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/SmallPtrSet.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/LLVMContext.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Target/TargetLowering.h"
36 #include "llvm/Target/TargetMachine.h"
37 #include "llvm/Target/TargetOptions.h"
38 #include "llvm/Target/TargetRegisterInfo.h"
39 #include "llvm/Target/TargetSubtargetInfo.h"
43 #define DEBUG_TYPE "dagcombine"
45 STATISTIC(NodesCombined , "Number of dag nodes combined");
46 STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created");
47 STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created");
48 STATISTIC(OpsNarrowed , "Number of load/op/store narrowed");
49 STATISTIC(LdStFP2Int , "Number of fp load/store pairs transformed to int");
50 STATISTIC(SlicedLoads, "Number of load sliced");
54 CombinerAA("combiner-alias-analysis", cl::Hidden,
55 cl::desc("Enable DAG combiner alias-analysis heuristics"));
58 CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden,
59 cl::desc("Enable DAG combiner's use of IR alias analysis"));
62 UseTBAA("combiner-use-tbaa", cl::Hidden, cl::init(true),
63 cl::desc("Enable DAG combiner's use of TBAA"));
66 static cl::opt<std::string>
67 CombinerAAOnlyFunc("combiner-aa-only-func", cl::Hidden,
68 cl::desc("Only use DAG-combiner alias analysis in this"
72 /// Hidden option to stress test load slicing, i.e., when this option
73 /// is enabled, load slicing bypasses most of its profitability guards.
75 StressLoadSlicing("combiner-stress-load-slicing", cl::Hidden,
76 cl::desc("Bypass the profitability model of load "
81 MaySplitLoadIndex("combiner-split-load-index", cl::Hidden, cl::init(true),
82 cl::desc("DAG combiner may split indexing from loads"));
84 //------------------------------ DAGCombiner ---------------------------------//
88 const TargetLowering &TLI;
90 CodeGenOpt::Level OptLevel;
95 /// \brief Worklist of all of the nodes that need to be simplified.
97 /// This must behave as a stack -- new nodes to process are pushed onto the
98 /// back and when processing we pop off of the back.
100 /// The worklist will not contain duplicates but may contain null entries
101 /// due to nodes being deleted from the underlying DAG.
102 SmallVector<SDNode *, 64> Worklist;
104 /// \brief Mapping from an SDNode to its position on the worklist.
106 /// This is used to find and remove nodes from the worklist (by nulling
107 /// them) when they are deleted from the underlying DAG. It relies on
108 /// stable indices of nodes within the worklist.
109 DenseMap<SDNode *, unsigned> WorklistMap;
111 /// \brief Set of nodes which have been combined (at least once).
113 /// This is used to allow us to reliably add any operands of a DAG node
114 /// which have not yet been combined to the worklist.
115 SmallPtrSet<SDNode *, 64> CombinedNodes;
117 // AA - Used for DAG load/store alias analysis.
120 /// When an instruction is simplified, add all users of the instruction to
121 /// the work lists because they might get more simplified now.
122 void AddUsersToWorklist(SDNode *N) {
123 for (SDNode *Node : N->uses())
127 /// Call the node-specific routine that folds each particular type of node.
128 SDValue visit(SDNode *N);
131 /// Add to the worklist making sure its instance is at the back (next to be
133 void AddToWorklist(SDNode *N) {
134 // Skip handle nodes as they can't usefully be combined and confuse the
135 // zero-use deletion strategy.
136 if (N->getOpcode() == ISD::HANDLENODE)
139 if (WorklistMap.insert(std::make_pair(N, Worklist.size())).second)
140 Worklist.push_back(N);
143 /// Remove all instances of N from the worklist.
144 void removeFromWorklist(SDNode *N) {
145 CombinedNodes.erase(N);
147 auto It = WorklistMap.find(N);
148 if (It == WorklistMap.end())
149 return; // Not in the worklist.
151 // Null out the entry rather than erasing it to avoid a linear operation.
152 Worklist[It->second] = nullptr;
153 WorklistMap.erase(It);
156 void deleteAndRecombine(SDNode *N);
157 bool recursivelyDeleteUnusedNodes(SDNode *N);
159 SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
162 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) {
163 return CombineTo(N, &Res, 1, AddTo);
166 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1,
168 SDValue To[] = { Res0, Res1 };
169 return CombineTo(N, To, 2, AddTo);
172 void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO);
176 /// Check the specified integer node value to see if it can be simplified or
177 /// if things it uses can be simplified by bit propagation.
178 /// If so, return true.
179 bool SimplifyDemandedBits(SDValue Op) {
180 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
181 APInt Demanded = APInt::getAllOnesValue(BitWidth);
182 return SimplifyDemandedBits(Op, Demanded);
185 bool SimplifyDemandedBits(SDValue Op, const APInt &Demanded);
187 bool CombineToPreIndexedLoadStore(SDNode *N);
188 bool CombineToPostIndexedLoadStore(SDNode *N);
189 SDValue SplitIndexingFromLoad(LoadSDNode *LD);
190 bool SliceUpLoad(SDNode *N);
192 /// \brief Replace an ISD::EXTRACT_VECTOR_ELT of a load with a narrowed
195 /// \param EVE ISD::EXTRACT_VECTOR_ELT to be replaced.
196 /// \param InVecVT type of the input vector to EVE with bitcasts resolved.
197 /// \param EltNo index of the vector element to load.
198 /// \param OriginalLoad load that EVE came from to be replaced.
199 /// \returns EVE on success SDValue() on failure.
200 SDValue ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
201 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad);
202 void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad);
203 SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace);
204 SDValue SExtPromoteOperand(SDValue Op, EVT PVT);
205 SDValue ZExtPromoteOperand(SDValue Op, EVT PVT);
206 SDValue PromoteIntBinOp(SDValue Op);
207 SDValue PromoteIntShiftOp(SDValue Op);
208 SDValue PromoteExtend(SDValue Op);
209 bool PromoteLoad(SDValue Op);
211 void ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
212 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
213 ISD::NodeType ExtType);
215 /// Call the node-specific routine that knows how to fold each
216 /// particular type of node. If that doesn't do anything, try the
217 /// target-specific DAG combines.
218 SDValue combine(SDNode *N);
220 // Visitation implementation - Implement dag node combining for different
221 // node types. The semantics are as follows:
223 // SDValue.getNode() == 0 - No change was made
224 // SDValue.getNode() == N - N was replaced, is dead and has been handled.
225 // otherwise - N should be replaced by the returned Operand.
227 SDValue visitTokenFactor(SDNode *N);
228 SDValue visitMERGE_VALUES(SDNode *N);
229 SDValue visitADD(SDNode *N);
230 SDValue visitSUB(SDNode *N);
231 SDValue visitADDC(SDNode *N);
232 SDValue visitSUBC(SDNode *N);
233 SDValue visitADDE(SDNode *N);
234 SDValue visitSUBE(SDNode *N);
235 SDValue visitMUL(SDNode *N);
236 SDValue visitSDIV(SDNode *N);
237 SDValue visitUDIV(SDNode *N);
238 SDValue visitSREM(SDNode *N);
239 SDValue visitUREM(SDNode *N);
240 SDValue visitMULHU(SDNode *N);
241 SDValue visitMULHS(SDNode *N);
242 SDValue visitSMUL_LOHI(SDNode *N);
243 SDValue visitUMUL_LOHI(SDNode *N);
244 SDValue visitSMULO(SDNode *N);
245 SDValue visitUMULO(SDNode *N);
246 SDValue visitSDIVREM(SDNode *N);
247 SDValue visitUDIVREM(SDNode *N);
248 SDValue visitAND(SDNode *N);
249 SDValue visitOR(SDNode *N);
250 SDValue visitXOR(SDNode *N);
251 SDValue SimplifyVBinOp(SDNode *N);
252 SDValue SimplifyVUnaryOp(SDNode *N);
253 SDValue visitSHL(SDNode *N);
254 SDValue visitSRA(SDNode *N);
255 SDValue visitSRL(SDNode *N);
256 SDValue visitRotate(SDNode *N);
257 SDValue visitCTLZ(SDNode *N);
258 SDValue visitCTLZ_ZERO_UNDEF(SDNode *N);
259 SDValue visitCTTZ(SDNode *N);
260 SDValue visitCTTZ_ZERO_UNDEF(SDNode *N);
261 SDValue visitCTPOP(SDNode *N);
262 SDValue visitSELECT(SDNode *N);
263 SDValue visitVSELECT(SDNode *N);
264 SDValue visitSELECT_CC(SDNode *N);
265 SDValue visitSETCC(SDNode *N);
266 SDValue visitSIGN_EXTEND(SDNode *N);
267 SDValue visitZERO_EXTEND(SDNode *N);
268 SDValue visitANY_EXTEND(SDNode *N);
269 SDValue visitSIGN_EXTEND_INREG(SDNode *N);
270 SDValue visitTRUNCATE(SDNode *N);
271 SDValue visitBITCAST(SDNode *N);
272 SDValue visitBUILD_PAIR(SDNode *N);
273 SDValue visitFADD(SDNode *N);
274 SDValue visitFSUB(SDNode *N);
275 SDValue visitFMUL(SDNode *N);
276 SDValue visitFMA(SDNode *N);
277 SDValue visitFDIV(SDNode *N);
278 SDValue visitFREM(SDNode *N);
279 SDValue visitFCOPYSIGN(SDNode *N);
280 SDValue visitSINT_TO_FP(SDNode *N);
281 SDValue visitUINT_TO_FP(SDNode *N);
282 SDValue visitFP_TO_SINT(SDNode *N);
283 SDValue visitFP_TO_UINT(SDNode *N);
284 SDValue visitFP_ROUND(SDNode *N);
285 SDValue visitFP_ROUND_INREG(SDNode *N);
286 SDValue visitFP_EXTEND(SDNode *N);
287 SDValue visitFNEG(SDNode *N);
288 SDValue visitFABS(SDNode *N);
289 SDValue visitFCEIL(SDNode *N);
290 SDValue visitFTRUNC(SDNode *N);
291 SDValue visitFFLOOR(SDNode *N);
292 SDValue visitBRCOND(SDNode *N);
293 SDValue visitBR_CC(SDNode *N);
294 SDValue visitLOAD(SDNode *N);
295 SDValue visitSTORE(SDNode *N);
296 SDValue visitINSERT_VECTOR_ELT(SDNode *N);
297 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
298 SDValue visitBUILD_VECTOR(SDNode *N);
299 SDValue visitCONCAT_VECTORS(SDNode *N);
300 SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
301 SDValue visitVECTOR_SHUFFLE(SDNode *N);
302 SDValue visitINSERT_SUBVECTOR(SDNode *N);
304 SDValue XformToShuffleWithZero(SDNode *N);
305 SDValue ReassociateOps(unsigned Opc, SDLoc DL, SDValue LHS, SDValue RHS);
307 SDValue visitShiftByConstant(SDNode *N, ConstantSDNode *Amt);
309 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
310 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
311 SDValue SimplifySelect(SDLoc DL, SDValue N0, SDValue N1, SDValue N2);
312 SDValue SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1, SDValue N2,
313 SDValue N3, ISD::CondCode CC,
314 bool NotExtCompare = false);
315 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
316 SDLoc DL, bool foldBooleans = true);
318 bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
320 bool isOneUseSetCC(SDValue N) const;
322 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
324 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
325 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
326 SDValue BuildSDIV(SDNode *N);
327 SDValue BuildSDIVPow2(SDNode *N);
328 SDValue BuildUDIV(SDNode *N);
329 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
330 bool DemandHighBits = true);
331 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
332 SDNode *MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg,
333 SDValue InnerPos, SDValue InnerNeg,
334 unsigned PosOpcode, unsigned NegOpcode,
336 SDNode *MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL);
337 SDValue ReduceLoadWidth(SDNode *N);
338 SDValue ReduceLoadOpStoreWidth(SDNode *N);
339 SDValue TransformFPLoadStorePair(SDNode *N);
340 SDValue reduceBuildVecExtToExtBuildVec(SDNode *N);
341 SDValue reduceBuildVecConvertToConvertBuildVec(SDNode *N);
343 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
345 /// Walk up chain skipping non-aliasing memory nodes,
346 /// looking for aliasing nodes and adding them to the Aliases vector.
347 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
348 SmallVectorImpl<SDValue> &Aliases);
350 /// Return true if there is any possibility that the two addresses overlap.
351 bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const;
353 /// Walk up chain skipping non-aliasing memory nodes, looking for a better
354 /// chain (aliasing node.)
355 SDValue FindBetterChain(SDNode *N, SDValue Chain);
357 /// Merge consecutive store operations into a wide store.
358 /// This optimization uses wide integers or vectors when possible.
359 /// \return True if some memory operations were changed.
360 bool MergeConsecutiveStores(StoreSDNode *N);
362 /// \brief Try to transform a truncation where C is a constant:
363 /// (trunc (and X, C)) -> (and (trunc X), (trunc C))
365 /// \p N needs to be a truncation and its first operand an AND. Other
366 /// requirements are checked by the function (e.g. that trunc is
367 /// single-use) and if missed an empty SDValue is returned.
368 SDValue distributeTruncateThroughAnd(SDNode *N);
371 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
372 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
373 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {
374 AttributeSet FnAttrs =
375 DAG.getMachineFunction().getFunction()->getAttributes();
377 FnAttrs.hasAttribute(AttributeSet::FunctionIndex,
378 Attribute::OptimizeForSize) ||
379 FnAttrs.hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize);
382 /// Runs the dag combiner on all nodes in the work list
383 void Run(CombineLevel AtLevel);
385 SelectionDAG &getDAG() const { return DAG; }
387 /// Returns a type large enough to hold any valid shift amount - before type
388 /// legalization these can be huge.
389 EVT getShiftAmountTy(EVT LHSTy) {
390 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
391 if (LHSTy.isVector())
393 return LegalTypes ? TLI.getScalarShiftAmountTy(LHSTy)
394 : TLI.getPointerTy();
397 /// This method returns true if we are running before type legalization or
398 /// if the specified VT is legal.
399 bool isTypeLegal(const EVT &VT) {
400 if (!LegalTypes) return true;
401 return TLI.isTypeLegal(VT);
404 /// Convenience wrapper around TargetLowering::getSetCCResultType
405 EVT getSetCCResultType(EVT VT) const {
406 return TLI.getSetCCResultType(*DAG.getContext(), VT);
413 /// This class is a DAGUpdateListener that removes any deleted
414 /// nodes from the worklist.
415 class WorklistRemover : public SelectionDAG::DAGUpdateListener {
418 explicit WorklistRemover(DAGCombiner &dc)
419 : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {}
421 void NodeDeleted(SDNode *N, SDNode *E) override {
422 DC.removeFromWorklist(N);
427 //===----------------------------------------------------------------------===//
428 // TargetLowering::DAGCombinerInfo implementation
429 //===----------------------------------------------------------------------===//
431 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
432 ((DAGCombiner*)DC)->AddToWorklist(N);
435 void TargetLowering::DAGCombinerInfo::RemoveFromWorklist(SDNode *N) {
436 ((DAGCombiner*)DC)->removeFromWorklist(N);
439 SDValue TargetLowering::DAGCombinerInfo::
440 CombineTo(SDNode *N, const std::vector<SDValue> &To, bool AddTo) {
441 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
444 SDValue TargetLowering::DAGCombinerInfo::
445 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
446 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
450 SDValue TargetLowering::DAGCombinerInfo::
451 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
452 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
455 void TargetLowering::DAGCombinerInfo::
456 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
457 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
460 //===----------------------------------------------------------------------===//
462 //===----------------------------------------------------------------------===//
464 void DAGCombiner::deleteAndRecombine(SDNode *N) {
465 removeFromWorklist(N);
467 // If the operands of this node are only used by the node, they will now be
468 // dead. Make sure to re-visit them and recursively delete dead nodes.
469 for (const SDValue &Op : N->ops())
470 // For an operand generating multiple values, one of the values may
471 // become dead allowing further simplification (e.g. split index
472 // arithmetic from an indexed load).
473 if (Op->hasOneUse() || Op->getNumValues() > 1)
474 AddToWorklist(Op.getNode());
479 /// Return 1 if we can compute the negated form of the specified expression for
480 /// the same cost as the expression itself, or 2 if we can compute the negated
481 /// form more cheaply than the expression itself.
482 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
483 const TargetLowering &TLI,
484 const TargetOptions *Options,
485 unsigned Depth = 0) {
486 // fneg is removable even if it has multiple uses.
487 if (Op.getOpcode() == ISD::FNEG) return 2;
489 // Don't allow anything with multiple uses.
490 if (!Op.hasOneUse()) return 0;
492 // Don't recurse exponentially.
493 if (Depth > 6) return 0;
495 switch (Op.getOpcode()) {
496 default: return false;
497 case ISD::ConstantFP:
498 // Don't invert constant FP values after legalize. The negated constant
499 // isn't necessarily legal.
500 return LegalOperations ? 0 : 1;
502 // FIXME: determine better conditions for this xform.
503 if (!Options->UnsafeFPMath) return 0;
505 // After operation legalization, it might not be legal to create new FSUBs.
506 if (LegalOperations &&
507 !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType()))
510 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
511 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
514 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
515 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
518 // We can't turn -(A-B) into B-A when we honor signed zeros.
519 if (!Options->UnsafeFPMath) return 0;
521 // fold (fneg (fsub A, B)) -> (fsub B, A)
526 if (Options->HonorSignDependentRoundingFPMath()) return 0;
528 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
529 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
533 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
539 return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options,
544 /// If isNegatibleForFree returns true, return the newly negated expression.
545 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
546 bool LegalOperations, unsigned Depth = 0) {
547 const TargetOptions &Options = DAG.getTarget().Options;
548 // fneg is removable even if it has multiple uses.
549 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
551 // Don't allow anything with multiple uses.
552 assert(Op.hasOneUse() && "Unknown reuse!");
554 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
555 switch (Op.getOpcode()) {
556 default: llvm_unreachable("Unknown code");
557 case ISD::ConstantFP: {
558 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
560 return DAG.getConstantFP(V, Op.getValueType());
563 // FIXME: determine better conditions for this xform.
564 assert(Options.UnsafeFPMath);
566 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
567 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
568 DAG.getTargetLoweringInfo(), &Options, Depth+1))
569 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
570 GetNegatedExpression(Op.getOperand(0), DAG,
571 LegalOperations, Depth+1),
573 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
574 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
575 GetNegatedExpression(Op.getOperand(1), DAG,
576 LegalOperations, Depth+1),
579 // We can't turn -(A-B) into B-A when we honor signed zeros.
580 assert(Options.UnsafeFPMath);
582 // fold (fneg (fsub 0, B)) -> B
583 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
584 if (N0CFP->getValueAPF().isZero())
585 return Op.getOperand(1);
587 // fold (fneg (fsub A, B)) -> (fsub B, A)
588 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
589 Op.getOperand(1), Op.getOperand(0));
593 assert(!Options.HonorSignDependentRoundingFPMath());
595 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
596 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
597 DAG.getTargetLoweringInfo(), &Options, Depth+1))
598 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
599 GetNegatedExpression(Op.getOperand(0), DAG,
600 LegalOperations, Depth+1),
603 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
604 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
606 GetNegatedExpression(Op.getOperand(1), DAG,
607 LegalOperations, Depth+1));
611 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
612 GetNegatedExpression(Op.getOperand(0), DAG,
613 LegalOperations, Depth+1));
615 return DAG.getNode(ISD::FP_ROUND, SDLoc(Op), Op.getValueType(),
616 GetNegatedExpression(Op.getOperand(0), DAG,
617 LegalOperations, Depth+1),
622 // Return true if this node is a setcc, or is a select_cc
623 // that selects between the target values used for true and false, making it
624 // equivalent to a setcc. Also, set the incoming LHS, RHS, and CC references to
625 // the appropriate nodes based on the type of node we are checking. This
626 // simplifies life a bit for the callers.
627 bool DAGCombiner::isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
629 if (N.getOpcode() == ISD::SETCC) {
630 LHS = N.getOperand(0);
631 RHS = N.getOperand(1);
632 CC = N.getOperand(2);
636 if (N.getOpcode() != ISD::SELECT_CC ||
637 !TLI.isConstTrueVal(N.getOperand(2).getNode()) ||
638 !TLI.isConstFalseVal(N.getOperand(3).getNode()))
641 LHS = N.getOperand(0);
642 RHS = N.getOperand(1);
643 CC = N.getOperand(4);
647 /// Return true if this is a SetCC-equivalent operation with only one use.
648 /// If this is true, it allows the users to invert the operation for free when
649 /// it is profitable to do so.
650 bool DAGCombiner::isOneUseSetCC(SDValue N) const {
652 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
657 /// Returns true if N is a BUILD_VECTOR node whose
658 /// elements are all the same constant or undefined.
659 static bool isConstantSplatVector(SDNode *N, APInt& SplatValue) {
660 BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(N);
665 unsigned SplatBitSize;
667 EVT EltVT = N->getValueType(0).getVectorElementType();
668 return (C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
670 EltVT.getSizeInBits() >= SplatBitSize);
673 // \brief Returns the SDNode if it is a constant BuildVector or constant.
674 static SDNode *isConstantBuildVectorOrConstantInt(SDValue N) {
675 if (isa<ConstantSDNode>(N))
677 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N);
678 if (BV && BV->isConstant())
683 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
685 static ConstantSDNode *isConstOrConstSplat(SDValue N) {
686 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N))
689 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
690 BitVector UndefElements;
691 ConstantSDNode *CN = BV->getConstantSplatNode(&UndefElements);
693 // BuildVectors can truncate their operands. Ignore that case here.
694 // FIXME: We blindly ignore splats which include undef which is overly
696 if (CN && UndefElements.none() &&
697 CN->getValueType(0) == N.getValueType().getScalarType())
704 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
706 static ConstantFPSDNode *isConstOrConstSplatFP(SDValue N) {
707 if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N))
710 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
711 BitVector UndefElements;
712 ConstantFPSDNode *CN = BV->getConstantFPSplatNode(&UndefElements);
714 if (CN && UndefElements.none())
721 SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL,
722 SDValue N0, SDValue N1) {
723 EVT VT = N0.getValueType();
724 if (N0.getOpcode() == Opc) {
725 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0.getOperand(1))) {
726 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1)) {
727 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
728 SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, L, R);
729 if (!OpNode.getNode())
731 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
733 if (N0.hasOneUse()) {
734 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one
736 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1);
737 if (!OpNode.getNode())
739 AddToWorklist(OpNode.getNode());
740 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
745 if (N1.getOpcode() == Opc) {
746 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1.getOperand(1))) {
747 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0)) {
748 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
749 SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, R, L);
750 if (!OpNode.getNode())
752 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
754 if (N1.hasOneUse()) {
755 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one
757 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N1.getOperand(0), N0);
758 if (!OpNode.getNode())
760 AddToWorklist(OpNode.getNode());
761 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
769 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
771 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
773 DEBUG(dbgs() << "\nReplacing.1 ";
775 dbgs() << "\nWith: ";
776 To[0].getNode()->dump(&DAG);
777 dbgs() << " and " << NumTo-1 << " other values\n";
778 for (unsigned i = 0, e = NumTo; i != e; ++i)
779 assert((!To[i].getNode() ||
780 N->getValueType(i) == To[i].getValueType()) &&
781 "Cannot combine value to value of different type!"));
782 WorklistRemover DeadNodes(*this);
783 DAG.ReplaceAllUsesWith(N, To);
785 // Push the new nodes and any users onto the worklist
786 for (unsigned i = 0, e = NumTo; i != e; ++i) {
787 if (To[i].getNode()) {
788 AddToWorklist(To[i].getNode());
789 AddUsersToWorklist(To[i].getNode());
794 // Finally, if the node is now dead, remove it from the graph. The node
795 // may not be dead if the replacement process recursively simplified to
796 // something else needing this node.
798 deleteAndRecombine(N);
799 return SDValue(N, 0);
803 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
804 // Replace all uses. If any nodes become isomorphic to other nodes and
805 // are deleted, make sure to remove them from our worklist.
806 WorklistRemover DeadNodes(*this);
807 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New);
809 // Push the new node and any (possibly new) users onto the worklist.
810 AddToWorklist(TLO.New.getNode());
811 AddUsersToWorklist(TLO.New.getNode());
813 // Finally, if the node is now dead, remove it from the graph. The node
814 // may not be dead if the replacement process recursively simplified to
815 // something else needing this node.
816 if (TLO.Old.getNode()->use_empty())
817 deleteAndRecombine(TLO.Old.getNode());
820 /// Check the specified integer node value to see if it can be simplified or if
821 /// things it uses can be simplified by bit propagation. If so, return true.
822 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
823 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
824 APInt KnownZero, KnownOne;
825 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
829 AddToWorklist(Op.getNode());
831 // Replace the old value with the new one.
833 DEBUG(dbgs() << "\nReplacing.2 ";
834 TLO.Old.getNode()->dump(&DAG);
835 dbgs() << "\nWith: ";
836 TLO.New.getNode()->dump(&DAG);
839 CommitTargetLoweringOpt(TLO);
843 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
845 EVT VT = Load->getValueType(0);
846 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
848 DEBUG(dbgs() << "\nReplacing.9 ";
850 dbgs() << "\nWith: ";
851 Trunc.getNode()->dump(&DAG);
853 WorklistRemover DeadNodes(*this);
854 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc);
855 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1));
856 deleteAndRecombine(Load);
857 AddToWorklist(Trunc.getNode());
860 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
863 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
864 EVT MemVT = LD->getMemoryVT();
865 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
866 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
868 : LD->getExtensionType();
870 return DAG.getExtLoad(ExtType, dl, PVT,
871 LD->getChain(), LD->getBasePtr(),
872 MemVT, LD->getMemOperand());
875 unsigned Opc = Op.getOpcode();
878 case ISD::AssertSext:
879 return DAG.getNode(ISD::AssertSext, dl, PVT,
880 SExtPromoteOperand(Op.getOperand(0), PVT),
882 case ISD::AssertZext:
883 return DAG.getNode(ISD::AssertZext, dl, PVT,
884 ZExtPromoteOperand(Op.getOperand(0), PVT),
886 case ISD::Constant: {
888 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
889 return DAG.getNode(ExtOpc, dl, PVT, Op);
893 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
895 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
898 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
899 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
901 EVT OldVT = Op.getValueType();
903 bool Replace = false;
904 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
905 if (!NewOp.getNode())
907 AddToWorklist(NewOp.getNode());
910 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
911 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
912 DAG.getValueType(OldVT));
915 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
916 EVT OldVT = Op.getValueType();
918 bool Replace = false;
919 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
920 if (!NewOp.getNode())
922 AddToWorklist(NewOp.getNode());
925 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
926 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
929 /// Promote the specified integer binary operation if the target indicates it is
930 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
931 /// i32 since i16 instructions are longer.
932 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
933 if (!LegalOperations)
936 EVT VT = Op.getValueType();
937 if (VT.isVector() || !VT.isInteger())
940 // If operation type is 'undesirable', e.g. i16 on x86, consider
942 unsigned Opc = Op.getOpcode();
943 if (TLI.isTypeDesirableForOp(Opc, VT))
947 // Consult target whether it is a good idea to promote this operation and
948 // what's the right type to promote it to.
949 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
950 assert(PVT != VT && "Don't know what type to promote to!");
952 bool Replace0 = false;
953 SDValue N0 = Op.getOperand(0);
954 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
958 bool Replace1 = false;
959 SDValue N1 = Op.getOperand(1);
964 NN1 = PromoteOperand(N1, PVT, Replace1);
969 AddToWorklist(NN0.getNode());
971 AddToWorklist(NN1.getNode());
974 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
976 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
978 DEBUG(dbgs() << "\nPromoting ";
979 Op.getNode()->dump(&DAG));
981 return DAG.getNode(ISD::TRUNCATE, dl, VT,
982 DAG.getNode(Opc, dl, PVT, NN0, NN1));
987 /// Promote the specified integer shift operation if the target indicates it is
988 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
989 /// i32 since i16 instructions are longer.
990 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
991 if (!LegalOperations)
994 EVT VT = Op.getValueType();
995 if (VT.isVector() || !VT.isInteger())
998 // If operation type is 'undesirable', e.g. i16 on x86, consider
1000 unsigned Opc = Op.getOpcode();
1001 if (TLI.isTypeDesirableForOp(Opc, VT))
1005 // Consult target whether it is a good idea to promote this operation and
1006 // what's the right type to promote it to.
1007 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1008 assert(PVT != VT && "Don't know what type to promote to!");
1010 bool Replace = false;
1011 SDValue N0 = Op.getOperand(0);
1012 if (Opc == ISD::SRA)
1013 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
1014 else if (Opc == ISD::SRL)
1015 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
1017 N0 = PromoteOperand(N0, PVT, Replace);
1021 AddToWorklist(N0.getNode());
1023 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
1025 DEBUG(dbgs() << "\nPromoting ";
1026 Op.getNode()->dump(&DAG));
1028 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1029 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
1034 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
1035 if (!LegalOperations)
1038 EVT VT = Op.getValueType();
1039 if (VT.isVector() || !VT.isInteger())
1042 // If operation type is 'undesirable', e.g. i16 on x86, consider
1044 unsigned Opc = Op.getOpcode();
1045 if (TLI.isTypeDesirableForOp(Opc, VT))
1049 // Consult target whether it is a good idea to promote this operation and
1050 // what's the right type to promote it to.
1051 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1052 assert(PVT != VT && "Don't know what type to promote to!");
1053 // fold (aext (aext x)) -> (aext x)
1054 // fold (aext (zext x)) -> (zext x)
1055 // fold (aext (sext x)) -> (sext x)
1056 DEBUG(dbgs() << "\nPromoting ";
1057 Op.getNode()->dump(&DAG));
1058 return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0));
1063 bool DAGCombiner::PromoteLoad(SDValue Op) {
1064 if (!LegalOperations)
1067 EVT VT = Op.getValueType();
1068 if (VT.isVector() || !VT.isInteger())
1071 // If operation type is 'undesirable', e.g. i16 on x86, consider
1073 unsigned Opc = Op.getOpcode();
1074 if (TLI.isTypeDesirableForOp(Opc, VT))
1078 // Consult target whether it is a good idea to promote this operation and
1079 // what's the right type to promote it to.
1080 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1081 assert(PVT != VT && "Don't know what type to promote to!");
1084 SDNode *N = Op.getNode();
1085 LoadSDNode *LD = cast<LoadSDNode>(N);
1086 EVT MemVT = LD->getMemoryVT();
1087 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
1088 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
1090 : LD->getExtensionType();
1091 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
1092 LD->getChain(), LD->getBasePtr(),
1093 MemVT, LD->getMemOperand());
1094 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
1096 DEBUG(dbgs() << "\nPromoting ";
1099 Result.getNode()->dump(&DAG);
1101 WorklistRemover DeadNodes(*this);
1102 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
1103 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1));
1104 deleteAndRecombine(N);
1105 AddToWorklist(Result.getNode());
1111 /// \brief Recursively delete a node which has no uses and any operands for
1112 /// which it is the only use.
1114 /// Note that this both deletes the nodes and removes them from the worklist.
1115 /// It also adds any nodes who have had a user deleted to the worklist as they
1116 /// may now have only one use and subject to other combines.
1117 bool DAGCombiner::recursivelyDeleteUnusedNodes(SDNode *N) {
1118 if (!N->use_empty())
1121 SmallSetVector<SDNode *, 16> Nodes;
1124 N = Nodes.pop_back_val();
1128 if (N->use_empty()) {
1129 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1130 Nodes.insert(N->getOperand(i).getNode());
1132 removeFromWorklist(N);
1137 } while (!Nodes.empty());
1141 //===----------------------------------------------------------------------===//
1142 // Main DAG Combiner implementation
1143 //===----------------------------------------------------------------------===//
1145 void DAGCombiner::Run(CombineLevel AtLevel) {
1146 // set the instance variables, so that the various visit routines may use it.
1148 LegalOperations = Level >= AfterLegalizeVectorOps;
1149 LegalTypes = Level >= AfterLegalizeTypes;
1151 // Add all the dag nodes to the worklist.
1152 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
1153 E = DAG.allnodes_end(); I != E; ++I)
1156 // Create a dummy node (which is not added to allnodes), that adds a reference
1157 // to the root node, preventing it from being deleted, and tracking any
1158 // changes of the root.
1159 HandleSDNode Dummy(DAG.getRoot());
1161 // while the worklist isn't empty, find a node and
1162 // try and combine it.
1163 while (!WorklistMap.empty()) {
1165 // The Worklist holds the SDNodes in order, but it may contain null entries.
1167 N = Worklist.pop_back_val();
1170 bool GoodWorklistEntry = WorklistMap.erase(N);
1171 (void)GoodWorklistEntry;
1172 assert(GoodWorklistEntry &&
1173 "Found a worklist entry without a corresponding map entry!");
1175 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1176 // N is deleted from the DAG, since they too may now be dead or may have a
1177 // reduced number of uses, allowing other xforms.
1178 if (recursivelyDeleteUnusedNodes(N))
1181 WorklistRemover DeadNodes(*this);
1183 // If this combine is running after legalizing the DAG, re-legalize any
1184 // nodes pulled off the worklist.
1185 if (Level == AfterLegalizeDAG) {
1186 SmallSetVector<SDNode *, 16> UpdatedNodes;
1187 bool NIsValid = DAG.LegalizeOp(N, UpdatedNodes);
1189 for (SDNode *LN : UpdatedNodes) {
1191 AddUsersToWorklist(LN);
1197 DEBUG(dbgs() << "\nCombining: "; N->dump(&DAG));
1199 // Add any operands of the new node which have not yet been combined to the
1200 // worklist as well. Because the worklist uniques things already, this
1201 // won't repeatedly process the same operand.
1202 CombinedNodes.insert(N);
1203 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1204 if (!CombinedNodes.count(N->getOperand(i).getNode()))
1205 AddToWorklist(N->getOperand(i).getNode());
1207 SDValue RV = combine(N);
1214 // If we get back the same node we passed in, rather than a new node or
1215 // zero, we know that the node must have defined multiple values and
1216 // CombineTo was used. Since CombineTo takes care of the worklist
1217 // mechanics for us, we have no work to do in this case.
1218 if (RV.getNode() == N)
1221 assert(N->getOpcode() != ISD::DELETED_NODE &&
1222 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1223 "Node was deleted but visit returned new node!");
1225 DEBUG(dbgs() << " ... into: ";
1226 RV.getNode()->dump(&DAG));
1228 // Transfer debug value.
1229 DAG.TransferDbgValues(SDValue(N, 0), RV);
1230 if (N->getNumValues() == RV.getNode()->getNumValues())
1231 DAG.ReplaceAllUsesWith(N, RV.getNode());
1233 assert(N->getValueType(0) == RV.getValueType() &&
1234 N->getNumValues() == 1 && "Type mismatch");
1236 DAG.ReplaceAllUsesWith(N, &OpV);
1239 // Push the new node and any users onto the worklist
1240 AddToWorklist(RV.getNode());
1241 AddUsersToWorklist(RV.getNode());
1243 // Finally, if the node is now dead, remove it from the graph. The node
1244 // may not be dead if the replacement process recursively simplified to
1245 // something else needing this node. This will also take care of adding any
1246 // operands which have lost a user to the worklist.
1247 recursivelyDeleteUnusedNodes(N);
1250 // If the root changed (e.g. it was a dead load, update the root).
1251 DAG.setRoot(Dummy.getValue());
1252 DAG.RemoveDeadNodes();
1255 SDValue DAGCombiner::visit(SDNode *N) {
1256 switch (N->getOpcode()) {
1258 case ISD::TokenFactor: return visitTokenFactor(N);
1259 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1260 case ISD::ADD: return visitADD(N);
1261 case ISD::SUB: return visitSUB(N);
1262 case ISD::ADDC: return visitADDC(N);
1263 case ISD::SUBC: return visitSUBC(N);
1264 case ISD::ADDE: return visitADDE(N);
1265 case ISD::SUBE: return visitSUBE(N);
1266 case ISD::MUL: return visitMUL(N);
1267 case ISD::SDIV: return visitSDIV(N);
1268 case ISD::UDIV: return visitUDIV(N);
1269 case ISD::SREM: return visitSREM(N);
1270 case ISD::UREM: return visitUREM(N);
1271 case ISD::MULHU: return visitMULHU(N);
1272 case ISD::MULHS: return visitMULHS(N);
1273 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1274 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1275 case ISD::SMULO: return visitSMULO(N);
1276 case ISD::UMULO: return visitUMULO(N);
1277 case ISD::SDIVREM: return visitSDIVREM(N);
1278 case ISD::UDIVREM: return visitUDIVREM(N);
1279 case ISD::AND: return visitAND(N);
1280 case ISD::OR: return visitOR(N);
1281 case ISD::XOR: return visitXOR(N);
1282 case ISD::SHL: return visitSHL(N);
1283 case ISD::SRA: return visitSRA(N);
1284 case ISD::SRL: return visitSRL(N);
1286 case ISD::ROTL: return visitRotate(N);
1287 case ISD::CTLZ: return visitCTLZ(N);
1288 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1289 case ISD::CTTZ: return visitCTTZ(N);
1290 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1291 case ISD::CTPOP: return visitCTPOP(N);
1292 case ISD::SELECT: return visitSELECT(N);
1293 case ISD::VSELECT: return visitVSELECT(N);
1294 case ISD::SELECT_CC: return visitSELECT_CC(N);
1295 case ISD::SETCC: return visitSETCC(N);
1296 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1297 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1298 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1299 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1300 case ISD::TRUNCATE: return visitTRUNCATE(N);
1301 case ISD::BITCAST: return visitBITCAST(N);
1302 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1303 case ISD::FADD: return visitFADD(N);
1304 case ISD::FSUB: return visitFSUB(N);
1305 case ISD::FMUL: return visitFMUL(N);
1306 case ISD::FMA: return visitFMA(N);
1307 case ISD::FDIV: return visitFDIV(N);
1308 case ISD::FREM: return visitFREM(N);
1309 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1310 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1311 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1312 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1313 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1314 case ISD::FP_ROUND: return visitFP_ROUND(N);
1315 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1316 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1317 case ISD::FNEG: return visitFNEG(N);
1318 case ISD::FABS: return visitFABS(N);
1319 case ISD::FFLOOR: return visitFFLOOR(N);
1320 case ISD::FCEIL: return visitFCEIL(N);
1321 case ISD::FTRUNC: return visitFTRUNC(N);
1322 case ISD::BRCOND: return visitBRCOND(N);
1323 case ISD::BR_CC: return visitBR_CC(N);
1324 case ISD::LOAD: return visitLOAD(N);
1325 case ISD::STORE: return visitSTORE(N);
1326 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1327 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1328 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1329 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1330 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1331 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1332 case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N);
1337 SDValue DAGCombiner::combine(SDNode *N) {
1338 SDValue RV = visit(N);
1340 // If nothing happened, try a target-specific DAG combine.
1341 if (!RV.getNode()) {
1342 assert(N->getOpcode() != ISD::DELETED_NODE &&
1343 "Node was deleted but visit returned NULL!");
1345 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1346 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1348 // Expose the DAG combiner to the target combiner impls.
1349 TargetLowering::DAGCombinerInfo
1350 DagCombineInfo(DAG, Level, false, this);
1352 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1356 // If nothing happened still, try promoting the operation.
1357 if (!RV.getNode()) {
1358 switch (N->getOpcode()) {
1366 RV = PromoteIntBinOp(SDValue(N, 0));
1371 RV = PromoteIntShiftOp(SDValue(N, 0));
1373 case ISD::SIGN_EXTEND:
1374 case ISD::ZERO_EXTEND:
1375 case ISD::ANY_EXTEND:
1376 RV = PromoteExtend(SDValue(N, 0));
1379 if (PromoteLoad(SDValue(N, 0)))
1385 // If N is a commutative binary node, try commuting it to enable more
1387 if (!RV.getNode() && SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1388 N->getNumValues() == 1) {
1389 SDValue N0 = N->getOperand(0);
1390 SDValue N1 = N->getOperand(1);
1392 // Constant operands are canonicalized to RHS.
1393 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1394 SDValue Ops[] = {N1, N0};
1396 if (const BinaryWithFlagsSDNode *BinNode =
1397 dyn_cast<BinaryWithFlagsSDNode>(N)) {
1398 CSENode = DAG.getNodeIfExists(
1399 N->getOpcode(), N->getVTList(), Ops, BinNode->hasNoUnsignedWrap(),
1400 BinNode->hasNoSignedWrap(), BinNode->isExact());
1402 CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops);
1405 return SDValue(CSENode, 0);
1412 /// Given a node, return its input chain if it has one, otherwise return a null
1414 static SDValue getInputChainForNode(SDNode *N) {
1415 if (unsigned NumOps = N->getNumOperands()) {
1416 if (N->getOperand(0).getValueType() == MVT::Other)
1417 return N->getOperand(0);
1418 if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1419 return N->getOperand(NumOps-1);
1420 for (unsigned i = 1; i < NumOps-1; ++i)
1421 if (N->getOperand(i).getValueType() == MVT::Other)
1422 return N->getOperand(i);
1427 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1428 // If N has two operands, where one has an input chain equal to the other,
1429 // the 'other' chain is redundant.
1430 if (N->getNumOperands() == 2) {
1431 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1432 return N->getOperand(0);
1433 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1434 return N->getOperand(1);
1437 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1438 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1439 SmallPtrSet<SDNode*, 16> SeenOps;
1440 bool Changed = false; // If we should replace this token factor.
1442 // Start out with this token factor.
1445 // Iterate through token factors. The TFs grows when new token factors are
1447 for (unsigned i = 0; i < TFs.size(); ++i) {
1448 SDNode *TF = TFs[i];
1450 // Check each of the operands.
1451 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1452 SDValue Op = TF->getOperand(i);
1454 switch (Op.getOpcode()) {
1455 case ISD::EntryToken:
1456 // Entry tokens don't need to be added to the list. They are
1461 case ISD::TokenFactor:
1462 if (Op.hasOneUse() &&
1463 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1464 // Queue up for processing.
1465 TFs.push_back(Op.getNode());
1466 // Clean up in case the token factor is removed.
1467 AddToWorklist(Op.getNode());
1474 // Only add if it isn't already in the list.
1475 if (SeenOps.insert(Op.getNode()))
1486 // If we've change things around then replace token factor.
1489 // The entry token is the only possible outcome.
1490 Result = DAG.getEntryNode();
1492 // New and improved token factor.
1493 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Ops);
1496 // Don't add users to work list.
1497 return CombineTo(N, Result, false);
1503 /// MERGE_VALUES can always be eliminated.
1504 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1505 WorklistRemover DeadNodes(*this);
1506 // Replacing results may cause a different MERGE_VALUES to suddenly
1507 // be CSE'd with N, and carry its uses with it. Iterate until no
1508 // uses remain, to ensure that the node can be safely deleted.
1509 // First add the users of this node to the work list so that they
1510 // can be tried again once they have new operands.
1511 AddUsersToWorklist(N);
1513 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1514 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1515 } while (!N->use_empty());
1516 deleteAndRecombine(N);
1517 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1520 SDValue DAGCombiner::visitADD(SDNode *N) {
1521 SDValue N0 = N->getOperand(0);
1522 SDValue N1 = N->getOperand(1);
1523 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1524 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1525 EVT VT = N0.getValueType();
1528 if (VT.isVector()) {
1529 SDValue FoldedVOp = SimplifyVBinOp(N);
1530 if (FoldedVOp.getNode()) return FoldedVOp;
1532 // fold (add x, 0) -> x, vector edition
1533 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1535 if (ISD::isBuildVectorAllZeros(N0.getNode()))
1539 // fold (add x, undef) -> undef
1540 if (N0.getOpcode() == ISD::UNDEF)
1542 if (N1.getOpcode() == ISD::UNDEF)
1544 // fold (add c1, c2) -> c1+c2
1546 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1547 // canonicalize constant to RHS
1549 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, N0);
1550 // fold (add x, 0) -> x
1551 if (N1C && N1C->isNullValue())
1553 // fold (add Sym, c) -> Sym+c
1554 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1555 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1556 GA->getOpcode() == ISD::GlobalAddress)
1557 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1559 (uint64_t)N1C->getSExtValue());
1560 // fold ((c1-A)+c2) -> (c1+c2)-A
1561 if (N1C && N0.getOpcode() == ISD::SUB)
1562 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1563 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1564 DAG.getConstant(N1C->getAPIntValue()+
1565 N0C->getAPIntValue(), VT),
1568 SDValue RADD = ReassociateOps(ISD::ADD, SDLoc(N), N0, N1);
1571 // fold ((0-A) + B) -> B-A
1572 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1573 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1574 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, N0.getOperand(1));
1575 // fold (A + (0-B)) -> A-B
1576 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1577 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1578 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1.getOperand(1));
1579 // fold (A+(B-A)) -> B
1580 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1581 return N1.getOperand(0);
1582 // fold ((B-A)+A) -> B
1583 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1584 return N0.getOperand(0);
1585 // fold (A+(B-(A+C))) to (B-C)
1586 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1587 N0 == N1.getOperand(1).getOperand(0))
1588 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1589 N1.getOperand(1).getOperand(1));
1590 // fold (A+(B-(C+A))) to (B-C)
1591 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1592 N0 == N1.getOperand(1).getOperand(1))
1593 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1594 N1.getOperand(1).getOperand(0));
1595 // fold (A+((B-A)+or-C)) to (B+or-C)
1596 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1597 N1.getOperand(0).getOpcode() == ISD::SUB &&
1598 N0 == N1.getOperand(0).getOperand(1))
1599 return DAG.getNode(N1.getOpcode(), SDLoc(N), VT,
1600 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1602 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1603 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1604 SDValue N00 = N0.getOperand(0);
1605 SDValue N01 = N0.getOperand(1);
1606 SDValue N10 = N1.getOperand(0);
1607 SDValue N11 = N1.getOperand(1);
1609 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1610 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1611 DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10),
1612 DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11));
1615 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1616 return SDValue(N, 0);
1618 // fold (a+b) -> (a|b) iff a and b share no bits.
1619 if (VT.isInteger() && !VT.isVector()) {
1620 APInt LHSZero, LHSOne;
1621 APInt RHSZero, RHSOne;
1622 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1624 if (LHSZero.getBoolValue()) {
1625 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1627 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1628 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1629 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero){
1630 if (!LegalOperations || TLI.isOperationLegal(ISD::OR, VT))
1631 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1);
1636 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1637 if (N1.getOpcode() == ISD::SHL &&
1638 N1.getOperand(0).getOpcode() == ISD::SUB)
1639 if (ConstantSDNode *C =
1640 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1641 if (C->getAPIntValue() == 0)
1642 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0,
1643 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1644 N1.getOperand(0).getOperand(1),
1646 if (N0.getOpcode() == ISD::SHL &&
1647 N0.getOperand(0).getOpcode() == ISD::SUB)
1648 if (ConstantSDNode *C =
1649 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1650 if (C->getAPIntValue() == 0)
1651 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1,
1652 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1653 N0.getOperand(0).getOperand(1),
1656 if (N1.getOpcode() == ISD::AND) {
1657 SDValue AndOp0 = N1.getOperand(0);
1658 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1659 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1660 unsigned DestBits = VT.getScalarType().getSizeInBits();
1662 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1663 // and similar xforms where the inner op is either ~0 or 0.
1664 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1666 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1670 // add (sext i1), X -> sub X, (zext i1)
1671 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1672 N0.getOperand(0).getValueType() == MVT::i1 &&
1673 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1675 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1676 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1682 SDValue DAGCombiner::visitADDC(SDNode *N) {
1683 SDValue N0 = N->getOperand(0);
1684 SDValue N1 = N->getOperand(1);
1685 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1686 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1687 EVT VT = N0.getValueType();
1689 // If the flag result is dead, turn this into an ADD.
1690 if (!N->hasAnyUseOfValue(1))
1691 return CombineTo(N, DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N1),
1692 DAG.getNode(ISD::CARRY_FALSE,
1693 SDLoc(N), MVT::Glue));
1695 // canonicalize constant to RHS.
1697 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N1, N0);
1699 // fold (addc x, 0) -> x + no carry out
1700 if (N1C && N1C->isNullValue())
1701 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1702 SDLoc(N), MVT::Glue));
1704 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1705 APInt LHSZero, LHSOne;
1706 APInt RHSZero, RHSOne;
1707 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1709 if (LHSZero.getBoolValue()) {
1710 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1712 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1713 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1714 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1715 return CombineTo(N, DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1),
1716 DAG.getNode(ISD::CARRY_FALSE,
1717 SDLoc(N), MVT::Glue));
1723 SDValue DAGCombiner::visitADDE(SDNode *N) {
1724 SDValue N0 = N->getOperand(0);
1725 SDValue N1 = N->getOperand(1);
1726 SDValue CarryIn = N->getOperand(2);
1727 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1728 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1730 // canonicalize constant to RHS
1732 return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(),
1735 // fold (adde x, y, false) -> (addc x, y)
1736 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1737 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1);
1742 // Since it may not be valid to emit a fold to zero for vector initializers
1743 // check if we can before folding.
1744 static SDValue tryFoldToZero(SDLoc DL, const TargetLowering &TLI, EVT VT,
1746 bool LegalOperations, bool LegalTypes) {
1748 return DAG.getConstant(0, VT);
1749 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
1750 return DAG.getConstant(0, VT);
1754 SDValue DAGCombiner::visitSUB(SDNode *N) {
1755 SDValue N0 = N->getOperand(0);
1756 SDValue N1 = N->getOperand(1);
1757 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1758 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1759 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? nullptr :
1760 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1761 EVT VT = N0.getValueType();
1764 if (VT.isVector()) {
1765 SDValue FoldedVOp = SimplifyVBinOp(N);
1766 if (FoldedVOp.getNode()) return FoldedVOp;
1768 // fold (sub x, 0) -> x, vector edition
1769 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1773 // fold (sub x, x) -> 0
1774 // FIXME: Refactor this and xor and other similar operations together.
1776 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
1777 // fold (sub c1, c2) -> c1-c2
1779 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1780 // fold (sub x, c) -> (add x, -c)
1782 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0,
1783 DAG.getConstant(-N1C->getAPIntValue(), VT));
1784 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1785 if (N0C && N0C->isAllOnesValue())
1786 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
1787 // fold A-(A-B) -> B
1788 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1789 return N1.getOperand(1);
1790 // fold (A+B)-A -> B
1791 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1792 return N0.getOperand(1);
1793 // fold (A+B)-B -> A
1794 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1795 return N0.getOperand(0);
1796 // fold C2-(A+C1) -> (C2-C1)-A
1797 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1798 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(),
1800 return DAG.getNode(ISD::SUB, SDLoc(N), VT, NewC,
1803 // fold ((A+(B+or-C))-B) -> A+or-C
1804 if (N0.getOpcode() == ISD::ADD &&
1805 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1806 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1807 N0.getOperand(1).getOperand(0) == N1)
1808 return DAG.getNode(N0.getOperand(1).getOpcode(), SDLoc(N), VT,
1809 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1810 // fold ((A+(C+B))-B) -> A+C
1811 if (N0.getOpcode() == ISD::ADD &&
1812 N0.getOperand(1).getOpcode() == ISD::ADD &&
1813 N0.getOperand(1).getOperand(1) == N1)
1814 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1815 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1816 // fold ((A-(B-C))-C) -> A-B
1817 if (N0.getOpcode() == ISD::SUB &&
1818 N0.getOperand(1).getOpcode() == ISD::SUB &&
1819 N0.getOperand(1).getOperand(1) == N1)
1820 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1821 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1823 // If either operand of a sub is undef, the result is undef
1824 if (N0.getOpcode() == ISD::UNDEF)
1826 if (N1.getOpcode() == ISD::UNDEF)
1829 // If the relocation model supports it, consider symbol offsets.
1830 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1831 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1832 // fold (sub Sym, c) -> Sym-c
1833 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1834 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1836 (uint64_t)N1C->getSExtValue());
1837 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1838 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1839 if (GA->getGlobal() == GB->getGlobal())
1840 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1847 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1848 SDValue N0 = N->getOperand(0);
1849 SDValue N1 = N->getOperand(1);
1850 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1851 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1852 EVT VT = N0.getValueType();
1854 // If the flag result is dead, turn this into an SUB.
1855 if (!N->hasAnyUseOfValue(1))
1856 return CombineTo(N, DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1),
1857 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1860 // fold (subc x, x) -> 0 + no borrow
1862 return CombineTo(N, DAG.getConstant(0, VT),
1863 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1866 // fold (subc x, 0) -> x + no borrow
1867 if (N1C && N1C->isNullValue())
1868 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1871 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1872 if (N0C && N0C->isAllOnesValue())
1873 return CombineTo(N, DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0),
1874 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1880 SDValue DAGCombiner::visitSUBE(SDNode *N) {
1881 SDValue N0 = N->getOperand(0);
1882 SDValue N1 = N->getOperand(1);
1883 SDValue CarryIn = N->getOperand(2);
1885 // fold (sube x, y, false) -> (subc x, y)
1886 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1887 return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1);
1892 SDValue DAGCombiner::visitMUL(SDNode *N) {
1893 SDValue N0 = N->getOperand(0);
1894 SDValue N1 = N->getOperand(1);
1895 EVT VT = N0.getValueType();
1897 // fold (mul x, undef) -> 0
1898 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1899 return DAG.getConstant(0, VT);
1901 bool N0IsConst = false;
1902 bool N1IsConst = false;
1903 APInt ConstValue0, ConstValue1;
1905 if (VT.isVector()) {
1906 SDValue FoldedVOp = SimplifyVBinOp(N);
1907 if (FoldedVOp.getNode()) return FoldedVOp;
1909 N0IsConst = isConstantSplatVector(N0.getNode(), ConstValue0);
1910 N1IsConst = isConstantSplatVector(N1.getNode(), ConstValue1);
1912 N0IsConst = dyn_cast<ConstantSDNode>(N0) != nullptr;
1913 ConstValue0 = N0IsConst ? (dyn_cast<ConstantSDNode>(N0))->getAPIntValue()
1915 N1IsConst = dyn_cast<ConstantSDNode>(N1) != nullptr;
1916 ConstValue1 = N1IsConst ? (dyn_cast<ConstantSDNode>(N1))->getAPIntValue()
1920 // fold (mul c1, c2) -> c1*c2
1921 if (N0IsConst && N1IsConst)
1922 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0.getNode(), N1.getNode());
1924 // canonicalize constant to RHS
1925 if (N0IsConst && !N1IsConst)
1926 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0);
1927 // fold (mul x, 0) -> 0
1928 if (N1IsConst && ConstValue1 == 0)
1930 // We require a splat of the entire scalar bit width for non-contiguous
1933 ConstValue1.getBitWidth() == VT.getScalarType().getSizeInBits();
1934 // fold (mul x, 1) -> x
1935 if (N1IsConst && ConstValue1 == 1 && IsFullSplat)
1937 // fold (mul x, -1) -> 0-x
1938 if (N1IsConst && ConstValue1.isAllOnesValue())
1939 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1940 DAG.getConstant(0, VT), N0);
1941 // fold (mul x, (1 << c)) -> x << c
1942 if (N1IsConst && ConstValue1.isPowerOf2() && IsFullSplat)
1943 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
1944 DAG.getConstant(ConstValue1.logBase2(),
1945 getShiftAmountTy(N0.getValueType())));
1946 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
1947 if (N1IsConst && (-ConstValue1).isPowerOf2() && IsFullSplat) {
1948 unsigned Log2Val = (-ConstValue1).logBase2();
1949 // FIXME: If the input is something that is easily negated (e.g. a
1950 // single-use add), we should put the negate there.
1951 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1952 DAG.getConstant(0, VT),
1953 DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
1954 DAG.getConstant(Log2Val,
1955 getShiftAmountTy(N0.getValueType()))));
1959 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
1960 if (N1IsConst && N0.getOpcode() == ISD::SHL &&
1961 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1962 isa<ConstantSDNode>(N0.getOperand(1)))) {
1963 SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT,
1964 N1, N0.getOperand(1));
1965 AddToWorklist(C3.getNode());
1966 return DAG.getNode(ISD::MUL, SDLoc(N), VT,
1967 N0.getOperand(0), C3);
1970 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
1973 SDValue Sh(nullptr,0), Y(nullptr,0);
1974 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
1975 if (N0.getOpcode() == ISD::SHL &&
1976 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1977 isa<ConstantSDNode>(N0.getOperand(1))) &&
1978 N0.getNode()->hasOneUse()) {
1980 } else if (N1.getOpcode() == ISD::SHL &&
1981 isa<ConstantSDNode>(N1.getOperand(1)) &&
1982 N1.getNode()->hasOneUse()) {
1987 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
1988 Sh.getOperand(0), Y);
1989 return DAG.getNode(ISD::SHL, SDLoc(N), VT,
1990 Mul, Sh.getOperand(1));
1994 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
1995 if (N1IsConst && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
1996 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1997 isa<ConstantSDNode>(N0.getOperand(1))))
1998 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1999 DAG.getNode(ISD::MUL, SDLoc(N0), VT,
2000 N0.getOperand(0), N1),
2001 DAG.getNode(ISD::MUL, SDLoc(N1), VT,
2002 N0.getOperand(1), N1));
2005 SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1);
2012 SDValue DAGCombiner::visitSDIV(SDNode *N) {
2013 SDValue N0 = N->getOperand(0);
2014 SDValue N1 = N->getOperand(1);
2015 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2016 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2017 EVT VT = N->getValueType(0);
2020 if (VT.isVector()) {
2021 SDValue FoldedVOp = SimplifyVBinOp(N);
2022 if (FoldedVOp.getNode()) return FoldedVOp;
2025 // fold (sdiv c1, c2) -> c1/c2
2026 if (N0C && N1C && !N1C->isNullValue())
2027 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
2028 // fold (sdiv X, 1) -> X
2029 if (N1C && N1C->getAPIntValue() == 1LL)
2031 // fold (sdiv X, -1) -> 0-X
2032 if (N1C && N1C->isAllOnesValue())
2033 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2034 DAG.getConstant(0, VT), N0);
2035 // If we know the sign bits of both operands are zero, strength reduce to a
2036 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
2037 if (!VT.isVector()) {
2038 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2039 return DAG.getNode(ISD::UDIV, SDLoc(N), N1.getValueType(),
2043 // fold (sdiv X, pow2) -> simple ops after legalize
2044 if (N1C && !N1C->isNullValue() && (N1C->getAPIntValue().isPowerOf2() ||
2045 (-N1C->getAPIntValue()).isPowerOf2())) {
2046 // If dividing by powers of two is cheap, then don't perform the following
2048 if (TLI.isPow2SDivCheap())
2051 // Target-specific implementation of sdiv x, pow2.
2052 SDValue Res = BuildSDIVPow2(N);
2056 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
2058 // Splat the sign bit into the register
2060 DAG.getNode(ISD::SRA, SDLoc(N), VT, N0,
2061 DAG.getConstant(VT.getScalarSizeInBits() - 1,
2062 getShiftAmountTy(N0.getValueType())));
2063 AddToWorklist(SGN.getNode());
2065 // Add (N0 < 0) ? abs2 - 1 : 0;
2067 DAG.getNode(ISD::SRL, SDLoc(N), VT, SGN,
2068 DAG.getConstant(VT.getScalarSizeInBits() - lg2,
2069 getShiftAmountTy(SGN.getValueType())));
2070 SDValue ADD = DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, SRL);
2071 AddToWorklist(SRL.getNode());
2072 AddToWorklist(ADD.getNode()); // Divide by pow2
2073 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), VT, ADD,
2074 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType())));
2076 // If we're dividing by a positive value, we're done. Otherwise, we must
2077 // negate the result.
2078 if (N1C->getAPIntValue().isNonNegative())
2081 AddToWorklist(SRA.getNode());
2082 return DAG.getNode(ISD::SUB, SDLoc(N), VT, DAG.getConstant(0, VT), SRA);
2085 // if integer divide is expensive and we satisfy the requirements, emit an
2086 // alternate sequence.
2087 if (N1C && !TLI.isIntDivCheap()) {
2088 SDValue Op = BuildSDIV(N);
2089 if (Op.getNode()) return Op;
2093 if (N0.getOpcode() == ISD::UNDEF)
2094 return DAG.getConstant(0, VT);
2095 // X / undef -> undef
2096 if (N1.getOpcode() == ISD::UNDEF)
2102 SDValue DAGCombiner::visitUDIV(SDNode *N) {
2103 SDValue N0 = N->getOperand(0);
2104 SDValue N1 = N->getOperand(1);
2105 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2106 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2107 EVT VT = N->getValueType(0);
2110 if (VT.isVector()) {
2111 SDValue FoldedVOp = SimplifyVBinOp(N);
2112 if (FoldedVOp.getNode()) return FoldedVOp;
2115 // fold (udiv c1, c2) -> c1/c2
2116 if (N0C && N1C && !N1C->isNullValue())
2117 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
2118 // fold (udiv x, (1 << c)) -> x >>u c
2119 if (N1C && N1C->getAPIntValue().isPowerOf2())
2120 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0,
2121 DAG.getConstant(N1C->getAPIntValue().logBase2(),
2122 getShiftAmountTy(N0.getValueType())));
2123 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
2124 if (N1.getOpcode() == ISD::SHL) {
2125 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2126 if (SHC->getAPIntValue().isPowerOf2()) {
2127 EVT ADDVT = N1.getOperand(1).getValueType();
2128 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N), ADDVT,
2130 DAG.getConstant(SHC->getAPIntValue()
2133 AddToWorklist(Add.getNode());
2134 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, Add);
2138 // fold (udiv x, c) -> alternate
2139 if (N1C && !TLI.isIntDivCheap()) {
2140 SDValue Op = BuildUDIV(N);
2141 if (Op.getNode()) return Op;
2145 if (N0.getOpcode() == ISD::UNDEF)
2146 return DAG.getConstant(0, VT);
2147 // X / undef -> undef
2148 if (N1.getOpcode() == ISD::UNDEF)
2154 SDValue DAGCombiner::visitSREM(SDNode *N) {
2155 SDValue N0 = N->getOperand(0);
2156 SDValue N1 = N->getOperand(1);
2157 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2158 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2159 EVT VT = N->getValueType(0);
2161 // fold (srem c1, c2) -> c1%c2
2162 if (N0C && N1C && !N1C->isNullValue())
2163 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
2164 // If we know the sign bits of both operands are zero, strength reduce to a
2165 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
2166 if (!VT.isVector()) {
2167 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2168 return DAG.getNode(ISD::UREM, SDLoc(N), VT, N0, N1);
2171 // If X/C can be simplified by the division-by-constant logic, lower
2172 // X%C to the equivalent of X-X/C*C.
2173 if (N1C && !N1C->isNullValue()) {
2174 SDValue Div = DAG.getNode(ISD::SDIV, SDLoc(N), VT, N0, N1);
2175 AddToWorklist(Div.getNode());
2176 SDValue OptimizedDiv = combine(Div.getNode());
2177 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2178 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2180 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2181 AddToWorklist(Mul.getNode());
2187 if (N0.getOpcode() == ISD::UNDEF)
2188 return DAG.getConstant(0, VT);
2189 // X % undef -> undef
2190 if (N1.getOpcode() == ISD::UNDEF)
2196 SDValue DAGCombiner::visitUREM(SDNode *N) {
2197 SDValue N0 = N->getOperand(0);
2198 SDValue N1 = N->getOperand(1);
2199 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2200 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2201 EVT VT = N->getValueType(0);
2203 // fold (urem c1, c2) -> c1%c2
2204 if (N0C && N1C && !N1C->isNullValue())
2205 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
2206 // fold (urem x, pow2) -> (and x, pow2-1)
2207 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
2208 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0,
2209 DAG.getConstant(N1C->getAPIntValue()-1,VT));
2210 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2211 if (N1.getOpcode() == ISD::SHL) {
2212 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2213 if (SHC->getAPIntValue().isPowerOf2()) {
2215 DAG.getNode(ISD::ADD, SDLoc(N), VT, N1,
2216 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
2218 AddToWorklist(Add.getNode());
2219 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, Add);
2224 // If X/C can be simplified by the division-by-constant logic, lower
2225 // X%C to the equivalent of X-X/C*C.
2226 if (N1C && !N1C->isNullValue()) {
2227 SDValue Div = DAG.getNode(ISD::UDIV, SDLoc(N), VT, N0, N1);
2228 AddToWorklist(Div.getNode());
2229 SDValue OptimizedDiv = combine(Div.getNode());
2230 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2231 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2233 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2234 AddToWorklist(Mul.getNode());
2240 if (N0.getOpcode() == ISD::UNDEF)
2241 return DAG.getConstant(0, VT);
2242 // X % undef -> undef
2243 if (N1.getOpcode() == ISD::UNDEF)
2249 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2250 SDValue N0 = N->getOperand(0);
2251 SDValue N1 = N->getOperand(1);
2252 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2253 EVT VT = N->getValueType(0);
2256 // fold (mulhs x, 0) -> 0
2257 if (N1C && N1C->isNullValue())
2259 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2260 if (N1C && N1C->getAPIntValue() == 1)
2261 return DAG.getNode(ISD::SRA, SDLoc(N), N0.getValueType(), N0,
2262 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2263 getShiftAmountTy(N0.getValueType())));
2264 // fold (mulhs x, undef) -> 0
2265 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2266 return DAG.getConstant(0, VT);
2268 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2270 if (VT.isSimple() && !VT.isVector()) {
2271 MVT Simple = VT.getSimpleVT();
2272 unsigned SimpleSize = Simple.getSizeInBits();
2273 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2274 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2275 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2276 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2277 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2278 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2279 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2280 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2287 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2288 SDValue N0 = N->getOperand(0);
2289 SDValue N1 = N->getOperand(1);
2290 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2291 EVT VT = N->getValueType(0);
2294 // fold (mulhu x, 0) -> 0
2295 if (N1C && N1C->isNullValue())
2297 // fold (mulhu x, 1) -> 0
2298 if (N1C && N1C->getAPIntValue() == 1)
2299 return DAG.getConstant(0, N0.getValueType());
2300 // fold (mulhu x, undef) -> 0
2301 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2302 return DAG.getConstant(0, VT);
2304 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2306 if (VT.isSimple() && !VT.isVector()) {
2307 MVT Simple = VT.getSimpleVT();
2308 unsigned SimpleSize = Simple.getSizeInBits();
2309 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2310 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2311 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2312 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2313 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2314 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2315 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2316 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2323 /// Perform optimizations common to nodes that compute two values. LoOp and HiOp
2324 /// give the opcodes for the two computations that are being performed. Return
2325 /// true if a simplification was made.
2326 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2328 // If the high half is not needed, just compute the low half.
2329 bool HiExists = N->hasAnyUseOfValue(1);
2331 (!LegalOperations ||
2332 TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) {
2333 SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2334 return CombineTo(N, Res, Res);
2337 // If the low half is not needed, just compute the high half.
2338 bool LoExists = N->hasAnyUseOfValue(0);
2340 (!LegalOperations ||
2341 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2342 SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2343 return CombineTo(N, Res, Res);
2346 // If both halves are used, return as it is.
2347 if (LoExists && HiExists)
2350 // If the two computed results can be simplified separately, separate them.
2352 SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2353 AddToWorklist(Lo.getNode());
2354 SDValue LoOpt = combine(Lo.getNode());
2355 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2356 (!LegalOperations ||
2357 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2358 return CombineTo(N, LoOpt, LoOpt);
2362 SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2363 AddToWorklist(Hi.getNode());
2364 SDValue HiOpt = combine(Hi.getNode());
2365 if (HiOpt.getNode() && HiOpt != Hi &&
2366 (!LegalOperations ||
2367 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2368 return CombineTo(N, HiOpt, HiOpt);
2374 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2375 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2376 if (Res.getNode()) return Res;
2378 EVT VT = N->getValueType(0);
2381 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2383 if (VT.isSimple() && !VT.isVector()) {
2384 MVT Simple = VT.getSimpleVT();
2385 unsigned SimpleSize = Simple.getSizeInBits();
2386 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2387 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2388 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2389 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2390 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2391 // Compute the high part as N1.
2392 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2393 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2394 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2395 // Compute the low part as N0.
2396 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2397 return CombineTo(N, Lo, Hi);
2404 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2405 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2406 if (Res.getNode()) return Res;
2408 EVT VT = N->getValueType(0);
2411 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2413 if (VT.isSimple() && !VT.isVector()) {
2414 MVT Simple = VT.getSimpleVT();
2415 unsigned SimpleSize = Simple.getSizeInBits();
2416 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2417 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2418 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2419 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2420 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2421 // Compute the high part as N1.
2422 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2423 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2424 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2425 // Compute the low part as N0.
2426 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2427 return CombineTo(N, Lo, Hi);
2434 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2435 // (smulo x, 2) -> (saddo x, x)
2436 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2437 if (C2->getAPIntValue() == 2)
2438 return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(),
2439 N->getOperand(0), N->getOperand(0));
2444 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2445 // (umulo x, 2) -> (uaddo x, x)
2446 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2447 if (C2->getAPIntValue() == 2)
2448 return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(),
2449 N->getOperand(0), N->getOperand(0));
2454 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2455 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2456 if (Res.getNode()) return Res;
2461 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2462 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2463 if (Res.getNode()) return Res;
2468 /// If this is a binary operator with two operands of the same opcode, try to
2470 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2471 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2472 EVT VT = N0.getValueType();
2473 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2475 // Bail early if none of these transforms apply.
2476 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2478 // For each of OP in AND/OR/XOR:
2479 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2480 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2481 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2482 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2484 // do not sink logical op inside of a vector extend, since it may combine
2486 EVT Op0VT = N0.getOperand(0).getValueType();
2487 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2488 N0.getOpcode() == ISD::SIGN_EXTEND ||
2489 // Avoid infinite looping with PromoteIntBinOp.
2490 (N0.getOpcode() == ISD::ANY_EXTEND &&
2491 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2492 (N0.getOpcode() == ISD::TRUNCATE &&
2493 (!TLI.isZExtFree(VT, Op0VT) ||
2494 !TLI.isTruncateFree(Op0VT, VT)) &&
2495 TLI.isTypeLegal(Op0VT))) &&
2497 Op0VT == N1.getOperand(0).getValueType() &&
2498 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2499 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2500 N0.getOperand(0).getValueType(),
2501 N0.getOperand(0), N1.getOperand(0));
2502 AddToWorklist(ORNode.getNode());
2503 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode);
2506 // For each of OP in SHL/SRL/SRA/AND...
2507 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2508 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2509 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2510 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2511 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2512 N0.getOperand(1) == N1.getOperand(1)) {
2513 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2514 N0.getOperand(0).getValueType(),
2515 N0.getOperand(0), N1.getOperand(0));
2516 AddToWorklist(ORNode.getNode());
2517 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
2518 ORNode, N0.getOperand(1));
2521 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2522 // Only perform this optimization after type legalization and before
2523 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2524 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2525 // we don't want to undo this promotion.
2526 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2528 if ((N0.getOpcode() == ISD::BITCAST ||
2529 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) &&
2530 Level == AfterLegalizeTypes) {
2531 SDValue In0 = N0.getOperand(0);
2532 SDValue In1 = N1.getOperand(0);
2533 EVT In0Ty = In0.getValueType();
2534 EVT In1Ty = In1.getValueType();
2536 // If both incoming values are integers, and the original types are the
2538 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2539 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1);
2540 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op);
2541 AddToWorklist(Op.getNode());
2546 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2547 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2548 // If both shuffles use the same mask, and both shuffle within a single
2549 // vector, then it is worthwhile to move the swizzle after the operation.
2550 // The type-legalizer generates this pattern when loading illegal
2551 // vector types from memory. In many cases this allows additional shuffle
2553 // There are other cases where moving the shuffle after the xor/and/or
2554 // is profitable even if shuffles don't perform a swizzle.
2555 // If both shuffles use the same mask, and both shuffles have the same first
2556 // or second operand, then it might still be profitable to move the shuffle
2557 // after the xor/and/or operation.
2558 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) {
2559 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2560 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2562 assert(N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() &&
2563 "Inputs to shuffles are not the same type");
2565 // Check that both shuffles use the same mask. The masks are known to be of
2566 // the same length because the result vector type is the same.
2567 // Check also that shuffles have only one use to avoid introducing extra
2569 if (SVN0->hasOneUse() && SVN1->hasOneUse() &&
2570 SVN0->getMask().equals(SVN1->getMask())) {
2571 SDValue ShOp = N0->getOperand(1);
2573 // Don't try to fold this node if it requires introducing a
2574 // build vector of all zeros that might be illegal at this stage.
2575 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2577 ShOp = DAG.getConstant(0, VT);
2582 // (AND (shuf (A, C), shuf (B, C)) -> shuf (AND (A, B), C)
2583 // (OR (shuf (A, C), shuf (B, C)) -> shuf (OR (A, B), C)
2584 // (XOR (shuf (A, C), shuf (B, C)) -> shuf (XOR (A, B), V_0)
2585 if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) {
2586 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2587 N0->getOperand(0), N1->getOperand(0));
2588 AddToWorklist(NewNode.getNode());
2589 return DAG.getVectorShuffle(VT, SDLoc(N), NewNode, ShOp,
2590 &SVN0->getMask()[0]);
2593 // Don't try to fold this node if it requires introducing a
2594 // build vector of all zeros that might be illegal at this stage.
2595 ShOp = N0->getOperand(0);
2596 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2598 ShOp = DAG.getConstant(0, VT);
2603 // (AND (shuf (C, A), shuf (C, B)) -> shuf (C, AND (A, B))
2604 // (OR (shuf (C, A), shuf (C, B)) -> shuf (C, OR (A, B))
2605 // (XOR (shuf (C, A), shuf (C, B)) -> shuf (V_0, XOR (A, B))
2606 if (N0->getOperand(0) == N1->getOperand(0) && ShOp.getNode()) {
2607 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2608 N0->getOperand(1), N1->getOperand(1));
2609 AddToWorklist(NewNode.getNode());
2610 return DAG.getVectorShuffle(VT, SDLoc(N), ShOp, NewNode,
2611 &SVN0->getMask()[0]);
2619 SDValue DAGCombiner::visitAND(SDNode *N) {
2620 SDValue N0 = N->getOperand(0);
2621 SDValue N1 = N->getOperand(1);
2622 SDValue LL, LR, RL, RR, CC0, CC1;
2623 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2624 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2625 EVT VT = N1.getValueType();
2626 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2629 if (VT.isVector()) {
2630 SDValue FoldedVOp = SimplifyVBinOp(N);
2631 if (FoldedVOp.getNode()) return FoldedVOp;
2633 // fold (and x, 0) -> 0, vector edition
2634 if (ISD::isBuildVectorAllZeros(N0.getNode()))
2635 // do not return N0, because undef node may exist in N0
2636 return DAG.getConstant(
2637 APInt::getNullValue(
2638 N0.getValueType().getScalarType().getSizeInBits()),
2640 if (ISD::isBuildVectorAllZeros(N1.getNode()))
2641 // do not return N1, because undef node may exist in N1
2642 return DAG.getConstant(
2643 APInt::getNullValue(
2644 N1.getValueType().getScalarType().getSizeInBits()),
2647 // fold (and x, -1) -> x, vector edition
2648 if (ISD::isBuildVectorAllOnes(N0.getNode()))
2650 if (ISD::isBuildVectorAllOnes(N1.getNode()))
2654 // fold (and x, undef) -> 0
2655 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2656 return DAG.getConstant(0, VT);
2657 // fold (and c1, c2) -> c1&c2
2659 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2660 // canonicalize constant to RHS
2662 return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0);
2663 // fold (and x, -1) -> x
2664 if (N1C && N1C->isAllOnesValue())
2666 // if (and x, c) is known to be zero, return 0
2667 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2668 APInt::getAllOnesValue(BitWidth)))
2669 return DAG.getConstant(0, VT);
2671 SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1);
2674 // fold (and (or x, C), D) -> D if (C & D) == D
2675 if (N1C && N0.getOpcode() == ISD::OR)
2676 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2677 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2679 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2680 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2681 SDValue N0Op0 = N0.getOperand(0);
2682 APInt Mask = ~N1C->getAPIntValue();
2683 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2684 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2685 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
2686 N0.getValueType(), N0Op0);
2688 // Replace uses of the AND with uses of the Zero extend node.
2691 // We actually want to replace all uses of the any_extend with the
2692 // zero_extend, to avoid duplicating things. This will later cause this
2693 // AND to be folded.
2694 CombineTo(N0.getNode(), Zext);
2695 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2698 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2699 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2700 // already be zero by virtue of the width of the base type of the load.
2702 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2704 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2705 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2706 N0.getOpcode() == ISD::LOAD) {
2707 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2708 N0 : N0.getOperand(0) );
2710 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2711 // This can be a pure constant or a vector splat, in which case we treat the
2712 // vector as a scalar and use the splat value.
2713 APInt Constant = APInt::getNullValue(1);
2714 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2715 Constant = C->getAPIntValue();
2716 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2717 APInt SplatValue, SplatUndef;
2718 unsigned SplatBitSize;
2720 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2721 SplatBitSize, HasAnyUndefs);
2723 // Undef bits can contribute to a possible optimisation if set, so
2725 SplatValue |= SplatUndef;
2727 // The splat value may be something like "0x00FFFFFF", which means 0 for
2728 // the first vector value and FF for the rest, repeating. We need a mask
2729 // that will apply equally to all members of the vector, so AND all the
2730 // lanes of the constant together.
2731 EVT VT = Vector->getValueType(0);
2732 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2734 // If the splat value has been compressed to a bitlength lower
2735 // than the size of the vector lane, we need to re-expand it to
2737 if (BitWidth > SplatBitSize)
2738 for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
2739 SplatBitSize < BitWidth;
2740 SplatBitSize = SplatBitSize * 2)
2741 SplatValue |= SplatValue.shl(SplatBitSize);
2743 Constant = APInt::getAllOnesValue(BitWidth);
2744 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
2745 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
2749 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
2750 // actually legal and isn't going to get expanded, else this is a false
2752 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
2753 Load->getMemoryVT());
2755 // Resize the constant to the same size as the original memory access before
2756 // extension. If it is still the AllOnesValue then this AND is completely
2759 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
2762 switch (Load->getExtensionType()) {
2763 default: B = false; break;
2764 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
2766 case ISD::NON_EXTLOAD: B = true; break;
2769 if (B && Constant.isAllOnesValue()) {
2770 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
2771 // preserve semantics once we get rid of the AND.
2772 SDValue NewLoad(Load, 0);
2773 if (Load->getExtensionType() == ISD::EXTLOAD) {
2774 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
2775 Load->getValueType(0), SDLoc(Load),
2776 Load->getChain(), Load->getBasePtr(),
2777 Load->getOffset(), Load->getMemoryVT(),
2778 Load->getMemOperand());
2779 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
2780 if (Load->getNumValues() == 3) {
2781 // PRE/POST_INC loads have 3 values.
2782 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
2783 NewLoad.getValue(2) };
2784 CombineTo(Load, To, 3, true);
2786 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
2790 // Fold the AND away, taking care not to fold to the old load node if we
2792 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
2794 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2797 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2798 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2799 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2800 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2802 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2803 LL.getValueType().isInteger()) {
2804 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2805 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
2806 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2807 LR.getValueType(), LL, RL);
2808 AddToWorklist(ORNode.getNode());
2809 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
2811 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2812 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2813 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(N0),
2814 LR.getValueType(), LL, RL);
2815 AddToWorklist(ANDNode.getNode());
2816 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
2818 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2819 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2820 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2821 LR.getValueType(), LL, RL);
2822 AddToWorklist(ORNode.getNode());
2823 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
2826 // Simplify (and (setne X, 0), (setne X, -1)) -> (setuge (add X, 1), 2)
2827 if (LL == RL && isa<ConstantSDNode>(LR) && isa<ConstantSDNode>(RR) &&
2828 Op0 == Op1 && LL.getValueType().isInteger() &&
2829 Op0 == ISD::SETNE && ((cast<ConstantSDNode>(LR)->isNullValue() &&
2830 cast<ConstantSDNode>(RR)->isAllOnesValue()) ||
2831 (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
2832 cast<ConstantSDNode>(RR)->isNullValue()))) {
2833 SDValue ADDNode = DAG.getNode(ISD::ADD, SDLoc(N0), LL.getValueType(),
2834 LL, DAG.getConstant(1, LL.getValueType()));
2835 AddToWorklist(ADDNode.getNode());
2836 return DAG.getSetCC(SDLoc(N), VT, ADDNode,
2837 DAG.getConstant(2, LL.getValueType()), ISD::SETUGE);
2839 // canonicalize equivalent to ll == rl
2840 if (LL == RR && LR == RL) {
2841 Op1 = ISD::getSetCCSwappedOperands(Op1);
2844 if (LL == RL && LR == RR) {
2845 bool isInteger = LL.getValueType().isInteger();
2846 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2847 if (Result != ISD::SETCC_INVALID &&
2848 (!LegalOperations ||
2849 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
2850 TLI.isOperationLegal(ISD::SETCC,
2851 getSetCCResultType(N0.getSimpleValueType())))))
2852 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
2857 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
2858 if (N0.getOpcode() == N1.getOpcode()) {
2859 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2860 if (Tmp.getNode()) return Tmp;
2863 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
2864 // fold (and (sra)) -> (and (srl)) when possible.
2865 if (!VT.isVector() &&
2866 SimplifyDemandedBits(SDValue(N, 0)))
2867 return SDValue(N, 0);
2869 // fold (zext_inreg (extload x)) -> (zextload x)
2870 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
2871 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2872 EVT MemVT = LN0->getMemoryVT();
2873 // If we zero all the possible extended bits, then we can turn this into
2874 // a zextload if we are running before legalize or the operation is legal.
2875 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2876 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2877 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2878 ((!LegalOperations && !LN0->isVolatile()) ||
2879 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2880 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2881 LN0->getChain(), LN0->getBasePtr(),
2882 MemVT, LN0->getMemOperand());
2884 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2885 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2888 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
2889 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
2891 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2892 EVT MemVT = LN0->getMemoryVT();
2893 // If we zero all the possible extended bits, then we can turn this into
2894 // a zextload if we are running before legalize or the operation is legal.
2895 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2896 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2897 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2898 ((!LegalOperations && !LN0->isVolatile()) ||
2899 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2900 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2901 LN0->getChain(), LN0->getBasePtr(),
2902 MemVT, LN0->getMemOperand());
2904 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2905 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2909 // fold (and (load x), 255) -> (zextload x, i8)
2910 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2911 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2912 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2913 (N0.getOpcode() == ISD::ANY_EXTEND &&
2914 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2915 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2916 LoadSDNode *LN0 = HasAnyExt
2917 ? cast<LoadSDNode>(N0.getOperand(0))
2918 : cast<LoadSDNode>(N0);
2919 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2920 LN0->isUnindexed() && N0.hasOneUse() && SDValue(LN0, 0).hasOneUse()) {
2921 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2922 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2923 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2924 EVT LoadedVT = LN0->getMemoryVT();
2926 if (ExtVT == LoadedVT &&
2927 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2928 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2931 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
2932 LN0->getChain(), LN0->getBasePtr(), ExtVT,
2933 LN0->getMemOperand());
2935 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
2936 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2939 // Do not change the width of a volatile load.
2940 // Do not generate loads of non-round integer types since these can
2941 // be expensive (and would be wrong if the type is not byte sized).
2942 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
2943 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2944 EVT PtrType = LN0->getOperand(1).getValueType();
2946 unsigned Alignment = LN0->getAlignment();
2947 SDValue NewPtr = LN0->getBasePtr();
2949 // For big endian targets, we need to add an offset to the pointer
2950 // to load the correct bytes. For little endian systems, we merely
2951 // need to read fewer bytes from the same pointer.
2952 if (TLI.isBigEndian()) {
2953 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
2954 unsigned EVTStoreBytes = ExtVT.getStoreSize();
2955 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
2956 NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0), PtrType,
2957 NewPtr, DAG.getConstant(PtrOff, PtrType));
2958 Alignment = MinAlign(Alignment, PtrOff);
2961 AddToWorklist(NewPtr.getNode());
2963 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2965 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
2966 LN0->getChain(), NewPtr,
2967 LN0->getPointerInfo(),
2968 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2969 LN0->isInvariant(), Alignment, LN0->getAAInfo());
2971 CombineTo(LN0, Load, Load.getValue(1));
2972 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2978 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
2979 VT.getSizeInBits() <= 64) {
2980 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2981 APInt ADDC = ADDI->getAPIntValue();
2982 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2983 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
2984 // immediate for an add, but it is legal if its top c2 bits are set,
2985 // transform the ADD so the immediate doesn't need to be materialized
2987 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
2988 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
2989 SRLI->getZExtValue());
2990 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
2992 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2994 DAG.getNode(ISD::ADD, SDLoc(N0), VT,
2995 N0.getOperand(0), DAG.getConstant(ADDC, VT));
2996 CombineTo(N0.getNode(), NewAdd);
2997 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3005 // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const)
3006 if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) {
3007 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
3008 N0.getOperand(1), false);
3009 if (BSwap.getNode())
3016 /// Match (a >> 8) | (a << 8) as (bswap a) >> 16.
3017 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
3018 bool DemandHighBits) {
3019 if (!LegalOperations)
3022 EVT VT = N->getValueType(0);
3023 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
3025 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3028 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
3029 bool LookPassAnd0 = false;
3030 bool LookPassAnd1 = false;
3031 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
3033 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
3035 if (N0.getOpcode() == ISD::AND) {
3036 if (!N0.getNode()->hasOneUse())
3038 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3039 if (!N01C || N01C->getZExtValue() != 0xFF00)
3041 N0 = N0.getOperand(0);
3042 LookPassAnd0 = true;
3045 if (N1.getOpcode() == ISD::AND) {
3046 if (!N1.getNode()->hasOneUse())
3048 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3049 if (!N11C || N11C->getZExtValue() != 0xFF)
3051 N1 = N1.getOperand(0);
3052 LookPassAnd1 = true;
3055 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
3057 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
3059 if (!N0.getNode()->hasOneUse() ||
3060 !N1.getNode()->hasOneUse())
3063 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3064 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3067 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
3070 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
3071 SDValue N00 = N0->getOperand(0);
3072 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
3073 if (!N00.getNode()->hasOneUse())
3075 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
3076 if (!N001C || N001C->getZExtValue() != 0xFF)
3078 N00 = N00.getOperand(0);
3079 LookPassAnd0 = true;
3082 SDValue N10 = N1->getOperand(0);
3083 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
3084 if (!N10.getNode()->hasOneUse())
3086 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
3087 if (!N101C || N101C->getZExtValue() != 0xFF00)
3089 N10 = N10.getOperand(0);
3090 LookPassAnd1 = true;
3096 // Make sure everything beyond the low halfword gets set to zero since the SRL
3097 // 16 will clear the top bits.
3098 unsigned OpSizeInBits = VT.getSizeInBits();
3099 if (DemandHighBits && OpSizeInBits > 16) {
3100 // If the left-shift isn't masked out then the only way this is a bswap is
3101 // if all bits beyond the low 8 are 0. In that case the entire pattern
3102 // reduces to a left shift anyway: leave it for other parts of the combiner.
3106 // However, if the right shift isn't masked out then it might be because
3107 // it's not needed. See if we can spot that too.
3108 if (!LookPassAnd1 &&
3109 !DAG.MaskedValueIsZero(
3110 N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16)))
3114 SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00);
3115 if (OpSizeInBits > 16)
3116 Res = DAG.getNode(ISD::SRL, SDLoc(N), VT, Res,
3117 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
3121 /// Return true if the specified node is an element that makes up a 32-bit
3122 /// packed halfword byteswap.
3123 /// ((x & 0x000000ff) << 8) |
3124 /// ((x & 0x0000ff00) >> 8) |
3125 /// ((x & 0x00ff0000) << 8) |
3126 /// ((x & 0xff000000) >> 8)
3127 static bool isBSwapHWordElement(SDValue N, SmallVectorImpl<SDNode *> &Parts) {
3128 if (!N.getNode()->hasOneUse())
3131 unsigned Opc = N.getOpcode();
3132 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
3135 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3140 switch (N1C->getZExtValue()) {
3143 case 0xFF: Num = 0; break;
3144 case 0xFF00: Num = 1; break;
3145 case 0xFF0000: Num = 2; break;
3146 case 0xFF000000: Num = 3; break;
3149 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
3150 SDValue N0 = N.getOperand(0);
3151 if (Opc == ISD::AND) {
3152 if (Num == 0 || Num == 2) {
3154 // (x >> 8) & 0xff0000
3155 if (N0.getOpcode() != ISD::SRL)
3157 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3158 if (!C || C->getZExtValue() != 8)
3161 // (x << 8) & 0xff00
3162 // (x << 8) & 0xff000000
3163 if (N0.getOpcode() != ISD::SHL)
3165 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3166 if (!C || C->getZExtValue() != 8)
3169 } else if (Opc == ISD::SHL) {
3171 // (x & 0xff0000) << 8
3172 if (Num != 0 && Num != 2)
3174 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3175 if (!C || C->getZExtValue() != 8)
3177 } else { // Opc == ISD::SRL
3178 // (x & 0xff00) >> 8
3179 // (x & 0xff000000) >> 8
3180 if (Num != 1 && Num != 3)
3182 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3183 if (!C || C->getZExtValue() != 8)
3190 Parts[Num] = N0.getOperand(0).getNode();
3194 /// Match a 32-bit packed halfword bswap. That is
3195 /// ((x & 0x000000ff) << 8) |
3196 /// ((x & 0x0000ff00) >> 8) |
3197 /// ((x & 0x00ff0000) << 8) |
3198 /// ((x & 0xff000000) >> 8)
3199 /// => (rotl (bswap x), 16)
3200 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
3201 if (!LegalOperations)
3204 EVT VT = N->getValueType(0);
3207 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3210 SmallVector<SDNode*,4> Parts(4, (SDNode*)nullptr);
3212 // (or (or (and), (and)), (or (and), (and)))
3213 // (or (or (or (and), (and)), (and)), (and))
3214 if (N0.getOpcode() != ISD::OR)
3216 SDValue N00 = N0.getOperand(0);
3217 SDValue N01 = N0.getOperand(1);
3219 if (N1.getOpcode() == ISD::OR &&
3220 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) {
3221 // (or (or (and), (and)), (or (and), (and)))
3222 SDValue N000 = N00.getOperand(0);
3223 if (!isBSwapHWordElement(N000, Parts))
3226 SDValue N001 = N00.getOperand(1);
3227 if (!isBSwapHWordElement(N001, Parts))
3229 SDValue N010 = N01.getOperand(0);
3230 if (!isBSwapHWordElement(N010, Parts))
3232 SDValue N011 = N01.getOperand(1);
3233 if (!isBSwapHWordElement(N011, Parts))
3236 // (or (or (or (and), (and)), (and)), (and))
3237 if (!isBSwapHWordElement(N1, Parts))
3239 if (!isBSwapHWordElement(N01, Parts))
3241 if (N00.getOpcode() != ISD::OR)
3243 SDValue N000 = N00.getOperand(0);
3244 if (!isBSwapHWordElement(N000, Parts))
3246 SDValue N001 = N00.getOperand(1);
3247 if (!isBSwapHWordElement(N001, Parts))
3251 // Make sure the parts are all coming from the same node.
3252 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
3255 SDValue BSwap = DAG.getNode(ISD::BSWAP, SDLoc(N), VT,
3256 SDValue(Parts[0],0));
3258 // Result of the bswap should be rotated by 16. If it's not legal, then
3259 // do (x << 16) | (x >> 16).
3260 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
3261 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3262 return DAG.getNode(ISD::ROTL, SDLoc(N), VT, BSwap, ShAmt);
3263 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3264 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, BSwap, ShAmt);
3265 return DAG.getNode(ISD::OR, SDLoc(N), VT,
3266 DAG.getNode(ISD::SHL, SDLoc(N), VT, BSwap, ShAmt),
3267 DAG.getNode(ISD::SRL, SDLoc(N), VT, BSwap, ShAmt));
3270 SDValue DAGCombiner::visitOR(SDNode *N) {
3271 SDValue N0 = N->getOperand(0);
3272 SDValue N1 = N->getOperand(1);
3273 SDValue LL, LR, RL, RR, CC0, CC1;
3274 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3275 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3276 EVT VT = N1.getValueType();
3279 if (VT.isVector()) {
3280 SDValue FoldedVOp = SimplifyVBinOp(N);
3281 if (FoldedVOp.getNode()) return FoldedVOp;
3283 // fold (or x, 0) -> x, vector edition
3284 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3286 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3289 // fold (or x, -1) -> -1, vector edition
3290 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3291 // do not return N0, because undef node may exist in N0
3292 return DAG.getConstant(
3293 APInt::getAllOnesValue(
3294 N0.getValueType().getScalarType().getSizeInBits()),
3296 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3297 // do not return N1, because undef node may exist in N1
3298 return DAG.getConstant(
3299 APInt::getAllOnesValue(
3300 N1.getValueType().getScalarType().getSizeInBits()),
3303 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask1)
3304 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf B, A, Mask2)
3305 // Do this only if the resulting shuffle is legal.
3306 if (isa<ShuffleVectorSDNode>(N0) &&
3307 isa<ShuffleVectorSDNode>(N1) &&
3308 // Avoid folding a node with illegal type.
3309 TLI.isTypeLegal(VT) &&
3310 N0->getOperand(1) == N1->getOperand(1) &&
3311 ISD::isBuildVectorAllZeros(N0.getOperand(1).getNode())) {
3312 bool CanFold = true;
3313 unsigned NumElts = VT.getVectorNumElements();
3314 const ShuffleVectorSDNode *SV0 = cast<ShuffleVectorSDNode>(N0);
3315 const ShuffleVectorSDNode *SV1 = cast<ShuffleVectorSDNode>(N1);
3316 // We construct two shuffle masks:
3317 // - Mask1 is a shuffle mask for a shuffle with N0 as the first operand
3318 // and N1 as the second operand.
3319 // - Mask2 is a shuffle mask for a shuffle with N1 as the first operand
3320 // and N0 as the second operand.
3321 // We do this because OR is commutable and therefore there might be
3322 // two ways to fold this node into a shuffle.
3323 SmallVector<int,4> Mask1;
3324 SmallVector<int,4> Mask2;
3326 for (unsigned i = 0; i != NumElts && CanFold; ++i) {
3327 int M0 = SV0->getMaskElt(i);
3328 int M1 = SV1->getMaskElt(i);
3330 // Both shuffle indexes are undef. Propagate Undef.
3331 if (M0 < 0 && M1 < 0) {
3332 Mask1.push_back(M0);
3333 Mask2.push_back(M0);
3337 if (M0 < 0 || M1 < 0 ||
3338 (M0 < (int)NumElts && M1 < (int)NumElts) ||
3339 (M0 >= (int)NumElts && M1 >= (int)NumElts)) {
3344 Mask1.push_back(M0 < (int)NumElts ? M0 : M1 + NumElts);
3345 Mask2.push_back(M1 < (int)NumElts ? M1 : M0 + NumElts);
3349 // Fold this sequence only if the resulting shuffle is 'legal'.
3350 if (TLI.isShuffleMaskLegal(Mask1, VT))
3351 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0),
3352 N1->getOperand(0), &Mask1[0]);
3353 if (TLI.isShuffleMaskLegal(Mask2, VT))
3354 return DAG.getVectorShuffle(VT, SDLoc(N), N1->getOperand(0),
3355 N0->getOperand(0), &Mask2[0]);
3360 // fold (or x, undef) -> -1
3361 if (!LegalOperations &&
3362 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3363 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3364 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
3366 // fold (or c1, c2) -> c1|c2
3368 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
3369 // canonicalize constant to RHS
3371 return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0);
3372 // fold (or x, 0) -> x
3373 if (N1C && N1C->isNullValue())
3375 // fold (or x, -1) -> -1
3376 if (N1C && N1C->isAllOnesValue())
3378 // fold (or x, c) -> c iff (x & ~c) == 0
3379 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3382 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3383 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
3384 if (BSwap.getNode())
3386 BSwap = MatchBSwapHWordLow(N, N0, N1);
3387 if (BSwap.getNode())
3391 SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1);
3394 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3395 // iff (c1 & c2) == 0.
3396 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3397 isa<ConstantSDNode>(N0.getOperand(1))) {
3398 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3399 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) {
3400 SDValue COR = DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1);
3403 return DAG.getNode(ISD::AND, SDLoc(N), VT,
3404 DAG.getNode(ISD::OR, SDLoc(N0), VT,
3405 N0.getOperand(0), N1), COR);
3408 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3409 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3410 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3411 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3413 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3414 LL.getValueType().isInteger()) {
3415 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3416 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3417 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3418 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3419 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(LR),
3420 LR.getValueType(), LL, RL);
3421 AddToWorklist(ORNode.getNode());
3422 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
3424 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3425 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3426 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3427 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3428 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(LR),
3429 LR.getValueType(), LL, RL);
3430 AddToWorklist(ANDNode.getNode());
3431 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
3434 // canonicalize equivalent to ll == rl
3435 if (LL == RR && LR == RL) {
3436 Op1 = ISD::getSetCCSwappedOperands(Op1);
3439 if (LL == RL && LR == RR) {
3440 bool isInteger = LL.getValueType().isInteger();
3441 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3442 if (Result != ISD::SETCC_INVALID &&
3443 (!LegalOperations ||
3444 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
3445 TLI.isOperationLegal(ISD::SETCC,
3446 getSetCCResultType(N0.getValueType())))))
3447 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
3452 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3453 if (N0.getOpcode() == N1.getOpcode()) {
3454 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3455 if (Tmp.getNode()) return Tmp;
3458 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3459 if (N0.getOpcode() == ISD::AND &&
3460 N1.getOpcode() == ISD::AND &&
3461 N0.getOperand(1).getOpcode() == ISD::Constant &&
3462 N1.getOperand(1).getOpcode() == ISD::Constant &&
3463 // Don't increase # computations.
3464 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3465 // We can only do this xform if we know that bits from X that are set in C2
3466 // but not in C1 are already zero. Likewise for Y.
3467 const APInt &LHSMask =
3468 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3469 const APInt &RHSMask =
3470 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3472 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3473 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3474 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3475 N0.getOperand(0), N1.getOperand(0));
3476 return DAG.getNode(ISD::AND, SDLoc(N), VT, X,
3477 DAG.getConstant(LHSMask | RHSMask, VT));
3481 // See if this is some rotate idiom.
3482 if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N)))
3483 return SDValue(Rot, 0);
3485 // Simplify the operands using demanded-bits information.
3486 if (!VT.isVector() &&
3487 SimplifyDemandedBits(SDValue(N, 0)))
3488 return SDValue(N, 0);
3493 /// Match "(X shl/srl V1) & V2" where V2 may not be present.
3494 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3495 if (Op.getOpcode() == ISD::AND) {
3496 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3497 Mask = Op.getOperand(1);
3498 Op = Op.getOperand(0);
3504 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3512 // Return true if we can prove that, whenever Neg and Pos are both in the
3513 // range [0, OpSize), Neg == (Pos == 0 ? 0 : OpSize - Pos). This means that
3514 // for two opposing shifts shift1 and shift2 and a value X with OpBits bits:
3516 // (or (shift1 X, Neg), (shift2 X, Pos))
3518 // reduces to a rotate in direction shift2 by Pos or (equivalently) a rotate
3519 // in direction shift1 by Neg. The range [0, OpSize) means that we only need
3520 // to consider shift amounts with defined behavior.
3521 static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned OpSize) {
3522 // If OpSize is a power of 2 then:
3524 // (a) (Pos == 0 ? 0 : OpSize - Pos) == (OpSize - Pos) & (OpSize - 1)
3525 // (b) Neg == Neg & (OpSize - 1) whenever Neg is in [0, OpSize).
3527 // So if OpSize is a power of 2 and Neg is (and Neg', OpSize-1), we check
3528 // for the stronger condition:
3530 // Neg & (OpSize - 1) == (OpSize - Pos) & (OpSize - 1) [A]
3532 // for all Neg and Pos. Since Neg & (OpSize - 1) == Neg' & (OpSize - 1)
3533 // we can just replace Neg with Neg' for the rest of the function.
3535 // In other cases we check for the even stronger condition:
3537 // Neg == OpSize - Pos [B]
3539 // for all Neg and Pos. Note that the (or ...) then invokes undefined
3540 // behavior if Pos == 0 (and consequently Neg == OpSize).
3542 // We could actually use [A] whenever OpSize is a power of 2, but the
3543 // only extra cases that it would match are those uninteresting ones
3544 // where Neg and Pos are never in range at the same time. E.g. for
3545 // OpSize == 32, using [A] would allow a Neg of the form (sub 64, Pos)
3546 // as well as (sub 32, Pos), but:
3548 // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos))
3550 // always invokes undefined behavior for 32-bit X.
3552 // Below, Mask == OpSize - 1 when using [A] and is all-ones otherwise.
3553 unsigned MaskLoBits = 0;
3554 if (Neg.getOpcode() == ISD::AND &&
3555 isPowerOf2_64(OpSize) &&
3556 Neg.getOperand(1).getOpcode() == ISD::Constant &&
3557 cast<ConstantSDNode>(Neg.getOperand(1))->getAPIntValue() == OpSize - 1) {
3558 Neg = Neg.getOperand(0);
3559 MaskLoBits = Log2_64(OpSize);
3562 // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1.
3563 if (Neg.getOpcode() != ISD::SUB)
3565 ConstantSDNode *NegC = dyn_cast<ConstantSDNode>(Neg.getOperand(0));
3568 SDValue NegOp1 = Neg.getOperand(1);
3570 // On the RHS of [A], if Pos is Pos' & (OpSize - 1), just replace Pos with
3571 // Pos'. The truncation is redundant for the purpose of the equality.
3573 Pos.getOpcode() == ISD::AND &&
3574 Pos.getOperand(1).getOpcode() == ISD::Constant &&
3575 cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() == OpSize - 1)
3576 Pos = Pos.getOperand(0);
3578 // The condition we need is now:
3580 // (NegC - NegOp1) & Mask == (OpSize - Pos) & Mask
3582 // If NegOp1 == Pos then we need:
3584 // OpSize & Mask == NegC & Mask
3586 // (because "x & Mask" is a truncation and distributes through subtraction).
3589 Width = NegC->getAPIntValue();
3590 // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC.
3591 // Then the condition we want to prove becomes:
3593 // (NegC - NegOp1) & Mask == (OpSize - (NegOp1 + PosC)) & Mask
3595 // which, again because "x & Mask" is a truncation, becomes:
3597 // NegC & Mask == (OpSize - PosC) & Mask
3598 // OpSize & Mask == (NegC + PosC) & Mask
3599 else if (Pos.getOpcode() == ISD::ADD &&
3600 Pos.getOperand(0) == NegOp1 &&
3601 Pos.getOperand(1).getOpcode() == ISD::Constant)
3602 Width = (cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() +
3603 NegC->getAPIntValue());
3607 // Now we just need to check that OpSize & Mask == Width & Mask.
3609 // Opsize & Mask is 0 since Mask is Opsize - 1.
3610 return Width.getLoBits(MaskLoBits) == 0;
3611 return Width == OpSize;
3614 // A subroutine of MatchRotate used once we have found an OR of two opposite
3615 // shifts of Shifted. If Neg == <operand size> - Pos then the OR reduces
3616 // to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the
3617 // former being preferred if supported. InnerPos and InnerNeg are Pos and
3618 // Neg with outer conversions stripped away.
3619 SDNode *DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos,
3620 SDValue Neg, SDValue InnerPos,
3621 SDValue InnerNeg, unsigned PosOpcode,
3622 unsigned NegOpcode, SDLoc DL) {
3623 // fold (or (shl x, (*ext y)),
3624 // (srl x, (*ext (sub 32, y)))) ->
3625 // (rotl x, y) or (rotr x, (sub 32, y))
3627 // fold (or (shl x, (*ext (sub 32, y))),
3628 // (srl x, (*ext y))) ->
3629 // (rotr x, y) or (rotl x, (sub 32, y))
3630 EVT VT = Shifted.getValueType();
3631 if (matchRotateSub(InnerPos, InnerNeg, VT.getSizeInBits())) {
3632 bool HasPos = TLI.isOperationLegalOrCustom(PosOpcode, VT);
3633 return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted,
3634 HasPos ? Pos : Neg).getNode();
3640 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3641 // idioms for rotate, and if the target supports rotation instructions, generate
3643 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL) {
3644 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3645 EVT VT = LHS.getValueType();
3646 if (!TLI.isTypeLegal(VT)) return nullptr;
3648 // The target must have at least one rotate flavor.
3649 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3650 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3651 if (!HasROTL && !HasROTR) return nullptr;
3653 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3654 SDValue LHSShift; // The shift.
3655 SDValue LHSMask; // AND value if any.
3656 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3657 return nullptr; // Not part of a rotate.
3659 SDValue RHSShift; // The shift.
3660 SDValue RHSMask; // AND value if any.
3661 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3662 return nullptr; // Not part of a rotate.
3664 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3665 return nullptr; // Not shifting the same value.
3667 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3668 return nullptr; // Shifts must disagree.
3670 // Canonicalize shl to left side in a shl/srl pair.
3671 if (RHSShift.getOpcode() == ISD::SHL) {
3672 std::swap(LHS, RHS);
3673 std::swap(LHSShift, RHSShift);
3674 std::swap(LHSMask , RHSMask );
3677 unsigned OpSizeInBits = VT.getSizeInBits();
3678 SDValue LHSShiftArg = LHSShift.getOperand(0);
3679 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3680 SDValue RHSShiftArg = RHSShift.getOperand(0);
3681 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3683 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3684 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3685 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3686 RHSShiftAmt.getOpcode() == ISD::Constant) {
3687 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3688 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3689 if ((LShVal + RShVal) != OpSizeInBits)
3692 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3693 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
3695 // If there is an AND of either shifted operand, apply it to the result.
3696 if (LHSMask.getNode() || RHSMask.getNode()) {
3697 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3699 if (LHSMask.getNode()) {
3700 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3701 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3703 if (RHSMask.getNode()) {
3704 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3705 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3708 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3711 return Rot.getNode();
3714 // If there is a mask here, and we have a variable shift, we can't be sure
3715 // that we're masking out the right stuff.
3716 if (LHSMask.getNode() || RHSMask.getNode())
3719 // If the shift amount is sign/zext/any-extended just peel it off.
3720 SDValue LExtOp0 = LHSShiftAmt;
3721 SDValue RExtOp0 = RHSShiftAmt;
3722 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3723 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3724 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3725 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3726 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3727 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3728 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3729 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3730 LExtOp0 = LHSShiftAmt.getOperand(0);
3731 RExtOp0 = RHSShiftAmt.getOperand(0);
3734 SDNode *TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt,
3735 LExtOp0, RExtOp0, ISD::ROTL, ISD::ROTR, DL);
3739 SDNode *TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt,
3740 RExtOp0, LExtOp0, ISD::ROTR, ISD::ROTL, DL);
3747 SDValue DAGCombiner::visitXOR(SDNode *N) {
3748 SDValue N0 = N->getOperand(0);
3749 SDValue N1 = N->getOperand(1);
3750 SDValue LHS, RHS, CC;
3751 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3752 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3753 EVT VT = N0.getValueType();
3756 if (VT.isVector()) {
3757 SDValue FoldedVOp = SimplifyVBinOp(N);
3758 if (FoldedVOp.getNode()) return FoldedVOp;
3760 // fold (xor x, 0) -> x, vector edition
3761 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3763 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3767 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3768 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3769 return DAG.getConstant(0, VT);
3770 // fold (xor x, undef) -> undef
3771 if (N0.getOpcode() == ISD::UNDEF)
3773 if (N1.getOpcode() == ISD::UNDEF)
3775 // fold (xor c1, c2) -> c1^c2
3777 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3778 // canonicalize constant to RHS
3780 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
3781 // fold (xor x, 0) -> x
3782 if (N1C && N1C->isNullValue())
3785 SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1);
3789 // fold !(x cc y) -> (x !cc y)
3790 if (N1C && N1C->getAPIntValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3791 bool isInt = LHS.getValueType().isInteger();
3792 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3795 if (!LegalOperations ||
3796 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) {
3797 switch (N0.getOpcode()) {
3799 llvm_unreachable("Unhandled SetCC Equivalent!");
3801 return DAG.getSetCC(SDLoc(N), VT, LHS, RHS, NotCC);
3802 case ISD::SELECT_CC:
3803 return DAG.getSelectCC(SDLoc(N), LHS, RHS, N0.getOperand(2),
3804 N0.getOperand(3), NotCC);
3809 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3810 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3811 N0.getNode()->hasOneUse() &&
3812 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3813 SDValue V = N0.getOperand(0);
3814 V = DAG.getNode(ISD::XOR, SDLoc(N0), V.getValueType(), V,
3815 DAG.getConstant(1, V.getValueType()));
3816 AddToWorklist(V.getNode());
3817 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V);
3820 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3821 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3822 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3823 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3824 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3825 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3826 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3827 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3828 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
3829 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3832 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3833 if (N1C && N1C->isAllOnesValue() &&
3834 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3835 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3836 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3837 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3838 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3839 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3840 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
3841 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3844 // fold (xor (and x, y), y) -> (and (not x), y)
3845 if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3846 N0->getOperand(1) == N1) {
3847 SDValue X = N0->getOperand(0);
3848 SDValue NotX = DAG.getNOT(SDLoc(X), X, VT);
3849 AddToWorklist(NotX.getNode());
3850 return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1);
3852 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3853 if (N1C && N0.getOpcode() == ISD::XOR) {
3854 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3855 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3857 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(1),
3858 DAG.getConstant(N1C->getAPIntValue() ^
3859 N00C->getAPIntValue(), VT));
3861 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(0),
3862 DAG.getConstant(N1C->getAPIntValue() ^
3863 N01C->getAPIntValue(), VT));
3865 // fold (xor x, x) -> 0
3867 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
3869 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
3870 if (N0.getOpcode() == N1.getOpcode()) {
3871 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3872 if (Tmp.getNode()) return Tmp;
3875 // Simplify the expression using non-local knowledge.
3876 if (!VT.isVector() &&
3877 SimplifyDemandedBits(SDValue(N, 0)))
3878 return SDValue(N, 0);
3883 /// Handle transforms common to the three shifts, when the shift amount is a
3885 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, ConstantSDNode *Amt) {
3886 // We can't and shouldn't fold opaque constants.
3887 if (Amt->isOpaque())
3890 SDNode *LHS = N->getOperand(0).getNode();
3891 if (!LHS->hasOneUse()) return SDValue();
3893 // We want to pull some binops through shifts, so that we have (and (shift))
3894 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
3895 // thing happens with address calculations, so it's important to canonicalize
3897 bool HighBitSet = false; // Can we transform this if the high bit is set?
3899 switch (LHS->getOpcode()) {
3900 default: return SDValue();
3903 HighBitSet = false; // We can only transform sra if the high bit is clear.
3906 HighBitSet = true; // We can only transform sra if the high bit is set.
3909 if (N->getOpcode() != ISD::SHL)
3910 return SDValue(); // only shl(add) not sr[al](add).
3911 HighBitSet = false; // We can only transform sra if the high bit is clear.
3915 // We require the RHS of the binop to be a constant and not opaque as well.
3916 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
3917 if (!BinOpCst || BinOpCst->isOpaque()) return SDValue();
3919 // FIXME: disable this unless the input to the binop is a shift by a constant.
3920 // If it is not a shift, it pessimizes some common cases like:
3922 // void foo(int *X, int i) { X[i & 1235] = 1; }
3923 // int bar(int *X, int i) { return X[i & 255]; }
3924 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
3925 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
3926 BinOpLHSVal->getOpcode() != ISD::SRA &&
3927 BinOpLHSVal->getOpcode() != ISD::SRL) ||
3928 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
3931 EVT VT = N->getValueType(0);
3933 // If this is a signed shift right, and the high bit is modified by the
3934 // logical operation, do not perform the transformation. The highBitSet
3935 // boolean indicates the value of the high bit of the constant which would
3936 // cause it to be modified for this operation.
3937 if (N->getOpcode() == ISD::SRA) {
3938 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
3939 if (BinOpRHSSignSet != HighBitSet)
3943 if (!TLI.isDesirableToCommuteWithShift(LHS))
3946 // Fold the constants, shifting the binop RHS by the shift amount.
3947 SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)),
3949 LHS->getOperand(1), N->getOperand(1));
3950 assert(isa<ConstantSDNode>(NewRHS) && "Folding was not successful!");
3952 // Create the new shift.
3953 SDValue NewShift = DAG.getNode(N->getOpcode(),
3954 SDLoc(LHS->getOperand(0)),
3955 VT, LHS->getOperand(0), N->getOperand(1));
3957 // Create the new binop.
3958 return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS);
3961 SDValue DAGCombiner::distributeTruncateThroughAnd(SDNode *N) {
3962 assert(N->getOpcode() == ISD::TRUNCATE);
3963 assert(N->getOperand(0).getOpcode() == ISD::AND);
3965 // (truncate:TruncVT (and N00, N01C)) -> (and (truncate:TruncVT N00), TruncC)
3966 if (N->hasOneUse() && N->getOperand(0).hasOneUse()) {
3967 SDValue N01 = N->getOperand(0).getOperand(1);
3969 if (ConstantSDNode *N01C = isConstOrConstSplat(N01)) {
3970 EVT TruncVT = N->getValueType(0);
3971 SDValue N00 = N->getOperand(0).getOperand(0);
3972 APInt TruncC = N01C->getAPIntValue();
3973 TruncC = TruncC.trunc(TruncVT.getScalarSizeInBits());
3975 return DAG.getNode(ISD::AND, SDLoc(N), TruncVT,
3976 DAG.getNode(ISD::TRUNCATE, SDLoc(N), TruncVT, N00),
3977 DAG.getConstant(TruncC, TruncVT));
3984 SDValue DAGCombiner::visitRotate(SDNode *N) {
3985 // fold (rot* x, (trunc (and y, c))) -> (rot* x, (and (trunc y), (trunc c))).
3986 if (N->getOperand(1).getOpcode() == ISD::TRUNCATE &&
3987 N->getOperand(1).getOperand(0).getOpcode() == ISD::AND) {
3988 SDValue NewOp1 = distributeTruncateThroughAnd(N->getOperand(1).getNode());
3989 if (NewOp1.getNode())
3990 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
3991 N->getOperand(0), NewOp1);
3996 SDValue DAGCombiner::visitSHL(SDNode *N) {
3997 SDValue N0 = N->getOperand(0);
3998 SDValue N1 = N->getOperand(1);
3999 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4000 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4001 EVT VT = N0.getValueType();
4002 unsigned OpSizeInBits = VT.getScalarSizeInBits();
4005 if (VT.isVector()) {
4006 SDValue FoldedVOp = SimplifyVBinOp(N);
4007 if (FoldedVOp.getNode()) return FoldedVOp;
4009 BuildVectorSDNode *N1CV = dyn_cast<BuildVectorSDNode>(N1);
4010 // If setcc produces all-one true value then:
4011 // (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<<N1CV)
4012 if (N1CV && N1CV->isConstant()) {
4013 if (N0.getOpcode() == ISD::AND) {
4014 SDValue N00 = N0->getOperand(0);
4015 SDValue N01 = N0->getOperand(1);
4016 BuildVectorSDNode *N01CV = dyn_cast<BuildVectorSDNode>(N01);
4018 if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC &&
4019 TLI.getBooleanContents(N00.getOperand(0).getValueType()) ==
4020 TargetLowering::ZeroOrNegativeOneBooleanContent) {
4021 SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, VT, N01CV, N1CV);
4023 return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C);
4026 N1C = isConstOrConstSplat(N1);
4031 // fold (shl c1, c2) -> c1<<c2
4033 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
4034 // fold (shl 0, x) -> 0
4035 if (N0C && N0C->isNullValue())
4037 // fold (shl x, c >= size(x)) -> undef
4038 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4039 return DAG.getUNDEF(VT);
4040 // fold (shl x, 0) -> x
4041 if (N1C && N1C->isNullValue())
4043 // fold (shl undef, x) -> 0
4044 if (N0.getOpcode() == ISD::UNDEF)
4045 return DAG.getConstant(0, VT);
4046 // if (shl x, c) is known to be zero, return 0
4047 if (DAG.MaskedValueIsZero(SDValue(N, 0),
4048 APInt::getAllOnesValue(OpSizeInBits)))
4049 return DAG.getConstant(0, VT);
4050 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
4051 if (N1.getOpcode() == ISD::TRUNCATE &&
4052 N1.getOperand(0).getOpcode() == ISD::AND) {
4053 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4054 if (NewOp1.getNode())
4055 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, NewOp1);
4058 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4059 return SDValue(N, 0);
4061 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
4062 if (N1C && N0.getOpcode() == ISD::SHL) {
4063 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4064 uint64_t c1 = N0C1->getZExtValue();
4065 uint64_t c2 = N1C->getZExtValue();
4066 if (c1 + c2 >= OpSizeInBits)
4067 return DAG.getConstant(0, VT);
4068 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4069 DAG.getConstant(c1 + c2, N1.getValueType()));
4073 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
4074 // For this to be valid, the second form must not preserve any of the bits
4075 // that are shifted out by the inner shift in the first form. This means
4076 // the outer shift size must be >= the number of bits added by the ext.
4077 // As a corollary, we don't care what kind of ext it is.
4078 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
4079 N0.getOpcode() == ISD::ANY_EXTEND ||
4080 N0.getOpcode() == ISD::SIGN_EXTEND) &&
4081 N0.getOperand(0).getOpcode() == ISD::SHL) {
4082 SDValue N0Op0 = N0.getOperand(0);
4083 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4084 uint64_t c1 = N0Op0C1->getZExtValue();
4085 uint64_t c2 = N1C->getZExtValue();
4086 EVT InnerShiftVT = N0Op0.getValueType();
4087 uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits();
4088 if (c2 >= OpSizeInBits - InnerShiftSize) {
4089 if (c1 + c2 >= OpSizeInBits)
4090 return DAG.getConstant(0, VT);
4091 return DAG.getNode(ISD::SHL, SDLoc(N0), VT,
4092 DAG.getNode(N0.getOpcode(), SDLoc(N0), VT,
4093 N0Op0->getOperand(0)),
4094 DAG.getConstant(c1 + c2, N1.getValueType()));
4099 // fold (shl (zext (srl x, C)), C) -> (zext (shl (srl x, C), C))
4100 // Only fold this if the inner zext has no other uses to avoid increasing
4101 // the total number of instructions.
4102 if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() &&
4103 N0.getOperand(0).getOpcode() == ISD::SRL) {
4104 SDValue N0Op0 = N0.getOperand(0);
4105 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4106 uint64_t c1 = N0Op0C1->getZExtValue();
4107 if (c1 < VT.getScalarSizeInBits()) {
4108 uint64_t c2 = N1C->getZExtValue();
4110 SDValue NewOp0 = N0.getOperand(0);
4111 EVT CountVT = NewOp0.getOperand(1).getValueType();
4112 SDValue NewSHL = DAG.getNode(ISD::SHL, SDLoc(N), NewOp0.getValueType(),
4113 NewOp0, DAG.getConstant(c2, CountVT));
4114 AddToWorklist(NewSHL.getNode());
4115 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
4121 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
4122 // (and (srl x, (sub c1, c2), MASK)
4123 // Only fold this if the inner shift has no other uses -- if it does, folding
4124 // this will increase the total number of instructions.
4125 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
4126 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4127 uint64_t c1 = N0C1->getZExtValue();
4128 if (c1 < OpSizeInBits) {
4129 uint64_t c2 = N1C->getZExtValue();
4130 APInt Mask = APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - c1);
4133 Mask = Mask.shl(c2 - c1);
4134 Shift = DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4135 DAG.getConstant(c2 - c1, N1.getValueType()));
4137 Mask = Mask.lshr(c1 - c2);
4138 Shift = DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4139 DAG.getConstant(c1 - c2, N1.getValueType()));
4141 return DAG.getNode(ISD::AND, SDLoc(N0), VT, Shift,
4142 DAG.getConstant(Mask, VT));
4146 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
4147 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
4148 unsigned BitSize = VT.getScalarSizeInBits();
4149 SDValue HiBitsMask =
4150 DAG.getConstant(APInt::getHighBitsSet(BitSize,
4151 BitSize - N1C->getZExtValue()), VT);
4152 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4156 // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
4157 // Variant of version done on multiply, except mul by a power of 2 is turned
4160 if (N1C && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
4161 (isa<ConstantSDNode>(N0.getOperand(1)) ||
4162 isConstantSplatVector(N0.getOperand(1).getNode(), Val))) {
4163 SDValue Shl0 = DAG.getNode(ISD::SHL, SDLoc(N0), VT, N0.getOperand(0), N1);
4164 SDValue Shl1 = DAG.getNode(ISD::SHL, SDLoc(N1), VT, N0.getOperand(1), N1);
4165 return DAG.getNode(ISD::ADD, SDLoc(N), VT, Shl0, Shl1);
4169 SDValue NewSHL = visitShiftByConstant(N, N1C);
4170 if (NewSHL.getNode())
4177 SDValue DAGCombiner::visitSRA(SDNode *N) {
4178 SDValue N0 = N->getOperand(0);
4179 SDValue N1 = N->getOperand(1);
4180 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4181 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4182 EVT VT = N0.getValueType();
4183 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4186 if (VT.isVector()) {
4187 SDValue FoldedVOp = SimplifyVBinOp(N);
4188 if (FoldedVOp.getNode()) return FoldedVOp;
4190 N1C = isConstOrConstSplat(N1);
4193 // fold (sra c1, c2) -> (sra c1, c2)
4195 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
4196 // fold (sra 0, x) -> 0
4197 if (N0C && N0C->isNullValue())
4199 // fold (sra -1, x) -> -1
4200 if (N0C && N0C->isAllOnesValue())
4202 // fold (sra x, (setge c, size(x))) -> undef
4203 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4204 return DAG.getUNDEF(VT);
4205 // fold (sra x, 0) -> x
4206 if (N1C && N1C->isNullValue())
4208 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
4210 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
4211 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
4212 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
4214 ExtVT = EVT::getVectorVT(*DAG.getContext(),
4215 ExtVT, VT.getVectorNumElements());
4216 if ((!LegalOperations ||
4217 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
4218 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
4219 N0.getOperand(0), DAG.getValueType(ExtVT));
4222 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
4223 if (N1C && N0.getOpcode() == ISD::SRA) {
4224 if (ConstantSDNode *C1 = isConstOrConstSplat(N0.getOperand(1))) {
4225 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
4226 if (Sum >= OpSizeInBits)
4227 Sum = OpSizeInBits - 1;
4228 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0.getOperand(0),
4229 DAG.getConstant(Sum, N1.getValueType()));
4233 // fold (sra (shl X, m), (sub result_size, n))
4234 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
4235 // result_size - n != m.
4236 // If truncate is free for the target sext(shl) is likely to result in better
4238 if (N0.getOpcode() == ISD::SHL && N1C) {
4239 // Get the two constanst of the shifts, CN0 = m, CN = n.
4240 const ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1));
4242 LLVMContext &Ctx = *DAG.getContext();
4243 // Determine what the truncate's result bitsize and type would be.
4244 EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - N1C->getZExtValue());
4247 TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorNumElements());
4249 // Determine the residual right-shift amount.
4250 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
4252 // If the shift is not a no-op (in which case this should be just a sign
4253 // extend already), the truncated to type is legal, sign_extend is legal
4254 // on that type, and the truncate to that type is both legal and free,
4255 // perform the transform.
4256 if ((ShiftAmt > 0) &&
4257 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
4258 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
4259 TLI.isTruncateFree(VT, TruncVT)) {
4261 SDValue Amt = DAG.getConstant(ShiftAmt,
4262 getShiftAmountTy(N0.getOperand(0).getValueType()));
4263 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0), VT,
4264 N0.getOperand(0), Amt);
4265 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), TruncVT,
4267 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N),
4268 N->getValueType(0), Trunc);
4273 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
4274 if (N1.getOpcode() == ISD::TRUNCATE &&
4275 N1.getOperand(0).getOpcode() == ISD::AND) {
4276 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4277 if (NewOp1.getNode())
4278 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, NewOp1);
4281 // fold (sra (trunc (srl x, c1)), c2) -> (trunc (sra x, c1 + c2))
4282 // if c1 is equal to the number of bits the trunc removes
4283 if (N0.getOpcode() == ISD::TRUNCATE &&
4284 (N0.getOperand(0).getOpcode() == ISD::SRL ||
4285 N0.getOperand(0).getOpcode() == ISD::SRA) &&
4286 N0.getOperand(0).hasOneUse() &&
4287 N0.getOperand(0).getOperand(1).hasOneUse() &&
4289 SDValue N0Op0 = N0.getOperand(0);
4290 if (ConstantSDNode *LargeShift = isConstOrConstSplat(N0Op0.getOperand(1))) {
4291 unsigned LargeShiftVal = LargeShift->getZExtValue();
4292 EVT LargeVT = N0Op0.getValueType();
4294 if (LargeVT.getScalarSizeInBits() - OpSizeInBits == LargeShiftVal) {
4296 DAG.getConstant(LargeShiftVal + N1C->getZExtValue(),
4297 getShiftAmountTy(N0Op0.getOperand(0).getValueType()));
4298 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), LargeVT,
4299 N0Op0.getOperand(0), Amt);
4300 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, SRA);
4305 // Simplify, based on bits shifted out of the LHS.
4306 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4307 return SDValue(N, 0);
4310 // If the sign bit is known to be zero, switch this to a SRL.
4311 if (DAG.SignBitIsZero(N0))
4312 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1);
4315 SDValue NewSRA = visitShiftByConstant(N, N1C);
4316 if (NewSRA.getNode())
4323 SDValue DAGCombiner::visitSRL(SDNode *N) {
4324 SDValue N0 = N->getOperand(0);
4325 SDValue N1 = N->getOperand(1);
4326 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4327 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4328 EVT VT = N0.getValueType();
4329 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4332 if (VT.isVector()) {
4333 SDValue FoldedVOp = SimplifyVBinOp(N);
4334 if (FoldedVOp.getNode()) return FoldedVOp;
4336 N1C = isConstOrConstSplat(N1);
4339 // fold (srl c1, c2) -> c1 >>u c2
4341 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
4342 // fold (srl 0, x) -> 0
4343 if (N0C && N0C->isNullValue())
4345 // fold (srl x, c >= size(x)) -> undef
4346 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4347 return DAG.getUNDEF(VT);
4348 // fold (srl x, 0) -> x
4349 if (N1C && N1C->isNullValue())
4351 // if (srl x, c) is known to be zero, return 0
4352 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
4353 APInt::getAllOnesValue(OpSizeInBits)))
4354 return DAG.getConstant(0, VT);
4356 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
4357 if (N1C && N0.getOpcode() == ISD::SRL) {
4358 if (ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1))) {
4359 uint64_t c1 = N01C->getZExtValue();
4360 uint64_t c2 = N1C->getZExtValue();
4361 if (c1 + c2 >= OpSizeInBits)
4362 return DAG.getConstant(0, VT);
4363 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4364 DAG.getConstant(c1 + c2, N1.getValueType()));
4368 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
4369 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
4370 N0.getOperand(0).getOpcode() == ISD::SRL &&
4371 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
4373 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
4374 uint64_t c2 = N1C->getZExtValue();
4375 EVT InnerShiftVT = N0.getOperand(0).getValueType();
4376 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
4377 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
4378 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
4379 if (c1 + OpSizeInBits == InnerShiftSize) {
4380 if (c1 + c2 >= InnerShiftSize)
4381 return DAG.getConstant(0, VT);
4382 return DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT,
4383 DAG.getNode(ISD::SRL, SDLoc(N0), InnerShiftVT,
4384 N0.getOperand(0)->getOperand(0),
4385 DAG.getConstant(c1 + c2, ShiftCountVT)));
4389 // fold (srl (shl x, c), c) -> (and x, cst2)
4390 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1) {
4391 unsigned BitSize = N0.getScalarValueSizeInBits();
4392 if (BitSize <= 64) {
4393 uint64_t ShAmt = N1C->getZExtValue() + 64 - BitSize;
4394 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4395 DAG.getConstant(~0ULL >> ShAmt, VT));
4399 // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask)
4400 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
4401 // Shifting in all undef bits?
4402 EVT SmallVT = N0.getOperand(0).getValueType();
4403 unsigned BitSize = SmallVT.getScalarSizeInBits();
4404 if (N1C->getZExtValue() >= BitSize)
4405 return DAG.getUNDEF(VT);
4407 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
4408 uint64_t ShiftAmt = N1C->getZExtValue();
4409 SDValue SmallShift = DAG.getNode(ISD::SRL, SDLoc(N0), SmallVT,
4411 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
4412 AddToWorklist(SmallShift.getNode());
4413 APInt Mask = APInt::getAllOnesValue(OpSizeInBits).lshr(ShiftAmt);
4414 return DAG.getNode(ISD::AND, SDLoc(N), VT,
4415 DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, SmallShift),
4416 DAG.getConstant(Mask, VT));
4420 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
4421 // bit, which is unmodified by sra.
4422 if (N1C && N1C->getZExtValue() + 1 == OpSizeInBits) {
4423 if (N0.getOpcode() == ISD::SRA)
4424 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1);
4427 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
4428 if (N1C && N0.getOpcode() == ISD::CTLZ &&
4429 N1C->getAPIntValue() == Log2_32(OpSizeInBits)) {
4430 APInt KnownZero, KnownOne;
4431 DAG.computeKnownBits(N0.getOperand(0), KnownZero, KnownOne);
4433 // If any of the input bits are KnownOne, then the input couldn't be all
4434 // zeros, thus the result of the srl will always be zero.
4435 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
4437 // If all of the bits input the to ctlz node are known to be zero, then
4438 // the result of the ctlz is "32" and the result of the shift is one.
4439 APInt UnknownBits = ~KnownZero;
4440 if (UnknownBits == 0) return DAG.getConstant(1, VT);
4442 // Otherwise, check to see if there is exactly one bit input to the ctlz.
4443 if ((UnknownBits & (UnknownBits - 1)) == 0) {
4444 // Okay, we know that only that the single bit specified by UnknownBits
4445 // could be set on input to the CTLZ node. If this bit is set, the SRL
4446 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
4447 // to an SRL/XOR pair, which is likely to simplify more.
4448 unsigned ShAmt = UnknownBits.countTrailingZeros();
4449 SDValue Op = N0.getOperand(0);
4452 Op = DAG.getNode(ISD::SRL, SDLoc(N0), VT, Op,
4453 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
4454 AddToWorklist(Op.getNode());
4457 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
4458 Op, DAG.getConstant(1, VT));
4462 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
4463 if (N1.getOpcode() == ISD::TRUNCATE &&
4464 N1.getOperand(0).getOpcode() == ISD::AND) {
4465 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4466 if (NewOp1.getNode())
4467 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, NewOp1);
4470 // fold operands of srl based on knowledge that the low bits are not
4472 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4473 return SDValue(N, 0);
4476 SDValue NewSRL = visitShiftByConstant(N, N1C);
4477 if (NewSRL.getNode())
4481 // Attempt to convert a srl of a load into a narrower zero-extending load.
4482 SDValue NarrowLoad = ReduceLoadWidth(N);
4483 if (NarrowLoad.getNode())
4486 // Here is a common situation. We want to optimize:
4489 // %b = and i32 %a, 2
4490 // %c = srl i32 %b, 1
4491 // brcond i32 %c ...
4497 // %c = setcc eq %b, 0
4500 // However when after the source operand of SRL is optimized into AND, the SRL
4501 // itself may not be optimized further. Look for it and add the BRCOND into
4503 if (N->hasOneUse()) {
4504 SDNode *Use = *N->use_begin();
4505 if (Use->getOpcode() == ISD::BRCOND)
4507 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
4508 // Also look pass the truncate.
4509 Use = *Use->use_begin();
4510 if (Use->getOpcode() == ISD::BRCOND)
4518 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
4519 SDValue N0 = N->getOperand(0);
4520 EVT VT = N->getValueType(0);
4522 // fold (ctlz c1) -> c2
4523 if (isa<ConstantSDNode>(N0))
4524 return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0);
4528 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4529 SDValue N0 = N->getOperand(0);
4530 EVT VT = N->getValueType(0);
4532 // fold (ctlz_zero_undef c1) -> c2
4533 if (isa<ConstantSDNode>(N0))
4534 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4538 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4539 SDValue N0 = N->getOperand(0);
4540 EVT VT = N->getValueType(0);
4542 // fold (cttz c1) -> c2
4543 if (isa<ConstantSDNode>(N0))
4544 return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0);
4548 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4549 SDValue N0 = N->getOperand(0);
4550 EVT VT = N->getValueType(0);
4552 // fold (cttz_zero_undef c1) -> c2
4553 if (isa<ConstantSDNode>(N0))
4554 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4558 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4559 SDValue N0 = N->getOperand(0);
4560 EVT VT = N->getValueType(0);
4562 // fold (ctpop c1) -> c2
4563 if (isa<ConstantSDNode>(N0))
4564 return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0);
4568 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4569 SDValue N0 = N->getOperand(0);
4570 SDValue N1 = N->getOperand(1);
4571 SDValue N2 = N->getOperand(2);
4572 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4573 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4574 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
4575 EVT VT = N->getValueType(0);
4576 EVT VT0 = N0.getValueType();
4578 // fold (select C, X, X) -> X
4581 // fold (select true, X, Y) -> X
4582 if (N0C && !N0C->isNullValue())
4584 // fold (select false, X, Y) -> Y
4585 if (N0C && N0C->isNullValue())
4587 // fold (select C, 1, X) -> (or C, X)
4588 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4589 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4590 // fold (select C, 0, 1) -> (xor C, 1)
4591 // We can't do this reliably if integer based booleans have different contents
4592 // to floating point based booleans. This is because we can't tell whether we
4593 // have an integer-based boolean or a floating-point-based boolean unless we
4594 // can find the SETCC that produced it and inspect its operands. This is
4595 // fairly easy if C is the SETCC node, but it can potentially be
4596 // undiscoverable (or not reasonably discoverable). For example, it could be
4597 // in another basic block or it could require searching a complicated
4599 if (VT.isInteger() &&
4600 (VT0 == MVT::i1 || (VT0.isInteger() &&
4601 TLI.getBooleanContents(false, false) ==
4602 TLI.getBooleanContents(false, true) &&
4603 TLI.getBooleanContents(false, false) ==
4604 TargetLowering::ZeroOrOneBooleanContent)) &&
4605 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4608 return DAG.getNode(ISD::XOR, SDLoc(N), VT0,
4609 N0, DAG.getConstant(1, VT0));
4610 XORNode = DAG.getNode(ISD::XOR, SDLoc(N0), VT0,
4611 N0, DAG.getConstant(1, VT0));
4612 AddToWorklist(XORNode.getNode());
4614 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, XORNode);
4615 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, XORNode);
4617 // fold (select C, 0, X) -> (and (not C), X)
4618 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4619 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4620 AddToWorklist(NOTNode.getNode());
4621 return DAG.getNode(ISD::AND, SDLoc(N), VT, NOTNode, N2);
4623 // fold (select C, X, 1) -> (or (not C), X)
4624 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4625 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4626 AddToWorklist(NOTNode.getNode());
4627 return DAG.getNode(ISD::OR, SDLoc(N), VT, NOTNode, N1);
4629 // fold (select C, X, 0) -> (and C, X)
4630 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4631 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4632 // fold (select X, X, Y) -> (or X, Y)
4633 // fold (select X, 1, Y) -> (or X, Y)
4634 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4635 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4636 // fold (select X, Y, X) -> (and X, Y)
4637 // fold (select X, Y, 0) -> (and X, Y)
4638 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4639 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4641 // If we can fold this based on the true/false value, do so.
4642 if (SimplifySelectOps(N, N1, N2))
4643 return SDValue(N, 0); // Don't revisit N.
4645 // fold selects based on a setcc into other things, such as min/max/abs
4646 if (N0.getOpcode() == ISD::SETCC) {
4647 if ((!LegalOperations &&
4648 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT)) ||
4649 TLI.isOperationLegal(ISD::SELECT_CC, VT))
4650 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT,
4651 N0.getOperand(0), N0.getOperand(1),
4652 N1, N2, N0.getOperand(2));
4653 return SimplifySelect(SDLoc(N), N0, N1, N2);
4660 std::pair<SDValue, SDValue> SplitVSETCC(const SDNode *N, SelectionDAG &DAG) {
4663 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
4665 // Split the inputs.
4666 SDValue Lo, Hi, LL, LH, RL, RH;
4667 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
4668 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
4670 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));
4671 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));
4673 return std::make_pair(Lo, Hi);
4676 // This function assumes all the vselect's arguments are CONCAT_VECTOR
4677 // nodes and that the condition is a BV of ConstantSDNodes (or undefs).
4678 static SDValue ConvertSelectToConcatVector(SDNode *N, SelectionDAG &DAG) {
4680 SDValue Cond = N->getOperand(0);
4681 SDValue LHS = N->getOperand(1);
4682 SDValue RHS = N->getOperand(2);
4683 EVT VT = N->getValueType(0);
4684 int NumElems = VT.getVectorNumElements();
4685 assert(LHS.getOpcode() == ISD::CONCAT_VECTORS &&
4686 RHS.getOpcode() == ISD::CONCAT_VECTORS &&
4687 Cond.getOpcode() == ISD::BUILD_VECTOR);
4689 // CONCAT_VECTOR can take an arbitrary number of arguments. We only care about
4690 // binary ones here.
4691 if (LHS->getNumOperands() != 2 || RHS->getNumOperands() != 2)
4694 // We're sure we have an even number of elements due to the
4695 // concat_vectors we have as arguments to vselect.
4696 // Skip BV elements until we find one that's not an UNDEF
4697 // After we find an UNDEF element, keep looping until we get to half the
4698 // length of the BV and see if all the non-undef nodes are the same.
4699 ConstantSDNode *BottomHalf = nullptr;
4700 for (int i = 0; i < NumElems / 2; ++i) {
4701 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4704 if (BottomHalf == nullptr)
4705 BottomHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4706 else if (Cond->getOperand(i).getNode() != BottomHalf)
4710 // Do the same for the second half of the BuildVector
4711 ConstantSDNode *TopHalf = nullptr;
4712 for (int i = NumElems / 2; i < NumElems; ++i) {
4713 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4716 if (TopHalf == nullptr)
4717 TopHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4718 else if (Cond->getOperand(i).getNode() != TopHalf)
4722 assert(TopHalf && BottomHalf &&
4723 "One half of the selector was all UNDEFs and the other was all the "
4724 "same value. This should have been addressed before this function.");
4726 ISD::CONCAT_VECTORS, dl, VT,
4727 BottomHalf->isNullValue() ? RHS->getOperand(0) : LHS->getOperand(0),
4728 TopHalf->isNullValue() ? RHS->getOperand(1) : LHS->getOperand(1));
4731 SDValue DAGCombiner::visitVSELECT(SDNode *N) {
4732 SDValue N0 = N->getOperand(0);
4733 SDValue N1 = N->getOperand(1);
4734 SDValue N2 = N->getOperand(2);
4737 // Canonicalize integer abs.
4738 // vselect (setg[te] X, 0), X, -X ->
4739 // vselect (setgt X, -1), X, -X ->
4740 // vselect (setl[te] X, 0), -X, X ->
4741 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
4742 if (N0.getOpcode() == ISD::SETCC) {
4743 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
4744 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
4746 bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
4748 if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
4749 (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) &&
4750 N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1))
4751 isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode());
4752 else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) &&
4753 N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1))
4754 isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode());
4757 EVT VT = LHS.getValueType();
4758 SDValue Shift = DAG.getNode(
4759 ISD::SRA, DL, VT, LHS,
4760 DAG.getConstant(VT.getScalarType().getSizeInBits() - 1, VT));
4761 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift);
4762 AddToWorklist(Shift.getNode());
4763 AddToWorklist(Add.getNode());
4764 return DAG.getNode(ISD::XOR, DL, VT, Add, Shift);
4768 // If the VSELECT result requires splitting and the mask is provided by a
4769 // SETCC, then split both nodes and its operands before legalization. This
4770 // prevents the type legalizer from unrolling SETCC into scalar comparisons
4771 // and enables future optimizations (e.g. min/max pattern matching on X86).
4772 if (N0.getOpcode() == ISD::SETCC) {
4773 EVT VT = N->getValueType(0);
4775 // Check if any splitting is required.
4776 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
4777 TargetLowering::TypeSplitVector)
4780 SDValue Lo, Hi, CCLo, CCHi, LL, LH, RL, RH;
4781 std::tie(CCLo, CCHi) = SplitVSETCC(N0.getNode(), DAG);
4782 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 1);
4783 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 2);
4785 Lo = DAG.getNode(N->getOpcode(), DL, LL.getValueType(), CCLo, LL, RL);
4786 Hi = DAG.getNode(N->getOpcode(), DL, LH.getValueType(), CCHi, LH, RH);
4788 // Add the new VSELECT nodes to the work list in case they need to be split
4790 AddToWorklist(Lo.getNode());
4791 AddToWorklist(Hi.getNode());
4793 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
4796 // Fold (vselect (build_vector all_ones), N1, N2) -> N1
4797 if (ISD::isBuildVectorAllOnes(N0.getNode()))
4799 // Fold (vselect (build_vector all_zeros), N1, N2) -> N2
4800 if (ISD::isBuildVectorAllZeros(N0.getNode()))
4803 // The ConvertSelectToConcatVector function is assuming both the above
4804 // checks for (vselect (build_vector all{ones,zeros) ...) have been made
4806 if (N1.getOpcode() == ISD::CONCAT_VECTORS &&
4807 N2.getOpcode() == ISD::CONCAT_VECTORS &&
4808 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
4809 SDValue CV = ConvertSelectToConcatVector(N, DAG);
4817 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
4818 SDValue N0 = N->getOperand(0);
4819 SDValue N1 = N->getOperand(1);
4820 SDValue N2 = N->getOperand(2);
4821 SDValue N3 = N->getOperand(3);
4822 SDValue N4 = N->getOperand(4);
4823 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
4825 // fold select_cc lhs, rhs, x, x, cc -> x
4829 // Determine if the condition we're dealing with is constant
4830 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
4831 N0, N1, CC, SDLoc(N), false);
4832 if (SCC.getNode()) {
4833 AddToWorklist(SCC.getNode());
4835 if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) {
4836 if (!SCCC->isNullValue())
4837 return N2; // cond always true -> true val
4839 return N3; // cond always false -> false val
4842 // Fold to a simpler select_cc
4843 if (SCC.getOpcode() == ISD::SETCC)
4844 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(),
4845 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
4849 // If we can fold this based on the true/false value, do so.
4850 if (SimplifySelectOps(N, N2, N3))
4851 return SDValue(N, 0); // Don't revisit N.
4853 // fold select_cc into other things, such as min/max/abs
4854 return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC);
4857 SDValue DAGCombiner::visitSETCC(SDNode *N) {
4858 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
4859 cast<CondCodeSDNode>(N->getOperand(2))->get(),
4863 // tryToFoldExtendOfConstant - Try to fold a sext/zext/aext
4864 // dag node into a ConstantSDNode or a build_vector of constants.
4865 // This function is called by the DAGCombiner when visiting sext/zext/aext
4866 // dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND).
4867 // Vector extends are not folded if operations are legal; this is to
4868 // avoid introducing illegal build_vector dag nodes.
4869 static SDNode *tryToFoldExtendOfConstant(SDNode *N, const TargetLowering &TLI,
4870 SelectionDAG &DAG, bool LegalTypes,
4871 bool LegalOperations) {
4872 unsigned Opcode = N->getOpcode();
4873 SDValue N0 = N->getOperand(0);
4874 EVT VT = N->getValueType(0);
4876 assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND ||
4877 Opcode == ISD::ANY_EXTEND) && "Expected EXTEND dag node in input!");
4879 // fold (sext c1) -> c1
4880 // fold (zext c1) -> c1
4881 // fold (aext c1) -> c1
4882 if (isa<ConstantSDNode>(N0))
4883 return DAG.getNode(Opcode, SDLoc(N), VT, N0).getNode();
4885 // fold (sext (build_vector AllConstants) -> (build_vector AllConstants)
4886 // fold (zext (build_vector AllConstants) -> (build_vector AllConstants)
4887 // fold (aext (build_vector AllConstants) -> (build_vector AllConstants)
4888 EVT SVT = VT.getScalarType();
4889 if (!(VT.isVector() &&
4890 (!LegalTypes || (!LegalOperations && TLI.isTypeLegal(SVT))) &&
4891 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())))
4894 // We can fold this node into a build_vector.
4895 unsigned VTBits = SVT.getSizeInBits();
4896 unsigned EVTBits = N0->getValueType(0).getScalarType().getSizeInBits();
4897 unsigned ShAmt = VTBits - EVTBits;
4898 SmallVector<SDValue, 8> Elts;
4899 unsigned NumElts = N0->getNumOperands();
4902 for (unsigned i=0; i != NumElts; ++i) {
4903 SDValue Op = N0->getOperand(i);
4904 if (Op->getOpcode() == ISD::UNDEF) {
4905 Elts.push_back(DAG.getUNDEF(SVT));
4909 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
4910 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
4911 if (Opcode == ISD::SIGN_EXTEND)
4912 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
4915 Elts.push_back(DAG.getConstant(C.shl(ShAmt).lshr(ShAmt).getZExtValue(),
4919 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Elts).getNode();
4922 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
4923 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
4924 // transformation. Returns true if extension are possible and the above
4925 // mentioned transformation is profitable.
4926 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
4928 SmallVectorImpl<SDNode *> &ExtendNodes,
4929 const TargetLowering &TLI) {
4930 bool HasCopyToRegUses = false;
4931 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
4932 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
4933 UE = N0.getNode()->use_end();
4938 if (UI.getUse().getResNo() != N0.getResNo())
4940 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
4941 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
4942 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
4943 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
4944 // Sign bits will be lost after a zext.
4947 for (unsigned i = 0; i != 2; ++i) {
4948 SDValue UseOp = User->getOperand(i);
4951 if (!isa<ConstantSDNode>(UseOp))
4956 ExtendNodes.push_back(User);
4959 // If truncates aren't free and there are users we can't
4960 // extend, it isn't worthwhile.
4963 // Remember if this value is live-out.
4964 if (User->getOpcode() == ISD::CopyToReg)
4965 HasCopyToRegUses = true;
4968 if (HasCopyToRegUses) {
4969 bool BothLiveOut = false;
4970 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
4972 SDUse &Use = UI.getUse();
4973 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
4979 // Both unextended and extended values are live out. There had better be
4980 // a good reason for the transformation.
4981 return ExtendNodes.size();
4986 void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
4987 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
4988 ISD::NodeType ExtType) {
4989 // Extend SetCC uses if necessary.
4990 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
4991 SDNode *SetCC = SetCCs[i];
4992 SmallVector<SDValue, 4> Ops;
4994 for (unsigned j = 0; j != 2; ++j) {
4995 SDValue SOp = SetCC->getOperand(j);
4997 Ops.push_back(ExtLoad);
4999 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
5002 Ops.push_back(SetCC->getOperand(2));
5003 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), Ops));
5007 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
5008 SDValue N0 = N->getOperand(0);
5009 EVT VT = N->getValueType(0);
5011 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5013 return SDValue(Res, 0);
5015 // fold (sext (sext x)) -> (sext x)
5016 // fold (sext (aext x)) -> (sext x)
5017 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5018 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT,
5021 if (N0.getOpcode() == ISD::TRUNCATE) {
5022 // fold (sext (truncate (load x))) -> (sext (smaller load x))
5023 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
5024 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5025 if (NarrowLoad.getNode()) {
5026 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5027 if (NarrowLoad.getNode() != N0.getNode()) {
5028 CombineTo(N0.getNode(), NarrowLoad);
5029 // CombineTo deleted the truncate, if needed, but not what's under it.
5032 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5035 // See if the value being truncated is already sign extended. If so, just
5036 // eliminate the trunc/sext pair.
5037 SDValue Op = N0.getOperand(0);
5038 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
5039 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
5040 unsigned DestBits = VT.getScalarType().getSizeInBits();
5041 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
5043 if (OpBits == DestBits) {
5044 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
5045 // bits, it is already ready.
5046 if (NumSignBits > DestBits-MidBits)
5048 } else if (OpBits < DestBits) {
5049 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
5050 // bits, just sext from i32.
5051 if (NumSignBits > OpBits-MidBits)
5052 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, Op);
5054 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
5055 // bits, just truncate to i32.
5056 if (NumSignBits > OpBits-MidBits)
5057 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5060 // fold (sext (truncate x)) -> (sextinreg x).
5061 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
5062 N0.getValueType())) {
5063 if (OpBits < DestBits)
5064 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op);
5065 else if (OpBits > DestBits)
5066 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op);
5067 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, Op,
5068 DAG.getValueType(N0.getValueType()));
5072 // fold (sext (load x)) -> (sext (truncate (sextload x)))
5073 // None of the supported targets knows how to perform load and sign extend
5074 // on vectors in one instruction. We only perform this transformation on
5076 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5077 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5078 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5079 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) {
5080 bool DoXform = true;
5081 SmallVector<SDNode*, 4> SetCCs;
5082 if (!N0.hasOneUse())
5083 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
5085 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5086 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5088 LN0->getBasePtr(), N0.getValueType(),
5089 LN0->getMemOperand());
5090 CombineTo(N, ExtLoad);
5091 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5092 N0.getValueType(), ExtLoad);
5093 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5094 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5096 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5100 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
5101 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
5102 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5103 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5104 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5105 EVT MemVT = LN0->getMemoryVT();
5106 if ((!LegalOperations && !LN0->isVolatile()) ||
5107 TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) {
5108 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5110 LN0->getBasePtr(), MemVT,
5111 LN0->getMemOperand());
5112 CombineTo(N, ExtLoad);
5113 CombineTo(N0.getNode(),
5114 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5115 N0.getValueType(), ExtLoad),
5116 ExtLoad.getValue(1));
5117 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5121 // fold (sext (and/or/xor (load x), cst)) ->
5122 // (and/or/xor (sextload x), (sext cst))
5123 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5124 N0.getOpcode() == ISD::XOR) &&
5125 isa<LoadSDNode>(N0.getOperand(0)) &&
5126 N0.getOperand(1).getOpcode() == ISD::Constant &&
5127 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()) &&
5128 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5129 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5130 if (LN0->getExtensionType() != ISD::ZEXTLOAD && LN0->isUnindexed()) {
5131 bool DoXform = true;
5132 SmallVector<SDNode*, 4> SetCCs;
5133 if (!N0.hasOneUse())
5134 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
5137 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN0), VT,
5138 LN0->getChain(), LN0->getBasePtr(),
5140 LN0->getMemOperand());
5141 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5142 Mask = Mask.sext(VT.getSizeInBits());
5143 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5144 ExtLoad, DAG.getConstant(Mask, VT));
5145 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5146 SDLoc(N0.getOperand(0)),
5147 N0.getOperand(0).getValueType(), ExtLoad);
5149 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5150 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5152 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5157 if (N0.getOpcode() == ISD::SETCC) {
5158 EVT N0VT = N0.getOperand(0).getValueType();
5159 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
5160 // Only do this before legalize for now.
5161 if (VT.isVector() && !LegalOperations &&
5162 TLI.getBooleanContents(N0VT) ==
5163 TargetLowering::ZeroOrNegativeOneBooleanContent) {
5164 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
5165 // of the same size as the compared operands. Only optimize sext(setcc())
5166 // if this is the case.
5167 EVT SVT = getSetCCResultType(N0VT);
5169 // We know that the # elements of the results is the same as the
5170 // # elements of the compare (and the # elements of the compare result
5171 // for that matter). Check to see that they are the same size. If so,
5172 // we know that the element size of the sext'd result matches the
5173 // element size of the compare operands.
5174 if (VT.getSizeInBits() == SVT.getSizeInBits())
5175 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5177 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5179 // If the desired elements are smaller or larger than the source
5180 // elements we can use a matching integer vector type and then
5181 // truncate/sign extend
5182 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5183 if (SVT == MatchingVectorType) {
5184 SDValue VsetCC = DAG.getSetCC(SDLoc(N), MatchingVectorType,
5185 N0.getOperand(0), N0.getOperand(1),
5186 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5187 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
5191 // sext(setcc x, y, cc) -> (select (setcc x, y, cc), -1, 0)
5192 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
5194 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
5196 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5197 NegOne, DAG.getConstant(0, VT),
5198 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5199 if (SCC.getNode()) return SCC;
5201 if (!VT.isVector()) {
5202 EVT SetCCVT = getSetCCResultType(N0.getOperand(0).getValueType());
5203 if (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, SetCCVT)) {
5205 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5206 SDValue SetCC = DAG.getSetCC(DL,
5208 N0.getOperand(0), N0.getOperand(1), CC);
5209 EVT SelectVT = getSetCCResultType(VT);
5210 return DAG.getSelect(DL, VT,
5211 DAG.getSExtOrTrunc(SetCC, DL, SelectVT),
5212 NegOne, DAG.getConstant(0, VT));
5218 // fold (sext x) -> (zext x) if the sign bit is known zero.
5219 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
5220 DAG.SignBitIsZero(N0))
5221 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0);
5226 // isTruncateOf - If N is a truncate of some other value, return true, record
5227 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
5228 // This function computes KnownZero to avoid a duplicated call to
5229 // computeKnownBits in the caller.
5230 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
5233 if (N->getOpcode() == ISD::TRUNCATE) {
5234 Op = N->getOperand(0);
5235 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5239 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
5240 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
5243 SDValue Op0 = N->getOperand(0);
5244 SDValue Op1 = N->getOperand(1);
5245 assert(Op0.getValueType() == Op1.getValueType());
5247 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
5248 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
5249 if (COp0 && COp0->isNullValue())
5251 else if (COp1 && COp1->isNullValue())
5256 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5258 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
5264 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
5265 SDValue N0 = N->getOperand(0);
5266 EVT VT = N->getValueType(0);
5268 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5270 return SDValue(Res, 0);
5272 // fold (zext (zext x)) -> (zext x)
5273 // fold (zext (aext x)) -> (zext x)
5274 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5275 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT,
5278 // fold (zext (truncate x)) -> (zext x) or
5279 // (zext (truncate x)) -> (truncate x)
5280 // This is valid when the truncated bits of x are already zero.
5281 // FIXME: We should extend this to work for vectors too.
5284 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
5285 APInt TruncatedBits =
5286 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
5287 APInt(Op.getValueSizeInBits(), 0) :
5288 APInt::getBitsSet(Op.getValueSizeInBits(),
5289 N0.getValueSizeInBits(),
5290 std::min(Op.getValueSizeInBits(),
5291 VT.getSizeInBits()));
5292 if (TruncatedBits == (KnownZero & TruncatedBits)) {
5293 if (VT.bitsGT(Op.getValueType()))
5294 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, Op);
5295 if (VT.bitsLT(Op.getValueType()))
5296 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5302 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5303 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
5304 if (N0.getOpcode() == ISD::TRUNCATE) {
5305 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5306 if (NarrowLoad.getNode()) {
5307 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5308 if (NarrowLoad.getNode() != N0.getNode()) {
5309 CombineTo(N0.getNode(), NarrowLoad);
5310 // CombineTo deleted the truncate, if needed, but not what's under it.
5313 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5317 // fold (zext (truncate x)) -> (and x, mask)
5318 if (N0.getOpcode() == ISD::TRUNCATE &&
5319 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
5321 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5322 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
5323 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5324 if (NarrowLoad.getNode()) {
5325 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5326 if (NarrowLoad.getNode() != N0.getNode()) {
5327 CombineTo(N0.getNode(), NarrowLoad);
5328 // CombineTo deleted the truncate, if needed, but not what's under it.
5331 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5334 SDValue Op = N0.getOperand(0);
5335 if (Op.getValueType().bitsLT(VT)) {
5336 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, Op);
5337 AddToWorklist(Op.getNode());
5338 } else if (Op.getValueType().bitsGT(VT)) {
5339 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5340 AddToWorklist(Op.getNode());
5342 return DAG.getZeroExtendInReg(Op, SDLoc(N),
5343 N0.getValueType().getScalarType());
5346 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
5347 // if either of the casts is not free.
5348 if (N0.getOpcode() == ISD::AND &&
5349 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5350 N0.getOperand(1).getOpcode() == ISD::Constant &&
5351 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5352 N0.getValueType()) ||
5353 !TLI.isZExtFree(N0.getValueType(), VT))) {
5354 SDValue X = N0.getOperand(0).getOperand(0);
5355 if (X.getValueType().bitsLT(VT)) {
5356 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(X), VT, X);
5357 } else if (X.getValueType().bitsGT(VT)) {
5358 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
5360 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5361 Mask = Mask.zext(VT.getSizeInBits());
5362 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5363 X, DAG.getConstant(Mask, VT));
5366 // fold (zext (load x)) -> (zext (truncate (zextload x)))
5367 // None of the supported targets knows how to perform load and vector_zext
5368 // on vectors in one instruction. We only perform this transformation on
5370 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5371 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5372 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5373 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
5374 bool DoXform = true;
5375 SmallVector<SDNode*, 4> SetCCs;
5376 if (!N0.hasOneUse())
5377 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
5379 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5380 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5382 LN0->getBasePtr(), N0.getValueType(),
5383 LN0->getMemOperand());
5384 CombineTo(N, ExtLoad);
5385 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5386 N0.getValueType(), ExtLoad);
5387 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5389 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5391 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5395 // fold (zext (and/or/xor (load x), cst)) ->
5396 // (and/or/xor (zextload x), (zext cst))
5397 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5398 N0.getOpcode() == ISD::XOR) &&
5399 isa<LoadSDNode>(N0.getOperand(0)) &&
5400 N0.getOperand(1).getOpcode() == ISD::Constant &&
5401 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()) &&
5402 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5403 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5404 if (LN0->getExtensionType() != ISD::SEXTLOAD && LN0->isUnindexed()) {
5405 bool DoXform = true;
5406 SmallVector<SDNode*, 4> SetCCs;
5407 if (!N0.hasOneUse())
5408 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
5411 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), VT,
5412 LN0->getChain(), LN0->getBasePtr(),
5414 LN0->getMemOperand());
5415 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5416 Mask = Mask.zext(VT.getSizeInBits());
5417 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5418 ExtLoad, DAG.getConstant(Mask, VT));
5419 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5420 SDLoc(N0.getOperand(0)),
5421 N0.getOperand(0).getValueType(), ExtLoad);
5423 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5424 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5426 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5431 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
5432 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
5433 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5434 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5435 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5436 EVT MemVT = LN0->getMemoryVT();
5437 if ((!LegalOperations && !LN0->isVolatile()) ||
5438 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) {
5439 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5441 LN0->getBasePtr(), MemVT,
5442 LN0->getMemOperand());
5443 CombineTo(N, ExtLoad);
5444 CombineTo(N0.getNode(),
5445 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(),
5447 ExtLoad.getValue(1));
5448 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5452 if (N0.getOpcode() == ISD::SETCC) {
5453 if (!LegalOperations && VT.isVector() &&
5454 N0.getValueType().getVectorElementType() == MVT::i1) {
5455 EVT N0VT = N0.getOperand(0).getValueType();
5456 if (getSetCCResultType(N0VT) == N0.getValueType())
5459 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
5460 // Only do this before legalize for now.
5461 EVT EltVT = VT.getVectorElementType();
5462 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
5463 DAG.getConstant(1, EltVT));
5464 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5465 // We know that the # elements of the results is the same as the
5466 // # elements of the compare (and the # elements of the compare result
5467 // for that matter). Check to see that they are the same size. If so,
5468 // we know that the element size of the sext'd result matches the
5469 // element size of the compare operands.
5470 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5471 DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5473 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
5474 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT,
5477 // If the desired elements are smaller or larger than the source
5478 // elements we can use a matching integer vector type and then
5479 // truncate/sign extend
5480 EVT MatchingElementType =
5481 EVT::getIntegerVT(*DAG.getContext(),
5482 N0VT.getScalarType().getSizeInBits());
5483 EVT MatchingVectorType =
5484 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
5485 N0VT.getVectorNumElements());
5487 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5489 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5490 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5491 DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT),
5492 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, OneOps));
5495 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5497 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5498 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5499 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5500 if (SCC.getNode()) return SCC;
5503 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
5504 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
5505 isa<ConstantSDNode>(N0.getOperand(1)) &&
5506 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
5508 SDValue ShAmt = N0.getOperand(1);
5509 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
5510 if (N0.getOpcode() == ISD::SHL) {
5511 SDValue InnerZExt = N0.getOperand(0);
5512 // If the original shl may be shifting out bits, do not perform this
5514 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
5515 InnerZExt.getOperand(0).getValueType().getSizeInBits();
5516 if (ShAmtVal > KnownZeroBits)
5522 // Ensure that the shift amount is wide enough for the shifted value.
5523 if (VT.getSizeInBits() >= 256)
5524 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
5526 return DAG.getNode(N0.getOpcode(), DL, VT,
5527 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
5534 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
5535 SDValue N0 = N->getOperand(0);
5536 EVT VT = N->getValueType(0);
5538 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5540 return SDValue(Res, 0);
5542 // fold (aext (aext x)) -> (aext x)
5543 // fold (aext (zext x)) -> (zext x)
5544 // fold (aext (sext x)) -> (sext x)
5545 if (N0.getOpcode() == ISD::ANY_EXTEND ||
5546 N0.getOpcode() == ISD::ZERO_EXTEND ||
5547 N0.getOpcode() == ISD::SIGN_EXTEND)
5548 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0));
5550 // fold (aext (truncate (load x))) -> (aext (smaller load x))
5551 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
5552 if (N0.getOpcode() == ISD::TRUNCATE) {
5553 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5554 if (NarrowLoad.getNode()) {
5555 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5556 if (NarrowLoad.getNode() != N0.getNode()) {
5557 CombineTo(N0.getNode(), NarrowLoad);
5558 // CombineTo deleted the truncate, if needed, but not what's under it.
5561 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5565 // fold (aext (truncate x))
5566 if (N0.getOpcode() == ISD::TRUNCATE) {
5567 SDValue TruncOp = N0.getOperand(0);
5568 if (TruncOp.getValueType() == VT)
5569 return TruncOp; // x iff x size == zext size.
5570 if (TruncOp.getValueType().bitsGT(VT))
5571 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, TruncOp);
5572 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, TruncOp);
5575 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
5576 // if the trunc is not free.
5577 if (N0.getOpcode() == ISD::AND &&
5578 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5579 N0.getOperand(1).getOpcode() == ISD::Constant &&
5580 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5581 N0.getValueType())) {
5582 SDValue X = N0.getOperand(0).getOperand(0);
5583 if (X.getValueType().bitsLT(VT)) {
5584 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, X);
5585 } else if (X.getValueType().bitsGT(VT)) {
5586 X = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, X);
5588 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5589 Mask = Mask.zext(VT.getSizeInBits());
5590 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5591 X, DAG.getConstant(Mask, VT));
5594 // fold (aext (load x)) -> (aext (truncate (extload x)))
5595 // None of the supported targets knows how to perform load and any_ext
5596 // on vectors in one instruction. We only perform this transformation on
5598 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5599 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5600 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType())) {
5601 bool DoXform = true;
5602 SmallVector<SDNode*, 4> SetCCs;
5603 if (!N0.hasOneUse())
5604 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
5606 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5607 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
5609 LN0->getBasePtr(), N0.getValueType(),
5610 LN0->getMemOperand());
5611 CombineTo(N, ExtLoad);
5612 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5613 N0.getValueType(), ExtLoad);
5614 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5615 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5617 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5621 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
5622 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
5623 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
5624 if (N0.getOpcode() == ISD::LOAD &&
5625 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5627 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5628 ISD::LoadExtType ExtType = LN0->getExtensionType();
5629 EVT MemVT = LN0->getMemoryVT();
5630 if (!LegalOperations || TLI.isLoadExtLegal(ExtType, MemVT)) {
5631 SDValue ExtLoad = DAG.getExtLoad(ExtType, SDLoc(N),
5632 VT, LN0->getChain(), LN0->getBasePtr(),
5633 MemVT, LN0->getMemOperand());
5634 CombineTo(N, ExtLoad);
5635 CombineTo(N0.getNode(),
5636 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5637 N0.getValueType(), ExtLoad),
5638 ExtLoad.getValue(1));
5639 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5643 if (N0.getOpcode() == ISD::SETCC) {
5645 // aext(setcc) -> vsetcc
5646 // aext(setcc) -> truncate(vsetcc)
5647 // aext(setcc) -> aext(vsetcc)
5648 // Only do this before legalize for now.
5649 if (VT.isVector() && !LegalOperations) {
5650 EVT N0VT = N0.getOperand(0).getValueType();
5651 // We know that the # elements of the results is the same as the
5652 // # elements of the compare (and the # elements of the compare result
5653 // for that matter). Check to see that they are the same size. If so,
5654 // we know that the element size of the sext'd result matches the
5655 // element size of the compare operands.
5656 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5657 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5659 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5660 // If the desired elements are smaller or larger than the source
5661 // elements we can use a matching integer vector type and then
5662 // truncate/any extend
5664 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5666 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5668 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5669 return DAG.getAnyExtOrTrunc(VsetCC, SDLoc(N), VT);
5673 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5675 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5676 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5677 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5685 /// See if the specified operand can be simplified with the knowledge that only
5686 /// the bits specified by Mask are used. If so, return the simpler operand,
5687 /// otherwise return a null SDValue.
5688 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
5689 switch (V.getOpcode()) {
5691 case ISD::Constant: {
5692 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
5693 assert(CV && "Const value should be ConstSDNode.");
5694 const APInt &CVal = CV->getAPIntValue();
5695 APInt NewVal = CVal & Mask;
5697 return DAG.getConstant(NewVal, V.getValueType());
5702 // If the LHS or RHS don't contribute bits to the or, drop them.
5703 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
5704 return V.getOperand(1);
5705 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
5706 return V.getOperand(0);
5709 // Only look at single-use SRLs.
5710 if (!V.getNode()->hasOneUse())
5712 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
5713 // See if we can recursively simplify the LHS.
5714 unsigned Amt = RHSC->getZExtValue();
5716 // Watch out for shift count overflow though.
5717 if (Amt >= Mask.getBitWidth()) break;
5718 APInt NewMask = Mask << Amt;
5719 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
5720 if (SimplifyLHS.getNode())
5721 return DAG.getNode(ISD::SRL, SDLoc(V), V.getValueType(),
5722 SimplifyLHS, V.getOperand(1));
5728 /// If the result of a wider load is shifted to right of N bits and then
5729 /// truncated to a narrower type and where N is a multiple of number of bits of
5730 /// the narrower type, transform it to a narrower load from address + N / num of
5731 /// bits of new type. If the result is to be extended, also fold the extension
5732 /// to form a extending load.
5733 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
5734 unsigned Opc = N->getOpcode();
5736 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
5737 SDValue N0 = N->getOperand(0);
5738 EVT VT = N->getValueType(0);
5741 // This transformation isn't valid for vector loads.
5745 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
5747 if (Opc == ISD::SIGN_EXTEND_INREG) {
5748 ExtType = ISD::SEXTLOAD;
5749 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
5750 } else if (Opc == ISD::SRL) {
5751 // Another special-case: SRL is basically zero-extending a narrower value.
5752 ExtType = ISD::ZEXTLOAD;
5754 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
5755 if (!N01) return SDValue();
5756 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
5757 VT.getSizeInBits() - N01->getZExtValue());
5759 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, ExtVT))
5762 unsigned EVTBits = ExtVT.getSizeInBits();
5764 // Do not generate loads of non-round integer types since these can
5765 // be expensive (and would be wrong if the type is not byte sized).
5766 if (!ExtVT.isRound())
5770 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
5771 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5772 ShAmt = N01->getZExtValue();
5773 // Is the shift amount a multiple of size of VT?
5774 if ((ShAmt & (EVTBits-1)) == 0) {
5775 N0 = N0.getOperand(0);
5776 // Is the load width a multiple of size of VT?
5777 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
5781 // At this point, we must have a load or else we can't do the transform.
5782 if (!isa<LoadSDNode>(N0)) return SDValue();
5784 // Because a SRL must be assumed to *need* to zero-extend the high bits
5785 // (as opposed to anyext the high bits), we can't combine the zextload
5786 // lowering of SRL and an sextload.
5787 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
5790 // If the shift amount is larger than the input type then we're not
5791 // accessing any of the loaded bytes. If the load was a zextload/extload
5792 // then the result of the shift+trunc is zero/undef (handled elsewhere).
5793 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
5798 // If the load is shifted left (and the result isn't shifted back right),
5799 // we can fold the truncate through the shift.
5800 unsigned ShLeftAmt = 0;
5801 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
5802 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
5803 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5804 ShLeftAmt = N01->getZExtValue();
5805 N0 = N0.getOperand(0);
5809 // If we haven't found a load, we can't narrow it. Don't transform one with
5810 // multiple uses, this would require adding a new load.
5811 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse())
5814 // Don't change the width of a volatile load.
5815 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5816 if (LN0->isVolatile())
5819 // Verify that we are actually reducing a load width here.
5820 if (LN0->getMemoryVT().getSizeInBits() < EVTBits)
5823 // For the transform to be legal, the load must produce only two values
5824 // (the value loaded and the chain). Don't transform a pre-increment
5825 // load, for example, which produces an extra value. Otherwise the
5826 // transformation is not equivalent, and the downstream logic to replace
5827 // uses gets things wrong.
5828 if (LN0->getNumValues() > 2)
5831 // If the load that we're shrinking is an extload and we're not just
5832 // discarding the extension we can't simply shrink the load. Bail.
5833 // TODO: It would be possible to merge the extensions in some cases.
5834 if (LN0->getExtensionType() != ISD::NON_EXTLOAD &&
5835 LN0->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits() + ShAmt)
5838 EVT PtrType = N0.getOperand(1).getValueType();
5840 if (PtrType == MVT::Untyped || PtrType.isExtended())
5841 // It's not possible to generate a constant of extended or untyped type.
5844 // For big endian targets, we need to adjust the offset to the pointer to
5845 // load the correct bytes.
5846 if (TLI.isBigEndian()) {
5847 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
5848 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
5849 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
5852 uint64_t PtrOff = ShAmt / 8;
5853 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
5854 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0),
5855 PtrType, LN0->getBasePtr(),
5856 DAG.getConstant(PtrOff, PtrType));
5857 AddToWorklist(NewPtr.getNode());
5860 if (ExtType == ISD::NON_EXTLOAD)
5861 Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr,
5862 LN0->getPointerInfo().getWithOffset(PtrOff),
5863 LN0->isVolatile(), LN0->isNonTemporal(),
5864 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
5866 Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(),NewPtr,
5867 LN0->getPointerInfo().getWithOffset(PtrOff),
5868 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
5869 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
5871 // Replace the old load's chain with the new load's chain.
5872 WorklistRemover DeadNodes(*this);
5873 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
5875 // Shift the result left, if we've swallowed a left shift.
5876 SDValue Result = Load;
5877 if (ShLeftAmt != 0) {
5878 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
5879 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
5881 // If the shift amount is as large as the result size (but, presumably,
5882 // no larger than the source) then the useful bits of the result are
5883 // zero; we can't simply return the shortened shift, because the result
5884 // of that operation is undefined.
5885 if (ShLeftAmt >= VT.getSizeInBits())
5886 Result = DAG.getConstant(0, VT);
5888 Result = DAG.getNode(ISD::SHL, SDLoc(N0), VT,
5889 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
5892 // Return the new loaded value.
5896 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
5897 SDValue N0 = N->getOperand(0);
5898 SDValue N1 = N->getOperand(1);
5899 EVT VT = N->getValueType(0);
5900 EVT EVT = cast<VTSDNode>(N1)->getVT();
5901 unsigned VTBits = VT.getScalarType().getSizeInBits();
5902 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
5904 // fold (sext_in_reg c1) -> c1
5905 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
5906 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1);
5908 // If the input is already sign extended, just drop the extension.
5909 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
5912 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
5913 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
5914 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT()))
5915 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
5916 N0.getOperand(0), N1);
5918 // fold (sext_in_reg (sext x)) -> (sext x)
5919 // fold (sext_in_reg (aext x)) -> (sext x)
5920 // if x is small enough.
5921 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
5922 SDValue N00 = N0.getOperand(0);
5923 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
5924 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
5925 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
5928 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
5929 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
5930 return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT);
5932 // fold operands of sext_in_reg based on knowledge that the top bits are not
5934 if (SimplifyDemandedBits(SDValue(N, 0)))
5935 return SDValue(N, 0);
5937 // fold (sext_in_reg (load x)) -> (smaller sextload x)
5938 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
5939 SDValue NarrowLoad = ReduceLoadWidth(N);
5940 if (NarrowLoad.getNode())
5943 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
5944 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
5945 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
5946 if (N0.getOpcode() == ISD::SRL) {
5947 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
5948 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
5949 // We can turn this into an SRA iff the input to the SRL is already sign
5951 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
5952 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
5953 return DAG.getNode(ISD::SRA, SDLoc(N), VT,
5954 N0.getOperand(0), N0.getOperand(1));
5958 // fold (sext_inreg (extload x)) -> (sextload x)
5959 if (ISD::isEXTLoad(N0.getNode()) &&
5960 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5961 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5962 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5963 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5964 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5965 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5967 LN0->getBasePtr(), EVT,
5968 LN0->getMemOperand());
5969 CombineTo(N, ExtLoad);
5970 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5971 AddToWorklist(ExtLoad.getNode());
5972 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5974 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
5975 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5977 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5978 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5979 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5980 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5981 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5983 LN0->getBasePtr(), EVT,
5984 LN0->getMemOperand());
5985 CombineTo(N, ExtLoad);
5986 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5987 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5990 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
5991 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
5992 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
5993 N0.getOperand(1), false);
5994 if (BSwap.getNode())
5995 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
5999 // Fold a sext_inreg of a build_vector of ConstantSDNodes or undefs
6000 // into a build_vector.
6001 if (ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
6002 SmallVector<SDValue, 8> Elts;
6003 unsigned NumElts = N0->getNumOperands();
6004 unsigned ShAmt = VTBits - EVTBits;
6006 for (unsigned i = 0; i != NumElts; ++i) {
6007 SDValue Op = N0->getOperand(i);
6008 if (Op->getOpcode() == ISD::UNDEF) {
6013 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
6014 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
6015 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
6016 Op.getValueType()));
6019 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Elts);
6025 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
6026 SDValue N0 = N->getOperand(0);
6027 EVT VT = N->getValueType(0);
6028 bool isLE = TLI.isLittleEndian();
6031 if (N0.getValueType() == N->getValueType(0))
6033 // fold (truncate c1) -> c1
6034 if (isa<ConstantSDNode>(N0))
6035 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0);
6036 // fold (truncate (truncate x)) -> (truncate x)
6037 if (N0.getOpcode() == ISD::TRUNCATE)
6038 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6039 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
6040 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
6041 N0.getOpcode() == ISD::SIGN_EXTEND ||
6042 N0.getOpcode() == ISD::ANY_EXTEND) {
6043 if (N0.getOperand(0).getValueType().bitsLT(VT))
6044 // if the source is smaller than the dest, we still need an extend
6045 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
6047 if (N0.getOperand(0).getValueType().bitsGT(VT))
6048 // if the source is larger than the dest, than we just need the truncate
6049 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6050 // if the source and dest are the same type, we can drop both the extend
6051 // and the truncate.
6052 return N0.getOperand(0);
6055 // Fold extract-and-trunc into a narrow extract. For example:
6056 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
6057 // i32 y = TRUNCATE(i64 x)
6059 // v16i8 b = BITCAST (v2i64 val)
6060 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
6062 // Note: We only run this optimization after type legalization (which often
6063 // creates this pattern) and before operation legalization after which
6064 // we need to be more careful about the vector instructions that we generate.
6065 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6066 LegalTypes && !LegalOperations && N0->hasOneUse() && VT != MVT::i1) {
6068 EVT VecTy = N0.getOperand(0).getValueType();
6069 EVT ExTy = N0.getValueType();
6070 EVT TrTy = N->getValueType(0);
6072 unsigned NumElem = VecTy.getVectorNumElements();
6073 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
6075 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
6076 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
6078 SDValue EltNo = N0->getOperand(1);
6079 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
6080 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
6081 EVT IndexTy = TLI.getVectorIdxTy();
6082 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
6084 SDValue V = DAG.getNode(ISD::BITCAST, SDLoc(N),
6085 NVT, N0.getOperand(0));
6087 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
6089 DAG.getConstant(Index, IndexTy));
6093 // trunc (select c, a, b) -> select c, (trunc a), (trunc b)
6094 if (N0.getOpcode() == ISD::SELECT) {
6095 EVT SrcVT = N0.getValueType();
6096 if ((!LegalOperations || TLI.isOperationLegal(ISD::SELECT, SrcVT)) &&
6097 TLI.isTruncateFree(SrcVT, VT)) {
6099 SDValue Cond = N0.getOperand(0);
6100 SDValue TruncOp0 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(1));
6101 SDValue TruncOp1 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(2));
6102 return DAG.getNode(ISD::SELECT, SDLoc(N), VT, Cond, TruncOp0, TruncOp1);
6106 // Fold a series of buildvector, bitcast, and truncate if possible.
6108 // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to
6109 // (2xi32 (buildvector x, y)).
6110 if (Level == AfterLegalizeVectorOps && VT.isVector() &&
6111 N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
6112 N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
6113 N0.getOperand(0).hasOneUse()) {
6115 SDValue BuildVect = N0.getOperand(0);
6116 EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType();
6117 EVT TruncVecEltTy = VT.getVectorElementType();
6119 // Check that the element types match.
6120 if (BuildVectEltTy == TruncVecEltTy) {
6121 // Now we only need to compute the offset of the truncated elements.
6122 unsigned BuildVecNumElts = BuildVect.getNumOperands();
6123 unsigned TruncVecNumElts = VT.getVectorNumElements();
6124 unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts;
6126 assert((BuildVecNumElts % TruncVecNumElts) == 0 &&
6127 "Invalid number of elements");
6129 SmallVector<SDValue, 8> Opnds;
6130 for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset)
6131 Opnds.push_back(BuildVect.getOperand(i));
6133 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
6137 // See if we can simplify the input to this truncate through knowledge that
6138 // only the low bits are being used.
6139 // For example "trunc (or (shl x, 8), y)" // -> trunc y
6140 // Currently we only perform this optimization on scalars because vectors
6141 // may have different active low bits.
6142 if (!VT.isVector()) {
6144 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
6145 VT.getSizeInBits()));
6146 if (Shorter.getNode())
6147 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter);
6149 // fold (truncate (load x)) -> (smaller load x)
6150 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
6151 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
6152 SDValue Reduced = ReduceLoadWidth(N);
6153 if (Reduced.getNode())
6155 // Handle the case where the load remains an extending load even
6156 // after truncation.
6157 if (N0.hasOneUse() && ISD::isUNINDEXEDLoad(N0.getNode())) {
6158 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6159 if (!LN0->isVolatile() &&
6160 LN0->getMemoryVT().getStoreSizeInBits() < VT.getSizeInBits()) {
6161 SDValue NewLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(LN0),
6162 VT, LN0->getChain(), LN0->getBasePtr(),
6164 LN0->getMemOperand());
6165 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLoad.getValue(1));
6170 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
6171 // where ... are all 'undef'.
6172 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
6173 SmallVector<EVT, 8> VTs;
6176 unsigned NumDefs = 0;
6178 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
6179 SDValue X = N0.getOperand(i);
6180 if (X.getOpcode() != ISD::UNDEF) {
6185 // Stop if more than one members are non-undef.
6188 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
6189 VT.getVectorElementType(),
6190 X.getValueType().getVectorNumElements()));
6194 return DAG.getUNDEF(VT);
6197 assert(V.getNode() && "The single defined operand is empty!");
6198 SmallVector<SDValue, 8> Opnds;
6199 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
6201 Opnds.push_back(DAG.getUNDEF(VTs[i]));
6204 SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V);
6205 AddToWorklist(NV.getNode());
6206 Opnds.push_back(NV);
6208 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Opnds);
6212 // Simplify the operands using demanded-bits information.
6213 if (!VT.isVector() &&
6214 SimplifyDemandedBits(SDValue(N, 0)))
6215 return SDValue(N, 0);
6220 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
6221 SDValue Elt = N->getOperand(i);
6222 if (Elt.getOpcode() != ISD::MERGE_VALUES)
6223 return Elt.getNode();
6224 return Elt.getOperand(Elt.getResNo()).getNode();
6227 /// build_pair (load, load) -> load
6228 /// if load locations are consecutive.
6229 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
6230 assert(N->getOpcode() == ISD::BUILD_PAIR);
6232 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
6233 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
6234 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
6235 LD1->getAddressSpace() != LD2->getAddressSpace())
6237 EVT LD1VT = LD1->getValueType(0);
6239 if (ISD::isNON_EXTLoad(LD2) &&
6241 // If both are volatile this would reduce the number of volatile loads.
6242 // If one is volatile it might be ok, but play conservative and bail out.
6243 !LD1->isVolatile() &&
6244 !LD2->isVolatile() &&
6245 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
6246 unsigned Align = LD1->getAlignment();
6247 unsigned NewAlign = TLI.getDataLayout()->
6248 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6250 if (NewAlign <= Align &&
6251 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
6252 return DAG.getLoad(VT, SDLoc(N), LD1->getChain(),
6253 LD1->getBasePtr(), LD1->getPointerInfo(),
6254 false, false, false, Align);
6260 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
6261 SDValue N0 = N->getOperand(0);
6262 EVT VT = N->getValueType(0);
6264 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
6265 // Only do this before legalize, since afterward the target may be depending
6266 // on the bitconvert.
6267 // First check to see if this is all constant.
6269 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
6271 bool isSimple = cast<BuildVectorSDNode>(N0)->isConstant();
6273 EVT DestEltVT = N->getValueType(0).getVectorElementType();
6274 assert(!DestEltVT.isVector() &&
6275 "Element type of vector ValueType must not be vector!");
6277 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
6280 // If the input is a constant, let getNode fold it.
6281 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
6282 SDValue Res = DAG.getNode(ISD::BITCAST, SDLoc(N), VT, N0);
6283 if (Res.getNode() != N) {
6284 if (!LegalOperations ||
6285 TLI.isOperationLegal(Res.getNode()->getOpcode(), VT))
6288 // Folding it resulted in an illegal node, and it's too late to
6289 // do that. Clean up the old node and forego the transformation.
6290 // Ideally this won't happen very often, because instcombine
6291 // and the earlier dagcombine runs (where illegal nodes are
6292 // permitted) should have folded most of them already.
6293 deleteAndRecombine(Res.getNode());
6297 // (conv (conv x, t1), t2) -> (conv x, t2)
6298 if (N0.getOpcode() == ISD::BITCAST)
6299 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT,
6302 // fold (conv (load x)) -> (load (conv*)x)
6303 // If the resultant load doesn't need a higher alignment than the original!
6304 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6305 // Do not change the width of a volatile load.
6306 !cast<LoadSDNode>(N0)->isVolatile() &&
6307 // Do not remove the cast if the types differ in endian layout.
6308 TLI.hasBigEndianPartOrdering(N0.getValueType()) ==
6309 TLI.hasBigEndianPartOrdering(VT) &&
6310 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)) &&
6311 TLI.isLoadBitCastBeneficial(N0.getValueType(), VT)) {
6312 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6313 unsigned Align = TLI.getDataLayout()->
6314 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6315 unsigned OrigAlign = LN0->getAlignment();
6317 if (Align <= OrigAlign) {
6318 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(),
6319 LN0->getBasePtr(), LN0->getPointerInfo(),
6320 LN0->isVolatile(), LN0->isNonTemporal(),
6321 LN0->isInvariant(), OrigAlign,
6323 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
6328 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
6329 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
6330 // This often reduces constant pool loads.
6331 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) ||
6332 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) &&
6333 N0.getNode()->hasOneUse() && VT.isInteger() &&
6334 !VT.isVector() && !N0.getValueType().isVector()) {
6335 SDValue NewConv = DAG.getNode(ISD::BITCAST, SDLoc(N0), VT,
6337 AddToWorklist(NewConv.getNode());
6339 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6340 if (N0.getOpcode() == ISD::FNEG)
6341 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
6342 NewConv, DAG.getConstant(SignBit, VT));
6343 assert(N0.getOpcode() == ISD::FABS);
6344 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6345 NewConv, DAG.getConstant(~SignBit, VT));
6348 // fold (bitconvert (fcopysign cst, x)) ->
6349 // (or (and (bitconvert x), sign), (and cst, (not sign)))
6350 // Note that we don't handle (copysign x, cst) because this can always be
6351 // folded to an fneg or fabs.
6352 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
6353 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
6354 VT.isInteger() && !VT.isVector()) {
6355 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
6356 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
6357 if (isTypeLegal(IntXVT)) {
6358 SDValue X = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6359 IntXVT, N0.getOperand(1));
6360 AddToWorklist(X.getNode());
6362 // If X has a different width than the result/lhs, sext it or truncate it.
6363 unsigned VTWidth = VT.getSizeInBits();
6364 if (OrigXWidth < VTWidth) {
6365 X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X);
6366 AddToWorklist(X.getNode());
6367 } else if (OrigXWidth > VTWidth) {
6368 // To get the sign bit in the right place, we have to shift it right
6369 // before truncating.
6370 X = DAG.getNode(ISD::SRL, SDLoc(X),
6371 X.getValueType(), X,
6372 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
6373 AddToWorklist(X.getNode());
6374 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
6375 AddToWorklist(X.getNode());
6378 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6379 X = DAG.getNode(ISD::AND, SDLoc(X), VT,
6380 X, DAG.getConstant(SignBit, VT));
6381 AddToWorklist(X.getNode());
6383 SDValue Cst = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6384 VT, N0.getOperand(0));
6385 Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT,
6386 Cst, DAG.getConstant(~SignBit, VT));
6387 AddToWorklist(Cst.getNode());
6389 return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst);
6393 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
6394 if (N0.getOpcode() == ISD::BUILD_PAIR) {
6395 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
6396 if (CombineLD.getNode())
6403 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
6404 EVT VT = N->getValueType(0);
6405 return CombineConsecutiveLoads(N, VT);
6408 /// We know that BV is a build_vector node with Constant, ConstantFP or Undef
6409 /// operands. DstEltVT indicates the destination element value type.
6410 SDValue DAGCombiner::
6411 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
6412 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
6414 // If this is already the right type, we're done.
6415 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
6417 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
6418 unsigned DstBitSize = DstEltVT.getSizeInBits();
6420 // If this is a conversion of N elements of one type to N elements of another
6421 // type, convert each element. This handles FP<->INT cases.
6422 if (SrcBitSize == DstBitSize) {
6423 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6424 BV->getValueType(0).getVectorNumElements());
6426 // Due to the FP element handling below calling this routine recursively,
6427 // we can end up with a scalar-to-vector node here.
6428 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
6429 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6430 DAG.getNode(ISD::BITCAST, SDLoc(BV),
6431 DstEltVT, BV->getOperand(0)));
6433 SmallVector<SDValue, 8> Ops;
6434 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6435 SDValue Op = BV->getOperand(i);
6436 // If the vector element type is not legal, the BUILD_VECTOR operands
6437 // are promoted and implicitly truncated. Make that explicit here.
6438 if (Op.getValueType() != SrcEltVT)
6439 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op);
6440 Ops.push_back(DAG.getNode(ISD::BITCAST, SDLoc(BV),
6442 AddToWorklist(Ops.back().getNode());
6444 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6447 // Otherwise, we're growing or shrinking the elements. To avoid having to
6448 // handle annoying details of growing/shrinking FP values, we convert them to
6450 if (SrcEltVT.isFloatingPoint()) {
6451 // Convert the input float vector to a int vector where the elements are the
6453 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
6454 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
6455 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
6459 // Now we know the input is an integer vector. If the output is a FP type,
6460 // convert to integer first, then to FP of the right size.
6461 if (DstEltVT.isFloatingPoint()) {
6462 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
6463 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
6464 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
6466 // Next, convert to FP elements of the same size.
6467 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
6470 // Okay, we know the src/dst types are both integers of differing types.
6471 // Handling growing first.
6472 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
6473 if (SrcBitSize < DstBitSize) {
6474 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
6476 SmallVector<SDValue, 8> Ops;
6477 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
6478 i += NumInputsPerOutput) {
6479 bool isLE = TLI.isLittleEndian();
6480 APInt NewBits = APInt(DstBitSize, 0);
6481 bool EltIsUndef = true;
6482 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
6483 // Shift the previously computed bits over.
6484 NewBits <<= SrcBitSize;
6485 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
6486 if (Op.getOpcode() == ISD::UNDEF) continue;
6489 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
6490 zextOrTrunc(SrcBitSize).zext(DstBitSize);
6494 Ops.push_back(DAG.getUNDEF(DstEltVT));
6496 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
6499 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
6500 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6503 // Finally, this must be the case where we are shrinking elements: each input
6504 // turns into multiple outputs.
6505 bool isS2V = ISD::isScalarToVector(BV);
6506 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
6507 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6508 NumOutputsPerInput*BV->getNumOperands());
6509 SmallVector<SDValue, 8> Ops;
6511 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6512 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
6513 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
6514 Ops.push_back(DAG.getUNDEF(DstEltVT));
6518 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
6519 getAPIntValue().zextOrTrunc(SrcBitSize);
6521 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
6522 APInt ThisVal = OpVal.trunc(DstBitSize);
6523 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
6524 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
6525 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
6526 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6528 OpVal = OpVal.lshr(DstBitSize);
6531 // For big endian targets, swap the order of the pieces of each element.
6532 if (TLI.isBigEndian())
6533 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
6536 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6539 SDValue DAGCombiner::visitFADD(SDNode *N) {
6540 SDValue N0 = N->getOperand(0);
6541 SDValue N1 = N->getOperand(1);
6542 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6543 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6544 EVT VT = N->getValueType(0);
6545 const TargetOptions &Options = DAG.getTarget().Options;
6548 if (VT.isVector()) {
6549 SDValue FoldedVOp = SimplifyVBinOp(N);
6550 if (FoldedVOp.getNode()) return FoldedVOp;
6553 // fold (fadd c1, c2) -> c1 + c2
6555 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N1);
6557 // canonicalize constant to RHS
6558 if (N0CFP && !N1CFP)
6559 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N0);
6561 // fold (fadd A, (fneg B)) -> (fsub A, B)
6562 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6563 isNegatibleForFree(N1, LegalOperations, TLI, &Options) == 2)
6564 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0,
6565 GetNegatedExpression(N1, DAG, LegalOperations));
6567 // fold (fadd (fneg A), B) -> (fsub B, A)
6568 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6569 isNegatibleForFree(N0, LegalOperations, TLI, &Options) == 2)
6570 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N1,
6571 GetNegatedExpression(N0, DAG, LegalOperations));
6573 // If 'unsafe math' is enabled, fold lots of things.
6574 if (Options.UnsafeFPMath) {
6575 // No FP constant should be created after legalization as Instruction
6576 // Selection pass has a hard time dealing with FP constants.
6577 bool AllowNewConst = (Level < AfterLegalizeDAG);
6579 // fold (fadd A, 0) -> A
6580 if (N1CFP && N1CFP->getValueAPF().isZero())
6583 // fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
6584 if (N1CFP && N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
6585 isa<ConstantFPSDNode>(N0.getOperand(1)))
6586 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0.getOperand(0),
6587 DAG.getNode(ISD::FADD, SDLoc(N), VT,
6588 N0.getOperand(1), N1));
6590 // If allowed, fold (fadd (fneg x), x) -> 0.0
6591 if (AllowNewConst && N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1)
6592 return DAG.getConstantFP(0.0, VT);
6594 // If allowed, fold (fadd x, (fneg x)) -> 0.0
6595 if (AllowNewConst && N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0)
6596 return DAG.getConstantFP(0.0, VT);
6598 // We can fold chains of FADD's of the same value into multiplications.
6599 // This transform is not safe in general because we are reducing the number
6600 // of rounding steps.
6601 if (TLI.isOperationLegalOrCustom(ISD::FMUL, VT) && !N0CFP && !N1CFP) {
6602 if (N0.getOpcode() == ISD::FMUL) {
6603 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
6604 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
6606 // (fadd (fmul x, c), x) -> (fmul x, c+1)
6607 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
6608 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6610 DAG.getConstantFP(1.0, VT));
6611 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, NewCFP);
6614 // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2)
6615 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
6616 N1.getOperand(0) == N1.getOperand(1) &&
6617 N0.getOperand(0) == N1.getOperand(0)) {
6618 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6620 DAG.getConstantFP(2.0, VT));
6621 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6622 N0.getOperand(0), NewCFP);
6626 if (N1.getOpcode() == ISD::FMUL) {
6627 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
6628 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
6630 // (fadd x, (fmul x, c)) -> (fmul x, c+1)
6631 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
6632 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6634 DAG.getConstantFP(1.0, VT));
6635 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, NewCFP);
6638 // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2)
6639 if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD &&
6640 N0.getOperand(0) == N0.getOperand(1) &&
6641 N1.getOperand(0) == N0.getOperand(0)) {
6642 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6644 DAG.getConstantFP(2.0, VT));
6645 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1.getOperand(0), NewCFP);
6649 if (N0.getOpcode() == ISD::FADD && AllowNewConst) {
6650 ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
6651 // (fadd (fadd x, x), x) -> (fmul x, 3.0)
6652 if (!CFP && N0.getOperand(0) == N0.getOperand(1) &&
6653 (N0.getOperand(0) == N1))
6654 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6655 N1, DAG.getConstantFP(3.0, VT));
6658 if (N1.getOpcode() == ISD::FADD && AllowNewConst) {
6659 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
6660 // (fadd x, (fadd x, x)) -> (fmul x, 3.0)
6661 if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) &&
6662 N1.getOperand(0) == N0)
6663 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6664 N0, DAG.getConstantFP(3.0, VT));
6667 // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0)
6668 if (AllowNewConst &&
6669 N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
6670 N0.getOperand(0) == N0.getOperand(1) &&
6671 N1.getOperand(0) == N1.getOperand(1) &&
6672 N0.getOperand(0) == N1.getOperand(0))
6673 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6674 N0.getOperand(0), DAG.getConstantFP(4.0, VT));
6676 } // enable-unsafe-fp-math
6678 // FADD -> FMA combines:
6679 if ((Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath) &&
6682 ->getTargetLowering()
6683 ->isFMAFasterThanFMulAndFAdd(VT) &&
6684 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
6686 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
6687 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse())
6688 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6689 N0.getOperand(0), N0.getOperand(1), N1);
6691 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
6692 // Note: Commutes FADD operands.
6693 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse())
6694 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6695 N1.getOperand(0), N1.getOperand(1), N0);
6701 SDValue DAGCombiner::visitFSUB(SDNode *N) {
6702 SDValue N0 = N->getOperand(0);
6703 SDValue N1 = N->getOperand(1);
6704 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
6705 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
6706 EVT VT = N->getValueType(0);
6708 const TargetOptions &Options = DAG.getTarget().Options;
6711 if (VT.isVector()) {
6712 SDValue FoldedVOp = SimplifyVBinOp(N);
6713 if (FoldedVOp.getNode()) return FoldedVOp;
6716 // fold (fsub c1, c2) -> c1-c2
6718 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0, N1);
6720 // fold (fsub A, (fneg B)) -> (fadd A, B)
6721 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
6722 return DAG.getNode(ISD::FADD, dl, VT, N0,
6723 GetNegatedExpression(N1, DAG, LegalOperations));
6725 // If 'unsafe math' is enabled, fold lots of things.
6726 if (Options.UnsafeFPMath) {
6728 if (N1CFP && N1CFP->getValueAPF().isZero())
6731 // (fsub 0, B) -> -B
6732 if (N0CFP && N0CFP->getValueAPF().isZero()) {
6733 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
6734 return GetNegatedExpression(N1, DAG, LegalOperations);
6735 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6736 return DAG.getNode(ISD::FNEG, dl, VT, N1);
6739 // (fsub x, x) -> 0.0
6741 return DAG.getConstantFP(0.0f, VT);
6743 // (fsub x, (fadd x, y)) -> (fneg y)
6744 // (fsub x, (fadd y, x)) -> (fneg y)
6745 if (N1.getOpcode() == ISD::FADD) {
6746 SDValue N10 = N1->getOperand(0);
6747 SDValue N11 = N1->getOperand(1);
6749 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI, &Options))
6750 return GetNegatedExpression(N11, DAG, LegalOperations);
6752 if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI, &Options))
6753 return GetNegatedExpression(N10, DAG, LegalOperations);
6757 // FSUB -> FMA combines:
6758 if ((Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath) &&
6759 DAG.getTarget().getSubtargetImpl()
6760 ->getTargetLowering()
6761 ->isFMAFasterThanFMulAndFAdd(VT) &&
6762 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
6764 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
6765 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse())
6766 return DAG.getNode(ISD::FMA, dl, VT,
6767 N0.getOperand(0), N0.getOperand(1),
6768 DAG.getNode(ISD::FNEG, dl, VT, N1));
6770 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
6771 // Note: Commutes FSUB operands.
6772 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse())
6773 return DAG.getNode(ISD::FMA, dl, VT,
6774 DAG.getNode(ISD::FNEG, dl, VT,
6776 N1.getOperand(1), N0);
6778 // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
6779 if (N0.getOpcode() == ISD::FNEG &&
6780 N0.getOperand(0).getOpcode() == ISD::FMUL &&
6781 N0->hasOneUse() && N0.getOperand(0).hasOneUse()) {
6782 SDValue N00 = N0.getOperand(0).getOperand(0);
6783 SDValue N01 = N0.getOperand(0).getOperand(1);
6784 return DAG.getNode(ISD::FMA, dl, VT,
6785 DAG.getNode(ISD::FNEG, dl, VT, N00), N01,
6786 DAG.getNode(ISD::FNEG, dl, VT, N1));
6793 SDValue DAGCombiner::visitFMUL(SDNode *N) {
6794 SDValue N0 = N->getOperand(0);
6795 SDValue N1 = N->getOperand(1);
6796 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
6797 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
6798 EVT VT = N->getValueType(0);
6799 const TargetOptions &Options = DAG.getTarget().Options;
6802 if (VT.isVector()) {
6803 // This just handles C1 * C2 for vectors. Other vector folds are below.
6804 SDValue FoldedVOp = SimplifyVBinOp(N);
6805 if (FoldedVOp.getNode())
6807 // Canonicalize vector constant to RHS.
6808 if (N0.getOpcode() == ISD::BUILD_VECTOR &&
6809 N1.getOpcode() != ISD::BUILD_VECTOR)
6810 if (auto *BV0 = dyn_cast<BuildVectorSDNode>(N0))
6811 if (BV0->isConstant())
6812 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, N1, N0);
6815 // fold (fmul c1, c2) -> c1*c2
6817 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, N1);
6819 // canonicalize constant to RHS
6820 if (N0CFP && !N1CFP)
6821 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, N0);
6823 // fold (fmul A, 1.0) -> A
6824 if (N1CFP && N1CFP->isExactlyValue(1.0))
6827 if (Options.UnsafeFPMath) {
6828 // fold (fmul A, 0) -> 0
6829 if (N1CFP && N1CFP->getValueAPF().isZero())
6832 // fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
6833 if (N0.getOpcode() == ISD::FMUL) {
6834 // Fold scalars or any vector constants (not just splats).
6835 // This fold is done in general by InstCombine, but extra fmul insts
6836 // may have been generated during lowering.
6837 SDValue N01 = N0.getOperand(1);
6838 auto *BV1 = dyn_cast<BuildVectorSDNode>(N1);
6839 auto *BV01 = dyn_cast<BuildVectorSDNode>(N01);
6840 if ((N1CFP && isConstOrConstSplatFP(N01)) ||
6841 (BV1 && BV01 && BV1->isConstant() && BV01->isConstant())) {
6843 SDValue MulConsts = DAG.getNode(ISD::FMUL, SL, VT, N01, N1);
6844 return DAG.getNode(ISD::FMUL, SL, VT, N0.getOperand(0), MulConsts);
6848 // fold (fmul (fadd x, x), c) -> (fmul x, (fmul 2.0, c))
6849 // Undo the fmul 2.0, x -> fadd x, x transformation, since if it occurs
6850 // during an early run of DAGCombiner can prevent folding with fmuls
6851 // inserted during lowering.
6852 if (N0.getOpcode() == ISD::FADD && N0.getOperand(0) == N0.getOperand(1)) {
6854 const SDValue Two = DAG.getConstantFP(2.0, VT);
6855 SDValue MulConsts = DAG.getNode(ISD::FMUL, SL, VT, Two, N1);
6856 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0.getOperand(0), MulConsts);
6860 // fold (fmul X, 2.0) -> (fadd X, X)
6861 if (N1CFP && N1CFP->isExactlyValue(+2.0))
6862 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N0);
6864 // fold (fmul X, -1.0) -> (fneg X)
6865 if (N1CFP && N1CFP->isExactlyValue(-1.0))
6866 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6867 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
6869 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
6870 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
6871 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
6872 // Both can be negated for free, check to see if at least one is cheaper
6874 if (LHSNeg == 2 || RHSNeg == 2)
6875 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6876 GetNegatedExpression(N0, DAG, LegalOperations),
6877 GetNegatedExpression(N1, DAG, LegalOperations));
6884 SDValue DAGCombiner::visitFMA(SDNode *N) {
6885 SDValue N0 = N->getOperand(0);
6886 SDValue N1 = N->getOperand(1);
6887 SDValue N2 = N->getOperand(2);
6888 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6889 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6890 EVT VT = N->getValueType(0);
6892 const TargetOptions &Options = DAG.getTarget().Options;
6894 // Constant fold FMA.
6895 if (isa<ConstantFPSDNode>(N0) &&
6896 isa<ConstantFPSDNode>(N1) &&
6897 isa<ConstantFPSDNode>(N2)) {
6898 return DAG.getNode(ISD::FMA, dl, VT, N0, N1, N2);
6901 if (Options.UnsafeFPMath) {
6902 if (N0CFP && N0CFP->isZero())
6904 if (N1CFP && N1CFP->isZero())
6907 if (N0CFP && N0CFP->isExactlyValue(1.0))
6908 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2);
6909 if (N1CFP && N1CFP->isExactlyValue(1.0))
6910 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2);
6912 // Canonicalize (fma c, x, y) -> (fma x, c, y)
6913 if (N0CFP && !N1CFP)
6914 return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2);
6916 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
6917 if (Options.UnsafeFPMath && N1CFP &&
6918 N2.getOpcode() == ISD::FMUL &&
6919 N0 == N2.getOperand(0) &&
6920 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
6921 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6922 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
6926 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
6927 if (Options.UnsafeFPMath &&
6928 N0.getOpcode() == ISD::FMUL && N1CFP &&
6929 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
6930 return DAG.getNode(ISD::FMA, dl, VT,
6932 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
6936 // (fma x, 1, y) -> (fadd x, y)
6937 // (fma x, -1, y) -> (fadd (fneg x), y)
6939 if (N1CFP->isExactlyValue(1.0))
6940 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
6942 if (N1CFP->isExactlyValue(-1.0) &&
6943 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
6944 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
6945 AddToWorklist(RHSNeg.getNode());
6946 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
6950 // (fma x, c, x) -> (fmul x, (c+1))
6951 if (Options.UnsafeFPMath && N1CFP && N0 == N2)
6952 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6953 DAG.getNode(ISD::FADD, dl, VT,
6954 N1, DAG.getConstantFP(1.0, VT)));
6956 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
6957 if (Options.UnsafeFPMath && N1CFP &&
6958 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0)
6959 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6960 DAG.getNode(ISD::FADD, dl, VT,
6961 N1, DAG.getConstantFP(-1.0, VT)));
6967 SDValue DAGCombiner::visitFDIV(SDNode *N) {
6968 SDValue N0 = N->getOperand(0);
6969 SDValue N1 = N->getOperand(1);
6970 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6971 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6972 EVT VT = N->getValueType(0);
6973 const TargetOptions &Options = DAG.getTarget().Options;
6976 if (VT.isVector()) {
6977 SDValue FoldedVOp = SimplifyVBinOp(N);
6978 if (FoldedVOp.getNode()) return FoldedVOp;
6981 // fold (fdiv c1, c2) -> c1/c2
6983 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1);
6985 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
6986 if (N1CFP && Options.UnsafeFPMath) {
6987 // Compute the reciprocal 1.0 / c2.
6988 APFloat N1APF = N1CFP->getValueAPF();
6989 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
6990 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
6991 // Only do the transform if the reciprocal is a legal fp immediate that
6992 // isn't too nasty (eg NaN, denormal, ...).
6993 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
6994 (!LegalOperations ||
6995 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
6996 // backend)... we should handle this gracefully after Legalize.
6997 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
6998 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
6999 TLI.isFPImmLegal(Recip, VT)))
7000 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0,
7001 DAG.getConstantFP(Recip, VT));
7004 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
7005 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
7006 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
7007 // Both can be negated for free, check to see if at least one is cheaper
7009 if (LHSNeg == 2 || RHSNeg == 2)
7010 return DAG.getNode(ISD::FDIV, SDLoc(N), VT,
7011 GetNegatedExpression(N0, DAG, LegalOperations),
7012 GetNegatedExpression(N1, DAG, LegalOperations));
7019 SDValue DAGCombiner::visitFREM(SDNode *N) {
7020 SDValue N0 = N->getOperand(0);
7021 SDValue N1 = N->getOperand(1);
7022 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7023 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7024 EVT VT = N->getValueType(0);
7026 // fold (frem c1, c2) -> fmod(c1,c2)
7028 return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1);
7033 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
7034 SDValue N0 = N->getOperand(0);
7035 SDValue N1 = N->getOperand(1);
7036 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7037 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7038 EVT VT = N->getValueType(0);
7040 if (N0CFP && N1CFP) // Constant fold
7041 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1);
7044 const APFloat& V = N1CFP->getValueAPF();
7045 // copysign(x, c1) -> fabs(x) iff ispos(c1)
7046 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
7047 if (!V.isNegative()) {
7048 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
7049 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7051 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7052 return DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7053 DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0));
7057 // copysign(fabs(x), y) -> copysign(x, y)
7058 // copysign(fneg(x), y) -> copysign(x, y)
7059 // copysign(copysign(x,z), y) -> copysign(x, y)
7060 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
7061 N0.getOpcode() == ISD::FCOPYSIGN)
7062 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7063 N0.getOperand(0), N1);
7065 // copysign(x, abs(y)) -> abs(x)
7066 if (N1.getOpcode() == ISD::FABS)
7067 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7069 // copysign(x, copysign(y,z)) -> copysign(x, z)
7070 if (N1.getOpcode() == ISD::FCOPYSIGN)
7071 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7072 N0, N1.getOperand(1));
7074 // copysign(x, fp_extend(y)) -> copysign(x, y)
7075 // copysign(x, fp_round(y)) -> copysign(x, y)
7076 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
7077 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7078 N0, N1.getOperand(0));
7083 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
7084 SDValue N0 = N->getOperand(0);
7085 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
7086 EVT VT = N->getValueType(0);
7087 EVT OpVT = N0.getValueType();
7089 // fold (sint_to_fp c1) -> c1fp
7091 // ...but only if the target supports immediate floating-point values
7092 (!LegalOperations ||
7093 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
7094 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
7096 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
7097 // but UINT_TO_FP is legal on this target, try to convert.
7098 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
7099 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
7100 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
7101 if (DAG.SignBitIsZero(N0))
7102 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
7105 // The next optimizations are desirable only if SELECT_CC can be lowered.
7106 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
7107 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7108 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
7110 (!LegalOperations ||
7111 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7113 { N0.getOperand(0), N0.getOperand(1),
7114 DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT),
7116 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7119 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
7120 // (select_cc x, y, 1.0, 0.0,, cc)
7121 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
7122 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
7123 (!LegalOperations ||
7124 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7126 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
7127 DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT),
7128 N0.getOperand(0).getOperand(2) };
7129 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7136 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
7137 SDValue N0 = N->getOperand(0);
7138 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
7139 EVT VT = N->getValueType(0);
7140 EVT OpVT = N0.getValueType();
7142 // fold (uint_to_fp c1) -> c1fp
7144 // ...but only if the target supports immediate floating-point values
7145 (!LegalOperations ||
7146 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
7147 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
7149 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
7150 // but SINT_TO_FP is legal on this target, try to convert.
7151 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
7152 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
7153 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
7154 if (DAG.SignBitIsZero(N0))
7155 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
7158 // The next optimizations are desirable only if SELECT_CC can be lowered.
7159 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
7160 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7162 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
7163 (!LegalOperations ||
7164 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7166 { N0.getOperand(0), N0.getOperand(1),
7167 DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT),
7169 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7176 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
7177 SDValue N0 = N->getOperand(0);
7178 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7179 EVT VT = N->getValueType(0);
7181 // fold (fp_to_sint c1fp) -> c1
7183 return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0);
7188 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
7189 SDValue N0 = N->getOperand(0);
7190 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7191 EVT VT = N->getValueType(0);
7193 // fold (fp_to_uint c1fp) -> c1
7195 return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0);
7200 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
7201 SDValue N0 = N->getOperand(0);
7202 SDValue N1 = N->getOperand(1);
7203 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7204 EVT VT = N->getValueType(0);
7206 // fold (fp_round c1fp) -> c1fp
7208 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1);
7210 // fold (fp_round (fp_extend x)) -> x
7211 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
7212 return N0.getOperand(0);
7214 // fold (fp_round (fp_round x)) -> (fp_round x)
7215 if (N0.getOpcode() == ISD::FP_ROUND) {
7216 // This is a value preserving truncation if both round's are.
7217 bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
7218 N0.getNode()->getConstantOperandVal(1) == 1;
7219 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0.getOperand(0),
7220 DAG.getIntPtrConstant(IsTrunc));
7223 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
7224 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
7225 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT,
7226 N0.getOperand(0), N1);
7227 AddToWorklist(Tmp.getNode());
7228 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7229 Tmp, N0.getOperand(1));
7235 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
7236 SDValue N0 = N->getOperand(0);
7237 EVT VT = N->getValueType(0);
7238 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
7239 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7241 // fold (fp_round_inreg c1fp) -> c1fp
7242 if (N0CFP && isTypeLegal(EVT)) {
7243 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
7244 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, Round);
7250 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
7251 SDValue N0 = N->getOperand(0);
7252 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7253 EVT VT = N->getValueType(0);
7255 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
7256 if (N->hasOneUse() &&
7257 N->use_begin()->getOpcode() == ISD::FP_ROUND)
7260 // fold (fp_extend c1fp) -> c1fp
7262 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0);
7264 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
7266 if (N0.getOpcode() == ISD::FP_ROUND
7267 && N0.getNode()->getConstantOperandVal(1) == 1) {
7268 SDValue In = N0.getOperand(0);
7269 if (In.getValueType() == VT) return In;
7270 if (VT.bitsLT(In.getValueType()))
7271 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT,
7272 In, N0.getOperand(1));
7273 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In);
7276 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
7277 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7278 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType())) {
7279 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7280 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
7282 LN0->getBasePtr(), N0.getValueType(),
7283 LN0->getMemOperand());
7284 CombineTo(N, ExtLoad);
7285 CombineTo(N0.getNode(),
7286 DAG.getNode(ISD::FP_ROUND, SDLoc(N0),
7287 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
7288 ExtLoad.getValue(1));
7289 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7295 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
7296 SDValue N0 = N->getOperand(0);
7297 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7298 EVT VT = N->getValueType(0);
7300 // fold (fceil c1) -> fceil(c1)
7302 return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0);
7307 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
7308 SDValue N0 = N->getOperand(0);
7309 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7310 EVT VT = N->getValueType(0);
7312 // fold (ftrunc c1) -> ftrunc(c1)
7314 return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0);
7319 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
7320 SDValue N0 = N->getOperand(0);
7321 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7322 EVT VT = N->getValueType(0);
7324 // fold (ffloor c1) -> ffloor(c1)
7326 return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0);
7331 // FIXME: FNEG and FABS have a lot in common; refactor.
7332 SDValue DAGCombiner::visitFNEG(SDNode *N) {
7333 SDValue N0 = N->getOperand(0);
7334 EVT VT = N->getValueType(0);
7336 if (VT.isVector()) {
7337 SDValue FoldedVOp = SimplifyVUnaryOp(N);
7338 if (FoldedVOp.getNode()) return FoldedVOp;
7341 // Constant fold FNEG.
7342 if (isa<ConstantFPSDNode>(N0))
7343 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N->getOperand(0));
7345 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
7346 &DAG.getTarget().Options))
7347 return GetNegatedExpression(N0, DAG, LegalOperations);
7349 // Transform fneg(bitconvert(x)) -> bitconvert(x ^ sign) to avoid loading
7350 // constant pool values.
7351 if (!TLI.isFNegFree(VT) &&
7352 N0.getOpcode() == ISD::BITCAST &&
7353 N0.getNode()->hasOneUse()) {
7354 SDValue Int = N0.getOperand(0);
7355 EVT IntVT = Int.getValueType();
7356 if (IntVT.isInteger() && !IntVT.isVector()) {
7358 if (N0.getValueType().isVector()) {
7359 // For a vector, get a mask such as 0x80... per scalar element
7361 SignMask = APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
7362 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
7364 // For a scalar, just generate 0x80...
7365 SignMask = APInt::getSignBit(IntVT.getSizeInBits());
7367 Int = DAG.getNode(ISD::XOR, SDLoc(N0), IntVT, Int,
7368 DAG.getConstant(SignMask, IntVT));
7369 AddToWorklist(Int.getNode());
7370 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Int);
7374 // (fneg (fmul c, x)) -> (fmul -c, x)
7375 if (N0.getOpcode() == ISD::FMUL) {
7376 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
7378 APFloat CVal = CFP1->getValueAPF();
7380 if (Level >= AfterLegalizeDAG &&
7381 (TLI.isFPImmLegal(CVal, N->getValueType(0)) ||
7382 TLI.isOperationLegal(ISD::ConstantFP, N->getValueType(0))))
7384 ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
7385 DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0.getOperand(1)));
7392 SDValue DAGCombiner::visitFABS(SDNode *N) {
7393 SDValue N0 = N->getOperand(0);
7394 EVT VT = N->getValueType(0);
7396 if (VT.isVector()) {
7397 SDValue FoldedVOp = SimplifyVUnaryOp(N);
7398 if (FoldedVOp.getNode()) return FoldedVOp;
7401 // fold (fabs c1) -> fabs(c1)
7402 if (isa<ConstantFPSDNode>(N0))
7403 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7405 // fold (fabs (fabs x)) -> (fabs x)
7406 if (N0.getOpcode() == ISD::FABS)
7407 return N->getOperand(0);
7409 // fold (fabs (fneg x)) -> (fabs x)
7410 // fold (fabs (fcopysign x, y)) -> (fabs x)
7411 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
7412 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0));
7414 // Transform fabs(bitconvert(x)) -> bitconvert(x & ~sign) to avoid loading
7415 // constant pool values.
7416 if (!TLI.isFAbsFree(VT) &&
7417 N0.getOpcode() == ISD::BITCAST &&
7418 N0.getNode()->hasOneUse()) {
7419 SDValue Int = N0.getOperand(0);
7420 EVT IntVT = Int.getValueType();
7421 if (IntVT.isInteger() && !IntVT.isVector()) {
7423 if (N0.getValueType().isVector()) {
7424 // For a vector, get a mask such as 0x7f... per scalar element
7426 SignMask = ~APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
7427 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
7429 // For a scalar, just generate 0x7f...
7430 SignMask = ~APInt::getSignBit(IntVT.getSizeInBits());
7432 Int = DAG.getNode(ISD::AND, SDLoc(N0), IntVT, Int,
7433 DAG.getConstant(SignMask, IntVT));
7434 AddToWorklist(Int.getNode());
7435 return DAG.getNode(ISD::BITCAST, SDLoc(N), N->getValueType(0), Int);
7442 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
7443 SDValue Chain = N->getOperand(0);
7444 SDValue N1 = N->getOperand(1);
7445 SDValue N2 = N->getOperand(2);
7447 // If N is a constant we could fold this into a fallthrough or unconditional
7448 // branch. However that doesn't happen very often in normal code, because
7449 // Instcombine/SimplifyCFG should have handled the available opportunities.
7450 // If we did this folding here, it would be necessary to update the
7451 // MachineBasicBlock CFG, which is awkward.
7453 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
7455 if (N1.getOpcode() == ISD::SETCC &&
7456 TLI.isOperationLegalOrCustom(ISD::BR_CC,
7457 N1.getOperand(0).getValueType())) {
7458 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
7459 Chain, N1.getOperand(2),
7460 N1.getOperand(0), N1.getOperand(1), N2);
7463 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
7464 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
7465 (N1.getOperand(0).hasOneUse() &&
7466 N1.getOperand(0).getOpcode() == ISD::SRL))) {
7467 SDNode *Trunc = nullptr;
7468 if (N1.getOpcode() == ISD::TRUNCATE) {
7469 // Look pass the truncate.
7470 Trunc = N1.getNode();
7471 N1 = N1.getOperand(0);
7474 // Match this pattern so that we can generate simpler code:
7477 // %b = and i32 %a, 2
7478 // %c = srl i32 %b, 1
7479 // brcond i32 %c ...
7484 // %b = and i32 %a, 2
7485 // %c = setcc eq %b, 0
7488 // This applies only when the AND constant value has one bit set and the
7489 // SRL constant is equal to the log2 of the AND constant. The back-end is
7490 // smart enough to convert the result into a TEST/JMP sequence.
7491 SDValue Op0 = N1.getOperand(0);
7492 SDValue Op1 = N1.getOperand(1);
7494 if (Op0.getOpcode() == ISD::AND &&
7495 Op1.getOpcode() == ISD::Constant) {
7496 SDValue AndOp1 = Op0.getOperand(1);
7498 if (AndOp1.getOpcode() == ISD::Constant) {
7499 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
7501 if (AndConst.isPowerOf2() &&
7502 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
7504 DAG.getSetCC(SDLoc(N),
7505 getSetCCResultType(Op0.getValueType()),
7506 Op0, DAG.getConstant(0, Op0.getValueType()),
7509 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, SDLoc(N),
7510 MVT::Other, Chain, SetCC, N2);
7511 // Don't add the new BRCond into the worklist or else SimplifySelectCC
7512 // will convert it back to (X & C1) >> C2.
7513 CombineTo(N, NewBRCond, false);
7514 // Truncate is dead.
7516 deleteAndRecombine(Trunc);
7517 // Replace the uses of SRL with SETCC
7518 WorklistRemover DeadNodes(*this);
7519 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
7520 deleteAndRecombine(N1.getNode());
7521 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7527 // Restore N1 if the above transformation doesn't match.
7528 N1 = N->getOperand(1);
7531 // Transform br(xor(x, y)) -> br(x != y)
7532 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
7533 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
7534 SDNode *TheXor = N1.getNode();
7535 SDValue Op0 = TheXor->getOperand(0);
7536 SDValue Op1 = TheXor->getOperand(1);
7537 if (Op0.getOpcode() == Op1.getOpcode()) {
7538 // Avoid missing important xor optimizations.
7539 SDValue Tmp = visitXOR(TheXor);
7540 if (Tmp.getNode()) {
7541 if (Tmp.getNode() != TheXor) {
7542 DEBUG(dbgs() << "\nReplacing.8 ";
7544 dbgs() << "\nWith: ";
7545 Tmp.getNode()->dump(&DAG);
7547 WorklistRemover DeadNodes(*this);
7548 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
7549 deleteAndRecombine(TheXor);
7550 return DAG.getNode(ISD::BRCOND, SDLoc(N),
7551 MVT::Other, Chain, Tmp, N2);
7554 // visitXOR has changed XOR's operands or replaced the XOR completely,
7556 return SDValue(N, 0);
7560 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
7562 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
7563 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
7564 Op0.getOpcode() == ISD::XOR) {
7565 TheXor = Op0.getNode();
7569 EVT SetCCVT = N1.getValueType();
7571 SetCCVT = getSetCCResultType(SetCCVT);
7572 SDValue SetCC = DAG.getSetCC(SDLoc(TheXor),
7575 Equal ? ISD::SETEQ : ISD::SETNE);
7576 // Replace the uses of XOR with SETCC
7577 WorklistRemover DeadNodes(*this);
7578 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
7579 deleteAndRecombine(N1.getNode());
7580 return DAG.getNode(ISD::BRCOND, SDLoc(N),
7581 MVT::Other, Chain, SetCC, N2);
7588 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
7590 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
7591 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
7592 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
7594 // If N is a constant we could fold this into a fallthrough or unconditional
7595 // branch. However that doesn't happen very often in normal code, because
7596 // Instcombine/SimplifyCFG should have handled the available opportunities.
7597 // If we did this folding here, it would be necessary to update the
7598 // MachineBasicBlock CFG, which is awkward.
7600 // Use SimplifySetCC to simplify SETCC's.
7601 SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()),
7602 CondLHS, CondRHS, CC->get(), SDLoc(N),
7604 if (Simp.getNode()) AddToWorklist(Simp.getNode());
7606 // fold to a simpler setcc
7607 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
7608 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
7609 N->getOperand(0), Simp.getOperand(2),
7610 Simp.getOperand(0), Simp.getOperand(1),
7616 /// Return true if 'Use' is a load or a store that uses N as its base pointer
7617 /// and that N may be folded in the load / store addressing mode.
7618 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
7620 const TargetLowering &TLI) {
7622 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
7623 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
7625 VT = Use->getValueType(0);
7626 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
7627 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
7629 VT = ST->getValue().getValueType();
7633 TargetLowering::AddrMode AM;
7634 if (N->getOpcode() == ISD::ADD) {
7635 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
7638 AM.BaseOffs = Offset->getSExtValue();
7642 } else if (N->getOpcode() == ISD::SUB) {
7643 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
7646 AM.BaseOffs = -Offset->getSExtValue();
7653 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
7656 /// Try turning a load/store into a pre-indexed load/store when the base
7657 /// pointer is an add or subtract and it has other uses besides the load/store.
7658 /// After the transformation, the new indexed load/store has effectively folded
7659 /// the add/subtract in and all of its other uses are redirected to the
7661 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
7662 if (Level < AfterLegalizeDAG)
7668 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7669 if (LD->isIndexed())
7671 VT = LD->getMemoryVT();
7672 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
7673 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
7675 Ptr = LD->getBasePtr();
7676 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7677 if (ST->isIndexed())
7679 VT = ST->getMemoryVT();
7680 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
7681 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
7683 Ptr = ST->getBasePtr();
7689 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
7690 // out. There is no reason to make this a preinc/predec.
7691 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
7692 Ptr.getNode()->hasOneUse())
7695 // Ask the target to do addressing mode selection.
7698 ISD::MemIndexedMode AM = ISD::UNINDEXED;
7699 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
7702 // Backends without true r+i pre-indexed forms may need to pass a
7703 // constant base with a variable offset so that constant coercion
7704 // will work with the patterns in canonical form.
7705 bool Swapped = false;
7706 if (isa<ConstantSDNode>(BasePtr)) {
7707 std::swap(BasePtr, Offset);
7711 // Don't create a indexed load / store with zero offset.
7712 if (isa<ConstantSDNode>(Offset) &&
7713 cast<ConstantSDNode>(Offset)->isNullValue())
7716 // Try turning it into a pre-indexed load / store except when:
7717 // 1) The new base ptr is a frame index.
7718 // 2) If N is a store and the new base ptr is either the same as or is a
7719 // predecessor of the value being stored.
7720 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
7721 // that would create a cycle.
7722 // 4) All uses are load / store ops that use it as old base ptr.
7724 // Check #1. Preinc'ing a frame index would require copying the stack pointer
7725 // (plus the implicit offset) to a register to preinc anyway.
7726 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
7731 SDValue Val = cast<StoreSDNode>(N)->getValue();
7732 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
7736 // If the offset is a constant, there may be other adds of constants that
7737 // can be folded with this one. We should do this to avoid having to keep
7738 // a copy of the original base pointer.
7739 SmallVector<SDNode *, 16> OtherUses;
7740 if (isa<ConstantSDNode>(Offset))
7741 for (SDNode *Use : BasePtr.getNode()->uses()) {
7742 if (Use == Ptr.getNode())
7745 if (Use->isPredecessorOf(N))
7748 if (Use->getOpcode() != ISD::ADD && Use->getOpcode() != ISD::SUB) {
7753 SDValue Op0 = Use->getOperand(0), Op1 = Use->getOperand(1);
7754 if (Op1.getNode() == BasePtr.getNode())
7755 std::swap(Op0, Op1);
7756 assert(Op0.getNode() == BasePtr.getNode() &&
7757 "Use of ADD/SUB but not an operand");
7759 if (!isa<ConstantSDNode>(Op1)) {
7764 // FIXME: In some cases, we can be smarter about this.
7765 if (Op1.getValueType() != Offset.getValueType()) {
7770 OtherUses.push_back(Use);
7774 std::swap(BasePtr, Offset);
7776 // Now check for #3 and #4.
7777 bool RealUse = false;
7779 // Caches for hasPredecessorHelper
7780 SmallPtrSet<const SDNode *, 32> Visited;
7781 SmallVector<const SDNode *, 16> Worklist;
7783 for (SDNode *Use : Ptr.getNode()->uses()) {
7786 if (N->hasPredecessorHelper(Use, Visited, Worklist))
7789 // If Ptr may be folded in addressing mode of other use, then it's
7790 // not profitable to do this transformation.
7791 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
7800 Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
7801 BasePtr, Offset, AM);
7803 Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
7804 BasePtr, Offset, AM);
7807 DEBUG(dbgs() << "\nReplacing.4 ";
7809 dbgs() << "\nWith: ";
7810 Result.getNode()->dump(&DAG);
7812 WorklistRemover DeadNodes(*this);
7814 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
7815 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
7817 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
7820 // Finally, since the node is now dead, remove it from the graph.
7821 deleteAndRecombine(N);
7824 std::swap(BasePtr, Offset);
7826 // Replace other uses of BasePtr that can be updated to use Ptr
7827 for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) {
7828 unsigned OffsetIdx = 1;
7829 if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode())
7831 assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() ==
7832 BasePtr.getNode() && "Expected BasePtr operand");
7834 // We need to replace ptr0 in the following expression:
7835 // x0 * offset0 + y0 * ptr0 = t0
7837 // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store)
7839 // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the
7840 // indexed load/store and the expresion that needs to be re-written.
7842 // Therefore, we have:
7843 // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1
7845 ConstantSDNode *CN =
7846 cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx));
7848 APInt Offset0 = CN->getAPIntValue();
7849 APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue();
7851 X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1;
7852 Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1;
7853 X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1;
7854 Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1;
7856 unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD;
7858 APInt CNV = Offset0;
7859 if (X0 < 0) CNV = -CNV;
7860 if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1;
7861 else CNV = CNV - Offset1;
7863 // We can now generate the new expression.
7864 SDValue NewOp1 = DAG.getConstant(CNV, CN->getValueType(0));
7865 SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0);
7867 SDValue NewUse = DAG.getNode(Opcode,
7868 SDLoc(OtherUses[i]),
7869 OtherUses[i]->getValueType(0), NewOp1, NewOp2);
7870 DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse);
7871 deleteAndRecombine(OtherUses[i]);
7874 // Replace the uses of Ptr with uses of the updated base value.
7875 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
7876 deleteAndRecombine(Ptr.getNode());
7881 /// Try to combine a load/store with a add/sub of the base pointer node into a
7882 /// post-indexed load/store. The transformation folded the add/subtract into the
7883 /// new indexed load/store effectively and all of its uses are redirected to the
7885 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
7886 if (Level < AfterLegalizeDAG)
7892 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7893 if (LD->isIndexed())
7895 VT = LD->getMemoryVT();
7896 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
7897 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
7899 Ptr = LD->getBasePtr();
7900 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7901 if (ST->isIndexed())
7903 VT = ST->getMemoryVT();
7904 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
7905 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
7907 Ptr = ST->getBasePtr();
7913 if (Ptr.getNode()->hasOneUse())
7916 for (SDNode *Op : Ptr.getNode()->uses()) {
7918 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
7923 ISD::MemIndexedMode AM = ISD::UNINDEXED;
7924 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
7925 // Don't create a indexed load / store with zero offset.
7926 if (isa<ConstantSDNode>(Offset) &&
7927 cast<ConstantSDNode>(Offset)->isNullValue())
7930 // Try turning it into a post-indexed load / store except when
7931 // 1) All uses are load / store ops that use it as base ptr (and
7932 // it may be folded as addressing mmode).
7933 // 2) Op must be independent of N, i.e. Op is neither a predecessor
7934 // nor a successor of N. Otherwise, if Op is folded that would
7937 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
7941 bool TryNext = false;
7942 for (SDNode *Use : BasePtr.getNode()->uses()) {
7943 if (Use == Ptr.getNode())
7946 // If all the uses are load / store addresses, then don't do the
7948 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
7949 bool RealUse = false;
7950 for (SDNode *UseUse : Use->uses()) {
7951 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
7966 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
7967 SDValue Result = isLoad
7968 ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
7969 BasePtr, Offset, AM)
7970 : DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
7971 BasePtr, Offset, AM);
7974 DEBUG(dbgs() << "\nReplacing.5 ";
7976 dbgs() << "\nWith: ";
7977 Result.getNode()->dump(&DAG);
7979 WorklistRemover DeadNodes(*this);
7981 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
7982 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
7984 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
7987 // Finally, since the node is now dead, remove it from the graph.
7988 deleteAndRecombine(N);
7990 // Replace the uses of Use with uses of the updated base value.
7991 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
7992 Result.getValue(isLoad ? 1 : 0));
7993 deleteAndRecombine(Op);
8002 /// \brief Return the base-pointer arithmetic from an indexed \p LD.
8003 SDValue DAGCombiner::SplitIndexingFromLoad(LoadSDNode *LD) {
8004 ISD::MemIndexedMode AM = LD->getAddressingMode();
8005 assert(AM != ISD::UNINDEXED);
8006 SDValue BP = LD->getOperand(1);
8007 SDValue Inc = LD->getOperand(2);
8009 // Some backends use TargetConstants for load offsets, but don't expect
8010 // TargetConstants in general ADD nodes. We can convert these constants into
8011 // regular Constants (if the constant is not opaque).
8012 assert((Inc.getOpcode() != ISD::TargetConstant ||
8013 !cast<ConstantSDNode>(Inc)->isOpaque()) &&
8014 "Cannot split out indexing using opaque target constants");
8015 if (Inc.getOpcode() == ISD::TargetConstant) {
8016 ConstantSDNode *ConstInc = cast<ConstantSDNode>(Inc);
8017 Inc = DAG.getConstant(*ConstInc->getConstantIntValue(),
8018 ConstInc->getValueType(0));
8022 (AM == ISD::PRE_INC || AM == ISD::POST_INC ? ISD::ADD : ISD::SUB);
8023 return DAG.getNode(Opc, SDLoc(LD), BP.getSimpleValueType(), BP, Inc);
8026 SDValue DAGCombiner::visitLOAD(SDNode *N) {
8027 LoadSDNode *LD = cast<LoadSDNode>(N);
8028 SDValue Chain = LD->getChain();
8029 SDValue Ptr = LD->getBasePtr();
8031 // If load is not volatile and there are no uses of the loaded value (and
8032 // the updated indexed value in case of indexed loads), change uses of the
8033 // chain value into uses of the chain input (i.e. delete the dead load).
8034 if (!LD->isVolatile()) {
8035 if (N->getValueType(1) == MVT::Other) {
8037 if (!N->hasAnyUseOfValue(0)) {
8038 // It's not safe to use the two value CombineTo variant here. e.g.
8039 // v1, chain2 = load chain1, loc
8040 // v2, chain3 = load chain2, loc
8042 // Now we replace use of chain2 with chain1. This makes the second load
8043 // isomorphic to the one we are deleting, and thus makes this load live.
8044 DEBUG(dbgs() << "\nReplacing.6 ";
8046 dbgs() << "\nWith chain: ";
8047 Chain.getNode()->dump(&DAG);
8049 WorklistRemover DeadNodes(*this);
8050 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
8053 deleteAndRecombine(N);
8055 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8059 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
8061 // If this load has an opaque TargetConstant offset, then we cannot split
8062 // the indexing into an add/sub directly (that TargetConstant may not be
8063 // valid for a different type of node, and we cannot convert an opaque
8064 // target constant into a regular constant).
8065 bool HasOTCInc = LD->getOperand(2).getOpcode() == ISD::TargetConstant &&
8066 cast<ConstantSDNode>(LD->getOperand(2))->isOpaque();
8068 if (!N->hasAnyUseOfValue(0) &&
8069 ((MaySplitLoadIndex && !HasOTCInc) || !N->hasAnyUseOfValue(1))) {
8070 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
8072 if (N->hasAnyUseOfValue(1) && MaySplitLoadIndex && !HasOTCInc) {
8073 Index = SplitIndexingFromLoad(LD);
8074 // Try to fold the base pointer arithmetic into subsequent loads and
8076 AddUsersToWorklist(N);
8078 Index = DAG.getUNDEF(N->getValueType(1));
8079 DEBUG(dbgs() << "\nReplacing.7 ";
8081 dbgs() << "\nWith: ";
8082 Undef.getNode()->dump(&DAG);
8083 dbgs() << " and 2 other values\n");
8084 WorklistRemover DeadNodes(*this);
8085 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
8086 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Index);
8087 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
8088 deleteAndRecombine(N);
8089 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8094 // If this load is directly stored, replace the load value with the stored
8096 // TODO: Handle store large -> read small portion.
8097 // TODO: Handle TRUNCSTORE/LOADEXT
8098 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
8099 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
8100 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
8101 if (PrevST->getBasePtr() == Ptr &&
8102 PrevST->getValue().getValueType() == N->getValueType(0))
8103 return CombineTo(N, Chain.getOperand(1), Chain);
8107 // Try to infer better alignment information than the load already has.
8108 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
8109 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
8110 if (Align > LD->getMemOperand()->getBaseAlignment()) {
8112 DAG.getExtLoad(LD->getExtensionType(), SDLoc(N),
8113 LD->getValueType(0),
8114 Chain, Ptr, LD->getPointerInfo(),
8116 LD->isVolatile(), LD->isNonTemporal(),
8117 LD->isInvariant(), Align, LD->getAAInfo());
8118 return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true);
8123 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA :
8124 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
8126 if (CombinerAAOnlyFunc.getNumOccurrences() &&
8127 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
8130 if (UseAA && LD->isUnindexed()) {
8131 // Walk up chain skipping non-aliasing memory nodes.
8132 SDValue BetterChain = FindBetterChain(N, Chain);
8134 // If there is a better chain.
8135 if (Chain != BetterChain) {
8138 // Replace the chain to void dependency.
8139 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
8140 ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD),
8141 BetterChain, Ptr, LD->getMemOperand());
8143 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD),
8144 LD->getValueType(0),
8145 BetterChain, Ptr, LD->getMemoryVT(),
8146 LD->getMemOperand());
8149 // Create token factor to keep old chain connected.
8150 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
8151 MVT::Other, Chain, ReplLoad.getValue(1));
8153 // Make sure the new and old chains are cleaned up.
8154 AddToWorklist(Token.getNode());
8156 // Replace uses with load result and token factor. Don't add users
8158 return CombineTo(N, ReplLoad.getValue(0), Token, false);
8162 // Try transforming N to an indexed load.
8163 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
8164 return SDValue(N, 0);
8166 // Try to slice up N to more direct loads if the slices are mapped to
8167 // different register banks or pairing can take place.
8169 return SDValue(N, 0);
8175 /// \brief Helper structure used to slice a load in smaller loads.
8176 /// Basically a slice is obtained from the following sequence:
8177 /// Origin = load Ty1, Base
8178 /// Shift = srl Ty1 Origin, CstTy Amount
8179 /// Inst = trunc Shift to Ty2
8181 /// Then, it will be rewriten into:
8182 /// Slice = load SliceTy, Base + SliceOffset
8183 /// [Inst = zext Slice to Ty2], only if SliceTy <> Ty2
8185 /// SliceTy is deduced from the number of bits that are actually used to
8187 struct LoadedSlice {
8188 /// \brief Helper structure used to compute the cost of a slice.
8190 /// Are we optimizing for code size.
8195 unsigned CrossRegisterBanksCopies;
8199 Cost(bool ForCodeSize = false)
8200 : ForCodeSize(ForCodeSize), Loads(0), Truncates(0),
8201 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {}
8203 /// \brief Get the cost of one isolated slice.
8204 Cost(const LoadedSlice &LS, bool ForCodeSize = false)
8205 : ForCodeSize(ForCodeSize), Loads(1), Truncates(0),
8206 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {
8207 EVT TruncType = LS.Inst->getValueType(0);
8208 EVT LoadedType = LS.getLoadedType();
8209 if (TruncType != LoadedType &&
8210 !LS.DAG->getTargetLoweringInfo().isZExtFree(LoadedType, TruncType))
8214 /// \brief Account for slicing gain in the current cost.
8215 /// Slicing provide a few gains like removing a shift or a
8216 /// truncate. This method allows to grow the cost of the original
8217 /// load with the gain from this slice.
8218 void addSliceGain(const LoadedSlice &LS) {
8219 // Each slice saves a truncate.
8220 const TargetLowering &TLI = LS.DAG->getTargetLoweringInfo();
8221 if (!TLI.isTruncateFree(LS.Inst->getValueType(0),
8222 LS.Inst->getOperand(0).getValueType()))
8224 // If there is a shift amount, this slice gets rid of it.
8227 // If this slice can merge a cross register bank copy, account for it.
8228 if (LS.canMergeExpensiveCrossRegisterBankCopy())
8229 ++CrossRegisterBanksCopies;
8232 Cost &operator+=(const Cost &RHS) {
8234 Truncates += RHS.Truncates;
8235 CrossRegisterBanksCopies += RHS.CrossRegisterBanksCopies;
8241 bool operator==(const Cost &RHS) const {
8242 return Loads == RHS.Loads && Truncates == RHS.Truncates &&
8243 CrossRegisterBanksCopies == RHS.CrossRegisterBanksCopies &&
8244 ZExts == RHS.ZExts && Shift == RHS.Shift;
8247 bool operator!=(const Cost &RHS) const { return !(*this == RHS); }
8249 bool operator<(const Cost &RHS) const {
8250 // Assume cross register banks copies are as expensive as loads.
8251 // FIXME: Do we want some more target hooks?
8252 unsigned ExpensiveOpsLHS = Loads + CrossRegisterBanksCopies;
8253 unsigned ExpensiveOpsRHS = RHS.Loads + RHS.CrossRegisterBanksCopies;
8254 // Unless we are optimizing for code size, consider the
8255 // expensive operation first.
8256 if (!ForCodeSize && ExpensiveOpsLHS != ExpensiveOpsRHS)
8257 return ExpensiveOpsLHS < ExpensiveOpsRHS;
8258 return (Truncates + ZExts + Shift + ExpensiveOpsLHS) <
8259 (RHS.Truncates + RHS.ZExts + RHS.Shift + ExpensiveOpsRHS);
8262 bool operator>(const Cost &RHS) const { return RHS < *this; }
8264 bool operator<=(const Cost &RHS) const { return !(RHS < *this); }
8266 bool operator>=(const Cost &RHS) const { return !(*this < RHS); }
8268 // The last instruction that represent the slice. This should be a
8269 // truncate instruction.
8271 // The original load instruction.
8273 // The right shift amount in bits from the original load.
8275 // The DAG from which Origin came from.
8276 // This is used to get some contextual information about legal types, etc.
8279 LoadedSlice(SDNode *Inst = nullptr, LoadSDNode *Origin = nullptr,
8280 unsigned Shift = 0, SelectionDAG *DAG = nullptr)
8281 : Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {}
8283 LoadedSlice(const LoadedSlice &LS)
8284 : Inst(LS.Inst), Origin(LS.Origin), Shift(LS.Shift), DAG(LS.DAG) {}
8286 /// \brief Get the bits used in a chunk of bits \p BitWidth large.
8287 /// \return Result is \p BitWidth and has used bits set to 1 and
8288 /// not used bits set to 0.
8289 APInt getUsedBits() const {
8290 // Reproduce the trunc(lshr) sequence:
8291 // - Start from the truncated value.
8292 // - Zero extend to the desired bit width.
8294 assert(Origin && "No original load to compare against.");
8295 unsigned BitWidth = Origin->getValueSizeInBits(0);
8296 assert(Inst && "This slice is not bound to an instruction");
8297 assert(Inst->getValueSizeInBits(0) <= BitWidth &&
8298 "Extracted slice is bigger than the whole type!");
8299 APInt UsedBits(Inst->getValueSizeInBits(0), 0);
8300 UsedBits.setAllBits();
8301 UsedBits = UsedBits.zext(BitWidth);
8306 /// \brief Get the size of the slice to be loaded in bytes.
8307 unsigned getLoadedSize() const {
8308 unsigned SliceSize = getUsedBits().countPopulation();
8309 assert(!(SliceSize & 0x7) && "Size is not a multiple of a byte.");
8310 return SliceSize / 8;
8313 /// \brief Get the type that will be loaded for this slice.
8314 /// Note: This may not be the final type for the slice.
8315 EVT getLoadedType() const {
8316 assert(DAG && "Missing context");
8317 LLVMContext &Ctxt = *DAG->getContext();
8318 return EVT::getIntegerVT(Ctxt, getLoadedSize() * 8);
8321 /// \brief Get the alignment of the load used for this slice.
8322 unsigned getAlignment() const {
8323 unsigned Alignment = Origin->getAlignment();
8324 unsigned Offset = getOffsetFromBase();
8326 Alignment = MinAlign(Alignment, Alignment + Offset);
8330 /// \brief Check if this slice can be rewritten with legal operations.
8331 bool isLegal() const {
8332 // An invalid slice is not legal.
8333 if (!Origin || !Inst || !DAG)
8336 // Offsets are for indexed load only, we do not handle that.
8337 if (Origin->getOffset().getOpcode() != ISD::UNDEF)
8340 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
8342 // Check that the type is legal.
8343 EVT SliceType = getLoadedType();
8344 if (!TLI.isTypeLegal(SliceType))
8347 // Check that the load is legal for this type.
8348 if (!TLI.isOperationLegal(ISD::LOAD, SliceType))
8351 // Check that the offset can be computed.
8352 // 1. Check its type.
8353 EVT PtrType = Origin->getBasePtr().getValueType();
8354 if (PtrType == MVT::Untyped || PtrType.isExtended())
8357 // 2. Check that it fits in the immediate.
8358 if (!TLI.isLegalAddImmediate(getOffsetFromBase()))
8361 // 3. Check that the computation is legal.
8362 if (!TLI.isOperationLegal(ISD::ADD, PtrType))
8365 // Check that the zext is legal if it needs one.
8366 EVT TruncateType = Inst->getValueType(0);
8367 if (TruncateType != SliceType &&
8368 !TLI.isOperationLegal(ISD::ZERO_EXTEND, TruncateType))
8374 /// \brief Get the offset in bytes of this slice in the original chunk of
8376 /// \pre DAG != nullptr.
8377 uint64_t getOffsetFromBase() const {
8378 assert(DAG && "Missing context.");
8380 DAG->getTargetLoweringInfo().getDataLayout()->isBigEndian();
8381 assert(!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported.");
8382 uint64_t Offset = Shift / 8;
8383 unsigned TySizeInBytes = Origin->getValueSizeInBits(0) / 8;
8384 assert(!(Origin->getValueSizeInBits(0) & 0x7) &&
8385 "The size of the original loaded type is not a multiple of a"
8387 // If Offset is bigger than TySizeInBytes, it means we are loading all
8388 // zeros. This should have been optimized before in the process.
8389 assert(TySizeInBytes > Offset &&
8390 "Invalid shift amount for given loaded size");
8392 Offset = TySizeInBytes - Offset - getLoadedSize();
8396 /// \brief Generate the sequence of instructions to load the slice
8397 /// represented by this object and redirect the uses of this slice to
8398 /// this new sequence of instructions.
8399 /// \pre this->Inst && this->Origin are valid Instructions and this
8400 /// object passed the legal check: LoadedSlice::isLegal returned true.
8401 /// \return The last instruction of the sequence used to load the slice.
8402 SDValue loadSlice() const {
8403 assert(Inst && Origin && "Unable to replace a non-existing slice.");
8404 const SDValue &OldBaseAddr = Origin->getBasePtr();
8405 SDValue BaseAddr = OldBaseAddr;
8406 // Get the offset in that chunk of bytes w.r.t. the endianess.
8407 int64_t Offset = static_cast<int64_t>(getOffsetFromBase());
8408 assert(Offset >= 0 && "Offset too big to fit in int64_t!");
8410 // BaseAddr = BaseAddr + Offset.
8411 EVT ArithType = BaseAddr.getValueType();
8412 BaseAddr = DAG->getNode(ISD::ADD, SDLoc(Origin), ArithType, BaseAddr,
8413 DAG->getConstant(Offset, ArithType));
8416 // Create the type of the loaded slice according to its size.
8417 EVT SliceType = getLoadedType();
8419 // Create the load for the slice.
8420 SDValue LastInst = DAG->getLoad(
8421 SliceType, SDLoc(Origin), Origin->getChain(), BaseAddr,
8422 Origin->getPointerInfo().getWithOffset(Offset), Origin->isVolatile(),
8423 Origin->isNonTemporal(), Origin->isInvariant(), getAlignment());
8424 // If the final type is not the same as the loaded type, this means that
8425 // we have to pad with zero. Create a zero extend for that.
8426 EVT FinalType = Inst->getValueType(0);
8427 if (SliceType != FinalType)
8429 DAG->getNode(ISD::ZERO_EXTEND, SDLoc(LastInst), FinalType, LastInst);
8433 /// \brief Check if this slice can be merged with an expensive cross register
8434 /// bank copy. E.g.,
8436 /// f = bitcast i32 i to float
8437 bool canMergeExpensiveCrossRegisterBankCopy() const {
8438 if (!Inst || !Inst->hasOneUse())
8440 SDNode *Use = *Inst->use_begin();
8441 if (Use->getOpcode() != ISD::BITCAST)
8443 assert(DAG && "Missing context");
8444 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
8445 EVT ResVT = Use->getValueType(0);
8446 const TargetRegisterClass *ResRC = TLI.getRegClassFor(ResVT.getSimpleVT());
8447 const TargetRegisterClass *ArgRC =
8448 TLI.getRegClassFor(Use->getOperand(0).getValueType().getSimpleVT());
8449 if (ArgRC == ResRC || !TLI.isOperationLegal(ISD::LOAD, ResVT))
8452 // At this point, we know that we perform a cross-register-bank copy.
8453 // Check if it is expensive.
8454 const TargetRegisterInfo *TRI =
8455 TLI.getTargetMachine().getSubtargetImpl()->getRegisterInfo();
8456 // Assume bitcasts are cheap, unless both register classes do not
8457 // explicitly share a common sub class.
8458 if (!TRI || TRI->getCommonSubClass(ArgRC, ResRC))
8461 // Check if it will be merged with the load.
8462 // 1. Check the alignment constraint.
8463 unsigned RequiredAlignment = TLI.getDataLayout()->getABITypeAlignment(
8464 ResVT.getTypeForEVT(*DAG->getContext()));
8466 if (RequiredAlignment > getAlignment())
8469 // 2. Check that the load is a legal operation for that type.
8470 if (!TLI.isOperationLegal(ISD::LOAD, ResVT))
8473 // 3. Check that we do not have a zext in the way.
8474 if (Inst->getValueType(0) != getLoadedType())
8482 /// \brief Check that all bits set in \p UsedBits form a dense region, i.e.,
8483 /// \p UsedBits looks like 0..0 1..1 0..0.
8484 static bool areUsedBitsDense(const APInt &UsedBits) {
8485 // If all the bits are one, this is dense!
8486 if (UsedBits.isAllOnesValue())
8489 // Get rid of the unused bits on the right.
8490 APInt NarrowedUsedBits = UsedBits.lshr(UsedBits.countTrailingZeros());
8491 // Get rid of the unused bits on the left.
8492 if (NarrowedUsedBits.countLeadingZeros())
8493 NarrowedUsedBits = NarrowedUsedBits.trunc(NarrowedUsedBits.getActiveBits());
8494 // Check that the chunk of bits is completely used.
8495 return NarrowedUsedBits.isAllOnesValue();
8498 /// \brief Check whether or not \p First and \p Second are next to each other
8499 /// in memory. This means that there is no hole between the bits loaded
8500 /// by \p First and the bits loaded by \p Second.
8501 static bool areSlicesNextToEachOther(const LoadedSlice &First,
8502 const LoadedSlice &Second) {
8503 assert(First.Origin == Second.Origin && First.Origin &&
8504 "Unable to match different memory origins.");
8505 APInt UsedBits = First.getUsedBits();
8506 assert((UsedBits & Second.getUsedBits()) == 0 &&
8507 "Slices are not supposed to overlap.");
8508 UsedBits |= Second.getUsedBits();
8509 return areUsedBitsDense(UsedBits);
8512 /// \brief Adjust the \p GlobalLSCost according to the target
8513 /// paring capabilities and the layout of the slices.
8514 /// \pre \p GlobalLSCost should account for at least as many loads as
8515 /// there is in the slices in \p LoadedSlices.
8516 static void adjustCostForPairing(SmallVectorImpl<LoadedSlice> &LoadedSlices,
8517 LoadedSlice::Cost &GlobalLSCost) {
8518 unsigned NumberOfSlices = LoadedSlices.size();
8519 // If there is less than 2 elements, no pairing is possible.
8520 if (NumberOfSlices < 2)
8523 // Sort the slices so that elements that are likely to be next to each
8524 // other in memory are next to each other in the list.
8525 std::sort(LoadedSlices.begin(), LoadedSlices.end(),
8526 [](const LoadedSlice &LHS, const LoadedSlice &RHS) {
8527 assert(LHS.Origin == RHS.Origin && "Different bases not implemented.");
8528 return LHS.getOffsetFromBase() < RHS.getOffsetFromBase();
8530 const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo();
8531 // First (resp. Second) is the first (resp. Second) potentially candidate
8532 // to be placed in a paired load.
8533 const LoadedSlice *First = nullptr;
8534 const LoadedSlice *Second = nullptr;
8535 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice,
8536 // Set the beginning of the pair.
8539 Second = &LoadedSlices[CurrSlice];
8541 // If First is NULL, it means we start a new pair.
8542 // Get to the next slice.
8546 EVT LoadedType = First->getLoadedType();
8548 // If the types of the slices are different, we cannot pair them.
8549 if (LoadedType != Second->getLoadedType())
8552 // Check if the target supplies paired loads for this type.
8553 unsigned RequiredAlignment = 0;
8554 if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) {
8555 // move to the next pair, this type is hopeless.
8559 // Check if we meet the alignment requirement.
8560 if (RequiredAlignment > First->getAlignment())
8563 // Check that both loads are next to each other in memory.
8564 if (!areSlicesNextToEachOther(*First, *Second))
8567 assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!");
8568 --GlobalLSCost.Loads;
8569 // Move to the next pair.
8574 /// \brief Check the profitability of all involved LoadedSlice.
8575 /// Currently, it is considered profitable if there is exactly two
8576 /// involved slices (1) which are (2) next to each other in memory, and
8577 /// whose cost (\see LoadedSlice::Cost) is smaller than the original load (3).
8579 /// Note: The order of the elements in \p LoadedSlices may be modified, but not
8580 /// the elements themselves.
8582 /// FIXME: When the cost model will be mature enough, we can relax
8583 /// constraints (1) and (2).
8584 static bool isSlicingProfitable(SmallVectorImpl<LoadedSlice> &LoadedSlices,
8585 const APInt &UsedBits, bool ForCodeSize) {
8586 unsigned NumberOfSlices = LoadedSlices.size();
8587 if (StressLoadSlicing)
8588 return NumberOfSlices > 1;
8591 if (NumberOfSlices != 2)
8595 if (!areUsedBitsDense(UsedBits))
8599 LoadedSlice::Cost OrigCost(ForCodeSize), GlobalSlicingCost(ForCodeSize);
8600 // The original code has one big load.
8602 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice) {
8603 const LoadedSlice &LS = LoadedSlices[CurrSlice];
8604 // Accumulate the cost of all the slices.
8605 LoadedSlice::Cost SliceCost(LS, ForCodeSize);
8606 GlobalSlicingCost += SliceCost;
8608 // Account as cost in the original configuration the gain obtained
8609 // with the current slices.
8610 OrigCost.addSliceGain(LS);
8613 // If the target supports paired load, adjust the cost accordingly.
8614 adjustCostForPairing(LoadedSlices, GlobalSlicingCost);
8615 return OrigCost > GlobalSlicingCost;
8618 /// \brief If the given load, \p LI, is used only by trunc or trunc(lshr)
8619 /// operations, split it in the various pieces being extracted.
8621 /// This sort of thing is introduced by SROA.
8622 /// This slicing takes care not to insert overlapping loads.
8623 /// \pre LI is a simple load (i.e., not an atomic or volatile load).
8624 bool DAGCombiner::SliceUpLoad(SDNode *N) {
8625 if (Level < AfterLegalizeDAG)
8628 LoadSDNode *LD = cast<LoadSDNode>(N);
8629 if (LD->isVolatile() || !ISD::isNormalLoad(LD) ||
8630 !LD->getValueType(0).isInteger())
8633 // Keep track of already used bits to detect overlapping values.
8634 // In that case, we will just abort the transformation.
8635 APInt UsedBits(LD->getValueSizeInBits(0), 0);
8637 SmallVector<LoadedSlice, 4> LoadedSlices;
8639 // Check if this load is used as several smaller chunks of bits.
8640 // Basically, look for uses in trunc or trunc(lshr) and record a new chain
8641 // of computation for each trunc.
8642 for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end();
8643 UI != UIEnd; ++UI) {
8644 // Skip the uses of the chain.
8645 if (UI.getUse().getResNo() != 0)
8651 // Check if this is a trunc(lshr).
8652 if (User->getOpcode() == ISD::SRL && User->hasOneUse() &&
8653 isa<ConstantSDNode>(User->getOperand(1))) {
8654 Shift = cast<ConstantSDNode>(User->getOperand(1))->getZExtValue();
8655 User = *User->use_begin();
8658 // At this point, User is a Truncate, iff we encountered, trunc or
8660 if (User->getOpcode() != ISD::TRUNCATE)
8663 // The width of the type must be a power of 2 and greater than 8-bits.
8664 // Otherwise the load cannot be represented in LLVM IR.
8665 // Moreover, if we shifted with a non-8-bits multiple, the slice
8666 // will be across several bytes. We do not support that.
8667 unsigned Width = User->getValueSizeInBits(0);
8668 if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7))
8671 // Build the slice for this chain of computations.
8672 LoadedSlice LS(User, LD, Shift, &DAG);
8673 APInt CurrentUsedBits = LS.getUsedBits();
8675 // Check if this slice overlaps with another.
8676 if ((CurrentUsedBits & UsedBits) != 0)
8678 // Update the bits used globally.
8679 UsedBits |= CurrentUsedBits;
8681 // Check if the new slice would be legal.
8685 // Record the slice.
8686 LoadedSlices.push_back(LS);
8689 // Abort slicing if it does not seem to be profitable.
8690 if (!isSlicingProfitable(LoadedSlices, UsedBits, ForCodeSize))
8695 // Rewrite each chain to use an independent load.
8696 // By construction, each chain can be represented by a unique load.
8698 // Prepare the argument for the new token factor for all the slices.
8699 SmallVector<SDValue, 8> ArgChains;
8700 for (SmallVectorImpl<LoadedSlice>::const_iterator
8701 LSIt = LoadedSlices.begin(),
8702 LSItEnd = LoadedSlices.end();
8703 LSIt != LSItEnd; ++LSIt) {
8704 SDValue SliceInst = LSIt->loadSlice();
8705 CombineTo(LSIt->Inst, SliceInst, true);
8706 if (SliceInst.getNode()->getOpcode() != ISD::LOAD)
8707 SliceInst = SliceInst.getOperand(0);
8708 assert(SliceInst->getOpcode() == ISD::LOAD &&
8709 "It takes more than a zext to get to the loaded slice!!");
8710 ArgChains.push_back(SliceInst.getValue(1));
8713 SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other,
8715 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
8719 /// Check to see if V is (and load (ptr), imm), where the load is having
8720 /// specific bytes cleared out. If so, return the byte size being masked out
8721 /// and the shift amount.
8722 static std::pair<unsigned, unsigned>
8723 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
8724 std::pair<unsigned, unsigned> Result(0, 0);
8726 // Check for the structure we're looking for.
8727 if (V->getOpcode() != ISD::AND ||
8728 !isa<ConstantSDNode>(V->getOperand(1)) ||
8729 !ISD::isNormalLoad(V->getOperand(0).getNode()))
8732 // Check the chain and pointer.
8733 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
8734 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
8736 // The store should be chained directly to the load or be an operand of a
8738 if (LD == Chain.getNode())
8740 else if (Chain->getOpcode() != ISD::TokenFactor)
8741 return Result; // Fail.
8744 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
8745 if (Chain->getOperand(i).getNode() == LD) {
8749 if (!isOk) return Result;
8752 // This only handles simple types.
8753 if (V.getValueType() != MVT::i16 &&
8754 V.getValueType() != MVT::i32 &&
8755 V.getValueType() != MVT::i64)
8758 // Check the constant mask. Invert it so that the bits being masked out are
8759 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
8760 // follow the sign bit for uniformity.
8761 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
8762 unsigned NotMaskLZ = countLeadingZeros(NotMask);
8763 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
8764 unsigned NotMaskTZ = countTrailingZeros(NotMask);
8765 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
8766 if (NotMaskLZ == 64) return Result; // All zero mask.
8768 // See if we have a continuous run of bits. If so, we have 0*1+0*
8769 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
8772 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
8773 if (V.getValueType() != MVT::i64 && NotMaskLZ)
8774 NotMaskLZ -= 64-V.getValueSizeInBits();
8776 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
8777 switch (MaskedBytes) {
8781 default: return Result; // All one mask, or 5-byte mask.
8784 // Verify that the first bit starts at a multiple of mask so that the access
8785 // is aligned the same as the access width.
8786 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
8788 Result.first = MaskedBytes;
8789 Result.second = NotMaskTZ/8;
8794 /// Check to see if IVal is something that provides a value as specified by
8795 /// MaskInfo. If so, replace the specified store with a narrower store of
8798 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
8799 SDValue IVal, StoreSDNode *St,
8801 unsigned NumBytes = MaskInfo.first;
8802 unsigned ByteShift = MaskInfo.second;
8803 SelectionDAG &DAG = DC->getDAG();
8805 // Check to see if IVal is all zeros in the part being masked in by the 'or'
8806 // that uses this. If not, this is not a replacement.
8807 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
8808 ByteShift*8, (ByteShift+NumBytes)*8);
8809 if (!DAG.MaskedValueIsZero(IVal, Mask)) return nullptr;
8811 // Check that it is legal on the target to do this. It is legal if the new
8812 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
8814 MVT VT = MVT::getIntegerVT(NumBytes*8);
8815 if (!DC->isTypeLegal(VT))
8818 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
8819 // shifted by ByteShift and truncated down to NumBytes.
8821 IVal = DAG.getNode(ISD::SRL, SDLoc(IVal), IVal.getValueType(), IVal,
8822 DAG.getConstant(ByteShift*8,
8823 DC->getShiftAmountTy(IVal.getValueType())));
8825 // Figure out the offset for the store and the alignment of the access.
8827 unsigned NewAlign = St->getAlignment();
8829 if (DAG.getTargetLoweringInfo().isLittleEndian())
8830 StOffset = ByteShift;
8832 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
8834 SDValue Ptr = St->getBasePtr();
8836 Ptr = DAG.getNode(ISD::ADD, SDLoc(IVal), Ptr.getValueType(),
8837 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
8838 NewAlign = MinAlign(NewAlign, StOffset);
8841 // Truncate down to the new size.
8842 IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal);
8845 return DAG.getStore(St->getChain(), SDLoc(St), IVal, Ptr,
8846 St->getPointerInfo().getWithOffset(StOffset),
8847 false, false, NewAlign).getNode();
8851 /// Look for sequence of load / op / store where op is one of 'or', 'xor', and
8852 /// 'and' of immediates. If 'op' is only touching some of the loaded bits, try
8853 /// narrowing the load and store if it would end up being a win for performance
8855 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
8856 StoreSDNode *ST = cast<StoreSDNode>(N);
8857 if (ST->isVolatile())
8860 SDValue Chain = ST->getChain();
8861 SDValue Value = ST->getValue();
8862 SDValue Ptr = ST->getBasePtr();
8863 EVT VT = Value.getValueType();
8865 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
8868 unsigned Opc = Value.getOpcode();
8870 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
8871 // is a byte mask indicating a consecutive number of bytes, check to see if
8872 // Y is known to provide just those bytes. If so, we try to replace the
8873 // load + replace + store sequence with a single (narrower) store, which makes
8875 if (Opc == ISD::OR) {
8876 std::pair<unsigned, unsigned> MaskedLoad;
8877 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
8878 if (MaskedLoad.first)
8879 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
8880 Value.getOperand(1), ST,this))
8881 return SDValue(NewST, 0);
8883 // Or is commutative, so try swapping X and Y.
8884 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
8885 if (MaskedLoad.first)
8886 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
8887 Value.getOperand(0), ST,this))
8888 return SDValue(NewST, 0);
8891 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
8892 Value.getOperand(1).getOpcode() != ISD::Constant)
8895 SDValue N0 = Value.getOperand(0);
8896 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
8897 Chain == SDValue(N0.getNode(), 1)) {
8898 LoadSDNode *LD = cast<LoadSDNode>(N0);
8899 if (LD->getBasePtr() != Ptr ||
8900 LD->getPointerInfo().getAddrSpace() !=
8901 ST->getPointerInfo().getAddrSpace())
8904 // Find the type to narrow it the load / op / store to.
8905 SDValue N1 = Value.getOperand(1);
8906 unsigned BitWidth = N1.getValueSizeInBits();
8907 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
8908 if (Opc == ISD::AND)
8909 Imm ^= APInt::getAllOnesValue(BitWidth);
8910 if (Imm == 0 || Imm.isAllOnesValue())
8912 unsigned ShAmt = Imm.countTrailingZeros();
8913 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
8914 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
8915 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
8916 while (NewBW < BitWidth &&
8917 !(TLI.isOperationLegalOrCustom(Opc, NewVT) &&
8918 TLI.isNarrowingProfitable(VT, NewVT))) {
8919 NewBW = NextPowerOf2(NewBW);
8920 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
8922 if (NewBW >= BitWidth)
8925 // If the lsb changed does not start at the type bitwidth boundary,
8926 // start at the previous one.
8928 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
8929 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt,
8930 std::min(BitWidth, ShAmt + NewBW));
8931 if ((Imm & Mask) == Imm) {
8932 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
8933 if (Opc == ISD::AND)
8934 NewImm ^= APInt::getAllOnesValue(NewBW);
8935 uint64_t PtrOff = ShAmt / 8;
8936 // For big endian targets, we need to adjust the offset to the pointer to
8937 // load the correct bytes.
8938 if (TLI.isBigEndian())
8939 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
8941 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
8942 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
8943 if (NewAlign < TLI.getDataLayout()->getABITypeAlignment(NewVTTy))
8946 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD),
8947 Ptr.getValueType(), Ptr,
8948 DAG.getConstant(PtrOff, Ptr.getValueType()));
8949 SDValue NewLD = DAG.getLoad(NewVT, SDLoc(N0),
8950 LD->getChain(), NewPtr,
8951 LD->getPointerInfo().getWithOffset(PtrOff),
8952 LD->isVolatile(), LD->isNonTemporal(),
8953 LD->isInvariant(), NewAlign,
8955 SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD,
8956 DAG.getConstant(NewImm, NewVT));
8957 SDValue NewST = DAG.getStore(Chain, SDLoc(N),
8959 ST->getPointerInfo().getWithOffset(PtrOff),
8960 false, false, NewAlign);
8962 AddToWorklist(NewPtr.getNode());
8963 AddToWorklist(NewLD.getNode());
8964 AddToWorklist(NewVal.getNode());
8965 WorklistRemover DeadNodes(*this);
8966 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
8975 /// For a given floating point load / store pair, if the load value isn't used
8976 /// by any other operations, then consider transforming the pair to integer
8977 /// load / store operations if the target deems the transformation profitable.
8978 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
8979 StoreSDNode *ST = cast<StoreSDNode>(N);
8980 SDValue Chain = ST->getChain();
8981 SDValue Value = ST->getValue();
8982 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
8983 Value.hasOneUse() &&
8984 Chain == SDValue(Value.getNode(), 1)) {
8985 LoadSDNode *LD = cast<LoadSDNode>(Value);
8986 EVT VT = LD->getMemoryVT();
8987 if (!VT.isFloatingPoint() ||
8988 VT != ST->getMemoryVT() ||
8989 LD->isNonTemporal() ||
8990 ST->isNonTemporal() ||
8991 LD->getPointerInfo().getAddrSpace() != 0 ||
8992 ST->getPointerInfo().getAddrSpace() != 0)
8995 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
8996 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
8997 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
8998 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
8999 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
9002 unsigned LDAlign = LD->getAlignment();
9003 unsigned STAlign = ST->getAlignment();
9004 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
9005 unsigned ABIAlign = TLI.getDataLayout()->getABITypeAlignment(IntVTTy);
9006 if (LDAlign < ABIAlign || STAlign < ABIAlign)
9009 SDValue NewLD = DAG.getLoad(IntVT, SDLoc(Value),
9010 LD->getChain(), LD->getBasePtr(),
9011 LD->getPointerInfo(),
9012 false, false, false, LDAlign);
9014 SDValue NewST = DAG.getStore(NewLD.getValue(1), SDLoc(N),
9015 NewLD, ST->getBasePtr(),
9016 ST->getPointerInfo(),
9017 false, false, STAlign);
9019 AddToWorklist(NewLD.getNode());
9020 AddToWorklist(NewST.getNode());
9021 WorklistRemover DeadNodes(*this);
9022 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
9030 /// Helper struct to parse and store a memory address as base + index + offset.
9031 /// We ignore sign extensions when it is safe to do so.
9032 /// The following two expressions are not equivalent. To differentiate we need
9033 /// to store whether there was a sign extension involved in the index
9035 /// (load (i64 add (i64 copyfromreg %c)
9036 /// (i64 signextend (add (i8 load %index)
9040 /// (load (i64 add (i64 copyfromreg %c)
9041 /// (i64 signextend (i32 add (i32 signextend (i8 load %index))
9043 struct BaseIndexOffset {
9047 bool IsIndexSignExt;
9049 BaseIndexOffset() : Offset(0), IsIndexSignExt(false) {}
9051 BaseIndexOffset(SDValue Base, SDValue Index, int64_t Offset,
9052 bool IsIndexSignExt) :
9053 Base(Base), Index(Index), Offset(Offset), IsIndexSignExt(IsIndexSignExt) {}
9055 bool equalBaseIndex(const BaseIndexOffset &Other) {
9056 return Other.Base == Base && Other.Index == Index &&
9057 Other.IsIndexSignExt == IsIndexSignExt;
9060 /// Parses tree in Ptr for base, index, offset addresses.
9061 static BaseIndexOffset match(SDValue Ptr) {
9062 bool IsIndexSignExt = false;
9064 // We only can pattern match BASE + INDEX + OFFSET. If Ptr is not an ADD
9065 // instruction, then it could be just the BASE or everything else we don't
9066 // know how to handle. Just use Ptr as BASE and give up.
9067 if (Ptr->getOpcode() != ISD::ADD)
9068 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
9070 // We know that we have at least an ADD instruction. Try to pattern match
9071 // the simple case of BASE + OFFSET.
9072 if (isa<ConstantSDNode>(Ptr->getOperand(1))) {
9073 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
9074 return BaseIndexOffset(Ptr->getOperand(0), SDValue(), Offset,
9078 // Inside a loop the current BASE pointer is calculated using an ADD and a
9079 // MUL instruction. In this case Ptr is the actual BASE pointer.
9080 // (i64 add (i64 %array_ptr)
9081 // (i64 mul (i64 %induction_var)
9082 // (i64 %element_size)))
9083 if (Ptr->getOperand(1)->getOpcode() == ISD::MUL)
9084 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
9086 // Look at Base + Index + Offset cases.
9087 SDValue Base = Ptr->getOperand(0);
9088 SDValue IndexOffset = Ptr->getOperand(1);
9090 // Skip signextends.
9091 if (IndexOffset->getOpcode() == ISD::SIGN_EXTEND) {
9092 IndexOffset = IndexOffset->getOperand(0);
9093 IsIndexSignExt = true;
9096 // Either the case of Base + Index (no offset) or something else.
9097 if (IndexOffset->getOpcode() != ISD::ADD)
9098 return BaseIndexOffset(Base, IndexOffset, 0, IsIndexSignExt);
9100 // Now we have the case of Base + Index + offset.
9101 SDValue Index = IndexOffset->getOperand(0);
9102 SDValue Offset = IndexOffset->getOperand(1);
9104 if (!isa<ConstantSDNode>(Offset))
9105 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
9107 // Ignore signextends.
9108 if (Index->getOpcode() == ISD::SIGN_EXTEND) {
9109 Index = Index->getOperand(0);
9110 IsIndexSignExt = true;
9111 } else IsIndexSignExt = false;
9113 int64_t Off = cast<ConstantSDNode>(Offset)->getSExtValue();
9114 return BaseIndexOffset(Base, Index, Off, IsIndexSignExt);
9118 /// Holds a pointer to an LSBaseSDNode as well as information on where it
9119 /// is located in a sequence of memory operations connected by a chain.
9121 MemOpLink (LSBaseSDNode *N, int64_t Offset, unsigned Seq):
9122 MemNode(N), OffsetFromBase(Offset), SequenceNum(Seq) { }
9123 // Ptr to the mem node.
9124 LSBaseSDNode *MemNode;
9125 // Offset from the base ptr.
9126 int64_t OffsetFromBase;
9127 // What is the sequence number of this mem node.
9128 // Lowest mem operand in the DAG starts at zero.
9129 unsigned SequenceNum;
9132 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
9133 EVT MemVT = St->getMemoryVT();
9134 int64_t ElementSizeBytes = MemVT.getSizeInBits()/8;
9135 bool NoVectors = DAG.getMachineFunction().getFunction()->getAttributes().
9136 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
9138 // Don't merge vectors into wider inputs.
9139 if (MemVT.isVector() || !MemVT.isSimple())
9142 // Perform an early exit check. Do not bother looking at stored values that
9143 // are not constants or loads.
9144 SDValue StoredVal = St->getValue();
9145 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
9146 if (!isa<ConstantSDNode>(StoredVal) && !isa<ConstantFPSDNode>(StoredVal) &&
9150 // Only look at ends of store sequences.
9151 SDValue Chain = SDValue(St, 0);
9152 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
9155 // This holds the base pointer, index, and the offset in bytes from the base
9157 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr());
9159 // We must have a base and an offset.
9160 if (!BasePtr.Base.getNode())
9163 // Do not handle stores to undef base pointers.
9164 if (BasePtr.Base.getOpcode() == ISD::UNDEF)
9167 // Save the LoadSDNodes that we find in the chain.
9168 // We need to make sure that these nodes do not interfere with
9169 // any of the store nodes.
9170 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
9172 // Save the StoreSDNodes that we find in the chain.
9173 SmallVector<MemOpLink, 8> StoreNodes;
9175 // Walk up the chain and look for nodes with offsets from the same
9176 // base pointer. Stop when reaching an instruction with a different kind
9177 // or instruction which has a different base pointer.
9179 StoreSDNode *Index = St;
9181 // If the chain has more than one use, then we can't reorder the mem ops.
9182 if (Index != St && !SDValue(Index, 0)->hasOneUse())
9185 // Find the base pointer and offset for this memory node.
9186 BaseIndexOffset Ptr = BaseIndexOffset::match(Index->getBasePtr());
9188 // Check that the base pointer is the same as the original one.
9189 if (!Ptr.equalBaseIndex(BasePtr))
9192 // Check that the alignment is the same.
9193 if (Index->getAlignment() != St->getAlignment())
9196 // The memory operands must not be volatile.
9197 if (Index->isVolatile() || Index->isIndexed())
9201 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
9202 if (St->isTruncatingStore())
9205 // The stored memory type must be the same.
9206 if (Index->getMemoryVT() != MemVT)
9209 // We do not allow unaligned stores because we want to prevent overriding
9211 if (Index->getAlignment()*8 != MemVT.getSizeInBits())
9214 // We found a potential memory operand to merge.
9215 StoreNodes.push_back(MemOpLink(Index, Ptr.Offset, Seq++));
9217 // Find the next memory operand in the chain. If the next operand in the
9218 // chain is a store then move up and continue the scan with the next
9219 // memory operand. If the next operand is a load save it and use alias
9220 // information to check if it interferes with anything.
9221 SDNode *NextInChain = Index->getChain().getNode();
9223 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
9224 // We found a store node. Use it for the next iteration.
9227 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
9228 if (Ldn->isVolatile()) {
9233 // Save the load node for later. Continue the scan.
9234 AliasLoadNodes.push_back(Ldn);
9235 NextInChain = Ldn->getChain().getNode();
9244 // Check if there is anything to merge.
9245 if (StoreNodes.size() < 2)
9248 // Sort the memory operands according to their distance from the base pointer.
9249 std::sort(StoreNodes.begin(), StoreNodes.end(),
9250 [](MemOpLink LHS, MemOpLink RHS) {
9251 return LHS.OffsetFromBase < RHS.OffsetFromBase ||
9252 (LHS.OffsetFromBase == RHS.OffsetFromBase &&
9253 LHS.SequenceNum > RHS.SequenceNum);
9256 // Scan the memory operations on the chain and find the first non-consecutive
9257 // store memory address.
9258 unsigned LastConsecutiveStore = 0;
9259 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
9260 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) {
9262 // Check that the addresses are consecutive starting from the second
9263 // element in the list of stores.
9265 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
9266 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
9271 // Check if this store interferes with any of the loads that we found.
9272 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld)
9273 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) {
9277 // We found a load that alias with this store. Stop the sequence.
9281 // Mark this node as useful.
9282 LastConsecutiveStore = i;
9285 // The node with the lowest store address.
9286 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
9288 // Store the constants into memory as one consecutive store.
9290 unsigned LastLegalType = 0;
9291 unsigned LastLegalVectorType = 0;
9292 bool NonZero = false;
9293 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
9294 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9295 SDValue StoredVal = St->getValue();
9297 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
9298 NonZero |= !C->isNullValue();
9299 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) {
9300 NonZero |= !C->getConstantFPValue()->isNullValue();
9306 // Find a legal type for the constant store.
9307 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
9308 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9309 if (TLI.isTypeLegal(StoreTy))
9310 LastLegalType = i+1;
9311 // Or check whether a truncstore is legal.
9312 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
9313 TargetLowering::TypePromoteInteger) {
9314 EVT LegalizedStoredValueTy =
9315 TLI.getTypeToTransformTo(*DAG.getContext(), StoredVal.getValueType());
9316 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy))
9317 LastLegalType = i+1;
9320 // Find a legal type for the vector store.
9321 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
9322 if (TLI.isTypeLegal(Ty))
9323 LastLegalVectorType = i + 1;
9326 // We only use vectors if the constant is known to be zero and the
9327 // function is not marked with the noimplicitfloat attribute.
9328 if (NonZero || NoVectors)
9329 LastLegalVectorType = 0;
9331 // Check if we found a legal integer type to store.
9332 if (LastLegalType == 0 && LastLegalVectorType == 0)
9335 bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors;
9336 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType;
9338 // Make sure we have something to merge.
9342 unsigned EarliestNodeUsed = 0;
9343 for (unsigned i=0; i < NumElem; ++i) {
9344 // Find a chain for the new wide-store operand. Notice that some
9345 // of the store nodes that we found may not be selected for inclusion
9346 // in the wide store. The chain we use needs to be the chain of the
9347 // earliest store node which is *used* and replaced by the wide store.
9348 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
9349 EarliestNodeUsed = i;
9352 // The earliest Node in the DAG.
9353 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
9354 SDLoc DL(StoreNodes[0].MemNode);
9358 // Find a legal type for the vector store.
9359 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
9360 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
9361 StoredVal = DAG.getConstant(0, Ty);
9363 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
9364 APInt StoreInt(StoreBW, 0);
9366 // Construct a single integer constant which is made of the smaller
9368 bool IsLE = TLI.isLittleEndian();
9369 for (unsigned i = 0; i < NumElem ; ++i) {
9370 unsigned Idx = IsLE ?(NumElem - 1 - i) : i;
9371 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
9372 SDValue Val = St->getValue();
9373 StoreInt<<=ElementSizeBytes*8;
9374 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
9375 StoreInt|=C->getAPIntValue().zext(StoreBW);
9376 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
9377 StoreInt|= C->getValueAPF().bitcastToAPInt().zext(StoreBW);
9379 assert(false && "Invalid constant element type");
9383 // Create the new Load and Store operations.
9384 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9385 StoredVal = DAG.getConstant(StoreInt, StoreTy);
9388 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), DL, StoredVal,
9389 FirstInChain->getBasePtr(),
9390 FirstInChain->getPointerInfo(),
9392 FirstInChain->getAlignment());
9394 // Replace the first store with the new store
9395 CombineTo(EarliestOp, NewStore);
9396 // Erase all other stores.
9397 for (unsigned i = 0; i < NumElem ; ++i) {
9398 if (StoreNodes[i].MemNode == EarliestOp)
9400 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9401 // ReplaceAllUsesWith will replace all uses that existed when it was
9402 // called, but graph optimizations may cause new ones to appear. For
9403 // example, the case in pr14333 looks like
9405 // St's chain -> St -> another store -> X
9407 // And the only difference from St to the other store is the chain.
9408 // When we change it's chain to be St's chain they become identical,
9409 // get CSEed and the net result is that X is now a use of St.
9410 // Since we know that St is redundant, just iterate.
9411 while (!St->use_empty())
9412 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain());
9413 deleteAndRecombine(St);
9419 // Below we handle the case of multiple consecutive stores that
9420 // come from multiple consecutive loads. We merge them into a single
9421 // wide load and a single wide store.
9423 // Look for load nodes which are used by the stored values.
9424 SmallVector<MemOpLink, 8> LoadNodes;
9426 // Find acceptable loads. Loads need to have the same chain (token factor),
9427 // must not be zext, volatile, indexed, and they must be consecutive.
9428 BaseIndexOffset LdBasePtr;
9429 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
9430 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9431 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
9434 // Loads must only have one use.
9435 if (!Ld->hasNUsesOfValue(1, 0))
9438 // Check that the alignment is the same as the stores.
9439 if (Ld->getAlignment() != St->getAlignment())
9442 // The memory operands must not be volatile.
9443 if (Ld->isVolatile() || Ld->isIndexed())
9446 // We do not accept ext loads.
9447 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
9450 // The stored memory type must be the same.
9451 if (Ld->getMemoryVT() != MemVT)
9454 BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld->getBasePtr());
9455 // If this is not the first ptr that we check.
9456 if (LdBasePtr.Base.getNode()) {
9457 // The base ptr must be the same.
9458 if (!LdPtr.equalBaseIndex(LdBasePtr))
9461 // Check that all other base pointers are the same as this one.
9465 // We found a potential memory operand to merge.
9466 LoadNodes.push_back(MemOpLink(Ld, LdPtr.Offset, 0));
9469 if (LoadNodes.size() < 2)
9472 // If we have load/store pair instructions and we only have two values,
9474 unsigned RequiredAlignment;
9475 if (LoadNodes.size() == 2 && TLI.hasPairedLoad(MemVT, RequiredAlignment) &&
9476 St->getAlignment() >= RequiredAlignment)
9479 // Scan the memory operations on the chain and find the first non-consecutive
9480 // load memory address. These variables hold the index in the store node
9482 unsigned LastConsecutiveLoad = 0;
9483 // This variable refers to the size and not index in the array.
9484 unsigned LastLegalVectorType = 0;
9485 unsigned LastLegalIntegerType = 0;
9486 StartAddress = LoadNodes[0].OffsetFromBase;
9487 SDValue FirstChain = LoadNodes[0].MemNode->getChain();
9488 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
9489 // All loads much share the same chain.
9490 if (LoadNodes[i].MemNode->getChain() != FirstChain)
9493 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
9494 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
9496 LastConsecutiveLoad = i;
9498 // Find a legal type for the vector store.
9499 EVT StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
9500 if (TLI.isTypeLegal(StoreTy))
9501 LastLegalVectorType = i + 1;
9503 // Find a legal type for the integer store.
9504 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
9505 StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9506 if (TLI.isTypeLegal(StoreTy))
9507 LastLegalIntegerType = i + 1;
9508 // Or check whether a truncstore and extload is legal.
9509 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
9510 TargetLowering::TypePromoteInteger) {
9511 EVT LegalizedStoredValueTy =
9512 TLI.getTypeToTransformTo(*DAG.getContext(), StoreTy);
9513 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
9514 TLI.isLoadExtLegal(ISD::ZEXTLOAD, StoreTy) &&
9515 TLI.isLoadExtLegal(ISD::SEXTLOAD, StoreTy) &&
9516 TLI.isLoadExtLegal(ISD::EXTLOAD, StoreTy))
9517 LastLegalIntegerType = i+1;
9521 // Only use vector types if the vector type is larger than the integer type.
9522 // If they are the same, use integers.
9523 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors;
9524 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType);
9526 // We add +1 here because the LastXXX variables refer to location while
9527 // the NumElem refers to array/index size.
9528 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1;
9529 NumElem = std::min(LastLegalType, NumElem);
9534 // The earliest Node in the DAG.
9535 unsigned EarliestNodeUsed = 0;
9536 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
9537 for (unsigned i=1; i<NumElem; ++i) {
9538 // Find a chain for the new wide-store operand. Notice that some
9539 // of the store nodes that we found may not be selected for inclusion
9540 // in the wide store. The chain we use needs to be the chain of the
9541 // earliest store node which is *used* and replaced by the wide store.
9542 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
9543 EarliestNodeUsed = i;
9546 // Find if it is better to use vectors or integers to load and store
9550 JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
9552 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
9553 JointMemOpVT = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9556 SDLoc LoadDL(LoadNodes[0].MemNode);
9557 SDLoc StoreDL(StoreNodes[0].MemNode);
9559 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
9560 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL,
9561 FirstLoad->getChain(),
9562 FirstLoad->getBasePtr(),
9563 FirstLoad->getPointerInfo(),
9564 false, false, false,
9565 FirstLoad->getAlignment());
9567 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), StoreDL, NewLoad,
9568 FirstInChain->getBasePtr(),
9569 FirstInChain->getPointerInfo(), false, false,
9570 FirstInChain->getAlignment());
9572 // Replace one of the loads with the new load.
9573 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode);
9574 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
9575 SDValue(NewLoad.getNode(), 1));
9577 // Remove the rest of the load chains.
9578 for (unsigned i = 1; i < NumElem ; ++i) {
9579 // Replace all chain users of the old load nodes with the chain of the new
9581 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
9582 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain());
9585 // Replace the first store with the new store.
9586 CombineTo(EarliestOp, NewStore);
9587 // Erase all other stores.
9588 for (unsigned i = 0; i < NumElem ; ++i) {
9589 // Remove all Store nodes.
9590 if (StoreNodes[i].MemNode == EarliestOp)
9592 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9593 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain());
9594 deleteAndRecombine(St);
9600 SDValue DAGCombiner::visitSTORE(SDNode *N) {
9601 StoreSDNode *ST = cast<StoreSDNode>(N);
9602 SDValue Chain = ST->getChain();
9603 SDValue Value = ST->getValue();
9604 SDValue Ptr = ST->getBasePtr();
9606 // If this is a store of a bit convert, store the input value if the
9607 // resultant store does not need a higher alignment than the original.
9608 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
9609 ST->isUnindexed()) {
9610 unsigned OrigAlign = ST->getAlignment();
9611 EVT SVT = Value.getOperand(0).getValueType();
9612 unsigned Align = TLI.getDataLayout()->
9613 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
9614 if (Align <= OrigAlign &&
9615 ((!LegalOperations && !ST->isVolatile()) ||
9616 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
9617 return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0),
9618 Ptr, ST->getPointerInfo(), ST->isVolatile(),
9619 ST->isNonTemporal(), OrigAlign,
9623 // Turn 'store undef, Ptr' -> nothing.
9624 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
9627 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
9628 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
9629 // NOTE: If the original store is volatile, this transform must not increase
9630 // the number of stores. For example, on x86-32 an f64 can be stored in one
9631 // processor operation but an i64 (which is not legal) requires two. So the
9632 // transform should not be done in this case.
9633 if (Value.getOpcode() != ISD::TargetConstantFP) {
9635 switch (CFP->getSimpleValueType(0).SimpleTy) {
9636 default: llvm_unreachable("Unknown FP type");
9637 case MVT::f16: // We don't do this for these yet.
9643 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
9644 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
9645 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
9646 bitcastToAPInt().getZExtValue(), MVT::i32);
9647 return DAG.getStore(Chain, SDLoc(N), Tmp,
9648 Ptr, ST->getMemOperand());
9652 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
9653 !ST->isVolatile()) ||
9654 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
9655 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
9656 getZExtValue(), MVT::i64);
9657 return DAG.getStore(Chain, SDLoc(N), Tmp,
9658 Ptr, ST->getMemOperand());
9661 if (!ST->isVolatile() &&
9662 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
9663 // Many FP stores are not made apparent until after legalize, e.g. for
9664 // argument passing. Since this is so common, custom legalize the
9665 // 64-bit integer store into two 32-bit stores.
9666 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
9667 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
9668 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
9669 if (TLI.isBigEndian()) std::swap(Lo, Hi);
9671 unsigned Alignment = ST->getAlignment();
9672 bool isVolatile = ST->isVolatile();
9673 bool isNonTemporal = ST->isNonTemporal();
9674 AAMDNodes AAInfo = ST->getAAInfo();
9676 SDValue St0 = DAG.getStore(Chain, SDLoc(ST), Lo,
9677 Ptr, ST->getPointerInfo(),
9678 isVolatile, isNonTemporal,
9679 ST->getAlignment(), AAInfo);
9680 Ptr = DAG.getNode(ISD::ADD, SDLoc(N), Ptr.getValueType(), Ptr,
9681 DAG.getConstant(4, Ptr.getValueType()));
9682 Alignment = MinAlign(Alignment, 4U);
9683 SDValue St1 = DAG.getStore(Chain, SDLoc(ST), Hi,
9684 Ptr, ST->getPointerInfo().getWithOffset(4),
9685 isVolatile, isNonTemporal,
9687 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other,
9696 // Try to infer better alignment information than the store already has.
9697 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
9698 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
9699 if (Align > ST->getAlignment())
9700 return DAG.getTruncStore(Chain, SDLoc(N), Value,
9701 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
9702 ST->isVolatile(), ST->isNonTemporal(), Align,
9707 // Try transforming a pair floating point load / store ops to integer
9708 // load / store ops.
9709 SDValue NewST = TransformFPLoadStorePair(N);
9710 if (NewST.getNode())
9713 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA :
9714 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
9716 if (CombinerAAOnlyFunc.getNumOccurrences() &&
9717 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
9720 if (UseAA && ST->isUnindexed()) {
9721 // Walk up chain skipping non-aliasing memory nodes.
9722 SDValue BetterChain = FindBetterChain(N, Chain);
9724 // If there is a better chain.
9725 if (Chain != BetterChain) {
9728 // Replace the chain to avoid dependency.
9729 if (ST->isTruncatingStore()) {
9730 ReplStore = DAG.getTruncStore(BetterChain, SDLoc(N), Value, Ptr,
9731 ST->getMemoryVT(), ST->getMemOperand());
9733 ReplStore = DAG.getStore(BetterChain, SDLoc(N), Value, Ptr,
9734 ST->getMemOperand());
9737 // Create token to keep both nodes around.
9738 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
9739 MVT::Other, Chain, ReplStore);
9741 // Make sure the new and old chains are cleaned up.
9742 AddToWorklist(Token.getNode());
9744 // Don't add users to work list.
9745 return CombineTo(N, Token, false);
9749 // Try transforming N to an indexed store.
9750 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
9751 return SDValue(N, 0);
9753 // FIXME: is there such a thing as a truncating indexed store?
9754 if (ST->isTruncatingStore() && ST->isUnindexed() &&
9755 Value.getValueType().isInteger()) {
9756 // See if we can simplify the input to this truncstore with knowledge that
9757 // only the low bits are being used. For example:
9758 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
9760 GetDemandedBits(Value,
9761 APInt::getLowBitsSet(
9762 Value.getValueType().getScalarType().getSizeInBits(),
9763 ST->getMemoryVT().getScalarType().getSizeInBits()));
9764 AddToWorklist(Value.getNode());
9765 if (Shorter.getNode())
9766 return DAG.getTruncStore(Chain, SDLoc(N), Shorter,
9767 Ptr, ST->getMemoryVT(), ST->getMemOperand());
9769 // Otherwise, see if we can simplify the operation with
9770 // SimplifyDemandedBits, which only works if the value has a single use.
9771 if (SimplifyDemandedBits(Value,
9772 APInt::getLowBitsSet(
9773 Value.getValueType().getScalarType().getSizeInBits(),
9774 ST->getMemoryVT().getScalarType().getSizeInBits())))
9775 return SDValue(N, 0);
9778 // If this is a load followed by a store to the same location, then the store
9780 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
9781 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
9782 ST->isUnindexed() && !ST->isVolatile() &&
9783 // There can't be any side effects between the load and store, such as
9785 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
9786 // The store is dead, remove it.
9791 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
9792 // truncating store. We can do this even if this is already a truncstore.
9793 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
9794 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
9795 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
9796 ST->getMemoryVT())) {
9797 return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0),
9798 Ptr, ST->getMemoryVT(), ST->getMemOperand());
9801 // Only perform this optimization before the types are legal, because we
9802 // don't want to perform this optimization on every DAGCombine invocation.
9804 bool EverChanged = false;
9807 // There can be multiple store sequences on the same chain.
9808 // Keep trying to merge store sequences until we are unable to do so
9809 // or until we merge the last store on the chain.
9810 bool Changed = MergeConsecutiveStores(ST);
9811 EverChanged |= Changed;
9812 if (!Changed) break;
9813 } while (ST->getOpcode() != ISD::DELETED_NODE);
9816 return SDValue(N, 0);
9819 return ReduceLoadOpStoreWidth(N);
9822 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
9823 SDValue InVec = N->getOperand(0);
9824 SDValue InVal = N->getOperand(1);
9825 SDValue EltNo = N->getOperand(2);
9828 // If the inserted element is an UNDEF, just use the input vector.
9829 if (InVal.getOpcode() == ISD::UNDEF)
9832 EVT VT = InVec.getValueType();
9834 // If we can't generate a legal BUILD_VECTOR, exit
9835 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
9838 // Check that we know which element is being inserted
9839 if (!isa<ConstantSDNode>(EltNo))
9841 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9843 // Canonicalize insert_vector_elt dag nodes.
9845 // (insert_vector_elt (insert_vector_elt A, Idx0), Idx1)
9846 // -> (insert_vector_elt (insert_vector_elt A, Idx1), Idx0)
9848 // Do this only if the child insert_vector node has one use; also
9849 // do this only if indices are both constants and Idx1 < Idx0.
9850 if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT && InVec.hasOneUse()
9851 && isa<ConstantSDNode>(InVec.getOperand(2))) {
9853 cast<ConstantSDNode>(InVec.getOperand(2))->getZExtValue();
9854 if (Elt < OtherElt) {
9856 SDValue NewOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(N), VT,
9857 InVec.getOperand(0), InVal, EltNo);
9858 AddToWorklist(NewOp.getNode());
9859 return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(InVec.getNode()),
9860 VT, NewOp, InVec.getOperand(1), InVec.getOperand(2));
9864 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
9865 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
9867 SmallVector<SDValue, 8> Ops;
9868 // Do not combine these two vectors if the output vector will not replace
9869 // the input vector.
9870 if (InVec.getOpcode() == ISD::BUILD_VECTOR && InVec.hasOneUse()) {
9871 Ops.append(InVec.getNode()->op_begin(),
9872 InVec.getNode()->op_end());
9873 } else if (InVec.getOpcode() == ISD::UNDEF) {
9874 unsigned NElts = VT.getVectorNumElements();
9875 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
9880 // Insert the element
9881 if (Elt < Ops.size()) {
9882 // All the operands of BUILD_VECTOR must have the same type;
9883 // we enforce that here.
9884 EVT OpVT = Ops[0].getValueType();
9885 if (InVal.getValueType() != OpVT)
9886 InVal = OpVT.bitsGT(InVal.getValueType()) ?
9887 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
9888 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
9892 // Return the new vector
9893 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
9896 SDValue DAGCombiner::ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
9897 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad) {
9898 EVT ResultVT = EVE->getValueType(0);
9899 EVT VecEltVT = InVecVT.getVectorElementType();
9900 unsigned Align = OriginalLoad->getAlignment();
9901 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
9902 VecEltVT.getTypeForEVT(*DAG.getContext()));
9904 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VecEltVT))
9909 SDValue NewPtr = OriginalLoad->getBasePtr();
9911 EVT PtrType = NewPtr.getValueType();
9912 MachinePointerInfo MPI;
9913 if (auto *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo)) {
9914 int Elt = ConstEltNo->getZExtValue();
9915 unsigned PtrOff = VecEltVT.getSizeInBits() * Elt / 8;
9916 if (TLI.isBigEndian())
9917 PtrOff = InVecVT.getSizeInBits() / 8 - PtrOff;
9918 Offset = DAG.getConstant(PtrOff, PtrType);
9919 MPI = OriginalLoad->getPointerInfo().getWithOffset(PtrOff);
9921 Offset = DAG.getNode(
9922 ISD::MUL, SDLoc(EVE), EltNo.getValueType(), EltNo,
9923 DAG.getConstant(VecEltVT.getStoreSize(), EltNo.getValueType()));
9924 if (TLI.isBigEndian())
9925 Offset = DAG.getNode(
9926 ISD::SUB, SDLoc(EVE), EltNo.getValueType(),
9927 DAG.getConstant(InVecVT.getStoreSize(), EltNo.getValueType()), Offset);
9928 MPI = OriginalLoad->getPointerInfo();
9930 NewPtr = DAG.getNode(ISD::ADD, SDLoc(EVE), PtrType, NewPtr, Offset);
9932 // The replacement we need to do here is a little tricky: we need to
9933 // replace an extractelement of a load with a load.
9934 // Use ReplaceAllUsesOfValuesWith to do the replacement.
9935 // Note that this replacement assumes that the extractvalue is the only
9936 // use of the load; that's okay because we don't want to perform this
9937 // transformation in other cases anyway.
9940 if (ResultVT.bitsGT(VecEltVT)) {
9941 // If the result type of vextract is wider than the load, then issue an
9942 // extending load instead.
9943 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, VecEltVT)
9946 Load = DAG.getExtLoad(
9947 ExtType, SDLoc(EVE), ResultVT, OriginalLoad->getChain(), NewPtr, MPI,
9948 VecEltVT, OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
9949 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
9950 Chain = Load.getValue(1);
9953 VecEltVT, SDLoc(EVE), OriginalLoad->getChain(), NewPtr, MPI,
9954 OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
9955 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
9956 Chain = Load.getValue(1);
9957 if (ResultVT.bitsLT(VecEltVT))
9958 Load = DAG.getNode(ISD::TRUNCATE, SDLoc(EVE), ResultVT, Load);
9960 Load = DAG.getNode(ISD::BITCAST, SDLoc(EVE), ResultVT, Load);
9962 WorklistRemover DeadNodes(*this);
9963 SDValue From[] = { SDValue(EVE, 0), SDValue(OriginalLoad, 1) };
9964 SDValue To[] = { Load, Chain };
9965 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
9966 // Since we're explicitly calling ReplaceAllUses, add the new node to the
9967 // worklist explicitly as well.
9968 AddToWorklist(Load.getNode());
9969 AddUsersToWorklist(Load.getNode()); // Add users too
9970 // Make sure to revisit this node to clean it up; it will usually be dead.
9973 return SDValue(EVE, 0);
9976 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
9977 // (vextract (scalar_to_vector val, 0) -> val
9978 SDValue InVec = N->getOperand(0);
9979 EVT VT = InVec.getValueType();
9980 EVT NVT = N->getValueType(0);
9982 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
9983 // Check if the result type doesn't match the inserted element type. A
9984 // SCALAR_TO_VECTOR may truncate the inserted element and the
9985 // EXTRACT_VECTOR_ELT may widen the extracted vector.
9986 SDValue InOp = InVec.getOperand(0);
9987 if (InOp.getValueType() != NVT) {
9988 assert(InOp.getValueType().isInteger() && NVT.isInteger());
9989 return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT);
9994 SDValue EltNo = N->getOperand(1);
9995 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
9997 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
9998 // We only perform this optimization before the op legalization phase because
9999 // we may introduce new vector instructions which are not backed by TD
10000 // patterns. For example on AVX, extracting elements from a wide vector
10001 // without using extract_subvector. However, if we can find an underlying
10002 // scalar value, then we can always use that.
10003 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
10005 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
10006 int NumElem = VT.getVectorNumElements();
10007 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
10008 // Find the new index to extract from.
10009 int OrigElt = SVOp->getMaskElt(Elt);
10011 // Extracting an undef index is undef.
10013 return DAG.getUNDEF(NVT);
10015 // Select the right vector half to extract from.
10017 if (OrigElt < NumElem) {
10018 SVInVec = InVec->getOperand(0);
10020 SVInVec = InVec->getOperand(1);
10021 OrigElt -= NumElem;
10024 if (SVInVec.getOpcode() == ISD::BUILD_VECTOR) {
10025 SDValue InOp = SVInVec.getOperand(OrigElt);
10026 if (InOp.getValueType() != NVT) {
10027 assert(InOp.getValueType().isInteger() && NVT.isInteger());
10028 InOp = DAG.getSExtOrTrunc(InOp, SDLoc(SVInVec), NVT);
10034 // FIXME: We should handle recursing on other vector shuffles and
10035 // scalar_to_vector here as well.
10037 if (!LegalOperations) {
10038 EVT IndexTy = TLI.getVectorIdxTy();
10039 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT,
10040 SVInVec, DAG.getConstant(OrigElt, IndexTy));
10044 bool BCNumEltsChanged = false;
10045 EVT ExtVT = VT.getVectorElementType();
10048 // If the result of load has to be truncated, then it's not necessarily
10050 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
10053 if (InVec.getOpcode() == ISD::BITCAST) {
10054 // Don't duplicate a load with other uses.
10055 if (!InVec.hasOneUse())
10058 EVT BCVT = InVec.getOperand(0).getValueType();
10059 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
10061 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
10062 BCNumEltsChanged = true;
10063 InVec = InVec.getOperand(0);
10064 ExtVT = BCVT.getVectorElementType();
10067 // (vextract (vN[if]M load $addr), i) -> ([if]M load $addr + i * size)
10068 if (!LegalOperations && !ConstEltNo && InVec.hasOneUse() &&
10069 ISD::isNormalLoad(InVec.getNode()) &&
10070 !N->getOperand(1)->hasPredecessor(InVec.getNode())) {
10071 SDValue Index = N->getOperand(1);
10072 if (LoadSDNode *OrigLoad = dyn_cast<LoadSDNode>(InVec))
10073 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, Index,
10077 // Perform only after legalization to ensure build_vector / vector_shuffle
10078 // optimizations have already been done.
10079 if (!LegalOperations) return SDValue();
10081 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
10082 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
10083 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
10086 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
10088 LoadSDNode *LN0 = nullptr;
10089 const ShuffleVectorSDNode *SVN = nullptr;
10090 if (ISD::isNormalLoad(InVec.getNode())) {
10091 LN0 = cast<LoadSDNode>(InVec);
10092 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
10093 InVec.getOperand(0).getValueType() == ExtVT &&
10094 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
10095 // Don't duplicate a load with other uses.
10096 if (!InVec.hasOneUse())
10099 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
10100 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
10101 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
10103 // (load $addr+1*size)
10105 // Don't duplicate a load with other uses.
10106 if (!InVec.hasOneUse())
10109 // If the bit convert changed the number of elements, it is unsafe
10110 // to examine the mask.
10111 if (BCNumEltsChanged)
10114 // Select the input vector, guarding against out of range extract vector.
10115 unsigned NumElems = VT.getVectorNumElements();
10116 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
10117 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
10119 if (InVec.getOpcode() == ISD::BITCAST) {
10120 // Don't duplicate a load with other uses.
10121 if (!InVec.hasOneUse())
10124 InVec = InVec.getOperand(0);
10126 if (ISD::isNormalLoad(InVec.getNode())) {
10127 LN0 = cast<LoadSDNode>(InVec);
10128 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
10129 EltNo = DAG.getConstant(Elt, EltNo.getValueType());
10133 // Make sure we found a non-volatile load and the extractelement is
10135 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
10138 // If Idx was -1 above, Elt is going to be -1, so just return undef.
10140 return DAG.getUNDEF(LVT);
10142 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, EltNo, LN0);
10148 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
10149 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
10150 // We perform this optimization post type-legalization because
10151 // the type-legalizer often scalarizes integer-promoted vectors.
10152 // Performing this optimization before may create bit-casts which
10153 // will be type-legalized to complex code sequences.
10154 // We perform this optimization only before the operation legalizer because we
10155 // may introduce illegal operations.
10156 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
10159 unsigned NumInScalars = N->getNumOperands();
10161 EVT VT = N->getValueType(0);
10163 // Check to see if this is a BUILD_VECTOR of a bunch of values
10164 // which come from any_extend or zero_extend nodes. If so, we can create
10165 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
10166 // optimizations. We do not handle sign-extend because we can't fill the sign
10168 EVT SourceType = MVT::Other;
10169 bool AllAnyExt = true;
10171 for (unsigned i = 0; i != NumInScalars; ++i) {
10172 SDValue In = N->getOperand(i);
10173 // Ignore undef inputs.
10174 if (In.getOpcode() == ISD::UNDEF) continue;
10176 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
10177 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
10179 // Abort if the element is not an extension.
10180 if (!ZeroExt && !AnyExt) {
10181 SourceType = MVT::Other;
10185 // The input is a ZeroExt or AnyExt. Check the original type.
10186 EVT InTy = In.getOperand(0).getValueType();
10188 // Check that all of the widened source types are the same.
10189 if (SourceType == MVT::Other)
10192 else if (InTy != SourceType) {
10193 // Multiple income types. Abort.
10194 SourceType = MVT::Other;
10198 // Check if all of the extends are ANY_EXTENDs.
10199 AllAnyExt &= AnyExt;
10202 // In order to have valid types, all of the inputs must be extended from the
10203 // same source type and all of the inputs must be any or zero extend.
10204 // Scalar sizes must be a power of two.
10205 EVT OutScalarTy = VT.getScalarType();
10206 bool ValidTypes = SourceType != MVT::Other &&
10207 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
10208 isPowerOf2_32(SourceType.getSizeInBits());
10210 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
10211 // turn into a single shuffle instruction.
10215 bool isLE = TLI.isLittleEndian();
10216 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
10217 assert(ElemRatio > 1 && "Invalid element size ratio");
10218 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
10219 DAG.getConstant(0, SourceType);
10221 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
10222 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
10224 // Populate the new build_vector
10225 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
10226 SDValue Cast = N->getOperand(i);
10227 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
10228 Cast.getOpcode() == ISD::ZERO_EXTEND ||
10229 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
10231 if (Cast.getOpcode() == ISD::UNDEF)
10232 In = DAG.getUNDEF(SourceType);
10234 In = Cast->getOperand(0);
10235 unsigned Index = isLE ? (i * ElemRatio) :
10236 (i * ElemRatio + (ElemRatio - 1));
10238 assert(Index < Ops.size() && "Invalid index");
10242 // The type of the new BUILD_VECTOR node.
10243 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
10244 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
10245 "Invalid vector size");
10246 // Check if the new vector type is legal.
10247 if (!isTypeLegal(VecVT)) return SDValue();
10249 // Make the new BUILD_VECTOR.
10250 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
10252 // The new BUILD_VECTOR node has the potential to be further optimized.
10253 AddToWorklist(BV.getNode());
10254 // Bitcast to the desired type.
10255 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
10258 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
10259 EVT VT = N->getValueType(0);
10261 unsigned NumInScalars = N->getNumOperands();
10264 EVT SrcVT = MVT::Other;
10265 unsigned Opcode = ISD::DELETED_NODE;
10266 unsigned NumDefs = 0;
10268 for (unsigned i = 0; i != NumInScalars; ++i) {
10269 SDValue In = N->getOperand(i);
10270 unsigned Opc = In.getOpcode();
10272 if (Opc == ISD::UNDEF)
10275 // If all scalar values are floats and converted from integers.
10276 if (Opcode == ISD::DELETED_NODE &&
10277 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
10284 EVT InVT = In.getOperand(0).getValueType();
10286 // If all scalar values are typed differently, bail out. It's chosen to
10287 // simplify BUILD_VECTOR of integer types.
10288 if (SrcVT == MVT::Other)
10295 // If the vector has just one element defined, it's not worth to fold it into
10296 // a vectorized one.
10300 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
10301 && "Should only handle conversion from integer to float.");
10302 assert(SrcVT != MVT::Other && "Cannot determine source type!");
10304 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
10306 if (!TLI.isOperationLegalOrCustom(Opcode, NVT))
10309 SmallVector<SDValue, 8> Opnds;
10310 for (unsigned i = 0; i != NumInScalars; ++i) {
10311 SDValue In = N->getOperand(i);
10313 if (In.getOpcode() == ISD::UNDEF)
10314 Opnds.push_back(DAG.getUNDEF(SrcVT));
10316 Opnds.push_back(In.getOperand(0));
10318 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Opnds);
10319 AddToWorklist(BV.getNode());
10321 return DAG.getNode(Opcode, dl, VT, BV);
10324 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
10325 unsigned NumInScalars = N->getNumOperands();
10327 EVT VT = N->getValueType(0);
10329 // A vector built entirely of undefs is undef.
10330 if (ISD::allOperandsUndef(N))
10331 return DAG.getUNDEF(VT);
10333 SDValue V = reduceBuildVecExtToExtBuildVec(N);
10337 V = reduceBuildVecConvertToConvertBuildVec(N);
10341 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
10342 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
10343 // at most two distinct vectors, turn this into a shuffle node.
10345 // May only combine to shuffle after legalize if shuffle is legal.
10346 if (LegalOperations && !TLI.isOperationLegal(ISD::VECTOR_SHUFFLE, VT))
10349 SDValue VecIn1, VecIn2;
10350 for (unsigned i = 0; i != NumInScalars; ++i) {
10351 // Ignore undef inputs.
10352 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
10354 // If this input is something other than a EXTRACT_VECTOR_ELT with a
10355 // constant index, bail out.
10356 if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
10357 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
10358 VecIn1 = VecIn2 = SDValue(nullptr, 0);
10362 // We allow up to two distinct input vectors.
10363 SDValue ExtractedFromVec = N->getOperand(i).getOperand(0);
10364 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
10367 if (!VecIn1.getNode()) {
10368 VecIn1 = ExtractedFromVec;
10369 } else if (!VecIn2.getNode()) {
10370 VecIn2 = ExtractedFromVec;
10372 // Too many inputs.
10373 VecIn1 = VecIn2 = SDValue(nullptr, 0);
10378 // If everything is good, we can make a shuffle operation.
10379 if (VecIn1.getNode()) {
10380 SmallVector<int, 8> Mask;
10381 for (unsigned i = 0; i != NumInScalars; ++i) {
10382 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
10383 Mask.push_back(-1);
10387 // If extracting from the first vector, just use the index directly.
10388 SDValue Extract = N->getOperand(i);
10389 SDValue ExtVal = Extract.getOperand(1);
10390 if (Extract.getOperand(0) == VecIn1) {
10391 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
10392 if (ExtIndex > VT.getVectorNumElements())
10395 Mask.push_back(ExtIndex);
10399 // Otherwise, use InIdx + VecSize
10400 unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue();
10401 Mask.push_back(Idx+NumInScalars);
10404 // We can't generate a shuffle node with mismatched input and output types.
10405 // Attempt to transform a single input vector to the correct type.
10406 if ((VT != VecIn1.getValueType())) {
10407 // We don't support shuffeling between TWO values of different types.
10408 if (VecIn2.getNode())
10411 // We only support widening of vectors which are half the size of the
10412 // output registers. For example XMM->YMM widening on X86 with AVX.
10413 if (VecIn1.getValueType().getSizeInBits()*2 != VT.getSizeInBits())
10416 // If the input vector type has a different base type to the output
10417 // vector type, bail out.
10418 if (VecIn1.getValueType().getVectorElementType() !=
10419 VT.getVectorElementType())
10422 // Widen the input vector by adding undef values.
10423 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
10424 VecIn1, DAG.getUNDEF(VecIn1.getValueType()));
10427 // If VecIn2 is unused then change it to undef.
10428 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
10430 // Check that we were able to transform all incoming values to the same
10432 if (VecIn2.getValueType() != VecIn1.getValueType() ||
10433 VecIn1.getValueType() != VT)
10436 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
10437 if (!isTypeLegal(VT))
10440 // Return the new VECTOR_SHUFFLE node.
10444 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]);
10450 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
10451 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
10452 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
10453 // inputs come from at most two distinct vectors, turn this into a shuffle
10456 // If we only have one input vector, we don't need to do any concatenation.
10457 if (N->getNumOperands() == 1)
10458 return N->getOperand(0);
10460 // Check if all of the operands are undefs.
10461 EVT VT = N->getValueType(0);
10462 if (ISD::allOperandsUndef(N))
10463 return DAG.getUNDEF(VT);
10465 // Optimize concat_vectors where one of the vectors is undef.
10466 if (N->getNumOperands() == 2 &&
10467 N->getOperand(1)->getOpcode() == ISD::UNDEF) {
10468 SDValue In = N->getOperand(0);
10469 assert(In.getValueType().isVector() && "Must concat vectors");
10471 // Transform: concat_vectors(scalar, undef) -> scalar_to_vector(sclr).
10472 if (In->getOpcode() == ISD::BITCAST &&
10473 !In->getOperand(0)->getValueType(0).isVector()) {
10474 SDValue Scalar = In->getOperand(0);
10475 EVT SclTy = Scalar->getValueType(0);
10477 if (!SclTy.isFloatingPoint() && !SclTy.isInteger())
10480 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SclTy,
10481 VT.getSizeInBits() / SclTy.getSizeInBits());
10482 if (!TLI.isTypeLegal(NVT) || !TLI.isTypeLegal(Scalar.getValueType()))
10485 SDLoc dl = SDLoc(N);
10486 SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NVT, Scalar);
10487 return DAG.getNode(ISD::BITCAST, dl, VT, Res);
10491 // fold (concat_vectors (BUILD_VECTOR A, B, ...), (BUILD_VECTOR C, D, ...))
10492 // -> (BUILD_VECTOR A, B, ..., C, D, ...)
10493 if (N->getNumOperands() == 2 &&
10494 N->getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
10495 N->getOperand(1).getOpcode() == ISD::BUILD_VECTOR) {
10496 EVT VT = N->getValueType(0);
10497 SDValue N0 = N->getOperand(0);
10498 SDValue N1 = N->getOperand(1);
10499 SmallVector<SDValue, 8> Opnds;
10500 unsigned BuildVecNumElts = N0.getNumOperands();
10502 EVT SclTy0 = N0.getOperand(0)->getValueType(0);
10503 EVT SclTy1 = N1.getOperand(0)->getValueType(0);
10504 if (SclTy0.isFloatingPoint()) {
10505 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10506 Opnds.push_back(N0.getOperand(i));
10507 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10508 Opnds.push_back(N1.getOperand(i));
10510 // If BUILD_VECTOR are from built from integer, they may have different
10511 // operand types. Get the smaller type and truncate all operands to it.
10512 EVT MinTy = SclTy0.bitsLE(SclTy1) ? SclTy0 : SclTy1;
10513 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10514 Opnds.push_back(DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinTy,
10515 N0.getOperand(i)));
10516 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10517 Opnds.push_back(DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinTy,
10518 N1.getOperand(i)));
10521 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
10524 // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
10525 // nodes often generate nop CONCAT_VECTOR nodes.
10526 // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that
10527 // place the incoming vectors at the exact same location.
10528 SDValue SingleSource = SDValue();
10529 unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements();
10531 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
10532 SDValue Op = N->getOperand(i);
10534 if (Op.getOpcode() == ISD::UNDEF)
10537 // Check if this is the identity extract:
10538 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
10541 // Find the single incoming vector for the extract_subvector.
10542 if (SingleSource.getNode()) {
10543 if (Op.getOperand(0) != SingleSource)
10546 SingleSource = Op.getOperand(0);
10548 // Check the source type is the same as the type of the result.
10549 // If not, this concat may extend the vector, so we can not
10550 // optimize it away.
10551 if (SingleSource.getValueType() != N->getValueType(0))
10555 unsigned IdentityIndex = i * PartNumElem;
10556 ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1));
10557 // The extract index must be constant.
10561 // Check that we are reading from the identity index.
10562 if (CS->getZExtValue() != IdentityIndex)
10566 if (SingleSource.getNode())
10567 return SingleSource;
10572 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
10573 EVT NVT = N->getValueType(0);
10574 SDValue V = N->getOperand(0);
10576 if (V->getOpcode() == ISD::CONCAT_VECTORS) {
10578 // (extract_subvec (concat V1, V2, ...), i)
10581 // Only operand 0 is checked as 'concat' assumes all inputs of the same
10583 if (V->getOperand(0).getValueType() != NVT)
10585 unsigned Idx = dyn_cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
10586 unsigned NumElems = NVT.getVectorNumElements();
10587 assert((Idx % NumElems) == 0 &&
10588 "IDX in concat is not a multiple of the result vector length.");
10589 return V->getOperand(Idx / NumElems);
10593 if (V->getOpcode() == ISD::BITCAST)
10594 V = V.getOperand(0);
10596 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
10598 // Handle only simple case where vector being inserted and vector
10599 // being extracted are of same type, and are half size of larger vectors.
10600 EVT BigVT = V->getOperand(0).getValueType();
10601 EVT SmallVT = V->getOperand(1).getValueType();
10602 if (!NVT.bitsEq(SmallVT) || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
10605 // Only handle cases where both indexes are constants with the same type.
10606 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
10607 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
10609 if (InsIdx && ExtIdx &&
10610 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
10611 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
10613 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
10615 // indices are equal or bit offsets are equal => V1
10616 // otherwise => (extract_subvec V1, ExtIdx)
10617 if (InsIdx->getZExtValue() * SmallVT.getScalarType().getSizeInBits() ==
10618 ExtIdx->getZExtValue() * NVT.getScalarType().getSizeInBits())
10619 return DAG.getNode(ISD::BITCAST, dl, NVT, V->getOperand(1));
10620 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT,
10621 DAG.getNode(ISD::BITCAST, dl,
10622 N->getOperand(0).getValueType(),
10623 V->getOperand(0)), N->getOperand(1));
10630 // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat.
10631 static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
10632 EVT VT = N->getValueType(0);
10633 unsigned NumElts = VT.getVectorNumElements();
10635 SDValue N0 = N->getOperand(0);
10636 SDValue N1 = N->getOperand(1);
10637 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
10639 SmallVector<SDValue, 4> Ops;
10640 EVT ConcatVT = N0.getOperand(0).getValueType();
10641 unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements();
10642 unsigned NumConcats = NumElts / NumElemsPerConcat;
10644 // Look at every vector that's inserted. We're looking for exact
10645 // subvector-sized copies from a concatenated vector
10646 for (unsigned I = 0; I != NumConcats; ++I) {
10647 // Make sure we're dealing with a copy.
10648 unsigned Begin = I * NumElemsPerConcat;
10649 bool AllUndef = true, NoUndef = true;
10650 for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) {
10651 if (SVN->getMaskElt(J) >= 0)
10658 if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0)
10661 for (unsigned J = 1; J != NumElemsPerConcat; ++J)
10662 if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J))
10665 unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat;
10666 if (FirstElt < N0.getNumOperands())
10667 Ops.push_back(N0.getOperand(FirstElt));
10669 Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands()));
10671 } else if (AllUndef) {
10672 Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType()));
10673 } else { // Mixed with general masks and undefs, can't do optimization.
10678 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops);
10681 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
10682 EVT VT = N->getValueType(0);
10683 unsigned NumElts = VT.getVectorNumElements();
10685 SDValue N0 = N->getOperand(0);
10686 SDValue N1 = N->getOperand(1);
10688 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
10690 // Canonicalize shuffle undef, undef -> undef
10691 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
10692 return DAG.getUNDEF(VT);
10694 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
10696 // Canonicalize shuffle v, v -> v, undef
10698 SmallVector<int, 8> NewMask;
10699 for (unsigned i = 0; i != NumElts; ++i) {
10700 int Idx = SVN->getMaskElt(i);
10701 if (Idx >= (int)NumElts) Idx -= NumElts;
10702 NewMask.push_back(Idx);
10704 return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT),
10708 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
10709 if (N0.getOpcode() == ISD::UNDEF) {
10710 SmallVector<int, 8> NewMask;
10711 for (unsigned i = 0; i != NumElts; ++i) {
10712 int Idx = SVN->getMaskElt(i);
10714 if (Idx >= (int)NumElts)
10717 Idx = -1; // remove reference to lhs
10719 NewMask.push_back(Idx);
10721 return DAG.getVectorShuffle(VT, SDLoc(N), N1, DAG.getUNDEF(VT),
10725 // Remove references to rhs if it is undef
10726 if (N1.getOpcode() == ISD::UNDEF) {
10727 bool Changed = false;
10728 SmallVector<int, 8> NewMask;
10729 for (unsigned i = 0; i != NumElts; ++i) {
10730 int Idx = SVN->getMaskElt(i);
10731 if (Idx >= (int)NumElts) {
10735 NewMask.push_back(Idx);
10738 return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, &NewMask[0]);
10741 // If it is a splat, check if the argument vector is another splat or a
10742 // build_vector with all scalar elements the same.
10743 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
10744 SDNode *V = N0.getNode();
10746 // If this is a bit convert that changes the element type of the vector but
10747 // not the number of vector elements, look through it. Be careful not to
10748 // look though conversions that change things like v4f32 to v2f64.
10749 if (V->getOpcode() == ISD::BITCAST) {
10750 SDValue ConvInput = V->getOperand(0);
10751 if (ConvInput.getValueType().isVector() &&
10752 ConvInput.getValueType().getVectorNumElements() == NumElts)
10753 V = ConvInput.getNode();
10756 if (V->getOpcode() == ISD::BUILD_VECTOR) {
10757 assert(V->getNumOperands() == NumElts &&
10758 "BUILD_VECTOR has wrong number of operands");
10760 bool AllSame = true;
10761 for (unsigned i = 0; i != NumElts; ++i) {
10762 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
10763 Base = V->getOperand(i);
10767 // Splat of <u, u, u, u>, return <u, u, u, u>
10768 if (!Base.getNode())
10770 for (unsigned i = 0; i != NumElts; ++i) {
10771 if (V->getOperand(i) != Base) {
10776 // Splat of <x, x, x, x>, return <x, x, x, x>
10782 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
10783 Level < AfterLegalizeVectorOps &&
10784 (N1.getOpcode() == ISD::UNDEF ||
10785 (N1.getOpcode() == ISD::CONCAT_VECTORS &&
10786 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) {
10787 SDValue V = partitionShuffleOfConcats(N, DAG);
10793 // If this shuffle node is simply a swizzle of another shuffle node,
10794 // then try to simplify it.
10795 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
10796 N1.getOpcode() == ISD::UNDEF) {
10798 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
10800 // The incoming shuffle must be of the same type as the result of the
10801 // current shuffle.
10802 assert(OtherSV->getOperand(0).getValueType() == VT &&
10803 "Shuffle types don't match");
10805 SmallVector<int, 4> Mask;
10806 // Compute the combined shuffle mask.
10807 for (unsigned i = 0; i != NumElts; ++i) {
10808 int Idx = SVN->getMaskElt(i);
10809 assert(Idx < (int)NumElts && "Index references undef operand");
10810 // Next, this index comes from the first value, which is the incoming
10811 // shuffle. Adopt the incoming index.
10813 Idx = OtherSV->getMaskElt(Idx);
10814 Mask.push_back(Idx);
10817 // Check if all indices in Mask are Undef. In case, propagate Undef.
10818 bool isUndefMask = true;
10819 for (unsigned i = 0; i != NumElts && isUndefMask; ++i)
10820 isUndefMask &= Mask[i] < 0;
10823 return DAG.getUNDEF(VT);
10825 bool CommuteOperands = false;
10826 if (N0.getOperand(1).getOpcode() != ISD::UNDEF) {
10827 // To be valid, the combine shuffle mask should only reference elements
10828 // from one of the two vectors in input to the inner shufflevector.
10829 bool IsValidMask = true;
10830 for (unsigned i = 0; i != NumElts && IsValidMask; ++i)
10831 // See if the combined mask only reference undefs or elements coming
10832 // from the first shufflevector operand.
10833 IsValidMask = Mask[i] < 0 || (unsigned)Mask[i] < NumElts;
10835 if (!IsValidMask) {
10836 IsValidMask = true;
10837 for (unsigned i = 0; i != NumElts && IsValidMask; ++i)
10838 // Check that all the elements come from the second shuffle operand.
10839 IsValidMask = Mask[i] < 0 || (unsigned)Mask[i] >= NumElts;
10840 CommuteOperands = IsValidMask;
10843 // Early exit if the combined shuffle mask is not valid.
10848 // See if this pair of shuffles can be safely folded according to either
10849 // of the following rules:
10850 // shuffle(shuffle(x, y), undef) -> x
10851 // shuffle(shuffle(x, undef), undef) -> x
10852 // shuffle(shuffle(x, y), undef) -> y
10853 bool IsIdentityMask = true;
10854 unsigned BaseMaskIndex = CommuteOperands ? NumElts : 0;
10855 for (unsigned i = 0; i != NumElts && IsIdentityMask; ++i) {
10860 // The combined shuffle must map each index to itself.
10861 IsIdentityMask = (unsigned)Mask[i] == i + BaseMaskIndex;
10864 if (IsIdentityMask) {
10865 if (CommuteOperands)
10866 // optimize shuffle(shuffle(x, y), undef) -> y.
10867 return OtherSV->getOperand(1);
10869 // optimize shuffle(shuffle(x, undef), undef) -> x
10870 // optimize shuffle(shuffle(x, y), undef) -> x
10871 return OtherSV->getOperand(0);
10874 // It may still be beneficial to combine the two shuffles if the
10875 // resulting shuffle is legal.
10876 if (TLI.isTypeLegal(VT)) {
10877 if (!CommuteOperands) {
10878 if (TLI.isShuffleMaskLegal(Mask, VT))
10879 // shuffle(shuffle(x, undef, M1), undef, M2) -> shuffle(x, undef, M3).
10880 // shuffle(shuffle(x, y, M1), undef, M2) -> shuffle(x, undef, M3)
10881 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0), N1,
10884 // Compute the commuted shuffle mask.
10885 for (unsigned i = 0; i != NumElts; ++i) {
10889 else if (idx < (int)NumElts)
10890 Mask[i] = idx + NumElts;
10892 Mask[i] = idx - NumElts;
10895 if (TLI.isShuffleMaskLegal(Mask, VT))
10896 // shuffle(shuffle(x, y, M1), undef, M2) -> shuffle(y, undef, M3)
10897 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(1), N1,
10903 // Canonicalize shuffles according to rules:
10904 // shuffle(A, shuffle(A, B)) -> shuffle(shuffle(A,B), A)
10905 // shuffle(B, shuffle(A, B)) -> shuffle(shuffle(A,B), B)
10906 // shuffle(B, shuffle(A, Undef)) -> shuffle(shuffle(A, Undef), B)
10907 if (N1.getOpcode() == ISD::VECTOR_SHUFFLE && N0.getOpcode() != ISD::UNDEF &&
10908 N0.getOpcode() != ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
10909 TLI.isTypeLegal(VT)) {
10910 // The incoming shuffle must be of the same type as the result of the
10911 // current shuffle.
10912 assert(N1->getOperand(0).getValueType() == VT &&
10913 "Shuffle types don't match");
10915 SDValue SV0 = N1->getOperand(0);
10916 SDValue SV1 = N1->getOperand(1);
10917 bool HasSameOp0 = N0 == SV0;
10918 bool IsSV1Undef = SV1.getOpcode() == ISD::UNDEF;
10919 if (HasSameOp0 || IsSV1Undef || N0 == SV1)
10920 // Commute the operands of this shuffle so that next rule
10922 return DAG.getCommutedVectorShuffle(*SVN);
10925 // Try to fold according to rules:
10926 // shuffle(shuffle(A, B, M0), B, M1) -> shuffle(A, B, M2)
10927 // shuffle(shuffle(A, B, M0), A, M1) -> shuffle(A, B, M2)
10928 // shuffle(shuffle(A, Undef, M0), B, M1) -> shuffle(A, B, M2)
10929 // shuffle(shuffle(A, Undef, M0), A, M1) -> shuffle(A, Undef, M2)
10930 // Don't try to fold shuffles with illegal type.
10931 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
10932 N1.getOpcode() != ISD::UNDEF && TLI.isTypeLegal(VT)) {
10933 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
10935 // The incoming shuffle must be of the same type as the result of the
10936 // current shuffle.
10937 assert(OtherSV->getOperand(0).getValueType() == VT &&
10938 "Shuffle types don't match");
10940 SDValue SV0 = OtherSV->getOperand(0);
10941 SDValue SV1 = OtherSV->getOperand(1);
10942 bool HasSameOp0 = N1 == SV0;
10943 bool IsSV1Undef = SV1.getOpcode() == ISD::UNDEF;
10944 if (!HasSameOp0 && !IsSV1Undef && N1 != SV1)
10948 SmallVector<int, 4> Mask;
10949 // Compute the combined shuffle mask for a shuffle with SV0 as the first
10950 // operand, and SV1 as the second operand.
10951 for (unsigned i = 0; i != NumElts; ++i) {
10952 int Idx = SVN->getMaskElt(i);
10954 // Propagate Undef.
10955 Mask.push_back(Idx);
10959 if (Idx < (int)NumElts) {
10960 Idx = OtherSV->getMaskElt(Idx);
10961 if (IsSV1Undef && Idx >= (int) NumElts)
10962 Idx = -1; // Propagate Undef.
10964 Idx = HasSameOp0 ? Idx - NumElts : Idx;
10966 Mask.push_back(Idx);
10969 // Check if all indices in Mask are Undef. In case, propagate Undef.
10970 bool isUndefMask = true;
10971 for (unsigned i = 0; i != NumElts && isUndefMask; ++i)
10972 isUndefMask &= Mask[i] < 0;
10975 return DAG.getUNDEF(VT);
10977 // Avoid introducing shuffles with illegal mask.
10978 if (TLI.isShuffleMaskLegal(Mask, VT)) {
10980 // shuffle(shuffle(A, Undef, M0), B, M1) -> shuffle(A, B, M2)
10981 // shuffle(shuffle(A, Undef, M0), A, M1) -> shuffle(A, Undef, M2)
10982 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, N1, &Mask[0]);
10983 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, &Mask[0]);
10990 SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
10991 SDValue N0 = N->getOperand(0);
10992 SDValue N2 = N->getOperand(2);
10994 // If the input vector is a concatenation, and the insert replaces
10995 // one of the halves, we can optimize into a single concat_vectors.
10996 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
10997 N0->getNumOperands() == 2 && N2.getOpcode() == ISD::Constant) {
10998 APInt InsIdx = cast<ConstantSDNode>(N2)->getAPIntValue();
10999 EVT VT = N->getValueType(0);
11001 // Lower half: fold (insert_subvector (concat_vectors X, Y), Z) ->
11002 // (concat_vectors Z, Y)
11004 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
11005 N->getOperand(1), N0.getOperand(1));
11007 // Upper half: fold (insert_subvector (concat_vectors X, Y), Z) ->
11008 // (concat_vectors X, Z)
11009 if (InsIdx == VT.getVectorNumElements()/2)
11010 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
11011 N0.getOperand(0), N->getOperand(1));
11017 /// Returns a vector_shuffle if it able to transform an AND to a vector_shuffle
11018 /// with the destination vector and a zero vector.
11019 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
11020 /// vector_shuffle V, Zero, <0, 4, 2, 4>
11021 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
11022 EVT VT = N->getValueType(0);
11024 SDValue LHS = N->getOperand(0);
11025 SDValue RHS = N->getOperand(1);
11026 if (N->getOpcode() == ISD::AND) {
11027 if (RHS.getOpcode() == ISD::BITCAST)
11028 RHS = RHS.getOperand(0);
11029 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
11030 SmallVector<int, 8> Indices;
11031 unsigned NumElts = RHS.getNumOperands();
11032 for (unsigned i = 0; i != NumElts; ++i) {
11033 SDValue Elt = RHS.getOperand(i);
11034 if (!isa<ConstantSDNode>(Elt))
11037 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
11038 Indices.push_back(i);
11039 else if (cast<ConstantSDNode>(Elt)->isNullValue())
11040 Indices.push_back(NumElts);
11045 // Let's see if the target supports this vector_shuffle.
11046 EVT RVT = RHS.getValueType();
11047 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
11050 // Return the new VECTOR_SHUFFLE node.
11051 EVT EltVT = RVT.getVectorElementType();
11052 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
11053 DAG.getConstant(0, EltVT));
11054 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), RVT, ZeroOps);
11055 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
11056 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
11057 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
11064 /// Visit a binary vector operation, like ADD.
11065 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
11066 assert(N->getValueType(0).isVector() &&
11067 "SimplifyVBinOp only works on vectors!");
11069 SDValue LHS = N->getOperand(0);
11070 SDValue RHS = N->getOperand(1);
11071 SDValue Shuffle = XformToShuffleWithZero(N);
11072 if (Shuffle.getNode()) return Shuffle;
11074 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
11076 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
11077 RHS.getOpcode() == ISD::BUILD_VECTOR) {
11078 // Check if both vectors are constants. If not bail out.
11079 if (!(cast<BuildVectorSDNode>(LHS)->isConstant() &&
11080 cast<BuildVectorSDNode>(RHS)->isConstant()))
11083 SmallVector<SDValue, 8> Ops;
11084 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
11085 SDValue LHSOp = LHS.getOperand(i);
11086 SDValue RHSOp = RHS.getOperand(i);
11088 // Can't fold divide by zero.
11089 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
11090 N->getOpcode() == ISD::FDIV) {
11091 if ((RHSOp.getOpcode() == ISD::Constant &&
11092 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
11093 (RHSOp.getOpcode() == ISD::ConstantFP &&
11094 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
11098 EVT VT = LHSOp.getValueType();
11099 EVT RVT = RHSOp.getValueType();
11101 // Integer BUILD_VECTOR operands may have types larger than the element
11102 // size (e.g., when the element type is not legal). Prior to type
11103 // legalization, the types may not match between the two BUILD_VECTORS.
11104 // Truncate one of the operands to make them match.
11105 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
11106 RHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, RHSOp);
11108 LHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), RVT, LHSOp);
11112 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(LHS), VT,
11114 if (FoldOp.getOpcode() != ISD::UNDEF &&
11115 FoldOp.getOpcode() != ISD::Constant &&
11116 FoldOp.getOpcode() != ISD::ConstantFP)
11118 Ops.push_back(FoldOp);
11119 AddToWorklist(FoldOp.getNode());
11122 if (Ops.size() == LHS.getNumOperands())
11123 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), LHS.getValueType(), Ops);
11126 // Type legalization might introduce new shuffles in the DAG.
11127 // Fold (VBinOp (shuffle (A, Undef, Mask)), (shuffle (B, Undef, Mask)))
11128 // -> (shuffle (VBinOp (A, B)), Undef, Mask).
11129 if (LegalTypes && isa<ShuffleVectorSDNode>(LHS) &&
11130 isa<ShuffleVectorSDNode>(RHS) && LHS.hasOneUse() && RHS.hasOneUse() &&
11131 LHS.getOperand(1).getOpcode() == ISD::UNDEF &&
11132 RHS.getOperand(1).getOpcode() == ISD::UNDEF) {
11133 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(LHS);
11134 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(RHS);
11136 if (SVN0->getMask().equals(SVN1->getMask())) {
11137 EVT VT = N->getValueType(0);
11138 SDValue UndefVector = LHS.getOperand(1);
11139 SDValue NewBinOp = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
11140 LHS.getOperand(0), RHS.getOperand(0));
11141 AddUsersToWorklist(N);
11142 return DAG.getVectorShuffle(VT, SDLoc(N), NewBinOp, UndefVector,
11143 &SVN0->getMask()[0]);
11150 /// Visit a binary vector operation, like FABS/FNEG.
11151 SDValue DAGCombiner::SimplifyVUnaryOp(SDNode *N) {
11152 assert(N->getValueType(0).isVector() &&
11153 "SimplifyVUnaryOp only works on vectors!");
11155 SDValue N0 = N->getOperand(0);
11157 if (N0.getOpcode() != ISD::BUILD_VECTOR)
11160 // Operand is a BUILD_VECTOR node, see if we can constant fold it.
11161 SmallVector<SDValue, 8> Ops;
11162 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
11163 SDValue Op = N0.getOperand(i);
11164 if (Op.getOpcode() != ISD::UNDEF &&
11165 Op.getOpcode() != ISD::ConstantFP)
11167 EVT EltVT = Op.getValueType();
11168 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(N0), EltVT, Op);
11169 if (FoldOp.getOpcode() != ISD::UNDEF &&
11170 FoldOp.getOpcode() != ISD::ConstantFP)
11172 Ops.push_back(FoldOp);
11173 AddToWorklist(FoldOp.getNode());
11176 if (Ops.size() != N0.getNumOperands())
11179 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), N0.getValueType(), Ops);
11182 SDValue DAGCombiner::SimplifySelect(SDLoc DL, SDValue N0,
11183 SDValue N1, SDValue N2){
11184 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
11186 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
11187 cast<CondCodeSDNode>(N0.getOperand(2))->get());
11189 // If we got a simplified select_cc node back from SimplifySelectCC, then
11190 // break it down into a new SETCC node, and a new SELECT node, and then return
11191 // the SELECT node, since we were called with a SELECT node.
11192 if (SCC.getNode()) {
11193 // Check to see if we got a select_cc back (to turn into setcc/select).
11194 // Otherwise, just return whatever node we got back, like fabs.
11195 if (SCC.getOpcode() == ISD::SELECT_CC) {
11196 SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0),
11198 SCC.getOperand(0), SCC.getOperand(1),
11199 SCC.getOperand(4));
11200 AddToWorklist(SETCC.getNode());
11201 return DAG.getSelect(SDLoc(SCC), SCC.getValueType(), SETCC,
11202 SCC.getOperand(2), SCC.getOperand(3));
11210 /// Given a SELECT or a SELECT_CC node, where LHS and RHS are the two values
11211 /// being selected between, see if we can simplify the select. Callers of this
11212 /// should assume that TheSelect is deleted if this returns true. As such, they
11213 /// should return the appropriate thing (e.g. the node) back to the top-level of
11214 /// the DAG combiner loop to avoid it being looked at.
11215 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
11218 // Cannot simplify select with vector condition
11219 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
11221 // If this is a select from two identical things, try to pull the operation
11222 // through the select.
11223 if (LHS.getOpcode() != RHS.getOpcode() ||
11224 !LHS.hasOneUse() || !RHS.hasOneUse())
11227 // If this is a load and the token chain is identical, replace the select
11228 // of two loads with a load through a select of the address to load from.
11229 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
11230 // constants have been dropped into the constant pool.
11231 if (LHS.getOpcode() == ISD::LOAD) {
11232 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
11233 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
11235 // Token chains must be identical.
11236 if (LHS.getOperand(0) != RHS.getOperand(0) ||
11237 // Do not let this transformation reduce the number of volatile loads.
11238 LLD->isVolatile() || RLD->isVolatile() ||
11239 // If this is an EXTLOAD, the VT's must match.
11240 LLD->getMemoryVT() != RLD->getMemoryVT() ||
11241 // If this is an EXTLOAD, the kind of extension must match.
11242 (LLD->getExtensionType() != RLD->getExtensionType() &&
11243 // The only exception is if one of the extensions is anyext.
11244 LLD->getExtensionType() != ISD::EXTLOAD &&
11245 RLD->getExtensionType() != ISD::EXTLOAD) ||
11246 // FIXME: this discards src value information. This is
11247 // over-conservative. It would be beneficial to be able to remember
11248 // both potential memory locations. Since we are discarding
11249 // src value info, don't do the transformation if the memory
11250 // locations are not in the default address space.
11251 LLD->getPointerInfo().getAddrSpace() != 0 ||
11252 RLD->getPointerInfo().getAddrSpace() != 0 ||
11253 !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(),
11254 LLD->getBasePtr().getValueType()))
11257 // Check that the select condition doesn't reach either load. If so,
11258 // folding this will induce a cycle into the DAG. If not, this is safe to
11259 // xform, so create a select of the addresses.
11261 if (TheSelect->getOpcode() == ISD::SELECT) {
11262 SDNode *CondNode = TheSelect->getOperand(0).getNode();
11263 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
11264 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
11266 // The loads must not depend on one another.
11267 if (LLD->isPredecessorOf(RLD) ||
11268 RLD->isPredecessorOf(LLD))
11270 Addr = DAG.getSelect(SDLoc(TheSelect),
11271 LLD->getBasePtr().getValueType(),
11272 TheSelect->getOperand(0), LLD->getBasePtr(),
11273 RLD->getBasePtr());
11274 } else { // Otherwise SELECT_CC
11275 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
11276 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
11278 if ((LLD->hasAnyUseOfValue(1) &&
11279 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
11280 (RLD->hasAnyUseOfValue(1) &&
11281 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
11284 Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect),
11285 LLD->getBasePtr().getValueType(),
11286 TheSelect->getOperand(0),
11287 TheSelect->getOperand(1),
11288 LLD->getBasePtr(), RLD->getBasePtr(),
11289 TheSelect->getOperand(4));
11293 // It is safe to replace the two loads if they have different alignments,
11294 // but the new load must be the minimum (most restrictive) alignment of the
11296 bool isInvariant = LLD->getAlignment() & RLD->getAlignment();
11297 unsigned Alignment = std::min(LLD->getAlignment(), RLD->getAlignment());
11298 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
11299 Load = DAG.getLoad(TheSelect->getValueType(0),
11301 // FIXME: Discards pointer and AA info.
11302 LLD->getChain(), Addr, MachinePointerInfo(),
11303 LLD->isVolatile(), LLD->isNonTemporal(),
11304 isInvariant, Alignment);
11306 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
11307 RLD->getExtensionType() : LLD->getExtensionType(),
11309 TheSelect->getValueType(0),
11310 // FIXME: Discards pointer and AA info.
11311 LLD->getChain(), Addr, MachinePointerInfo(),
11312 LLD->getMemoryVT(), LLD->isVolatile(),
11313 LLD->isNonTemporal(), isInvariant, Alignment);
11316 // Users of the select now use the result of the load.
11317 CombineTo(TheSelect, Load);
11319 // Users of the old loads now use the new load's chain. We know the
11320 // old-load value is dead now.
11321 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
11322 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
11329 /// Simplify an expression of the form (N0 cond N1) ? N2 : N3
11330 /// where 'cond' is the comparison specified by CC.
11331 SDValue DAGCombiner::SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1,
11332 SDValue N2, SDValue N3,
11333 ISD::CondCode CC, bool NotExtCompare) {
11334 // (x ? y : y) -> y.
11335 if (N2 == N3) return N2;
11337 EVT VT = N2.getValueType();
11338 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
11339 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
11340 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
11342 // Determine if the condition we're dealing with is constant
11343 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
11344 N0, N1, CC, DL, false);
11345 if (SCC.getNode()) AddToWorklist(SCC.getNode());
11346 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
11348 // fold select_cc true, x, y -> x
11349 if (SCCC && !SCCC->isNullValue())
11351 // fold select_cc false, x, y -> y
11352 if (SCCC && SCCC->isNullValue())
11355 // Check to see if we can simplify the select into an fabs node
11356 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
11357 // Allow either -0.0 or 0.0
11358 if (CFP->getValueAPF().isZero()) {
11359 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
11360 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
11361 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
11362 N2 == N3.getOperand(0))
11363 return DAG.getNode(ISD::FABS, DL, VT, N0);
11365 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
11366 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
11367 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
11368 N2.getOperand(0) == N3)
11369 return DAG.getNode(ISD::FABS, DL, VT, N3);
11373 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
11374 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
11375 // in it. This is a win when the constant is not otherwise available because
11376 // it replaces two constant pool loads with one. We only do this if the FP
11377 // type is known to be legal, because if it isn't, then we are before legalize
11378 // types an we want the other legalization to happen first (e.g. to avoid
11379 // messing with soft float) and if the ConstantFP is not legal, because if
11380 // it is legal, we may not need to store the FP constant in a constant pool.
11381 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
11382 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
11383 if (TLI.isTypeLegal(N2.getValueType()) &&
11384 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
11385 TargetLowering::Legal &&
11386 !TLI.isFPImmLegal(TV->getValueAPF(), TV->getValueType(0)) &&
11387 !TLI.isFPImmLegal(FV->getValueAPF(), FV->getValueType(0))) &&
11388 // If both constants have multiple uses, then we won't need to do an
11389 // extra load, they are likely around in registers for other users.
11390 (TV->hasOneUse() || FV->hasOneUse())) {
11391 Constant *Elts[] = {
11392 const_cast<ConstantFP*>(FV->getConstantFPValue()),
11393 const_cast<ConstantFP*>(TV->getConstantFPValue())
11395 Type *FPTy = Elts[0]->getType();
11396 const DataLayout &TD = *TLI.getDataLayout();
11398 // Create a ConstantArray of the two constants.
11399 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
11400 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
11401 TD.getPrefTypeAlignment(FPTy));
11402 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
11404 // Get the offsets to the 0 and 1 element of the array so that we can
11405 // select between them.
11406 SDValue Zero = DAG.getIntPtrConstant(0);
11407 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
11408 SDValue One = DAG.getIntPtrConstant(EltSize);
11410 SDValue Cond = DAG.getSetCC(DL,
11411 getSetCCResultType(N0.getValueType()),
11413 AddToWorklist(Cond.getNode());
11414 SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(),
11416 AddToWorklist(CstOffset.getNode());
11417 CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx,
11419 AddToWorklist(CPIdx.getNode());
11420 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
11421 MachinePointerInfo::getConstantPool(), false,
11422 false, false, Alignment);
11427 // Check to see if we can perform the "gzip trick", transforming
11428 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
11429 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
11430 (N1C->isNullValue() || // (a < 0) ? b : 0
11431 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
11432 EVT XType = N0.getValueType();
11433 EVT AType = N2.getValueType();
11434 if (XType.bitsGE(AType)) {
11435 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
11436 // single-bit constant.
11437 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
11438 unsigned ShCtV = N2C->getAPIntValue().logBase2();
11439 ShCtV = XType.getSizeInBits()-ShCtV-1;
11440 SDValue ShCt = DAG.getConstant(ShCtV,
11441 getShiftAmountTy(N0.getValueType()));
11442 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0),
11444 AddToWorklist(Shift.getNode());
11446 if (XType.bitsGT(AType)) {
11447 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
11448 AddToWorklist(Shift.getNode());
11451 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
11454 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0),
11456 DAG.getConstant(XType.getSizeInBits()-1,
11457 getShiftAmountTy(N0.getValueType())));
11458 AddToWorklist(Shift.getNode());
11460 if (XType.bitsGT(AType)) {
11461 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
11462 AddToWorklist(Shift.getNode());
11465 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
11469 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
11470 // where y is has a single bit set.
11471 // A plaintext description would be, we can turn the SELECT_CC into an AND
11472 // when the condition can be materialized as an all-ones register. Any
11473 // single bit-test can be materialized as an all-ones register with
11474 // shift-left and shift-right-arith.
11475 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
11476 N0->getValueType(0) == VT &&
11477 N1C && N1C->isNullValue() &&
11478 N2C && N2C->isNullValue()) {
11479 SDValue AndLHS = N0->getOperand(0);
11480 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
11481 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
11482 // Shift the tested bit over the sign bit.
11483 APInt AndMask = ConstAndRHS->getAPIntValue();
11485 DAG.getConstant(AndMask.countLeadingZeros(),
11486 getShiftAmountTy(AndLHS.getValueType()));
11487 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt);
11489 // Now arithmetic right shift it all the way over, so the result is either
11490 // all-ones, or zero.
11492 DAG.getConstant(AndMask.getBitWidth()-1,
11493 getShiftAmountTy(Shl.getValueType()));
11494 SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt);
11496 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
11500 // fold select C, 16, 0 -> shl C, 4
11501 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
11502 TLI.getBooleanContents(N0.getValueType()) ==
11503 TargetLowering::ZeroOrOneBooleanContent) {
11505 // If the caller doesn't want us to simplify this into a zext of a compare,
11507 if (NotExtCompare && N2C->getAPIntValue() == 1)
11510 // Get a SetCC of the condition
11511 // NOTE: Don't create a SETCC if it's not legal on this target.
11512 if (!LegalOperations ||
11513 TLI.isOperationLegal(ISD::SETCC,
11514 LegalTypes ? getSetCCResultType(N0.getValueType()) : MVT::i1)) {
11516 // cast from setcc result type to select result type
11518 SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()),
11520 if (N2.getValueType().bitsLT(SCC.getValueType()))
11521 Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2),
11522 N2.getValueType());
11524 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
11525 N2.getValueType(), SCC);
11527 SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC);
11528 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
11529 N2.getValueType(), SCC);
11532 AddToWorklist(SCC.getNode());
11533 AddToWorklist(Temp.getNode());
11535 if (N2C->getAPIntValue() == 1)
11538 // shl setcc result by log2 n2c
11539 return DAG.getNode(
11540 ISD::SHL, DL, N2.getValueType(), Temp,
11541 DAG.getConstant(N2C->getAPIntValue().logBase2(),
11542 getShiftAmountTy(Temp.getValueType())));
11546 // Check to see if this is the equivalent of setcc
11547 // FIXME: Turn all of these into setcc if setcc if setcc is legal
11548 // otherwise, go ahead with the folds.
11549 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
11550 EVT XType = N0.getValueType();
11551 if (!LegalOperations ||
11552 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(XType))) {
11553 SDValue Res = DAG.getSetCC(DL, getSetCCResultType(XType), N0, N1, CC);
11554 if (Res.getValueType() != VT)
11555 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
11559 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
11560 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
11561 (!LegalOperations ||
11562 TLI.isOperationLegal(ISD::CTLZ, XType))) {
11563 SDValue Ctlz = DAG.getNode(ISD::CTLZ, SDLoc(N0), XType, N0);
11564 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
11565 DAG.getConstant(Log2_32(XType.getSizeInBits()),
11566 getShiftAmountTy(Ctlz.getValueType())));
11568 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
11569 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
11570 SDValue NegN0 = DAG.getNode(ISD::SUB, SDLoc(N0),
11571 XType, DAG.getConstant(0, XType), N0);
11572 SDValue NotN0 = DAG.getNOT(SDLoc(N0), N0, XType);
11573 return DAG.getNode(ISD::SRL, DL, XType,
11574 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
11575 DAG.getConstant(XType.getSizeInBits()-1,
11576 getShiftAmountTy(XType)));
11578 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
11579 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
11580 SDValue Sign = DAG.getNode(ISD::SRL, SDLoc(N0), XType, N0,
11581 DAG.getConstant(XType.getSizeInBits()-1,
11582 getShiftAmountTy(N0.getValueType())));
11583 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
11587 // Check to see if this is an integer abs.
11588 // select_cc setg[te] X, 0, X, -X ->
11589 // select_cc setgt X, -1, X, -X ->
11590 // select_cc setl[te] X, 0, -X, X ->
11591 // select_cc setlt X, 1, -X, X ->
11592 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
11594 ConstantSDNode *SubC = nullptr;
11595 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
11596 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
11597 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
11598 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
11599 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
11600 (N1C->isOne() && CC == ISD::SETLT)) &&
11601 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
11602 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
11604 EVT XType = N0.getValueType();
11605 if (SubC && SubC->isNullValue() && XType.isInteger()) {
11606 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0), XType,
11608 DAG.getConstant(XType.getSizeInBits()-1,
11609 getShiftAmountTy(N0.getValueType())));
11610 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0),
11612 AddToWorklist(Shift.getNode());
11613 AddToWorklist(Add.getNode());
11614 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
11621 /// This is a stub for TargetLowering::SimplifySetCC.
11622 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
11623 SDValue N1, ISD::CondCode Cond,
11624 SDLoc DL, bool foldBooleans) {
11625 TargetLowering::DAGCombinerInfo
11626 DagCombineInfo(DAG, Level, false, this);
11627 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
11630 /// Given an ISD::SDIV node expressing a divide by constant, return
11631 /// a DAG expression to select that will generate the same value by multiplying
11632 /// by a magic number. See:
11633 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
11634 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
11635 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
11639 // Avoid division by zero.
11640 if (!C->getAPIntValue())
11643 std::vector<SDNode*> Built;
11645 TLI.BuildSDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
11647 for (SDNode *N : Built)
11652 /// Given an ISD::SDIV node expressing a divide by constant power of 2, return a
11653 /// DAG expression that will generate the same value by right shifting.
11654 SDValue DAGCombiner::BuildSDIVPow2(SDNode *N) {
11655 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
11659 // Avoid division by zero.
11660 if (!C->getAPIntValue())
11663 std::vector<SDNode *> Built;
11664 SDValue S = TLI.BuildSDIVPow2(N, C->getAPIntValue(), DAG, &Built);
11666 for (SDNode *N : Built)
11671 /// Given an ISD::UDIV node expressing a divide by constant, return a DAG
11672 /// expression that will generate the same value by multiplying by a magic
11674 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
11675 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
11676 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
11680 // Avoid division by zero.
11681 if (!C->getAPIntValue())
11684 std::vector<SDNode*> Built;
11686 TLI.BuildUDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
11688 for (SDNode *N : Built)
11693 /// Return true if base is a frame index, which is known not to alias with
11694 /// anything but itself. Provides base object and offset as results.
11695 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
11696 const GlobalValue *&GV, const void *&CV) {
11697 // Assume it is a primitive operation.
11698 Base = Ptr; Offset = 0; GV = nullptr; CV = nullptr;
11700 // If it's an adding a simple constant then integrate the offset.
11701 if (Base.getOpcode() == ISD::ADD) {
11702 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
11703 Base = Base.getOperand(0);
11704 Offset += C->getZExtValue();
11708 // Return the underlying GlobalValue, and update the Offset. Return false
11709 // for GlobalAddressSDNode since the same GlobalAddress may be represented
11710 // by multiple nodes with different offsets.
11711 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
11712 GV = G->getGlobal();
11713 Offset += G->getOffset();
11717 // Return the underlying Constant value, and update the Offset. Return false
11718 // for ConstantSDNodes since the same constant pool entry may be represented
11719 // by multiple nodes with different offsets.
11720 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
11721 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
11722 : (const void *)C->getConstVal();
11723 Offset += C->getOffset();
11726 // If it's any of the following then it can't alias with anything but itself.
11727 return isa<FrameIndexSDNode>(Base);
11730 /// Return true if there is any possibility that the two addresses overlap.
11731 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const {
11732 // If they are the same then they must be aliases.
11733 if (Op0->getBasePtr() == Op1->getBasePtr()) return true;
11735 // If they are both volatile then they cannot be reordered.
11736 if (Op0->isVolatile() && Op1->isVolatile()) return true;
11738 // Gather base node and offset information.
11739 SDValue Base1, Base2;
11740 int64_t Offset1, Offset2;
11741 const GlobalValue *GV1, *GV2;
11742 const void *CV1, *CV2;
11743 bool isFrameIndex1 = FindBaseOffset(Op0->getBasePtr(),
11744 Base1, Offset1, GV1, CV1);
11745 bool isFrameIndex2 = FindBaseOffset(Op1->getBasePtr(),
11746 Base2, Offset2, GV2, CV2);
11748 // If they have a same base address then check to see if they overlap.
11749 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
11750 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
11751 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
11753 // It is possible for different frame indices to alias each other, mostly
11754 // when tail call optimization reuses return address slots for arguments.
11755 // To catch this case, look up the actual index of frame indices to compute
11756 // the real alias relationship.
11757 if (isFrameIndex1 && isFrameIndex2) {
11758 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11759 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
11760 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
11761 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
11762 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
11765 // Otherwise, if we know what the bases are, and they aren't identical, then
11766 // we know they cannot alias.
11767 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
11770 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
11771 // compared to the size and offset of the access, we may be able to prove they
11772 // do not alias. This check is conservative for now to catch cases created by
11773 // splitting vector types.
11774 if ((Op0->getOriginalAlignment() == Op1->getOriginalAlignment()) &&
11775 (Op0->getSrcValueOffset() != Op1->getSrcValueOffset()) &&
11776 (Op0->getMemoryVT().getSizeInBits() >> 3 ==
11777 Op1->getMemoryVT().getSizeInBits() >> 3) &&
11778 (Op0->getOriginalAlignment() > Op0->getMemoryVT().getSizeInBits()) >> 3) {
11779 int64_t OffAlign1 = Op0->getSrcValueOffset() % Op0->getOriginalAlignment();
11780 int64_t OffAlign2 = Op1->getSrcValueOffset() % Op1->getOriginalAlignment();
11782 // There is no overlap between these relatively aligned accesses of similar
11783 // size, return no alias.
11784 if ((OffAlign1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign2 ||
11785 (OffAlign2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign1)
11789 bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0 ? CombinerGlobalAA :
11790 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
11792 if (CombinerAAOnlyFunc.getNumOccurrences() &&
11793 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
11797 Op0->getMemOperand()->getValue() && Op1->getMemOperand()->getValue()) {
11798 // Use alias analysis information.
11799 int64_t MinOffset = std::min(Op0->getSrcValueOffset(),
11800 Op1->getSrcValueOffset());
11801 int64_t Overlap1 = (Op0->getMemoryVT().getSizeInBits() >> 3) +
11802 Op0->getSrcValueOffset() - MinOffset;
11803 int64_t Overlap2 = (Op1->getMemoryVT().getSizeInBits() >> 3) +
11804 Op1->getSrcValueOffset() - MinOffset;
11805 AliasAnalysis::AliasResult AAResult =
11806 AA.alias(AliasAnalysis::Location(Op0->getMemOperand()->getValue(),
11808 UseTBAA ? Op0->getAAInfo() : AAMDNodes()),
11809 AliasAnalysis::Location(Op1->getMemOperand()->getValue(),
11811 UseTBAA ? Op1->getAAInfo() : AAMDNodes()));
11812 if (AAResult == AliasAnalysis::NoAlias)
11816 // Otherwise we have to assume they alias.
11820 /// Walk up chain skipping non-aliasing memory nodes,
11821 /// looking for aliasing nodes and adding them to the Aliases vector.
11822 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
11823 SmallVectorImpl<SDValue> &Aliases) {
11824 SmallVector<SDValue, 8> Chains; // List of chains to visit.
11825 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
11827 // Get alias information for node.
11828 bool IsLoad = isa<LoadSDNode>(N) && !cast<LSBaseSDNode>(N)->isVolatile();
11831 Chains.push_back(OriginalChain);
11832 unsigned Depth = 0;
11834 // Look at each chain and determine if it is an alias. If so, add it to the
11835 // aliases list. If not, then continue up the chain looking for the next
11837 while (!Chains.empty()) {
11838 SDValue Chain = Chains.back();
11841 // For TokenFactor nodes, look at each operand and only continue up the
11842 // chain until we find two aliases. If we've seen two aliases, assume we'll
11843 // find more and revert to original chain since the xform is unlikely to be
11846 // FIXME: The depth check could be made to return the last non-aliasing
11847 // chain we found before we hit a tokenfactor rather than the original
11849 if (Depth > 6 || Aliases.size() == 2) {
11851 Aliases.push_back(OriginalChain);
11855 // Don't bother if we've been before.
11856 if (!Visited.insert(Chain.getNode()))
11859 switch (Chain.getOpcode()) {
11860 case ISD::EntryToken:
11861 // Entry token is ideal chain operand, but handled in FindBetterChain.
11866 // Get alias information for Chain.
11867 bool IsOpLoad = isa<LoadSDNode>(Chain.getNode()) &&
11868 !cast<LSBaseSDNode>(Chain.getNode())->isVolatile();
11870 // If chain is alias then stop here.
11871 if (!(IsLoad && IsOpLoad) &&
11872 isAlias(cast<LSBaseSDNode>(N), cast<LSBaseSDNode>(Chain.getNode()))) {
11873 Aliases.push_back(Chain);
11875 // Look further up the chain.
11876 Chains.push_back(Chain.getOperand(0));
11882 case ISD::TokenFactor:
11883 // We have to check each of the operands of the token factor for "small"
11884 // token factors, so we queue them up. Adding the operands to the queue
11885 // (stack) in reverse order maintains the original order and increases the
11886 // likelihood that getNode will find a matching token factor (CSE.)
11887 if (Chain.getNumOperands() > 16) {
11888 Aliases.push_back(Chain);
11891 for (unsigned n = Chain.getNumOperands(); n;)
11892 Chains.push_back(Chain.getOperand(--n));
11897 // For all other instructions we will just have to take what we can get.
11898 Aliases.push_back(Chain);
11903 // We need to be careful here to also search for aliases through the
11904 // value operand of a store, etc. Consider the following situation:
11906 // L1 = load Token1, %52
11907 // S1 = store Token1, L1, %51
11908 // L2 = load Token1, %52+8
11909 // S2 = store Token1, L2, %51+8
11910 // Token2 = Token(S1, S2)
11911 // L3 = load Token2, %53
11912 // S3 = store Token2, L3, %52
11913 // L4 = load Token2, %53+8
11914 // S4 = store Token2, L4, %52+8
11915 // If we search for aliases of S3 (which loads address %52), and we look
11916 // only through the chain, then we'll miss the trivial dependence on L1
11917 // (which also loads from %52). We then might change all loads and
11918 // stores to use Token1 as their chain operand, which could result in
11919 // copying %53 into %52 before copying %52 into %51 (which should
11922 // The problem is, however, that searching for such data dependencies
11923 // can become expensive, and the cost is not directly related to the
11924 // chain depth. Instead, we'll rule out such configurations here by
11925 // insisting that we've visited all chain users (except for users
11926 // of the original chain, which is not necessary). When doing this,
11927 // we need to look through nodes we don't care about (otherwise, things
11928 // like register copies will interfere with trivial cases).
11930 SmallVector<const SDNode *, 16> Worklist;
11931 for (const SDNode *N : Visited)
11932 if (N != OriginalChain.getNode())
11933 Worklist.push_back(N);
11935 while (!Worklist.empty()) {
11936 const SDNode *M = Worklist.pop_back_val();
11938 // We have already visited M, and want to make sure we've visited any uses
11939 // of M that we care about. For uses that we've not visisted, and don't
11940 // care about, queue them to the worklist.
11942 for (SDNode::use_iterator UI = M->use_begin(),
11943 UIE = M->use_end(); UI != UIE; ++UI)
11944 if (UI.getUse().getValueType() == MVT::Other && Visited.insert(*UI)) {
11945 if (isa<MemIntrinsicSDNode>(*UI) || isa<MemSDNode>(*UI)) {
11946 // We've not visited this use, and we care about it (it could have an
11947 // ordering dependency with the original node).
11949 Aliases.push_back(OriginalChain);
11953 // We've not visited this use, but we don't care about it. Mark it as
11954 // visited and enqueue it to the worklist.
11955 Worklist.push_back(*UI);
11960 /// Walk up chain skipping non-aliasing memory nodes, looking for a better chain
11961 /// (aliasing node.)
11962 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
11963 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
11965 // Accumulate all the aliases to this node.
11966 GatherAllAliases(N, OldChain, Aliases);
11968 // If no operands then chain to entry token.
11969 if (Aliases.size() == 0)
11970 return DAG.getEntryNode();
11972 // If a single operand then chain to it. We don't need to revisit it.
11973 if (Aliases.size() == 1)
11976 // Construct a custom tailored token factor.
11977 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Aliases);
11980 /// This is the entry point for the file.
11981 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
11982 CodeGenOpt::Level OptLevel) {
11983 /// This is the main entry point to this class.
11984 DAGCombiner(*this, AA, OptLevel).Run(Level);