode();
1121 if (TLI.isTypeDesirableForOp(Opc, VT))
1125 // Consult target whether it is a good idea to promote this operation and
1126 // what's the right type to promote it to.
1127 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1128 assert(PVT != VT && "Don't know what type to promote to!");
1131 SDNode *N = Op.getNode();
1132 LoadSDNode *LD = cast<LoadSDNode>(N);
1133 EVT MemVT = LD->getMemoryVT();
1134 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
1135 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, PVT, MemVT) ? ISD::ZEXTLOAD
1137 : LD->getExtensionType();
1138 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
1139 LD->getChain(), LD->getBasePtr(),
1140 MemVT, LD->getMemOperand());
1141 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
1143 DEBUG(dbgs() << "\nPromoting ";
1146 Result.getNode()->dump(&DAG);
1148 WorklistRemover DeadNodes(*this);
1149 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
1150 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1));
1151 deleteAndRecombine(N);
1152 AddToWorklist(Result.getNode());
1158 /// \brief Recursively delete a node which has no uses and any operands for
1159 /// which it is the only use.
1161 /// Note that this both deletes the nodes and removes them from the worklist.
1162 /// It also adds any nodes who have had a user deleted to the worklist as they
1163 /// may now have only one use and subject to other combines.
1164 bool DAGCombiner::recursivelyDeleteUnusedNodes(SDNode *N) {
1165 if (!N->use_empty())
1168 SmallSetVector<SDNode *, 16> Nodes;
1171 N = Nodes.pop_back_val();
1175 if (N->use_empty()) {
1176 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1177 Nodes.insert(N->getOperand(i).getNode());
1179 removeFromWorklist(N);
1184 } while (!Nodes.empty());
1188 //===----------------------------------------------------------------------===//
1189 // Main DAG Combiner implementation
1190 //===----------------------------------------------------------------------===//
1192 void DAGCombiner::Run(CombineLevel AtLevel) {
1193 // set the instance variables, so that the various visit routines may use it.
1195 LegalOperations = Level >= AfterLegalizeVectorOps;
1196 LegalTypes = Level >= AfterLegalizeTypes;
1198 // Add all the dag nodes to the worklist.
1199 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
1200 E = DAG.allnodes_end(); I != E; ++I)
1203 // Create a dummy node (which is not added to allnodes), that adds a reference
1204 // to the root node, preventing it from being deleted, and tracking any
1205 // changes of the root.
1206 HandleSDNode Dummy(DAG.getRoot());
1208 // while the worklist isn't empty, find a node and
1209 // try and combine it.
1210 while (!WorklistMap.empty()) {
1212 // The Worklist holds the SDNodes in order, but it may contain null entries.
1214 N = Worklist.pop_back_val();
1217 bool GoodWorklistEntry = WorklistMap.erase(N);
1218 (void)GoodWorklistEntry;
1219 assert(GoodWorklistEntry &&
1220 "Found a worklist entry without a corresponding map entry!");
1222 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1223 // N is deleted from the DAG, since they too may now be dead or may have a
1224 // reduced number of uses, allowing other xforms.
1225 if (recursivelyDeleteUnusedNodes(N))
1228 WorklistRemover DeadNodes(*this);
1230 // If this combine is running after legalizing the DAG, re-legalize any
1231 // nodes pulled off the worklist.
1232 if (Level == AfterLegalizeDAG) {
1233 SmallSetVector<SDNode *, 16> UpdatedNodes;
1234 bool NIsValid = DAG.LegalizeOp(N, UpdatedNodes);
1236 for (SDNode *LN : UpdatedNodes) {
1238 AddUsersToWorklist(LN);
1244 DEBUG(dbgs() << "\nCombining: "; N->dump(&DAG));
1246 // Add any operands of the new node which have not yet been combined to the
1247 // worklist as well. Because the worklist uniques things already, this
1248 // won't repeatedly process the same operand.
1249 CombinedNodes.insert(N);
1250 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1251 if (!CombinedNodes.count(N->getOperand(i).getNode()))
1252 AddToWorklist(N->getOperand(i).getNode());
1254 SDValue RV = combine(N);
1261 // If we get back the same node we passed in, rather than a new node or
1262 // zero, we know that the node must have defined multiple values and
1263 // CombineTo was used. Since CombineTo takes care of the worklist
1264 // mechanics for us, we have no work to do in this case.
1265 if (RV.getNode() == N)
1268 assert(N->getOpcode() != ISD::DELETED_NODE &&
1269 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1270 "Node was deleted but visit returned new node!");
1272 DEBUG(dbgs() << " ... into: ";
1273 RV.getNode()->dump(&DAG));
1275 // Transfer debug value.
1276 DAG.TransferDbgValues(SDValue(N, 0), RV);
1277 if (N->getNumValues() == RV.getNode()->getNumValues())
1278 DAG.ReplaceAllUsesWith(N, RV.getNode());
1280 assert(N->getValueType(0) == RV.getValueType() &&
1281 N->getNumValues() == 1 && "Type mismatch");
1283 DAG.ReplaceAllUsesWith(N, &OpV);
1286 // Push the new node and any users onto the worklist
1287 AddToWorklist(RV.getNode());
1288 AddUsersToWorklist(RV.getNode());
1290 // Finally, if the node is now dead, remove it from the graph. The node
1291 // may not be dead if the replacement process recursively simplified to
1292 // something else needing this node. This will also take care of adding any
1293 // operands which have lost a user to the worklist.
1294 recursivelyDeleteUnusedNodes(N);
1297 // If the root changed (e.g. it was a dead load, update the root).
1298 DAG.setRoot(Dummy.getValue());
1299 DAG.RemoveDeadNodes();
1302 SDValue DAGCombiner::visit(SDNode *N) {
1303 switch (N->getOpcode()) {
1305 case ISD::TokenFactor: return visitTokenFactor(N);
1306 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1307 case ISD::ADD: return visitADD(N);
1308 case ISD::SUB: return visitSUB(N);
1309 case ISD::ADDC: return visitADDC(N);
1310 case ISD::SUBC: return visitSUBC(N);
1311 case ISD::ADDE: return visitADDE(N);
1312 case ISD::SUBE: return visitSUBE(N);
1313 case ISD::MUL: return visitMUL(N);
1314 case ISD::SDIV: return visitSDIV(N);
1315 case ISD::UDIV: return visitUDIV(N);
1316 case ISD::SREM: return visitSREM(N);
1317 case ISD::UREM: return visitUREM(N);
1318 case ISD::MULHU: return visitMULHU(N);
1319 case ISD::MULHS: return visitMULHS(N);
1320 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1321 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1322 case ISD::SMULO: return visitSMULO(N);
1323 case ISD::UMULO: return visitUMULO(N);
1324 case ISD::SDIVREM: return visitSDIVREM(N);
1325 case ISD::UDIVREM: return visitUDIVREM(N);
1326 case ISD::AND: return visitAND(N);
1327 case ISD::OR: return visitOR(N);
1328 case ISD::XOR: return visitXOR(N);
1329 case ISD::SHL: return visitSHL(N);
1330 case ISD::SRA: return visitSRA(N);
1331 case ISD::SRL: return visitSRL(N);
1333 case ISD::ROTL: return visitRotate(N);
1334 case ISD::CTLZ: return visitCTLZ(N);
1335 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1336 case ISD::CTTZ: return visitCTTZ(N);
1337 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1338 case ISD::CTPOP: return visitCTPOP(N);
1339 case ISD::SELECT: return visitSELECT(N);
1340 case ISD::VSELECT: return visitVSELECT(N);
1341 case ISD::SELECT_CC: return visitSELECT_CC(N);
1342 case ISD::SETCC: return visitSETCC(N);
1343 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1344 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1345 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1346 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1347 case ISD::TRUNCATE: return visitTRUNCATE(N);
1348 case ISD::BITCAST: return visitBITCAST(N);
1349 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1350 case ISD::FADD: return visitFADD(N);
1351 case ISD::FSUB: return visitFSUB(N);
1352 case ISD::FMUL: return visitFMUL(N);
1353 case ISD::FMA: return visitFMA(N);
1354 case ISD::FDIV: return visitFDIV(N);
1355 case ISD::FREM: return visitFREM(N);
1356 case ISD::FSQRT: return visitFSQRT(N);
1357 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1358 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1359 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1360 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1361 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1362 case ISD::FP_ROUND: return visitFP_ROUND(N);
1363 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1364 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1365 case ISD::FNEG: return visitFNEG(N);
1366 case ISD::FABS: return visitFABS(N);
1367 case ISD::FFLOOR: return visitFFLOOR(N);
1368 case ISD::FMINNUM: return visitFMINNUM(N);
1369 case ISD::FMAXNUM: return visitFMAXNUM(N);
1370 case ISD::FCEIL: return visitFCEIL(N);
1371 case ISD::FTRUNC: return visitFTRUNC(N);
1372 case ISD::BRCOND: return visitBRCOND(N);
1373 case ISD::BR_CC: return visitBR_CC(N);
1374 case ISD::LOAD: return visitLOAD(N);
1375 case ISD::STORE: return visitSTORE(N);
1376 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1377 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1378 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1379 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1380 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1381 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1382 case ISD::SCALAR_TO_VECTOR: return visitSCALAR_TO_VECTOR(N);
1383 case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N);
1384 case ISD::MLOAD: return visitMLOAD(N);
1385 case ISD::MSTORE: return visitMSTORE(N);
1390 SDValue DAGCombiner::combine(SDNode *N) {
1391 SDValue RV = visit(N);
1393 // If nothing happened, try a target-specific DAG combine.
1394 if (!RV.getNode()) {
1395 assert(N->getOpcode() != ISD::DELETED_NODE &&
1396 "Node was deleted but visit returned NULL!");
1398 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1399 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1401 // Expose the DAG combiner to the target combiner impls.
1402 TargetLowering::DAGCombinerInfo
1403 DagCombineInfo(DAG, Level, false, this);
1405 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1409 // If nothing happened still, try promoting the operation.
1410 if (!RV.getNode()) {
1411 switch (N->getOpcode()) {
1419 RV = PromoteIntBinOp(SDValue(N, 0));
1424 RV = PromoteIntShiftOp(SDValue(N, 0));
1426 case ISD::SIGN_EXTEND:
1427 case ISD::ZERO_EXTEND:
1428 case ISD::ANY_EXTEND:
1429 RV = PromoteExtend(SDValue(N, 0));
1432 if (PromoteLoad(SDValue(N, 0)))
1438 // If N is a commutative binary node, try commuting it to enable more
1440 if (!RV.getNode() && SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1441 N->getNumValues() == 1) {
1442 SDValue N0 = N->getOperand(0);
1443 SDValue N1 = N->getOperand(1);
1445 // Constant operands are canonicalized to RHS.
1446 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1447 SDValue Ops[] = {N1, N0};
1449 if (const BinaryWithFlagsSDNode *BinNode =
1450 dyn_cast<BinaryWithFlagsSDNode>(N)) {
1451 CSENode = DAG.getNodeIfExists(
1452 N->getOpcode(), N->getVTList(), Ops, BinNode->hasNoUnsignedWrap(),
1453 BinNode->hasNoSignedWrap(), BinNode->isExact());
1455 CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops);
1458 return SDValue(CSENode, 0);
1465 /// Given a node, return its input chain if it has one, otherwise return a null
1467 static SDValue getInputChainForNode(SDNode *N) {
1468 if (unsigned NumOps = N->getNumOperands()) {
1469 if (N->getOperand(0).getValueType() == MVT::Other)
1470 return N->getOperand(0);
1471 if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1472 return N->getOperand(NumOps-1);
1473 for (unsigned i = 1; i < NumOps-1; ++i)
1474 if (N->getOperand(i).getValueType() == MVT::Other)
1475 return N->getOperand(i);
1480 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1481 // If N has two operands, where one has an input chain equal to the other,
1482 // the 'other' chain is redundant.
1483 if (N->getNumOperands() == 2) {
1484 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1485 return N->getOperand(0);
1486 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1487 return N->getOperand(1);
1490 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1491 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1492 SmallPtrSet<SDNode*, 16> SeenOps;
1493 bool Changed = false; // If we should replace this token factor.
1495 // Start out with this token factor.
1498 // Iterate through token factors. The TFs grows when new token factors are
1500 for (unsigned i = 0; i < TFs.size(); ++i) {
1501 SDNode *TF = TFs[i];
1503 // Check each of the operands.
1504 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1505 SDValue Op = TF->getOperand(i);
1507 switch (Op.getOpcode()) {
1508 case ISD::EntryToken:
1509 // Entry tokens don't need to be added to the list. They are
1514 case ISD::TokenFactor:
1515 if (Op.hasOneUse() &&
1516 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1517 // Queue up for processing.
1518 TFs.push_back(Op.getNode());
1519 // Clean up in case the token factor is removed.
1520 AddToWorklist(Op.getNode());
1527 // Only add if it isn't already in the list.
1528 if (SeenOps.insert(Op.getNode()).second)
1539 // If we've changed things around then replace token factor.
1542 // The entry token is the only possible outcome.
1543 Result = DAG.getEntryNode();
1545 // New and improved token factor.
1546 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Ops);
1549 // Add users to worklist if AA is enabled, since it may introduce
1550 // a lot of new chained token factors while removing memory deps.
1551 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
1552 : DAG.getSubtarget().useAA();
1553 return CombineTo(N, Result, UseAA /*add to worklist*/);
1559 /// MERGE_VALUES can always be eliminated.
1560 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1561 WorklistRemover DeadNodes(*this);
1562 // Replacing results may cause a different MERGE_VALUES to suddenly
1563 // be CSE'd with N, and carry its uses with it. Iterate until no
1564 // uses remain, to ensure that the node can be safely deleted.
1565 // First add the users of this node to the work list so that they
1566 // can be tried again once they have new operands.
1567 AddUsersToWorklist(N);
1569 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1570 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1571 } while (!N->use_empty());
1572 deleteAndRecombine(N);
1573 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1576 SDValue DAGCombiner::visitADD(SDNode *N) {
1577 SDValue N0 = N->getOperand(0);
1578 SDValue N1 = N->getOperand(1);
1579 EVT VT = N0.getValueType();
1582 if (VT.isVector()) {
1583 if (SDValue FoldedVOp = SimplifyVBinOp(N))
1586 // fold (add x, 0) -> x, vector edition
1587 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1589 if (ISD::isBuildVectorAllZeros(N0.getNode()))
1593 // fold (add x, undef) -> undef
1594 if (N0.getOpcode() == ISD::UNDEF)
1596 if (N1.getOpcode() == ISD::UNDEF)
1598 // fold (add c1, c2) -> c1+c2
1599 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1600 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1602 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1603 // canonicalize constant to RHS
1604 if (isConstantIntBuildVectorOrConstantInt(N0) &&
1605 !isConstantIntBuildVectorOrConstantInt(N1))
1606 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, N0);
1607 // fold (add x, 0) -> x
1608 if (N1C && N1C->isNullValue())
1610 // fold (add Sym, c) -> Sym+c
1611 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1612 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1613 GA->getOpcode() == ISD::GlobalAddress)
1614 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1616 (uint64_t)N1C->getSExtValue());
1617 // fold ((c1-A)+c2) -> (c1+c2)-A
1618 if (N1C && N0.getOpcode() == ISD::SUB)
1619 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1620 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1621 DAG.getConstant(N1C->getAPIntValue()+
1622 N0C->getAPIntValue(), VT),
1625 if (SDValue RADD = ReassociateOps(ISD::ADD, SDLoc(N), N0, N1))
1627 // fold ((0-A) + B) -> B-A
1628 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1629 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1630 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, N0.getOperand(1));
1631 // fold (A + (0-B)) -> A-B
1632 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1633 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1634 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1.getOperand(1));
1635 // fold (A+(B-A)) -> B
1636 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1637 return N1.getOperand(0);
1638 // fold ((B-A)+A) -> B
1639 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1640 return N0.getOperand(0);
1641 // fold (A+(B-(A+C))) to (B-C)
1642 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1643 N0 == N1.getOperand(1).getOperand(0))
1644 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1645 N1.getOperand(1).getOperand(1));
1646 // fold (A+(B-(C+A))) to (B-C)
1647 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1648 N0 == N1.getOperand(1).getOperand(1))
1649 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1650 N1.getOperand(1).getOperand(0));
1651 // fold (A+((B-A)+or-C)) to (B+or-C)
1652 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1653 N1.getOperand(0).getOpcode() == ISD::SUB &&
1654 N0 == N1.getOperand(0).getOperand(1))
1655 return DAG.getNode(N1.getOpcode(), SDLoc(N), VT,
1656 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1658 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1659 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1660 SDValue N00 = N0.getOperand(0);
1661 SDValue N01 = N0.getOperand(1);
1662 SDValue N10 = N1.getOperand(0);
1663 SDValue N11 = N1.getOperand(1);
1665 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1666 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1667 DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10),
1668 DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11));
1671 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1672 return SDValue(N, 0);
1674 // fold (a+b) -> (a|b) iff a and b share no bits.
1675 if (VT.isInteger() && !VT.isVector()) {
1676 APInt LHSZero, LHSOne;
1677 APInt RHSZero, RHSOne;
1678 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1680 if (LHSZero.getBoolValue()) {
1681 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1683 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1684 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1685 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero){
1686 if (!LegalOperations || TLI.isOperationLegal(ISD::OR, VT))
1687 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1);
1692 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1693 if (N1.getOpcode() == ISD::SHL &&
1694 N1.getOperand(0).getOpcode() == ISD::SUB)
1695 if (ConstantSDNode *C =
1696 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1697 if (C->getAPIntValue() == 0)
1698 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0,
1699 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1700 N1.getOperand(0).getOperand(1),
1702 if (N0.getOpcode() == ISD::SHL &&
1703 N0.getOperand(0).getOpcode() == ISD::SUB)
1704 if (ConstantSDNode *C =
1705 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1706 if (C->getAPIntValue() == 0)
1707 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1,
1708 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1709 N0.getOperand(0).getOperand(1),
1712 if (N1.getOpcode() == ISD::AND) {
1713 SDValue AndOp0 = N1.getOperand(0);
1714 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1715 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1716 unsigned DestBits = VT.getScalarType().getSizeInBits();
1718 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1719 // and similar xforms where the inner op is either ~0 or 0.
1720 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1722 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1726 // add (sext i1), X -> sub X, (zext i1)
1727 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1728 N0.getOperand(0).getValueType() == MVT::i1 &&
1729 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1731 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1732 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1735 // add X, (sextinreg Y i1) -> sub X, (and Y 1)
1736 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
1737 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
1738 if (TN->getVT() == MVT::i1) {
1740 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
1741 DAG.getConstant(1, VT));
1742 return DAG.getNode(ISD::SUB, DL, VT, N0, ZExt);
1749 SDValue DAGCombiner::visitADDC(SDNode *N) {
1750 SDValue N0 = N->getOperand(0);
1751 SDValue N1 = N->getOperand(1);
1752 EVT VT = N0.getValueType();
1754 // If the flag result is dead, turn this into an ADD.
1755 if (!N->hasAnyUseOfValue(1))
1756 return CombineTo(N, DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N1),
1757 DAG.getNode(ISD::CARRY_FALSE,
1758 SDLoc(N), MVT::Glue));
1760 // canonicalize constant to RHS.
1761 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1762 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1764 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N1, N0);
1766 // fold (addc x, 0) -> x + no carry out
1767 if (N1C && N1C->isNullValue())
1768 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1769 SDLoc(N), MVT::Glue));
1771 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1772 APInt LHSZero, LHSOne;
1773 APInt RHSZero, RHSOne;
1774 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1776 if (LHSZero.getBoolValue()) {
1777 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1779 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1780 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1781 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1782 return CombineTo(N, DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1),
1783 DAG.getNode(ISD::CARRY_FALSE,
1784 SDLoc(N), MVT::Glue));
1790 SDValue DAGCombiner::visitADDE(SDNode *N) {
1791 SDValue N0 = N->getOperand(0);
1792 SDValue N1 = N->getOperand(1);
1793 SDValue CarryIn = N->getOperand(2);
1795 // canonicalize constant to RHS
1796 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1797 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1799 return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(),
1802 // fold (adde x, y, false) -> (addc x, y)
1803 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1804 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1);
1809 // Since it may not be valid to emit a fold to zero for vector initializers
1810 // check if we can before folding.
1811 static SDValue tryFoldToZero(SDLoc DL, const TargetLowering &TLI, EVT VT,
1813 bool LegalOperations, bool LegalTypes) {
1815 return DAG.getConstant(0, VT);
1816 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
1817 return DAG.getConstant(0, VT);
1821 SDValue DAGCombiner::visitSUB(SDNode *N) {
1822 SDValue N0 = N->getOperand(0);
1823 SDValue N1 = N->getOperand(1);
1824 EVT VT = N0.getValueType();
1827 if (VT.isVector()) {
1828 if (SDValue FoldedVOp = SimplifyVBinOp(N))
1831 // fold (sub x, 0) -> x, vector edition
1832 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1836 // fold (sub x, x) -> 0
1837 // FIXME: Refactor this and xor and other similar operations together.
1839 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
1840 // fold (sub c1, c2) -> c1-c2
1841 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1842 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1844 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1845 // fold (sub x, c) -> (add x, -c)
1847 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0,
1848 DAG.getConstant(-N1C->getAPIntValue(), VT));
1849 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1850 if (N0C && N0C->isAllOnesValue())
1851 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
1852 // fold A-(A-B) -> B
1853 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1854 return N1.getOperand(1);
1855 // fold (A+B)-A -> B
1856 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1857 return N0.getOperand(1);
1858 // fold (A+B)-B -> A
1859 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1860 return N0.getOperand(0);
1861 // fold C2-(A+C1) -> (C2-C1)-A
1862 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? nullptr :
1863 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1864 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1865 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(),
1867 return DAG.getNode(ISD::SUB, SDLoc(N), VT, NewC,
1870 // fold ((A+(B+or-C))-B) -> A+or-C
1871 if (N0.getOpcode() == ISD::ADD &&
1872 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1873 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1874 N0.getOperand(1).getOperand(0) == N1)
1875 return DAG.getNode(N0.getOperand(1).getOpcode(), SDLoc(N), VT,
1876 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1877 // fold ((A+(C+B))-B) -> A+C
1878 if (N0.getOpcode() == ISD::ADD &&
1879 N0.getOperand(1).getOpcode() == ISD::ADD &&
1880 N0.getOperand(1).getOperand(1) == N1)
1881 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1882 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1883 // fold ((A-(B-C))-C) -> A-B
1884 if (N0.getOpcode() == ISD::SUB &&
1885 N0.getOperand(1).getOpcode() == ISD::SUB &&
1886 N0.getOperand(1).getOperand(1) == N1)
1887 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1888 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1890 // If either operand of a sub is undef, the result is undef
1891 if (N0.getOpcode() == ISD::UNDEF)
1893 if (N1.getOpcode() == ISD::UNDEF)
1896 // If the relocation model supports it, consider symbol offsets.
1897 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1898 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1899 // fold (sub Sym, c) -> Sym-c
1900 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1901 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1903 (uint64_t)N1C->getSExtValue());
1904 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1905 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1906 if (GA->getGlobal() == GB->getGlobal())
1907 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1911 // sub X, (sextinreg Y i1) -> add X, (and Y 1)
1912 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
1913 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
1914 if (TN->getVT() == MVT::i1) {
1916 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
1917 DAG.getConstant(1, VT));
1918 return DAG.getNode(ISD::ADD, DL, VT, N0, ZExt);
1925 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1926 SDValue N0 = N->getOperand(0);
1927 SDValue N1 = N->getOperand(1);
1928 EVT VT = N0.getValueType();
1930 // If the flag result is dead, turn this into an SUB.
1931 if (!N->hasAnyUseOfValue(1))
1932 return CombineTo(N, DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1),
1933 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1936 // fold (subc x, x) -> 0 + no borrow
1938 return CombineTo(N, DAG.getConstant(0, VT),
1939 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1942 // fold (subc x, 0) -> x + no borrow
1943 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1944 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1945 if (N1C && N1C->isNullValue())
1946 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1949 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1950 if (N0C && N0C->isAllOnesValue())
1951 return CombineTo(N, DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0),
1952 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1958 SDValue DAGCombiner::visitSUBE(SDNode *N) {
1959 SDValue N0 = N->getOperand(0);
1960 SDValue N1 = N->getOperand(1);
1961 SDValue CarryIn = N->getOperand(2);
1963 // fold (sube x, y, false) -> (subc x, y)
1964 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1965 return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1);
1970 SDValue DAGCombiner::visitMUL(SDNode *N) {
1971 SDValue N0 = N->getOperand(0);
1972 SDValue N1 = N->getOperand(1);
1973 EVT VT = N0.getValueType();
1975 // fold (mul x, undef) -> 0
1976 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1977 return DAG.getConstant(0, VT);
1979 bool N0IsConst = false;
1980 bool N1IsConst = false;
1981 APInt ConstValue0, ConstValue1;
1983 if (VT.isVector()) {
1984 if (SDValue FoldedVOp = SimplifyVBinOp(N))
1987 N0IsConst = isConstantSplatVector(N0.getNode(), ConstValue0);
1988 N1IsConst = isConstantSplatVector(N1.getNode(), ConstValue1);
1990 N0IsConst = dyn_cast<ConstantSDNode>(N0) != nullptr;
1991 ConstValue0 = N0IsConst ? (dyn_cast<ConstantSDNode>(N0))->getAPIntValue()
1993 N1IsConst = dyn_cast<ConstantSDNode>(N1) != nullptr;
1994 ConstValue1 = N1IsConst ? (dyn_cast<ConstantSDNode>(N1))->getAPIntValue()
1998 // fold (mul c1, c2) -> c1*c2
1999 if (N0IsConst && N1IsConst)
2000 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0.getNode(), N1.getNode());
2002 // canonicalize constant to RHS (vector doesn't have to splat)
2003 if (isConstantIntBuildVectorOrConstantInt(N0) &&
2004 !isConstantIntBuildVectorOrConstantInt(N1))
2005 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0);
2006 // fold (mul x, 0) -> 0
2007 if (N1IsConst && ConstValue1 == 0)
2009 // We require a splat of the entire scalar bit width for non-contiguous
2012 ConstValue1.getBitWidth() == VT.getScalarType().getSizeInBits();
2013 // fold (mul x, 1) -> x
2014 if (N1IsConst && ConstValue1 == 1 && IsFullSplat)
2016 // fold (mul x, -1) -> 0-x
2017 if (N1IsConst && ConstValue1.isAllOnesValue())
2018 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2019 DAG.getConstant(0, VT), N0);
2020 // fold (mul x, (1 << c)) -> x << c
2021 if (N1IsConst && ConstValue1.isPowerOf2() && IsFullSplat)
2022 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
2023 DAG.getConstant(ConstValue1.logBase2(),
2024 getShiftAmountTy(N0.getValueType())));
2025 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
2026 if (N1IsConst && (-ConstValue1).isPowerOf2() && IsFullSplat) {
2027 unsigned Log2Val = (-ConstValue1).logBase2();
2028 // FIXME: If the input is something that is easily negated (e.g. a
2029 // single-use add), we should put the negate there.
2030 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2031 DAG.getConstant(0, VT),
2032 DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
2033 DAG.getConstant(Log2Val,
2034 getShiftAmountTy(N0.getValueType()))));
2038 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
2039 if (N1IsConst && N0.getOpcode() == ISD::SHL &&
2040 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2041 isa<ConstantSDNode>(N0.getOperand(1)))) {
2042 SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT,
2043 N1, N0.getOperand(1));
2044 AddToWorklist(C3.getNode());
2045 return DAG.getNode(ISD::MUL, SDLoc(N), VT,
2046 N0.getOperand(0), C3);
2049 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
2052 SDValue Sh(nullptr,0), Y(nullptr,0);
2053 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
2054 if (N0.getOpcode() == ISD::SHL &&
2055 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2056 isa<ConstantSDNode>(N0.getOperand(1))) &&
2057 N0.getNode()->hasOneUse()) {
2059 } else if (N1.getOpcode() == ISD::SHL &&
2060 isa<ConstantSDNode>(N1.getOperand(1)) &&
2061 N1.getNode()->hasOneUse()) {
2066 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2067 Sh.getOperand(0), Y);
2068 return DAG.getNode(ISD::SHL, SDLoc(N), VT,
2069 Mul, Sh.getOperand(1));
2073 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
2074 if (N1IsConst && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
2075 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2076 isa<ConstantSDNode>(N0.getOperand(1))))
2077 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
2078 DAG.getNode(ISD::MUL, SDLoc(N0), VT,
2079 N0.getOperand(0), N1),
2080 DAG.getNode(ISD::MUL, SDLoc(N1), VT,
2081 N0.getOperand(1), N1));
2084 if (SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1))
2090 SDValue DAGCombiner::visitSDIV(SDNode *N) {
2091 SDValue N0 = N->getOperand(0);
2092 SDValue N1 = N->getOperand(1);
2093 EVT VT = N->getValueType(0);
2097 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2100 // fold (sdiv c1, c2) -> c1/c2
2101 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2102 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2103 if (N0C && N1C && !N1C->isNullValue())
2104 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
2105 // fold (sdiv X, 1) -> X
2106 if (N1C && N1C->getAPIntValue() == 1LL)
2108 // fold (sdiv X, -1) -> 0-X
2109 if (N1C && N1C->isAllOnesValue())
2110 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2111 DAG.getConstant(0, VT), N0);
2112 // If we know the sign bits of both operands are zero, strength reduce to a
2113 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
2114 if (!VT.isVector()) {
2115 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2116 return DAG.getNode(ISD::UDIV, SDLoc(N), N1.getValueType(),
2120 // fold (sdiv X, pow2) -> simple ops after legalize
2121 if (N1C && !N1C->isNullValue() && (N1C->getAPIntValue().isPowerOf2() ||
2122 (-N1C->getAPIntValue()).isPowerOf2())) {
2123 // If dividing by powers of two is cheap, then don't perform the following
2125 if (TLI.isPow2SDivCheap())
2128 // Target-specific implementation of sdiv x, pow2.
2129 SDValue Res = BuildSDIVPow2(N);
2133 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
2135 // Splat the sign bit into the register
2137 DAG.getNode(ISD::SRA, SDLoc(N), VT, N0,
2138 DAG.getConstant(VT.getScalarSizeInBits() - 1,
2139 getShiftAmountTy(N0.getValueType())));
2140 AddToWorklist(SGN.getNode());
2142 // Add (N0 < 0) ? abs2 - 1 : 0;
2144 DAG.getNode(ISD::SRL, SDLoc(N), VT, SGN,
2145 DAG.getConstant(VT.getScalarSizeInBits() - lg2,
2146 getShiftAmountTy(SGN.getValueType())));
2147 SDValue ADD = DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, SRL);
2148 AddToWorklist(SRL.getNode());
2149 AddToWorklist(ADD.getNode()); // Divide by pow2
2150 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), VT, ADD,
2151 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType())));
2153 // If we're dividing by a positive value, we're done. Otherwise, we must
2154 // negate the result.
2155 if (N1C->getAPIntValue().isNonNegative())
2158 AddToWorklist(SRA.getNode());
2159 return DAG.getNode(ISD::SUB, SDLoc(N), VT, DAG.getConstant(0, VT), SRA);
2162 // If integer divide is expensive and we satisfy the requirements, emit an
2163 // alternate sequence.
2164 if (N1C && !TLI.isIntDivCheap()) {
2165 SDValue Op = BuildSDIV(N);
2166 if (Op.getNode()) return Op;
2170 if (N0.getOpcode() == ISD::UNDEF)
2171 return DAG.getConstant(0, VT);
2172 // X / undef -> undef
2173 if (N1.getOpcode() == ISD::UNDEF)
2179 SDValue DAGCombiner::visitUDIV(SDNode *N) {
2180 SDValue N0 = N->getOperand(0);
2181 SDValue N1 = N->getOperand(1);
2182 EVT VT = N->getValueType(0);
2186 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2189 // fold (udiv c1, c2) -> c1/c2
2190 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2191 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2192 if (N0C && N1C && !N1C->isNullValue())
2193 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
2194 // fold (udiv x, (1 << c)) -> x >>u c
2195 if (N1C && N1C->getAPIntValue().isPowerOf2())
2196 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0,
2197 DAG.getConstant(N1C->getAPIntValue().logBase2(),
2198 getShiftAmountTy(N0.getValueType())));
2199 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
2200 if (N1.getOpcode() == ISD::SHL) {
2201 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2202 if (SHC->getAPIntValue().isPowerOf2()) {
2203 EVT ADDVT = N1.getOperand(1).getValueType();
2204 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N), ADDVT,
2206 DAG.getConstant(SHC->getAPIntValue()
2209 AddToWorklist(Add.getNode());
2210 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, Add);
2214 // fold (udiv x, c) -> alternate
2215 if (N1C && !TLI.isIntDivCheap()) {
2216 SDValue Op = BuildUDIV(N);
2217 if (Op.getNode()) return Op;
2221 if (N0.getOpcode() == ISD::UNDEF)
2222 return DAG.getConstant(0, VT);
2223 // X / undef -> undef
2224 if (N1.getOpcode() == ISD::UNDEF)
2230 SDValue DAGCombiner::visitSREM(SDNode *N) {
2231 SDValue N0 = N->getOperand(0);
2232 SDValue N1 = N->getOperand(1);
2233 EVT VT = N->getValueType(0);
2235 // fold (srem c1, c2) -> c1%c2
2236 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2237 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2238 if (N0C && N1C && !N1C->isNullValue())
2239 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
2240 // If we know the sign bits of both operands are zero, strength reduce to a
2241 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
2242 if (!VT.isVector()) {
2243 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2244 return DAG.getNode(ISD::UREM, SDLoc(N), VT, N0, N1);
2247 // If X/C can be simplified by the division-by-constant logic, lower
2248 // X%C to the equivalent of X-X/C*C.
2249 if (N1C && !N1C->isNullValue()) {
2250 SDValue Div = DAG.getNode(ISD::SDIV, SDLoc(N), VT, N0, N1);
2251 AddToWorklist(Div.getNode());
2252 SDValue OptimizedDiv = combine(Div.getNode());
2253 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2254 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2256 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2257 AddToWorklist(Mul.getNode());
2263 if (N0.getOpcode() == ISD::UNDEF)
2264 return DAG.getConstant(0, VT);
2265 // X % undef -> undef
2266 if (N1.getOpcode() == ISD::UNDEF)
2272 SDValue DAGCombiner::visitUREM(SDNode *N) {
2273 SDValue N0 = N->getOperand(0);
2274 SDValue N1 = N->getOperand(1);
2275 EVT VT = N->getValueType(0);
2277 // fold (urem c1, c2) -> c1%c2
2278 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2279 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2280 if (N0C && N1C && !N1C->isNullValue())
2281 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
2282 // fold (urem x, pow2) -> (and x, pow2-1)
2283 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
2284 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0,
2285 DAG.getConstant(N1C->getAPIntValue()-1,VT));
2286 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2287 if (N1.getOpcode() == ISD::SHL) {
2288 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2289 if (SHC->getAPIntValue().isPowerOf2()) {
2291 DAG.getNode(ISD::ADD, SDLoc(N), VT, N1,
2292 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
2294 AddToWorklist(Add.getNode());
2295 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, Add);
2300 // If X/C can be simplified by the division-by-constant logic, lower
2301 // X%C to the equivalent of X-X/C*C.
2302 if (N1C && !N1C->isNullValue()) {
2303 SDValue Div = DAG.getNode(ISD::UDIV, SDLoc(N), VT, N0, N1);
2304 AddToWorklist(Div.getNode());
2305 SDValue OptimizedDiv = combine(Div.getNode());
2306 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2307 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2309 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2310 AddToWorklist(Mul.getNode());
2316 if (N0.getOpcode() == ISD::UNDEF)
2317 return DAG.getConstant(0, VT);
2318 // X % undef -> undef
2319 if (N1.getOpcode() == ISD::UNDEF)
2325 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2326 SDValue N0 = N->getOperand(0);
2327 SDValue N1 = N->getOperand(1);
2328 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2329 EVT VT = N->getValueType(0);
2332 // fold (mulhs x, 0) -> 0
2333 if (N1C && N1C->isNullValue())
2335 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2336 if (N1C && N1C->getAPIntValue() == 1)
2337 return DAG.getNode(ISD::SRA, SDLoc(N), N0.getValueType(), N0,
2338 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2339 getShiftAmountTy(N0.getValueType())));
2340 // fold (mulhs x, undef) -> 0
2341 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2342 return DAG.getConstant(0, VT);
2344 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2346 if (VT.isSimple() && !VT.isVector()) {
2347 MVT Simple = VT.getSimpleVT();
2348 unsigned SimpleSize = Simple.getSizeInBits();
2349 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2350 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2351 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2352 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2353 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2354 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2355 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2356 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2363 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2364 SDValue N0 = N->getOperand(0);
2365 SDValue N1 = N->getOperand(1);
2366 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2367 EVT VT = N->getValueType(0);
2370 // fold (mulhu x, 0) -> 0
2371 if (N1C && N1C->isNullValue())
2373 // fold (mulhu x, 1) -> 0
2374 if (N1C && N1C->getAPIntValue() == 1)
2375 return DAG.getConstant(0, N0.getValueType());
2376 // fold (mulhu x, undef) -> 0
2377 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2378 return DAG.getConstant(0, VT);
2380 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2382 if (VT.isSimple() && !VT.isVector()) {
2383 MVT Simple = VT.getSimpleVT();
2384 unsigned SimpleSize = Simple.getSizeInBits();
2385 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2386 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2387 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2388 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2389 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2390 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2391 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2392 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2399 /// Perform optimizations common to nodes that compute two values. LoOp and HiOp
2400 /// give the opcodes for the two computations that are being performed. Return
2401 /// true if a simplification was made.
2402 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2404 // If the high half is not needed, just compute the low half.
2405 bool HiExists = N->hasAnyUseOfValue(1);
2407 (!LegalOperations ||
2408 TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) {
2409 SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2410 return CombineTo(N, Res, Res);
2413 // If the low half is not needed, just compute the high half.
2414 bool LoExists = N->hasAnyUseOfValue(0);
2416 (!LegalOperations ||
2417 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2418 SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2419 return CombineTo(N, Res, Res);
2422 // If both halves are used, return as it is.
2423 if (LoExists && HiExists)
2426 // If the two computed results can be simplified separately, separate them.
2428 SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2429 AddToWorklist(Lo.getNode());
2430 SDValue LoOpt = combine(Lo.getNode());
2431 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2432 (!LegalOperations ||
2433 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2434 return CombineTo(N, LoOpt, LoOpt);
2438 SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2439 AddToWorklist(Hi.getNode());
2440 SDValue HiOpt = combine(Hi.getNode());
2441 if (HiOpt.getNode() && HiOpt != Hi &&
2442 (!LegalOperations ||
2443 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2444 return CombineTo(N, HiOpt, HiOpt);
2450 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2451 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2452 if (Res.getNode()) return Res;
2454 EVT VT = N->getValueType(0);
2457 // If the type is twice as wide is legal, transform the mulhu to a wider
2458 // multiply plus a shift.
2459 if (VT.isSimple() && !VT.isVector()) {
2460 MVT Simple = VT.getSimpleVT();
2461 unsigned SimpleSize = Simple.getSizeInBits();
2462 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2463 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2464 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2465 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2466 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2467 // Compute the high part as N1.
2468 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2469 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2470 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2471 // Compute the low part as N0.
2472 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2473 return CombineTo(N, Lo, Hi);
2480 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2481 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2482 if (Res.getNode()) return Res;
2484 EVT VT = N->getValueType(0);
2487 // If the type is twice as wide is legal, transform the mulhu to a wider
2488 // multiply plus a shift.
2489 if (VT.isSimple() && !VT.isVector()) {
2490 MVT Simple = VT.getSimpleVT();
2491 unsigned SimpleSize = Simple.getSizeInBits();
2492 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2493 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2494 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2495 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2496 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2497 // Compute the high part as N1.
2498 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2499 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2500 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2501 // Compute the low part as N0.
2502 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2503 return CombineTo(N, Lo, Hi);
2510 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2511 // (smulo x, 2) -> (saddo x, x)
2512 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2513 if (C2->getAPIntValue() == 2)
2514 return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(),
2515 N->getOperand(0), N->getOperand(0));
2520 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2521 // (umulo x, 2) -> (uaddo x, x)
2522 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2523 if (C2->getAPIntValue() == 2)
2524 return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(),
2525 N->getOperand(0), N->getOperand(0));
2530 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2531 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2532 if (Res.getNode()) return Res;
2537 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2538 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2539 if (Res.getNode()) return Res;
2544 /// If this is a binary operator with two operands of the same opcode, try to
2546 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2547 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2548 EVT VT = N0.getValueType();
2549 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2551 // Bail early if none of these transforms apply.
2552 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2554 // For each of OP in AND/OR/XOR:
2555 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2556 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2557 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2558 // fold (OP (bswap x), (bswap y)) -> (bswap (OP x, y))
2559 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2561 // do not sink logical op inside of a vector extend, since it may combine
2563 EVT Op0VT = N0.getOperand(0).getValueType();
2564 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2565 N0.getOpcode() == ISD::SIGN_EXTEND ||
2566 N0.getOpcode() == ISD::BSWAP ||
2567 // Avoid infinite looping with PromoteIntBinOp.
2568 (N0.getOpcode() == ISD::ANY_EXTEND &&
2569 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2570 (N0.getOpcode() == ISD::TRUNCATE &&
2571 (!TLI.isZExtFree(VT, Op0VT) ||
2572 !TLI.isTruncateFree(Op0VT, VT)) &&
2573 TLI.isTypeLegal(Op0VT))) &&
2575 Op0VT == N1.getOperand(0).getValueType() &&
2576 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2577 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2578 N0.getOperand(0).getValueType(),
2579 N0.getOperand(0), N1.getOperand(0));
2580 AddToWorklist(ORNode.getNode());
2581 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode);
2584 // For each of OP in SHL/SRL/SRA/AND...
2585 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2586 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2587 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2588 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2589 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2590 N0.getOperand(1) == N1.getOperand(1)) {
2591 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2592 N0.getOperand(0).getValueType(),
2593 N0.getOperand(0), N1.getOperand(0));
2594 AddToWorklist(ORNode.getNode());
2595 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
2596 ORNode, N0.getOperand(1));
2599 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2600 // Only perform this optimization after type legalization and before
2601 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2602 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2603 // we don't want to undo this promotion.
2604 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2606 if ((N0.getOpcode() == ISD::BITCAST ||
2607 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) &&
2608 Level == AfterLegalizeTypes) {
2609 SDValue In0 = N0.getOperand(0);
2610 SDValue In1 = N1.getOperand(0);
2611 EVT In0Ty = In0.getValueType();
2612 EVT In1Ty = In1.getValueType();
2614 // If both incoming values are integers, and the original types are the
2616 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2617 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1);
2618 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op);
2619 AddToWorklist(Op.getNode());
2624 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2625 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2626 // If both shuffles use the same mask, and both shuffle within a single
2627 // vector, then it is worthwhile to move the swizzle after the operation.
2628 // The type-legalizer generates this pattern when loading illegal
2629 // vector types from memory. In many cases this allows additional shuffle
2631 // There are other cases where moving the shuffle after the xor/and/or
2632 // is profitable even if shuffles don't perform a swizzle.
2633 // If both shuffles use the same mask, and both shuffles have the same first
2634 // or second operand, then it might still be profitable to move the shuffle
2635 // after the xor/and/or operation.
2636 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) {
2637 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2638 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2640 assert(N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() &&
2641 "Inputs to shuffles are not the same type");
2643 // Check that both shuffles use the same mask. The masks are known to be of
2644 // the same length because the result vector type is the same.
2645 // Check also that shuffles have only one use to avoid introducing extra
2647 if (SVN0->hasOneUse() && SVN1->hasOneUse() &&
2648 SVN0->getMask().equals(SVN1->getMask())) {
2649 SDValue ShOp = N0->getOperand(1);
2651 // Don't try to fold this node if it requires introducing a
2652 // build vector of all zeros that might be illegal at this stage.
2653 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2655 ShOp = DAG.getConstant(0, VT);
2660 // (AND (shuf (A, C), shuf (B, C)) -> shuf (AND (A, B), C)
2661 // (OR (shuf (A, C), shuf (B, C)) -> shuf (OR (A, B), C)
2662 // (XOR (shuf (A, C), shuf (B, C)) -> shuf (XOR (A, B), V_0)
2663 if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) {
2664 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2665 N0->getOperand(0), N1->getOperand(0));
2666 AddToWorklist(NewNode.getNode());
2667 return DAG.getVectorShuffle(VT, SDLoc(N), NewNode, ShOp,
2668 &SVN0->getMask()[0]);
2671 // Don't try to fold this node if it requires introducing a
2672 // build vector of all zeros that might be illegal at this stage.
2673 ShOp = N0->getOperand(0);
2674 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2676 ShOp = DAG.getConstant(0, VT);
2681 // (AND (shuf (C, A), shuf (C, B)) -> shuf (C, AND (A, B))
2682 // (OR (shuf (C, A), shuf (C, B)) -> shuf (C, OR (A, B))
2683 // (XOR (shuf (C, A), shuf (C, B)) -> shuf (V_0, XOR (A, B))
2684 if (N0->getOperand(0) == N1->getOperand(0) && ShOp.getNode()) {
2685 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2686 N0->getOperand(1), N1->getOperand(1));
2687 AddToWorklist(NewNode.getNode());
2688 return DAG.getVectorShuffle(VT, SDLoc(N), ShOp, NewNode,
2689 &SVN0->getMask()[0]);
2697 /// This contains all DAGCombine rules which reduce two values combined by
2698 /// an And operation to a single value. This makes them reusable in the context
2699 /// of visitSELECT(). Rules involving constants are not included as
2700 /// visitSELECT() already handles those cases.
2701 SDValue DAGCombiner::visitANDLike(SDValue N0, SDValue N1,
2702 SDNode *LocReference) {
2703 EVT VT = N1.getValueType();
2705 // fold (and x, undef) -> 0
2706 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2707 return DAG.getConstant(0, VT);
2708 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2709 SDValue LL, LR, RL, RR, CC0, CC1;
2710 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2711 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2712 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2714 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2715 LL.getValueType().isInteger()) {
2716 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2717 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
2718 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2719 LR.getValueType(), LL, RL);
2720 AddToWorklist(ORNode.getNode());
2721 return DAG.getSetCC(SDLoc(LocReference), VT, ORNode, LR, Op1);
2723 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2724 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2725 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(N0),
2726 LR.getValueType(), LL, RL);
2727 AddToWorklist(ANDNode.getNode());
2728 return DAG.getSetCC(SDLoc(LocReference), VT, ANDNode, LR, Op1);
2730 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2731 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2732 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2733 LR.getValueType(), LL, RL);
2734 AddToWorklist(ORNode.getNode());
2735 return DAG.getSetCC(SDLoc(LocReference), VT, ORNode, LR, Op1);
2738 // Simplify (and (setne X, 0), (setne X, -1)) -> (setuge (add X, 1), 2)
2739 if (LL == RL && isa<ConstantSDNode>(LR) && isa<ConstantSDNode>(RR) &&
2740 Op0 == Op1 && LL.getValueType().isInteger() &&
2741 Op0 == ISD::SETNE && ((cast<ConstantSDNode>(LR)->isNullValue() &&
2742 cast<ConstantSDNode>(RR)->isAllOnesValue()) ||
2743 (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
2744 cast<ConstantSDNode>(RR)->isNullValue()))) {
2745 SDValue ADDNode = DAG.getNode(ISD::ADD, SDLoc(N0), LL.getValueType(),
2746 LL, DAG.getConstant(1, LL.getValueType()));
2747 AddToWorklist(ADDNode.getNode());
2748 return DAG.getSetCC(SDLoc(LocReference), VT, ADDNode,
2749 DAG.getConstant(2, LL.getValueType()), ISD::SETUGE);
2751 // canonicalize equivalent to ll == rl
2752 if (LL == RR && LR == RL) {
2753 Op1 = ISD::getSetCCSwappedOperands(Op1);
2756 if (LL == RL && LR == RR) {
2757 bool isInteger = LL.getValueType().isInteger();
2758 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2759 if (Result != ISD::SETCC_INVALID &&
2760 (!LegalOperations ||
2761 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
2762 TLI.isOperationLegal(ISD::SETCC,
2763 getSetCCResultType(N0.getSimpleValueType())))))
2764 return DAG.getSetCC(SDLoc(LocReference), N0.getValueType(),
2769 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
2770 VT.getSizeInBits() <= 64) {
2771 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2772 APInt ADDC = ADDI->getAPIntValue();
2773 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2774 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
2775 // immediate for an add, but it is legal if its top c2 bits are set,
2776 // transform the ADD so the immediate doesn't need to be materialized
2778 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
2779 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
2780 SRLI->getZExtValue());
2781 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
2783 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2785 DAG.getNode(ISD::ADD, SDLoc(N0), VT,
2786 N0.getOperand(0), DAG.getConstant(ADDC, VT));
2787 CombineTo(N0.getNode(), NewAdd);
2788 // Return N so it doesn't get rechecked!
2789 return SDValue(LocReference, 0);
2800 SDValue DAGCombiner::visitAND(SDNode *N) {
2801 SDValue N0 = N->getOperand(0);
2802 SDValue N1 = N->getOperand(1);
2803 EVT VT = N1.getValueType();
2806 if (VT.isVector()) {
2807 if (SDValue FoldedVOp = SimplifyVBinOp(N))
2810 // fold (and x, 0) -> 0, vector edition
2811 if (ISD::isBuildVectorAllZeros(N0.getNode()))
2812 // do not return N0, because undef node may exist in N0
2813 return DAG.getConstant(
2814 APInt::getNullValue(
2815 N0.getValueType().getScalarType().getSizeInBits()),
2817 if (ISD::isBuildVectorAllZeros(N1.getNode()))
2818 // do not return N1, because undef node may exist in N1
2819 return DAG.getConstant(
2820 APInt::getNullValue(
2821 N1.getValueType().getScalarType().getSizeInBits()),
2824 // fold (and x, -1) -> x, vector edition
2825 if (ISD::isBuildVectorAllOnes(N0.getNode()))
2827 if (ISD::isBuildVectorAllOnes(N1.getNode()))
2831 // fold (and c1, c2) -> c1&c2
2832 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2833 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2835 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2836 // canonicalize constant to RHS
2837 if (isConstantIntBuildVectorOrConstantInt(N0) &&
2838 !isConstantIntBuildVectorOrConstantInt(N1))
2839 return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0);
2840 // fold (and x, -1) -> x
2841 if (N1C && N1C->isAllOnesValue())
2843 // if (and x, c) is known to be zero, return 0
2844 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2845 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2846 APInt::getAllOnesValue(BitWidth)))
2847 return DAG.getConstant(0, VT);
2849 if (SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1))
2851 // fold (and (or x, C), D) -> D if (C & D) == D
2852 if (N1C && N0.getOpcode() == ISD::OR)
2853 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2854 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2856 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2857 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2858 SDValue N0Op0 = N0.getOperand(0);
2859 APInt Mask = ~N1C->getAPIntValue();
2860 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2861 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2862 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
2863 N0.getValueType(), N0Op0);
2865 // Replace uses of the AND with uses of the Zero extend node.
2868 // We actually want to replace all uses of the any_extend with the
2869 // zero_extend, to avoid duplicating things. This will later cause this
2870 // AND to be folded.
2871 CombineTo(N0.getNode(), Zext);
2872 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2875 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2876 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2877 // already be zero by virtue of the width of the base type of the load.
2879 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2881 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2882 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2883 N0.getOpcode() == ISD::LOAD) {
2884 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2885 N0 : N0.getOperand(0) );
2887 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2888 // This can be a pure constant or a vector splat, in which case we treat the
2889 // vector as a scalar and use the splat value.
2890 APInt Constant = APInt::getNullValue(1);
2891 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2892 Constant = C->getAPIntValue();
2893 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2894 APInt SplatValue, SplatUndef;
2895 unsigned SplatBitSize;
2897 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2898 SplatBitSize, HasAnyUndefs);
2900 // Undef bits can contribute to a possible optimisation if set, so
2902 SplatValue |= SplatUndef;
2904 // The splat value may be something like "0x00FFFFFF", which means 0 for
2905 // the first vector value and FF for the rest, repeating. We need a mask
2906 // that will apply equally to all members of the vector, so AND all the
2907 // lanes of the constant together.
2908 EVT VT = Vector->getValueType(0);
2909 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2911 // If the splat value has been compressed to a bitlength lower
2912 // than the size of the vector lane, we need to re-expand it to
2914 if (BitWidth > SplatBitSize)
2915 for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
2916 SplatBitSize < BitWidth;
2917 SplatBitSize = SplatBitSize * 2)
2918 SplatValue |= SplatValue.shl(SplatBitSize);
2920 // Make sure that variable 'Constant' is only set if 'SplatBitSize' is a
2921 // multiple of 'BitWidth'. Otherwise, we could propagate a wrong value.
2922 if (SplatBitSize % BitWidth == 0) {
2923 Constant = APInt::getAllOnesValue(BitWidth);
2924 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
2925 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
2930 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
2931 // actually legal and isn't going to get expanded, else this is a false
2933 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
2934 Load->getValueType(0),
2935 Load->getMemoryVT());
2937 // Resize the constant to the same size as the original memory access before
2938 // extension. If it is still the AllOnesValue then this AND is completely
2941 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
2944 switch (Load->getExtensionType()) {
2945 default: B = false; break;
2946 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
2948 case ISD::NON_EXTLOAD: B = true; break;
2951 if (B && Constant.isAllOnesValue()) {
2952 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
2953 // preserve semantics once we get rid of the AND.
2954 SDValue NewLoad(Load, 0);
2955 if (Load->getExtensionType() == ISD::EXTLOAD) {
2956 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
2957 Load->getValueType(0), SDLoc(Load),
2958 Load->getChain(), Load->getBasePtr(),
2959 Load->getOffset(), Load->getMemoryVT(),
2960 Load->getMemOperand());
2961 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
2962 if (Load->getNumValues() == 3) {
2963 // PRE/POST_INC loads have 3 values.
2964 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
2965 NewLoad.getValue(2) };
2966 CombineTo(Load, To, 3, true);
2968 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
2972 // Fold the AND away, taking care not to fold to the old load node if we
2974 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
2976 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2980 // fold (and (load x), 255) -> (zextload x, i8)
2981 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2982 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2983 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2984 (N0.getOpcode() == ISD::ANY_EXTEND &&
2985 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2986 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2987 LoadSDNode *LN0 = HasAnyExt
2988 ? cast<LoadSDNode>(N0.getOperand(0))
2989 : cast<LoadSDNode>(N0);
2990 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2991 LN0->isUnindexed() && N0.hasOneUse() && SDValue(LN0, 0).hasOneUse()) {
2992 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2993 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2994 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2995 EVT LoadedVT = LN0->getMemoryVT();
2996 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2998 if (ExtVT == LoadedVT &&
2999 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy,
3003 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
3004 LN0->getChain(), LN0->getBasePtr(), ExtVT,
3005 LN0->getMemOperand());
3007 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
3008 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3011 // Do not change the width of a volatile load.
3012 // Do not generate loads of non-round integer types since these can
3013 // be expensive (and would be wrong if the type is not byte sized).
3014 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
3015 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy,
3017 EVT PtrType = LN0->getOperand(1).getValueType();
3019 unsigned Alignment = LN0->getAlignment();
3020 SDValue NewPtr = LN0->getBasePtr();
3022 // For big endian targets, we need to add an offset to the pointer
3023 // to load the correct bytes. For little endian systems, we merely
3024 // need to read fewer bytes from the same pointer.
3025 if (TLI.isBigEndian()) {
3026 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
3027 unsigned EVTStoreBytes = ExtVT.getStoreSize();
3028 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
3029 NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0), PtrType,
3030 NewPtr, DAG.getConstant(PtrOff, PtrType));
3031 Alignment = MinAlign(Alignment, PtrOff);
3034 AddToWorklist(NewPtr.getNode());
3037 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
3038 LN0->getChain(), NewPtr,
3039 LN0->getPointerInfo(),
3040 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
3041 LN0->isInvariant(), Alignment, LN0->getAAInfo());
3043 CombineTo(LN0, Load, Load.getValue(1));
3044 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3050 if (SDValue Combined = visitANDLike(N0, N1, N))
3053 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
3054 if (N0.getOpcode() == N1.getOpcode()) {
3055 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3056 if (Tmp.getNode()) return Tmp;
3059 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
3060 // fold (and (sra)) -> (and (srl)) when possible.
3061 if (!VT.isVector() &&
3062 SimplifyDemandedBits(SDValue(N, 0)))
3063 return SDValue(N, 0);
3065 // fold (zext_inreg (extload x)) -> (zextload x)
3066 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
3067 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3068 EVT MemVT = LN0->getMemoryVT();
3069 // If we zero all the possible extended bits, then we can turn this into
3070 // a zextload if we are running before legalize or the operation is legal.
3071 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
3072 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
3073 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
3074 ((!LegalOperations && !LN0->isVolatile()) ||
3075 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) {
3076 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
3077 LN0->getChain(), LN0->getBasePtr(),
3078 MemVT, LN0->getMemOperand());
3080 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
3081 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3084 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
3085 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
3087 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
3088 EVT MemVT = LN0->getMemoryVT();
3089 // If we zero all the possible extended bits, then we can turn this into
3090 // a zextload if we are running before legalize or the operation is legal.
3091 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
3092 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
3093 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
3094 ((!LegalOperations && !LN0->isVolatile()) ||
3095 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) {
3096 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
3097 LN0->getChain(), LN0->getBasePtr(),
3098 MemVT, LN0->getMemOperand());
3100 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
3101 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3104 // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const)
3105 if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) {
3106 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
3107 N0.getOperand(1), false);
3108 if (BSwap.getNode())
3115 /// Match (a >> 8) | (a << 8) as (bswap a) >> 16.
3116 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
3117 bool DemandHighBits) {
3118 if (!LegalOperations)
3121 EVT VT = N->getValueType(0);
3122 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
3124 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3127 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
3128 bool LookPassAnd0 = false;
3129 bool LookPassAnd1 = false;
3130 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
3132 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
3134 if (N0.getOpcode() == ISD::AND) {
3135 if (!N0.getNode()->hasOneUse())
3137 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3138 if (!N01C || N01C->getZExtValue() != 0xFF00)
3140 N0 = N0.getOperand(0);
3141 LookPassAnd0 = true;
3144 if (N1.getOpcode() == ISD::AND) {
3145 if (!N1.getNode()->hasOneUse())
3147 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3148 if (!N11C || N11C->getZExtValue() != 0xFF)
3150 N1 = N1.getOperand(0);
3151 LookPassAnd1 = true;
3154 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
3156 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
3158 if (!N0.getNode()->hasOneUse() ||
3159 !N1.getNode()->hasOneUse())
3162 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3163 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3166 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
3169 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
3170 SDValue N00 = N0->getOperand(0);
3171 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
3172 if (!N00.getNode()->hasOneUse())
3174 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
3175 if (!N001C || N001C->getZExtValue() != 0xFF)
3177 N00 = N00.getOperand(0);
3178 LookPassAnd0 = true;
3181 SDValue N10 = N1->getOperand(0);
3182 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
3183 if (!N10.getNode()->hasOneUse())
3185 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
3186 if (!N101C || N101C->getZExtValue() != 0xFF00)
3188 N10 = N10.getOperand(0);
3189 LookPassAnd1 = true;
3195 // Make sure everything beyond the low halfword gets set to zero since the SRL
3196 // 16 will clear the top bits.
3197 unsigned OpSizeInBits = VT.getSizeInBits();
3198 if (DemandHighBits && OpSizeInBits > 16) {
3199 // If the left-shift isn't masked out then the only way this is a bswap is
3200 // if all bits beyond the low 8 are 0. In that case the entire pattern
3201 // reduces to a left shift anyway: leave it for other parts of the combiner.
3205 // However, if the right shift isn't masked out then it might be because
3206 // it's not needed. See if we can spot that too.
3207 if (!LookPassAnd1 &&
3208 !DAG.MaskedValueIsZero(
3209 N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16)))
3213 SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00);
3214 if (OpSizeInBits > 16)
3215 Res = DAG.getNode(ISD::SRL, SDLoc(N), VT, Res,
3216 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
3220 /// Return true if the specified node is an element that makes up a 32-bit
3221 /// packed halfword byteswap.
3222 /// ((x & 0x000000ff) << 8) |
3223 /// ((x & 0x0000ff00) >> 8) |
3224 /// ((x & 0x00ff0000) << 8) |
3225 /// ((x & 0xff000000) >> 8)
3226 static bool isBSwapHWordElement(SDValue N, MutableArrayRef<SDNode *> Parts) {
3227 if (!N.getNode()->hasOneUse())
3230 unsigned Opc = N.getOpcode();
3231 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
3234 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3239 switch (N1C->getZExtValue()) {
3242 case 0xFF: Num = 0; break;
3243 case 0xFF00: Num = 1; break;
3244 case 0xFF0000: Num = 2; break;
3245 case 0xFF000000: Num = 3; break;
3248 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
3249 SDValue N0 = N.getOperand(0);
3250 if (Opc == ISD::AND) {
3251 if (Num == 0 || Num == 2) {
3253 // (x >> 8) & 0xff0000
3254 if (N0.getOpcode() != ISD::SRL)
3256 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3257 if (!C || C->getZExtValue() != 8)
3260 // (x << 8) & 0xff00
3261 // (x << 8) & 0xff000000
3262 if (N0.getOpcode() != ISD::SHL)
3264 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3265 if (!C || C->getZExtValue() != 8)
3268 } else if (Opc == ISD::SHL) {
3270 // (x & 0xff0000) << 8
3271 if (Num != 0 && Num != 2)
3273 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3274 if (!C || C->getZExtValue() != 8)
3276 } else { // Opc == ISD::SRL
3277 // (x & 0xff00) >> 8
3278 // (x & 0xff000000) >> 8
3279 if (Num != 1 && Num != 3)
3281 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3282 if (!C || C->getZExtValue() != 8)
3289 Parts[Num] = N0.getOperand(0).getNode();
3293 /// Match a 32-bit packed halfword bswap. That is
3294 /// ((x & 0x000000ff) << 8) |
3295 /// ((x & 0x0000ff00) >> 8) |
3296 /// ((x & 0x00ff0000) << 8) |
3297 /// ((x & 0xff000000) >> 8)
3298 /// => (rotl (bswap x), 16)
3299 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
3300 if (!LegalOperations)
3303 EVT VT = N->getValueType(0);
3306 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3310 // (or (or (and), (and)), (or (and), (and)))
3311 // (or (or (or (and), (and)), (and)), (and))
3312 if (N0.getOpcode() != ISD::OR)
3314 SDValue N00 = N0.getOperand(0);
3315 SDValue N01 = N0.getOperand(1);
3316 SDNode *Parts[4] = {};
3318 if (N1.getOpcode() == ISD::OR &&
3319 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) {
3320 // (or (or (and), (and)), (or (and), (and)))
3321 SDValue N000 = N00.getOperand(0);
3322 if (!isBSwapHWordElement(N000, Parts))
3325 SDValue N001 = N00.getOperand(1);
3326 if (!isBSwapHWordElement(N001, Parts))
3328 SDValue N010 = N01.getOperand(0);
3329 if (!isBSwapHWordElement(N010, Parts))
3331 SDValue N011 = N01.getOperand(1);
3332 if (!isBSwapHWordElement(N011, Parts))
3335 // (or (or (or (and), (and)), (and)), (and))
3336 if (!isBSwapHWordElement(N1, Parts))
3338 if (!isBSwapHWordElement(N01, Parts))
3340 if (N00.getOpcode() != ISD::OR)
3342 SDValue N000 = N00.getOperand(0);
3343 if (!isBSwapHWordElement(N000, Parts))
3345 SDValue N001 = N00.getOperand(1);
3346 if (!isBSwapHWordElement(N001, Parts))
3350 // Make sure the parts are all coming from the same node.
3351 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
3354 SDValue BSwap = DAG.getNode(ISD::BSWAP, SDLoc(N), VT,
3355 SDValue(Parts[0],0));
3357 // Result of the bswap should be rotated by 16. If it's not legal, then
3358 // do (x << 16) | (x >> 16).
3359 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
3360 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3361 return DAG.getNode(ISD::ROTL, SDLoc(N), VT, BSwap, ShAmt);
3362 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3363 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, BSwap, ShAmt);
3364 return DAG.getNode(ISD::OR, SDLoc(N), VT,
3365 DAG.getNode(ISD::SHL, SDLoc(N), VT, BSwap, ShAmt),
3366 DAG.getNode(ISD::SRL, SDLoc(N), VT, BSwap, ShAmt));
3369 /// This contains all DAGCombine rules which reduce two values combined by
3370 /// an Or operation to a single value \see visitANDLike().
3371 SDValue DAGCombiner::visitORLike(SDValue N0, SDValue N1, SDNode *LocReference) {
3372 EVT VT = N1.getValueType();
3373 // fold (or x, undef) -> -1
3374 if (!LegalOperations &&
3375 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3376 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3377 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
3379 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3380 SDValue LL, LR, RL, RR, CC0, CC1;
3381 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3382 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3383 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3385 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3386 LL.getValueType().isInteger()) {
3387 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3388 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3389 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3390 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3391 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(LR),
3392 LR.getValueType(), LL, RL);
3393 AddToWorklist(ORNode.getNode());
3394 return DAG.getSetCC(SDLoc(LocReference), VT, ORNode, LR, Op1);
3396 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3397 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3398 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3399 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3400 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(LR),
3401 LR.getValueType(), LL, RL);
3402 AddToWorklist(ANDNode.getNode());
3403 return DAG.getSetCC(SDLoc(LocReference), VT, ANDNode, LR, Op1);
3406 // canonicalize equivalent to ll == rl
3407 if (LL == RR && LR == RL) {
3408 Op1 = ISD::getSetCCSwappedOperands(Op1);
3411 if (LL == RL && LR == RR) {
3412 bool isInteger = LL.getValueType().isInteger();
3413 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3414 if (Result != ISD::SETCC_INVALID &&
3415 (!LegalOperations ||
3416 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
3417 TLI.isOperationLegal(ISD::SETCC,
3418 getSetCCResultType(N0.getValueType())))))
3419 return DAG.getSetCC(SDLoc(LocReference), N0.getValueType(),
3424 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3425 if (N0.getOpcode() == ISD::AND &&
3426 N1.getOpcode() == ISD::AND &&
3427 N0.getOperand(1).getOpcode() == ISD::Constant &&
3428 N1.getOperand(1).getOpcode() == ISD::Constant &&
3429 // Don't increase # computations.
3430 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3431 // We can only do this xform if we know that bits from X that are set in C2
3432 // but not in C1 are already zero. Likewise for Y.
3433 const APInt &LHSMask =
3434 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3435 const APInt &RHSMask =
3436 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3438 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3439 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3440 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3441 N0.getOperand(0), N1.getOperand(0));
3442 return DAG.getNode(ISD::AND, SDLoc(LocReference), VT, X,
3443 DAG.getConstant(LHSMask | RHSMask, VT));
3447 // (or (and X, M), (and X, N)) -> (and X, (or M, N))
3448 if (N0.getOpcode() == ISD::AND &&
3449 N1.getOpcode() == ISD::AND &&
3450 N0.getOperand(0) == N1.getOperand(0) &&
3451 // Don't increase # computations.
3452 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3453 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3454 N0.getOperand(1), N1.getOperand(1));
3455 return DAG.getNode(ISD::AND, SDLoc(LocReference), VT, N0.getOperand(0), X);
3461 SDValue DAGCombiner::visitOR(SDNode *N) {
3462 SDValue N0 = N->getOperand(0);
3463 SDValue N1 = N->getOperand(1);
3464 EVT VT = N1.getValueType();
3467 if (VT.isVector()) {
3468 if (SDValue FoldedVOp = SimplifyVBinOp(N))
3471 // fold (or x, 0) -> x, vector edition
3472 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3474 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3477 // fold (or x, -1) -> -1, vector edition
3478 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3479 // do not return N0, because undef node may exist in N0
3480 return DAG.getConstant(
3481 APInt::getAllOnesValue(
3482 N0.getValueType().getScalarType().getSizeInBits()),
3484 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3485 // do not return N1, because undef node may exist in N1
3486 return DAG.getConstant(
3487 APInt::getAllOnesValue(
3488 N1.getValueType().getScalarType().getSizeInBits()),
3491 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask1)
3492 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf B, A, Mask2)
3493 // Do this only if the resulting shuffle is legal.
3494 if (isa<ShuffleVectorSDNode>(N0) &&
3495 isa<ShuffleVectorSDNode>(N1) &&
3496 // Avoid folding a node with illegal type.
3497 TLI.isTypeLegal(VT) &&
3498 N0->getOperand(1) == N1->getOperand(1) &&
3499 ISD::isBuildVectorAllZeros(N0.getOperand(1).getNode())) {
3500 bool CanFold = true;
3501 unsigned NumElts = VT.getVectorNumElements();
3502 const ShuffleVectorSDNode *SV0 = cast<ShuffleVectorSDNode>(N0);
3503 const ShuffleVectorSDNode *SV1 = cast<ShuffleVectorSDNode>(N1);
3504 // We construct two shuffle masks:
3505 // - Mask1 is a shuffle mask for a shuffle with N0 as the first operand
3506 // and N1 as the second operand.
3507 // - Mask2 is a shuffle mask for a shuffle with N1 as the first operand
3508 // and N0 as the second operand.
3509 // We do this because OR is commutable and therefore there might be
3510 // two ways to fold this node into a shuffle.
3511 SmallVector<int,4> Mask1;
3512 SmallVector<int,4> Mask2;
3514 for (unsigned i = 0; i != NumElts && CanFold; ++i) {
3515 int M0 = SV0->getMaskElt(i);
3516 int M1 = SV1->getMaskElt(i);
3518 // Both shuffle indexes are undef. Propagate Undef.
3519 if (M0 < 0 && M1 < 0) {
3520 Mask1.push_back(M0);
3521 Mask2.push_back(M0);
3525 if (M0 < 0 || M1 < 0 ||
3526 (M0 < (int)NumElts && M1 < (int)NumElts) ||
3527 (M0 >= (int)NumElts && M1 >= (int)NumElts)) {
3532 Mask1.push_back(M0 < (int)NumElts ? M0 : M1 + NumElts);
3533 Mask2.push_back(M1 < (int)NumElts ? M1 : M0 + NumElts);
3537 // Fold this sequence only if the resulting shuffle is 'legal'.
3538 if (TLI.isShuffleMaskLegal(Mask1, VT))
3539 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0),
3540 N1->getOperand(0), &Mask1[0]);
3541 if (TLI.isShuffleMaskLegal(Mask2, VT))
3542 return DAG.getVectorShuffle(VT, SDLoc(N), N1->getOperand(0),
3543 N0->getOperand(0), &Mask2[0]);
3548 // fold (or c1, c2) -> c1|c2
3549 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3550 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3552 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
3553 // canonicalize constant to RHS
3554 if (isConstantIntBuildVectorOrConstantInt(N0) &&
3555 !isConstantIntBuildVectorOrConstantInt(N1))
3556 return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0);
3557 // fold (or x, 0) -> x
3558 if (N1C && N1C->isNullValue())
3560 // fold (or x, -1) -> -1
3561 if (N1C && N1C->isAllOnesValue())
3563 // fold (or x, c) -> c iff (x & ~c) == 0
3564 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3567 if (SDValue Combined = visitORLike(N0, N1, N))
3570 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3571 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
3572 if (BSwap.getNode())
3574 BSwap = MatchBSwapHWordLow(N, N0, N1);
3575 if (BSwap.getNode())
3579 if (SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1))
3581 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3582 // iff (c1 & c2) == 0.
3583 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3584 isa<ConstantSDNode>(N0.getOperand(1))) {
3585 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3586 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) {
3587 if (SDValue COR = DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1))
3589 ISD::AND, SDLoc(N), VT,
3590 DAG.getNode(ISD::OR, SDLoc(N0), VT, N0.getOperand(0), N1), COR);
3594 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3595 if (N0.getOpcode() == N1.getOpcode()) {
3596 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3597 if (Tmp.getNode()) return Tmp;
3600 // See if this is some rotate idiom.
3601 if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N)))
3602 return SDValue(Rot, 0);
3604 // Simplify the operands using demanded-bits information.
3605 if (!VT.isVector() &&
3606 SimplifyDemandedBits(SDValue(N, 0)))
3607 return SDValue(N, 0);
3612 /// Match "(X shl/srl V1) & V2" where V2 may not be present.
3613 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3614 if (Op.getOpcode() == ISD::AND) {
3615 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3616 Mask = Op.getOperand(1);
3617 Op = Op.getOperand(0);
3623 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3631 // Return true if we can prove that, whenever Neg and Pos are both in the
3632 // range [0, OpSize), Neg == (Pos == 0 ? 0 : OpSize - Pos). This means that
3633 // for two opposing shifts shift1 and shift2 and a value X with OpBits bits:
3635 // (or (shift1 X, Neg), (shift2 X, Pos))
3637 // reduces to a rotate in direction shift2 by Pos or (equivalently) a rotate
3638 // in direction shift1 by Neg. The range [0, OpSize) means that we only need
3639 // to consider shift amounts with defined behavior.
3640 static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned OpSize) {
3641 // If OpSize is a power of 2 then:
3643 // (a) (Pos == 0 ? 0 : OpSize - Pos) == (OpSize - Pos) & (OpSize - 1)
3644 // (b) Neg == Neg & (OpSize - 1) whenever Neg is in [0, OpSize).
3646 // So if OpSize is a power of 2 and Neg is (and Neg', OpSize-1), we check
3647 // for the stronger condition:
3649 // Neg & (OpSize - 1) == (OpSize - Pos) & (OpSize - 1) [A]
3651 // for all Neg and Pos. Since Neg & (OpSize - 1) == Neg' & (OpSize - 1)
3652 // we can just replace Neg with Neg' for the rest of the function.
3654 // In other cases we check for the even stronger condition:
3656 // Neg == OpSize - Pos [B]
3658 // for all Neg and Pos. Note that the (or ...) then invokes undefined
3659 // behavior if Pos == 0 (and consequently Neg == OpSize).
3661 // We could actually use [A] whenever OpSize is a power of 2, but the
3662 // only extra cases that it would match are those uninteresting ones
3663 // where Neg and Pos are never in range at the same time. E.g. for
3664 // OpSize == 32, using [A] would allow a Neg of the form (sub 64, Pos)
3665 // as well as (sub 32, Pos), but:
3667 // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos))
3669 // always invokes undefined behavior for 32-bit X.
3671 // Below, Mask == OpSize - 1 when using [A] and is all-ones otherwise.
3672 unsigned MaskLoBits = 0;
3673 if (Neg.getOpcode() == ISD::AND &&
3674 isPowerOf2_64(OpSize) &&
3675 Neg.getOperand(1).getOpcode() == ISD::Constant &&
3676 cast<ConstantSDNode>(Neg.getOperand(1))->getAPIntValue() == OpSize - 1) {
3677 Neg = Neg.getOperand(0);
3678 MaskLoBits = Log2_64(OpSize);
3681 // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1.
3682 if (Neg.getOpcode() != ISD::SUB)
3684 ConstantSDNode *NegC = dyn_cast<ConstantSDNode>(Neg.getOperand(0));
3687 SDValue NegOp1 = Neg.getOperand(1);
3689 // On the RHS of [A], if Pos is Pos' & (OpSize - 1), just replace Pos with
3690 // Pos'. The truncation is redundant for the purpose of the equality.
3692 Pos.getOpcode() == ISD::AND &&
3693 Pos.getOperand(1).getOpcode() == ISD::Constant &&
3694 cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() == OpSize - 1)
3695 Pos = Pos.getOperand(0);
3697 // The condition we need is now:
3699 // (NegC - NegOp1) & Mask == (OpSize - Pos) & Mask
3701 // If NegOp1 == Pos then we need:
3703 // OpSize & Mask == NegC & Mask
3705 // (because "x & Mask" is a truncation and distributes through subtraction).
3708 Width = NegC->getAPIntValue();
3709 // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC.
3710 // Then the condition we want to prove becomes:
3712 // (NegC - NegOp1) & Mask == (OpSize - (NegOp1 + PosC)) & Mask
3714 // which, again because "x & Mask" is a truncation, becomes:
3716 // NegC & Mask == (OpSize - PosC) & Mask
3717 // OpSize & Mask == (NegC + PosC) & Mask
3718 else if (Pos.getOpcode() == ISD::ADD &&
3719 Pos.getOperand(0) == NegOp1 &&
3720 Pos.getOperand(1).getOpcode() == ISD::Constant)
3721 Width = (cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() +
3722 NegC->getAPIntValue());
3726 // Now we just need to check that OpSize & Mask == Width & Mask.
3728 // Opsize & Mask is 0 since Mask is Opsize - 1.
3729 return Width.getLoBits(MaskLoBits) == 0;
3730 return Width == OpSize;
3733 // A subroutine of MatchRotate used once we have found an OR of two opposite
3734 // shifts of Shifted. If Neg == <operand size> - Pos then the OR reduces
3735 // to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the
3736 // former being preferred if supported. InnerPos and InnerNeg are Pos and
3737 // Neg with outer conversions stripped away.
3738 SDNode *DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos,
3739 SDValue Neg, SDValue InnerPos,
3740 SDValue InnerNeg, unsigned PosOpcode,
3741 unsigned NegOpcode, SDLoc DL) {
3742 // fold (or (shl x, (*ext y)),
3743 // (srl x, (*ext (sub 32, y)))) ->
3744 // (rotl x, y) or (rotr x, (sub 32, y))
3746 // fold (or (shl x, (*ext (sub 32, y))),
3747 // (srl x, (*ext y))) ->
3748 // (rotr x, y) or (rotl x, (sub 32, y))
3749 EVT VT = Shifted.getValueType();
3750 if (matchRotateSub(InnerPos, InnerNeg, VT.getSizeInBits())) {
3751 bool HasPos = TLI.isOperationLegalOrCustom(PosOpcode, VT);
3752 return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted,
3753 HasPos ? Pos : Neg).getNode();
3759 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3760 // idioms for rotate, and if the target supports rotation instructions, generate
3762 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL) {
3763 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3764 EVT VT = LHS.getValueType();
3765 if (!TLI.isTypeLegal(VT)) return nullptr;
3767 // The target must have at least one rotate flavor.
3768 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3769 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3770 if (!HasROTL && !HasROTR) return nullptr;
3772 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3773 SDValue LHSShift; // The shift.
3774 SDValue LHSMask; // AND value if any.
3775 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3776 return nullptr; // Not part of a rotate.
3778 SDValue RHSShift; // The shift.
3779 SDValue RHSMask; // AND value if any.
3780 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3781 return nullptr; // Not part of a rotate.
3783 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3784 return nullptr; // Not shifting the same value.
3786 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3787 return nullptr; // Shifts must disagree.
3789 // Canonicalize shl to left side in a shl/srl pair.
3790 if (RHSShift.getOpcode() == ISD::SHL) {
3791 std::swap(LHS, RHS);
3792 std::swap(LHSShift, RHSShift);
3793 std::swap(LHSMask , RHSMask );
3796 unsigned OpSizeInBits = VT.getSizeInBits();
3797 SDValue LHSShiftArg = LHSShift.getOperand(0);
3798 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3799 SDValue RHSShiftArg = RHSShift.getOperand(0);
3800 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3802 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3803 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3804 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3805 RHSShiftAmt.getOpcode() == ISD::Constant) {
3806 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3807 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3808 if ((LShVal + RShVal) != OpSizeInBits)
3811 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3812 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
3814 // If there is an AND of either shifted operand, apply it to the result.
3815 if (LHSMask.getNode() || RHSMask.getNode()) {
3816 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3818 if (LHSMask.getNode()) {
3819 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3820 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3822 if (RHSMask.getNode()) {
3823 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3824 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3827 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3830 return Rot.getNode();
3833 // If there is a mask here, and we have a variable shift, we can't be sure
3834 // that we're masking out the right stuff.
3835 if (LHSMask.getNode() || RHSMask.getNode())
3838 // If the shift amount is sign/zext/any-extended just peel it off.
3839 SDValue LExtOp0 = LHSShiftAmt;
3840 SDValue RExtOp0 = RHSShiftAmt;
3841 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3842 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3843 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3844 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3845 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3846 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3847 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3848 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3849 LExtOp0 = LHSShiftAmt.getOperand(0);
3850 RExtOp0 = RHSShiftAmt.getOperand(0);
3853 SDNode *TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt,
3854 LExtOp0, RExtOp0, ISD::ROTL, ISD::ROTR, DL);
3858 SDNode *TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt,
3859 RExtOp0, LExtOp0, ISD::ROTR, ISD::ROTL, DL);
3866 SDValue DAGCombiner::visitXOR(SDNode *N) {
3867 SDValue N0 = N->getOperand(0);
3868 SDValue N1 = N->getOperand(1);
3869 EVT VT = N0.getValueType();
3872 if (VT.isVector()) {
3873 if (SDValue FoldedVOp = SimplifyVBinOp(N))
3876 // fold (xor x, 0) -> x, vector edition
3877 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3879 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3883 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3884 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3885 return DAG.getConstant(0, VT);
3886 // fold (xor x, undef) -> undef
3887 if (N0.getOpcode() == ISD::UNDEF)
3889 if (N1.getOpcode() == ISD::UNDEF)
3891 // fold (xor c1, c2) -> c1^c2
3892 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3893 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3895 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3896 // canonicalize constant to RHS
3897 if (isConstantIntBuildVectorOrConstantInt(N0) &&
3898 !isConstantIntBuildVectorOrConstantInt(N1))
3899 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
3900 // fold (xor x, 0) -> x
3901 if (N1C && N1C->isNullValue())
3904 if (SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1))
3907 // fold !(x cc y) -> (x !cc y)
3908 SDValue LHS, RHS, CC;
3909 if (TLI.isConstTrueVal(N1.getNode()) && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3910 bool isInt = LHS.getValueType().isInteger();
3911 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3914 if (!LegalOperations ||
3915 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) {
3916 switch (N0.getOpcode()) {
3918 llvm_unreachable("Unhandled SetCC Equivalent!");
3920 return DAG.getSetCC(SDLoc(N), VT, LHS, RHS, NotCC);
3921 case ISD::SELECT_CC:
3922 return DAG.getSelectCC(SDLoc(N), LHS, RHS, N0.getOperand(2),
3923 N0.getOperand(3), NotCC);
3928 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3929 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3930 N0.getNode()->hasOneUse() &&
3931 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3932 SDValue V = N0.getOperand(0);
3933 V = DAG.getNode(ISD::XOR, SDLoc(N0), V.getValueType(), V,
3934 DAG.getConstant(1, V.getValueType()));
3935 AddToWorklist(V.getNode());
3936 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V);
3939 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3940 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3941 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3942 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3943 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3944 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3945 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3946 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3947 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
3948 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3951 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3952 if (N1C && N1C->isAllOnesValue() &&
3953 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3954 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3955 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3956 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3957 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3958 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3959 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
3960 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3963 // fold (xor (and x, y), y) -> (and (not x), y)
3964 if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3965 N0->getOperand(1) == N1) {
3966 SDValue X = N0->getOperand(0);
3967 SDValue NotX = DAG.getNOT(SDLoc(X), X, VT);
3968 AddToWorklist(NotX.getNode());
3969 return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1);
3971 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3972 if (N1C && N0.getOpcode() == ISD::XOR) {
3973 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3974 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3976 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(1),
3977 DAG.getConstant(N1C->getAPIntValue() ^
3978 N00C->getAPIntValue(), VT));
3980 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(0),
3981 DAG.getConstant(N1C->getAPIntValue() ^
3982 N01C->getAPIntValue(), VT));
3984 // fold (xor x, x) -> 0
3986 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
3988 // fold (xor (shl 1, x), -1) -> (rotl ~1, x)
3989 // Here is a concrete example of this equivalence:
3991 // i16 shl == 1 << 14 == 16384 == 0b0100000000000000
3992 // i16 xor == ~(1 << 14) == 49151 == 0b1011111111111111
3996 // i16 ~1 == 0b1111111111111110
3997 // i16 rol(~1, 14) == 0b1011111111111111
3999 // Some additional tips to help conceptualize this transform:
4000 // - Try to see the operation as placing a single zero in a value of all ones.
4001 // - There exists no value for x which would allow the result to contain zero.
4002 // - Values of x larger than the bitwidth are undefined and do not require a
4003 // consistent result.
4004 // - Pushing the zero left requires shifting one bits in from the right.
4005 // A rotate left of ~1 is a nice way of achieving the desired result.
4006 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
4007 if (auto *N1C = dyn_cast<ConstantSDNode>(N1.getNode()))
4008 if (N0.getOpcode() == ISD::SHL)
4009 if (auto *ShlLHS = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
4010 if (N1C->isAllOnesValue() && ShlLHS->isOne())
4011 return DAG.getNode(ISD::ROTL, SDLoc(N), VT, DAG.getConstant(~1, VT),
4014 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
4015 if (N0.getOpcode() == N1.getOpcode()) {
4016 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
4017 if (Tmp.getNode()) return Tmp;
4020 // Simplify the expression using non-local knowledge.
4021 if (!VT.isVector() &&
4022 SimplifyDemandedBits(SDValue(N, 0)))
4023 return SDValue(N, 0);
4028 /// Handle transforms common to the three shifts, when the shift amount is a
4030 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, ConstantSDNode *Amt) {
4031 // We can't and shouldn't fold opaque constants.
4032 if (Amt->isOpaque())
4035 SDNode *LHS = N->getOperand(0).getNode();
4036 if (!LHS->hasOneUse()) return SDValue();
4038 // We want to pull some binops through shifts, so that we have (and (shift))
4039 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
4040 // thing happens with address calculations, so it's important to canonicalize
4042 bool HighBitSet = false; // Can we transform this if the high bit is set?
4044 switch (LHS->getOpcode()) {
4045 default: return SDValue();
4048 HighBitSet = false; // We can only transform sra if the high bit is clear.
4051 HighBitSet = true; // We can only transform sra if the high bit is set.
4054 if (N->getOpcode() != ISD::SHL)
4055 return SDValue(); // only shl(add) not sr[al](add).
4056 HighBitSet = false; // We can only transform sra if the high bit is clear.
4060 // We require the RHS of the binop to be a constant and not opaque as well.
4061 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
4062 if (!BinOpCst || BinOpCst->isOpaque()) return SDValue();
4064 // FIXME: disable this unless the input to the binop is a shift by a constant.
4065 // If it is not a shift, it pessimizes some common cases like:
4067 // void foo(int *X, int i) { X[i & 1235] = 1; }
4068 // int bar(int *X, int i) { return X[i & 255]; }
4069 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
4070 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
4071 BinOpLHSVal->getOpcode() != ISD::SRA &&
4072 BinOpLHSVal->getOpcode() != ISD::SRL) ||
4073 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
4076 EVT VT = N->getValueType(0);
4078 // If this is a signed shift right, and the high bit is modified by the
4079 // logical operation, do not perform the transformation. The highBitSet
4080 // boolean indicates the value of the high bit of the constant which would
4081 // cause it to be modified for this operation.
4082 if (N->getOpcode() == ISD::SRA) {
4083 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
4084 if (BinOpRHSSignSet != HighBitSet)
4088 if (!TLI.isDesirableToCommuteWithShift(LHS))
4091 // Fold the constants, shifting the binop RHS by the shift amount.
4092 SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)),
4094 LHS->getOperand(1), N->getOperand(1));
4095 assert(isa<ConstantSDNode>(NewRHS) && "Folding was not successful!");
4097 // Create the new shift.
4098 SDValue NewShift = DAG.getNode(N->getOpcode(),
4099 SDLoc(LHS->getOperand(0)),
4100 VT, LHS->getOperand(0), N->getOperand(1));
4102 // Create the new binop.
4103 return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS);
4106 SDValue DAGCombiner::distributeTruncateThroughAnd(SDNode *N) {
4107 assert(N->getOpcode() == ISD::TRUNCATE);
4108 assert(N->getOperand(0).getOpcode() == ISD::AND);
4110 // (truncate:TruncVT (and N00, N01C)) -> (and (truncate:TruncVT N00), TruncC)
4111 if (N->hasOneUse() && N->getOperand(0).hasOneUse()) {
4112 SDValue N01 = N->getOperand(0).getOperand(1);
4114 if (ConstantSDNode *N01C = isConstOrConstSplat(N01)) {
4115 EVT TruncVT = N->getValueType(0);
4116 SDValue N00 = N->getOperand(0).getOperand(0);
4117 APInt TruncC = N01C->getAPIntValue();
4118 TruncC = TruncC.trunc(TruncVT.getScalarSizeInBits());
4120 return DAG.getNode(ISD::AND, SDLoc(N), TruncVT,
4121 DAG.getNode(ISD::TRUNCATE, SDLoc(N), TruncVT, N00),
4122 DAG.getConstant(TruncC, TruncVT));
4129 SDValue DAGCombiner::visitRotate(SDNode *N) {
4130 // fold (rot* x, (trunc (and y, c))) -> (rot* x, (and (trunc y), (trunc c))).
4131 if (N->getOperand(1).getOpcode() == ISD::TRUNCATE &&
4132 N->getOperand(1).getOperand(0).getOpcode() == ISD::AND) {
4133 SDValue NewOp1 = distributeTruncateThroughAnd(N->getOperand(1).getNode());
4134 if (NewOp1.getNode())
4135 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
4136 N->getOperand(0), NewOp1);
4141 SDValue DAGCombiner::visitSHL(SDNode *N) {
4142 SDValue N0 = N->getOperand(0);
4143 SDValue N1 = N->getOperand(1);
4144 EVT VT = N0.getValueType();
4145 unsigned OpSizeInBits = VT.getScalarSizeInBits();
4148 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4149 if (VT.isVector()) {
4150 if (SDValue FoldedVOp = SimplifyVBinOp(N))
4153 BuildVectorSDNode *N1CV = dyn_cast<BuildVectorSDNode>(N1);
4154 // If setcc produces all-one true value then:
4155 // (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<<N1CV)
4156 if (N1CV && N1CV->isConstant()) {
4157 if (N0.getOpcode() == ISD::AND) {
4158 SDValue N00 = N0->getOperand(0);
4159 SDValue N01 = N0->getOperand(1);
4160 BuildVectorSDNode *N01CV = dyn_cast<BuildVectorSDNode>(N01);
4162 if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC &&
4163 TLI.getBooleanContents(N00.getOperand(0).getValueType()) ==
4164 TargetLowering::ZeroOrNegativeOneBooleanContent) {
4165 if (SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, VT, N01CV, N1CV))
4166 return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C);
4169 N1C = isConstOrConstSplat(N1);
4174 // fold (shl c1, c2) -> c1<<c2
4175 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4177 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
4178 // fold (shl 0, x) -> 0
4179 if (N0C && N0C->isNullValue())
4181 // fold (shl x, c >= size(x)) -> undef
4182 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4183 return DAG.getUNDEF(VT);
4184 // fold (shl x, 0) -> x
4185 if (N1C && N1C->isNullValue())
4187 // fold (shl undef, x) -> 0
4188 if (N0.getOpcode() == ISD::UNDEF)
4189 return DAG.getConstant(0, VT);
4190 // if (shl x, c) is known to be zero, return 0
4191 if (DAG.MaskedValueIsZero(SDValue(N, 0),
4192 APInt::getAllOnesValue(OpSizeInBits)))
4193 return DAG.getConstant(0, VT);
4194 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
4195 if (N1.getOpcode() == ISD::TRUNCATE &&
4196 N1.getOperand(0).getOpcode() == ISD::AND) {
4197 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4198 if (NewOp1.getNode())
4199 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, NewOp1);
4202 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4203 return SDValue(N, 0);
4205 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
4206 if (N1C && N0.getOpcode() == ISD::SHL) {
4207 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4208 uint64_t c1 = N0C1->getZExtValue();
4209 uint64_t c2 = N1C->getZExtValue();
4210 if (c1 + c2 >= OpSizeInBits)
4211 return DAG.getConstant(0, VT);
4212 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4213 DAG.getConstant(c1 + c2, N1.getValueType()));
4217 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
4218 // For this to be valid, the second form must not preserve any of the bits
4219 // that are shifted out by the inner shift in the first form. This means
4220 // the outer shift size must be >= the number of bits added by the ext.
4221 // As a corollary, we don't care what kind of ext it is.
4222 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
4223 N0.getOpcode() == ISD::ANY_EXTEND ||
4224 N0.getOpcode() == ISD::SIGN_EXTEND) &&
4225 N0.getOperand(0).getOpcode() == ISD::SHL) {
4226 SDValue N0Op0 = N0.getOperand(0);
4227 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4228 uint64_t c1 = N0Op0C1->getZExtValue();
4229 uint64_t c2 = N1C->getZExtValue();
4230 EVT InnerShiftVT = N0Op0.getValueType();
4231 uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits();
4232 if (c2 >= OpSizeInBits - InnerShiftSize) {
4233 if (c1 + c2 >= OpSizeInBits)
4234 return DAG.getConstant(0, VT);
4235 return DAG.getNode(ISD::SHL, SDLoc(N0), VT,
4236 DAG.getNode(N0.getOpcode(), SDLoc(N0), VT,
4237 N0Op0->getOperand(0)),
4238 DAG.getConstant(c1 + c2, N1.getValueType()));
4243 // fold (shl (zext (srl x, C)), C) -> (zext (shl (srl x, C), C))
4244 // Only fold this if the inner zext has no other uses to avoid increasing
4245 // the total number of instructions.
4246 if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() &&
4247 N0.getOperand(0).getOpcode() == ISD::SRL) {
4248 SDValue N0Op0 = N0.getOperand(0);
4249 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4250 uint64_t c1 = N0Op0C1->getZExtValue();
4251 if (c1 < VT.getScalarSizeInBits()) {
4252 uint64_t c2 = N1C->getZExtValue();
4254 SDValue NewOp0 = N0.getOperand(0);
4255 EVT CountVT = NewOp0.getOperand(1).getValueType();
4256 SDValue NewSHL = DAG.getNode(ISD::SHL, SDLoc(N), NewOp0.getValueType(),
4257 NewOp0, DAG.getConstant(c2, CountVT));
4258 AddToWorklist(NewSHL.getNode());
4259 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
4265 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
4266 // (and (srl x, (sub c1, c2), MASK)
4267 // Only fold this if the inner shift has no other uses -- if it does, folding
4268 // this will increase the total number of instructions.
4269 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
4270 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4271 uint64_t c1 = N0C1->getZExtValue();
4272 if (c1 < OpSizeInBits) {
4273 uint64_t c2 = N1C->getZExtValue();
4274 APInt Mask = APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - c1);
4277 Mask = Mask.shl(c2 - c1);
4278 Shift = DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4279 DAG.getConstant(c2 - c1, N1.getValueType()));
4281 Mask = Mask.lshr(c1 - c2);
4282 Shift = DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4283 DAG.getConstant(c1 - c2, N1.getValueType()));
4285 return DAG.getNode(ISD::AND, SDLoc(N0), VT, Shift,
4286 DAG.getConstant(Mask, VT));
4290 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
4291 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
4292 unsigned BitSize = VT.getScalarSizeInBits();
4293 SDValue HiBitsMask =
4294 DAG.getConstant(APInt::getHighBitsSet(BitSize,
4295 BitSize - N1C->getZExtValue()), VT);
4296 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4300 // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
4301 // Variant of version done on multiply, except mul by a power of 2 is turned
4304 if (N1C && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
4305 (isa<ConstantSDNode>(N0.getOperand(1)) ||
4306 isConstantSplatVector(N0.getOperand(1).getNode(), Val))) {
4307 SDValue Shl0 = DAG.getNode(ISD::SHL, SDLoc(N0), VT, N0.getOperand(0), N1);
4308 SDValue Shl1 = DAG.getNode(ISD::SHL, SDLoc(N1), VT, N0.getOperand(1), N1);
4309 return DAG.getNode(ISD::ADD, SDLoc(N), VT, Shl0, Shl1);
4313 SDValue NewSHL = visitShiftByConstant(N, N1C);
4314 if (NewSHL.getNode())
4321 SDValue DAGCombiner::visitSRA(SDNode *N) {
4322 SDValue N0 = N->getOperand(0);
4323 SDValue N1 = N->getOperand(1);
4324 EVT VT = N0.getValueType();
4325 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4328 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4329 if (VT.isVector()) {
4330 if (SDValue FoldedVOp = SimplifyVBinOp(N))
4333 N1C = isConstOrConstSplat(N1);
4336 // fold (sra c1, c2) -> (sra c1, c2)
4337 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4339 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
4340 // fold (sra 0, x) -> 0
4341 if (N0C && N0C->isNullValue())
4343 // fold (sra -1, x) -> -1
4344 if (N0C && N0C->isAllOnesValue())
4346 // fold (sra x, (setge c, size(x))) -> undef
4347 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4348 return DAG.getUNDEF(VT);
4349 // fold (sra x, 0) -> x
4350 if (N1C && N1C->isNullValue())
4352 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
4354 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
4355 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
4356 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
4358 ExtVT = EVT::getVectorVT(*DAG.getContext(),
4359 ExtVT, VT.getVectorNumElements());
4360 if ((!LegalOperations ||
4361 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
4362 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
4363 N0.getOperand(0), DAG.getValueType(ExtVT));
4366 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
4367 if (N1C && N0.getOpcode() == ISD::SRA) {
4368 if (ConstantSDNode *C1 = isConstOrConstSplat(N0.getOperand(1))) {
4369 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
4370 if (Sum >= OpSizeInBits)
4371 Sum = OpSizeInBits - 1;
4372 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0.getOperand(0),
4373 DAG.getConstant(Sum, N1.getValueType()));
4377 // fold (sra (shl X, m), (sub result_size, n))
4378 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
4379 // result_size - n != m.
4380 // If truncate is free for the target sext(shl) is likely to result in better
4382 if (N0.getOpcode() == ISD::SHL && N1C) {
4383 // Get the two constanst of the shifts, CN0 = m, CN = n.
4384 const ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1));
4386 LLVMContext &Ctx = *DAG.getContext();
4387 // Determine what the truncate's result bitsize and type would be.
4388 EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - N1C->getZExtValue());
4391 TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorNumElements());
4393 // Determine the residual right-shift amount.
4394 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
4396 // If the shift is not a no-op (in which case this should be just a sign
4397 // extend already), the truncated to type is legal, sign_extend is legal
4398 // on that type, and the truncate to that type is both legal and free,
4399 // perform the transform.
4400 if ((ShiftAmt > 0) &&
4401 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
4402 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
4403 TLI.isTruncateFree(VT, TruncVT)) {
4405 SDValue Amt = DAG.getConstant(ShiftAmt,
4406 getShiftAmountTy(N0.getOperand(0).getValueType()));
4407 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0), VT,
4408 N0.getOperand(0), Amt);
4409 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), TruncVT,
4411 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N),
4412 N->getValueType(0), Trunc);
4417 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
4418 if (N1.getOpcode() == ISD::TRUNCATE &&
4419 N1.getOperand(0).getOpcode() == ISD::AND) {
4420 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4421 if (NewOp1.getNode())
4422 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, NewOp1);
4425 // fold (sra (trunc (srl x, c1)), c2) -> (trunc (sra x, c1 + c2))
4426 // if c1 is equal to the number of bits the trunc removes
4427 if (N0.getOpcode() == ISD::TRUNCATE &&
4428 (N0.getOperand(0).getOpcode() == ISD::SRL ||
4429 N0.getOperand(0).getOpcode() == ISD::SRA) &&
4430 N0.getOperand(0).hasOneUse() &&
4431 N0.getOperand(0).getOperand(1).hasOneUse() &&
4433 SDValue N0Op0 = N0.getOperand(0);
4434 if (ConstantSDNode *LargeShift = isConstOrConstSplat(N0Op0.getOperand(1))) {
4435 unsigned LargeShiftVal = LargeShift->getZExtValue();
4436 EVT LargeVT = N0Op0.getValueType();
4438 if (LargeVT.getScalarSizeInBits() - OpSizeInBits == LargeShiftVal) {
4440 DAG.getConstant(LargeShiftVal + N1C->getZExtValue(),
4441 getShiftAmountTy(N0Op0.getOperand(0).getValueType()));
4442 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), LargeVT,
4443 N0Op0.getOperand(0), Amt);
4444 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, SRA);
4449 // Simplify, based on bits shifted out of the LHS.
4450 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4451 return SDValue(N, 0);
4454 // If the sign bit is known to be zero, switch this to a SRL.
4455 if (DAG.SignBitIsZero(N0))
4456 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1);
4459 SDValue NewSRA = visitShiftByConstant(N, N1C);
4460 if (NewSRA.getNode())
4467 SDValue DAGCombiner::visitSRL(SDNode *N) {
4468 SDValue N0 = N->getOperand(0);
4469 SDValue N1 = N->getOperand(1);
4470 EVT VT = N0.getValueType();
4471 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4474 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4475 if (VT.isVector()) {
4476 if (SDValue FoldedVOp = SimplifyVBinOp(N))
4479 N1C = isConstOrConstSplat(N1);
4482 // fold (srl c1, c2) -> c1 >>u c2
4483 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4485 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
4486 // fold (srl 0, x) -> 0
4487 if (N0C && N0C->isNullValue())
4489 // fold (srl x, c >= size(x)) -> undef
4490 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4491 return DAG.getUNDEF(VT);
4492 // fold (srl x, 0) -> x
4493 if (N1C && N1C->isNullValue())
4495 // if (srl x, c) is known to be zero, return 0
4496 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
4497 APInt::getAllOnesValue(OpSizeInBits)))
4498 return DAG.getConstant(0, VT);
4500 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
4501 if (N1C && N0.getOpcode() == ISD::SRL) {
4502 if (ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1))) {
4503 uint64_t c1 = N01C->getZExtValue();
4504 uint64_t c2 = N1C->getZExtValue();
4505 if (c1 + c2 >= OpSizeInBits)
4506 return DAG.getConstant(0, VT);
4507 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4508 DAG.getConstant(c1 + c2, N1.getValueType()));
4512 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
4513 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
4514 N0.getOperand(0).getOpcode() == ISD::SRL &&
4515 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
4517 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
4518 uint64_t c2 = N1C->getZExtValue();
4519 EVT InnerShiftVT = N0.getOperand(0).getValueType();
4520 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
4521 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
4522 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
4523 if (c1 + OpSizeInBits == InnerShiftSize) {
4524 if (c1 + c2 >= InnerShiftSize)
4525 return DAG.getConstant(0, VT);
4526 return DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT,
4527 DAG.getNode(ISD::SRL, SDLoc(N0), InnerShiftVT,
4528 N0.getOperand(0)->getOperand(0),
4529 DAG.getConstant(c1 + c2, ShiftCountVT)));
4533 // fold (srl (shl x, c), c) -> (and x, cst2)
4534 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1) {
4535 unsigned BitSize = N0.getScalarValueSizeInBits();
4536 if (BitSize <= 64) {
4537 uint64_t ShAmt = N1C->getZExtValue() + 64 - BitSize;
4538 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4539 DAG.getConstant(~0ULL >> ShAmt, VT));
4543 // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask)
4544 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
4545 // Shifting in all undef bits?
4546 EVT SmallVT = N0.getOperand(0).getValueType();
4547 unsigned BitSize = SmallVT.getScalarSizeInBits();
4548 if (N1C->getZExtValue() >= BitSize)
4549 return DAG.getUNDEF(VT);
4551 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
4552 uint64_t ShiftAmt = N1C->getZExtValue();
4553 SDValue SmallShift = DAG.getNode(ISD::SRL, SDLoc(N0), SmallVT,
4555 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
4556 AddToWorklist(SmallShift.getNode());
4557 APInt Mask = APInt::getAllOnesValue(OpSizeInBits).lshr(ShiftAmt);
4558 return DAG.getNode(ISD::AND, SDLoc(N), VT,
4559 DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, SmallShift),
4560 DAG.getConstant(Mask, VT));
4564 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
4565 // bit, which is unmodified by sra.
4566 if (N1C && N1C->getZExtValue() + 1 == OpSizeInBits) {
4567 if (N0.getOpcode() == ISD::SRA)
4568 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1);
4571 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
4572 if (N1C && N0.getOpcode() == ISD::CTLZ &&
4573 N1C->getAPIntValue() == Log2_32(OpSizeInBits)) {
4574 APInt KnownZero, KnownOne;
4575 DAG.computeKnownBits(N0.getOperand(0), KnownZero, KnownOne);
4577 // If any of the input bits are KnownOne, then the input couldn't be all
4578 // zeros, thus the result of the srl will always be zero.
4579 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
4581 // If all of the bits input the to ctlz node are known to be zero, then
4582 // the result of the ctlz is "32" and the result of the shift is one.
4583 APInt UnknownBits = ~KnownZero;
4584 if (UnknownBits == 0) return DAG.getConstant(1, VT);
4586 // Otherwise, check to see if there is exactly one bit input to the ctlz.
4587 if ((UnknownBits & (UnknownBits - 1)) == 0) {
4588 // Okay, we know that only that the single bit specified by UnknownBits
4589 // could be set on input to the CTLZ node. If this bit is set, the SRL
4590 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
4591 // to an SRL/XOR pair, which is likely to simplify more.
4592 unsigned ShAmt = UnknownBits.countTrailingZeros();
4593 SDValue Op = N0.getOperand(0);
4596 Op = DAG.getNode(ISD::SRL, SDLoc(N0), VT, Op,
4597 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
4598 AddToWorklist(Op.getNode());
4601 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
4602 Op, DAG.getConstant(1, VT));
4606 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
4607 if (N1.getOpcode() == ISD::TRUNCATE &&
4608 N1.getOperand(0).getOpcode() == ISD::AND) {
4609 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4610 if (NewOp1.getNode())
4611 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, NewOp1);
4614 // fold operands of srl based on knowledge that the low bits are not
4616 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4617 return SDValue(N, 0);
4620 SDValue NewSRL = visitShiftByConstant(N, N1C);
4621 if (NewSRL.getNode())
4625 // Attempt to convert a srl of a load into a narrower zero-extending load.
4626 SDValue NarrowLoad = ReduceLoadWidth(N);
4627 if (NarrowLoad.getNode())
4630 // Here is a common situation. We want to optimize:
4633 // %b = and i32 %a, 2
4634 // %c = srl i32 %b, 1
4635 // brcond i32 %c ...
4641 // %c = setcc eq %b, 0
4644 // However when after the source operand of SRL is optimized into AND, the SRL
4645 // itself may not be optimized further. Look for it and add the BRCOND into
4647 if (N->hasOneUse()) {
4648 SDNode *Use = *N->use_begin();
4649 if (Use->getOpcode() == ISD::BRCOND)
4651 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
4652 // Also look pass the truncate.
4653 Use = *Use->use_begin();
4654 if (Use->getOpcode() == ISD::BRCOND)
4662 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
4663 SDValue N0 = N->getOperand(0);
4664 EVT VT = N->getValueType(0);
4666 // fold (ctlz c1) -> c2
4667 if (isa<ConstantSDNode>(N0))
4668 return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0);
4672 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4673 SDValue N0 = N->getOperand(0);
4674 EVT VT = N->getValueType(0);
4676 // fold (ctlz_zero_undef c1) -> c2
4677 if (isa<ConstantSDNode>(N0))
4678 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4682 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4683 SDValue N0 = N->getOperand(0);
4684 EVT VT = N->getValueType(0);
4686 // fold (cttz c1) -> c2
4687 if (isa<ConstantSDNode>(N0))
4688 return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0);
4692 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4693 SDValue N0 = N->getOperand(0);
4694 EVT VT = N->getValueType(0);
4696 // fold (cttz_zero_undef c1) -> c2
4697 if (isa<ConstantSDNode>(N0))
4698 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4702 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4703 SDValue N0 = N->getOperand(0);
4704 EVT VT = N->getValueType(0);
4706 // fold (ctpop c1) -> c2
4707 if (isa<ConstantSDNode>(N0))
4708 return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0);
4713 /// \brief Generate Min/Max node
4714 static SDValue combineMinNumMaxNum(SDLoc DL, EVT VT, SDValue LHS, SDValue RHS,
4715 SDValue True, SDValue False,
4716 ISD::CondCode CC, const TargetLowering &TLI,
4717 SelectionDAG &DAG) {
4718 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
4728 unsigned Opcode = (LHS == True) ? ISD::FMINNUM : ISD::FMAXNUM;
4729 if (TLI.isOperationLegal(Opcode, VT))
4730 return DAG.getNode(Opcode, DL, VT, LHS, RHS);
4739 unsigned Opcode = (LHS == True) ? ISD::FMAXNUM : ISD::FMINNUM;
4740 if (TLI.isOperationLegal(Opcode, VT))
4741 return DAG.getNode(Opcode, DL, VT, LHS, RHS);
4749 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4750 SDValue N0 = N->getOperand(0);
4751 SDValue N1 = N->getOperand(1);
4752 SDValue N2 = N->getOperand(2);
4753 EVT VT = N->getValueType(0);
4754 EVT VT0 = N0.getValueType();
4756 // fold (select C, X, X) -> X
4759 // fold (select true, X, Y) -> X
4760 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4761 if (N0C && !N0C->isNullValue())
4763 // fold (select false, X, Y) -> Y
4764 if (N0C && N0C->isNullValue())
4766 // fold (select C, 1, X) -> (or C, X)
4767 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4768 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4769 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4770 // fold (select C, 0, 1) -> (xor C, 1)
4771 // We can't do this reliably if integer based booleans have different contents
4772 // to floating point based booleans. This is because we can't tell whether we
4773 // have an integer-based boolean or a floating-point-based boolean unless we
4774 // can find the SETCC that produced it and inspect its operands. This is
4775 // fairly easy if C is the SETCC node, but it can potentially be
4776 // undiscoverable (or not reasonably discoverable). For example, it could be
4777 // in another basic block or it could require searching a complicated
4779 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
4780 if (VT.isInteger() &&
4781 (VT0 == MVT::i1 || (VT0.isInteger() &&
4782 TLI.getBooleanContents(false, false) ==
4783 TLI.getBooleanContents(false, true) &&
4784 TLI.getBooleanContents(false, false) ==
4785 TargetLowering::ZeroOrOneBooleanContent)) &&
4786 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4789 return DAG.getNode(ISD::XOR, SDLoc(N), VT0,
4790 N0, DAG.getConstant(1, VT0));
4791 XORNode = DAG.getNode(ISD::XOR, SDLoc(N0), VT0,
4792 N0, DAG.getConstant(1, VT0));
4793 AddToWorklist(XORNode.getNode());
4795 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, XORNode);
4796 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, XORNode);
4798 // fold (select C, 0, X) -> (and (not C), X)
4799 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4800 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4801 AddToWorklist(NOTNode.getNode());
4802 return DAG.getNode(ISD::AND, SDLoc(N), VT, NOTNode, N2);
4804 // fold (select C, X, 1) -> (or (not C), X)
4805 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4806 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4807 AddToWorklist(NOTNode.getNode());
4808 return DAG.getNode(ISD::OR, SDLoc(N), VT, NOTNode, N1);
4810 // fold (select C, X, 0) -> (and C, X)
4811 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4812 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4813 // fold (select X, X, Y) -> (or X, Y)
4814 // fold (select X, 1, Y) -> (or X, Y)
4815 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4816 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4817 // fold (select X, Y, X) -> (and X, Y)
4818 // fold (select X, Y, 0) -> (and X, Y)
4819 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4820 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4822 // If we can fold this based on the true/false value, do so.
4823 if (SimplifySelectOps(N, N1, N2))
4824 return SDValue(N, 0); // Don't revisit N.
4826 // fold selects based on a setcc into other things, such as min/max/abs
4827 if (N0.getOpcode() == ISD::SETCC) {
4828 // select x, y (fcmp lt x, y) -> fminnum x, y
4829 // select x, y (fcmp gt x, y) -> fmaxnum x, y
4831 // This is OK if we don't care about what happens if either operand is a
4835 // FIXME: Instead of testing for UnsafeFPMath, this should be checking for
4836 // no signed zeros as well as no nans.
4837 const TargetOptions &Options = DAG.getTarget().Options;
4838 if (Options.UnsafeFPMath &&
4839 VT.isFloatingPoint() && N0.hasOneUse() &&
4840 DAG.isKnownNeverNaN(N1) && DAG.isKnownNeverNaN(N2)) {
4841 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
4844 combineMinNumMaxNum(SDLoc(N), VT, N0.getOperand(0), N0.getOperand(1),
4845 N1, N2, CC, TLI, DAG);
4850 if ((!LegalOperations &&
4851 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT)) ||
4852 TLI.isOperationLegal(ISD::SELECT_CC, VT))
4853 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT,
4854 N0.getOperand(0), N0.getOperand(1),
4855 N1, N2, N0.getOperand(2));
4856 return SimplifySelect(SDLoc(N), N0, N1, N2);
4859 if (VT0 == MVT::i1) {
4860 if (TLI.shouldNormalizeToSelectSequence(*DAG.getContext(), VT)) {
4861 // select (and Cond0, Cond1), X, Y
4862 // -> select Cond0, (select Cond1, X, Y), Y
4863 if (N0->getOpcode() == ISD::AND && N0->hasOneUse()) {
4864 SDValue Cond0 = N0->getOperand(0);
4865 SDValue Cond1 = N0->getOperand(1);
4866 SDValue InnerSelect = DAG.getNode(ISD::SELECT, SDLoc(N),
4867 N1.getValueType(), Cond1, N1, N2);
4868 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Cond0,
4871 // select (or Cond0, Cond1), X, Y -> select Cond0, X, (select Cond1, X, Y)
4872 if (N0->getOpcode() == ISD::OR && N0->hasOneUse()) {
4873 SDValue Cond0 = N0->getOperand(0);
4874 SDValue Cond1 = N0->getOperand(1);
4875 SDValue InnerSelect = DAG.getNode(ISD::SELECT, SDLoc(N),
4876 N1.getValueType(), Cond1, N1, N2);
4877 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Cond0, N1,
4882 // select Cond0, (select Cond1, X, Y), Y -> select (and Cond0, Cond1), X, Y
4883 if (N1->getOpcode() == ISD::SELECT) {
4884 SDValue N1_0 = N1->getOperand(0);
4885 SDValue N1_1 = N1->getOperand(1);
4886 SDValue N1_2 = N1->getOperand(2);
4888 // Create the actual and node if we can generate good code for it.
4889 if (!TLI.shouldNormalizeToSelectSequence(*DAG.getContext(), VT)) {
4890 SDValue And = DAG.getNode(ISD::AND, SDLoc(N), N0.getValueType(),
4892 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), And,
4895 // Otherwise see if we can optimize the "and" to a better pattern.
4896 if (SDValue Combined = visitANDLike(N0, N1_0, N))
4897 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Combined,
4901 // select Cond0, X, (select Cond1, X, Y) -> select (or Cond0, Cond1), X, Y
4902 if (N2->getOpcode() == ISD::SELECT) {
4903 SDValue N2_0 = N2->getOperand(0);
4904 SDValue N2_1 = N2->getOperand(1);
4905 SDValue N2_2 = N2->getOperand(2);
4907 // Create the actual or node if we can generate good code for it.
4908 if (!TLI.shouldNormalizeToSelectSequence(*DAG.getContext(), VT)) {
4909 SDValue Or = DAG.getNode(ISD::OR, SDLoc(N), N0.getValueType(),
4911 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Or,
4914 // Otherwise see if we can optimize to a better pattern.
4915 if (SDValue Combined = visitORLike(N0, N2_0, N))
4916 return DAG.getNode(ISD::SELECT, SDLoc(N), N1.getValueType(), Combined,
4926 std::pair<SDValue, SDValue> SplitVSETCC(const SDNode *N, SelectionDAG &DAG) {
4929 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
4931 // Split the inputs.
4932 SDValue Lo, Hi, LL, LH, RL, RH;
4933 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
4934 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
4936 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));
4937 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));
4939 return std::make_pair(Lo, Hi);
4942 // This function assumes all the vselect's arguments are CONCAT_VECTOR
4943 // nodes and that the condition is a BV of ConstantSDNodes (or undefs).
4944 static SDValue ConvertSelectToConcatVector(SDNode *N, SelectionDAG &DAG) {
4946 SDValue Cond = N->getOperand(0);
4947 SDValue LHS = N->getOperand(1);
4948 SDValue RHS = N->getOperand(2);
4949 EVT VT = N->getValueType(0);
4950 int NumElems = VT.getVectorNumElements();
4951 assert(LHS.getOpcode() == ISD::CONCAT_VECTORS &&
4952 RHS.getOpcode() == ISD::CONCAT_VECTORS &&
4953 Cond.getOpcode() == ISD::BUILD_VECTOR);
4955 // CONCAT_VECTOR can take an arbitrary number of arguments. We only care about
4956 // binary ones here.
4957 if (LHS->getNumOperands() != 2 || RHS->getNumOperands() != 2)
4960 // We're sure we have an even number of elements due to the
4961 // concat_vectors we have as arguments to vselect.
4962 // Skip BV elements until we find one that's not an UNDEF
4963 // After we find an UNDEF element, keep looping until we get to half the
4964 // length of the BV and see if all the non-undef nodes are the same.
4965 ConstantSDNode *BottomHalf = nullptr;
4966 for (int i = 0; i < NumElems / 2; ++i) {
4967 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4970 if (BottomHalf == nullptr)
4971 BottomHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4972 else if (Cond->getOperand(i).getNode() != BottomHalf)
4976 // Do the same for the second half of the BuildVector
4977 ConstantSDNode *TopHalf = nullptr;
4978 for (int i = NumElems / 2; i < NumElems; ++i) {
4979 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4982 if (TopHalf == nullptr)
4983 TopHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4984 else if (Cond->getOperand(i).getNode() != TopHalf)
4988 assert(TopHalf && BottomHalf &&
4989 "One half of the selector was all UNDEFs and the other was all the "
4990 "same value. This should have been addressed before this function.");
4992 ISD::CONCAT_VECTORS, dl, VT,
4993 BottomHalf->isNullValue() ? RHS->getOperand(0) : LHS->getOperand(0),
4994 TopHalf->isNullValue() ? RHS->getOperand(1) : LHS->getOperand(1));
4997 SDValue DAGCombiner::visitMSTORE(SDNode *N) {
4999 if (Level >= AfterLegalizeTypes)
5002 MaskedStoreSDNode *MST = dyn_cast<MaskedStoreSDNode>(N);
5003 SDValue Mask = MST->getMask();
5004 SDValue Data = MST->getValue();
5007 // If the MSTORE data type requires splitting and the mask is provided by a
5008 // SETCC, then split both nodes and its operands before legalization. This
5009 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5010 // and enables future optimizations (e.g. min/max pattern matching on X86).
5011 if (Mask.getOpcode() == ISD::SETCC) {
5013 // Check if any splitting is required.
5014 if (TLI.getTypeAction(*DAG.getContext(), Data.getValueType()) !=
5015 TargetLowering::TypeSplitVector)
5018 SDValue MaskLo, MaskHi, Lo, Hi;
5019 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
5022 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MST->getValueType(0));
5024 SDValue Chain = MST->getChain();
5025 SDValue Ptr = MST->getBasePtr();
5027 EVT MemoryVT = MST->getMemoryVT();
5028 unsigned Alignment = MST->getOriginalAlignment();
5030 // if Alignment is equal to the vector size,
5031 // take the half of it for the second part
5032 unsigned SecondHalfAlignment =
5033 (Alignment == Data->getValueType(0).getSizeInBits()/8) ?
5034 Alignment/2 : Alignment;
5036 EVT LoMemVT, HiMemVT;
5037 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
5039 SDValue DataLo, DataHi;
5040 std::tie(DataLo, DataHi) = DAG.SplitVector(Data, DL);
5042 MachineMemOperand *MMO = DAG.getMachineFunction().
5043 getMachineMemOperand(MST->getPointerInfo(),
5044 MachineMemOperand::MOStore, LoMemVT.getStoreSize(),
5045 Alignment, MST->getAAInfo(), MST->getRanges());
5047 Lo = DAG.getMaskedStore(Chain, DL, DataLo, Ptr, MaskLo, LoMemVT, MMO,
5048 MST->isTruncatingStore());
5050 unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
5051 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5052 DAG.getConstant(IncrementSize, Ptr.getValueType()));
5054 MMO = DAG.getMachineFunction().
5055 getMachineMemOperand(MST->getPointerInfo(),
5056 MachineMemOperand::MOStore, HiMemVT.getStoreSize(),
5057 SecondHalfAlignment, MST->getAAInfo(),
5060 Hi = DAG.getMaskedStore(Chain, DL, DataHi, Ptr, MaskHi, HiMemVT, MMO,
5061 MST->isTruncatingStore());
5063 AddToWorklist(Lo.getNode());
5064 AddToWorklist(Hi.getNode());
5066 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);
5071 SDValue DAGCombiner::visitMLOAD(SDNode *N) {
5073 if (Level >= AfterLegalizeTypes)
5076 MaskedLoadSDNode *MLD = dyn_cast<MaskedLoadSDNode>(N);
5077 SDValue Mask = MLD->getMask();
5080 // If the MLOAD result requires splitting and the mask is provided by a
5081 // SETCC, then split both nodes and its operands before legalization. This
5082 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5083 // and enables future optimizations (e.g. min/max pattern matching on X86).
5085 if (Mask.getOpcode() == ISD::SETCC) {
5086 EVT VT = N->getValueType(0);
5088 // Check if any splitting is required.
5089 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
5090 TargetLowering::TypeSplitVector)
5093 SDValue MaskLo, MaskHi, Lo, Hi;
5094 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
5096 SDValue Src0 = MLD->getSrc0();
5097 SDValue Src0Lo, Src0Hi;
5098 std::tie(Src0Lo, Src0Hi) = DAG.SplitVector(Src0, DL);
5101 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MLD->getValueType(0));
5103 SDValue Chain = MLD->getChain();
5104 SDValue Ptr = MLD->getBasePtr();
5105 EVT MemoryVT = MLD->getMemoryVT();
5106 unsigned Alignment = MLD->getOriginalAlignment();
5108 // if Alignment is equal to the vector size,
5109 // take the half of it for the second part
5110 unsigned SecondHalfAlignment =
5111 (Alignment == MLD->getValueType(0).getSizeInBits()/8) ?
5112 Alignment/2 : Alignment;
5114 EVT LoMemVT, HiMemVT;
5115 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
5117 MachineMemOperand *MMO = DAG.getMachineFunction().
5118 getMachineMemOperand(MLD->getPointerInfo(),
5119 MachineMemOperand::MOLoad, LoMemVT.getStoreSize(),
5120 Alignment, MLD->getAAInfo(), MLD->getRanges());
5122 Lo = DAG.getMaskedLoad(LoVT, DL, Chain, Ptr, MaskLo, Src0Lo, LoMemVT, MMO,
5125 unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
5126 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5127 DAG.getConstant(IncrementSize, Ptr.getValueType()));
5129 MMO = DAG.getMachineFunction().
5130 getMachineMemOperand(MLD->getPointerInfo(),
5131 MachineMemOperand::MOLoad, HiMemVT.getStoreSize(),
5132 SecondHalfAlignment, MLD->getAAInfo(), MLD->getRanges());
5134 Hi = DAG.getMaskedLoad(HiVT, DL, Chain, Ptr, MaskHi, Src0Hi, HiMemVT, MMO,
5137 AddToWorklist(Lo.getNode());
5138 AddToWorklist(Hi.getNode());
5140 // Build a factor node to remember that this load is independent of the
5142 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo.getValue(1),
5145 // Legalized the chain result - switch anything that used the old chain to
5147 DAG.ReplaceAllUsesOfValueWith(SDValue(MLD, 1), Chain);
5149 SDValue LoadRes = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
5151 SDValue RetOps[] = { LoadRes, Chain };
5152 return DAG.getMergeValues(RetOps, DL);
5157 SDValue DAGCombiner::visitVSELECT(SDNode *N) {
5158 SDValue N0 = N->getOperand(0);
5159 SDValue N1 = N->getOperand(1);
5160 SDValue N2 = N->getOperand(2);
5163 // Canonicalize integer abs.
5164 // vselect (setg[te] X, 0), X, -X ->
5165 // vselect (setgt X, -1), X, -X ->
5166 // vselect (setl[te] X, 0), -X, X ->
5167 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
5168 if (N0.getOpcode() == ISD::SETCC) {
5169 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
5170 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5172 bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
5174 if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
5175 (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) &&
5176 N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1))
5177 isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode());
5178 else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) &&
5179 N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1))
5180 isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode());
5183 EVT VT = LHS.getValueType();
5184 SDValue Shift = DAG.getNode(
5185 ISD::SRA, DL, VT, LHS,
5186 DAG.getConstant(VT.getScalarType().getSizeInBits() - 1, VT));
5187 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift);
5188 AddToWorklist(Shift.getNode());
5189 AddToWorklist(Add.getNode());
5190 return DAG.getNode(ISD::XOR, DL, VT, Add, Shift);
5194 // If the VSELECT result requires splitting and the mask is provided by a
5195 // SETCC, then split both nodes and its operands before legalization. This
5196 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5197 // and enables future optimizations (e.g. min/max pattern matching on X86).
5198 if (N0.getOpcode() == ISD::SETCC) {
5199 EVT VT = N->getValueType(0);
5201 // Check if any splitting is required.
5202 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
5203 TargetLowering::TypeSplitVector)
5206 SDValue Lo, Hi, CCLo, CCHi, LL, LH, RL, RH;
5207 std::tie(CCLo, CCHi) = SplitVSETCC(N0.getNode(), DAG);
5208 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 1);
5209 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 2);
5211 Lo = DAG.getNode(N->getOpcode(), DL, LL.getValueType(), CCLo, LL, RL);
5212 Hi = DAG.getNode(N->getOpcode(), DL, LH.getValueType(), CCHi, LH, RH);
5214 // Add the new VSELECT nodes to the work list in case they need to be split
5216 AddToWorklist(Lo.getNode());
5217 AddToWorklist(Hi.getNode());
5219 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
5222 // Fold (vselect (build_vector all_ones), N1, N2) -> N1
5223 if (ISD::isBuildVectorAllOnes(N0.getNode()))
5225 // Fold (vselect (build_vector all_zeros), N1, N2) -> N2
5226 if (ISD::isBuildVectorAllZeros(N0.getNode()))
5229 // The ConvertSelectToConcatVector function is assuming both the above
5230 // checks for (vselect (build_vector all{ones,zeros) ...) have been made
5232 if (N1.getOpcode() == ISD::CONCAT_VECTORS &&
5233 N2.getOpcode() == ISD::CONCAT_VECTORS &&
5234 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
5235 SDValue CV = ConvertSelectToConcatVector(N, DAG);
5243 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
5244 SDValue N0 = N->getOperand(0);
5245 SDValue N1 = N->getOperand(1);
5246 SDValue N2 = N->getOperand(2);
5247 SDValue N3 = N->getOperand(3);
5248 SDValue N4 = N->getOperand(4);
5249 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
5251 // fold select_cc lhs, rhs, x, x, cc -> x
5255 // Determine if the condition we're dealing with is constant
5256 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
5257 N0, N1, CC, SDLoc(N), false);
5258 if (SCC.getNode()) {
5259 AddToWorklist(SCC.getNode());
5261 if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) {
5262 if (!SCCC->isNullValue())
5263 return N2; // cond always true -> true val
5265 return N3; // cond always false -> false val
5266 } else if (SCC->getOpcode() == ISD::UNDEF) {
5267 // When the condition is UNDEF, just return the first operand. This is
5268 // coherent the DAG creation, no setcc node is created in this case
5270 } else if (SCC.getOpcode() == ISD::SETCC) {
5271 // Fold to a simpler select_cc
5272 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(),
5273 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
5278 // If we can fold this based on the true/false value, do so.
5279 if (SimplifySelectOps(N, N2, N3))
5280 return SDValue(N, 0); // Don't revisit N.
5282 // fold select_cc into other things, such as min/max/abs
5283 return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC);
5286 SDValue DAGCombiner::visitSETCC(SDNode *N) {
5287 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
5288 cast<CondCodeSDNode>(N->getOperand(2))->get(),
5292 // tryToFoldExtendOfConstant - Try to fold a sext/zext/aext
5293 // dag node into a ConstantSDNode or a build_vector of constants.
5294 // This function is called by the DAGCombiner when visiting sext/zext/aext
5295 // dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND).
5296 // Vector extends are not folded if operations are legal; this is to
5297 // avoid introducing illegal build_vector dag nodes.
5298 static SDNode *tryToFoldExtendOfConstant(SDNode *N, const TargetLowering &TLI,
5299 SelectionDAG &DAG, bool LegalTypes,
5300 bool LegalOperations) {
5301 unsigned Opcode = N->getOpcode();
5302 SDValue N0 = N->getOperand(0);
5303 EVT VT = N->getValueType(0);
5305 assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND ||
5306 Opcode == ISD::ANY_EXTEND) && "Expected EXTEND dag node in input!");
5308 // fold (sext c1) -> c1
5309 // fold (zext c1) -> c1
5310 // fold (aext c1) -> c1
5311 if (isa<ConstantSDNode>(N0))
5312 return DAG.getNode(Opcode, SDLoc(N), VT, N0).getNode();
5314 // fold (sext (build_vector AllConstants) -> (build_vector AllConstants)
5315 // fold (zext (build_vector AllConstants) -> (build_vector AllConstants)
5316 // fold (aext (build_vector AllConstants) -> (build_vector AllConstants)
5317 EVT SVT = VT.getScalarType();
5318 if (!(VT.isVector() &&
5319 (!LegalTypes || (!LegalOperations && TLI.isTypeLegal(SVT))) &&
5320 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())))
5323 // We can fold this node into a build_vector.
5324 unsigned VTBits = SVT.getSizeInBits();
5325 unsigned EVTBits = N0->getValueType(0).getScalarType().getSizeInBits();
5326 unsigned ShAmt = VTBits - EVTBits;
5327 SmallVector<SDValue, 8> Elts;
5328 unsigned NumElts = N0->getNumOperands();
5331 for (unsigned i=0; i != NumElts; ++i) {
5332 SDValue Op = N0->getOperand(i);
5333 if (Op->getOpcode() == ISD::UNDEF) {
5334 Elts.push_back(DAG.getUNDEF(SVT));
5338 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
5339 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
5340 if (Opcode == ISD::SIGN_EXTEND)
5341 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
5344 Elts.push_back(DAG.getConstant(C.shl(ShAmt).lshr(ShAmt).getZExtValue(),
5348 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Elts).getNode();
5351 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
5352 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
5353 // transformation. Returns true if extension are possible and the above
5354 // mentioned transformation is profitable.
5355 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
5357 SmallVectorImpl<SDNode *> &ExtendNodes,
5358 const TargetLowering &TLI) {
5359 bool HasCopyToRegUses = false;
5360 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
5361 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
5362 UE = N0.getNode()->use_end();
5367 if (UI.getUse().getResNo() != N0.getResNo())
5369 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
5370 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
5371 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
5372 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
5373 // Sign bits will be lost after a zext.
5376 for (unsigned i = 0; i != 2; ++i) {
5377 SDValue UseOp = User->getOperand(i);
5380 if (!isa<ConstantSDNode>(UseOp))
5385 ExtendNodes.push_back(User);
5388 // If truncates aren't free and there are users we can't
5389 // extend, it isn't worthwhile.
5392 // Remember if this value is live-out.
5393 if (User->getOpcode() == ISD::CopyToReg)
5394 HasCopyToRegUses = true;
5397 if (HasCopyToRegUses) {
5398 bool BothLiveOut = false;
5399 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
5401 SDUse &Use = UI.getUse();
5402 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
5408 // Both unextended and extended values are live out. There had better be
5409 // a good reason for the transformation.
5410 return ExtendNodes.size();
5415 void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
5416 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
5417 ISD::NodeType ExtType) {
5418 // Extend SetCC uses if necessary.
5419 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
5420 SDNode *SetCC = SetCCs[i];
5421 SmallVector<SDValue, 4> Ops;
5423 for (unsigned j = 0; j != 2; ++j) {
5424 SDValue SOp = SetCC->getOperand(j);
5426 Ops.push_back(ExtLoad);
5428 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
5431 Ops.push_back(SetCC->getOperand(2));
5432 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), Ops));
5436 // FIXME: Bring more similar combines here, common to sext/zext (maybe aext?).
5437 SDValue DAGCombiner::CombineExtLoad(SDNode *N) {
5438 SDValue N0 = N->getOperand(0);
5439 EVT DstVT = N->getValueType(0);
5440 EVT SrcVT = N0.getValueType();
5442 assert((N->getOpcode() == ISD::SIGN_EXTEND ||
5443 N->getOpcode() == ISD::ZERO_EXTEND) &&
5444 "Unexpected node type (not an extend)!");
5446 // fold (sext (load x)) to multiple smaller sextloads; same for zext.
5447 // For example, on a target with legal v4i32, but illegal v8i32, turn:
5448 // (v8i32 (sext (v8i16 (load x))))
5450 // (v8i32 (concat_vectors (v4i32 (sextload x)),
5451 // (v4i32 (sextload (x + 16)))))
5452 // Where uses of the original load, i.e.:
5454 // are replaced with:
5456 // (v8i32 (concat_vectors (v4i32 (sextload x)),
5457 // (v4i32 (sextload (x + 16)))))))
5459 // This combine is only applicable to illegal, but splittable, vectors.
5460 // All legal types, and illegal non-vector types, are handled elsewhere.
5461 // This combine is controlled by TargetLowering::isVectorLoadExtDesirable.
5463 if (N0->getOpcode() != ISD::LOAD)
5466 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5468 if (!ISD::isNON_EXTLoad(LN0) || !ISD::isUNINDEXEDLoad(LN0) ||
5469 !N0.hasOneUse() || LN0->isVolatile() || !DstVT.isVector() ||
5470 !DstVT.isPow2VectorType() || !TLI.isVectorLoadExtDesirable(SDValue(N, 0)))
5473 SmallVector<SDNode *, 4> SetCCs;
5474 if (!ExtendUsesToFormExtLoad(N, N0, N->getOpcode(), SetCCs, TLI))
5477 ISD::LoadExtType ExtType =
5478 N->getOpcode() == ISD::SIGN_EXTEND ? ISD::SEXTLOAD : ISD::ZEXTLOAD;
5480 // Try to split the vector types to get down to legal types.
5481 EVT SplitSrcVT = SrcVT;
5482 EVT SplitDstVT = DstVT;
5483 while (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT) &&
5484 SplitSrcVT.getVectorNumElements() > 1) {
5485 SplitDstVT = DAG.GetSplitDestVTs(SplitDstVT).first;
5486 SplitSrcVT = DAG.GetSplitDestVTs(SplitSrcVT).first;
5489 if (!TLI.isLoadExtLegalOrCustom(ExtType, SplitDstVT, SplitSrcVT))
5493 const unsigned NumSplits =
5494 DstVT.getVectorNumElements() / SplitDstVT.getVectorNumElements();
5495 const unsigned Stride = SplitSrcVT.getStoreSize();
5496 SmallVector<SDValue, 4> Loads;
5497 SmallVector<SDValue, 4> Chains;
5499 SDValue BasePtr = LN0->getBasePtr();
5500 for (unsigned Idx = 0; Idx < NumSplits; Idx++) {
5501 const unsigned Offset = Idx * Stride;
5502 const unsigned Align = MinAlign(LN0->getAlignment(), Offset);
5504 SDValue SplitLoad = DAG.getExtLoad(
5505 ExtType, DL, SplitDstVT, LN0->getChain(), BasePtr,
5506 LN0->getPointerInfo().getWithOffset(Offset), SplitSrcVT,
5507 LN0->isVolatile(), LN0->isNonTemporal(), LN0->isInvariant(),
5508 Align, LN0->getAAInfo());
5510 BasePtr = DAG.getNode(ISD::ADD, DL, BasePtr.getValueType(), BasePtr,
5511 DAG.getConstant(Stride, BasePtr.getValueType()));
5513 Loads.push_back(SplitLoad.getValue(0));
5514 Chains.push_back(SplitLoad.getValue(1));
5517 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
5518 SDValue NewValue = DAG.getNode(ISD::CONCAT_VECTORS, DL, DstVT, Loads);
5520 CombineTo(N, NewValue);
5522 // Replace uses of the original load (before extension)
5523 // with a truncate of the concatenated sextloaded vectors.
5525 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(), NewValue);
5526 CombineTo(N0.getNode(), Trunc, NewChain);
5527 ExtendSetCCUses(SetCCs, Trunc, NewValue, DL,
5528 (ISD::NodeType)N->getOpcode());
5529 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5532 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
5533 SDValue N0 = N->getOperand(0);
5534 EVT VT = N->getValueType(0);
5536 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5538 return SDValue(Res, 0);
5540 // fold (sext (sext x)) -> (sext x)
5541 // fold (sext (aext x)) -> (sext x)
5542 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5543 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT,
5546 if (N0.getOpcode() == ISD::TRUNCATE) {
5547 // fold (sext (truncate (load x))) -> (sext (smaller load x))
5548 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
5549 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5550 if (NarrowLoad.getNode()) {
5551 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5552 if (NarrowLoad.getNode() != N0.getNode()) {
5553 CombineTo(N0.getNode(), NarrowLoad);
5554 // CombineTo deleted the truncate, if needed, but not what's under it.
5557 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5560 // See if the value being truncated is already sign extended. If so, just
5561 // eliminate the trunc/sext pair.
5562 SDValue Op = N0.getOperand(0);
5563 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
5564 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
5565 unsigned DestBits = VT.getScalarType().getSizeInBits();
5566 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
5568 if (OpBits == DestBits) {
5569 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
5570 // bits, it is already ready.
5571 if (NumSignBits > DestBits-MidBits)
5573 } else if (OpBits < DestBits) {
5574 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
5575 // bits, just sext from i32.
5576 if (NumSignBits > OpBits-MidBits)
5577 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, Op);
5579 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
5580 // bits, just truncate to i32.
5581 if (NumSignBits > OpBits-MidBits)
5582 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5585 // fold (sext (truncate x)) -> (sextinreg x).
5586 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
5587 N0.getValueType())) {
5588 if (OpBits < DestBits)
5589 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op);
5590 else if (OpBits > DestBits)
5591 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op);
5592 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, Op,
5593 DAG.getValueType(N0.getValueType()));
5597 // fold (sext (load x)) -> (sext (truncate (sextload x)))
5598 // Only generate vector extloads when 1) they're legal, and 2) they are
5599 // deemed desirable by the target.
5600 if (ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5601 ((!LegalOperations && !VT.isVector() &&
5602 !cast<LoadSDNode>(N0)->isVolatile()) ||
5603 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, N0.getValueType()))) {
5604 bool DoXform = true;
5605 SmallVector<SDNode*, 4> SetCCs;
5606 if (!N0.hasOneUse())
5607 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
5609 DoXform &= TLI.isVectorLoadExtDesirable(SDValue(N, 0));
5611 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5612 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5614 LN0->getBasePtr(), N0.getValueType(),
5615 LN0->getMemOperand());
5616 CombineTo(N, ExtLoad);
5617 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5618 N0.getValueType(), ExtLoad);
5619 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5620 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5622 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5626 // fold (sext (load x)) to multiple smaller sextloads.
5627 // Only on illegal but splittable vectors.
5628 if (SDValue ExtLoad = CombineExtLoad(N))
5631 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
5632 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
5633 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5634 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5635 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5636 EVT MemVT = LN0->getMemoryVT();
5637 if ((!LegalOperations && !LN0->isVolatile()) ||
5638 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, MemVT)) {
5639 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5641 LN0->getBasePtr(), MemVT,
5642 LN0->getMemOperand());
5643 CombineTo(N, ExtLoad);
5644 CombineTo(N0.getNode(),
5645 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5646 N0.getValueType(), ExtLoad),
5647 ExtLoad.getValue(1));
5648 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5652 // fold (sext (and/or/xor (load x), cst)) ->
5653 // (and/or/xor (sextload x), (sext cst))
5654 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5655 N0.getOpcode() == ISD::XOR) &&
5656 isa<LoadSDNode>(N0.getOperand(0)) &&
5657 N0.getOperand(1).getOpcode() == ISD::Constant &&
5658 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, N0.getValueType()) &&
5659 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5660 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5661 if (LN0->getExtensionType() != ISD::ZEXTLOAD && LN0->isUnindexed()) {
5662 bool DoXform = true;
5663 SmallVector<SDNode*, 4> SetCCs;
5664 if (!N0.hasOneUse())
5665 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
5668 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN0), VT,
5669 LN0->getChain(), LN0->getBasePtr(),
5671 LN0->getMemOperand());
5672 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5673 Mask = Mask.sext(VT.getSizeInBits());
5674 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5675 ExtLoad, DAG.getConstant(Mask, VT));
5676 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5677 SDLoc(N0.getOperand(0)),
5678 N0.getOperand(0).getValueType(), ExtLoad);
5680 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5681 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5683 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5688 if (N0.getOpcode() == ISD::SETCC) {
5689 EVT N0VT = N0.getOperand(0).getValueType();
5690 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
5691 // Only do this before legalize for now.
5692 if (VT.isVector() && !LegalOperations &&
5693 TLI.getBooleanContents(N0VT) ==
5694 TargetLowering::ZeroOrNegativeOneBooleanContent) {
5695 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
5696 // of the same size as the compared operands. Only optimize sext(setcc())
5697 // if this is the case.
5698 EVT SVT = getSetCCResultType(N0VT);
5700 // We know that the # elements of the results is the same as the
5701 // # elements of the compare (and the # elements of the compare result
5702 // for that matter). Check to see that they are the same size. If so,
5703 // we know that the element size of the sext'd result matches the
5704 // element size of the compare operands.
5705 if (VT.getSizeInBits() == SVT.getSizeInBits())
5706 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5708 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5710 // If the desired elements are smaller or larger than the source
5711 // elements we can use a matching integer vector type and then
5712 // truncate/sign extend
5713 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5714 if (SVT == MatchingVectorType) {
5715 SDValue VsetCC = DAG.getSetCC(SDLoc(N), MatchingVectorType,
5716 N0.getOperand(0), N0.getOperand(1),
5717 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5718 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
5722 // sext(setcc x, y, cc) -> (select (setcc x, y, cc), -1, 0)
5723 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
5725 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
5727 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5728 NegOne, DAG.getConstant(0, VT),
5729 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5730 if (SCC.getNode()) return SCC;
5732 if (!VT.isVector()) {
5733 EVT SetCCVT = getSetCCResultType(N0.getOperand(0).getValueType());
5734 if (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, SetCCVT)) {
5736 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5737 SDValue SetCC = DAG.getSetCC(DL, SetCCVT,
5738 N0.getOperand(0), N0.getOperand(1), CC);
5739 return DAG.getSelect(DL, VT, SetCC,
5740 NegOne, DAG.getConstant(0, VT));
5745 // fold (sext x) -> (zext x) if the sign bit is known zero.
5746 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
5747 DAG.SignBitIsZero(N0))
5748 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0);
5753 // isTruncateOf - If N is a truncate of some other value, return true, record
5754 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
5755 // This function computes KnownZero to avoid a duplicated call to
5756 // computeKnownBits in the caller.
5757 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
5760 if (N->getOpcode() == ISD::TRUNCATE) {
5761 Op = N->getOperand(0);
5762 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5766 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
5767 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
5770 SDValue Op0 = N->getOperand(0);
5771 SDValue Op1 = N->getOperand(1);
5772 assert(Op0.getValueType() == Op1.getValueType());
5774 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
5775 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
5776 if (COp0 && COp0->isNullValue())
5778 else if (COp1 && COp1->isNullValue())
5783 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5785 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
5791 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
5792 SDValue N0 = N->getOperand(0);
5793 EVT VT = N->getValueType(0);
5795 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5797 return SDValue(Res, 0);
5799 // fold (zext (zext x)) -> (zext x)
5800 // fold (zext (aext x)) -> (zext x)
5801 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5802 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT,
5805 // fold (zext (truncate x)) -> (zext x) or
5806 // (zext (truncate x)) -> (truncate x)
5807 // This is valid when the truncated bits of x are already zero.
5808 // FIXME: We should extend this to work for vectors too.
5811 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
5812 APInt TruncatedBits =
5813 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
5814 APInt(Op.getValueSizeInBits(), 0) :
5815 APInt::getBitsSet(Op.getValueSizeInBits(),
5816 N0.getValueSizeInBits(),
5817 std::min(Op.getValueSizeInBits(),
5818 VT.getSizeInBits()));
5819 if (TruncatedBits == (KnownZero & TruncatedBits)) {
5820 if (VT.bitsGT(Op.getValueType()))
5821 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, Op);
5822 if (VT.bitsLT(Op.getValueType()))
5823 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5829 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5830 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
5831 if (N0.getOpcode() == ISD::TRUNCATE) {
5832 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5833 if (NarrowLoad.getNode()) {
5834 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5835 if (NarrowLoad.getNode() != N0.getNode()) {
5836 CombineTo(N0.getNode(), NarrowLoad);
5837 // CombineTo deleted the truncate, if needed, but not what's under it.
5840 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5844 // fold (zext (truncate x)) -> (and x, mask)
5845 if (N0.getOpcode() == ISD::TRUNCATE &&
5846 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
5848 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5849 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
5850 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5851 if (NarrowLoad.getNode()) {
5852 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5853 if (NarrowLoad.getNode() != N0.getNode()) {
5854 CombineTo(N0.getNode(), NarrowLoad);
5855 // CombineTo deleted the truncate, if needed, but not what's under it.
5858 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5861 SDValue Op = N0.getOperand(0);
5862 if (Op.getValueType().bitsLT(VT)) {
5863 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, Op);
5864 AddToWorklist(Op.getNode());
5865 } else if (Op.getValueType().bitsGT(VT)) {
5866 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5867 AddToWorklist(Op.getNode());
5869 return DAG.getZeroExtendInReg(Op, SDLoc(N),
5870 N0.getValueType().getScalarType());
5873 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
5874 // if either of the casts is not free.
5875 if (N0.getOpcode() == ISD::AND &&
5876 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5877 N0.getOperand(1).getOpcode() == ISD::Constant &&
5878 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5879 N0.getValueType()) ||
5880 !TLI.isZExtFree(N0.getValueType(), VT))) {
5881 SDValue X = N0.getOperand(0).getOperand(0);
5882 if (X.getValueType().bitsLT(VT)) {
5883 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(X), VT, X);
5884 } else if (X.getValueType().bitsGT(VT)) {
5885 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
5887 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5888 Mask = Mask.zext(VT.getSizeInBits());
5889 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5890 X, DAG.getConstant(Mask, VT));
5893 // fold (zext (load x)) -> (zext (truncate (zextload x)))
5894 // Only generate vector extloads when 1) they're legal, and 2) they are
5895 // deemed desirable by the target.
5896 if (ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5897 ((!LegalOperations && !VT.isVector() &&
5898 !cast<LoadSDNode>(N0)->isVolatile()) ||
5899 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, N0.getValueType()))) {
5900 bool DoXform = true;
5901 SmallVector<SDNode*, 4> SetCCs;
5902 if (!N0.hasOneUse())
5903 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
5905 DoXform &= TLI.isVectorLoadExtDesirable(SDValue(N, 0));
5907 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5908 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5910 LN0->getBasePtr(), N0.getValueType(),
5911 LN0->getMemOperand());
5912 CombineTo(N, ExtLoad);
5913 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5914 N0.getValueType(), ExtLoad);
5915 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5917 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5919 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5923 // fold (zext (load x)) to multiple smaller zextloads.
5924 // Only on illegal but splittable vectors.
5925 if (SDValue ExtLoad = CombineExtLoad(N))
5928 // fold (zext (and/or/xor (load x), cst)) ->
5929 // (and/or/xor (zextload x), (zext cst))
5930 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5931 N0.getOpcode() == ISD::XOR) &&
5932 isa<LoadSDNode>(N0.getOperand(0)) &&
5933 N0.getOperand(1).getOpcode() == ISD::Constant &&
5934 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, N0.getValueType()) &&
5935 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5936 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5937 if (LN0->getExtensionType() != ISD::SEXTLOAD && LN0->isUnindexed()) {
5938 bool DoXform = true;
5939 SmallVector<SDNode*, 4> SetCCs;
5940 if (!N0.hasOneUse())
5941 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
5944 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), VT,
5945 LN0->getChain(), LN0->getBasePtr(),
5947 LN0->getMemOperand());
5948 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5949 Mask = Mask.zext(VT.getSizeInBits());
5950 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5951 ExtLoad, DAG.getConstant(Mask, VT));
5952 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5953 SDLoc(N0.getOperand(0)),
5954 N0.getOperand(0).getValueType(), ExtLoad);
5956 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5957 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5959 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5964 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
5965 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
5966 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5967 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5968 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5969 EVT MemVT = LN0->getMemoryVT();
5970 if ((!LegalOperations && !LN0->isVolatile()) ||
5971 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT)) {
5972 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5974 LN0->getBasePtr(), MemVT,
5975 LN0->getMemOperand());
5976 CombineTo(N, ExtLoad);
5977 CombineTo(N0.getNode(),
5978 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(),
5980 ExtLoad.getValue(1));
5981 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5985 if (N0.getOpcode() == ISD::SETCC) {
5986 if (!LegalOperations && VT.isVector() &&
5987 N0.getValueType().getVectorElementType() == MVT::i1) {
5988 EVT N0VT = N0.getOperand(0).getValueType();
5989 if (getSetCCResultType(N0VT) == N0.getValueType())
5992 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
5993 // Only do this before legalize for now.
5994 EVT EltVT = VT.getVectorElementType();
5995 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
5996 DAG.getConstant(1, EltVT));
5997 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5998 // We know that the # elements of the results is the same as the
5999 // # elements of the compare (and the # elements of the compare result
6000 // for that matter). Check to see that they are the same size. If so,
6001 // we know that the element size of the sext'd result matches the
6002 // element size of the compare operands.
6003 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6004 DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
6006 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
6007 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT,
6010 // If the desired elements are smaller or larger than the source
6011 // elements we can use a matching integer vector type and then
6012 // truncate/sign extend
6013 EVT MatchingElementType =
6014 EVT::getIntegerVT(*DAG.getContext(),
6015 N0VT.getScalarType().getSizeInBits());
6016 EVT MatchingVectorType =
6017 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
6018 N0VT.getVectorNumElements());
6020 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
6022 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6023 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6024 DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT),
6025 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, OneOps));
6028 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
6030 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
6031 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
6032 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
6033 if (SCC.getNode()) return SCC;
6036 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
6037 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
6038 isa<ConstantSDNode>(N0.getOperand(1)) &&
6039 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
6041 SDValue ShAmt = N0.getOperand(1);
6042 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
6043 if (N0.getOpcode() == ISD::SHL) {
6044 SDValue InnerZExt = N0.getOperand(0);
6045 // If the original shl may be shifting out bits, do not perform this
6047 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
6048 InnerZExt.getOperand(0).getValueType().getSizeInBits();
6049 if (ShAmtVal > KnownZeroBits)
6055 // Ensure that the shift amount is wide enough for the shifted value.
6056 if (VT.getSizeInBits() >= 256)
6057 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
6059 return DAG.getNode(N0.getOpcode(), DL, VT,
6060 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
6067 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
6068 SDValue N0 = N->getOperand(0);
6069 EVT VT = N->getValueType(0);
6071 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
6073 return SDValue(Res, 0);
6075 // fold (aext (aext x)) -> (aext x)
6076 // fold (aext (zext x)) -> (zext x)
6077 // fold (aext (sext x)) -> (sext x)
6078 if (N0.getOpcode() == ISD::ANY_EXTEND ||
6079 N0.getOpcode() == ISD::ZERO_EXTEND ||
6080 N0.getOpcode() == ISD::SIGN_EXTEND)
6081 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0));
6083 // fold (aext (truncate (load x))) -> (aext (smaller load x))
6084 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
6085 if (N0.getOpcode() == ISD::TRUNCATE) {
6086 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
6087 if (NarrowLoad.getNode()) {
6088 SDNode* oye = N0.getNode()->getOperand(0).getNode();
6089 if (NarrowLoad.getNode() != N0.getNode()) {
6090 CombineTo(N0.getNode(), NarrowLoad);
6091 // CombineTo deleted the truncate, if needed, but not what's under it.
6094 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6098 // fold (aext (truncate x))
6099 if (N0.getOpcode() == ISD::TRUNCATE) {
6100 SDValue TruncOp = N0.getOperand(0);
6101 if (TruncOp.getValueType() == VT)
6102 return TruncOp; // x iff x size == zext size.
6103 if (TruncOp.getValueType().bitsGT(VT))
6104 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, TruncOp);
6105 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, TruncOp);
6108 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
6109 // if the trunc is not free.
6110 if (N0.getOpcode() == ISD::AND &&
6111 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
6112 N0.getOperand(1).getOpcode() == ISD::Constant &&
6113 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
6114 N0.getValueType())) {
6115 SDValue X = N0.getOperand(0).getOperand(0);
6116 if (X.getValueType().bitsLT(VT)) {
6117 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, X);
6118 } else if (X.getValueType().bitsGT(VT)) {
6119 X = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, X);
6121 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
6122 Mask = Mask.zext(VT.getSizeInBits());
6123 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6124 X, DAG.getConstant(Mask, VT));
6127 // fold (aext (load x)) -> (aext (truncate (extload x)))
6128 // None of the supported targets knows how to perform load and any_ext
6129 // on vectors in one instruction. We only perform this transformation on
6131 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
6132 ISD::isUNINDEXEDLoad(N0.getNode()) &&
6133 TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) {
6134 bool DoXform = true;
6135 SmallVector<SDNode*, 4> SetCCs;
6136 if (!N0.hasOneUse())
6137 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
6139 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6140 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
6142 LN0->getBasePtr(), N0.getValueType(),
6143 LN0->getMemOperand());
6144 CombineTo(N, ExtLoad);
6145 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
6146 N0.getValueType(), ExtLoad);
6147 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
6148 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
6150 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6154 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
6155 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
6156 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
6157 if (N0.getOpcode() == ISD::LOAD &&
6158 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
6160 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6161 ISD::LoadExtType ExtType = LN0->getExtensionType();
6162 EVT MemVT = LN0->getMemoryVT();
6163 if (!LegalOperations || TLI.isLoadExtLegal(ExtType, VT, MemVT)) {
6164 SDValue ExtLoad = DAG.getExtLoad(ExtType, SDLoc(N),
6165 VT, LN0->getChain(), LN0->getBasePtr(),
6166 MemVT, LN0->getMemOperand());
6167 CombineTo(N, ExtLoad);
6168 CombineTo(N0.getNode(),
6169 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
6170 N0.getValueType(), ExtLoad),
6171 ExtLoad.getValue(1));
6172 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6176 if (N0.getOpcode() == ISD::SETCC) {
6178 // aext(setcc) -> vsetcc
6179 // aext(setcc) -> truncate(vsetcc)
6180 // aext(setcc) -> aext(vsetcc)
6181 // Only do this before legalize for now.
6182 if (VT.isVector() && !LegalOperations) {
6183 EVT N0VT = N0.getOperand(0).getValueType();
6184 // We know that the # elements of the results is the same as the
6185 // # elements of the compare (and the # elements of the compare result
6186 // for that matter). Check to see that they are the same size. If so,
6187 // we know that the element size of the sext'd result matches the
6188 // element size of the compare operands.
6189 if (VT.getSizeInBits() == N0VT.getSizeInBits())
6190 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
6192 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6193 // If the desired elements are smaller or larger than the source
6194 // elements we can use a matching integer vector type and then
6195 // truncate/any extend
6197 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
6199 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
6201 cast<CondCodeSDNode>(N0.getOperand(2))->get());
6202 return DAG.getAnyExtOrTrunc(VsetCC, SDLoc(N), VT);
6206 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
6208 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
6209 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
6210 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
6218 /// See if the specified operand can be simplified with the knowledge that only
6219 /// the bits specified by Mask are used. If so, return the simpler operand,
6220 /// otherwise return a null SDValue.
6221 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
6222 switch (V.getOpcode()) {
6224 case ISD::Constant: {
6225 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
6226 assert(CV && "Const value should be ConstSDNode.");
6227 const APInt &CVal = CV->getAPIntValue();
6228 APInt NewVal = CVal & Mask;
6230 return DAG.getConstant(NewVal, V.getValueType());
6235 // If the LHS or RHS don't contribute bits to the or, drop them.
6236 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
6237 return V.getOperand(1);
6238 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
6239 return V.getOperand(0);
6242 // Only look at single-use SRLs.
6243 if (!V.getNode()->hasOneUse())
6245 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
6246 // See if we can recursively simplify the LHS.
6247 unsigned Amt = RHSC->getZExtValue();
6249 // Watch out for shift count overflow though.
6250 if (Amt >= Mask.getBitWidth()) break;
6251 APInt NewMask = Mask << Amt;
6252 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
6253 if (SimplifyLHS.getNode())
6254 return DAG.getNode(ISD::SRL, SDLoc(V), V.getValueType(),
6255 SimplifyLHS, V.getOperand(1));
6261 /// If the result of a wider load is shifted to right of N bits and then
6262 /// truncated to a narrower type and where N is a multiple of number of bits of
6263 /// the narrower type, transform it to a narrower load from address + N / num of
6264 /// bits of new type. If the result is to be extended, also fold the extension
6265 /// to form a extending load.
6266 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
6267 unsigned Opc = N->getOpcode();
6269 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
6270 SDValue N0 = N->getOperand(0);
6271 EVT VT = N->getValueType(0);
6274 // This transformation isn't valid for vector loads.
6278 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
6280 if (Opc == ISD::SIGN_EXTEND_INREG) {
6281 ExtType = ISD::SEXTLOAD;
6282 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
6283 } else if (Opc == ISD::SRL) {
6284 // Another special-case: SRL is basically zero-extending a narrower value.
6285 ExtType = ISD::ZEXTLOAD;
6287 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
6288 if (!N01) return SDValue();
6289 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
6290 VT.getSizeInBits() - N01->getZExtValue());
6292 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, VT, ExtVT))
6295 unsigned EVTBits = ExtVT.getSizeInBits();
6297 // Do not generate loads of non-round integer types since these can
6298 // be expensive (and would be wrong if the type is not byte sized).
6299 if (!ExtVT.isRound())
6303 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
6304 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
6305 ShAmt = N01->getZExtValue();
6306 // Is the shift amount a multiple of size of VT?
6307 if ((ShAmt & (EVTBits-1)) == 0) {
6308 N0 = N0.getOperand(0);
6309 // Is the load width a multiple of size of VT?
6310 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
6314 // At this point, we must have a load or else we can't do the transform.
6315 if (!isa<LoadSDNode>(N0)) return SDValue();
6317 // Because a SRL must be assumed to *need* to zero-extend the high bits
6318 // (as opposed to anyext the high bits), we can't combine the zextload
6319 // lowering of SRL and an sextload.
6320 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
6323 // If the shift amount is larger than the input type then we're not
6324 // accessing any of the loaded bytes. If the load was a zextload/extload
6325 // then the result of the shift+trunc is zero/undef (handled elsewhere).
6326 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
6331 // If the load is shifted left (and the result isn't shifted back right),
6332 // we can fold the truncate through the shift.
6333 unsigned ShLeftAmt = 0;
6334 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
6335 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
6336 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
6337 ShLeftAmt = N01->getZExtValue();
6338 N0 = N0.getOperand(0);
6342 // If we haven't found a load, we can't narrow it. Don't transform one with
6343 // multiple uses, this would require adding a new load.
6344 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse())
6347 // Don't change the width of a volatile load.
6348 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6349 if (LN0->isVolatile())
6352 // Verify that we are actually reducing a load width here.
6353 if (LN0->getMemoryVT().getSizeInBits() < EVTBits)
6356 // For the transform to be legal, the load must produce only two values
6357 // (the value loaded and the chain). Don't transform a pre-increment
6358 // load, for example, which produces an extra value. Otherwise the
6359 // transformation is not equivalent, and the downstream logic to replace
6360 // uses gets things wrong.
6361 if (LN0->getNumValues() > 2)
6364 // If the load that we're shrinking is an extload and we're not just
6365 // discarding the extension we can't simply shrink the load. Bail.
6366 // TODO: It would be possible to merge the extensions in some cases.
6367 if (LN0->getExtensionType() != ISD::NON_EXTLOAD &&
6368 LN0->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits() + ShAmt)
6371 if (!TLI.shouldReduceLoadWidth(LN0, ExtType, ExtVT))
6374 EVT PtrType = N0.getOperand(1).getValueType();
6376 if (PtrType == MVT::Untyped || PtrType.isExtended())
6377 // It's not possible to generate a constant of extended or untyped type.
6380 // For big endian targets, we need to adjust the offset to the pointer to
6381 // load the correct bytes.
6382 if (TLI.isBigEndian()) {
6383 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
6384 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
6385 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
6388 uint64_t PtrOff = ShAmt / 8;
6389 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
6390 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0),
6391 PtrType, LN0->getBasePtr(),
6392 DAG.getConstant(PtrOff, PtrType));
6393 AddToWorklist(NewPtr.getNode());
6396 if (ExtType == ISD::NON_EXTLOAD)
6397 Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr,
6398 LN0->getPointerInfo().getWithOffset(PtrOff),
6399 LN0->isVolatile(), LN0->isNonTemporal(),
6400 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
6402 Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(),NewPtr,
6403 LN0->getPointerInfo().getWithOffset(PtrOff),
6404 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
6405 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
6407 // Replace the old load's chain with the new load's chain.
6408 WorklistRemover DeadNodes(*this);
6409 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
6411 // Shift the result left, if we've swallowed a left shift.
6412 SDValue Result = Load;
6413 if (ShLeftAmt != 0) {
6414 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
6415 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
6417 // If the shift amount is as large as the result size (but, presumably,
6418 // no larger than the source) then the useful bits of the result are
6419 // zero; we can't simply return the shortened shift, because the result
6420 // of that operation is undefined.
6421 if (ShLeftAmt >= VT.getSizeInBits())
6422 Result = DAG.getConstant(0, VT);
6424 Result = DAG.getNode(ISD::SHL, SDLoc(N0), VT,
6425 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
6428 // Return the new loaded value.
6432 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
6433 SDValue N0 = N->getOperand(0);
6434 SDValue N1 = N->getOperand(1);
6435 EVT VT = N->getValueType(0);
6436 EVT EVT = cast<VTSDNode>(N1)->getVT();
6437 unsigned VTBits = VT.getScalarType().getSizeInBits();
6438 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
6440 // fold (sext_in_reg c1) -> c1
6441 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
6442 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1);
6444 // If the input is already sign extended, just drop the extension.
6445 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
6448 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
6449 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
6450 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT()))
6451 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
6452 N0.getOperand(0), N1);
6454 // fold (sext_in_reg (sext x)) -> (sext x)
6455 // fold (sext_in_reg (aext x)) -> (sext x)
6456 // if x is small enough.
6457 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
6458 SDValue N00 = N0.getOperand(0);
6459 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
6460 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
6461 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
6464 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
6465 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
6466 return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT);
6468 // fold operands of sext_in_reg based on knowledge that the top bits are not
6470 if (SimplifyDemandedBits(SDValue(N, 0)))
6471 return SDValue(N, 0);
6473 // fold (sext_in_reg (load x)) -> (smaller sextload x)
6474 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
6475 SDValue NarrowLoad = ReduceLoadWidth(N);
6476 if (NarrowLoad.getNode())
6479 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
6480 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
6481 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
6482 if (N0.getOpcode() == ISD::SRL) {
6483 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
6484 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
6485 // We can turn this into an SRA iff the input to the SRL is already sign
6487 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
6488 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
6489 return DAG.getNode(ISD::SRA, SDLoc(N), VT,
6490 N0.getOperand(0), N0.getOperand(1));
6494 // fold (sext_inreg (extload x)) -> (sextload x)
6495 if (ISD::isEXTLoad(N0.getNode()) &&
6496 ISD::isUNINDEXEDLoad(N0.getNode()) &&
6497 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
6498 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6499 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) {
6500 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6501 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
6503 LN0->getBasePtr(), EVT,
6504 LN0->getMemOperand());
6505 CombineTo(N, ExtLoad);
6506 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
6507 AddToWorklist(ExtLoad.getNode());
6508 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6510 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
6511 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
6513 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
6514 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6515 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) {
6516 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6517 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
6519 LN0->getBasePtr(), EVT,
6520 LN0->getMemOperand());
6521 CombineTo(N, ExtLoad);
6522 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
6523 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6526 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
6527 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
6528 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
6529 N0.getOperand(1), false);
6530 if (BSwap.getNode())
6531 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
6535 // Fold a sext_inreg of a build_vector of ConstantSDNodes or undefs
6536 // into a build_vector.
6537 if (ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
6538 SmallVector<SDValue, 8> Elts;
6539 unsigned NumElts = N0->getNumOperands();
6540 unsigned ShAmt = VTBits - EVTBits;
6542 for (unsigned i = 0; i != NumElts; ++i) {
6543 SDValue Op = N0->getOperand(i);
6544 if (Op->getOpcode() == ISD::UNDEF) {
6549 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
6550 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
6551 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
6552 Op.getValueType()));
6555 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Elts);
6561 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
6562 SDValue N0 = N->getOperand(0);
6563 EVT VT = N->getValueType(0);
6564 bool isLE = TLI.isLittleEndian();
6567 if (N0.getValueType() == N->getValueType(0))
6569 // fold (truncate c1) -> c1
6570 if (isConstantIntBuildVectorOrConstantInt(N0))
6571 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0);
6572 // fold (truncate (truncate x)) -> (truncate x)
6573 if (N0.getOpcode() == ISD::TRUNCATE)
6574 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6575 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
6576 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
6577 N0.getOpcode() == ISD::SIGN_EXTEND ||
6578 N0.getOpcode() == ISD::ANY_EXTEND) {
6579 if (N0.getOperand(0).getValueType().bitsLT(VT))
6580 // if the source is smaller than the dest, we still need an extend
6581 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
6583 if (N0.getOperand(0).getValueType().bitsGT(VT))
6584 // if the source is larger than the dest, than we just need the truncate
6585 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6586 // if the source and dest are the same type, we can drop both the extend
6587 // and the truncate.
6588 return N0.getOperand(0);
6591 // Fold extract-and-trunc into a narrow extract. For example:
6592 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
6593 // i32 y = TRUNCATE(i64 x)
6595 // v16i8 b = BITCAST (v2i64 val)
6596 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
6598 // Note: We only run this optimization after type legalization (which often
6599 // creates this pattern) and before operation legalization after which
6600 // we need to be more careful about the vector instructions that we generate.
6601 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6602 LegalTypes && !LegalOperations && N0->hasOneUse() && VT != MVT::i1) {
6604 EVT VecTy = N0.getOperand(0).getValueType();
6605 EVT ExTy = N0.getValueType();
6606 EVT TrTy = N->getValueType(0);
6608 unsigned NumElem = VecTy.getVectorNumElements();
6609 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
6611 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
6612 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
6614 SDValue EltNo = N0->getOperand(1);
6615 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
6616 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
6617 EVT IndexTy = TLI.getVectorIdxTy();
6618 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
6620 SDValue V = DAG.getNode(ISD::BITCAST, SDLoc(N),
6621 NVT, N0.getOperand(0));
6623 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
6625 DAG.getConstant(Index, IndexTy));
6629 // trunc (select c, a, b) -> select c, (trunc a), (trunc b)
6630 if (N0.getOpcode() == ISD::SELECT) {
6631 EVT SrcVT = N0.getValueType();
6632 if ((!LegalOperations || TLI.isOperationLegal(ISD::SELECT, SrcVT)) &&
6633 TLI.isTruncateFree(SrcVT, VT)) {
6635 SDValue Cond = N0.getOperand(0);
6636 SDValue TruncOp0 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(1));
6637 SDValue TruncOp1 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(2));
6638 return DAG.getNode(ISD::SELECT, SDLoc(N), VT, Cond, TruncOp0, TruncOp1);
6642 // Fold a series of buildvector, bitcast, and truncate if possible.
6644 // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to
6645 // (2xi32 (buildvector x, y)).
6646 if (Level == AfterLegalizeVectorOps && VT.isVector() &&
6647 N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
6648 N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
6649 N0.getOperand(0).hasOneUse()) {
6651 SDValue BuildVect = N0.getOperand(0);
6652 EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType();
6653 EVT TruncVecEltTy = VT.getVectorElementType();
6655 // Check that the element types match.
6656 if (BuildVectEltTy == TruncVecEltTy) {
6657 // Now we only need to compute the offset of the truncated elements.
6658 unsigned BuildVecNumElts = BuildVect.getNumOperands();
6659 unsigned TruncVecNumElts = VT.getVectorNumElements();
6660 unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts;
6662 assert((BuildVecNumElts % TruncVecNumElts) == 0 &&
6663 "Invalid number of elements");
6665 SmallVector<SDValue, 8> Opnds;
6666 for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset)
6667 Opnds.push_back(BuildVect.getOperand(i));
6669 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
6673 // See if we can simplify the input to this truncate through knowledge that
6674 // only the low bits are being used.
6675 // For example "trunc (or (shl x, 8), y)" // -> trunc y
6676 // Currently we only perform this optimization on scalars because vectors
6677 // may have different active low bits.
6678 if (!VT.isVector()) {
6680 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
6681 VT.getSizeInBits()));
6682 if (Shorter.getNode())
6683 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter);
6685 // fold (truncate (load x)) -> (smaller load x)
6686 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
6687 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
6688 SDValue Reduced = ReduceLoadWidth(N);
6689 if (Reduced.getNode())
6691 // Handle the case where the load remains an extending load even
6692 // after truncation.
6693 if (N0.hasOneUse() && ISD::isUNINDEXEDLoad(N0.getNode())) {
6694 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6695 if (!LN0->isVolatile() &&
6696 LN0->getMemoryVT().getStoreSizeInBits() < VT.getSizeInBits()) {
6697 SDValue NewLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(LN0),
6698 VT, LN0->getChain(), LN0->getBasePtr(),
6700 LN0->getMemOperand());
6701 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLoad.getValue(1));
6706 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
6707 // where ... are all 'undef'.
6708 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
6709 SmallVector<EVT, 8> VTs;
6712 unsigned NumDefs = 0;
6714 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
6715 SDValue X = N0.getOperand(i);
6716 if (X.getOpcode() != ISD::UNDEF) {
6721 // Stop if more than one members are non-undef.
6724 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
6725 VT.getVectorElementType(),
6726 X.getValueType().getVectorNumElements()));
6730 return DAG.getUNDEF(VT);
6733 assert(V.getNode() && "The single defined operand is empty!");
6734 SmallVector<SDValue, 8> Opnds;
6735 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
6737 Opnds.push_back(DAG.getUNDEF(VTs[i]));
6740 SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V);
6741 AddToWorklist(NV.getNode());
6742 Opnds.push_back(NV);
6744 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Opnds);
6748 // Simplify the operands using demanded-bits information.
6749 if (!VT.isVector() &&
6750 SimplifyDemandedBits(SDValue(N, 0)))
6751 return SDValue(N, 0);
6756 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
6757 SDValue Elt = N->getOperand(i);
6758 if (Elt.getOpcode() != ISD::MERGE_VALUES)
6759 return Elt.getNode();
6760 return Elt.getOperand(Elt.getResNo()).getNode();
6763 /// build_pair (load, load) -> load
6764 /// if load locations are consecutive.
6765 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
6766 assert(N->getOpcode() == ISD::BUILD_PAIR);
6768 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
6769 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
6770 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
6771 LD1->getAddressSpace() != LD2->getAddressSpace())
6773 EVT LD1VT = LD1->getValueType(0);
6775 if (ISD::isNON_EXTLoad(LD2) &&
6777 // If both are volatile this would reduce the number of volatile loads.
6778 // If one is volatile it might be ok, but play conservative and bail out.
6779 !LD1->isVolatile() &&
6780 !LD2->isVolatile() &&
6781 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
6782 unsigned Align = LD1->getAlignment();
6783 unsigned NewAlign = TLI.getDataLayout()->
6784 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6786 if (NewAlign <= Align &&
6787 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
6788 return DAG.getLoad(VT, SDLoc(N), LD1->getChain(),
6789 LD1->getBasePtr(), LD1->getPointerInfo(),
6790 false, false, false, Align);
6796 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
6797 SDValue N0 = N->getOperand(0);
6798 EVT VT = N->getValueType(0);
6800 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
6801 // Only do this before legalize, since afterward the target may be depending
6802 // on the bitconvert.
6803 // First check to see if this is all constant.
6805 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
6807 bool isSimple = cast<BuildVectorSDNode>(N0)->isConstant();
6809 EVT DestEltVT = N->getValueType(0).getVectorElementType();
6810 assert(!DestEltVT.isVector() &&
6811 "Element type of vector ValueType must not be vector!");
6813 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
6816 // If the input is a constant, let getNode fold it.
6817 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
6818 // If we can't allow illegal operations, we need to check that this is just
6819 // a fp -> int or int -> conversion and that the resulting operation will
6821 if (!LegalOperations ||
6822 (isa<ConstantSDNode>(N0) && VT.isFloatingPoint() && !VT.isVector() &&
6823 TLI.isOperationLegal(ISD::ConstantFP, VT)) ||
6824 (isa<ConstantFPSDNode>(N0) && VT.isInteger() && !VT.isVector() &&
6825 TLI.isOperationLegal(ISD::Constant, VT)))
6826 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, N0);
6829 // (conv (conv x, t1), t2) -> (conv x, t2)
6830 if (N0.getOpcode() == ISD::BITCAST)
6831 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT,
6834 // fold (conv (load x)) -> (load (conv*)x)
6835 // If the resultant load doesn't need a higher alignment than the original!
6836 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6837 // Do not change the width of a volatile load.
6838 !cast<LoadSDNode>(N0)->isVolatile() &&
6839 // Do not remove the cast if the types differ in endian layout.
6840 TLI.hasBigEndianPartOrdering(N0.getValueType()) ==
6841 TLI.hasBigEndianPartOrdering(VT) &&
6842 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)) &&
6843 TLI.isLoadBitCastBeneficial(N0.getValueType(), VT)) {
6844 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6845 unsigned Align = TLI.getDataLayout()->
6846 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6847 unsigned OrigAlign = LN0->getAlignment();
6849 if (Align <= OrigAlign) {
6850 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(),
6851 LN0->getBasePtr(), LN0->getPointerInfo(),
6852 LN0->isVolatile(), LN0->isNonTemporal(),
6853 LN0->isInvariant(), OrigAlign,
6855 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
6860 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
6861 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
6862 // This often reduces constant pool loads.
6863 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) ||
6864 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) &&
6865 N0.getNode()->hasOneUse() && VT.isInteger() &&
6866 !VT.isVector() && !N0.getValueType().isVector()) {
6867 SDValue NewConv = DAG.getNode(ISD::BITCAST, SDLoc(N0), VT,
6869 AddToWorklist(NewConv.getNode());
6871 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6872 if (N0.getOpcode() == ISD::FNEG)
6873 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
6874 NewConv, DAG.getConstant(SignBit, VT));
6875 assert(N0.getOpcode() == ISD::FABS);
6876 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6877 NewConv, DAG.getConstant(~SignBit, VT));
6880 // fold (bitconvert (fcopysign cst, x)) ->
6881 // (or (and (bitconvert x), sign), (and cst, (not sign)))
6882 // Note that we don't handle (copysign x, cst) because this can always be
6883 // folded to an fneg or fabs.
6884 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
6885 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
6886 VT.isInteger() && !VT.isVector()) {
6887 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
6888 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
6889 if (isTypeLegal(IntXVT)) {
6890 SDValue X = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6891 IntXVT, N0.getOperand(1));
6892 AddToWorklist(X.getNode());
6894 // If X has a different width than the result/lhs, sext it or truncate it.
6895 unsigned VTWidth = VT.getSizeInBits();
6896 if (OrigXWidth < VTWidth) {
6897 X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X);
6898 AddToWorklist(X.getNode());
6899 } else if (OrigXWidth > VTWidth) {
6900 // To get the sign bit in the right place, we have to shift it right
6901 // before truncating.
6902 X = DAG.getNode(ISD::SRL, SDLoc(X),
6903 X.getValueType(), X,
6904 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
6905 AddToWorklist(X.getNode());
6906 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
6907 AddToWorklist(X.getNode());
6910 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6911 X = DAG.getNode(ISD::AND, SDLoc(X), VT,
6912 X, DAG.getConstant(SignBit, VT));
6913 AddToWorklist(X.getNode());
6915 SDValue Cst = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6916 VT, N0.getOperand(0));
6917 Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT,
6918 Cst, DAG.getConstant(~SignBit, VT));
6919 AddToWorklist(Cst.getNode());
6921 return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst);
6925 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
6926 if (N0.getOpcode() == ISD::BUILD_PAIR) {
6927 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
6928 if (CombineLD.getNode())
6935 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
6936 EVT VT = N->getValueType(0);
6937 return CombineConsecutiveLoads(N, VT);
6940 /// We know that BV is a build_vector node with Constant, ConstantFP or Undef
6941 /// operands. DstEltVT indicates the destination element value type.
6942 SDValue DAGCombiner::
6943 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
6944 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
6946 // If this is already the right type, we're done.
6947 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
6949 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
6950 unsigned DstBitSize = DstEltVT.getSizeInBits();
6952 // If this is a conversion of N elements of one type to N elements of another
6953 // type, convert each element. This handles FP<->INT cases.
6954 if (SrcBitSize == DstBitSize) {
6955 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6956 BV->getValueType(0).getVectorNumElements());
6958 // Due to the FP element handling below calling this routine recursively,
6959 // we can end up with a scalar-to-vector node here.
6960 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
6961 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6962 DAG.getNode(ISD::BITCAST, SDLoc(BV),
6963 DstEltVT, BV->getOperand(0)));
6965 SmallVector<SDValue, 8> Ops;
6966 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6967 SDValue Op = BV->getOperand(i);
6968 // If the vector element type is not legal, the BUILD_VECTOR operands
6969 // are promoted and implicitly truncated. Make that explicit here.
6970 if (Op.getValueType() != SrcEltVT)
6971 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op);
6972 Ops.push_back(DAG.getNode(ISD::BITCAST, SDLoc(BV),
6974 AddToWorklist(Ops.back().getNode());
6976 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6979 // Otherwise, we're growing or shrinking the elements. To avoid having to
6980 // handle annoying details of growing/shrinking FP values, we convert them to
6982 if (SrcEltVT.isFloatingPoint()) {
6983 // Convert the input float vector to a int vector where the elements are the
6985 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
6986 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
6990 // Now we know the input is an integer vector. If the output is a FP type,
6991 // convert to integer first, then to FP of the right size.
6992 if (DstEltVT.isFloatingPoint()) {
6993 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
6994 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
6996 // Next, convert to FP elements of the same size.
6997 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
7000 // Okay, we know the src/dst types are both integers of differing types.
7001 // Handling growing first.
7002 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
7003 if (SrcBitSize < DstBitSize) {
7004 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
7006 SmallVector<SDValue, 8> Ops;
7007 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
7008 i += NumInputsPerOutput) {
7009 bool isLE = TLI.isLittleEndian();
7010 APInt NewBits = APInt(DstBitSize, 0);
7011 bool EltIsUndef = true;
7012 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
7013 // Shift the previously computed bits over.
7014 NewBits <<= SrcBitSize;
7015 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
7016 if (Op.getOpcode() == ISD::UNDEF) continue;
7019 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
7020 zextOrTrunc(SrcBitSize).zext(DstBitSize);
7024 Ops.push_back(DAG.getUNDEF(DstEltVT));
7026 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
7029 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
7030 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
7033 // Finally, this must be the case where we are shrinking elements: each input
7034 // turns into multiple outputs.
7035 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
7036 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
7037 NumOutputsPerInput*BV->getNumOperands());
7038 SmallVector<SDValue, 8> Ops;
7040 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
7041 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
7042 Ops.append(NumOutputsPerInput, DAG.getUNDEF(DstEltVT));
7046 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
7047 getAPIntValue().zextOrTrunc(SrcBitSize);
7049 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
7050 APInt ThisVal = OpVal.trunc(DstBitSize);
7051 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
7052 OpVal = OpVal.lshr(DstBitSize);
7055 // For big endian targets, swap the order of the pieces of each element.
7056 if (TLI.isBigEndian())
7057 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
7060 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
7063 /// Try to perform FMA combining on a given FADD node.
7064 SDValue DAGCombiner::visitFADDForFMACombine(SDNode *N) {
7065 SDValue N0 = N->getOperand(0);
7066 SDValue N1 = N->getOperand(1);
7067 EVT VT = N->getValueType(0);
7070 const TargetOptions &Options = DAG.getTarget().Options;
7071 bool UnsafeFPMath = (Options.AllowFPOpFusion == FPOpFusion::Fast ||
7072 Options.UnsafeFPMath);
7074 // Floating-point multiply-add with intermediate rounding.
7075 bool HasFMAD = (LegalOperations &&
7076 TLI.isOperationLegal(ISD::FMAD, VT));
7078 // Floating-point multiply-add without intermediate rounding.
7079 bool HasFMA = ((!LegalOperations ||
7080 TLI.isOperationLegalOrCustom(ISD::FMA, VT)) &&
7081 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
7084 // No valid opcode, do not combine.
7085 if (!HasFMAD && !HasFMA)
7088 // Always prefer FMAD to FMA for precision.
7089 unsigned int PreferredFusedOpcode = HasFMAD ? ISD::FMAD : ISD::FMA;
7090 bool Aggressive = TLI.enableAggressiveFMAFusion(VT);
7091 bool LookThroughFPExt = TLI.isFPExtFree(VT);
7093 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
7094 if (N0.getOpcode() == ISD::FMUL &&
7095 (Aggressive || N0->hasOneUse())) {
7096 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7097 N0.getOperand(0), N0.getOperand(1), N1);
7100 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
7101 // Note: Commutes FADD operands.
7102 if (N1.getOpcode() == ISD::FMUL &&
7103 (Aggressive || N1->hasOneUse())) {
7104 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7105 N1.getOperand(0), N1.getOperand(1), N0);
7108 // Look through FP_EXTEND nodes to do more combining.
7109 if (UnsafeFPMath && LookThroughFPExt) {
7110 // fold (fadd (fpext (fmul x, y)), z) -> (fma (fpext x), (fpext y), z)
7111 if (N0.getOpcode() == ISD::FP_EXTEND) {
7112 SDValue N00 = N0.getOperand(0);
7113 if (N00.getOpcode() == ISD::FMUL)
7114 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7115 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7117 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7118 N00.getOperand(1)), N1);
7121 // fold (fadd x, (fpext (fmul y, z))) -> (fma (fpext y), (fpext z), x)
7122 // Note: Commutes FADD operands.
7123 if (N1.getOpcode() == ISD::FP_EXTEND) {
7124 SDValue N10 = N1.getOperand(0);
7125 if (N10.getOpcode() == ISD::FMUL)
7126 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7127 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7129 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7130 N10.getOperand(1)), N0);
7134 // More folding opportunities when target permits.
7135 if (UnsafeFPMath && Aggressive) {
7136 // fold (fadd (fma x, y, (fmul u, v)), z) -> (fma x, y (fma u, v, z))
7137 if (N0.getOpcode() == PreferredFusedOpcode &&
7138 N0.getOperand(2).getOpcode() == ISD::FMUL) {
7139 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7140 N0.getOperand(0), N0.getOperand(1),
7141 DAG.getNode(PreferredFusedOpcode, SL, VT,
7142 N0.getOperand(2).getOperand(0),
7143 N0.getOperand(2).getOperand(1),
7147 // fold (fadd x, (fma y, z, (fmul u, v)) -> (fma y, z (fma u, v, x))
7148 if (N1->getOpcode() == PreferredFusedOpcode &&
7149 N1.getOperand(2).getOpcode() == ISD::FMUL) {
7150 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7151 N1.getOperand(0), N1.getOperand(1),
7152 DAG.getNode(PreferredFusedOpcode, SL, VT,
7153 N1.getOperand(2).getOperand(0),
7154 N1.getOperand(2).getOperand(1),
7158 if (LookThroughFPExt) {
7159 // fold (fadd (fma x, y, (fpext (fmul u, v))), z)
7160 // -> (fma x, y, (fma (fpext u), (fpext v), z))
7161 auto FoldFAddFMAFPExtFMul = [&] (
7162 SDValue X, SDValue Y, SDValue U, SDValue V, SDValue Z) {
7163 return DAG.getNode(PreferredFusedOpcode, SL, VT, X, Y,
7164 DAG.getNode(PreferredFusedOpcode, SL, VT,
7165 DAG.getNode(ISD::FP_EXTEND, SL, VT, U),
7166 DAG.getNode(ISD::FP_EXTEND, SL, VT, V),
7169 if (N0.getOpcode() == PreferredFusedOpcode) {
7170 SDValue N02 = N0.getOperand(2);
7171 if (N02.getOpcode() == ISD::FP_EXTEND) {
7172 SDValue N020 = N02.getOperand(0);
7173 if (N020.getOpcode() == ISD::FMUL)
7174 return FoldFAddFMAFPExtFMul(N0.getOperand(0), N0.getOperand(1),
7175 N020.getOperand(0), N020.getOperand(1),
7180 // fold (fadd (fpext (fma x, y, (fmul u, v))), z)
7181 // -> (fma (fpext x), (fpext y), (fma (fpext u), (fpext v), z))
7182 // FIXME: This turns two single-precision and one double-precision
7183 // operation into two double-precision operations, which might not be
7184 // interesting for all targets, especially GPUs.
7185 auto FoldFAddFPExtFMAFMul = [&] (
7186 SDValue X, SDValue Y, SDValue U, SDValue V, SDValue Z) {
7187 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7188 DAG.getNode(ISD::FP_EXTEND, SL, VT, X),
7189 DAG.getNode(ISD::FP_EXTEND, SL, VT, Y),
7190 DAG.getNode(PreferredFusedOpcode, SL, VT,
7191 DAG.getNode(ISD::FP_EXTEND, SL, VT, U),
7192 DAG.getNode(ISD::FP_EXTEND, SL, VT, V),
7195 if (N0.getOpcode() == ISD::FP_EXTEND) {
7196 SDValue N00 = N0.getOperand(0);
7197 if (N00.getOpcode() == PreferredFusedOpcode) {
7198 SDValue N002 = N00.getOperand(2);
7199 if (N002.getOpcode() == ISD::FMUL)
7200 return FoldFAddFPExtFMAFMul(N00.getOperand(0), N00.getOperand(1),
7201 N002.getOperand(0), N002.getOperand(1),
7206 // fold (fadd x, (fma y, z, (fpext (fmul u, v)))
7207 // -> (fma y, z, (fma (fpext u), (fpext v), x))
7208 if (N1.getOpcode() == PreferredFusedOpcode) {
7209 SDValue N12 = N1.getOperand(2);
7210 if (N12.getOpcode() == ISD::FP_EXTEND) {
7211 SDValue N120 = N12.getOperand(0);
7212 if (N120.getOpcode() == ISD::FMUL)
7213 return FoldFAddFMAFPExtFMul(N1.getOperand(0), N1.getOperand(1),
7214 N120.getOperand(0), N120.getOperand(1),
7219 // fold (fadd x, (fpext (fma y, z, (fmul u, v)))
7220 // -> (fma (fpext y), (fpext z), (fma (fpext u), (fpext v), x))
7221 // FIXME: This turns two single-precision and one double-precision
7222 // operation into two double-precision operations, which might not be
7223 // interesting for all targets, especially GPUs.
7224 if (N1.getOpcode() == ISD::FP_EXTEND) {
7225 SDValue N10 = N1.getOperand(0);
7226 if (N10.getOpcode() == PreferredFusedOpcode) {
7227 SDValue N102 = N10.getOperand(2);
7228 if (N102.getOpcode() == ISD::FMUL)
7229 return FoldFAddFPExtFMAFMul(N10.getOperand(0), N10.getOperand(1),
7230 N102.getOperand(0), N102.getOperand(1),
7240 /// Try to perform FMA combining on a given FSUB node.
7241 SDValue DAGCombiner::visitFSUBForFMACombine(SDNode *N) {
7242 SDValue N0 = N->getOperand(0);
7243 SDValue N1 = N->getOperand(1);
7244 EVT VT = N->getValueType(0);
7247 const TargetOptions &Options = DAG.getTarget().Options;
7248 bool UnsafeFPMath = (Options.AllowFPOpFusion == FPOpFusion::Fast ||
7249 Options.UnsafeFPMath);
7251 // Floating-point multiply-add with intermediate rounding.
7252 bool HasFMAD = (LegalOperations &&
7253 TLI.isOperationLegal(ISD::FMAD, VT));
7255 // Floating-point multiply-add without intermediate rounding.
7256 bool HasFMA = ((!LegalOperations ||
7257 TLI.isOperationLegalOrCustom(ISD::FMA, VT)) &&
7258 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
7261 // No valid opcode, do not combine.
7262 if (!HasFMAD && !HasFMA)
7265 // Always prefer FMAD to FMA for precision.
7266 unsigned int PreferredFusedOpcode = HasFMAD ? ISD::FMAD : ISD::FMA;
7267 bool Aggressive = TLI.enableAggressiveFMAFusion(VT);
7268 bool LookThroughFPExt = TLI.isFPExtFree(VT);
7270 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
7271 if (N0.getOpcode() == ISD::FMUL &&
7272 (Aggressive || N0->hasOneUse())) {
7273 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7274 N0.getOperand(0), N0.getOperand(1),
7275 DAG.getNode(ISD::FNEG, SL, VT, N1));
7278 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
7279 // Note: Commutes FSUB operands.
7280 if (N1.getOpcode() == ISD::FMUL &&
7281 (Aggressive || N1->hasOneUse()))
7282 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7283 DAG.getNode(ISD::FNEG, SL, VT,
7285 N1.getOperand(1), N0);
7287 // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
7288 if (N0.getOpcode() == ISD::FNEG &&
7289 N0.getOperand(0).getOpcode() == ISD::FMUL &&
7290 (Aggressive || (N0->hasOneUse() && N0.getOperand(0).hasOneUse()))) {
7291 SDValue N00 = N0.getOperand(0).getOperand(0);
7292 SDValue N01 = N0.getOperand(0).getOperand(1);
7293 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7294 DAG.getNode(ISD::FNEG, SL, VT, N00), N01,
7295 DAG.getNode(ISD::FNEG, SL, VT, N1));
7298 // Look through FP_EXTEND nodes to do more combining.
7299 if (UnsafeFPMath && LookThroughFPExt) {
7300 // fold (fsub (fpext (fmul x, y)), z)
7301 // -> (fma (fpext x), (fpext y), (fneg z))
7302 if (N0.getOpcode() == ISD::FP_EXTEND) {
7303 SDValue N00 = N0.getOperand(0);
7304 if (N00.getOpcode() == ISD::FMUL)
7305 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7306 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7308 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7310 DAG.getNode(ISD::FNEG, SL, VT, N1));
7313 // fold (fsub x, (fpext (fmul y, z)))
7314 // -> (fma (fneg (fpext y)), (fpext z), x)
7315 // Note: Commutes FSUB operands.
7316 if (N1.getOpcode() == ISD::FP_EXTEND) {
7317 SDValue N10 = N1.getOperand(0);
7318 if (N10.getOpcode() == ISD::FMUL)
7319 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7320 DAG.getNode(ISD::FNEG, SL, VT,
7321 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7322 N10.getOperand(0))),
7323 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7328 // fold (fsub (fpext (fneg (fmul, x, y))), z)
7329 // -> (fneg (fma (fpext x), (fpext y), z))
7330 // Note: This could be removed with appropriate canonicalization of the
7331 // input expression into (fneg (fadd (fpext (fmul, x, y)), z). However, the
7332 // orthogonal flags -fp-contract=fast and -enable-unsafe-fp-math prevent
7333 // from implementing the canonicalization in visitFSUB.
7334 if (N0.getOpcode() == ISD::FP_EXTEND) {
7335 SDValue N00 = N0.getOperand(0);
7336 if (N00.getOpcode() == ISD::FNEG) {
7337 SDValue N000 = N00.getOperand(0);
7338 if (N000.getOpcode() == ISD::FMUL) {
7339 return DAG.getNode(ISD::FNEG, SL, VT,
7340 DAG.getNode(PreferredFusedOpcode, SL, VT,
7341 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7342 N000.getOperand(0)),
7343 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7344 N000.getOperand(1)),
7350 // fold (fsub (fneg (fpext (fmul, x, y))), z)
7351 // -> (fneg (fma (fpext x)), (fpext y), z)
7352 // Note: This could be removed with appropriate canonicalization of the
7353 // input expression into (fneg (fadd (fpext (fmul, x, y)), z). However, the
7354 // orthogonal flags -fp-contract=fast and -enable-unsafe-fp-math prevent
7355 // from implementing the canonicalization in visitFSUB.
7356 if (N0.getOpcode() == ISD::FNEG) {
7357 SDValue N00 = N0.getOperand(0);
7358 if (N00.getOpcode() == ISD::FP_EXTEND) {
7359 SDValue N000 = N00.getOperand(0);
7360 if (N000.getOpcode() == ISD::FMUL) {
7361 return DAG.getNode(ISD::FNEG, SL, VT,
7362 DAG.getNode(PreferredFusedOpcode, SL, VT,
7363 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7364 N000.getOperand(0)),
7365 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7366 N000.getOperand(1)),
7374 // More folding opportunities when target permits.
7375 if (UnsafeFPMath && Aggressive) {
7376 // fold (fsub (fma x, y, (fmul u, v)), z)
7377 // -> (fma x, y (fma u, v, (fneg z)))
7378 if (N0.getOpcode() == PreferredFusedOpcode &&
7379 N0.getOperand(2).getOpcode() == ISD::FMUL) {
7380 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7381 N0.getOperand(0), N0.getOperand(1),
7382 DAG.getNode(PreferredFusedOpcode, SL, VT,
7383 N0.getOperand(2).getOperand(0),
7384 N0.getOperand(2).getOperand(1),
7385 DAG.getNode(ISD::FNEG, SL, VT,
7389 // fold (fsub x, (fma y, z, (fmul u, v)))
7390 // -> (fma (fneg y), z, (fma (fneg u), v, x))
7391 if (N1.getOpcode() == PreferredFusedOpcode &&
7392 N1.getOperand(2).getOpcode() == ISD::FMUL) {
7393 SDValue N20 = N1.getOperand(2).getOperand(0);
7394 SDValue N21 = N1.getOperand(2).getOperand(1);
7395 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7396 DAG.getNode(ISD::FNEG, SL, VT,
7399 DAG.getNode(PreferredFusedOpcode, SL, VT,
7400 DAG.getNode(ISD::FNEG, SL, VT, N20),
7404 if (LookThroughFPExt) {
7405 // fold (fsub (fma x, y, (fpext (fmul u, v))), z)
7406 // -> (fma x, y (fma (fpext u), (fpext v), (fneg z)))
7407 if (N0.getOpcode() == PreferredFusedOpcode) {
7408 SDValue N02 = N0.getOperand(2);
7409 if (N02.getOpcode() == ISD::FP_EXTEND) {
7410 SDValue N020 = N02.getOperand(0);
7411 if (N020.getOpcode() == ISD::FMUL)
7412 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7413 N0.getOperand(0), N0.getOperand(1),
7414 DAG.getNode(PreferredFusedOpcode, SL, VT,
7415 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7416 N020.getOperand(0)),
7417 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7418 N020.getOperand(1)),
7419 DAG.getNode(ISD::FNEG, SL, VT,
7424 // fold (fsub (fpext (fma x, y, (fmul u, v))), z)
7425 // -> (fma (fpext x), (fpext y),
7426 // (fma (fpext u), (fpext v), (fneg z)))
7427 // FIXME: This turns two single-precision and one double-precision
7428 // operation into two double-precision operations, which might not be
7429 // interesting for all targets, especially GPUs.
7430 if (N0.getOpcode() == ISD::FP_EXTEND) {
7431 SDValue N00 = N0.getOperand(0);
7432 if (N00.getOpcode() == PreferredFusedOpcode) {
7433 SDValue N002 = N00.getOperand(2);
7434 if (N002.getOpcode() == ISD::FMUL)
7435 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7436 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7438 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7440 DAG.getNode(PreferredFusedOpcode, SL, VT,
7441 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7442 N002.getOperand(0)),
7443 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7444 N002.getOperand(1)),
7445 DAG.getNode(ISD::FNEG, SL, VT,
7450 // fold (fsub x, (fma y, z, (fpext (fmul u, v))))
7451 // -> (fma (fneg y), z, (fma (fneg (fpext u)), (fpext v), x))
7452 if (N1.getOpcode() == PreferredFusedOpcode &&
7453 N1.getOperand(2).getOpcode() == ISD::FP_EXTEND) {
7454 SDValue N120 = N1.getOperand(2).getOperand(0);
7455 if (N120.getOpcode() == ISD::FMUL) {
7456 SDValue N1200 = N120.getOperand(0);
7457 SDValue N1201 = N120.getOperand(1);
7458 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7459 DAG.getNode(ISD::FNEG, SL, VT, N1.getOperand(0)),
7461 DAG.getNode(PreferredFusedOpcode, SL, VT,
7462 DAG.getNode(ISD::FNEG, SL, VT,
7463 DAG.getNode(ISD::FP_EXTEND, SL,
7465 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7471 // fold (fsub x, (fpext (fma y, z, (fmul u, v))))
7472 // -> (fma (fneg (fpext y)), (fpext z),
7473 // (fma (fneg (fpext u)), (fpext v), x))
7474 // FIXME: This turns two single-precision and one double-precision
7475 // operation into two double-precision operations, which might not be
7476 // interesting for all targets, especially GPUs.
7477 if (N1.getOpcode() == ISD::FP_EXTEND &&
7478 N1.getOperand(0).getOpcode() == PreferredFusedOpcode) {
7479 SDValue N100 = N1.getOperand(0).getOperand(0);
7480 SDValue N101 = N1.getOperand(0).getOperand(1);
7481 SDValue N102 = N1.getOperand(0).getOperand(2);
7482 if (N102.getOpcode() == ISD::FMUL) {
7483 SDValue N1020 = N102.getOperand(0);
7484 SDValue N1021 = N102.getOperand(1);
7485 return DAG.getNode(PreferredFusedOpcode, SL, VT,
7486 DAG.getNode(ISD::FNEG, SL, VT,
7487 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7489 DAG.getNode(ISD::FP_EXTEND, SL, VT, N101),
7490 DAG.getNode(PreferredFusedOpcode, SL, VT,
7491 DAG.getNode(ISD::FNEG, SL, VT,
7492 DAG.getNode(ISD::FP_EXTEND, SL,
7494 DAG.getNode(ISD::FP_EXTEND, SL, VT,
7505 SDValue DAGCombiner::visitFADD(SDNode *N) {
7506 SDValue N0 = N->getOperand(0);
7507 SDValue N1 = N->getOperand(1);
7508 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7509 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7510 EVT VT = N->getValueType(0);
7511 const TargetOptions &Options = DAG.getTarget().Options;
7515 if (SDValue FoldedVOp = SimplifyVBinOp(N))
7518 // fold (fadd c1, c2) -> c1 + c2
7520 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N1);
7522 // canonicalize constant to RHS
7523 if (N0CFP && !N1CFP)
7524 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N0);
7526 // fold (fadd A, (fneg B)) -> (fsub A, B)
7527 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
7528 isNegatibleForFree(N1, LegalOperations, TLI, &Options) == 2)
7529 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0,
7530 GetNegatedExpression(N1, DAG, LegalOperations));
7532 // fold (fadd (fneg A), B) -> (fsub B, A)
7533 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
7534 isNegatibleForFree(N0, LegalOperations, TLI, &Options) == 2)
7535 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N1,
7536 GetNegatedExpression(N0, DAG, LegalOperations));
7538 // If 'unsafe math' is enabled, fold lots of things.
7539 if (Options.UnsafeFPMath) {
7540 // No FP constant should be created after legalization as Instruction
7541 // Selection pass has a hard time dealing with FP constants.
7542 bool AllowNewConst = (Level < AfterLegalizeDAG);
7544 // fold (fadd A, 0) -> A
7545 if (N1CFP && N1CFP->getValueAPF().isZero())
7548 // fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
7549 if (N1CFP && N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
7550 isa<ConstantFPSDNode>(N0.getOperand(1)))
7551 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0.getOperand(0),
7552 DAG.getNode(ISD::FADD, SDLoc(N), VT,
7553 N0.getOperand(1), N1));
7555 // If allowed, fold (fadd (fneg x), x) -> 0.0
7556 if (AllowNewConst && N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1)
7557 return DAG.getConstantFP(0.0, VT);
7559 // If allowed, fold (fadd x, (fneg x)) -> 0.0
7560 if (AllowNewConst && N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0)
7561 return DAG.getConstantFP(0.0, VT);
7563 // We can fold chains of FADD's of the same value into multiplications.
7564 // This transform is not safe in general because we are reducing the number
7565 // of rounding steps.
7566 if (TLI.isOperationLegalOrCustom(ISD::FMUL, VT) && !N0CFP && !N1CFP) {
7567 if (N0.getOpcode() == ISD::FMUL) {
7568 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
7569 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
7571 // (fadd (fmul x, c), x) -> (fmul x, c+1)
7572 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
7573 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
7575 DAG.getConstantFP(1.0, VT));
7576 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, NewCFP);
7579 // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2)
7580 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
7581 N1.getOperand(0) == N1.getOperand(1) &&
7582 N0.getOperand(0) == N1.getOperand(0)) {
7583 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
7585 DAG.getConstantFP(2.0, VT));
7586 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7587 N0.getOperand(0), NewCFP);
7591 if (N1.getOpcode() == ISD::FMUL) {
7592 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
7593 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
7595 // (fadd x, (fmul x, c)) -> (fmul x, c+1)
7596 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
7597 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
7599 DAG.getConstantFP(1.0, VT));
7600 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, NewCFP);
7603 // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2)
7604 if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD &&
7605 N0.getOperand(0) == N0.getOperand(1) &&
7606 N1.getOperand(0) == N0.getOperand(0)) {
7607 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
7609 DAG.getConstantFP(2.0, VT));
7610 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1.getOperand(0), NewCFP);
7614 if (N0.getOpcode() == ISD::FADD && AllowNewConst) {
7615 ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
7616 // (fadd (fadd x, x), x) -> (fmul x, 3.0)
7617 if (!CFP && N0.getOperand(0) == N0.getOperand(1) &&
7618 (N0.getOperand(0) == N1))
7619 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7620 N1, DAG.getConstantFP(3.0, VT));
7623 if (N1.getOpcode() == ISD::FADD && AllowNewConst) {
7624 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
7625 // (fadd x, (fadd x, x)) -> (fmul x, 3.0)
7626 if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) &&
7627 N1.getOperand(0) == N0)
7628 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7629 N0, DAG.getConstantFP(3.0, VT));
7632 // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0)
7633 if (AllowNewConst &&
7634 N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
7635 N0.getOperand(0) == N0.getOperand(1) &&
7636 N1.getOperand(0) == N1.getOperand(1) &&
7637 N0.getOperand(0) == N1.getOperand(0))
7638 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7639 N0.getOperand(0), DAG.getConstantFP(4.0, VT));
7641 } // enable-unsafe-fp-math
7643 // FADD -> FMA combines:
7644 SDValue Fused = visitFADDForFMACombine(N);
7646 AddToWorklist(Fused.getNode());
7653 SDValue DAGCombiner::visitFSUB(SDNode *N) {
7654 SDValue N0 = N->getOperand(0);
7655 SDValue N1 = N->getOperand(1);
7656 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
7657 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
7658 EVT VT = N->getValueType(0);
7660 const TargetOptions &Options = DAG.getTarget().Options;
7664 if (SDValue FoldedVOp = SimplifyVBinOp(N))
7667 // fold (fsub c1, c2) -> c1-c2
7669 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0, N1);
7671 // fold (fsub A, (fneg B)) -> (fadd A, B)
7672 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
7673 return DAG.getNode(ISD::FADD, dl, VT, N0,
7674 GetNegatedExpression(N1, DAG, LegalOperations));
7676 // If 'unsafe math' is enabled, fold lots of things.
7677 if (Options.UnsafeFPMath) {
7679 if (N1CFP && N1CFP->getValueAPF().isZero())
7682 // (fsub 0, B) -> -B
7683 if (N0CFP && N0CFP->getValueAPF().isZero()) {
7684 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
7685 return GetNegatedExpression(N1, DAG, LegalOperations);
7686 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7687 return DAG.getNode(ISD::FNEG, dl, VT, N1);
7690 // (fsub x, x) -> 0.0
7692 return DAG.getConstantFP(0.0f, VT);
7694 // (fsub x, (fadd x, y)) -> (fneg y)
7695 // (fsub x, (fadd y, x)) -> (fneg y)
7696 if (N1.getOpcode() == ISD::FADD) {
7697 SDValue N10 = N1->getOperand(0);
7698 SDValue N11 = N1->getOperand(1);
7700 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI, &Options))
7701 return GetNegatedExpression(N11, DAG, LegalOperations);
7703 if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI, &Options))
7704 return GetNegatedExpression(N10, DAG, LegalOperations);
7708 // FSUB -> FMA combines:
7709 SDValue Fused = visitFSUBForFMACombine(N);
7711 AddToWorklist(Fused.getNode());
7718 SDValue DAGCombiner::visitFMUL(SDNode *N) {
7719 SDValue N0 = N->getOperand(0);
7720 SDValue N1 = N->getOperand(1);
7721 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
7722 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
7723 EVT VT = N->getValueType(0);
7724 const TargetOptions &Options = DAG.getTarget().Options;
7727 if (VT.isVector()) {
7728 // This just handles C1 * C2 for vectors. Other vector folds are below.
7729 if (SDValue FoldedVOp = SimplifyVBinOp(N))
7733 // fold (fmul c1, c2) -> c1*c2
7735 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, N1);
7737 // canonicalize constant to RHS
7738 if (isConstantFPBuildVectorOrConstantFP(N0) &&
7739 !isConstantFPBuildVectorOrConstantFP(N1))
7740 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, N0);
7742 // fold (fmul A, 1.0) -> A
7743 if (N1CFP && N1CFP->isExactlyValue(1.0))
7746 if (Options.UnsafeFPMath) {
7747 // fold (fmul A, 0) -> 0
7748 if (N1CFP && N1CFP->getValueAPF().isZero())
7751 // fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
7752 if (N0.getOpcode() == ISD::FMUL) {
7753 // Fold scalars or any vector constants (not just splats).
7754 // This fold is done in general by InstCombine, but extra fmul insts
7755 // may have been generated during lowering.
7756 SDValue N00 = N0.getOperand(0);
7757 SDValue N01 = N0.getOperand(1);
7758 auto *BV1 = dyn_cast<BuildVectorSDNode>(N1);
7759 auto *BV00 = dyn_cast<BuildVectorSDNode>(N00);
7760 auto *BV01 = dyn_cast<BuildVectorSDNode>(N01);
7762 // Check 1: Make sure that the first operand of the inner multiply is NOT
7763 // a constant. Otherwise, we may induce infinite looping.
7764 if (!(isConstOrConstSplatFP(N00) || (BV00 && BV00->isConstant()))) {
7765 // Check 2: Make sure that the second operand of the inner multiply and
7766 // the second operand of the outer multiply are constants.
7767 if ((N1CFP && isConstOrConstSplatFP(N01)) ||
7768 (BV1 && BV01 && BV1->isConstant() && BV01->isConstant())) {
7770 SDValue MulConsts = DAG.getNode(ISD::FMUL, SL, VT, N01, N1);
7771 return DAG.getNode(ISD::FMUL, SL, VT, N00, MulConsts);
7776 // fold (fmul (fadd x, x), c) -> (fmul x, (fmul 2.0, c))
7777 // Undo the fmul 2.0, x -> fadd x, x transformation, since if it occurs
7778 // during an early run of DAGCombiner can prevent folding with fmuls
7779 // inserted during lowering.
7780 if (N0.getOpcode() == ISD::FADD && N0.getOperand(0) == N0.getOperand(1)) {
7782 const SDValue Two = DAG.getConstantFP(2.0, VT);
7783 SDValue MulConsts = DAG.getNode(ISD::FMUL, SL, VT, Two, N1);
7784 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0.getOperand(0), MulConsts);
7788 // fold (fmul X, 2.0) -> (fadd X, X)
7789 if (N1CFP && N1CFP->isExactlyValue(+2.0))
7790 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N0);
7792 // fold (fmul X, -1.0) -> (fneg X)
7793 if (N1CFP && N1CFP->isExactlyValue(-1.0))
7794 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7795 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
7797 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
7798 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
7799 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
7800 // Both can be negated for free, check to see if at least one is cheaper
7802 if (LHSNeg == 2 || RHSNeg == 2)
7803 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7804 GetNegatedExpression(N0, DAG, LegalOperations),
7805 GetNegatedExpression(N1, DAG, LegalOperations));
7812 SDValue DAGCombiner::visitFMA(SDNode *N) {
7813 SDValue N0 = N->getOperand(0);
7814 SDValue N1 = N->getOperand(1);
7815 SDValue N2 = N->getOperand(2);
7816 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7817 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7818 EVT VT = N->getValueType(0);
7820 const TargetOptions &Options = DAG.getTarget().Options;
7822 // Constant fold FMA.
7823 if (isa<ConstantFPSDNode>(N0) &&
7824 isa<ConstantFPSDNode>(N1) &&
7825 isa<ConstantFPSDNode>(N2)) {
7826 return DAG.getNode(ISD::FMA, dl, VT, N0, N1, N2);
7829 if (Options.UnsafeFPMath) {
7830 if (N0CFP && N0CFP->isZero())
7832 if (N1CFP && N1CFP->isZero())
7835 if (N0CFP && N0CFP->isExactlyValue(1.0))
7836 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2);
7837 if (N1CFP && N1CFP->isExactlyValue(1.0))
7838 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2);
7840 // Canonicalize (fma c, x, y) -> (fma x, c, y)
7841 if (N0CFP && !N1CFP)
7842 return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2);
7844 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
7845 if (Options.UnsafeFPMath && N1CFP &&
7846 N2.getOpcode() == ISD::FMUL &&
7847 N0 == N2.getOperand(0) &&
7848 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
7849 return DAG.getNode(ISD::FMUL, dl, VT, N0,
7850 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
7854 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
7855 if (Options.UnsafeFPMath &&
7856 N0.getOpcode() == ISD::FMUL && N1CFP &&
7857 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
7858 return DAG.getNode(ISD::FMA, dl, VT,
7860 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
7864 // (fma x, 1, y) -> (fadd x, y)
7865 // (fma x, -1, y) -> (fadd (fneg x), y)
7867 if (N1CFP->isExactlyValue(1.0))
7868 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
7870 if (N1CFP->isExactlyValue(-1.0) &&
7871 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
7872 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
7873 AddToWorklist(RHSNeg.getNode());
7874 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
7878 // (fma x, c, x) -> (fmul x, (c+1))
7879 if (Options.UnsafeFPMath && N1CFP && N0 == N2)
7880 return DAG.getNode(ISD::FMUL, dl, VT, N0,
7881 DAG.getNode(ISD::FADD, dl, VT,
7882 N1, DAG.getConstantFP(1.0, VT)));
7884 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
7885 if (Options.UnsafeFPMath && N1CFP &&
7886 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0)
7887 return DAG.getNode(ISD::FMUL, dl, VT, N0,
7888 DAG.getNode(ISD::FADD, dl, VT,
7889 N1, DAG.getConstantFP(-1.0, VT)));
7895 SDValue DAGCombiner::visitFDIV(SDNode *N) {
7896 SDValue N0 = N->getOperand(0);
7897 SDValue N1 = N->getOperand(1);
7898 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7899 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7900 EVT VT = N->getValueType(0);
7902 const TargetOptions &Options = DAG.getTarget().Options;
7906 if (SDValue FoldedVOp = SimplifyVBinOp(N))
7909 // fold (fdiv c1, c2) -> c1/c2
7911 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1);
7913 if (Options.UnsafeFPMath) {
7914 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
7916 // Compute the reciprocal 1.0 / c2.
7917 APFloat N1APF = N1CFP->getValueAPF();
7918 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
7919 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
7920 // Only do the transform if the reciprocal is a legal fp immediate that
7921 // isn't too nasty (eg NaN, denormal, ...).
7922 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
7923 (!LegalOperations ||
7924 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
7925 // backend)... we should handle this gracefully after Legalize.
7926 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
7927 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
7928 TLI.isFPImmLegal(Recip, VT)))
7929 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0,
7930 DAG.getConstantFP(Recip, VT));
7933 // If this FDIV is part of a reciprocal square root, it may be folded
7934 // into a target-specific square root estimate instruction.
7935 if (N1.getOpcode() == ISD::FSQRT) {
7936 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0))) {
7937 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7939 } else if (N1.getOpcode() == ISD::FP_EXTEND &&
7940 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
7941 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0).getOperand(0))) {
7942 RV = DAG.getNode(ISD::FP_EXTEND, SDLoc(N1), VT, RV);
7943 AddToWorklist(RV.getNode());
7944 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7946 } else if (N1.getOpcode() == ISD::FP_ROUND &&
7947 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
7948 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0).getOperand(0))) {
7949 RV = DAG.getNode(ISD::FP_ROUND, SDLoc(N1), VT, RV, N1.getOperand(1));
7950 AddToWorklist(RV.getNode());
7951 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7953 } else if (N1.getOpcode() == ISD::FMUL) {
7954 // Look through an FMUL. Even though this won't remove the FDIV directly,
7955 // it's still worthwhile to get rid of the FSQRT if possible.
7958 if (N1.getOperand(0).getOpcode() == ISD::FSQRT) {
7959 SqrtOp = N1.getOperand(0);
7960 OtherOp = N1.getOperand(1);
7961 } else if (N1.getOperand(1).getOpcode() == ISD::FSQRT) {
7962 SqrtOp = N1.getOperand(1);
7963 OtherOp = N1.getOperand(0);
7965 if (SqrtOp.getNode()) {
7966 // We found a FSQRT, so try to make this fold:
7967 // x / (y * sqrt(z)) -> x * (rsqrt(z) / y)
7968 if (SDValue RV = BuildRsqrtEstimate(SqrtOp.getOperand(0))) {
7969 RV = DAG.getNode(ISD::FDIV, SDLoc(N1), VT, RV, OtherOp);
7970 AddToWorklist(RV.getNode());
7971 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7976 // Fold into a reciprocal estimate and multiply instead of a real divide.
7977 if (SDValue RV = BuildReciprocalEstimate(N1)) {
7978 AddToWorklist(RV.getNode());
7979 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7983 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
7984 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
7985 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
7986 // Both can be negated for free, check to see if at least one is cheaper
7988 if (LHSNeg == 2 || RHSNeg == 2)
7989 return DAG.getNode(ISD::FDIV, SDLoc(N), VT,
7990 GetNegatedExpression(N0, DAG, LegalOperations),
7991 GetNegatedExpression(N1, DAG, LegalOperations));
7995 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
7997 // E.g., (a / D; b / D;) -> (recip = 1.0 / D; a * recip; b * recip)
7998 // Notice that this is not always beneficial. One reason is different target
7999 // may have different costs for FDIV and FMUL, so sometimes the cost of two
8000 // FDIVs may be lower than the cost of one FDIV and two FMULs. Another reason
8001 // is the critical path is increased from "one FDIV" to "one FDIV + one FMUL".
8002 if (Options.UnsafeFPMath) {
8003 // Skip if current node is a reciprocal.
8004 if (N0CFP && N0CFP->isExactlyValue(1.0))
8007 SmallVector<SDNode *, 4> Users;
8008 // Find all FDIV users of the same divisor.
8009 for (SDNode::use_iterator UI = N1.getNode()->use_begin(),
8010 UE = N1.getNode()->use_end();
8012 SDNode *User = UI.getUse().getUser();
8013 if (User->getOpcode() == ISD::FDIV && User->getOperand(1) == N1)
8014 Users.push_back(User);
8017 if (TLI.combineRepeatedFPDivisors(Users.size())) {
8018 SDValue FPOne = DAG.getConstantFP(1.0, VT); // floating point 1.0
8019 SDValue Reciprocal = DAG.getNode(ISD::FDIV, SDLoc(N), VT, FPOne, N1);
8021 // Dividend / Divisor -> Dividend * Reciprocal
8022 for (auto I = Users.begin(), E = Users.end(); I != E; ++I) {
8023 if ((*I)->getOperand(0) != FPOne) {
8024 SDValue NewNode = DAG.getNode(ISD::FMUL, SDLoc(*I), VT,
8025 (*I)->getOperand(0), Reciprocal);
8026 DAG.ReplaceAllUsesWith(*I, NewNode.getNode());
8036 SDValue DAGCombiner::visitFREM(SDNode *N) {
8037 SDValue N0 = N->getOperand(0);
8038 SDValue N1 = N->getOperand(1);
8039 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8040 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8041 EVT VT = N->getValueType(0);
8043 // fold (frem c1, c2) -> fmod(c1,c2)
8045 return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1);
8050 SDValue DAGCombiner::visitFSQRT(SDNode *N) {
8051 if (DAG.getTarget().Options.UnsafeFPMath &&
8052 !TLI.isFsqrtCheap()) {
8053 // Compute this as X * (1/sqrt(X)) = X * (X ** -0.5)
8054 if (SDValue RV = BuildRsqrtEstimate(N->getOperand(0))) {
8055 EVT VT = RV.getValueType();
8056 RV = DAG.getNode(ISD::FMUL, SDLoc(N), VT, N->getOperand(0), RV);
8057 AddToWorklist(RV.getNode());
8059 // Unfortunately, RV is now NaN if the input was exactly 0.
8060 // Select out this case and force the answer to 0.
8061 SDValue Zero = DAG.getConstantFP(0.0, VT);
8063 DAG.getSetCC(SDLoc(N), TLI.getSetCCResultType(*DAG.getContext(), VT),
8064 N->getOperand(0), Zero, ISD::SETEQ);
8065 AddToWorklist(ZeroCmp.getNode());
8066 AddToWorklist(RV.getNode());
8068 RV = DAG.getNode(VT.isVector() ? ISD::VSELECT : ISD::SELECT,
8069 SDLoc(N), VT, ZeroCmp, Zero, RV);
8076 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
8077 SDValue N0 = N->getOperand(0);
8078 SDValue N1 = N->getOperand(1);
8079 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8080 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8081 EVT VT = N->getValueType(0);
8083 if (N0CFP && N1CFP) // Constant fold
8084 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1);
8087 const APFloat& V = N1CFP->getValueAPF();
8088 // copysign(x, c1) -> fabs(x) iff ispos(c1)
8089 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
8090 if (!V.isNegative()) {
8091 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
8092 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
8094 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
8095 return DAG.getNode(ISD::FNEG, SDLoc(N), VT,
8096 DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0));
8100 // copysign(fabs(x), y) -> copysign(x, y)
8101 // copysign(fneg(x), y) -> copysign(x, y)
8102 // copysign(copysign(x,z), y) -> copysign(x, y)
8103 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
8104 N0.getOpcode() == ISD::FCOPYSIGN)
8105 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
8106 N0.getOperand(0), N1);
8108 // copysign(x, abs(y)) -> abs(x)
8109 if (N1.getOpcode() == ISD::FABS)
8110 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
8112 // copysign(x, copysign(y,z)) -> copysign(x, z)
8113 if (N1.getOpcode() == ISD::FCOPYSIGN)
8114 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
8115 N0, N1.getOperand(1));
8117 // copysign(x, fp_extend(y)) -> copysign(x, y)
8118 // copysign(x, fp_round(y)) -> copysign(x, y)
8119 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
8120 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
8121 N0, N1.getOperand(0));
8126 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
8127 SDValue N0 = N->getOperand(0);
8128 EVT VT = N->getValueType(0);
8129 EVT OpVT = N0.getValueType();
8131 // fold (sint_to_fp c1) -> c1fp
8132 if (isConstantIntBuildVectorOrConstantInt(N0) &&
8133 // ...but only if the target supports immediate floating-point values
8134 (!LegalOperations ||
8135 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
8136 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
8138 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
8139 // but UINT_TO_FP is legal on this target, try to convert.
8140 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
8141 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
8142 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
8143 if (DAG.SignBitIsZero(N0))
8144 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
8147 // The next optimizations are desirable only if SELECT_CC can be lowered.
8148 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
8149 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
8150 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
8152 (!LegalOperations ||
8153 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
8155 { N0.getOperand(0), N0.getOperand(1),
8156 DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT),
8158 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
8161 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
8162 // (select_cc x, y, 1.0, 0.0,, cc)
8163 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
8164 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
8165 (!LegalOperations ||
8166 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
8168 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
8169 DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT),
8170 N0.getOperand(0).getOperand(2) };
8171 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
8178 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
8179 SDValue N0 = N->getOperand(0);
8180 EVT VT = N->getValueType(0);
8181 EVT OpVT = N0.getValueType();
8183 // fold (uint_to_fp c1) -> c1fp
8184 if (isConstantIntBuildVectorOrConstantInt(N0) &&
8185 // ...but only if the target supports immediate floating-point values
8186 (!LegalOperations ||
8187 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
8188 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
8190 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
8191 // but SINT_TO_FP is legal on this target, try to convert.
8192 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
8193 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
8194 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
8195 if (DAG.SignBitIsZero(N0))
8196 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
8199 // The next optimizations are desirable only if SELECT_CC can be lowered.
8200 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
8201 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
8203 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
8204 (!LegalOperations ||
8205 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
8207 { N0.getOperand(0), N0.getOperand(1),
8208 DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT),
8210 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
8217 // Fold (fp_to_{s/u}int ({s/u}int_to_fpx)) -> zext x, sext x, trunc x, or x
8218 static SDValue FoldIntToFPToInt(SDNode *N, SelectionDAG &DAG) {
8219 SDValue N0 = N->getOperand(0);
8220 EVT VT = N->getValueType(0);
8222 if (N0.getOpcode() != ISD::UINT_TO_FP && N0.getOpcode() != ISD::SINT_TO_FP)
8225 SDValue Src = N0.getOperand(0);
8226 EVT SrcVT = Src.getValueType();
8227 bool IsInputSigned = N0.getOpcode() == ISD::SINT_TO_FP;
8228 bool IsOutputSigned = N->getOpcode() == ISD::FP_TO_SINT;
8230 // We can safely assume the conversion won't overflow the output range,
8231 // because (for example) (uint8_t)18293.f is undefined behavior.
8233 // Since we can assume the conversion won't overflow, our decision as to
8234 // whether the input will fit in the float should depend on the minimum
8235 // of the input range and output range.
8237 // This means this is also safe for a signed input and unsigned output, since
8238 // a negative input would lead to undefined behavior.
8239 unsigned InputSize = (int)SrcVT.getScalarSizeInBits() - IsInputSigned;
8240 unsigned OutputSize = (int)VT.getScalarSizeInBits() - IsOutputSigned;
8241 unsigned ActualSize = std::min(InputSize, OutputSize);
8242 const fltSemantics &sem = DAG.EVTToAPFloatSemantics(N0.getValueType());
8244 // We can only fold away the float conversion if the input range can be
8245 // represented exactly in the float range.
8246 if (APFloat::semanticsPrecision(sem) >= ActualSize) {
8247 if (VT.getScalarSizeInBits() > SrcVT.getScalarSizeInBits()) {
8248 unsigned ExtOp = IsInputSigned && IsOutputSigned ? ISD::SIGN_EXTEND
8250 return DAG.getNode(ExtOp, SDLoc(N), VT, Src);
8252 if (VT.getScalarSizeInBits() < SrcVT.getScalarSizeInBits())
8253 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Src);
8256 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Src);
8261 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
8262 SDValue N0 = N->getOperand(0);
8263 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8264 EVT VT = N->getValueType(0);
8266 // fold (fp_to_sint c1fp) -> c1
8268 return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0);
8270 return FoldIntToFPToInt(N, DAG);
8273 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
8274 SDValue N0 = N->getOperand(0);
8275 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8276 EVT VT = N->getValueType(0);
8278 // fold (fp_to_uint c1fp) -> c1
8280 return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0);
8282 return FoldIntToFPToInt(N, DAG);
8285 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
8286 SDValue N0 = N->getOperand(0);
8287 SDValue N1 = N->getOperand(1);
8288 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8289 EVT VT = N->getValueType(0);
8291 // fold (fp_round c1fp) -> c1fp
8293 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1);
8295 // fold (fp_round (fp_extend x)) -> x
8296 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
8297 return N0.getOperand(0);
8299 // fold (fp_round (fp_round x)) -> (fp_round x)
8300 if (N0.getOpcode() == ISD::FP_ROUND) {
8301 const bool NIsTrunc = N->getConstantOperandVal(1) == 1;
8302 const bool N0IsTrunc = N0.getNode()->getConstantOperandVal(1) == 1;
8303 // If the first fp_round isn't a value preserving truncation, it might
8304 // introduce a tie in the second fp_round, that wouldn't occur in the
8305 // single-step fp_round we want to fold to.
8306 // In other words, double rounding isn't the same as rounding.
8307 // Also, this is a value preserving truncation iff both fp_round's are.
8308 if (DAG.getTarget().Options.UnsafeFPMath || N0IsTrunc)
8309 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0.getOperand(0),
8310 DAG.getIntPtrConstant(NIsTrunc && N0IsTrunc));
8313 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
8314 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
8315 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT,
8316 N0.getOperand(0), N1);
8317 AddToWorklist(Tmp.getNode());
8318 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
8319 Tmp, N0.getOperand(1));
8325 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
8326 SDValue N0 = N->getOperand(0);
8327 EVT VT = N->getValueType(0);
8328 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
8329 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8331 // fold (fp_round_inreg c1fp) -> c1fp
8332 if (N0CFP && isTypeLegal(EVT)) {
8333 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
8334 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, Round);
8340 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
8341 SDValue N0 = N->getOperand(0);
8342 EVT VT = N->getValueType(0);
8344 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
8345 if (N->hasOneUse() &&
8346 N->use_begin()->getOpcode() == ISD::FP_ROUND)
8349 // fold (fp_extend c1fp) -> c1fp
8350 if (isConstantFPBuildVectorOrConstantFP(N0))
8351 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0);
8353 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
8355 if (N0.getOpcode() == ISD::FP_ROUND
8356 && N0.getNode()->getConstantOperandVal(1) == 1) {
8357 SDValue In = N0.getOperand(0);
8358 if (In.getValueType() == VT) return In;
8359 if (VT.bitsLT(In.getValueType()))
8360 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT,
8361 In, N0.getOperand(1));
8362 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In);
8365 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
8366 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
8367 TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) {
8368 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
8369 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
8371 LN0->getBasePtr(), N0.getValueType(),
8372 LN0->getMemOperand());
8373 CombineTo(N, ExtLoad);
8374 CombineTo(N0.getNode(),
8375 DAG.getNode(ISD::FP_ROUND, SDLoc(N0),
8376 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
8377 ExtLoad.getValue(1));
8378 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8384 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
8385 SDValue N0 = N->getOperand(0);
8386 EVT VT = N->getValueType(0);
8388 // fold (fceil c1) -> fceil(c1)
8389 if (isConstantFPBuildVectorOrConstantFP(N0))
8390 return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0);
8395 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
8396 SDValue N0 = N->getOperand(0);
8397 EVT VT = N->getValueType(0);
8399 // fold (ftrunc c1) -> ftrunc(c1)
8400 if (isConstantFPBuildVectorOrConstantFP(N0))
8401 return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0);
8406 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
8407 SDValue N0 = N->getOperand(0);
8408 EVT VT = N->getValueType(0);
8410 // fold (ffloor c1) -> ffloor(c1)
8411 if (isConstantFPBuildVectorOrConstantFP(N0))
8412 return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0);
8417 // FIXME: FNEG and FABS have a lot in common; refactor.
8418 SDValue DAGCombiner::visitFNEG(SDNode *N) {
8419 SDValue N0 = N->getOperand(0);
8420 EVT VT = N->getValueType(0);
8422 // Constant fold FNEG.
8423 if (isConstantFPBuildVectorOrConstantFP(N0))
8424 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
8426 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
8427 &DAG.getTarget().Options))
8428 return GetNegatedExpression(N0, DAG, LegalOperations);
8430 // Transform fneg(bitconvert(x)) -> bitconvert(x ^ sign) to avoid loading
8431 // constant pool values.
8432 if (!TLI.isFNegFree(VT) &&
8433 N0.getOpcode() == ISD::BITCAST &&
8434 N0.getNode()->hasOneUse()) {
8435 SDValue Int = N0.getOperand(0);
8436 EVT IntVT = Int.getValueType();
8437 if (IntVT.isInteger() && !IntVT.isVector()) {
8439 if (N0.getValueType().isVector()) {
8440 // For a vector, get a mask such as 0x80... per scalar element
8442 SignMask = APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
8443 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
8445 // For a scalar, just generate 0x80...
8446 SignMask = APInt::getSignBit(IntVT.getSizeInBits());
8448 Int = DAG.getNode(ISD::XOR, SDLoc(N0), IntVT, Int,
8449 DAG.getConstant(SignMask, IntVT));
8450 AddToWorklist(Int.getNode());
8451 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Int);
8455 // (fneg (fmul c, x)) -> (fmul -c, x)
8456 if (N0.getOpcode() == ISD::FMUL) {
8457 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
8459 APFloat CVal = CFP1->getValueAPF();
8461 if (Level >= AfterLegalizeDAG &&
8462 (TLI.isFPImmLegal(CVal, N->getValueType(0)) ||
8463 TLI.isOperationLegal(ISD::ConstantFP, N->getValueType(0))))
8465 ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
8466 DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0.getOperand(1)));
8473 SDValue DAGCombiner::visitFMINNUM(SDNode *N) {
8474 SDValue N0 = N->getOperand(0);
8475 SDValue N1 = N->getOperand(1);
8476 const ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8477 const ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8479 if (N0CFP && N1CFP) {
8480 const APFloat &C0 = N0CFP->getValueAPF();
8481 const APFloat &C1 = N1CFP->getValueAPF();
8482 return DAG.getConstantFP(minnum(C0, C1), N->getValueType(0));
8486 EVT VT = N->getValueType(0);
8487 // Canonicalize to constant on RHS.
8488 return DAG.getNode(ISD::FMINNUM, SDLoc(N), VT, N1, N0);
8494 SDValue DAGCombiner::visitFMAXNUM(SDNode *N) {
8495 SDValue N0 = N->getOperand(0);
8496 SDValue N1 = N->getOperand(1);
8497 const ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
8498 const ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
8500 if (N0CFP && N1CFP) {
8501 const APFloat &C0 = N0CFP->getValueAPF();
8502 const APFloat &C1 = N1CFP->getValueAPF();
8503 return DAG.getConstantFP(maxnum(C0, C1), N->getValueType(0));
8507 EVT VT = N->getValueType(0);
8508 // Canonicalize to constant on RHS.
8509 return DAG.getNode(ISD::FMAXNUM, SDLoc(N), VT, N1, N0);
8515 SDValue DAGCombiner::visitFABS(SDNode *N) {
8516 SDValue N0 = N->getOperand(0);
8517 EVT VT = N->getValueType(0);
8519 // fold (fabs c1) -> fabs(c1)
8520 if (isConstantFPBuildVectorOrConstantFP(N0))
8521 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
8523 // fold (fabs (fabs x)) -> (fabs x)
8524 if (N0.getOpcode() == ISD::FABS)
8525 return N->getOperand(0);
8527 // fold (fabs (fneg x)) -> (fabs x)
8528 // fold (fabs (fcopysign x, y)) -> (fabs x)
8529 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
8530 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0));
8532 // Transform fabs(bitconvert(x)) -> bitconvert(x & ~sign) to avoid loading
8533 // constant pool values.
8534 if (!TLI.isFAbsFree(VT) &&
8535 N0.getOpcode() == ISD::BITCAST &&
8536 N0.getNode()->hasOneUse()) {
8537 SDValue Int = N0.getOperand(0);
8538 EVT IntVT = Int.getValueType();
8539 if (IntVT.isInteger() && !IntVT.isVector()) {
8541 if (N0.getValueType().isVector()) {
8542 // For a vector, get a mask such as 0x7f... per scalar element
8544 SignMask = ~APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
8545 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
8547 // For a scalar, just generate 0x7f...
8548 SignMask = ~APInt::getSignBit(IntVT.getSizeInBits());
8550 Int = DAG.getNode(ISD::AND, SDLoc(N0), IntVT, Int,
8551 DAG.getConstant(SignMask, IntVT));
8552 AddToWorklist(Int.getNode());
8553 return DAG.getNode(ISD::BITCAST, SDLoc(N), N->getValueType(0), Int);
8560 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
8561 SDValue Chain = N->getOperand(0);
8562 SDValue N1 = N->getOperand(1);
8563 SDValue N2 = N->getOperand(2);
8565 // If N is a constant we could fold this into a fallthrough or unconditional
8566 // branch. However that doesn't happen very often in normal code, because
8567 // Instcombine/SimplifyCFG should have handled the available opportunities.
8568 // If we did this folding here, it would be necessary to update the
8569 // MachineBasicBlock CFG, which is awkward.
8571 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
8573 if (N1.getOpcode() == ISD::SETCC &&
8574 TLI.isOperationLegalOrCustom(ISD::BR_CC,
8575 N1.getOperand(0).getValueType())) {
8576 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
8577 Chain, N1.getOperand(2),
8578 N1.getOperand(0), N1.getOperand(1), N2);
8581 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
8582 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
8583 (N1.getOperand(0).hasOneUse() &&
8584 N1.getOperand(0).getOpcode() == ISD::SRL))) {
8585 SDNode *Trunc = nullptr;
8586 if (N1.getOpcode() == ISD::TRUNCATE) {
8587 // Look pass the truncate.
8588 Trunc = N1.getNode();
8589 N1 = N1.getOperand(0);
8592 // Match this pattern so that we can generate simpler code:
8595 // %b = and i32 %a, 2
8596 // %c = srl i32 %b, 1
8597 // brcond i32 %c ...
8602 // %b = and i32 %a, 2
8603 // %c = setcc eq %b, 0
8606 // This applies only when the AND constant value has one bit set and the
8607 // SRL constant is equal to the log2 of the AND constant. The back-end is
8608 // smart enough to convert the result into a TEST/JMP sequence.
8609 SDValue Op0 = N1.getOperand(0);
8610 SDValue Op1 = N1.getOperand(1);
8612 if (Op0.getOpcode() == ISD::AND &&
8613 Op1.getOpcode() == ISD::Constant) {
8614 SDValue AndOp1 = Op0.getOperand(1);
8616 if (AndOp1.getOpcode() == ISD::Constant) {
8617 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
8619 if (AndConst.isPowerOf2() &&
8620 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
8622 DAG.getSetCC(SDLoc(N),
8623 getSetCCResultType(Op0.getValueType()),
8624 Op0, DAG.getConstant(0, Op0.getValueType()),
8627 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, SDLoc(N),
8628 MVT::Other, Chain, SetCC, N2);
8629 // Don't add the new BRCond into the worklist or else SimplifySelectCC
8630 // will convert it back to (X & C1) >> C2.
8631 CombineTo(N, NewBRCond, false);
8632 // Truncate is dead.
8634 deleteAndRecombine(Trunc);
8635 // Replace the uses of SRL with SETCC
8636 WorklistRemover DeadNodes(*this);
8637 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
8638 deleteAndRecombine(N1.getNode());
8639 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8645 // Restore N1 if the above transformation doesn't match.
8646 N1 = N->getOperand(1);
8649 // Transform br(xor(x, y)) -> br(x != y)
8650 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
8651 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
8652 SDNode *TheXor = N1.getNode();
8653 SDValue Op0 = TheXor->getOperand(0);
8654 SDValue Op1 = TheXor->getOperand(1);
8655 if (Op0.getOpcode() == Op1.getOpcode()) {
8656 // Avoid missing important xor optimizations.
8657 SDValue Tmp = visitXOR(TheXor);
8658 if (Tmp.getNode()) {
8659 if (Tmp.getNode() != TheXor) {
8660 DEBUG(dbgs() << "\nReplacing.8 ";
8662 dbgs() << "\nWith: ";
8663 Tmp.getNode()->dump(&DAG);
8665 WorklistRemover DeadNodes(*this);
8666 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
8667 deleteAndRecombine(TheXor);
8668 return DAG.getNode(ISD::BRCOND, SDLoc(N),
8669 MVT::Other, Chain, Tmp, N2);
8672 // visitXOR has changed XOR's operands or replaced the XOR completely,
8674 return SDValue(N, 0);
8678 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
8680 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
8681 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
8682 Op0.getOpcode() == ISD::XOR) {
8683 TheXor = Op0.getNode();
8687 EVT SetCCVT = N1.getValueType();
8689 SetCCVT = getSetCCResultType(SetCCVT);
8690 SDValue SetCC = DAG.getSetCC(SDLoc(TheXor),
8693 Equal ? ISD::SETEQ : ISD::SETNE);
8694 // Replace the uses of XOR with SETCC
8695 WorklistRemover DeadNodes(*this);
8696 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
8697 deleteAndRecombine(N1.getNode());
8698 return DAG.getNode(ISD::BRCOND, SDLoc(N),
8699 MVT::Other, Chain, SetCC, N2);
8706 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
8708 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
8709 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
8710 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
8712 // If N is a constant we could fold this into a fallthrough or unconditional
8713 // branch. However that doesn't happen very often in normal code, because
8714 // Instcombine/SimplifyCFG should have handled the available opportunities.
8715 // If we did this folding here, it would be necessary to update the
8716 // MachineBasicBlock CFG, which is awkward.
8718 // Use SimplifySetCC to simplify SETCC's.
8719 SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()),
8720 CondLHS, CondRHS, CC->get(), SDLoc(N),
8722 if (Simp.getNode()) AddToWorklist(Simp.getNode());
8724 // fold to a simpler setcc
8725 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
8726 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
8727 N->getOperand(0), Simp.getOperand(2),
8728 Simp.getOperand(0), Simp.getOperand(1),
8734 /// Return true if 'Use' is a load or a store that uses N as its base pointer
8735 /// and that N may be folded in the load / store addressing mode.
8736 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
8738 const TargetLowering &TLI) {
8740 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
8741 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
8743 VT = Use->getValueType(0);
8744 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
8745 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
8747 VT = ST->getValue().getValueType();
8751 TargetLowering::AddrMode AM;
8752 if (N->getOpcode() == ISD::ADD) {
8753 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
8756 AM.BaseOffs = Offset->getSExtValue();
8760 } else if (N->getOpcode() == ISD::SUB) {
8761 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
8764 AM.BaseOffs = -Offset->getSExtValue();
8771 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
8774 /// Try turning a load/store into a pre-indexed load/store when the base
8775 /// pointer is an add or subtract and it has other uses besides the load/store.
8776 /// After the transformation, the new indexed load/store has effectively folded
8777 /// the add/subtract in and all of its other uses are redirected to the
8779 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
8780 if (Level < AfterLegalizeDAG)
8786 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8787 if (LD->isIndexed())
8789 VT = LD->getMemoryVT();
8790 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
8791 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
8793 Ptr = LD->getBasePtr();
8794 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8795 if (ST->isIndexed())
8797 VT = ST->getMemoryVT();
8798 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
8799 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
8801 Ptr = ST->getBasePtr();
8807 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
8808 // out. There is no reason to make this a preinc/predec.
8809 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
8810 Ptr.getNode()->hasOneUse())
8813 // Ask the target to do addressing mode selection.
8816 ISD::MemIndexedMode AM = ISD::UNINDEXED;
8817 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
8820 // Backends without true r+i pre-indexed forms may need to pass a
8821 // constant base with a variable offset so that constant coercion
8822 // will work with the patterns in canonical form.
8823 bool Swapped = false;
8824 if (isa<ConstantSDNode>(BasePtr)) {
8825 std::swap(BasePtr, Offset);
8829 // Don't create a indexed load / store with zero offset.
8830 if (isa<ConstantSDNode>(Offset) &&
8831 cast<ConstantSDNode>(Offset)->isNullValue())
8834 // Try turning it into a pre-indexed load / store except when:
8835 // 1) The new base ptr is a frame index.
8836 // 2) If N is a store and the new base ptr is either the same as or is a
8837 // predecessor of the value being stored.
8838 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
8839 // that would create a cycle.
8840 // 4) All uses are load / store ops that use it as old base ptr.
8842 // Check #1. Preinc'ing a frame index would require copying the stack pointer
8843 // (plus the implicit offset) to a register to preinc anyway.
8844 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
8849 SDValue Val = cast<StoreSDNode>(N)->getValue();
8850 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
8854 // If the offset is a constant, there may be other adds of constants that
8855 // can be folded with this one. We should do this to avoid having to keep
8856 // a copy of the original base pointer.
8857 SmallVector<SDNode *, 16> OtherUses;
8858 if (isa<ConstantSDNode>(Offset))
8859 for (SDNode *Use : BasePtr.getNode()->uses()) {
8860 if (Use == Ptr.getNode())
8863 if (Use->isPredecessorOf(N))
8866 if (Use->getOpcode() != ISD::ADD && Use->getOpcode() != ISD::SUB) {
8871 SDValue Op0 = Use->getOperand(0), Op1 = Use->getOperand(1);
8872 if (Op1.getNode() == BasePtr.getNode())
8873 std::swap(Op0, Op1);
8874 assert(Op0.getNode() == BasePtr.getNode() &&
8875 "Use of ADD/SUB but not an operand");
8877 if (!isa<ConstantSDNode>(Op1)) {
8882 // FIXME: In some cases, we can be smarter about this.
8883 if (Op1.getValueType() != Offset.getValueType()) {
8888 OtherUses.push_back(Use);
8892 std::swap(BasePtr, Offset);
8894 // Now check for #3 and #4.
8895 bool RealUse = false;
8897 // Caches for hasPredecessorHelper
8898 SmallPtrSet<const SDNode *, 32> Visited;
8899 SmallVector<const SDNode *, 16> Worklist;
8901 for (SDNode *Use : Ptr.getNode()->uses()) {
8904 if (N->hasPredecessorHelper(Use, Visited, Worklist))
8907 // If Ptr may be folded in addressing mode of other use, then it's
8908 // not profitable to do this transformation.
8909 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
8918 Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
8919 BasePtr, Offset, AM);
8921 Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
8922 BasePtr, Offset, AM);
8925 DEBUG(dbgs() << "\nReplacing.4 ";
8927 dbgs() << "\nWith: ";
8928 Result.getNode()->dump(&DAG);
8930 WorklistRemover DeadNodes(*this);
8932 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
8933 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
8935 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
8938 // Finally, since the node is now dead, remove it from the graph.
8939 deleteAndRecombine(N);
8942 std::swap(BasePtr, Offset);
8944 // Replace other uses of BasePtr that can be updated to use Ptr
8945 for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) {
8946 unsigned OffsetIdx = 1;
8947 if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode())
8949 assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() ==
8950 BasePtr.getNode() && "Expected BasePtr operand");
8952 // We need to replace ptr0 in the following expression:
8953 // x0 * offset0 + y0 * ptr0 = t0
8955 // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store)
8957 // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the
8958 // indexed load/store and the expresion that needs to be re-written.
8960 // Therefore, we have:
8961 // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1
8963 ConstantSDNode *CN =
8964 cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx));
8966 APInt Offset0 = CN->getAPIntValue();
8967 APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue();
8969 X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1;
8970 Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1;
8971 X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1;
8972 Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1;
8974 unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD;
8976 APInt CNV = Offset0;
8977 if (X0 < 0) CNV = -CNV;
8978 if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1;
8979 else CNV = CNV - Offset1;
8981 // We can now generate the new expression.
8982 SDValue NewOp1 = DAG.getConstant(CNV, CN->getValueType(0));
8983 SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0);
8985 SDValue NewUse = DAG.getNode(Opcode,
8986 SDLoc(OtherUses[i]),
8987 OtherUses[i]->getValueType(0), NewOp1, NewOp2);
8988 DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse);
8989 deleteAndRecombine(OtherUses[i]);
8992 // Replace the uses of Ptr with uses of the updated base value.
8993 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
8994 deleteAndRecombine(Ptr.getNode());
8999 /// Try to combine a load/store with a add/sub of the base pointer node into a
9000 /// post-indexed load/store. The transformation folded the add/subtract into the
9001 /// new indexed load/store effectively and all of its uses are redirected to the
9003 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
9004 if (Level < AfterLegalizeDAG)
9010 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
9011 if (LD->isIndexed())
9013 VT = LD->getMemoryVT();
9014 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
9015 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
9017 Ptr = LD->getBasePtr();
9018 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
9019 if (ST->isIndexed())
9021 VT = ST->getMemoryVT();
9022 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
9023 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
9025 Ptr = ST->getBasePtr();
9031 if (Ptr.getNode()->hasOneUse())
9034 for (SDNode *Op : Ptr.getNode()->uses()) {
9036 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
9041 ISD::MemIndexedMode AM = ISD::UNINDEXED;
9042 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
9043 // Don't create a indexed load / store with zero offset.
9044 if (isa<ConstantSDNode>(Offset) &&
9045 cast<ConstantSDNode>(Offset)->isNullValue())
9048 // Try turning it into a post-indexed load / store except when
9049 // 1) All uses are load / store ops that use it as base ptr (and
9050 // it may be folded as addressing mmode).
9051 // 2) Op must be independent of N, i.e. Op is neither a predecessor
9052 // nor a successor of N. Otherwise, if Op is folded that would
9055 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
9059 bool TryNext = false;
9060 for (SDNode *Use : BasePtr.getNode()->uses()) {
9061 if (Use == Ptr.getNode())
9064 // If all the uses are load / store addresses, then don't do the
9066 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
9067 bool RealUse = false;
9068 for (SDNode *UseUse : Use->uses()) {
9069 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
9084 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
9085 SDValue Result = isLoad
9086 ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
9087 BasePtr, Offset, AM)
9088 : DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
9089 BasePtr, Offset, AM);
9092 DEBUG(dbgs() << "\nReplacing.5 ";
9094 dbgs() << "\nWith: ";
9095 Result.getNode()->dump(&DAG);
9097 WorklistRemover DeadNodes(*this);
9099 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
9100 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
9102 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
9105 // Finally, since the node is now dead, remove it from the graph.
9106 deleteAndRecombine(N);
9108 // Replace the uses of Use with uses of the updated base value.
9109 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
9110 Result.getValue(isLoad ? 1 : 0));
9111 deleteAndRecombine(Op);
9120 /// \brief Return the base-pointer arithmetic from an indexed \p LD.
9121 SDValue DAGCombiner::SplitIndexingFromLoad(LoadSDNode *LD) {
9122 ISD::MemIndexedMode AM = LD->getAddressingMode();
9123 assert(AM != ISD::UNINDEXED);
9124 SDValue BP = LD->getOperand(1);
9125 SDValue Inc = LD->getOperand(2);
9127 // Some backends use TargetConstants for load offsets, but don't expect
9128 // TargetConstants in general ADD nodes. We can convert these constants into
9129 // regular Constants (if the constant is not opaque).
9130 assert((Inc.getOpcode() != ISD::TargetConstant ||
9131 !cast<ConstantSDNode>(Inc)->isOpaque()) &&
9132 "Cannot split out indexing using opaque target constants");
9133 if (Inc.getOpcode() == ISD::TargetConstant) {
9134 ConstantSDNode *ConstInc = cast<ConstantSDNode>(Inc);
9135 Inc = DAG.getConstant(*ConstInc->getConstantIntValue(),
9136 ConstInc->getValueType(0));
9140 (AM == ISD::PRE_INC || AM == ISD::POST_INC ? ISD::ADD : ISD::SUB);
9141 return DAG.getNode(Opc, SDLoc(LD), BP.getSimpleValueType(), BP, Inc);
9144 SDValue DAGCombiner::visitLOAD(SDNode *N) {
9145 LoadSDNode *LD = cast<LoadSDNode>(N);
9146 SDValue Chain = LD->getChain();
9147 SDValue Ptr = LD->getBasePtr();
9149 // If load is not volatile and there are no uses of the loaded value (and
9150 // the updated indexed value in case of indexed loads), change uses of the
9151 // chain value into uses of the chain input (i.e. delete the dead load).
9152 if (!LD->isVolatile()) {
9153 if (N->getValueType(1) == MVT::Other) {
9155 if (!N->hasAnyUseOfValue(0)) {
9156 // It's not safe to use the two value CombineTo variant here. e.g.
9157 // v1, chain2 = load chain1, loc
9158 // v2, chain3 = load chain2, loc
9160 // Now we replace use of chain2 with chain1. This makes the second load
9161 // isomorphic to the one we are deleting, and thus makes this load live.
9162 DEBUG(dbgs() << "\nReplacing.6 ";
9164 dbgs() << "\nWith chain: ";
9165 Chain.getNode()->dump(&DAG);
9167 WorklistRemover DeadNodes(*this);
9168 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
9171 deleteAndRecombine(N);
9173 return SDValue(N, 0); // Return N so it doesn't get rechecked!
9177 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
9179 // If this load has an opaque TargetConstant offset, then we cannot split
9180 // the indexing into an add/sub directly (that TargetConstant may not be
9181 // valid for a different type of node, and we cannot convert an opaque
9182 // target constant into a regular constant).
9183 bool HasOTCInc = LD->getOperand(2).getOpcode() == ISD::TargetConstant &&
9184 cast<ConstantSDNode>(LD->getOperand(2))->isOpaque();
9186 if (!N->hasAnyUseOfValue(0) &&
9187 ((MaySplitLoadIndex && !HasOTCInc) || !N->hasAnyUseOfValue(1))) {
9188 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
9190 if (N->hasAnyUseOfValue(1) && MaySplitLoadIndex && !HasOTCInc) {
9191 Index = SplitIndexingFromLoad(LD);
9192 // Try to fold the base pointer arithmetic into subsequent loads and
9194 AddUsersToWorklist(N);
9196 Index = DAG.getUNDEF(N->getValueType(1));
9197 DEBUG(dbgs() << "\nReplacing.7 ";
9199 dbgs() << "\nWith: ";
9200 Undef.getNode()->dump(&DAG);
9201 dbgs() << " and 2 other values\n");
9202 WorklistRemover DeadNodes(*this);
9203 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
9204 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Index);
9205 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
9206 deleteAndRecombine(N);
9207 return SDValue(N, 0); // Return N so it doesn't get rechecked!
9212 // If this load is directly stored, replace the load value with the stored
9214 // TODO: Handle store large -> read small portion.
9215 // TODO: Handle TRUNCSTORE/LOADEXT
9216 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
9217 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
9218 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
9219 if (PrevST->getBasePtr() == Ptr &&
9220 PrevST->getValue().getValueType() == N->getValueType(0))
9221 return CombineTo(N, Chain.getOperand(1), Chain);
9225 // Try to infer better alignment information than the load already has.
9226 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
9227 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
9228 if (Align > LD->getMemOperand()->getBaseAlignment()) {
9230 DAG.getExtLoad(LD->getExtensionType(), SDLoc(N),
9231 LD->getValueType(0),
9232 Chain, Ptr, LD->getPointerInfo(),
9234 LD->isVolatile(), LD->isNonTemporal(),
9235 LD->isInvariant(), Align, LD->getAAInfo());
9236 if (NewLoad.getNode() != N)
9237 return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true);
9242 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
9243 : DAG.getSubtarget().useAA();
9245 if (CombinerAAOnlyFunc.getNumOccurrences() &&
9246 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
9249 if (UseAA && LD->isUnindexed()) {
9250 // Walk up chain skipping non-aliasing memory nodes.
9251 SDValue BetterChain = FindBetterChain(N, Chain);
9253 // If there is a better chain.
9254 if (Chain != BetterChain) {
9257 // Replace the chain to void dependency.
9258 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
9259 ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD),
9260 BetterChain, Ptr, LD->getMemOperand());
9262 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD),
9263 LD->getValueType(0),
9264 BetterChain, Ptr, LD->getMemoryVT(),
9265 LD->getMemOperand());
9268 // Create token factor to keep old chain connected.
9269 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
9270 MVT::Other, Chain, ReplLoad.getValue(1));
9272 // Make sure the new and old chains are cleaned up.
9273 AddToWorklist(Token.getNode());
9275 // Replace uses with load result and token factor. Don't add users
9277 return CombineTo(N, ReplLoad.getValue(0), Token, false);
9281 // Try transforming N to an indexed load.
9282 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
9283 return SDValue(N, 0);
9285 // Try to slice up N to more direct loads if the slices are mapped to
9286 // different register banks or pairing can take place.
9288 return SDValue(N, 0);
9294 /// \brief Helper structure used to slice a load in smaller loads.
9295 /// Basically a slice is obtained from the following sequence:
9296 /// Origin = load Ty1, Base
9297 /// Shift = srl Ty1 Origin, CstTy Amount
9298 /// Inst = trunc Shift to Ty2
9300 /// Then, it will be rewriten into:
9301 /// Slice = load SliceTy, Base + SliceOffset
9302 /// [Inst = zext Slice to Ty2], only if SliceTy <> Ty2
9304 /// SliceTy is deduced from the number of bits that are actually used to
9306 struct LoadedSlice {
9307 /// \brief Helper structure used to compute the cost of a slice.
9309 /// Are we optimizing for code size.
9314 unsigned CrossRegisterBanksCopies;
9318 Cost(bool ForCodeSize = false)
9319 : ForCodeSize(ForCodeSize), Loads(0), Truncates(0),
9320 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {}
9322 /// \brief Get the cost of one isolated slice.
9323 Cost(const LoadedSlice &LS, bool ForCodeSize = false)
9324 : ForCodeSize(ForCodeSize), Loads(1), Truncates(0),
9325 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {
9326 EVT TruncType = LS.Inst->getValueType(0);
9327 EVT LoadedType = LS.getLoadedType();
9328 if (TruncType != LoadedType &&
9329 !LS.DAG->getTargetLoweringInfo().isZExtFree(LoadedType, TruncType))
9333 /// \brief Account for slicing gain in the current cost.
9334 /// Slicing provide a few gains like removing a shift or a
9335 /// truncate. This method allows to grow the cost of the original
9336 /// load with the gain from this slice.
9337 void addSliceGain(const LoadedSlice &LS) {
9338 // Each slice saves a truncate.
9339 const TargetLowering &TLI = LS.DAG->getTargetLoweringInfo();
9340 if (!TLI.isTruncateFree(LS.Inst->getValueType(0),
9341 LS.Inst->getOperand(0).getValueType()))
9343 // If there is a shift amount, this slice gets rid of it.
9346 // If this slice can merge a cross register bank copy, account for it.
9347 if (LS.canMergeExpensiveCrossRegisterBankCopy())
9348 ++CrossRegisterBanksCopies;
9351 Cost &operator+=(const Cost &RHS) {
9353 Truncates += RHS.Truncates;
9354 CrossRegisterBanksCopies += RHS.CrossRegisterBanksCopies;
9360 bool operator==(const Cost &RHS) const {
9361 return Loads == RHS.Loads && Truncates == RHS.Truncates &&
9362 CrossRegisterBanksCopies == RHS.CrossRegisterBanksCopies &&
9363 ZExts == RHS.ZExts && Shift == RHS.Shift;
9366 bool operator!=(const Cost &RHS) const { return !(*this == RHS); }
9368 bool operator<(const Cost &RHS) const {
9369 // Assume cross register banks copies are as expensive as loads.
9370 // FIXME: Do we want some more target hooks?
9371 unsigned ExpensiveOpsLHS = Loads + CrossRegisterBanksCopies;
9372 unsigned ExpensiveOpsRHS = RHS.Loads + RHS.CrossRegisterBanksCopies;
9373 // Unless we are optimizing for code size, consider the
9374 // expensive operation first.
9375 if (!ForCodeSize && ExpensiveOpsLHS != ExpensiveOpsRHS)
9376 return ExpensiveOpsLHS < ExpensiveOpsRHS;
9377 return (Truncates + ZExts + Shift + ExpensiveOpsLHS) <
9378 (RHS.Truncates + RHS.ZExts + RHS.Shift + ExpensiveOpsRHS);
9381 bool operator>(const Cost &RHS) const { return RHS < *this; }
9383 bool operator<=(const Cost &RHS) const { return !(RHS < *this); }
9385 bool operator>=(const Cost &RHS) const { return !(*this < RHS); }
9387 // The last instruction that represent the slice. This should be a
9388 // truncate instruction.
9390 // The original load instruction.
9392 // The right shift amount in bits from the original load.
9394 // The DAG from which Origin came from.
9395 // This is used to get some contextual information about legal types, etc.
9398 LoadedSlice(SDNode *Inst = nullptr, LoadSDNode *Origin = nullptr,
9399 unsigned Shift = 0, SelectionDAG *DAG = nullptr)
9400 : Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {}
9402 /// \brief Get the bits used in a chunk of bits \p BitWidth large.
9403 /// \return Result is \p BitWidth and has used bits set to 1 and
9404 /// not used bits set to 0.
9405 APInt getUsedBits() const {
9406 // Reproduce the trunc(lshr) sequence:
9407 // - Start from the truncated value.
9408 // - Zero extend to the desired bit width.
9410 assert(Origin && "No original load to compare against.");
9411 unsigned BitWidth = Origin->getValueSizeInBits(0);
9412 assert(Inst && "This slice is not bound to an instruction");
9413 assert(Inst->getValueSizeInBits(0) <= BitWidth &&
9414 "Extracted slice is bigger than the whole type!");
9415 APInt UsedBits(Inst->getValueSizeInBits(0), 0);
9416 UsedBits.setAllBits();
9417 UsedBits = UsedBits.zext(BitWidth);
9422 /// \brief Get the size of the slice to be loaded in bytes.
9423 unsigned getLoadedSize() const {
9424 unsigned SliceSize = getUsedBits().countPopulation();
9425 assert(!(SliceSize & 0x7) && "Size is not a multiple of a byte.");
9426 return SliceSize / 8;
9429 /// \brief Get the type that will be loaded for this slice.
9430 /// Note: This may not be the final type for the slice.
9431 EVT getLoadedType() const {
9432 assert(DAG && "Missing context");
9433 LLVMContext &Ctxt = *DAG->getContext();
9434 return EVT::getIntegerVT(Ctxt, getLoadedSize() * 8);
9437 /// \brief Get the alignment of the load used for this slice.
9438 unsigned getAlignment() const {
9439 unsigned Alignment = Origin->getAlignment();
9440 unsigned Offset = getOffsetFromBase();
9442 Alignment = MinAlign(Alignment, Alignment + Offset);
9446 /// \brief Check if this slice can be rewritten with legal operations.
9447 bool isLegal() const {
9448 // An invalid slice is not legal.
9449 if (!Origin || !Inst || !DAG)
9452 // Offsets are for indexed load only, we do not handle that.
9453 if (Origin->getOffset().getOpcode() != ISD::UNDEF)
9456 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
9458 // Check that the type is legal.
9459 EVT SliceType = getLoadedType();
9460 if (!TLI.isTypeLegal(SliceType))
9463 // Check that the load is legal for this type.
9464 if (!TLI.isOperationLegal(ISD::LOAD, SliceType))
9467 // Check that the offset can be computed.
9468 // 1. Check its type.
9469 EVT PtrType = Origin->getBasePtr().getValueType();
9470 if (PtrType == MVT::Untyped || PtrType.isExtended())
9473 // 2. Check that it fits in the immediate.
9474 if (!TLI.isLegalAddImmediate(getOffsetFromBase()))
9477 // 3. Check that the computation is legal.
9478 if (!TLI.isOperationLegal(ISD::ADD, PtrType))
9481 // Check that the zext is legal if it needs one.
9482 EVT TruncateType = Inst->getValueType(0);
9483 if (TruncateType != SliceType &&
9484 !TLI.isOperationLegal(ISD::ZERO_EXTEND, TruncateType))
9490 /// \brief Get the offset in bytes of this slice in the original chunk of
9492 /// \pre DAG != nullptr.
9493 uint64_t getOffsetFromBase() const {
9494 assert(DAG && "Missing context.");
9496 DAG->getTargetLoweringInfo().getDataLayout()->isBigEndian();
9497 assert(!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported.");
9498 uint64_t Offset = Shift / 8;
9499 unsigned TySizeInBytes = Origin->getValueSizeInBits(0) / 8;
9500 assert(!(Origin->getValueSizeInBits(0) & 0x7) &&
9501 "The size of the original loaded type is not a multiple of a"
9503 // If Offset is bigger than TySizeInBytes, it means we are loading all
9504 // zeros. This should have been optimized before in the process.
9505 assert(TySizeInBytes > Offset &&
9506 "Invalid shift amount for given loaded size");
9508 Offset = TySizeInBytes - Offset - getLoadedSize();
9512 /// \brief Generate the sequence of instructions to load the slice
9513 /// represented by this object and redirect the uses of this slice to
9514 /// this new sequence of instructions.
9515 /// \pre this->Inst && this->Origin are valid Instructions and this
9516 /// object passed the legal check: LoadedSlice::isLegal returned true.
9517 /// \return The last instruction of the sequence used to load the slice.
9518 SDValue loadSlice() const {
9519 assert(Inst && Origin && "Unable to replace a non-existing slice.");
9520 const SDValue &OldBaseAddr = Origin->getBasePtr();
9521 SDValue BaseAddr = OldBaseAddr;
9522 // Get the offset in that chunk of bytes w.r.t. the endianess.
9523 int64_t Offset = static_cast<int64_t>(getOffsetFromBase());
9524 assert(Offset >= 0 && "Offset too big to fit in int64_t!");
9526 // BaseAddr = BaseAddr + Offset.
9527 EVT ArithType = BaseAddr.getValueType();
9528 BaseAddr = DAG->getNode(ISD::ADD, SDLoc(Origin), ArithType, BaseAddr,
9529 DAG->getConstant(Offset, ArithType));
9532 // Create the type of the loaded slice according to its size.
9533 EVT SliceType = getLoadedType();
9535 // Create the load for the slice.
9536 SDValue LastInst = DAG->getLoad(
9537 SliceType, SDLoc(Origin), Origin->getChain(), BaseAddr,
9538 Origin->getPointerInfo().getWithOffset(Offset), Origin->isVolatile(),
9539 Origin->isNonTemporal(), Origin->isInvariant(), getAlignment());
9540 // If the final type is not the same as the loaded type, this means that
9541 // we have to pad with zero. Create a zero extend for that.
9542 EVT FinalType = Inst->getValueType(0);
9543 if (SliceType != FinalType)
9545 DAG->getNode(ISD::ZERO_EXTEND, SDLoc(LastInst), FinalType, LastInst);
9549 /// \brief Check if this slice can be merged with an expensive cross register
9550 /// bank copy. E.g.,
9552 /// f = bitcast i32 i to float
9553 bool canMergeExpensiveCrossRegisterBankCopy() const {
9554 if (!Inst || !Inst->hasOneUse())
9556 SDNode *Use = *Inst->use_begin();
9557 if (Use->getOpcode() != ISD::BITCAST)
9559 assert(DAG && "Missing context");
9560 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
9561 EVT ResVT = Use->getValueType(0);
9562 const TargetRegisterClass *ResRC = TLI.getRegClassFor(ResVT.getSimpleVT());
9563 const TargetRegisterClass *ArgRC =
9564 TLI.getRegClassFor(Use->getOperand(0).getValueType().getSimpleVT());
9565 if (ArgRC == ResRC || !TLI.isOperationLegal(ISD::LOAD, ResVT))
9568 // At this point, we know that we perform a cross-register-bank copy.
9569 // Check if it is expensive.
9570 const TargetRegisterInfo *TRI = DAG->getSubtarget().getRegisterInfo();
9571 // Assume bitcasts are cheap, unless both register classes do not
9572 // explicitly share a common sub class.
9573 if (!TRI || TRI->getCommonSubClass(ArgRC, ResRC))
9576 // Check if it will be merged with the load.
9577 // 1. Check the alignment constraint.
9578 unsigned RequiredAlignment = TLI.getDataLayout()->getABITypeAlignment(
9579 ResVT.getTypeForEVT(*DAG->getContext()));
9581 if (RequiredAlignment > getAlignment())
9584 // 2. Check that the load is a legal operation for that type.
9585 if (!TLI.isOperationLegal(ISD::LOAD, ResVT))
9588 // 3. Check that we do not have a zext in the way.
9589 if (Inst->getValueType(0) != getLoadedType())
9597 /// \brief Check that all bits set in \p UsedBits form a dense region, i.e.,
9598 /// \p UsedBits looks like 0..0 1..1 0..0.
9599 static bool areUsedBitsDense(const APInt &UsedBits) {
9600 // If all the bits are one, this is dense!
9601 if (UsedBits.isAllOnesValue())
9604 // Get rid of the unused bits on the right.
9605 APInt NarrowedUsedBits = UsedBits.lshr(UsedBits.countTrailingZeros());
9606 // Get rid of the unused bits on the left.
9607 if (NarrowedUsedBits.countLeadingZeros())
9608 NarrowedUsedBits = NarrowedUsedBits.trunc(NarrowedUsedBits.getActiveBits());
9609 // Check that the chunk of bits is completely used.
9610 return NarrowedUsedBits.isAllOnesValue();
9613 /// \brief Check whether or not \p First and \p Second are next to each other
9614 /// in memory. This means that there is no hole between the bits loaded
9615 /// by \p First and the bits loaded by \p Second.
9616 static bool areSlicesNextToEachOther(const LoadedSlice &First,
9617 const LoadedSlice &Second) {
9618 assert(First.Origin == Second.Origin && First.Origin &&
9619 "Unable to match different memory origins.");
9620 APInt UsedBits = First.getUsedBits();
9621 assert((UsedBits & Second.getUsedBits()) == 0 &&
9622 "Slices are not supposed to overlap.");
9623 UsedBits |= Second.getUsedBits();
9624 return areUsedBitsDense(UsedBits);
9627 /// \brief Adjust the \p GlobalLSCost according to the target
9628 /// paring capabilities and the layout of the slices.
9629 /// \pre \p GlobalLSCost should account for at least as many loads as
9630 /// there is in the slices in \p LoadedSlices.
9631 static void adjustCostForPairing(SmallVectorImpl<LoadedSlice> &LoadedSlices,
9632 LoadedSlice::Cost &GlobalLSCost) {
9633 unsigned NumberOfSlices = LoadedSlices.size();
9634 // If there is less than 2 elements, no pairing is possible.
9635 if (NumberOfSlices < 2)
9638 // Sort the slices so that elements that are likely to be next to each
9639 // other in memory are next to each other in the list.
9640 std::sort(LoadedSlices.begin(), LoadedSlices.end(),
9641 [](const LoadedSlice &LHS, const LoadedSlice &RHS) {
9642 assert(LHS.Origin == RHS.Origin && "Different bases not implemented.");
9643 return LHS.getOffsetFromBase() < RHS.getOffsetFromBase();
9645 const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo();
9646 // First (resp. Second) is the first (resp. Second) potentially candidate
9647 // to be placed in a paired load.
9648 const LoadedSlice *First = nullptr;
9649 const LoadedSlice *Second = nullptr;
9650 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice,
9651 // Set the beginning of the pair.
9654 Second = &LoadedSlices[CurrSlice];
9656 // If First is NULL, it means we start a new pair.
9657 // Get to the next slice.
9661 EVT LoadedType = First->getLoadedType();
9663 // If the types of the slices are different, we cannot pair them.
9664 if (LoadedType != Second->getLoadedType())
9667 // Check if the target supplies paired loads for this type.
9668 unsigned RequiredAlignment = 0;
9669 if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) {
9670 // move to the next pair, this type is hopeless.
9674 // Check if we meet the alignment requirement.
9675 if (RequiredAlignment > First->getAlignment())
9678 // Check that both loads are next to each other in memory.
9679 if (!areSlicesNextToEachOther(*First, *Second))
9682 assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!");
9683 --GlobalLSCost.Loads;
9684 // Move to the next pair.
9689 /// \brief Check the profitability of all involved LoadedSlice.
9690 /// Currently, it is considered profitable if there is exactly two
9691 /// involved slices (1) which are (2) next to each other in memory, and
9692 /// whose cost (\see LoadedSlice::Cost) is smaller than the original load (3).
9694 /// Note: The order of the elements in \p LoadedSlices may be modified, but not
9695 /// the elements themselves.
9697 /// FIXME: When the cost model will be mature enough, we can relax
9698 /// constraints (1) and (2).
9699 static bool isSlicingProfitable(SmallVectorImpl<LoadedSlice> &LoadedSlices,
9700 const APInt &UsedBits, bool ForCodeSize) {
9701 unsigned NumberOfSlices = LoadedSlices.size();
9702 if (StressLoadSlicing)
9703 return NumberOfSlices > 1;
9706 if (NumberOfSlices != 2)
9710 if (!areUsedBitsDense(UsedBits))
9714 LoadedSlice::Cost OrigCost(ForCodeSize), GlobalSlicingCost(ForCodeSize);
9715 // The original code has one big load.
9717 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice) {
9718 const LoadedSlice &LS = LoadedSlices[CurrSlice];
9719 // Accumulate the cost of all the slices.
9720 LoadedSlice::Cost SliceCost(LS, ForCodeSize);
9721 GlobalSlicingCost += SliceCost;
9723 // Account as cost in the original configuration the gain obtained
9724 // with the current slices.
9725 OrigCost.addSliceGain(LS);
9728 // If the target supports paired load, adjust the cost accordingly.
9729 adjustCostForPairing(LoadedSlices, GlobalSlicingCost);
9730 return OrigCost > GlobalSlicingCost;
9733 /// \brief If the given load, \p LI, is used only by trunc or trunc(lshr)
9734 /// operations, split it in the various pieces being extracted.
9736 /// This sort of thing is introduced by SROA.
9737 /// This slicing takes care not to insert overlapping loads.
9738 /// \pre LI is a simple load (i.e., not an atomic or volatile load).
9739 bool DAGCombiner::SliceUpLoad(SDNode *N) {
9740 if (Level < AfterLegalizeDAG)
9743 LoadSDNode *LD = cast<LoadSDNode>(N);
9744 if (LD->isVolatile() || !ISD::isNormalLoad(LD) ||
9745 !LD->getValueType(0).isInteger())
9748 // Keep track of already used bits to detect overlapping values.
9749 // In that case, we will just abort the transformation.
9750 APInt UsedBits(LD->getValueSizeInBits(0), 0);
9752 SmallVector<LoadedSlice, 4> LoadedSlices;
9754 // Check if this load is used as several smaller chunks of bits.
9755 // Basically, look for uses in trunc or trunc(lshr) and record a new chain
9756 // of computation for each trunc.
9757 for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end();
9758 UI != UIEnd; ++UI) {
9759 // Skip the uses of the chain.
9760 if (UI.getUse().getResNo() != 0)
9766 // Check if this is a trunc(lshr).
9767 if (User->getOpcode() == ISD::SRL && User->hasOneUse() &&
9768 isa<ConstantSDNode>(User->getOperand(1))) {
9769 Shift = cast<ConstantSDNode>(User->getOperand(1))->getZExtValue();
9770 User = *User->use_begin();
9773 // At this point, User is a Truncate, iff we encountered, trunc or
9775 if (User->getOpcode() != ISD::TRUNCATE)
9778 // The width of the type must be a power of 2 and greater than 8-bits.
9779 // Otherwise the load cannot be represented in LLVM IR.
9780 // Moreover, if we shifted with a non-8-bits multiple, the slice
9781 // will be across several bytes. We do not support that.
9782 unsigned Width = User->getValueSizeInBits(0);
9783 if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7))
9786 // Build the slice for this chain of computations.
9787 LoadedSlice LS(User, LD, Shift, &DAG);
9788 APInt CurrentUsedBits = LS.getUsedBits();
9790 // Check if this slice overlaps with another.
9791 if ((CurrentUsedBits & UsedBits) != 0)
9793 // Update the bits used globally.
9794 UsedBits |= CurrentUsedBits;
9796 // Check if the new slice would be legal.
9800 // Record the slice.
9801 LoadedSlices.push_back(LS);
9804 // Abort slicing if it does not seem to be profitable.
9805 if (!isSlicingProfitable(LoadedSlices, UsedBits, ForCodeSize))
9810 // Rewrite each chain to use an independent load.
9811 // By construction, each chain can be represented by a unique load.
9813 // Prepare the argument for the new token factor for all the slices.
9814 SmallVector<SDValue, 8> ArgChains;
9815 for (SmallVectorImpl<LoadedSlice>::const_iterator
9816 LSIt = LoadedSlices.begin(),
9817 LSItEnd = LoadedSlices.end();
9818 LSIt != LSItEnd; ++LSIt) {
9819 SDValue SliceInst = LSIt->loadSlice();
9820 CombineTo(LSIt->Inst, SliceInst, true);
9821 if (SliceInst.getNode()->getOpcode() != ISD::LOAD)
9822 SliceInst = SliceInst.getOperand(0);
9823 assert(SliceInst->getOpcode() == ISD::LOAD &&
9824 "It takes more than a zext to get to the loaded slice!!");
9825 ArgChains.push_back(SliceInst.getValue(1));
9828 SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other,
9830 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
9834 /// Check to see if V is (and load (ptr), imm), where the load is having
9835 /// specific bytes cleared out. If so, return the byte size being masked out
9836 /// and the shift amount.
9837 static std::pair<unsigned, unsigned>
9838 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
9839 std::pair<unsigned, unsigned> Result(0, 0);
9841 // Check for the structure we're looking for.
9842 if (V->getOpcode() != ISD::AND ||
9843 !isa<ConstantSDNode>(V->getOperand(1)) ||
9844 !ISD::isNormalLoad(V->getOperand(0).getNode()))
9847 // Check the chain and pointer.
9848 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
9849 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
9851 // The store should be chained directly to the load or be an operand of a
9853 if (LD == Chain.getNode())
9855 else if (Chain->getOpcode() != ISD::TokenFactor)
9856 return Result; // Fail.
9859 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
9860 if (Chain->getOperand(i).getNode() == LD) {
9864 if (!isOk) return Result;
9867 // This only handles simple types.
9868 if (V.getValueType() != MVT::i16 &&
9869 V.getValueType() != MVT::i32 &&
9870 V.getValueType() != MVT::i64)
9873 // Check the constant mask. Invert it so that the bits being masked out are
9874 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
9875 // follow the sign bit for uniformity.
9876 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
9877 unsigned NotMaskLZ = countLeadingZeros(NotMask);
9878 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
9879 unsigned NotMaskTZ = countTrailingZeros(NotMask);
9880 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
9881 if (NotMaskLZ == 64) return Result; // All zero mask.
9883 // See if we have a continuous run of bits. If so, we have 0*1+0*
9884 if (countTrailingOnes(NotMask >> NotMaskTZ) + NotMaskTZ + NotMaskLZ != 64)
9887 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
9888 if (V.getValueType() != MVT::i64 && NotMaskLZ)
9889 NotMaskLZ -= 64-V.getValueSizeInBits();
9891 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
9892 switch (MaskedBytes) {
9896 default: return Result; // All one mask, or 5-byte mask.
9899 // Verify that the first bit starts at a multiple of mask so that the access
9900 // is aligned the same as the access width.
9901 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
9903 Result.first = MaskedBytes;
9904 Result.second = NotMaskTZ/8;
9909 /// Check to see if IVal is something that provides a value as specified by
9910 /// MaskInfo. If so, replace the specified store with a narrower store of
9913 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
9914 SDValue IVal, StoreSDNode *St,
9916 unsigned NumBytes = MaskInfo.first;
9917 unsigned ByteShift = MaskInfo.second;
9918 SelectionDAG &DAG = DC->getDAG();
9920 // Check to see if IVal is all zeros in the part being masked in by the 'or'
9921 // that uses this. If not, this is not a replacement.
9922 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
9923 ByteShift*8, (ByteShift+NumBytes)*8);
9924 if (!DAG.MaskedValueIsZero(IVal, Mask)) return nullptr;
9926 // Check that it is legal on the target to do this. It is legal if the new
9927 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
9929 MVT VT = MVT::getIntegerVT(NumBytes*8);
9930 if (!DC->isTypeLegal(VT))
9933 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
9934 // shifted by ByteShift and truncated down to NumBytes.
9936 IVal = DAG.getNode(ISD::SRL, SDLoc(IVal), IVal.getValueType(), IVal,
9937 DAG.getConstant(ByteShift*8,
9938 DC->getShiftAmountTy(IVal.getValueType())));
9940 // Figure out the offset for the store and the alignment of the access.
9942 unsigned NewAlign = St->getAlignment();
9944 if (DAG.getTargetLoweringInfo().isLittleEndian())
9945 StOffset = ByteShift;
9947 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
9949 SDValue Ptr = St->getBasePtr();
9951 Ptr = DAG.getNode(ISD::ADD, SDLoc(IVal), Ptr.getValueType(),
9952 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
9953 NewAlign = MinAlign(NewAlign, StOffset);
9956 // Truncate down to the new size.
9957 IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal);
9960 return DAG.getStore(St->getChain(), SDLoc(St), IVal, Ptr,
9961 St->getPointerInfo().getWithOffset(StOffset),
9962 false, false, NewAlign).getNode();
9966 /// Look for sequence of load / op / store where op is one of 'or', 'xor', and
9967 /// 'and' of immediates. If 'op' is only touching some of the loaded bits, try
9968 /// narrowing the load and store if it would end up being a win for performance
9970 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
9971 StoreSDNode *ST = cast<StoreSDNode>(N);
9972 if (ST->isVolatile())
9975 SDValue Chain = ST->getChain();
9976 SDValue Value = ST->getValue();
9977 SDValue Ptr = ST->getBasePtr();
9978 EVT VT = Value.getValueType();
9980 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
9983 unsigned Opc = Value.getOpcode();
9985 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
9986 // is a byte mask indicating a consecutive number of bytes, check to see if
9987 // Y is known to provide just those bytes. If so, we try to replace the
9988 // load + replace + store sequence with a single (narrower) store, which makes
9990 if (Opc == ISD::OR) {
9991 std::pair<unsigned, unsigned> MaskedLoad;
9992 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
9993 if (MaskedLoad.first)
9994 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
9995 Value.getOperand(1), ST,this))
9996 return SDValue(NewST, 0);
9998 // Or is commutative, so try swapping X and Y.
9999 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
10000 if (MaskedLoad.first)
10001 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
10002 Value.getOperand(0), ST,this))
10003 return SDValue(NewST, 0);
10006 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
10007 Value.getOperand(1).getOpcode() != ISD::Constant)
10010 SDValue N0 = Value.getOperand(0);
10011 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
10012 Chain == SDValue(N0.getNode(), 1)) {
10013 LoadSDNode *LD = cast<LoadSDNode>(N0);
10014 if (LD->getBasePtr() != Ptr ||
10015 LD->getPointerInfo().getAddrSpace() !=
10016 ST->getPointerInfo().getAddrSpace())
10019 // Find the type to narrow it the load / op / store to.
10020 SDValue N1 = Value.getOperand(1);
10021 unsigned BitWidth = N1.getValueSizeInBits();
10022 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
10023 if (Opc == ISD::AND)
10024 Imm ^= APInt::getAllOnesValue(BitWidth);
10025 if (Imm == 0 || Imm.isAllOnesValue())
10027 unsigned ShAmt = Imm.countTrailingZeros();
10028 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
10029 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
10030 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
10031 // The narrowing should be profitable, the load/store operation should be
10032 // legal (or custom) and the store size should be equal to the NewVT width.
10033 while (NewBW < BitWidth &&
10034 (NewVT.getStoreSizeInBits() != NewBW ||
10035 !TLI.isOperationLegalOrCustom(Opc, NewVT) ||
10036 !TLI.isNarrowingProfitable(VT, NewVT))) {
10037 NewBW = NextPowerOf2(NewBW);
10038 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
10040 if (NewBW >= BitWidth)
10043 // If the lsb changed does not start at the type bitwidth boundary,
10044 // start at the previous one.
10046 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
10047 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt,
10048 std::min(BitWidth, ShAmt + NewBW));
10049 if ((Imm & Mask) == Imm) {
10050 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
10051 if (Opc == ISD::AND)
10052 NewImm ^= APInt::getAllOnesValue(NewBW);
10053 uint64_t PtrOff = ShAmt / 8;
10054 // For big endian targets, we need to adjust the offset to the pointer to
10055 // load the correct bytes.
10056 if (TLI.isBigEndian())
10057 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
10059 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
10060 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
10061 if (NewAlign < TLI.getDataLayout()->getABITypeAlignment(NewVTTy))
10064 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD),
10065 Ptr.getValueType(), Ptr,
10066 DAG.getConstant(PtrOff, Ptr.getValueType()));
10067 SDValue NewLD = DAG.getLoad(NewVT, SDLoc(N0),
10068 LD->getChain(), NewPtr,
10069 LD->getPointerInfo().getWithOffset(PtrOff),
10070 LD->isVolatile(), LD->isNonTemporal(),
10071 LD->isInvariant(), NewAlign,
10073 SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD,
10074 DAG.getConstant(NewImm, NewVT));
10075 SDValue NewST = DAG.getStore(Chain, SDLoc(N),
10077 ST->getPointerInfo().getWithOffset(PtrOff),
10078 false, false, NewAlign);
10080 AddToWorklist(NewPtr.getNode());
10081 AddToWorklist(NewLD.getNode());
10082 AddToWorklist(NewVal.getNode());
10083 WorklistRemover DeadNodes(*this);
10084 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
10093 /// For a given floating point load / store pair, if the load value isn't used
10094 /// by any other operations, then consider transforming the pair to integer
10095 /// load / store operations if the target deems the transformation profitable.
10096 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
10097 StoreSDNode *ST = cast<StoreSDNode>(N);
10098 SDValue Chain = ST->getChain();
10099 SDValue Value = ST->getValue();
10100 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
10101 Value.hasOneUse() &&
10102 Chain == SDValue(Value.getNode(), 1)) {
10103 LoadSDNode *LD = cast<LoadSDNode>(Value);
10104 EVT VT = LD->getMemoryVT();
10105 if (!VT.isFloatingPoint() ||
10106 VT != ST->getMemoryVT() ||
10107 LD->isNonTemporal() ||
10108 ST->isNonTemporal() ||
10109 LD->getPointerInfo().getAddrSpace() != 0 ||
10110 ST->getPointerInfo().getAddrSpace() != 0)
10113 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
10114 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
10115 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
10116 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
10117 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
10120 unsigned LDAlign = LD->getAlignment();
10121 unsigned STAlign = ST->getAlignment();
10122 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
10123 unsigned ABIAlign = TLI.getDataLayout()->getABITypeAlignment(IntVTTy);
10124 if (LDAlign < ABIAlign || STAlign < ABIAlign)
10127 SDValue NewLD = DAG.getLoad(IntVT, SDLoc(Value),
10128 LD->getChain(), LD->getBasePtr(),
10129 LD->getPointerInfo(),
10130 false, false, false, LDAlign);
10132 SDValue NewST = DAG.getStore(NewLD.getValue(1), SDLoc(N),
10133 NewLD, ST->getBasePtr(),
10134 ST->getPointerInfo(),
10135 false, false, STAlign);
10137 AddToWorklist(NewLD.getNode());
10138 AddToWorklist(NewST.getNode());
10139 WorklistRemover DeadNodes(*this);
10140 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
10149 /// Helper struct to parse and store a memory address as base + index + offset.
10150 /// We ignore sign extensions when it is safe to do so.
10151 /// The following two expressions are not equivalent. To differentiate we need
10152 /// to store whether there was a sign extension involved in the index
10154 /// (load (i64 add (i64 copyfromreg %c)
10155 /// (i64 signextend (add (i8 load %index)
10159 /// (load (i64 add (i64 copyfromreg %c)
10160 /// (i64 signextend (i32 add (i32 signextend (i8 load %index))
10162 struct BaseIndexOffset {
10166 bool IsIndexSignExt;
10168 BaseIndexOffset() : Offset(0), IsIndexSignExt(false) {}
10170 BaseIndexOffset(SDValue Base, SDValue Index, int64_t Offset,
10171 bool IsIndexSignExt) :
10172 Base(Base), Index(Index), Offset(Offset), IsIndexSignExt(IsIndexSignExt) {}
10174 bool equalBaseIndex(const BaseIndexOffset &Other) {
10175 return Other.Base == Base && Other.Index == Index &&
10176 Other.IsIndexSignExt == IsIndexSignExt;
10179 /// Parses tree in Ptr for base, index, offset addresses.
10180 static BaseIndexOffset match(SDValue Ptr) {
10181 bool IsIndexSignExt = false;
10183 // We only can pattern match BASE + INDEX + OFFSET. If Ptr is not an ADD
10184 // instruction, then it could be just the BASE or everything else we don't
10185 // know how to handle. Just use Ptr as BASE and give up.
10186 if (Ptr->getOpcode() != ISD::ADD)
10187 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
10189 // We know that we have at least an ADD instruction. Try to pattern match
10190 // the simple case of BASE + OFFSET.
10191 if (isa<ConstantSDNode>(Ptr->getOperand(1))) {
10192 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
10193 return BaseIndexOffset(Ptr->getOperand(0), SDValue(), Offset,
10197 // Inside a loop the current BASE pointer is calculated using an ADD and a
10198 // MUL instruction. In this case Ptr is the actual BASE pointer.
10199 // (i64 add (i64 %array_ptr)
10200 // (i64 mul (i64 %induction_var)
10201 // (i64 %element_size)))
10202 if (Ptr->getOperand(1)->getOpcode() == ISD::MUL)
10203 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
10205 // Look at Base + Index + Offset cases.
10206 SDValue Base = Ptr->getOperand(0);
10207 SDValue IndexOffset = Ptr->getOperand(1);
10209 // Skip signextends.
10210 if (IndexOffset->getOpcode() == ISD::SIGN_EXTEND) {
10211 IndexOffset = IndexOffset->getOperand(0);
10212 IsIndexSignExt = true;
10215 // Either the case of Base + Index (no offset) or something else.
10216 if (IndexOffset->getOpcode() != ISD::ADD)
10217 return BaseIndexOffset(Base, IndexOffset, 0, IsIndexSignExt);
10219 // Now we have the case of Base + Index + offset.
10220 SDValue Index = IndexOffset->getOperand(0);
10221 SDValue Offset = IndexOffset->getOperand(1);
10223 if (!isa<ConstantSDNode>(Offset))
10224 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
10226 // Ignore signextends.
10227 if (Index->getOpcode() == ISD::SIGN_EXTEND) {
10228 Index = Index->getOperand(0);
10229 IsIndexSignExt = true;
10230 } else IsIndexSignExt = false;
10232 int64_t Off = cast<ConstantSDNode>(Offset)->getSExtValue();
10233 return BaseIndexOffset(Base, Index, Off, IsIndexSignExt);
10238 bool DAGCombiner::MergeStoresOfConstantsOrVecElts(
10239 SmallVectorImpl<MemOpLink> &StoreNodes, EVT MemVT,
10240 unsigned NumElem, bool IsConstantSrc, bool UseVector) {
10241 // Make sure we have something to merge.
10245 int64_t ElementSizeBytes = MemVT.getSizeInBits() / 8;
10246 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
10247 unsigned LatestNodeUsed = 0;
10249 for (unsigned i=0; i < NumElem; ++i) {
10250 // Find a chain for the new wide-store operand. Notice that some
10251 // of the store nodes that we found may not be selected for inclusion
10252 // in the wide store. The chain we use needs to be the chain of the
10253 // latest store node which is *used* and replaced by the wide store.
10254 if (StoreNodes[i].SequenceNum < StoreNodes[LatestNodeUsed].SequenceNum)
10255 LatestNodeUsed = i;
10258 // The latest Node in the DAG.
10259 LSBaseSDNode *LatestOp = StoreNodes[LatestNodeUsed].MemNode;
10260 SDLoc DL(StoreNodes[0].MemNode);
10264 // Find a legal type for the vector store.
10265 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
10266 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
10267 if (IsConstantSrc) {
10268 // A vector store with a constant source implies that the constant is
10269 // zero; we only handle merging stores of constant zeros because the zero
10270 // can be materialized without a load.
10271 // It may be beneficial to loosen this restriction to allow non-zero
10273 StoredVal = DAG.getConstant(0, Ty);
10275 SmallVector<SDValue, 8> Ops;
10276 for (unsigned i = 0; i < NumElem ; ++i) {
10277 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10278 SDValue Val = St->getValue();
10279 // All of the operands of a BUILD_VECTOR must have the same type.
10280 if (Val.getValueType() != MemVT)
10282 Ops.push_back(Val);
10285 // Build the extracted vector elements back into a vector.
10286 StoredVal = DAG.getNode(ISD::BUILD_VECTOR, DL, Ty, Ops);
10289 // We should always use a vector store when merging extracted vector
10290 // elements, so this path implies a store of constants.
10291 assert(IsConstantSrc && "Merged vector elements should use vector store");
10293 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
10294 APInt StoreInt(StoreBW, 0);
10296 // Construct a single integer constant which is made of the smaller
10297 // constant inputs.
10298 bool IsLE = TLI.isLittleEndian();
10299 for (unsigned i = 0; i < NumElem ; ++i) {
10300 unsigned Idx = IsLE ? (NumElem - 1 - i) : i;
10301 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
10302 SDValue Val = St->getValue();
10303 StoreInt <<= ElementSizeBytes*8;
10304 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
10305 StoreInt |= C->getAPIntValue().zext(StoreBW);
10306 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
10307 StoreInt |= C->getValueAPF().bitcastToAPInt().zext(StoreBW);
10309 llvm_unreachable("Invalid constant element type");
10313 // Create the new Load and Store operations.
10314 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
10315 StoredVal = DAG.getConstant(StoreInt, StoreTy);
10318 SDValue NewStore = DAG.getStore(LatestOp->getChain(), DL, StoredVal,
10319 FirstInChain->getBasePtr(),
10320 FirstInChain->getPointerInfo(),
10322 FirstInChain->getAlignment());
10324 // Replace the last store with the new store
10325 CombineTo(LatestOp, NewStore);
10326 // Erase all other stores.
10327 for (unsigned i = 0; i < NumElem ; ++i) {
10328 if (StoreNodes[i].MemNode == LatestOp)
10330 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10331 // ReplaceAllUsesWith will replace all uses that existed when it was
10332 // called, but graph optimizations may cause new ones to appear. For
10333 // example, the case in pr14333 looks like
10335 // St's chain -> St -> another store -> X
10337 // And the only difference from St to the other store is the chain.
10338 // When we change it's chain to be St's chain they become identical,
10339 // get CSEed and the net result is that X is now a use of St.
10340 // Since we know that St is redundant, just iterate.
10341 while (!St->use_empty())
10342 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain());
10343 deleteAndRecombine(St);
10349 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
10350 if (OptLevel == CodeGenOpt::None)
10353 EVT MemVT = St->getMemoryVT();
10354 int64_t ElementSizeBytes = MemVT.getSizeInBits()/8;
10355 bool NoVectors = DAG.getMachineFunction().getFunction()->hasFnAttribute(
10356 Attribute::NoImplicitFloat);
10358 // Don't merge vectors into wider inputs.
10359 if (MemVT.isVector() || !MemVT.isSimple())
10362 // Perform an early exit check. Do not bother looking at stored values that
10363 // are not constants, loads, or extracted vector elements.
10364 SDValue StoredVal = St->getValue();
10365 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
10366 bool IsConstantSrc = isa<ConstantSDNode>(StoredVal) ||
10367 isa<ConstantFPSDNode>(StoredVal);
10368 bool IsExtractVecEltSrc = (StoredVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT);
10370 if (!IsConstantSrc && !IsLoadSrc && !IsExtractVecEltSrc)
10373 // Only look at ends of store sequences.
10374 SDValue Chain = SDValue(St, 0);
10375 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
10378 // This holds the base pointer, index, and the offset in bytes from the base
10380 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr());
10382 // We must have a base and an offset.
10383 if (!BasePtr.Base.getNode())
10386 // Do not handle stores to undef base pointers.
10387 if (BasePtr.Base.getOpcode() == ISD::UNDEF)
10390 // Save the LoadSDNodes that we find in the chain.
10391 // We need to make sure that these nodes do not interfere with
10392 // any of the store nodes.
10393 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
10395 // Save the StoreSDNodes that we find in the chain.
10396 SmallVector<MemOpLink, 8> StoreNodes;
10398 // Walk up the chain and look for nodes with offsets from the same
10399 // base pointer. Stop when reaching an instruction with a different kind
10400 // or instruction which has a different base pointer.
10402 StoreSDNode *Index = St;
10404 // If the chain has more than one use, then we can't reorder the mem ops.
10405 if (Index != St && !SDValue(Index, 0)->hasOneUse())
10408 // Find the base pointer and offset for this memory node.
10409 BaseIndexOffset Ptr = BaseIndexOffset::match(Index->getBasePtr());
10411 // Check that the base pointer is the same as the original one.
10412 if (!Ptr.equalBaseIndex(BasePtr))
10415 // Check that the alignment is the same.
10416 if (Index->getAlignment() != St->getAlignment())
10419 // The memory operands must not be volatile.
10420 if (Index->isVolatile() || Index->isIndexed())
10424 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
10425 if (St->isTruncatingStore())
10428 // The stored memory type must be the same.
10429 if (Index->getMemoryVT() != MemVT)
10432 // We do not allow unaligned stores because we want to prevent overriding
10434 if (Index->getAlignment()*8 != MemVT.getSizeInBits())
10437 // We found a potential memory operand to merge.
10438 StoreNodes.push_back(MemOpLink(Index, Ptr.Offset, Seq++));
10440 // Find the next memory operand in the chain. If the next operand in the
10441 // chain is a store then move up and continue the scan with the next
10442 // memory operand. If the next operand is a load save it and use alias
10443 // information to check if it interferes with anything.
10444 SDNode *NextInChain = Index->getChain().getNode();
10446 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
10447 // We found a store node. Use it for the next iteration.
10450 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
10451 if (Ldn->isVolatile()) {
10456 // Save the load node for later. Continue the scan.
10457 AliasLoadNodes.push_back(Ldn);
10458 NextInChain = Ldn->getChain().getNode();
10467 // Check if there is anything to merge.
10468 if (StoreNodes.size() < 2)
10471 // Sort the memory operands according to their distance from the base pointer.
10472 std::sort(StoreNodes.begin(), StoreNodes.end(),
10473 [](MemOpLink LHS, MemOpLink RHS) {
10474 return LHS.OffsetFromBase < RHS.OffsetFromBase ||
10475 (LHS.OffsetFromBase == RHS.OffsetFromBase &&
10476 LHS.SequenceNum > RHS.SequenceNum);
10479 // Scan the memory operations on the chain and find the first non-consecutive
10480 // store memory address.
10481 unsigned LastConsecutiveStore = 0;
10482 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
10483 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) {
10485 // Check that the addresses are consecutive starting from the second
10486 // element in the list of stores.
10488 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
10489 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
10493 bool Alias = false;
10494 // Check if this store interferes with any of the loads that we found.
10495 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld)
10496 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) {
10500 // We found a load that alias with this store. Stop the sequence.
10504 // Mark this node as useful.
10505 LastConsecutiveStore = i;
10508 // The node with the lowest store address.
10509 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
10511 // Store the constants into memory as one consecutive store.
10512 if (IsConstantSrc) {
10513 unsigned LastLegalType = 0;
10514 unsigned LastLegalVectorType = 0;
10515 bool NonZero = false;
10516 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
10517 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10518 SDValue StoredVal = St->getValue();
10520 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
10521 NonZero |= !C->isNullValue();
10522 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) {
10523 NonZero |= !C->getConstantFPValue()->isNullValue();
10529 // Find a legal type for the constant store.
10530 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
10531 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
10532 if (TLI.isTypeLegal(StoreTy))
10533 LastLegalType = i+1;
10534 // Or check whether a truncstore is legal.
10535 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
10536 TargetLowering::TypePromoteInteger) {
10537 EVT LegalizedStoredValueTy =
10538 TLI.getTypeToTransformTo(*DAG.getContext(), StoredVal.getValueType());
10539 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy))
10540 LastLegalType = i+1;
10543 // Find a legal type for the vector store.
10544 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
10545 if (TLI.isTypeLegal(Ty))
10546 LastLegalVectorType = i + 1;
10549 // We only use vectors if the constant is known to be zero and the
10550 // function is not marked with the noimplicitfloat attribute.
10551 if (NonZero || NoVectors)
10552 LastLegalVectorType = 0;
10554 // Check if we found a legal integer type to store.
10555 if (LastLegalType == 0 && LastLegalVectorType == 0)
10558 bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors;
10559 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType;
10561 return MergeStoresOfConstantsOrVecElts(StoreNodes, MemVT, NumElem,
10565 // When extracting multiple vector elements, try to store them
10566 // in one vector store rather than a sequence of scalar stores.
10567 if (IsExtractVecEltSrc) {
10568 unsigned NumElem = 0;
10569 for (unsigned i = 0; i < LastConsecutiveStore + 1; ++i) {
10570 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10571 SDValue StoredVal = St->getValue();
10572 // This restriction could be loosened.
10573 // Bail out if any stored values are not elements extracted from a vector.
10574 // It should be possible to handle mixed sources, but load sources need
10575 // more careful handling (see the block of code below that handles
10576 // consecutive loads).
10577 if (StoredVal.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
10580 // Find a legal type for the vector store.
10581 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
10582 if (TLI.isTypeLegal(Ty))
10586 return MergeStoresOfConstantsOrVecElts(StoreNodes, MemVT, NumElem,
10590 // Below we handle the case of multiple consecutive stores that
10591 // come from multiple consecutive loads. We merge them into a single
10592 // wide load and a single wide store.
10594 // Look for load nodes which are used by the stored values.
10595 SmallVector<MemOpLink, 8> LoadNodes;
10597 // Find acceptable loads. Loads need to have the same chain (token factor),
10598 // must not be zext, volatile, indexed, and they must be consecutive.
10599 BaseIndexOffset LdBasePtr;
10600 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
10601 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10602 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
10605 // Loads must only have one use.
10606 if (!Ld->hasNUsesOfValue(1, 0))
10609 // Check that the alignment is the same as the stores.
10610 if (Ld->getAlignment() != St->getAlignment())
10613 // The memory operands must not be volatile.
10614 if (Ld->isVolatile() || Ld->isIndexed())
10617 // We do not accept ext loads.
10618 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
10621 // The stored memory type must be the same.
10622 if (Ld->getMemoryVT() != MemVT)
10625 BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld->getBasePtr());
10626 // If this is not the first ptr that we check.
10627 if (LdBasePtr.Base.getNode()) {
10628 // The base ptr must be the same.
10629 if (!LdPtr.equalBaseIndex(LdBasePtr))
10632 // Check that all other base pointers are the same as this one.
10636 // We found a potential memory operand to merge.
10637 LoadNodes.push_back(MemOpLink(Ld, LdPtr.Offset, 0));
10640 if (LoadNodes.size() < 2)
10643 // If we have load/store pair instructions and we only have two values,
10645 unsigned RequiredAlignment;
10646 if (LoadNodes.size() == 2 && TLI.hasPairedLoad(MemVT, RequiredAlignment) &&
10647 St->getAlignment() >= RequiredAlignment)
10650 // Scan the memory operations on the chain and find the first non-consecutive
10651 // load memory address. These variables hold the index in the store node
10653 unsigned LastConsecutiveLoad = 0;
10654 // This variable refers to the size and not index in the array.
10655 unsigned LastLegalVectorType = 0;
10656 unsigned LastLegalIntegerType = 0;
10657 StartAddress = LoadNodes[0].OffsetFromBase;
10658 SDValue FirstChain = LoadNodes[0].MemNode->getChain();
10659 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
10660 // All loads much share the same chain.
10661 if (LoadNodes[i].MemNode->getChain() != FirstChain)
10664 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
10665 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
10667 LastConsecutiveLoad = i;
10669 // Find a legal type for the vector store.
10670 EVT StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
10671 if (TLI.isTypeLegal(StoreTy))
10672 LastLegalVectorType = i + 1;
10674 // Find a legal type for the integer store.
10675 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
10676 StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
10677 if (TLI.isTypeLegal(StoreTy))
10678 LastLegalIntegerType = i + 1;
10679 // Or check whether a truncstore and extload is legal.
10680 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
10681 TargetLowering::TypePromoteInteger) {
10682 EVT LegalizedStoredValueTy =
10683 TLI.getTypeToTransformTo(*DAG.getContext(), StoreTy);
10684 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
10685 TLI.isLoadExtLegal(ISD::ZEXTLOAD, LegalizedStoredValueTy, StoreTy) &&
10686 TLI.isLoadExtLegal(ISD::SEXTLOAD, LegalizedStoredValueTy, StoreTy) &&
10687 TLI.isLoadExtLegal(ISD::EXTLOAD, LegalizedStoredValueTy, StoreTy))
10688 LastLegalIntegerType = i+1;
10692 // Only use vector types if the vector type is larger than the integer type.
10693 // If they are the same, use integers.
10694 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors;
10695 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType);
10697 // We add +1 here because the LastXXX variables refer to location while
10698 // the NumElem refers to array/index size.
10699 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1;
10700 NumElem = std::min(LastLegalType, NumElem);
10705 // The latest Node in the DAG.
10706 unsigned LatestNodeUsed = 0;
10707 for (unsigned i=1; i<NumElem; ++i) {
10708 // Find a chain for the new wide-store operand. Notice that some
10709 // of the store nodes that we found may not be selected for inclusion
10710 // in the wide store. The chain we use needs to be the chain of the
10711 // latest store node which is *used* and replaced by the wide store.
10712 if (StoreNodes[i].SequenceNum < StoreNodes[LatestNodeUsed].SequenceNum)
10713 LatestNodeUsed = i;
10716 LSBaseSDNode *LatestOp = StoreNodes[LatestNodeUsed].MemNode;
10718 // Find if it is better to use vectors or integers to load and store
10722 JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
10724 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
10725 JointMemOpVT = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
10728 SDLoc LoadDL(LoadNodes[0].MemNode);
10729 SDLoc StoreDL(StoreNodes[0].MemNode);
10731 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
10732 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL,
10733 FirstLoad->getChain(),
10734 FirstLoad->getBasePtr(),
10735 FirstLoad->getPointerInfo(),
10736 false, false, false,
10737 FirstLoad->getAlignment());
10739 SDValue NewStore = DAG.getStore(LatestOp->getChain(), StoreDL, NewLoad,
10740 FirstInChain->getBasePtr(),
10741 FirstInChain->getPointerInfo(), false, false,
10742 FirstInChain->getAlignment());
10744 // Replace one of the loads with the new load.
10745 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode);
10746 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
10747 SDValue(NewLoad.getNode(), 1));
10749 // Remove the rest of the load chains.
10750 for (unsigned i = 1; i < NumElem ; ++i) {
10751 // Replace all chain users of the old load nodes with the chain of the new
10753 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
10754 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain());
10757 // Replace the last store with the new store.
10758 CombineTo(LatestOp, NewStore);
10759 // Erase all other stores.
10760 for (unsigned i = 0; i < NumElem ; ++i) {
10761 // Remove all Store nodes.
10762 if (StoreNodes[i].MemNode == LatestOp)
10764 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10765 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain());
10766 deleteAndRecombine(St);
10772 SDValue DAGCombiner::visitSTORE(SDNode *N) {
10773 StoreSDNode *ST = cast<StoreSDNode>(N);
10774 SDValue Chain = ST->getChain();
10775 SDValue Value = ST->getValue();
10776 SDValue Ptr = ST->getBasePtr();
10778 // If this is a store of a bit convert, store the input value if the
10779 // resultant store does not need a higher alignment than the original.
10780 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
10781 ST->isUnindexed()) {
10782 unsigned OrigAlign = ST->getAlignment();
10783 EVT SVT = Value.getOperand(0).getValueType();
10784 unsigned Align = TLI.getDataLayout()->
10785 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
10786 if (Align <= OrigAlign &&
10787 ((!LegalOperations && !ST->isVolatile()) ||
10788 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
10789 return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0),
10790 Ptr, ST->getPointerInfo(), ST->isVolatile(),
10791 ST->isNonTemporal(), OrigAlign,
10795 // Turn 'store undef, Ptr' -> nothing.
10796 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
10799 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
10800 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
10801 // NOTE: If the original store is volatile, this transform must not increase
10802 // the number of stores. For example, on x86-32 an f64 can be stored in one
10803 // processor operation but an i64 (which is not legal) requires two. So the
10804 // transform should not be done in this case.
10805 if (Value.getOpcode() != ISD::TargetConstantFP) {
10807 switch (CFP->getSimpleValueType(0).SimpleTy) {
10808 default: llvm_unreachable("Unknown FP type");
10809 case MVT::f16: // We don't do this for these yet.
10815 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
10816 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
10817 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
10818 bitcastToAPInt().getZExtValue(), MVT::i32);
10819 return DAG.getStore(Chain, SDLoc(N), Tmp,
10820 Ptr, ST->getMemOperand());
10824 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
10825 !ST->isVolatile()) ||
10826 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
10827 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
10828 getZExtValue(), MVT::i64);
10829 return DAG.getStore(Chain, SDLoc(N), Tmp,
10830 Ptr, ST->getMemOperand());
10833 if (!ST->isVolatile() &&
10834 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
10835 // Many FP stores are not made apparent until after legalize, e.g. for
10836 // argument passing. Since this is so common, custom legalize the
10837 // 64-bit integer store into two 32-bit stores.
10838 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
10839 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
10840 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
10841 if (TLI.isBigEndian()) std::swap(Lo, Hi);
10843 unsigned Alignment = ST->getAlignment();
10844 bool isVolatile = ST->isVolatile();
10845 bool isNonTemporal = ST->isNonTemporal();
10846 AAMDNodes AAInfo = ST->getAAInfo();
10848 SDValue St0 = DAG.getStore(Chain, SDLoc(ST), Lo,
10849 Ptr, ST->getPointerInfo(),
10850 isVolatile, isNonTemporal,
10851 ST->getAlignment(), AAInfo);
10852 Ptr = DAG.getNode(ISD::ADD, SDLoc(N), Ptr.getValueType(), Ptr,
10853 DAG.getConstant(4, Ptr.getValueType()));
10854 Alignment = MinAlign(Alignment, 4U);
10855 SDValue St1 = DAG.getStore(Chain, SDLoc(ST), Hi,
10856 Ptr, ST->getPointerInfo().getWithOffset(4),
10857 isVolatile, isNonTemporal,
10858 Alignment, AAInfo);
10859 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other,
10868 // Try to infer better alignment information than the store already has.
10869 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
10870 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
10871 if (Align > ST->getAlignment()) {
10873 DAG.getTruncStore(Chain, SDLoc(N), Value,
10874 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
10875 ST->isVolatile(), ST->isNonTemporal(), Align,
10877 if (NewStore.getNode() != N)
10878 return CombineTo(ST, NewStore, true);
10883 // Try transforming a pair floating point load / store ops to integer
10884 // load / store ops.
10885 SDValue NewST = TransformFPLoadStorePair(N);
10886 if (NewST.getNode())
10889 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
10890 : DAG.getSubtarget().useAA();
10892 if (CombinerAAOnlyFunc.getNumOccurrences() &&
10893 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
10896 if (UseAA && ST->isUnindexed()) {
10897 // Walk up chain skipping non-aliasing memory nodes.
10898 SDValue BetterChain = FindBetterChain(N, Chain);
10900 // If there is a better chain.
10901 if (Chain != BetterChain) {
10904 // Replace the chain to avoid dependency.
10905 if (ST->isTruncatingStore()) {
10906 ReplStore = DAG.getTruncStore(BetterChain, SDLoc(N), Value, Ptr,
10907 ST->getMemoryVT(), ST->getMemOperand());
10909 ReplStore = DAG.getStore(BetterChain, SDLoc(N), Value, Ptr,
10910 ST->getMemOperand());
10913 // Create token to keep both nodes around.
10914 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
10915 MVT::Other, Chain, ReplStore);
10917 // Make sure the new and old chains are cleaned up.
10918 AddToWorklist(Token.getNode());
10920 // Don't add users to work list.
10921 return CombineTo(N, Token, false);
10925 // Try transforming N to an indexed store.
10926 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
10927 return SDValue(N, 0);
10929 // FIXME: is there such a thing as a truncating indexed store?
10930 if (ST->isTruncatingStore() && ST->isUnindexed() &&
10931 Value.getValueType().isInteger()) {
10932 // See if we can simplify the input to this truncstore with knowledge that
10933 // only the low bits are being used. For example:
10934 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
10936 GetDemandedBits(Value,
10937 APInt::getLowBitsSet(
10938 Value.getValueType().getScalarType().getSizeInBits(),
10939 ST->getMemoryVT().getScalarType().getSizeInBits()));
10940 AddToWorklist(Value.getNode());
10941 if (Shorter.getNode())
10942 return DAG.getTruncStore(Chain, SDLoc(N), Shorter,
10943 Ptr, ST->getMemoryVT(), ST->getMemOperand());
10945 // Otherwise, see if we can simplify the operation with
10946 // SimplifyDemandedBits, which only works if the value has a single use.
10947 if (SimplifyDemandedBits(Value,
10948 APInt::getLowBitsSet(
10949 Value.getValueType().getScalarType().getSizeInBits(),
10950 ST->getMemoryVT().getScalarType().getSizeInBits())))
10951 return SDValue(N, 0);
10954 // If this is a load followed by a store to the same location, then the store
10956 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
10957 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
10958 ST->isUnindexed() && !ST->isVolatile() &&
10959 // There can't be any side effects between the load and store, such as
10960 // a call or store.
10961 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
10962 // The store is dead, remove it.
10967 // If this is a store followed by a store with the same value to the same
10968 // location, then the store is dead/noop.
10969 if (StoreSDNode *ST1 = dyn_cast<StoreSDNode>(Chain)) {
10970 if (ST1->getBasePtr() == Ptr && ST->getMemoryVT() == ST1->getMemoryVT() &&
10971 ST1->getValue() == Value && ST->isUnindexed() && !ST->isVolatile() &&
10972 ST1->isUnindexed() && !ST1->isVolatile()) {
10973 // The store is dead, remove it.
10978 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
10979 // truncating store. We can do this even if this is already a truncstore.
10980 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
10981 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
10982 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
10983 ST->getMemoryVT())) {
10984 return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0),
10985 Ptr, ST->getMemoryVT(), ST->getMemOperand());
10988 // Only perform this optimization before the types are legal, because we
10989 // don't want to perform this optimization on every DAGCombine invocation.
10991 bool EverChanged = false;
10994 // There can be multiple store sequences on the same chain.
10995 // Keep trying to merge store sequences until we are unable to do so
10996 // or until we merge the last store on the chain.
10997 bool Changed = MergeConsecutiveStores(ST);
10998 EverChanged |= Changed;
10999 if (!Changed) break;
11000 } while (ST->getOpcode() != ISD::DELETED_NODE);
11003 return SDValue(N, 0);
11006 return ReduceLoadOpStoreWidth(N);
11009 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
11010 SDValue InVec = N->getOperand(0);
11011 SDValue InVal = N->getOperand(1);
11012 SDValue EltNo = N->getOperand(2);
11015 // If the inserted element is an UNDEF, just use the input vector.
11016 if (InVal.getOpcode() == ISD::UNDEF)
11019 EVT VT = InVec.getValueType();
11021 // If we can't generate a legal BUILD_VECTOR, exit
11022 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
11025 // Check that we know which element is being inserted
11026 if (!isa<ConstantSDNode>(EltNo))
11028 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
11030 // Canonicalize insert_vector_elt dag nodes.
11032 // (insert_vector_elt (insert_vector_elt A, Idx0), Idx1)
11033 // -> (insert_vector_elt (insert_vector_elt A, Idx1), Idx0)
11035 // Do this only if the child insert_vector node has one use; also
11036 // do this only if indices are both constants and Idx1 < Idx0.
11037 if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT && InVec.hasOneUse()
11038 && isa<ConstantSDNode>(InVec.getOperand(2))) {
11039 unsigned OtherElt =
11040 cast<ConstantSDNode>(InVec.getOperand(2))->getZExtValue();
11041 if (Elt < OtherElt) {
11043 SDValue NewOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(N), VT,
11044 InVec.getOperand(0), InVal, EltNo);
11045 AddToWorklist(NewOp.getNode());
11046 return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(InVec.getNode()),
11047 VT, NewOp, InVec.getOperand(1), InVec.getOperand(2));
11051 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
11052 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
11053 // vector elements.
11054 SmallVector<SDValue, 8> Ops;
11055 // Do not combine these two vectors if the output vector will not replace
11056 // the input vector.
11057 if (InVec.getOpcode() == ISD::BUILD_VECTOR && InVec.hasOneUse()) {
11058 Ops.append(InVec.getNode()->op_begin(),
11059 InVec.getNode()->op_end());
11060 } else if (InVec.getOpcode() == ISD::UNDEF) {
11061 unsigned NElts = VT.getVectorNumElements();
11062 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
11067 // Insert the element
11068 if (Elt < Ops.size()) {
11069 // All the operands of BUILD_VECTOR must have the same type;
11070 // we enforce that here.
11071 EVT OpVT = Ops[0].getValueType();
11072 if (InVal.getValueType() != OpVT)
11073 InVal = OpVT.bitsGT(InVal.getValueType()) ?
11074 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
11075 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
11079 // Return the new vector
11080 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
11083 SDValue DAGCombiner::ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
11084 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad) {
11085 EVT ResultVT = EVE->getValueType(0);
11086 EVT VecEltVT = InVecVT.getVectorElementType();
11087 unsigned Align = OriginalLoad->getAlignment();
11088 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
11089 VecEltVT.getTypeForEVT(*DAG.getContext()));
11091 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VecEltVT))
11096 SDValue NewPtr = OriginalLoad->getBasePtr();
11098 EVT PtrType = NewPtr.getValueType();
11099 MachinePointerInfo MPI;
11100 if (auto *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo)) {
11101 int Elt = ConstEltNo->getZExtValue();
11102 unsigned PtrOff = VecEltVT.getSizeInBits() * Elt / 8;
11103 if (TLI.isBigEndian())
11104 PtrOff = InVecVT.getSizeInBits() / 8 - PtrOff;
11105 Offset = DAG.getConstant(PtrOff, PtrType);
11106 MPI = OriginalLoad->getPointerInfo().getWithOffset(PtrOff);
11108 Offset = DAG.getNode(
11109 ISD::MUL, SDLoc(EVE), EltNo.getValueType(), EltNo,
11110 DAG.getConstant(VecEltVT.getStoreSize(), EltNo.getValueType()));
11111 if (TLI.isBigEndian())
11112 Offset = DAG.getNode(
11113 ISD::SUB, SDLoc(EVE), EltNo.getValueType(),
11114 DAG.getConstant(InVecVT.getStoreSize(), EltNo.getValueType()), Offset);
11115 MPI = OriginalLoad->getPointerInfo();
11117 NewPtr = DAG.getNode(ISD::ADD, SDLoc(EVE), PtrType, NewPtr, Offset);
11119 // The replacement we need to do here is a little tricky: we need to
11120 // replace an extractelement of a load with a load.
11121 // Use ReplaceAllUsesOfValuesWith to do the replacement.
11122 // Note that this replacement assumes that the extractvalue is the only
11123 // use of the load; that's okay because we don't want to perform this
11124 // transformation in other cases anyway.
11127 if (ResultVT.bitsGT(VecEltVT)) {
11128 // If the result type of vextract is wider than the load, then issue an
11129 // extending load instead.
11130 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, ResultVT,
11134 Load = DAG.getExtLoad(
11135 ExtType, SDLoc(EVE), ResultVT, OriginalLoad->getChain(), NewPtr, MPI,
11136 VecEltVT, OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
11137 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
11138 Chain = Load.getValue(1);
11140 Load = DAG.getLoad(
11141 VecEltVT, SDLoc(EVE), OriginalLoad->getChain(), NewPtr, MPI,
11142 OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
11143 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
11144 Chain = Load.getValue(1);
11145 if (ResultVT.bitsLT(VecEltVT))
11146 Load = DAG.getNode(ISD::TRUNCATE, SDLoc(EVE), ResultVT, Load);
11148 Load = DAG.getNode(ISD::BITCAST, SDLoc(EVE), ResultVT, Load);
11150 WorklistRemover DeadNodes(*this);
11151 SDValue From[] = { SDValue(EVE, 0), SDValue(OriginalLoad, 1) };
11152 SDValue To[] = { Load, Chain };
11153 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
11154 // Since we're explicitly calling ReplaceAllUses, add the new node to the
11155 // worklist explicitly as well.
11156 AddToWorklist(Load.getNode());
11157 AddUsersToWorklist(Load.getNode()); // Add users too
11158 // Make sure to revisit this node to clean it up; it will usually be dead.
11159 AddToWorklist(EVE);
11161 return SDValue(EVE, 0);
11164 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
11165 // (vextract (scalar_to_vector val, 0) -> val
11166 SDValue InVec = N->getOperand(0);
11167 EVT VT = InVec.getValueType();
11168 EVT NVT = N->getValueType(0);
11170 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
11171 // Check if the result type doesn't match the inserted element type. A
11172 // SCALAR_TO_VECTOR may truncate the inserted element and the
11173 // EXTRACT_VECTOR_ELT may widen the extracted vector.
11174 SDValue InOp = InVec.getOperand(0);
11175 if (InOp.getValueType() != NVT) {
11176 assert(InOp.getValueType().isInteger() && NVT.isInteger());
11177 return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT);
11182 SDValue EltNo = N->getOperand(1);
11183 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
11185 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
11186 // We only perform this optimization before the op legalization phase because
11187 // we may introduce new vector instructions which are not backed by TD
11188 // patterns. For example on AVX, extracting elements from a wide vector
11189 // without using extract_subvector. However, if we can find an underlying
11190 // scalar value, then we can always use that.
11191 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
11193 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
11194 int NumElem = VT.getVectorNumElements();
11195 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
11196 // Find the new index to extract from.
11197 int OrigElt = SVOp->getMaskElt(Elt);
11199 // Extracting an undef index is undef.
11201 return DAG.getUNDEF(NVT);
11203 // Select the right vector half to extract from.
11205 if (OrigElt < NumElem) {
11206 SVInVec = InVec->getOperand(0);
11208 SVInVec = InVec->getOperand(1);
11209 OrigElt -= NumElem;
11212 if (SVInVec.getOpcode() == ISD::BUILD_VECTOR) {
11213 SDValue InOp = SVInVec.getOperand(OrigElt);
11214 if (InOp.getValueType() != NVT) {
11215 assert(InOp.getValueType().isInteger() && NVT.isInteger());
11216 InOp = DAG.getSExtOrTrunc(InOp, SDLoc(SVInVec), NVT);
11222 // FIXME: We should handle recursing on other vector shuffles and
11223 // scalar_to_vector here as well.
11225 if (!LegalOperations) {
11226 EVT IndexTy = TLI.getVectorIdxTy();
11227 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT,
11228 SVInVec, DAG.getConstant(OrigElt, IndexTy));
11232 bool BCNumEltsChanged = false;
11233 EVT ExtVT = VT.getVectorElementType();
11236 // If the result of load has to be truncated, then it's not necessarily
11238 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
11241 if (InVec.getOpcode() == ISD::BITCAST) {
11242 // Don't duplicate a load with other uses.
11243 if (!InVec.hasOneUse())
11246 EVT BCVT = InVec.getOperand(0).getValueType();
11247 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
11249 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
11250 BCNumEltsChanged = true;
11251 InVec = InVec.getOperand(0);
11252 ExtVT = BCVT.getVectorElementType();
11255 // (vextract (vN[if]M load $addr), i) -> ([if]M load $addr + i * size)
11256 if (!LegalOperations && !ConstEltNo && InVec.hasOneUse() &&
11257 ISD::isNormalLoad(InVec.getNode()) &&
11258 !N->getOperand(1)->hasPredecessor(InVec.getNode())) {
11259 SDValue Index = N->getOperand(1);
11260 if (LoadSDNode *OrigLoad = dyn_cast<LoadSDNode>(InVec))
11261 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, Index,
11265 // Perform only after legalization to ensure build_vector / vector_shuffle
11266 // optimizations have already been done.
11267 if (!LegalOperations) return SDValue();
11269 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
11270 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
11271 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
11274 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
11276 LoadSDNode *LN0 = nullptr;
11277 const ShuffleVectorSDNode *SVN = nullptr;
11278 if (ISD::isNormalLoad(InVec.getNode())) {
11279 LN0 = cast<LoadSDNode>(InVec);
11280 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
11281 InVec.getOperand(0).getValueType() == ExtVT &&
11282 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
11283 // Don't duplicate a load with other uses.
11284 if (!InVec.hasOneUse())
11287 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
11288 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
11289 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
11291 // (load $addr+1*size)
11293 // Don't duplicate a load with other uses.
11294 if (!InVec.hasOneUse())
11297 // If the bit convert changed the number of elements, it is unsafe
11298 // to examine the mask.
11299 if (BCNumEltsChanged)
11302 // Select the input vector, guarding against out of range extract vector.
11303 unsigned NumElems = VT.getVectorNumElements();
11304 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
11305 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
11307 if (InVec.getOpcode() == ISD::BITCAST) {
11308 // Don't duplicate a load with other uses.
11309 if (!InVec.hasOneUse())
11312 InVec = InVec.getOperand(0);
11314 if (ISD::isNormalLoad(InVec.getNode())) {
11315 LN0 = cast<LoadSDNode>(InVec);
11316 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
11317 EltNo = DAG.getConstant(Elt, EltNo.getValueType());
11321 // Make sure we found a non-volatile load and the extractelement is
11323 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
11326 // If Idx was -1 above, Elt is going to be -1, so just return undef.
11328 return DAG.getUNDEF(LVT);
11330 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, EltNo, LN0);
11336 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
11337 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
11338 // We perform this optimization post type-legalization because
11339 // the type-legalizer often scalarizes integer-promoted vectors.
11340 // Performing this optimization before may create bit-casts which
11341 // will be type-legalized to complex code sequences.
11342 // We perform this optimization only before the operation legalizer because we
11343 // may introduce illegal operations.
11344 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
11347 unsigned NumInScalars = N->getNumOperands();
11349 EVT VT = N->getValueType(0);
11351 // Check to see if this is a BUILD_VECTOR of a bunch of values
11352 // which come from any_extend or zero_extend nodes. If so, we can create
11353 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
11354 // optimizations. We do not handle sign-extend because we can't fill the sign
11356 EVT SourceType = MVT::Other;
11357 bool AllAnyExt = true;
11359 for (unsigned i = 0; i != NumInScalars; ++i) {
11360 SDValue In = N->getOperand(i);
11361 // Ignore undef inputs.
11362 if (In.getOpcode() == ISD::UNDEF) continue;
11364 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
11365 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
11367 // Abort if the element is not an extension.
11368 if (!ZeroExt && !AnyExt) {
11369 SourceType = MVT::Other;
11373 // The input is a ZeroExt or AnyExt. Check the original type.
11374 EVT InTy = In.getOperand(0).getValueType();
11376 // Check that all of the widened source types are the same.
11377 if (SourceType == MVT::Other)
11380 else if (InTy != SourceType) {
11381 // Multiple income types. Abort.
11382 SourceType = MVT::Other;
11386 // Check if all of the extends are ANY_EXTENDs.
11387 AllAnyExt &= AnyExt;
11390 // In order to have valid types, all of the inputs must be extended from the
11391 // same source type and all of the inputs must be any or zero extend.
11392 // Scalar sizes must be a power of two.
11393 EVT OutScalarTy = VT.getScalarType();
11394 bool ValidTypes = SourceType != MVT::Other &&
11395 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
11396 isPowerOf2_32(SourceType.getSizeInBits());
11398 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
11399 // turn into a single shuffle instruction.
11403 bool isLE = TLI.isLittleEndian();
11404 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
11405 assert(ElemRatio > 1 && "Invalid element size ratio");
11406 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
11407 DAG.getConstant(0, SourceType);
11409 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
11410 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
11412 // Populate the new build_vector
11413 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
11414 SDValue Cast = N->getOperand(i);
11415 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
11416 Cast.getOpcode() == ISD::ZERO_EXTEND ||
11417 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
11419 if (Cast.getOpcode() == ISD::UNDEF)
11420 In = DAG.getUNDEF(SourceType);
11422 In = Cast->getOperand(0);
11423 unsigned Index = isLE ? (i * ElemRatio) :
11424 (i * ElemRatio + (ElemRatio - 1));
11426 assert(Index < Ops.size() && "Invalid index");
11430 // The type of the new BUILD_VECTOR node.
11431 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
11432 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
11433 "Invalid vector size");
11434 // Check if the new vector type is legal.
11435 if (!isTypeLegal(VecVT)) return SDValue();
11437 // Make the new BUILD_VECTOR.
11438 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
11440 // The new BUILD_VECTOR node has the potential to be further optimized.
11441 AddToWorklist(BV.getNode());
11442 // Bitcast to the desired type.
11443 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
11446 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
11447 EVT VT = N->getValueType(0);
11449 unsigned NumInScalars = N->getNumOperands();
11452 EVT SrcVT = MVT::Other;
11453 unsigned Opcode = ISD::DELETED_NODE;
11454 unsigned NumDefs = 0;
11456 for (unsigned i = 0; i != NumInScalars; ++i) {
11457 SDValue In = N->getOperand(i);
11458 unsigned Opc = In.getOpcode();
11460 if (Opc == ISD::UNDEF)
11463 // If all scalar values are floats and converted from integers.
11464 if (Opcode == ISD::DELETED_NODE &&
11465 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
11472 EVT InVT = In.getOperand(0).getValueType();
11474 // If all scalar values are typed differently, bail out. It's chosen to
11475 // simplify BUILD_VECTOR of integer types.
11476 if (SrcVT == MVT::Other)
11483 // If the vector has just one element defined, it's not worth to fold it into
11484 // a vectorized one.
11488 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
11489 && "Should only handle conversion from integer to float.");
11490 assert(SrcVT != MVT::Other && "Cannot determine source type!");
11492 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
11494 if (!TLI.isOperationLegalOrCustom(Opcode, NVT))
11497 // Just because the floating-point vector type is legal does not necessarily
11498 // mean that the corresponding integer vector type is.
11499 if (!isTypeLegal(NVT))
11502 SmallVector<SDValue, 8> Opnds;
11503 for (unsigned i = 0; i != NumInScalars; ++i) {
11504 SDValue In = N->getOperand(i);
11506 if (In.getOpcode() == ISD::UNDEF)
11507 Opnds.push_back(DAG.getUNDEF(SrcVT));
11509 Opnds.push_back(In.getOperand(0));
11511 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Opnds);
11512 AddToWorklist(BV.getNode());
11514 return DAG.getNode(Opcode, dl, VT, BV);
11517 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
11518 unsigned NumInScalars = N->getNumOperands();
11520 EVT VT = N->getValueType(0);
11522 // A vector built entirely of undefs is undef.
11523 if (ISD::allOperandsUndef(N))
11524 return DAG.getUNDEF(VT);
11526 if (SDValue V = reduceBuildVecExtToExtBuildVec(N))
11529 if (SDValue V = reduceBuildVecConvertToConvertBuildVec(N))
11532 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
11533 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
11534 // at most two distinct vectors, turn this into a shuffle node.
11536 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
11537 if (!isTypeLegal(VT))
11540 // May only combine to shuffle after legalize if shuffle is legal.
11541 if (LegalOperations && !TLI.isOperationLegal(ISD::VECTOR_SHUFFLE, VT))
11544 SDValue VecIn1, VecIn2;
11545 bool UsesZeroVector = false;
11546 for (unsigned i = 0; i != NumInScalars; ++i) {
11547 SDValue Op = N->getOperand(i);
11548 // Ignore undef inputs.
11549 if (Op.getOpcode() == ISD::UNDEF) continue;
11551 // See if we can combine this build_vector into a blend with a zero vector.
11552 if (!VecIn2.getNode() && ((Op.getOpcode() == ISD::Constant &&
11553 cast<ConstantSDNode>(Op.getNode())->isNullValue()) ||
11554 (Op.getOpcode() == ISD::ConstantFP &&
11555 cast<ConstantFPSDNode>(Op.getNode())->getValueAPF().isZero()))) {
11556 UsesZeroVector = true;
11560 // If this input is something other than a EXTRACT_VECTOR_ELT with a
11561 // constant index, bail out.
11562 if (Op.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
11563 !isa<ConstantSDNode>(Op.getOperand(1))) {
11564 VecIn1 = VecIn2 = SDValue(nullptr, 0);
11568 // We allow up to two distinct input vectors.
11569 SDValue ExtractedFromVec = Op.getOperand(0);
11570 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
11573 if (!VecIn1.getNode()) {
11574 VecIn1 = ExtractedFromVec;
11575 } else if (!VecIn2.getNode() && !UsesZeroVector) {
11576 VecIn2 = ExtractedFromVec;
11578 // Too many inputs.
11579 VecIn1 = VecIn2 = SDValue(nullptr, 0);
11584 // If everything is good, we can make a shuffle operation.
11585 if (VecIn1.getNode()) {
11586 unsigned InNumElements = VecIn1.getValueType().getVectorNumElements();
11587 SmallVector<int, 8> Mask;
11588 for (unsigned i = 0; i != NumInScalars; ++i) {
11589 unsigned Opcode = N->getOperand(i).getOpcode();
11590 if (Opcode == ISD::UNDEF) {
11591 Mask.push_back(-1);
11595 // Operands can also be zero.
11596 if (Opcode != ISD::EXTRACT_VECTOR_ELT) {
11597 assert(UsesZeroVector &&
11598 (Opcode == ISD::Constant || Opcode == ISD::ConstantFP) &&
11599 "Unexpected node found!");
11600 Mask.push_back(NumInScalars+i);
11604 // If extracting from the first vector, just use the index directly.
11605 SDValue Extract = N->getOperand(i);
11606 SDValue ExtVal = Extract.getOperand(1);
11607 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
11608 if (Extract.getOperand(0) == VecIn1) {
11609 Mask.push_back(ExtIndex);
11613 // Otherwise, use InIdx + InputVecSize
11614 Mask.push_back(InNumElements + ExtIndex);
11617 // Avoid introducing illegal shuffles with zero.
11618 if (UsesZeroVector && !TLI.isVectorClearMaskLegal(Mask, VT))
11621 // We can't generate a shuffle node with mismatched input and output types.
11622 // Attempt to transform a single input vector to the correct type.
11623 if ((VT != VecIn1.getValueType())) {
11624 // If the input vector type has a different base type to the output
11625 // vector type, bail out.
11626 EVT VTElemType = VT.getVectorElementType();
11627 if ((VecIn1.getValueType().getVectorElementType() != VTElemType) ||
11628 (VecIn2.getNode() &&
11629 (VecIn2.getValueType().getVectorElementType() != VTElemType)))
11632 // If the input vector is too small, widen it.
11633 // We only support widening of vectors which are half the size of the
11634 // output registers. For example XMM->YMM widening on X86 with AVX.
11635 EVT VecInT = VecIn1.getValueType();
11636 if (VecInT.getSizeInBits() * 2 == VT.getSizeInBits()) {
11637 // If we only have one small input, widen it by adding undef values.
11638 if (!VecIn2.getNode())
11639 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, VecIn1,
11640 DAG.getUNDEF(VecIn1.getValueType()));
11641 else if (VecIn1.getValueType() == VecIn2.getValueType()) {
11642 // If we have two small inputs of the same type, try to concat them.
11643 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, VecIn1, VecIn2);
11644 VecIn2 = SDValue(nullptr, 0);
11647 } else if (VecInT.getSizeInBits() == VT.getSizeInBits() * 2) {
11648 // If the input vector is too large, try to split it.
11649 // We don't support having two input vectors that are too large.
11650 // If the zero vector was used, we can not split the vector,
11651 // since we'd need 3 inputs.
11652 if (UsesZeroVector || VecIn2.getNode())
11655 if (!TLI.isExtractSubvectorCheap(VT, VT.getVectorNumElements()))
11658 // Try to replace VecIn1 with two extract_subvectors
11659 // No need to update the masks, they should still be correct.
11660 VecIn2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, VecIn1,
11661 DAG.getConstant(VT.getVectorNumElements(), TLI.getVectorIdxTy()));
11662 VecIn1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, VecIn1,
11663 DAG.getConstant(0, TLI.getVectorIdxTy()));
11668 if (UsesZeroVector)
11669 VecIn2 = VT.isInteger() ? DAG.getConstant(0, VT) :
11670 DAG.getConstantFP(0.0, VT);
11672 // If VecIn2 is unused then change it to undef.
11673 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
11675 // Check that we were able to transform all incoming values to the same
11677 if (VecIn2.getValueType() != VecIn1.getValueType() ||
11678 VecIn1.getValueType() != VT)
11681 // Return the new VECTOR_SHUFFLE node.
11685 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]);
11691 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
11692 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
11693 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
11694 // inputs come from at most two distinct vectors, turn this into a shuffle
11697 // If we only have one input vector, we don't need to do any concatenation.
11698 if (N->getNumOperands() == 1)
11699 return N->getOperand(0);
11701 // Check if all of the operands are undefs.
11702 EVT VT = N->getValueType(0);
11703 if (ISD::allOperandsUndef(N))
11704 return DAG.getUNDEF(VT);
11706 // Optimize concat_vectors where one of the vectors is undef.
11707 if (N->getNumOperands() == 2 &&
11708 N->getOperand(1)->getOpcode() == ISD::UNDEF) {
11709 SDValue In = N->getOperand(0);
11710 assert(In.getValueType().isVector() && "Must concat vectors");
11712 // Transform: concat_vectors(scalar, undef) -> scalar_to_vector(sclr).
11713 if (In->getOpcode() == ISD::BITCAST &&
11714 !In->getOperand(0)->getValueType(0).isVector()) {
11715 SDValue Scalar = In->getOperand(0);
11716 EVT SclTy = Scalar->getValueType(0);
11718 if (!SclTy.isFloatingPoint() && !SclTy.isInteger())
11721 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SclTy,
11722 VT.getSizeInBits() / SclTy.getSizeInBits());
11723 if (!TLI.isTypeLegal(NVT) || !TLI.isTypeLegal(Scalar.getValueType()))
11726 SDLoc dl = SDLoc(N);
11727 SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NVT, Scalar);
11728 return DAG.getNode(ISD::BITCAST, dl, VT, Res);
11732 // Fold any combination of BUILD_VECTOR or UNDEF nodes into one BUILD_VECTOR.
11733 // We have already tested above for an UNDEF only concatenation.
11734 // fold (concat_vectors (BUILD_VECTOR A, B, ...), (BUILD_VECTOR C, D, ...))
11735 // -> (BUILD_VECTOR A, B, ..., C, D, ...)
11736 auto IsBuildVectorOrUndef = [](const SDValue &Op) {
11737 return ISD::UNDEF == Op.getOpcode() || ISD::BUILD_VECTOR == Op.getOpcode();
11739 bool AllBuildVectorsOrUndefs =
11740 std::all_of(N->op_begin(), N->op_end(), IsBuildVectorOrUndef);
11741 if (AllBuildVectorsOrUndefs) {
11742 SmallVector<SDValue, 8> Opnds;
11743 EVT SVT = VT.getScalarType();
11746 if (!SVT.isFloatingPoint()) {
11747 // If BUILD_VECTOR are from built from integer, they may have different
11748 // operand types. Get the smallest type and truncate all operands to it.
11749 bool FoundMinVT = false;
11750 for (const SDValue &Op : N->ops())
11751 if (ISD::BUILD_VECTOR == Op.getOpcode()) {
11752 EVT OpSVT = Op.getOperand(0)->getValueType(0);
11753 MinVT = (!FoundMinVT || OpSVT.bitsLE(MinVT)) ? OpSVT : MinVT;
11756 assert(FoundMinVT && "Concat vector type mismatch");
11759 for (const SDValue &Op : N->ops()) {
11760 EVT OpVT = Op.getValueType();
11761 unsigned NumElts = OpVT.getVectorNumElements();
11763 if (ISD::UNDEF == Op.getOpcode())
11764 Opnds.append(NumElts, DAG.getUNDEF(MinVT));
11766 if (ISD::BUILD_VECTOR == Op.getOpcode()) {
11767 if (SVT.isFloatingPoint()) {
11768 assert(SVT == OpVT.getScalarType() && "Concat vector type mismatch");
11769 Opnds.append(Op->op_begin(), Op->op_begin() + NumElts);
11771 for (unsigned i = 0; i != NumElts; ++i)
11773 DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinVT, Op.getOperand(i)));
11778 assert(VT.getVectorNumElements() == Opnds.size() &&
11779 "Concat vector type mismatch");
11780 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
11783 // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
11784 // nodes often generate nop CONCAT_VECTOR nodes.
11785 // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that
11786 // place the incoming vectors at the exact same location.
11787 SDValue SingleSource = SDValue();
11788 unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements();
11790 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
11791 SDValue Op = N->getOperand(i);
11793 if (Op.getOpcode() == ISD::UNDEF)
11796 // Check if this is the identity extract:
11797 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
11800 // Find the single incoming vector for the extract_subvector.
11801 if (SingleSource.getNode()) {
11802 if (Op.getOperand(0) != SingleSource)
11805 SingleSource = Op.getOperand(0);
11807 // Check the source type is the same as the type of the result.
11808 // If not, this concat may extend the vector, so we can not
11809 // optimize it away.
11810 if (SingleSource.getValueType() != N->getValueType(0))
11814 unsigned IdentityIndex = i * PartNumElem;
11815 ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1));
11816 // The extract index must be constant.
11820 // Check that we are reading from the identity index.
11821 if (CS->getZExtValue() != IdentityIndex)
11825 if (SingleSource.getNode())
11826 return SingleSource;
11831 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
11832 EVT NVT = N->getValueType(0);
11833 SDValue V = N->getOperand(0);
11835 if (V->getOpcode() == ISD::CONCAT_VECTORS) {
11837 // (extract_subvec (concat V1, V2, ...), i)
11840 // Only operand 0 is checked as 'concat' assumes all inputs of the same
11842 if (V->getOperand(0).getValueType() != NVT)
11844 unsigned Idx = dyn_cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
11845 unsigned NumElems = NVT.getVectorNumElements();
11846 assert((Idx % NumElems) == 0 &&
11847 "IDX in concat is not a multiple of the result vector length.");
11848 return V->getOperand(Idx / NumElems);
11852 if (V->getOpcode() == ISD::BITCAST)
11853 V = V.getOperand(0);
11855 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
11857 // Handle only simple case where vector being inserted and vector
11858 // being extracted are of same type, and are half size of larger vectors.
11859 EVT BigVT = V->getOperand(0).getValueType();
11860 EVT SmallVT = V->getOperand(1).getValueType();
11861 if (!NVT.bitsEq(SmallVT) || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
11864 // Only handle cases where both indexes are constants with the same type.
11865 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
11866 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
11868 if (InsIdx && ExtIdx &&
11869 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
11870 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
11872 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
11874 // indices are equal or bit offsets are equal => V1
11875 // otherwise => (extract_subvec V1, ExtIdx)
11876 if (InsIdx->getZExtValue() * SmallVT.getScalarType().getSizeInBits() ==
11877 ExtIdx->getZExtValue() * NVT.getScalarType().getSizeInBits())
11878 return DAG.getNode(ISD::BITCAST, dl, NVT, V->getOperand(1));
11879 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT,
11880 DAG.getNode(ISD::BITCAST, dl,
11881 N->getOperand(0).getValueType(),
11882 V->getOperand(0)), N->getOperand(1));
11889 static SDValue simplifyShuffleOperandRecursively(SmallBitVector &UsedElements,
11890 SDValue V, SelectionDAG &DAG) {
11892 EVT VT = V.getValueType();
11894 switch (V.getOpcode()) {
11898 case ISD::CONCAT_VECTORS: {
11899 EVT OpVT = V->getOperand(0).getValueType();
11900 int OpSize = OpVT.getVectorNumElements();
11901 SmallBitVector OpUsedElements(OpSize, false);
11902 bool FoundSimplification = false;
11903 SmallVector<SDValue, 4> NewOps;
11904 NewOps.reserve(V->getNumOperands());
11905 for (int i = 0, NumOps = V->getNumOperands(); i < NumOps; ++i) {
11906 SDValue Op = V->getOperand(i);
11907 bool OpUsed = false;
11908 for (int j = 0; j < OpSize; ++j)
11909 if (UsedElements[i * OpSize + j]) {
11910 OpUsedElements[j] = true;
11914 OpUsed ? simplifyShuffleOperandRecursively(OpUsedElements, Op, DAG)
11915 : DAG.getUNDEF(OpVT));
11916 FoundSimplification |= Op == NewOps.back();
11917 OpUsedElements.reset();
11919 if (FoundSimplification)
11920 V = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, NewOps);
11924 case ISD::INSERT_SUBVECTOR: {
11925 SDValue BaseV = V->getOperand(0);
11926 SDValue SubV = V->getOperand(1);
11927 auto *IdxN = dyn_cast<ConstantSDNode>(V->getOperand(2));
11931 int SubSize = SubV.getValueType().getVectorNumElements();
11932 int Idx = IdxN->getZExtValue();
11933 bool SubVectorUsed = false;
11934 SmallBitVector SubUsedElements(SubSize, false);
11935 for (int i = 0; i < SubSize; ++i)
11936 if (UsedElements[i + Idx]) {
11937 SubVectorUsed = true;
11938 SubUsedElements[i] = true;
11939 UsedElements[i + Idx] = false;
11942 // Now recurse on both the base and sub vectors.
11943 SDValue SimplifiedSubV =
11945 ? simplifyShuffleOperandRecursively(SubUsedElements, SubV, DAG)
11946 : DAG.getUNDEF(SubV.getValueType());
11947 SDValue SimplifiedBaseV = simplifyShuffleOperandRecursively(UsedElements, BaseV, DAG);
11948 if (SimplifiedSubV != SubV || SimplifiedBaseV != BaseV)
11949 V = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT,
11950 SimplifiedBaseV, SimplifiedSubV, V->getOperand(2));
11956 static SDValue simplifyShuffleOperands(ShuffleVectorSDNode *SVN, SDValue N0,
11957 SDValue N1, SelectionDAG &DAG) {
11958 EVT VT = SVN->getValueType(0);
11959 int NumElts = VT.getVectorNumElements();
11960 SmallBitVector N0UsedElements(NumElts, false), N1UsedElements(NumElts, false);
11961 for (int M : SVN->getMask())
11962 if (M >= 0 && M < NumElts)
11963 N0UsedElements[M] = true;
11964 else if (M >= NumElts)
11965 N1UsedElements[M - NumElts] = true;
11967 SDValue S0 = simplifyShuffleOperandRecursively(N0UsedElements, N0, DAG);
11968 SDValue S1 = simplifyShuffleOperandRecursively(N1UsedElements, N1, DAG);
11969 if (S0 == N0 && S1 == N1)
11972 return DAG.getVectorShuffle(VT, SDLoc(SVN), S0, S1, SVN->getMask());
11975 // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat,
11976 // or turn a shuffle of a single concat into simpler shuffle then concat.
11977 static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
11978 EVT VT = N->getValueType(0);
11979 unsigned NumElts = VT.getVectorNumElements();
11981 SDValue N0 = N->getOperand(0);
11982 SDValue N1 = N->getOperand(1);
11983 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
11985 SmallVector<SDValue, 4> Ops;
11986 EVT ConcatVT = N0.getOperand(0).getValueType();
11987 unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements();
11988 unsigned NumConcats = NumElts / NumElemsPerConcat;
11990 // Special case: shuffle(concat(A,B)) can be more efficiently represented
11991 // as concat(shuffle(A,B),UNDEF) if the shuffle doesn't set any of the high
11992 // half vector elements.
11993 if (NumElemsPerConcat * 2 == NumElts && N1.getOpcode() == ISD::UNDEF &&
11994 std::all_of(SVN->getMask().begin() + NumElemsPerConcat,
11995 SVN->getMask().end(), [](int i) { return i == -1; })) {
11996 N0 = DAG.getVectorShuffle(ConcatVT, SDLoc(N), N0.getOperand(0), N0.getOperand(1),
11997 ArrayRef<int>(SVN->getMask().begin(), NumElemsPerConcat));
11998 N1 = DAG.getUNDEF(ConcatVT);
11999 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, N0, N1);
12002 // Look at every vector that's inserted. We're looking for exact
12003 // subvector-sized copies from a concatenated vector
12004 for (unsigned I = 0; I != NumConcats; ++I) {
12005 // Make sure we're dealing with a copy.
12006 unsigned Begin = I * NumElemsPerConcat;
12007 bool AllUndef = true, NoUndef = true;
12008 for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) {
12009 if (SVN->getMaskElt(J) >= 0)
12016 if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0)
12019 for (unsigned J = 1; J != NumElemsPerConcat; ++J)
12020 if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J))
12023 unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat;
12024 if (FirstElt < N0.getNumOperands())
12025 Ops.push_back(N0.getOperand(FirstElt));
12027 Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands()));
12029 } else if (AllUndef) {
12030 Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType()));
12031 } else { // Mixed with general masks and undefs, can't do optimization.
12036 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops);
12039 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
12040 EVT VT = N->getValueType(0);
12041 unsigned NumElts = VT.getVectorNumElements();
12043 SDValue N0 = N->getOperand(0);
12044 SDValue N1 = N->getOperand(1);
12046 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
12048 // Canonicalize shuffle undef, undef -> undef
12049 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
12050 return DAG.getUNDEF(VT);
12052 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
12054 // Canonicalize shuffle v, v -> v, undef
12056 SmallVector<int, 8> NewMask;
12057 for (unsigned i = 0; i != NumElts; ++i) {
12058 int Idx = SVN->getMaskElt(i);
12059 if (Idx >= (int)NumElts) Idx -= NumElts;
12060 NewMask.push_back(Idx);
12062 return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT),
12066 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
12067 if (N0.getOpcode() == ISD::UNDEF) {
12068 SmallVector<int, 8> NewMask;
12069 for (unsigned i = 0; i != NumElts; ++i) {
12070 int Idx = SVN->getMaskElt(i);
12072 if (Idx >= (int)NumElts)
12075 Idx = -1; // remove reference to lhs
12077 NewMask.push_back(Idx);
12079 return DAG.getVectorShuffle(VT, SDLoc(N), N1, DAG.getUNDEF(VT),
12083 // Remove references to rhs if it is undef
12084 if (N1.getOpcode() == ISD::UNDEF) {
12085 bool Changed = false;
12086 SmallVector<int, 8> NewMask;
12087 for (unsigned i = 0; i != NumElts; ++i) {
12088 int Idx = SVN->getMaskElt(i);
12089 if (Idx >= (int)NumElts) {
12093 NewMask.push_back(Idx);
12096 return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, &NewMask[0]);
12099 // If it is a splat, check if the argument vector is another splat or a
12101 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
12102 SDNode *V = N0.getNode();
12104 // If this is a bit convert that changes the element type of the vector but
12105 // not the number of vector elements, look through it. Be careful not to
12106 // look though conversions that change things like v4f32 to v2f64.
12107 if (V->getOpcode() == ISD::BITCAST) {
12108 SDValue ConvInput = V->getOperand(0);
12109 if (ConvInput.getValueType().isVector() &&
12110 ConvInput.getValueType().getVectorNumElements() == NumElts)
12111 V = ConvInput.getNode();
12114 if (V->getOpcode() == ISD::BUILD_VECTOR) {
12115 assert(V->getNumOperands() == NumElts &&
12116 "BUILD_VECTOR has wrong number of operands");
12118 bool AllSame = true;
12119 for (unsigned i = 0; i != NumElts; ++i) {
12120 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
12121 Base = V->getOperand(i);
12125 // Splat of <u, u, u, u>, return <u, u, u, u>
12126 if (!Base.getNode())
12128 for (unsigned i = 0; i != NumElts; ++i) {
12129 if (V->getOperand(i) != Base) {
12134 // Splat of <x, x, x, x>, return <x, x, x, x>
12138 // Canonicalize any other splat as a build_vector.
12139 const SDValue &Splatted = V->getOperand(SVN->getSplatIndex());
12140 SmallVector<SDValue, 8> Ops(NumElts, Splatted);
12141 SDValue NewBV = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
12142 V->getValueType(0), Ops);
12144 // We may have jumped through bitcasts, so the type of the
12145 // BUILD_VECTOR may not match the type of the shuffle.
12146 if (V->getValueType(0) != VT)
12147 NewBV = DAG.getNode(ISD::BITCAST, SDLoc(N), VT, NewBV);
12152 // There are various patterns used to build up a vector from smaller vectors,
12153 // subvectors, or elements. Scan chains of these and replace unused insertions
12154 // or components with undef.
12155 if (SDValue S = simplifyShuffleOperands(SVN, N0, N1, DAG))
12158 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
12159 Level < AfterLegalizeVectorOps &&
12160 (N1.getOpcode() == ISD::UNDEF ||
12161 (N1.getOpcode() == ISD::CONCAT_VECTORS &&
12162 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) {
12163 SDValue V = partitionShuffleOfConcats(N, DAG);
12169 // Attempt to combine a shuffle of 2 inputs of 'scalar sources' -
12170 // BUILD_VECTOR or SCALAR_TO_VECTOR into a single BUILD_VECTOR.
12171 if (Level < AfterLegalizeVectorOps && TLI.isTypeLegal(VT)) {
12172 SmallVector<SDValue, 8> Ops;
12173 for (int M : SVN->getMask()) {
12174 SDValue Op = DAG.getUNDEF(VT.getScalarType());
12176 int Idx = M % NumElts;
12177 SDValue &S = (M < (int)NumElts ? N0 : N1);
12178 if (S.getOpcode() == ISD::BUILD_VECTOR && S.hasOneUse()) {
12179 Op = S.getOperand(Idx);
12180 } else if (S.getOpcode() == ISD::SCALAR_TO_VECTOR && S.hasOneUse()) {
12182 Op = S.getOperand(0);
12184 // Operand can't be combined - bail out.
12190 if (Ops.size() == VT.getVectorNumElements()) {
12191 // BUILD_VECTOR requires all inputs to be of the same type, find the
12192 // maximum type and extend them all.
12193 EVT SVT = VT.getScalarType();
12194 if (SVT.isInteger())
12195 for (SDValue &Op : Ops)
12196 SVT = (SVT.bitsLT(Op.getValueType()) ? Op.getValueType() : SVT);
12197 if (SVT != VT.getScalarType())
12198 for (SDValue &Op : Ops)
12199 Op = TLI.isZExtFree(Op.getValueType(), SVT)
12200 ? DAG.getZExtOrTrunc(Op, SDLoc(N), SVT)
12201 : DAG.getSExtOrTrunc(Op, SDLoc(N), SVT);
12202 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Ops);
12206 // If this shuffle only has a single input that is a bitcasted shuffle,
12207 // attempt to merge the 2 shuffles and suitably bitcast the inputs/output
12208 // back to their original types.
12209 if (N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
12210 N1.getOpcode() == ISD::UNDEF && Level < AfterLegalizeVectorOps &&
12211 TLI.isTypeLegal(VT)) {
12213 // Peek through the bitcast only if there is one user.
12215 while (BC0.getOpcode() == ISD::BITCAST) {
12216 if (!BC0.hasOneUse())
12218 BC0 = BC0.getOperand(0);
12221 auto ScaleShuffleMask = [](ArrayRef<int> Mask, int Scale) {
12223 return SmallVector<int, 8>(Mask.begin(), Mask.end());
12225 SmallVector<int, 8> NewMask;
12227 for (int s = 0; s != Scale; ++s)
12228 NewMask.push_back(M < 0 ? -1 : Scale * M + s);
12232 if (BC0.getOpcode() == ISD::VECTOR_SHUFFLE && BC0.hasOneUse()) {
12233 EVT SVT = VT.getScalarType();
12234 EVT InnerVT = BC0->getValueType(0);
12235 EVT InnerSVT = InnerVT.getScalarType();
12237 // Determine which shuffle works with the smaller scalar type.
12238 EVT ScaleVT = SVT.bitsLT(InnerSVT) ? VT : InnerVT;
12239 EVT ScaleSVT = ScaleVT.getScalarType();
12241 if (TLI.isTypeLegal(ScaleVT) &&
12242 0 == (InnerSVT.getSizeInBits() % ScaleSVT.getSizeInBits()) &&
12243 0 == (SVT.getSizeInBits() % ScaleSVT.getSizeInBits())) {
12245 int InnerScale = InnerSVT.getSizeInBits() / ScaleSVT.getSizeInBits();
12246 int OuterScale = SVT.getSizeInBits() / ScaleSVT.getSizeInBits();
12248 // Scale the shuffle masks to the smaller scalar type.
12249 ShuffleVectorSDNode *InnerSVN = cast<ShuffleVectorSDNode>(BC0);
12250 SmallVector<int, 8> InnerMask =
12251 ScaleShuffleMask(InnerSVN->getMask(), InnerScale);
12252 SmallVector<int, 8> OuterMask =
12253 ScaleShuffleMask(SVN->getMask(), OuterScale);
12255 // Merge the shuffle masks.
12256 SmallVector<int, 8> NewMask;
12257 for (int M : OuterMask)
12258 NewMask.push_back(M < 0 ? -1 : InnerMask[M]);
12260 // Test for shuffle mask legality over both commutations.
12261 SDValue SV0 = BC0->getOperand(0);
12262 SDValue SV1 = BC0->getOperand(1);
12263 bool LegalMask = TLI.isShuffleMaskLegal(NewMask, ScaleVT);
12265 std::swap(SV0, SV1);
12266 ShuffleVectorSDNode::commuteMask(NewMask);
12267 LegalMask = TLI.isShuffleMaskLegal(NewMask, ScaleVT);
12271 SV0 = DAG.getNode(ISD::BITCAST, SDLoc(N), ScaleVT, SV0);
12272 SV1 = DAG.getNode(ISD::BITCAST, SDLoc(N), ScaleVT, SV1);
12273 return DAG.getNode(
12274 ISD::BITCAST, SDLoc(N), VT,
12275 DAG.getVectorShuffle(ScaleVT, SDLoc(N), SV0, SV1, NewMask));
12281 // Canonicalize shuffles according to rules:
12282 // shuffle(A, shuffle(A, B)) -> shuffle(shuffle(A,B), A)
12283 // shuffle(B, shuffle(A, B)) -> shuffle(shuffle(A,B), B)
12284 // shuffle(B, shuffle(A, Undef)) -> shuffle(shuffle(A, Undef), B)
12285 if (N1.getOpcode() == ISD::VECTOR_SHUFFLE &&
12286 N0.getOpcode() != ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
12287 TLI.isTypeLegal(VT)) {
12288 // The incoming shuffle must be of the same type as the result of the
12289 // current shuffle.
12290 assert(N1->getOperand(0).getValueType() == VT &&
12291 "Shuffle types don't match");
12293 SDValue SV0 = N1->getOperand(0);
12294 SDValue SV1 = N1->getOperand(1);
12295 bool HasSameOp0 = N0 == SV0;
12296 bool IsSV1Undef = SV1.getOpcode() == ISD::UNDEF;
12297 if (HasSameOp0 || IsSV1Undef || N0 == SV1)
12298 // Commute the operands of this shuffle so that next rule
12300 return DAG.getCommutedVectorShuffle(*SVN);
12303 // Try to fold according to rules:
12304 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2)
12305 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2)
12306 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2)
12307 // Don't try to fold shuffles with illegal type.
12308 // Only fold if this shuffle is the only user of the other shuffle.
12309 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && N->isOnlyUserOf(N0.getNode()) &&
12310 Level < AfterLegalizeDAG && TLI.isTypeLegal(VT)) {
12311 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
12313 // The incoming shuffle must be of the same type as the result of the
12314 // current shuffle.
12315 assert(OtherSV->getOperand(0).getValueType() == VT &&
12316 "Shuffle types don't match");
12319 SmallVector<int, 4> Mask;
12320 // Compute the combined shuffle mask for a shuffle with SV0 as the first
12321 // operand, and SV1 as the second operand.
12322 for (unsigned i = 0; i != NumElts; ++i) {
12323 int Idx = SVN->getMaskElt(i);
12325 // Propagate Undef.
12326 Mask.push_back(Idx);
12330 SDValue CurrentVec;
12331 if (Idx < (int)NumElts) {
12332 // This shuffle index refers to the inner shuffle N0. Lookup the inner
12333 // shuffle mask to identify which vector is actually referenced.
12334 Idx = OtherSV->getMaskElt(Idx);
12336 // Propagate Undef.
12337 Mask.push_back(Idx);
12341 CurrentVec = (Idx < (int) NumElts) ? OtherSV->getOperand(0)
12342 : OtherSV->getOperand(1);
12344 // This shuffle index references an element within N1.
12348 // Simple case where 'CurrentVec' is UNDEF.
12349 if (CurrentVec.getOpcode() == ISD::UNDEF) {
12350 Mask.push_back(-1);
12354 // Canonicalize the shuffle index. We don't know yet if CurrentVec
12355 // will be the first or second operand of the combined shuffle.
12356 Idx = Idx % NumElts;
12357 if (!SV0.getNode() || SV0 == CurrentVec) {
12358 // Ok. CurrentVec is the left hand side.
12359 // Update the mask accordingly.
12361 Mask.push_back(Idx);
12365 // Bail out if we cannot convert the shuffle pair into a single shuffle.
12366 if (SV1.getNode() && SV1 != CurrentVec)
12369 // Ok. CurrentVec is the right hand side.
12370 // Update the mask accordingly.
12372 Mask.push_back(Idx + NumElts);
12375 // Check if all indices in Mask are Undef. In case, propagate Undef.
12376 bool isUndefMask = true;
12377 for (unsigned i = 0; i != NumElts && isUndefMask; ++i)
12378 isUndefMask &= Mask[i] < 0;
12381 return DAG.getUNDEF(VT);
12383 if (!SV0.getNode())
12384 SV0 = DAG.getUNDEF(VT);
12385 if (!SV1.getNode())
12386 SV1 = DAG.getUNDEF(VT);
12388 // Avoid introducing shuffles with illegal mask.
12389 if (!TLI.isShuffleMaskLegal(Mask, VT)) {
12390 ShuffleVectorSDNode::commuteMask(Mask);
12392 if (!TLI.isShuffleMaskLegal(Mask, VT))
12395 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, A, M2)
12396 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, A, M2)
12397 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, B, M2)
12398 std::swap(SV0, SV1);
12401 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2)
12402 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2)
12403 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2)
12404 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, &Mask[0]);
12410 SDValue DAGCombiner::visitSCALAR_TO_VECTOR(SDNode *N) {
12411 SDValue InVal = N->getOperand(0);
12412 EVT VT = N->getValueType(0);
12414 // Replace a SCALAR_TO_VECTOR(EXTRACT_VECTOR_ELT(V,C0)) pattern
12415 // with a VECTOR_SHUFFLE.
12416 if (InVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
12417 SDValue InVec = InVal->getOperand(0);
12418 SDValue EltNo = InVal->getOperand(1);
12420 // FIXME: We could support implicit truncation if the shuffle can be
12421 // scaled to a smaller vector scalar type.
12422 ConstantSDNode *C0 = dyn_cast<ConstantSDNode>(EltNo);
12423 if (C0 && VT == InVec.getValueType() &&
12424 VT.getScalarType() == InVal.getValueType()) {
12425 SmallVector<int, 8> NewMask(VT.getVectorNumElements(), -1);
12426 int Elt = C0->getZExtValue();
12429 if (TLI.isShuffleMaskLegal(NewMask, VT))
12430 return DAG.getVectorShuffle(VT, SDLoc(N), InVec, DAG.getUNDEF(VT),
12438 SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
12439 SDValue N0 = N->getOperand(0);
12440 SDValue N2 = N->getOperand(2);
12442 // If the input vector is a concatenation, and the insert replaces
12443 // one of the halves, we can optimize into a single concat_vectors.
12444 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
12445 N0->getNumOperands() == 2 && N2.getOpcode() == ISD::Constant) {
12446 APInt InsIdx = cast<ConstantSDNode>(N2)->getAPIntValue();
12447 EVT VT = N->getValueType(0);
12449 // Lower half: fold (insert_subvector (concat_vectors X, Y), Z) ->
12450 // (concat_vectors Z, Y)
12452 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
12453 N->getOperand(1), N0.getOperand(1));
12455 // Upper half: fold (insert_subvector (concat_vectors X, Y), Z) ->
12456 // (concat_vectors X, Z)
12457 if (InsIdx == VT.getVectorNumElements()/2)
12458 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
12459 N0.getOperand(0), N->getOperand(1));
12465 /// Returns a vector_shuffle if it able to transform an AND to a vector_shuffle
12466 /// with the destination vector and a zero vector.
12467 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
12468 /// vector_shuffle V, Zero, <0, 4, 2, 4>
12469 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
12470 EVT VT = N->getValueType(0);
12471 SDValue LHS = N->getOperand(0);
12472 SDValue RHS = N->getOperand(1);
12475 // Make sure we're not running after operation legalization where it
12476 // may have custom lowered the vector shuffles.
12477 if (LegalOperations)
12480 if (N->getOpcode() != ISD::AND)
12483 if (RHS.getOpcode() == ISD::BITCAST)
12484 RHS = RHS.getOperand(0);
12486 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
12487 SmallVector<int, 8> Indices;
12488 unsigned NumElts = RHS.getNumOperands();
12490 for (unsigned i = 0; i != NumElts; ++i) {
12491 SDValue Elt = RHS.getOperand(i);
12492 if (!isa<ConstantSDNode>(Elt))
12495 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
12496 Indices.push_back(i);
12497 else if (cast<ConstantSDNode>(Elt)->isNullValue())
12498 Indices.push_back(NumElts+i);
12503 // Let's see if the target supports this vector_shuffle.
12504 EVT RVT = RHS.getValueType();
12505 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
12508 // Return the new VECTOR_SHUFFLE node.
12509 EVT EltVT = RVT.getVectorElementType();
12510 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
12511 DAG.getConstant(0, EltVT));
12512 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), RVT, ZeroOps);
12513 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
12514 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
12515 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
12521 /// Visit a binary vector operation, like ADD.
12522 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
12523 assert(N->getValueType(0).isVector() &&
12524 "SimplifyVBinOp only works on vectors!");
12526 SDValue LHS = N->getOperand(0);
12527 SDValue RHS = N->getOperand(1);
12529 if (SDValue Shuffle = XformToShuffleWithZero(N))
12532 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
12534 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
12535 RHS.getOpcode() == ISD::BUILD_VECTOR) {
12536 // Check if both vectors are constants. If not bail out.
12537 if (!(cast<BuildVectorSDNode>(LHS)->isConstant() &&
12538 cast<BuildVectorSDNode>(RHS)->isConstant()))
12541 SmallVector<SDValue, 8> Ops;
12542 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
12543 SDValue LHSOp = LHS.getOperand(i);
12544 SDValue RHSOp = RHS.getOperand(i);
12546 // Can't fold divide by zero.
12547 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
12548 N->getOpcode() == ISD::FDIV) {
12549 if ((RHSOp.getOpcode() == ISD::Constant &&
12550 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
12551 (RHSOp.getOpcode() == ISD::ConstantFP &&
12552 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
12556 EVT VT = LHSOp.getValueType();
12557 EVT RVT = RHSOp.getValueType();
12559 // Integer BUILD_VECTOR operands may have types larger than the element
12560 // size (e.g., when the element type is not legal). Prior to type
12561 // legalization, the types may not match between the two BUILD_VECTORS.
12562 // Truncate one of the operands to make them match.
12563 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
12564 RHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, RHSOp);
12566 LHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), RVT, LHSOp);
12570 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(LHS), VT,
12572 if (FoldOp.getOpcode() != ISD::UNDEF &&
12573 FoldOp.getOpcode() != ISD::Constant &&
12574 FoldOp.getOpcode() != ISD::ConstantFP)
12576 Ops.push_back(FoldOp);
12577 AddToWorklist(FoldOp.getNode());
12580 if (Ops.size() == LHS.getNumOperands())
12581 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), LHS.getValueType(), Ops);
12584 // Type legalization might introduce new shuffles in the DAG.
12585 // Fold (VBinOp (shuffle (A, Undef, Mask)), (shuffle (B, Undef, Mask)))
12586 // -> (shuffle (VBinOp (A, B)), Undef, Mask).
12587 if (LegalTypes && isa<ShuffleVectorSDNode>(LHS) &&
12588 isa<ShuffleVectorSDNode>(RHS) && LHS.hasOneUse() && RHS.hasOneUse() &&
12589 LHS.getOperand(1).getOpcode() == ISD::UNDEF &&
12590 RHS.getOperand(1).getOpcode() == ISD::UNDEF) {
12591 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(LHS);
12592 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(RHS);
12594 if (SVN0->getMask().equals(SVN1->getMask())) {
12595 EVT VT = N->getValueType(0);
12596 SDValue UndefVector = LHS.getOperand(1);
12597 SDValue NewBinOp = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
12598 LHS.getOperand(0), RHS.getOperand(0));
12599 AddUsersToWorklist(N);
12600 return DAG.getVectorShuffle(VT, SDLoc(N), NewBinOp, UndefVector,
12601 &SVN0->getMask()[0]);
12608 SDValue DAGCombiner::SimplifySelect(SDLoc DL, SDValue N0,
12609 SDValue N1, SDValue N2){
12610 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
12612 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
12613 cast<CondCodeSDNode>(N0.getOperand(2))->get());
12615 // If we got a simplified select_cc node back from SimplifySelectCC, then
12616 // break it down into a new SETCC node, and a new SELECT node, and then return
12617 // the SELECT node, since we were called with a SELECT node.
12618 if (SCC.getNode()) {
12619 // Check to see if we got a select_cc back (to turn into setcc/select).
12620 // Otherwise, just return whatever node we got back, like fabs.
12621 if (SCC.getOpcode() == ISD::SELECT_CC) {
12622 SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0),
12624 SCC.getOperand(0), SCC.getOperand(1),
12625 SCC.getOperand(4));
12626 AddToWorklist(SETCC.getNode());
12627 return DAG.getSelect(SDLoc(SCC), SCC.getValueType(), SETCC,
12628 SCC.getOperand(2), SCC.getOperand(3));
12636 /// Given a SELECT or a SELECT_CC node, where LHS and RHS are the two values
12637 /// being selected between, see if we can simplify the select. Callers of this
12638 /// should assume that TheSelect is deleted if this returns true. As such, they
12639 /// should return the appropriate thing (e.g. the node) back to the top-level of
12640 /// the DAG combiner loop to avoid it being looked at.
12641 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
12644 // Cannot simplify select with vector condition
12645 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
12647 // If this is a select from two identical things, try to pull the operation
12648 // through the select.
12649 if (LHS.getOpcode() != RHS.getOpcode() ||
12650 !LHS.hasOneUse() || !RHS.hasOneUse())
12653 // If this is a load and the token chain is identical, replace the select
12654 // of two loads with a load through a select of the address to load from.
12655 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
12656 // constants have been dropped into the constant pool.
12657 if (LHS.getOpcode() == ISD::LOAD) {
12658 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
12659 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
12661 // Token chains must be identical.
12662 if (LHS.getOperand(0) != RHS.getOperand(0) ||
12663 // Do not let this transformation reduce the number of volatile loads.
12664 LLD->isVolatile() || RLD->isVolatile() ||
12665 // If this is an EXTLOAD, the VT's must match.
12666 LLD->getMemoryVT() != RLD->getMemoryVT() ||
12667 // If this is an EXTLOAD, the kind of extension must match.
12668 (LLD->getExtensionType() != RLD->getExtensionType() &&
12669 // The only exception is if one of the extensions is anyext.
12670 LLD->getExtensionType() != ISD::EXTLOAD &&
12671 RLD->getExtensionType() != ISD::EXTLOAD) ||
12672 // FIXME: this discards src value information. This is
12673 // over-conservative. It would be beneficial to be able to remember
12674 // both potential memory locations. Since we are discarding
12675 // src value info, don't do the transformation if the memory
12676 // locations are not in the default address space.
12677 LLD->getPointerInfo().getAddrSpace() != 0 ||
12678 RLD->getPointerInfo().getAddrSpace() != 0 ||
12679 !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(),
12680 LLD->getBasePtr().getValueType()))
12683 // Check that the select condition doesn't reach either load. If so,
12684 // folding this will induce a cycle into the DAG. If not, this is safe to
12685 // xform, so create a select of the addresses.
12687 if (TheSelect->getOpcode() == ISD::SELECT) {
12688 SDNode *CondNode = TheSelect->getOperand(0).getNode();
12689 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
12690 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
12692 // The loads must not depend on one another.
12693 if (LLD->isPredecessorOf(RLD) ||
12694 RLD->isPredecessorOf(LLD))
12696 Addr = DAG.getSelect(SDLoc(TheSelect),
12697 LLD->getBasePtr().getValueType(),
12698 TheSelect->getOperand(0), LLD->getBasePtr(),
12699 RLD->getBasePtr());
12700 } else { // Otherwise SELECT_CC
12701 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
12702 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
12704 if ((LLD->hasAnyUseOfValue(1) &&
12705 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
12706 (RLD->hasAnyUseOfValue(1) &&
12707 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
12710 Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect),
12711 LLD->getBasePtr().getValueType(),
12712 TheSelect->getOperand(0),
12713 TheSelect->getOperand(1),
12714 LLD->getBasePtr(), RLD->getBasePtr(),
12715 TheSelect->getOperand(4));
12719 // It is safe to replace the two loads if they have different alignments,
12720 // but the new load must be the minimum (most restrictive) alignment of the
12722 bool isInvariant = LLD->isInvariant() & RLD->isInvariant();
12723 unsigned Alignment = std::min(LLD->getAlignment(), RLD->getAlignment());
12724 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
12725 Load = DAG.getLoad(TheSelect->getValueType(0),
12727 // FIXME: Discards pointer and AA info.
12728 LLD->getChain(), Addr, MachinePointerInfo(),
12729 LLD->isVolatile(), LLD->isNonTemporal(),
12730 isInvariant, Alignment);
12732 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
12733 RLD->getExtensionType() : LLD->getExtensionType(),
12735 TheSelect->getValueType(0),
12736 // FIXME: Discards pointer and AA info.
12737 LLD->getChain(), Addr, MachinePointerInfo(),
12738 LLD->getMemoryVT(), LLD->isVolatile(),
12739 LLD->isNonTemporal(), isInvariant, Alignment);
12742 // Users of the select now use the result of the load.
12743 CombineTo(TheSelect, Load);
12745 // Users of the old loads now use the new load's chain. We know the
12746 // old-load value is dead now.
12747 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
12748 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
12755 /// Simplify an expression of the form (N0 cond N1) ? N2 : N3
12756 /// where 'cond' is the comparison specified by CC.
12757 SDValue DAGCombiner::SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1,
12758 SDValue N2, SDValue N3,
12759 ISD::CondCode CC, bool NotExtCompare) {
12760 // (x ? y : y) -> y.
12761 if (N2 == N3) return N2;
12763 EVT VT = N2.getValueType();
12764 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
12765 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
12766 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
12768 // Determine if the condition we're dealing with is constant
12769 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
12770 N0, N1, CC, DL, false);
12771 if (SCC.getNode()) AddToWorklist(SCC.getNode());
12772 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
12774 // fold select_cc true, x, y -> x
12775 if (SCCC && !SCCC->isNullValue())
12777 // fold select_cc false, x, y -> y
12778 if (SCCC && SCCC->isNullValue())
12781 // Check to see if we can simplify the select into an fabs node
12782 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
12783 // Allow either -0.0 or 0.0
12784 if (CFP->getValueAPF().isZero()) {
12785 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
12786 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
12787 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
12788 N2 == N3.getOperand(0))
12789 return DAG.getNode(ISD::FABS, DL, VT, N0);
12791 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
12792 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
12793 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
12794 N2.getOperand(0) == N3)
12795 return DAG.getNode(ISD::FABS, DL, VT, N3);
12799 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
12800 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
12801 // in it. This is a win when the constant is not otherwise available because
12802 // it replaces two constant pool loads with one. We only do this if the FP
12803 // type is known to be legal, because if it isn't, then we are before legalize
12804 // types an we want the other legalization to happen first (e.g. to avoid
12805 // messing with soft float) and if the ConstantFP is not legal, because if
12806 // it is legal, we may not need to store the FP constant in a constant pool.
12807 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
12808 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
12809 if (TLI.isTypeLegal(N2.getValueType()) &&
12810 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
12811 TargetLowering::Legal &&
12812 !TLI.isFPImmLegal(TV->getValueAPF(), TV->getValueType(0)) &&
12813 !TLI.isFPImmLegal(FV->getValueAPF(), FV->getValueType(0))) &&
12814 // If both constants have multiple uses, then we won't need to do an
12815 // extra load, they are likely around in registers for other users.
12816 (TV->hasOneUse() || FV->hasOneUse())) {
12817 Constant *Elts[] = {
12818 const_cast<ConstantFP*>(FV->getConstantFPValue()),
12819 const_cast<ConstantFP*>(TV->getConstantFPValue())
12821 Type *FPTy = Elts[0]->getType();
12822 const DataLayout &TD = *TLI.getDataLayout();
12824 // Create a ConstantArray of the two constants.
12825 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
12826 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
12827 TD.getPrefTypeAlignment(FPTy));
12828 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
12830 // Get the offsets to the 0 and 1 element of the array so that we can
12831 // select between them.
12832 SDValue Zero = DAG.getIntPtrConstant(0);
12833 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
12834 SDValue One = DAG.getIntPtrConstant(EltSize);
12836 SDValue Cond = DAG.getSetCC(DL,
12837 getSetCCResultType(N0.getValueType()),
12839 AddToWorklist(Cond.getNode());
12840 SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(),
12842 AddToWorklist(CstOffset.getNode());
12843 CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx,
12845 AddToWorklist(CPIdx.getNode());
12846 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
12847 MachinePointerInfo::getConstantPool(), false,
12848 false, false, Alignment);
12853 // Check to see if we can perform the "gzip trick", transforming
12854 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
12855 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
12856 (N1C->isNullValue() || // (a < 0) ? b : 0
12857 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
12858 EVT XType = N0.getValueType();
12859 EVT AType = N2.getValueType();
12860 if (XType.bitsGE(AType)) {
12861 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
12862 // single-bit constant.
12863 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
12864 unsigned ShCtV = N2C->getAPIntValue().logBase2();
12865 ShCtV = XType.getSizeInBits()-ShCtV-1;
12866 SDValue ShCt = DAG.getConstant(ShCtV,
12867 getShiftAmountTy(N0.getValueType()));
12868 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0),
12870 AddToWorklist(Shift.getNode());
12872 if (XType.bitsGT(AType)) {
12873 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
12874 AddToWorklist(Shift.getNode());
12877 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
12880 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0),
12882 DAG.getConstant(XType.getSizeInBits()-1,
12883 getShiftAmountTy(N0.getValueType())));
12884 AddToWorklist(Shift.getNode());
12886 if (XType.bitsGT(AType)) {
12887 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
12888 AddToWorklist(Shift.getNode());
12891 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
12895 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
12896 // where y is has a single bit set.
12897 // A plaintext description would be, we can turn the SELECT_CC into an AND
12898 // when the condition can be materialized as an all-ones register. Any
12899 // single bit-test can be materialized as an all-ones register with
12900 // shift-left and shift-right-arith.
12901 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
12902 N0->getValueType(0) == VT &&
12903 N1C && N1C->isNullValue() &&
12904 N2C && N2C->isNullValue()) {
12905 SDValue AndLHS = N0->getOperand(0);
12906 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
12907 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
12908 // Shift the tested bit over the sign bit.
12909 APInt AndMask = ConstAndRHS->getAPIntValue();
12911 DAG.getConstant(AndMask.countLeadingZeros(),
12912 getShiftAmountTy(AndLHS.getValueType()));
12913 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt);
12915 // Now arithmetic right shift it all the way over, so the result is either
12916 // all-ones, or zero.
12918 DAG.getConstant(AndMask.getBitWidth()-1,
12919 getShiftAmountTy(Shl.getValueType()));
12920 SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt);
12922 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
12926 // fold select C, 16, 0 -> shl C, 4
12927 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
12928 TLI.getBooleanContents(N0.getValueType()) ==
12929 TargetLowering::ZeroOrOneBooleanContent) {
12931 // If the caller doesn't want us to simplify this into a zext of a compare,
12933 if (NotExtCompare && N2C->getAPIntValue() == 1)
12936 // Get a SetCC of the condition
12937 // NOTE: Don't create a SETCC if it's not legal on this target.
12938 if (!LegalOperations ||
12939 TLI.isOperationLegal(ISD::SETCC,
12940 LegalTypes ? getSetCCResultType(N0.getValueType()) : MVT::i1)) {
12942 // cast from setcc result type to select result type
12944 SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()),
12946 if (N2.getValueType().bitsLT(SCC.getValueType()))
12947 Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2),
12948 N2.getValueType());
12950 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
12951 N2.getValueType(), SCC);
12953 SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC);
12954 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
12955 N2.getValueType(), SCC);
12958 AddToWorklist(SCC.getNode());
12959 AddToWorklist(Temp.getNode());
12961 if (N2C->getAPIntValue() == 1)
12964 // shl setcc result by log2 n2c
12965 return DAG.getNode(
12966 ISD::SHL, DL, N2.getValueType(), Temp,
12967 DAG.getConstant(N2C->getAPIntValue().logBase2(),
12968 getShiftAmountTy(Temp.getValueType())));
12972 // Check to see if this is the equivalent of setcc
12973 // FIXME: Turn all of these into setcc if setcc if setcc is legal
12974 // otherwise, go ahead with the folds.
12975 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
12976 EVT XType = N0.getValueType();
12977 if (!LegalOperations ||
12978 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(XType))) {
12979 SDValue Res = DAG.getSetCC(DL, getSetCCResultType(XType), N0, N1, CC);
12980 if (Res.getValueType() != VT)
12981 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
12985 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
12986 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
12987 (!LegalOperations ||
12988 TLI.isOperationLegal(ISD::CTLZ, XType))) {
12989 SDValue Ctlz = DAG.getNode(ISD::CTLZ, SDLoc(N0), XType, N0);
12990 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
12991 DAG.getConstant(Log2_32(XType.getSizeInBits()),
12992 getShiftAmountTy(Ctlz.getValueType())));
12994 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
12995 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
12996 SDValue NegN0 = DAG.getNode(ISD::SUB, SDLoc(N0),
12997 XType, DAG.getConstant(0, XType), N0);
12998 SDValue NotN0 = DAG.getNOT(SDLoc(N0), N0, XType);
12999 return DAG.getNode(ISD::SRL, DL, XType,
13000 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
13001 DAG.getConstant(XType.getSizeInBits()-1,
13002 getShiftAmountTy(XType)));
13004 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
13005 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
13006 SDValue Sign = DAG.getNode(ISD::SRL, SDLoc(N0), XType, N0,
13007 DAG.getConstant(XType.getSizeInBits()-1,
13008 getShiftAmountTy(N0.getValueType())));
13009 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
13013 // Check to see if this is an integer abs.
13014 // select_cc setg[te] X, 0, X, -X ->
13015 // select_cc setgt X, -1, X, -X ->
13016 // select_cc setl[te] X, 0, -X, X ->
13017 // select_cc setlt X, 1, -X, X ->
13018 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
13020 ConstantSDNode *SubC = nullptr;
13021 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
13022 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
13023 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
13024 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
13025 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
13026 (N1C->isOne() && CC == ISD::SETLT)) &&
13027 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
13028 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
13030 EVT XType = N0.getValueType();
13031 if (SubC && SubC->isNullValue() && XType.isInteger()) {
13032 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0), XType,
13034 DAG.getConstant(XType.getSizeInBits()-1,
13035 getShiftAmountTy(N0.getValueType())));
13036 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0),
13038 AddToWorklist(Shift.getNode());
13039 AddToWorklist(Add.getNode());
13040 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
13047 /// This is a stub for TargetLowering::SimplifySetCC.
13048 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
13049 SDValue N1, ISD::CondCode Cond,
13050 SDLoc DL, bool foldBooleans) {
13051 TargetLowering::DAGCombinerInfo
13052 DagCombineInfo(DAG, Level, false, this);
13053 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
13056 /// Given an ISD::SDIV node expressing a divide by constant, return
13057 /// a DAG expression to select that will generate the same value by multiplying
13058 /// by a magic number.
13059 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
13060 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
13061 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
13065 // Avoid division by zero.
13066 if (!C->getAPIntValue())
13069 std::vector<SDNode*> Built;
13071 TLI.BuildSDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
13073 for (SDNode *N : Built)
13078 /// Given an ISD::SDIV node expressing a divide by constant power of 2, return a
13079 /// DAG expression that will generate the same value by right shifting.
13080 SDValue DAGCombiner::BuildSDIVPow2(SDNode *N) {
13081 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
13085 // Avoid division by zero.
13086 if (!C->getAPIntValue())
13089 std::vector<SDNode *> Built;
13090 SDValue S = TLI.BuildSDIVPow2(N, C->getAPIntValue(), DAG, &Built);
13092 for (SDNode *N : Built)
13097 /// Given an ISD::UDIV node expressing a divide by constant, return a DAG
13098 /// expression that will generate the same value by multiplying by a magic
13100 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
13101 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
13102 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
13106 // Avoid division by zero.
13107 if (!C->getAPIntValue())
13110 std::vector<SDNode*> Built;
13112 TLI.BuildUDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
13114 for (SDNode *N : Built)
13119 SDValue DAGCombiner::BuildReciprocalEstimate(SDValue Op) {
13120 if (Level >= AfterLegalizeDAG)
13123 // Expose the DAG combiner to the target combiner implementations.
13124 TargetLowering::DAGCombinerInfo DCI(DAG, Level, false, this);
13126 unsigned Iterations = 0;
13127 if (SDValue Est = TLI.getRecipEstimate(Op, DCI, Iterations)) {
13129 // Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
13130 // For the reciprocal, we need to find the zero of the function:
13131 // F(X) = A X - 1 [which has a zero at X = 1/A]
13133 // X_{i+1} = X_i (2 - A X_i) = X_i + X_i (1 - A X_i) [this second form
13134 // does not require additional intermediate precision]
13135 EVT VT = Op.getValueType();
13137 SDValue FPOne = DAG.getConstantFP(1.0, VT);
13139 AddToWorklist(Est.getNode());
13141 // Newton iterations: Est = Est + Est (1 - Arg * Est)
13142 for (unsigned i = 0; i < Iterations; ++i) {
13143 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Op, Est);
13144 AddToWorklist(NewEst.getNode());
13146 NewEst = DAG.getNode(ISD::FSUB, DL, VT, FPOne, NewEst);
13147 AddToWorklist(NewEst.getNode());
13149 NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst);
13150 AddToWorklist(NewEst.getNode());
13152 Est = DAG.getNode(ISD::FADD, DL, VT, Est, NewEst);
13153 AddToWorklist(Est.getNode());
13162 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
13163 /// For the reciprocal sqrt, we need to find the zero of the function:
13164 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
13166 /// X_{i+1} = X_i (1.5 - A X_i^2 / 2)
13167 /// As a result, we precompute A/2 prior to the iteration loop.
13168 SDValue DAGCombiner::BuildRsqrtNROneConst(SDValue Arg, SDValue Est,
13169 unsigned Iterations) {
13170 EVT VT = Arg.getValueType();
13172 SDValue ThreeHalves = DAG.getConstantFP(1.5, VT);
13174 // We now need 0.5 * Arg which we can write as (1.5 * Arg - Arg) so that
13175 // this entire sequence requires only one FP constant.
13176 SDValue HalfArg = DAG.getNode(ISD::FMUL, DL, VT, ThreeHalves, Arg);
13177 AddToWorklist(HalfArg.getNode());
13179 HalfArg = DAG.getNode(ISD::FSUB, DL, VT, HalfArg, Arg);
13180 AddToWorklist(HalfArg.getNode());
13182 // Newton iterations: Est = Est * (1.5 - HalfArg * Est * Est)
13183 for (unsigned i = 0; i < Iterations; ++i) {
13184 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, Est);
13185 AddToWorklist(NewEst.getNode());
13187 NewEst = DAG.getNode(ISD::FMUL, DL, VT, HalfArg, NewEst);
13188 AddToWorklist(NewEst.getNode());
13190 NewEst = DAG.getNode(ISD::FSUB, DL, VT, ThreeHalves, NewEst);
13191 AddToWorklist(NewEst.getNode());
13193 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst);
13194 AddToWorklist(Est.getNode());
13199 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
13200 /// For the reciprocal sqrt, we need to find the zero of the function:
13201 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
13203 /// X_{i+1} = (-0.5 * X_i) * (A * X_i * X_i + (-3.0))
13204 SDValue DAGCombiner::BuildRsqrtNRTwoConst(SDValue Arg, SDValue Est,
13205 unsigned Iterations) {
13206 EVT VT = Arg.getValueType();
13208 SDValue MinusThree = DAG.getConstantFP(-3.0, VT);
13209 SDValue MinusHalf = DAG.getConstantFP(-0.5, VT);
13211 // Newton iterations: Est = -0.5 * Est * (-3.0 + Arg * Est * Est)
13212 for (unsigned i = 0; i < Iterations; ++i) {
13213 SDValue HalfEst = DAG.getNode(ISD::FMUL, DL, VT, Est, MinusHalf);
13214 AddToWorklist(HalfEst.getNode());
13216 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Est);
13217 AddToWorklist(Est.getNode());
13219 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Arg);
13220 AddToWorklist(Est.getNode());
13222 Est = DAG.getNode(ISD::FADD, DL, VT, Est, MinusThree);
13223 AddToWorklist(Est.getNode());
13225 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, HalfEst);
13226 AddToWorklist(Est.getNode());
13231 SDValue DAGCombiner::BuildRsqrtEstimate(SDValue Op) {
13232 if (Level >= AfterLegalizeDAG)
13235 // Expose the DAG combiner to the target combiner implementations.
13236 TargetLowering::DAGCombinerInfo DCI(DAG, Level, false, this);
13237 unsigned Iterations = 0;
13238 bool UseOneConstNR = false;
13239 if (SDValue Est = TLI.getRsqrtEstimate(Op, DCI, Iterations, UseOneConstNR)) {
13240 AddToWorklist(Est.getNode());
13242 Est = UseOneConstNR ?
13243 BuildRsqrtNROneConst(Op, Est, Iterations) :
13244 BuildRsqrtNRTwoConst(Op, Est, Iterations);
13252 /// Return true if base is a frame index, which is known not to alias with
13253 /// anything but itself. Provides base object and offset as results.
13254 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
13255 const GlobalValue *&GV, const void *&CV) {
13256 // Assume it is a primitive operation.
13257 Base = Ptr; Offset = 0; GV = nullptr; CV = nullptr;
13259 // If it's an adding a simple constant then integrate the offset.
13260 if (Base.getOpcode() == ISD::ADD) {
13261 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
13262 Base = Base.getOperand(0);
13263 Offset += C->getZExtValue();
13267 // Return the underlying GlobalValue, and update the Offset. Return false
13268 // for GlobalAddressSDNode since the same GlobalAddress may be represented
13269 // by multiple nodes with different offsets.
13270 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
13271 GV = G->getGlobal();
13272 Offset += G->getOffset();
13276 // Return the underlying Constant value, and update the Offset. Return false
13277 // for ConstantSDNodes since the same constant pool entry may be represented
13278 // by multiple nodes with different offsets.
13279 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
13280 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
13281 : (const void *)C->getConstVal();
13282 Offset += C->getOffset();
13285 // If it's any of the following then it can't alias with anything but itself.
13286 return isa<FrameIndexSDNode>(Base);
13289 /// Return true if there is any possibility that the two addresses overlap.
13290 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const {
13291 // If they are the same then they must be aliases.
13292 if (Op0->getBasePtr() == Op1->getBasePtr()) return true;
13294 // If they are both volatile then they cannot be reordered.
13295 if (Op0->isVolatile() && Op1->isVolatile()) return true;
13297 // Gather base node and offset information.
13298 SDValue Base1, Base2;
13299 int64_t Offset1, Offset2;
13300 const GlobalValue *GV1, *GV2;
13301 const void *CV1, *CV2;
13302 bool isFrameIndex1 = FindBaseOffset(Op0->getBasePtr(),
13303 Base1, Offset1, GV1, CV1);
13304 bool isFrameIndex2 = FindBaseOffset(Op1->getBasePtr(),
13305 Base2, Offset2, GV2, CV2);
13307 // If they have a same base address then check to see if they overlap.
13308 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
13309 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
13310 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
13312 // It is possible for different frame indices to alias each other, mostly
13313 // when tail call optimization reuses return address slots for arguments.
13314 // To catch this case, look up the actual index of frame indices to compute
13315 // the real alias relationship.
13316 if (isFrameIndex1 && isFrameIndex2) {
13317 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
13318 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
13319 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
13320 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
13321 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
13324 // Otherwise, if we know what the bases are, and they aren't identical, then
13325 // we know they cannot alias.
13326 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
13329 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
13330 // compared to the size and offset of the access, we may be able to prove they
13331 // do not alias. This check is conservative for now to catch cases created by
13332 // splitting vector types.
13333 if ((Op0->getOriginalAlignment() == Op1->getOriginalAlignment()) &&
13334 (Op0->getSrcValueOffset() != Op1->getSrcValueOffset()) &&
13335 (Op0->getMemoryVT().getSizeInBits() >> 3 ==
13336 Op1->getMemoryVT().getSizeInBits() >> 3) &&
13337 (Op0->getOriginalAlignment() > Op0->getMemoryVT().getSizeInBits()) >> 3) {
13338 int64_t OffAlign1 = Op0->getSrcValueOffset() % Op0->getOriginalAlignment();
13339 int64_t OffAlign2 = Op1->getSrcValueOffset() % Op1->getOriginalAlignment();
13341 // There is no overlap between these relatively aligned accesses of similar
13342 // size, return no alias.
13343 if ((OffAlign1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign2 ||
13344 (OffAlign2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign1)
13348 bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0
13350 : DAG.getSubtarget().useAA();
13352 if (CombinerAAOnlyFunc.getNumOccurrences() &&
13353 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
13357 Op0->getMemOperand()->getValue() && Op1->getMemOperand()->getValue()) {
13358 // Use alias analysis information.
13359 int64_t MinOffset = std::min(Op0->getSrcValueOffset(),
13360 Op1->getSrcValueOffset());
13361 int64_t Overlap1 = (Op0->getMemoryVT().getSizeInBits() >> 3) +
13362 Op0->getSrcValueOffset() - MinOffset;
13363 int64_t Overlap2 = (Op1->getMemoryVT().getSizeInBits() >> 3) +
13364 Op1->getSrcValueOffset() - MinOffset;
13365 AliasAnalysis::AliasResult AAResult =
13366 AA.alias(AliasAnalysis::Location(Op0->getMemOperand()->getValue(),
13368 UseTBAA ? Op0->getAAInfo() : AAMDNodes()),
13369 AliasAnalysis::Location(Op1->getMemOperand()->getValue(),
13371 UseTBAA ? Op1->getAAInfo() : AAMDNodes()));
13372 if (AAResult == AliasAnalysis::NoAlias)
13376 // Otherwise we have to assume they alias.
13380 /// Walk up chain skipping non-aliasing memory nodes,
13381 /// looking for aliasing nodes and adding them to the Aliases vector.
13382 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
13383 SmallVectorImpl<SDValue> &Aliases) {
13384 SmallVector<SDValue, 8> Chains; // List of chains to visit.
13385 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
13387 // Get alias information for node.
13388 bool IsLoad = isa<LoadSDNode>(N) && !cast<LSBaseSDNode>(N)->isVolatile();
13391 Chains.push_back(OriginalChain);
13392 unsigned Depth = 0;
13394 // Look at each chain and determine if it is an alias. If so, add it to the
13395 // aliases list. If not, then continue up the chain looking for the next
13397 while (!Chains.empty()) {
13398 SDValue Chain = Chains.back();
13401 // For TokenFactor nodes, look at each operand and only continue up the
13402 // chain until we find two aliases. If we've seen two aliases, assume we'll
13403 // find more and revert to original chain since the xform is unlikely to be
13406 // FIXME: The depth check could be made to return the last non-aliasing
13407 // chain we found before we hit a tokenfactor rather than the original
13409 if (Depth > 6 || Aliases.size() == 2) {
13411 Aliases.push_back(OriginalChain);
13415 // Don't bother if we've been before.
13416 if (!Visited.insert(Chain.getNode()).second)
13419 switch (Chain.getOpcode()) {
13420 case ISD::EntryToken:
13421 // Entry token is ideal chain operand, but handled in FindBetterChain.
13426 // Get alias information for Chain.
13427 bool IsOpLoad = isa<LoadSDNode>(Chain.getNode()) &&
13428 !cast<LSBaseSDNode>(Chain.getNode())->isVolatile();
13430 // If chain is alias then stop here.
13431 if (!(IsLoad && IsOpLoad) &&
13432 isAlias(cast<LSBaseSDNode>(N), cast<LSBaseSDNode>(Chain.getNode()))) {
13433 Aliases.push_back(Chain);
13435 // Look further up the chain.
13436 Chains.push_back(Chain.getOperand(0));
13442 case ISD::TokenFactor:
13443 // We have to check each of the operands of the token factor for "small"
13444 // token factors, so we queue them up. Adding the operands to the queue
13445 // (stack) in reverse order maintains the original order and increases the
13446 // likelihood that getNode will find a matching token factor (CSE.)
13447 if (Chain.getNumOperands() > 16) {
13448 Aliases.push_back(Chain);
13451 for (unsigned n = Chain.getNumOperands(); n;)
13452 Chains.push_back(Chain.getOperand(--n));
13457 // For all other instructions we will just have to take what we can get.
13458 Aliases.push_back(Chain);
13463 // We need to be careful here to also search for aliases through the
13464 // value operand of a store, etc. Consider the following situation:
13466 // L1 = load Token1, %52
13467 // S1 = store Token1, L1, %51
13468 // L2 = load Token1, %52+8
13469 // S2 = store Token1, L2, %51+8
13470 // Token2 = Token(S1, S2)
13471 // L3 = load Token2, %53
13472 // S3 = store Token2, L3, %52
13473 // L4 = load Token2, %53+8
13474 // S4 = store Token2, L4, %52+8
13475 // If we search for aliases of S3 (which loads address %52), and we look
13476 // only through the chain, then we'll miss the trivial dependence on L1
13477 // (which also loads from %52). We then might change all loads and
13478 // stores to use Token1 as their chain operand, which could result in
13479 // copying %53 into %52 before copying %52 into %51 (which should
13482 // The problem is, however, that searching for such data dependencies
13483 // can become expensive, and the cost is not directly related to the
13484 // chain depth. Instead, we'll rule out such configurations here by
13485 // insisting that we've visited all chain users (except for users
13486 // of the original chain, which is not necessary). When doing this,
13487 // we need to look through nodes we don't care about (otherwise, things
13488 // like register copies will interfere with trivial cases).
13490 SmallVector<const SDNode *, 16> Worklist;
13491 for (const SDNode *N : Visited)
13492 if (N != OriginalChain.getNode())
13493 Worklist.push_back(N);
13495 while (!Worklist.empty()) {
13496 const SDNode *M = Worklist.pop_back_val();
13498 // We have already visited M, and want to make sure we've visited any uses
13499 // of M that we care about. For uses that we've not visisted, and don't
13500 // care about, queue them to the worklist.
13502 for (SDNode::use_iterator UI = M->use_begin(),
13503 UIE = M->use_end(); UI != UIE; ++UI)
13504 if (UI.getUse().getValueType() == MVT::Other &&
13505 Visited.insert(*UI).second) {
13506 if (isa<MemIntrinsicSDNode>(*UI) || isa<MemSDNode>(*UI)) {
13507 // We've not visited this use, and we care about it (it could have an
13508 // ordering dependency with the original node).
13510 Aliases.push_back(OriginalChain);
13514 // We've not visited this use, but we don't care about it. Mark it as
13515 // visited and enqueue it to the worklist.
13516 Worklist.push_back(*UI);
13521 /// Walk up chain skipping non-aliasing memory nodes, looking for a better chain
13522 /// (aliasing node.)
13523 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
13524 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
13526 // Accumulate all the aliases to this node.
13527 GatherAllAliases(N, OldChain, Aliases);
13529 // If no operands then chain to entry token.
13530 if (Aliases.size() == 0)
13531 return DAG.getEntryNode();
13533 // If a single operand then chain to it. We don't need to revisit it.
13534 if (Aliases.size() == 1)
13537 // Construct a custom tailored token factor.
13538 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Aliases);
13541 /// This is the entry point for the file.
13542 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
13543 CodeGenOpt::Level OptLevel) {
13544 /// This is the main entry point to this class.
13545 DAGCombiner(*this, AA, OptLevel).Run(Level);