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/SmallBitVector.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/SetVector.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/Analysis/AliasAnalysis.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/DerivedTypes.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/LLVMContext.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Target/TargetLowering.h"
37 #include "llvm/Target/TargetOptions.h"
38 #include "llvm/Target/TargetRegisterInfo.h"
39 #include "llvm/Target/TargetSubtargetInfo.h"
43 #define DEBUG_TYPE "dagcombine"
45 STATISTIC(NodesCombined , "Number of dag nodes combined");
46 STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created");
47 STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created");
48 STATISTIC(OpsNarrowed , "Number of load/op/store narrowed");
49 STATISTIC(LdStFP2Int , "Number of fp load/store pairs transformed to int");
50 STATISTIC(SlicedLoads, "Number of load sliced");
54 CombinerAA("combiner-alias-analysis", cl::Hidden,
55 cl::desc("Enable DAG combiner alias-analysis heuristics"));
58 CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden,
59 cl::desc("Enable DAG combiner's use of IR alias analysis"));
62 UseTBAA("combiner-use-tbaa", cl::Hidden, cl::init(true),
63 cl::desc("Enable DAG combiner's use of TBAA"));
66 static cl::opt<std::string>
67 CombinerAAOnlyFunc("combiner-aa-only-func", cl::Hidden,
68 cl::desc("Only use DAG-combiner alias analysis in this"
72 /// Hidden option to stress test load slicing, i.e., when this option
73 /// is enabled, load slicing bypasses most of its profitability guards.
75 StressLoadSlicing("combiner-stress-load-slicing", cl::Hidden,
76 cl::desc("Bypass the profitability model of load "
81 MaySplitLoadIndex("combiner-split-load-index", cl::Hidden, cl::init(true),
82 cl::desc("DAG combiner may split indexing from loads"));
84 //------------------------------ DAGCombiner ---------------------------------//
88 const TargetLowering &TLI;
90 CodeGenOpt::Level OptLevel;
95 /// \brief Worklist of all of the nodes that need to be simplified.
97 /// This must behave as a stack -- new nodes to process are pushed onto the
98 /// back and when processing we pop off of the back.
100 /// The worklist will not contain duplicates but may contain null entries
101 /// due to nodes being deleted from the underlying DAG.
102 SmallVector<SDNode *, 64> Worklist;
104 /// \brief Mapping from an SDNode to its position on the worklist.
106 /// This is used to find and remove nodes from the worklist (by nulling
107 /// them) when they are deleted from the underlying DAG. It relies on
108 /// stable indices of nodes within the worklist.
109 DenseMap<SDNode *, unsigned> WorklistMap;
111 /// \brief Set of nodes which have been combined (at least once).
113 /// This is used to allow us to reliably add any operands of a DAG node
114 /// which have not yet been combined to the worklist.
115 SmallPtrSet<SDNode *, 64> CombinedNodes;
117 // AA - Used for DAG load/store alias analysis.
120 /// When an instruction is simplified, add all users of the instruction to
121 /// the work lists because they might get more simplified now.
122 void AddUsersToWorklist(SDNode *N) {
123 for (SDNode *Node : N->uses())
127 /// Call the node-specific routine that folds each particular type of node.
128 SDValue visit(SDNode *N);
131 /// Add to the worklist making sure its instance is at the back (next to be
133 void AddToWorklist(SDNode *N) {
134 // Skip handle nodes as they can't usefully be combined and confuse the
135 // zero-use deletion strategy.
136 if (N->getOpcode() == ISD::HANDLENODE)
139 if (WorklistMap.insert(std::make_pair(N, Worklist.size())).second)
140 Worklist.push_back(N);
143 /// Remove all instances of N from the worklist.
144 void removeFromWorklist(SDNode *N) {
145 CombinedNodes.erase(N);
147 auto It = WorklistMap.find(N);
148 if (It == WorklistMap.end())
149 return; // Not in the worklist.
151 // Null out the entry rather than erasing it to avoid a linear operation.
152 Worklist[It->second] = nullptr;
153 WorklistMap.erase(It);
156 void deleteAndRecombine(SDNode *N);
157 bool recursivelyDeleteUnusedNodes(SDNode *N);
159 SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
162 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) {
163 return CombineTo(N, &Res, 1, AddTo);
166 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1,
168 SDValue To[] = { Res0, Res1 };
169 return CombineTo(N, To, 2, AddTo);
172 void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO);
176 /// Check the specified integer node value to see if it can be simplified or
177 /// if things it uses can be simplified by bit propagation.
178 /// If so, return true.
179 bool SimplifyDemandedBits(SDValue Op) {
180 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
181 APInt Demanded = APInt::getAllOnesValue(BitWidth);
182 return SimplifyDemandedBits(Op, Demanded);
185 bool SimplifyDemandedBits(SDValue Op, const APInt &Demanded);
187 bool CombineToPreIndexedLoadStore(SDNode *N);
188 bool CombineToPostIndexedLoadStore(SDNode *N);
189 SDValue SplitIndexingFromLoad(LoadSDNode *LD);
190 bool SliceUpLoad(SDNode *N);
192 /// \brief Replace an ISD::EXTRACT_VECTOR_ELT of a load with a narrowed
195 /// \param EVE ISD::EXTRACT_VECTOR_ELT to be replaced.
196 /// \param InVecVT type of the input vector to EVE with bitcasts resolved.
197 /// \param EltNo index of the vector element to load.
198 /// \param OriginalLoad load that EVE came from to be replaced.
199 /// \returns EVE on success SDValue() on failure.
200 SDValue ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
201 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad);
202 void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad);
203 SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace);
204 SDValue SExtPromoteOperand(SDValue Op, EVT PVT);
205 SDValue ZExtPromoteOperand(SDValue Op, EVT PVT);
206 SDValue PromoteIntBinOp(SDValue Op);
207 SDValue PromoteIntShiftOp(SDValue Op);
208 SDValue PromoteExtend(SDValue Op);
209 bool PromoteLoad(SDValue Op);
211 void ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
212 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
213 ISD::NodeType ExtType);
215 /// Call the node-specific routine that knows how to fold each
216 /// particular type of node. If that doesn't do anything, try the
217 /// target-specific DAG combines.
218 SDValue combine(SDNode *N);
220 // Visitation implementation - Implement dag node combining for different
221 // node types. The semantics are as follows:
223 // SDValue.getNode() == 0 - No change was made
224 // SDValue.getNode() == N - N was replaced, is dead and has been handled.
225 // otherwise - N should be replaced by the returned Operand.
227 SDValue visitTokenFactor(SDNode *N);
228 SDValue visitMERGE_VALUES(SDNode *N);
229 SDValue visitADD(SDNode *N);
230 SDValue visitSUB(SDNode *N);
231 SDValue visitADDC(SDNode *N);
232 SDValue visitSUBC(SDNode *N);
233 SDValue visitADDE(SDNode *N);
234 SDValue visitSUBE(SDNode *N);
235 SDValue visitMUL(SDNode *N);
236 SDValue visitSDIV(SDNode *N);
237 SDValue visitUDIV(SDNode *N);
238 SDValue visitSREM(SDNode *N);
239 SDValue visitUREM(SDNode *N);
240 SDValue visitMULHU(SDNode *N);
241 SDValue visitMULHS(SDNode *N);
242 SDValue visitSMUL_LOHI(SDNode *N);
243 SDValue visitUMUL_LOHI(SDNode *N);
244 SDValue visitSMULO(SDNode *N);
245 SDValue visitUMULO(SDNode *N);
246 SDValue visitSDIVREM(SDNode *N);
247 SDValue visitUDIVREM(SDNode *N);
248 SDValue visitAND(SDNode *N);
249 SDValue 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 visitFSQRT(SDNode *N);
280 SDValue visitFCOPYSIGN(SDNode *N);
281 SDValue visitSINT_TO_FP(SDNode *N);
282 SDValue visitUINT_TO_FP(SDNode *N);
283 SDValue visitFP_TO_SINT(SDNode *N);
284 SDValue visitFP_TO_UINT(SDNode *N);
285 SDValue visitFP_ROUND(SDNode *N);
286 SDValue visitFP_ROUND_INREG(SDNode *N);
287 SDValue visitFP_EXTEND(SDNode *N);
288 SDValue visitFNEG(SDNode *N);
289 SDValue visitFABS(SDNode *N);
290 SDValue visitFCEIL(SDNode *N);
291 SDValue visitFTRUNC(SDNode *N);
292 SDValue visitFFLOOR(SDNode *N);
293 SDValue visitFMINNUM(SDNode *N);
294 SDValue visitFMAXNUM(SDNode *N);
295 SDValue visitBRCOND(SDNode *N);
296 SDValue visitBR_CC(SDNode *N);
297 SDValue visitLOAD(SDNode *N);
298 SDValue visitSTORE(SDNode *N);
299 SDValue visitINSERT_VECTOR_ELT(SDNode *N);
300 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
301 SDValue visitBUILD_VECTOR(SDNode *N);
302 SDValue visitCONCAT_VECTORS(SDNode *N);
303 SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
304 SDValue visitVECTOR_SHUFFLE(SDNode *N);
305 SDValue visitINSERT_SUBVECTOR(SDNode *N);
307 SDValue XformToShuffleWithZero(SDNode *N);
308 SDValue ReassociateOps(unsigned Opc, SDLoc DL, SDValue LHS, SDValue RHS);
310 SDValue visitShiftByConstant(SDNode *N, ConstantSDNode *Amt);
312 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
313 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
314 SDValue SimplifySelect(SDLoc DL, SDValue N0, SDValue N1, SDValue N2);
315 SDValue SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1, SDValue N2,
316 SDValue N3, ISD::CondCode CC,
317 bool NotExtCompare = false);
318 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
319 SDLoc DL, bool foldBooleans = true);
321 bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
323 bool isOneUseSetCC(SDValue N) const;
325 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
327 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
328 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
329 SDValue BuildSDIV(SDNode *N);
330 SDValue BuildSDIVPow2(SDNode *N);
331 SDValue BuildUDIV(SDNode *N);
332 SDValue BuildReciprocalEstimate(SDValue Op);
333 SDValue BuildRsqrtEstimate(SDValue Op);
334 SDValue BuildRsqrtNROneConst(SDValue Op, SDValue Est, unsigned Iterations);
335 SDValue BuildRsqrtNRTwoConst(SDValue Op, SDValue Est, unsigned Iterations);
336 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
337 bool DemandHighBits = true);
338 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
339 SDNode *MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg,
340 SDValue InnerPos, SDValue InnerNeg,
341 unsigned PosOpcode, unsigned NegOpcode,
343 SDNode *MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL);
344 SDValue ReduceLoadWidth(SDNode *N);
345 SDValue ReduceLoadOpStoreWidth(SDNode *N);
346 SDValue TransformFPLoadStorePair(SDNode *N);
347 SDValue reduceBuildVecExtToExtBuildVec(SDNode *N);
348 SDValue reduceBuildVecConvertToConvertBuildVec(SDNode *N);
350 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
352 /// Walk up chain skipping non-aliasing memory nodes,
353 /// looking for aliasing nodes and adding them to the Aliases vector.
354 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
355 SmallVectorImpl<SDValue> &Aliases);
357 /// Return true if there is any possibility that the two addresses overlap.
358 bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const;
360 /// Walk up chain skipping non-aliasing memory nodes, looking for a better
361 /// chain (aliasing node.)
362 SDValue FindBetterChain(SDNode *N, SDValue Chain);
364 /// Merge consecutive store operations into a wide store.
365 /// This optimization uses wide integers or vectors when possible.
366 /// \return True if some memory operations were changed.
367 bool MergeConsecutiveStores(StoreSDNode *N);
369 /// \brief Try to transform a truncation where C is a constant:
370 /// (trunc (and X, C)) -> (and (trunc X), (trunc C))
372 /// \p N needs to be a truncation and its first operand an AND. Other
373 /// requirements are checked by the function (e.g. that trunc is
374 /// single-use) and if missed an empty SDValue is returned.
375 SDValue distributeTruncateThroughAnd(SDNode *N);
378 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
379 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
380 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {
381 AttributeSet FnAttrs =
382 DAG.getMachineFunction().getFunction()->getAttributes();
384 FnAttrs.hasAttribute(AttributeSet::FunctionIndex,
385 Attribute::OptimizeForSize) ||
386 FnAttrs.hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize);
389 /// Runs the dag combiner on all nodes in the work list
390 void Run(CombineLevel AtLevel);
392 SelectionDAG &getDAG() const { return DAG; }
394 /// Returns a type large enough to hold any valid shift amount - before type
395 /// legalization these can be huge.
396 EVT getShiftAmountTy(EVT LHSTy) {
397 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
398 if (LHSTy.isVector())
400 return LegalTypes ? TLI.getScalarShiftAmountTy(LHSTy)
401 : TLI.getPointerTy();
404 /// This method returns true if we are running before type legalization or
405 /// if the specified VT is legal.
406 bool isTypeLegal(const EVT &VT) {
407 if (!LegalTypes) return true;
408 return TLI.isTypeLegal(VT);
411 /// Convenience wrapper around TargetLowering::getSetCCResultType
412 EVT getSetCCResultType(EVT VT) const {
413 return TLI.getSetCCResultType(*DAG.getContext(), VT);
420 /// This class is a DAGUpdateListener that removes any deleted
421 /// nodes from the worklist.
422 class WorklistRemover : public SelectionDAG::DAGUpdateListener {
425 explicit WorklistRemover(DAGCombiner &dc)
426 : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {}
428 void NodeDeleted(SDNode *N, SDNode *E) override {
429 DC.removeFromWorklist(N);
434 //===----------------------------------------------------------------------===//
435 // TargetLowering::DAGCombinerInfo implementation
436 //===----------------------------------------------------------------------===//
438 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
439 ((DAGCombiner*)DC)->AddToWorklist(N);
442 void TargetLowering::DAGCombinerInfo::RemoveFromWorklist(SDNode *N) {
443 ((DAGCombiner*)DC)->removeFromWorklist(N);
446 SDValue TargetLowering::DAGCombinerInfo::
447 CombineTo(SDNode *N, const std::vector<SDValue> &To, bool AddTo) {
448 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
451 SDValue TargetLowering::DAGCombinerInfo::
452 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
453 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
457 SDValue TargetLowering::DAGCombinerInfo::
458 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
459 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
462 void TargetLowering::DAGCombinerInfo::
463 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
464 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
467 //===----------------------------------------------------------------------===//
469 //===----------------------------------------------------------------------===//
471 void DAGCombiner::deleteAndRecombine(SDNode *N) {
472 removeFromWorklist(N);
474 // If the operands of this node are only used by the node, they will now be
475 // dead. Make sure to re-visit them and recursively delete dead nodes.
476 for (const SDValue &Op : N->ops())
477 // For an operand generating multiple values, one of the values may
478 // become dead allowing further simplification (e.g. split index
479 // arithmetic from an indexed load).
480 if (Op->hasOneUse() || Op->getNumValues() > 1)
481 AddToWorklist(Op.getNode());
486 /// Return 1 if we can compute the negated form of the specified expression for
487 /// the same cost as the expression itself, or 2 if we can compute the negated
488 /// form more cheaply than the expression itself.
489 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
490 const TargetLowering &TLI,
491 const TargetOptions *Options,
492 unsigned Depth = 0) {
493 // fneg is removable even if it has multiple uses.
494 if (Op.getOpcode() == ISD::FNEG) return 2;
496 // Don't allow anything with multiple uses.
497 if (!Op.hasOneUse()) return 0;
499 // Don't recurse exponentially.
500 if (Depth > 6) return 0;
502 switch (Op.getOpcode()) {
503 default: return false;
504 case ISD::ConstantFP:
505 // Don't invert constant FP values after legalize. The negated constant
506 // isn't necessarily legal.
507 return LegalOperations ? 0 : 1;
509 // FIXME: determine better conditions for this xform.
510 if (!Options->UnsafeFPMath) return 0;
512 // After operation legalization, it might not be legal to create new FSUBs.
513 if (LegalOperations &&
514 !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType()))
517 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
518 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
521 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
522 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
525 // We can't turn -(A-B) into B-A when we honor signed zeros.
526 if (!Options->UnsafeFPMath) return 0;
528 // fold (fneg (fsub A, B)) -> (fsub B, A)
533 if (Options->HonorSignDependentRoundingFPMath()) return 0;
535 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
536 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
540 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
546 return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options,
551 /// If isNegatibleForFree returns true, return the newly negated expression.
552 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
553 bool LegalOperations, unsigned Depth = 0) {
554 const TargetOptions &Options = DAG.getTarget().Options;
555 // fneg is removable even if it has multiple uses.
556 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
558 // Don't allow anything with multiple uses.
559 assert(Op.hasOneUse() && "Unknown reuse!");
561 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
562 switch (Op.getOpcode()) {
563 default: llvm_unreachable("Unknown code");
564 case ISD::ConstantFP: {
565 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
567 return DAG.getConstantFP(V, Op.getValueType());
570 // FIXME: determine better conditions for this xform.
571 assert(Options.UnsafeFPMath);
573 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
574 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
575 DAG.getTargetLoweringInfo(), &Options, Depth+1))
576 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
577 GetNegatedExpression(Op.getOperand(0), DAG,
578 LegalOperations, Depth+1),
580 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
581 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
582 GetNegatedExpression(Op.getOperand(1), DAG,
583 LegalOperations, Depth+1),
586 // We can't turn -(A-B) into B-A when we honor signed zeros.
587 assert(Options.UnsafeFPMath);
589 // fold (fneg (fsub 0, B)) -> B
590 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
591 if (N0CFP->getValueAPF().isZero())
592 return Op.getOperand(1);
594 // fold (fneg (fsub A, B)) -> (fsub B, A)
595 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
596 Op.getOperand(1), Op.getOperand(0));
600 assert(!Options.HonorSignDependentRoundingFPMath());
602 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
603 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
604 DAG.getTargetLoweringInfo(), &Options, Depth+1))
605 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
606 GetNegatedExpression(Op.getOperand(0), DAG,
607 LegalOperations, Depth+1),
610 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
611 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
613 GetNegatedExpression(Op.getOperand(1), DAG,
614 LegalOperations, Depth+1));
618 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
619 GetNegatedExpression(Op.getOperand(0), DAG,
620 LegalOperations, Depth+1));
622 return DAG.getNode(ISD::FP_ROUND, SDLoc(Op), Op.getValueType(),
623 GetNegatedExpression(Op.getOperand(0), DAG,
624 LegalOperations, Depth+1),
629 // Return true if this node is a setcc, or is a select_cc
630 // that selects between the target values used for true and false, making it
631 // equivalent to a setcc. Also, set the incoming LHS, RHS, and CC references to
632 // the appropriate nodes based on the type of node we are checking. This
633 // simplifies life a bit for the callers.
634 bool DAGCombiner::isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
636 if (N.getOpcode() == ISD::SETCC) {
637 LHS = N.getOperand(0);
638 RHS = N.getOperand(1);
639 CC = N.getOperand(2);
643 if (N.getOpcode() != ISD::SELECT_CC ||
644 !TLI.isConstTrueVal(N.getOperand(2).getNode()) ||
645 !TLI.isConstFalseVal(N.getOperand(3).getNode()))
648 LHS = N.getOperand(0);
649 RHS = N.getOperand(1);
650 CC = N.getOperand(4);
654 /// Return true if this is a SetCC-equivalent operation with only one use.
655 /// If this is true, it allows the users to invert the operation for free when
656 /// it is profitable to do so.
657 bool DAGCombiner::isOneUseSetCC(SDValue N) const {
659 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
664 /// Returns true if N is a BUILD_VECTOR node whose
665 /// elements are all the same constant or undefined.
666 static bool isConstantSplatVector(SDNode *N, APInt& SplatValue) {
667 BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(N);
672 unsigned SplatBitSize;
674 EVT EltVT = N->getValueType(0).getVectorElementType();
675 return (C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
677 EltVT.getSizeInBits() >= SplatBitSize);
680 // \brief Returns the SDNode if it is a constant BuildVector or constant.
681 static SDNode *isConstantBuildVectorOrConstantInt(SDValue N) {
682 if (isa<ConstantSDNode>(N))
684 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N);
685 if (BV && BV->isConstant())
690 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
692 static ConstantSDNode *isConstOrConstSplat(SDValue N) {
693 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N))
696 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
697 BitVector UndefElements;
698 ConstantSDNode *CN = BV->getConstantSplatNode(&UndefElements);
700 // BuildVectors can truncate their operands. Ignore that case here.
701 // FIXME: We blindly ignore splats which include undef which is overly
703 if (CN && UndefElements.none() &&
704 CN->getValueType(0) == N.getValueType().getScalarType())
711 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
713 static ConstantFPSDNode *isConstOrConstSplatFP(SDValue N) {
714 if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N))
717 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
718 BitVector UndefElements;
719 ConstantFPSDNode *CN = BV->getConstantFPSplatNode(&UndefElements);
721 if (CN && UndefElements.none())
728 SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL,
729 SDValue N0, SDValue N1) {
730 EVT VT = N0.getValueType();
731 if (N0.getOpcode() == Opc) {
732 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0.getOperand(1))) {
733 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1)) {
734 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
735 SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, L, R);
736 if (!OpNode.getNode())
738 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
740 if (N0.hasOneUse()) {
741 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one
743 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1);
744 if (!OpNode.getNode())
746 AddToWorklist(OpNode.getNode());
747 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
752 if (N1.getOpcode() == Opc) {
753 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1.getOperand(1))) {
754 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0)) {
755 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
756 SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, R, L);
757 if (!OpNode.getNode())
759 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
761 if (N1.hasOneUse()) {
762 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one
764 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N1.getOperand(0), N0);
765 if (!OpNode.getNode())
767 AddToWorklist(OpNode.getNode());
768 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
776 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
778 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
780 DEBUG(dbgs() << "\nReplacing.1 ";
782 dbgs() << "\nWith: ";
783 To[0].getNode()->dump(&DAG);
784 dbgs() << " and " << NumTo-1 << " other values\n";
785 for (unsigned i = 0, e = NumTo; i != e; ++i)
786 assert((!To[i].getNode() ||
787 N->getValueType(i) == To[i].getValueType()) &&
788 "Cannot combine value to value of different type!"));
789 WorklistRemover DeadNodes(*this);
790 DAG.ReplaceAllUsesWith(N, To);
792 // Push the new nodes and any users onto the worklist
793 for (unsigned i = 0, e = NumTo; i != e; ++i) {
794 if (To[i].getNode()) {
795 AddToWorklist(To[i].getNode());
796 AddUsersToWorklist(To[i].getNode());
801 // Finally, if the node is now dead, remove it from the graph. The node
802 // may not be dead if the replacement process recursively simplified to
803 // something else needing this node.
805 deleteAndRecombine(N);
806 return SDValue(N, 0);
810 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
811 // Replace all uses. If any nodes become isomorphic to other nodes and
812 // are deleted, make sure to remove them from our worklist.
813 WorklistRemover DeadNodes(*this);
814 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New);
816 // Push the new node and any (possibly new) users onto the worklist.
817 AddToWorklist(TLO.New.getNode());
818 AddUsersToWorklist(TLO.New.getNode());
820 // Finally, if the node is now dead, remove it from the graph. The node
821 // may not be dead if the replacement process recursively simplified to
822 // something else needing this node.
823 if (TLO.Old.getNode()->use_empty())
824 deleteAndRecombine(TLO.Old.getNode());
827 /// Check the specified integer node value to see if it can be simplified or if
828 /// things it uses can be simplified by bit propagation. If so, return true.
829 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
830 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
831 APInt KnownZero, KnownOne;
832 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
836 AddToWorklist(Op.getNode());
838 // Replace the old value with the new one.
840 DEBUG(dbgs() << "\nReplacing.2 ";
841 TLO.Old.getNode()->dump(&DAG);
842 dbgs() << "\nWith: ";
843 TLO.New.getNode()->dump(&DAG);
846 CommitTargetLoweringOpt(TLO);
850 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
852 EVT VT = Load->getValueType(0);
853 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
855 DEBUG(dbgs() << "\nReplacing.9 ";
857 dbgs() << "\nWith: ";
858 Trunc.getNode()->dump(&DAG);
860 WorklistRemover DeadNodes(*this);
861 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc);
862 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1));
863 deleteAndRecombine(Load);
864 AddToWorklist(Trunc.getNode());
867 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
870 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
871 EVT MemVT = LD->getMemoryVT();
872 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
873 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
875 : LD->getExtensionType();
877 return DAG.getExtLoad(ExtType, dl, PVT,
878 LD->getChain(), LD->getBasePtr(),
879 MemVT, LD->getMemOperand());
882 unsigned Opc = Op.getOpcode();
885 case ISD::AssertSext:
886 return DAG.getNode(ISD::AssertSext, dl, PVT,
887 SExtPromoteOperand(Op.getOperand(0), PVT),
889 case ISD::AssertZext:
890 return DAG.getNode(ISD::AssertZext, dl, PVT,
891 ZExtPromoteOperand(Op.getOperand(0), PVT),
893 case ISD::Constant: {
895 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
896 return DAG.getNode(ExtOpc, dl, PVT, Op);
900 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
902 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
905 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
906 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
908 EVT OldVT = Op.getValueType();
910 bool Replace = false;
911 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
912 if (!NewOp.getNode())
914 AddToWorklist(NewOp.getNode());
917 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
918 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
919 DAG.getValueType(OldVT));
922 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
923 EVT OldVT = Op.getValueType();
925 bool Replace = false;
926 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
927 if (!NewOp.getNode())
929 AddToWorklist(NewOp.getNode());
932 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
933 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
936 /// Promote the specified integer binary operation if the target indicates it is
937 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
938 /// i32 since i16 instructions are longer.
939 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
940 if (!LegalOperations)
943 EVT VT = Op.getValueType();
944 if (VT.isVector() || !VT.isInteger())
947 // If operation type is 'undesirable', e.g. i16 on x86, consider
949 unsigned Opc = Op.getOpcode();
950 if (TLI.isTypeDesirableForOp(Opc, VT))
954 // Consult target whether it is a good idea to promote this operation and
955 // what's the right type to promote it to.
956 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
957 assert(PVT != VT && "Don't know what type to promote to!");
959 bool Replace0 = false;
960 SDValue N0 = Op.getOperand(0);
961 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
965 bool Replace1 = false;
966 SDValue N1 = Op.getOperand(1);
971 NN1 = PromoteOperand(N1, PVT, Replace1);
976 AddToWorklist(NN0.getNode());
978 AddToWorklist(NN1.getNode());
981 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
983 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
985 DEBUG(dbgs() << "\nPromoting ";
986 Op.getNode()->dump(&DAG));
988 return DAG.getNode(ISD::TRUNCATE, dl, VT,
989 DAG.getNode(Opc, dl, PVT, NN0, NN1));
994 /// Promote the specified integer shift operation if the target indicates it is
995 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
996 /// i32 since i16 instructions are longer.
997 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
998 if (!LegalOperations)
1001 EVT VT = Op.getValueType();
1002 if (VT.isVector() || !VT.isInteger())
1005 // If operation type is 'undesirable', e.g. i16 on x86, consider
1007 unsigned Opc = Op.getOpcode();
1008 if (TLI.isTypeDesirableForOp(Opc, VT))
1012 // Consult target whether it is a good idea to promote this operation and
1013 // what's the right type to promote it to.
1014 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1015 assert(PVT != VT && "Don't know what type to promote to!");
1017 bool Replace = false;
1018 SDValue N0 = Op.getOperand(0);
1019 if (Opc == ISD::SRA)
1020 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
1021 else if (Opc == ISD::SRL)
1022 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
1024 N0 = PromoteOperand(N0, PVT, Replace);
1028 AddToWorklist(N0.getNode());
1030 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
1032 DEBUG(dbgs() << "\nPromoting ";
1033 Op.getNode()->dump(&DAG));
1035 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1036 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
1041 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
1042 if (!LegalOperations)
1045 EVT VT = Op.getValueType();
1046 if (VT.isVector() || !VT.isInteger())
1049 // If operation type is 'undesirable', e.g. i16 on x86, consider
1051 unsigned Opc = Op.getOpcode();
1052 if (TLI.isTypeDesirableForOp(Opc, VT))
1056 // Consult target whether it is a good idea to promote this operation and
1057 // what's the right type to promote it to.
1058 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1059 assert(PVT != VT && "Don't know what type to promote to!");
1060 // fold (aext (aext x)) -> (aext x)
1061 // fold (aext (zext x)) -> (zext x)
1062 // fold (aext (sext x)) -> (sext x)
1063 DEBUG(dbgs() << "\nPromoting ";
1064 Op.getNode()->dump(&DAG));
1065 return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0));
1070 bool DAGCombiner::PromoteLoad(SDValue Op) {
1071 if (!LegalOperations)
1074 EVT VT = Op.getValueType();
1075 if (VT.isVector() || !VT.isInteger())
1078 // If operation type is 'undesirable', e.g. i16 on x86, consider
1080 unsigned Opc = Op.getOpcode();
1081 if (TLI.isTypeDesirableForOp(Opc, VT))
1085 // Consult target whether it is a good idea to promote this operation and
1086 // what's the right type to promote it to.
1087 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1088 assert(PVT != VT && "Don't know what type to promote to!");
1091 SDNode *N = Op.getNode();
1092 LoadSDNode *LD = cast<LoadSDNode>(N);
1093 EVT MemVT = LD->getMemoryVT();
1094 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
1095 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
1097 : LD->getExtensionType();
1098 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
1099 LD->getChain(), LD->getBasePtr(),
1100 MemVT, LD->getMemOperand());
1101 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
1103 DEBUG(dbgs() << "\nPromoting ";
1106 Result.getNode()->dump(&DAG);
1108 WorklistRemover DeadNodes(*this);
1109 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
1110 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1));
1111 deleteAndRecombine(N);
1112 AddToWorklist(Result.getNode());
1118 /// \brief Recursively delete a node which has no uses and any operands for
1119 /// which it is the only use.
1121 /// Note that this both deletes the nodes and removes them from the worklist.
1122 /// It also adds any nodes who have had a user deleted to the worklist as they
1123 /// may now have only one use and subject to other combines.
1124 bool DAGCombiner::recursivelyDeleteUnusedNodes(SDNode *N) {
1125 if (!N->use_empty())
1128 SmallSetVector<SDNode *, 16> Nodes;
1131 N = Nodes.pop_back_val();
1135 if (N->use_empty()) {
1136 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1137 Nodes.insert(N->getOperand(i).getNode());
1139 removeFromWorklist(N);
1144 } while (!Nodes.empty());
1148 //===----------------------------------------------------------------------===//
1149 // Main DAG Combiner implementation
1150 //===----------------------------------------------------------------------===//
1152 void DAGCombiner::Run(CombineLevel AtLevel) {
1153 // set the instance variables, so that the various visit routines may use it.
1155 LegalOperations = Level >= AfterLegalizeVectorOps;
1156 LegalTypes = Level >= AfterLegalizeTypes;
1158 // Add all the dag nodes to the worklist.
1159 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
1160 E = DAG.allnodes_end(); I != E; ++I)
1163 // Create a dummy node (which is not added to allnodes), that adds a reference
1164 // to the root node, preventing it from being deleted, and tracking any
1165 // changes of the root.
1166 HandleSDNode Dummy(DAG.getRoot());
1168 // while the worklist isn't empty, find a node and
1169 // try and combine it.
1170 while (!WorklistMap.empty()) {
1172 // The Worklist holds the SDNodes in order, but it may contain null entries.
1174 N = Worklist.pop_back_val();
1177 bool GoodWorklistEntry = WorklistMap.erase(N);
1178 (void)GoodWorklistEntry;
1179 assert(GoodWorklistEntry &&
1180 "Found a worklist entry without a corresponding map entry!");
1182 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1183 // N is deleted from the DAG, since they too may now be dead or may have a
1184 // reduced number of uses, allowing other xforms.
1185 if (recursivelyDeleteUnusedNodes(N))
1188 WorklistRemover DeadNodes(*this);
1190 // If this combine is running after legalizing the DAG, re-legalize any
1191 // nodes pulled off the worklist.
1192 if (Level == AfterLegalizeDAG) {
1193 SmallSetVector<SDNode *, 16> UpdatedNodes;
1194 bool NIsValid = DAG.LegalizeOp(N, UpdatedNodes);
1196 for (SDNode *LN : UpdatedNodes) {
1198 AddUsersToWorklist(LN);
1204 DEBUG(dbgs() << "\nCombining: "; N->dump(&DAG));
1206 // Add any operands of the new node which have not yet been combined to the
1207 // worklist as well. Because the worklist uniques things already, this
1208 // won't repeatedly process the same operand.
1209 CombinedNodes.insert(N);
1210 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1211 if (!CombinedNodes.count(N->getOperand(i).getNode()))
1212 AddToWorklist(N->getOperand(i).getNode());
1214 SDValue RV = combine(N);
1221 // If we get back the same node we passed in, rather than a new node or
1222 // zero, we know that the node must have defined multiple values and
1223 // CombineTo was used. Since CombineTo takes care of the worklist
1224 // mechanics for us, we have no work to do in this case.
1225 if (RV.getNode() == N)
1228 assert(N->getOpcode() != ISD::DELETED_NODE &&
1229 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1230 "Node was deleted but visit returned new node!");
1232 DEBUG(dbgs() << " ... into: ";
1233 RV.getNode()->dump(&DAG));
1235 // Transfer debug value.
1236 DAG.TransferDbgValues(SDValue(N, 0), RV);
1237 if (N->getNumValues() == RV.getNode()->getNumValues())
1238 DAG.ReplaceAllUsesWith(N, RV.getNode());
1240 assert(N->getValueType(0) == RV.getValueType() &&
1241 N->getNumValues() == 1 && "Type mismatch");
1243 DAG.ReplaceAllUsesWith(N, &OpV);
1246 // Push the new node and any users onto the worklist
1247 AddToWorklist(RV.getNode());
1248 AddUsersToWorklist(RV.getNode());
1250 // Finally, if the node is now dead, remove it from the graph. The node
1251 // may not be dead if the replacement process recursively simplified to
1252 // something else needing this node. This will also take care of adding any
1253 // operands which have lost a user to the worklist.
1254 recursivelyDeleteUnusedNodes(N);
1257 // If the root changed (e.g. it was a dead load, update the root).
1258 DAG.setRoot(Dummy.getValue());
1259 DAG.RemoveDeadNodes();
1262 SDValue DAGCombiner::visit(SDNode *N) {
1263 switch (N->getOpcode()) {
1265 case ISD::TokenFactor: return visitTokenFactor(N);
1266 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1267 case ISD::ADD: return visitADD(N);
1268 case ISD::SUB: return visitSUB(N);
1269 case ISD::ADDC: return visitADDC(N);
1270 case ISD::SUBC: return visitSUBC(N);
1271 case ISD::ADDE: return visitADDE(N);
1272 case ISD::SUBE: return visitSUBE(N);
1273 case ISD::MUL: return visitMUL(N);
1274 case ISD::SDIV: return visitSDIV(N);
1275 case ISD::UDIV: return visitUDIV(N);
1276 case ISD::SREM: return visitSREM(N);
1277 case ISD::UREM: return visitUREM(N);
1278 case ISD::MULHU: return visitMULHU(N);
1279 case ISD::MULHS: return visitMULHS(N);
1280 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1281 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1282 case ISD::SMULO: return visitSMULO(N);
1283 case ISD::UMULO: return visitUMULO(N);
1284 case ISD::SDIVREM: return visitSDIVREM(N);
1285 case ISD::UDIVREM: return visitUDIVREM(N);
1286 case ISD::AND: return visitAND(N);
1287 case ISD::OR: return visitOR(N);
1288 case ISD::XOR: return visitXOR(N);
1289 case ISD::SHL: return visitSHL(N);
1290 case ISD::SRA: return visitSRA(N);
1291 case ISD::SRL: return visitSRL(N);
1293 case ISD::ROTL: return visitRotate(N);
1294 case ISD::CTLZ: return visitCTLZ(N);
1295 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1296 case ISD::CTTZ: return visitCTTZ(N);
1297 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1298 case ISD::CTPOP: return visitCTPOP(N);
1299 case ISD::SELECT: return visitSELECT(N);
1300 case ISD::VSELECT: return visitVSELECT(N);
1301 case ISD::SELECT_CC: return visitSELECT_CC(N);
1302 case ISD::SETCC: return visitSETCC(N);
1303 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1304 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1305 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1306 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1307 case ISD::TRUNCATE: return visitTRUNCATE(N);
1308 case ISD::BITCAST: return visitBITCAST(N);
1309 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1310 case ISD::FADD: return visitFADD(N);
1311 case ISD::FSUB: return visitFSUB(N);
1312 case ISD::FMUL: return visitFMUL(N);
1313 case ISD::FMA: return visitFMA(N);
1314 case ISD::FDIV: return visitFDIV(N);
1315 case ISD::FREM: return visitFREM(N);
1316 case ISD::FSQRT: return visitFSQRT(N);
1317 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1318 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1319 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1320 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1321 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1322 case ISD::FP_ROUND: return visitFP_ROUND(N);
1323 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1324 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1325 case ISD::FNEG: return visitFNEG(N);
1326 case ISD::FABS: return visitFABS(N);
1327 case ISD::FFLOOR: return visitFFLOOR(N);
1328 case ISD::FMINNUM: return visitFMINNUM(N);
1329 case ISD::FMAXNUM: return visitFMAXNUM(N);
1330 case ISD::FCEIL: return visitFCEIL(N);
1331 case ISD::FTRUNC: return visitFTRUNC(N);
1332 case ISD::BRCOND: return visitBRCOND(N);
1333 case ISD::BR_CC: return visitBR_CC(N);
1334 case ISD::LOAD: return visitLOAD(N);
1335 case ISD::STORE: return visitSTORE(N);
1336 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1337 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1338 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1339 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1340 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1341 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1342 case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N);
1347 SDValue DAGCombiner::combine(SDNode *N) {
1348 SDValue RV = visit(N);
1350 // If nothing happened, try a target-specific DAG combine.
1351 if (!RV.getNode()) {
1352 assert(N->getOpcode() != ISD::DELETED_NODE &&
1353 "Node was deleted but visit returned NULL!");
1355 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1356 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1358 // Expose the DAG combiner to the target combiner impls.
1359 TargetLowering::DAGCombinerInfo
1360 DagCombineInfo(DAG, Level, false, this);
1362 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1366 // If nothing happened still, try promoting the operation.
1367 if (!RV.getNode()) {
1368 switch (N->getOpcode()) {
1376 RV = PromoteIntBinOp(SDValue(N, 0));
1381 RV = PromoteIntShiftOp(SDValue(N, 0));
1383 case ISD::SIGN_EXTEND:
1384 case ISD::ZERO_EXTEND:
1385 case ISD::ANY_EXTEND:
1386 RV = PromoteExtend(SDValue(N, 0));
1389 if (PromoteLoad(SDValue(N, 0)))
1395 // If N is a commutative binary node, try commuting it to enable more
1397 if (!RV.getNode() && SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1398 N->getNumValues() == 1) {
1399 SDValue N0 = N->getOperand(0);
1400 SDValue N1 = N->getOperand(1);
1402 // Constant operands are canonicalized to RHS.
1403 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1404 SDValue Ops[] = {N1, N0};
1406 if (const BinaryWithFlagsSDNode *BinNode =
1407 dyn_cast<BinaryWithFlagsSDNode>(N)) {
1408 CSENode = DAG.getNodeIfExists(
1409 N->getOpcode(), N->getVTList(), Ops, BinNode->hasNoUnsignedWrap(),
1410 BinNode->hasNoSignedWrap(), BinNode->isExact());
1412 CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops);
1415 return SDValue(CSENode, 0);
1422 /// Given a node, return its input chain if it has one, otherwise return a null
1424 static SDValue getInputChainForNode(SDNode *N) {
1425 if (unsigned NumOps = N->getNumOperands()) {
1426 if (N->getOperand(0).getValueType() == MVT::Other)
1427 return N->getOperand(0);
1428 if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1429 return N->getOperand(NumOps-1);
1430 for (unsigned i = 1; i < NumOps-1; ++i)
1431 if (N->getOperand(i).getValueType() == MVT::Other)
1432 return N->getOperand(i);
1437 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1438 // If N has two operands, where one has an input chain equal to the other,
1439 // the 'other' chain is redundant.
1440 if (N->getNumOperands() == 2) {
1441 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1442 return N->getOperand(0);
1443 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1444 return N->getOperand(1);
1447 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1448 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1449 SmallPtrSet<SDNode*, 16> SeenOps;
1450 bool Changed = false; // If we should replace this token factor.
1452 // Start out with this token factor.
1455 // Iterate through token factors. The TFs grows when new token factors are
1457 for (unsigned i = 0; i < TFs.size(); ++i) {
1458 SDNode *TF = TFs[i];
1460 // Check each of the operands.
1461 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1462 SDValue Op = TF->getOperand(i);
1464 switch (Op.getOpcode()) {
1465 case ISD::EntryToken:
1466 // Entry tokens don't need to be added to the list. They are
1471 case ISD::TokenFactor:
1472 if (Op.hasOneUse() &&
1473 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1474 // Queue up for processing.
1475 TFs.push_back(Op.getNode());
1476 // Clean up in case the token factor is removed.
1477 AddToWorklist(Op.getNode());
1484 // Only add if it isn't already in the list.
1485 if (SeenOps.insert(Op.getNode()))
1496 // If we've change things around then replace token factor.
1499 // The entry token is the only possible outcome.
1500 Result = DAG.getEntryNode();
1502 // New and improved token factor.
1503 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Ops);
1506 // Don't add users to work list.
1507 return CombineTo(N, Result, false);
1513 /// MERGE_VALUES can always be eliminated.
1514 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1515 WorklistRemover DeadNodes(*this);
1516 // Replacing results may cause a different MERGE_VALUES to suddenly
1517 // be CSE'd with N, and carry its uses with it. Iterate until no
1518 // uses remain, to ensure that the node can be safely deleted.
1519 // First add the users of this node to the work list so that they
1520 // can be tried again once they have new operands.
1521 AddUsersToWorklist(N);
1523 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1524 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1525 } while (!N->use_empty());
1526 deleteAndRecombine(N);
1527 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1530 SDValue DAGCombiner::visitADD(SDNode *N) {
1531 SDValue N0 = N->getOperand(0);
1532 SDValue N1 = N->getOperand(1);
1533 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1534 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1535 EVT VT = N0.getValueType();
1538 if (VT.isVector()) {
1539 SDValue FoldedVOp = SimplifyVBinOp(N);
1540 if (FoldedVOp.getNode()) return FoldedVOp;
1542 // fold (add x, 0) -> x, vector edition
1543 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1545 if (ISD::isBuildVectorAllZeros(N0.getNode()))
1549 // fold (add x, undef) -> undef
1550 if (N0.getOpcode() == ISD::UNDEF)
1552 if (N1.getOpcode() == ISD::UNDEF)
1554 // fold (add c1, c2) -> c1+c2
1556 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1557 // canonicalize constant to RHS
1559 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, N0);
1560 // fold (add x, 0) -> x
1561 if (N1C && N1C->isNullValue())
1563 // fold (add Sym, c) -> Sym+c
1564 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1565 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1566 GA->getOpcode() == ISD::GlobalAddress)
1567 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1569 (uint64_t)N1C->getSExtValue());
1570 // fold ((c1-A)+c2) -> (c1+c2)-A
1571 if (N1C && N0.getOpcode() == ISD::SUB)
1572 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1573 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1574 DAG.getConstant(N1C->getAPIntValue()+
1575 N0C->getAPIntValue(), VT),
1578 SDValue RADD = ReassociateOps(ISD::ADD, SDLoc(N), N0, N1);
1581 // fold ((0-A) + B) -> B-A
1582 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1583 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1584 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, N0.getOperand(1));
1585 // fold (A + (0-B)) -> A-B
1586 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1587 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1588 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1.getOperand(1));
1589 // fold (A+(B-A)) -> B
1590 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1591 return N1.getOperand(0);
1592 // fold ((B-A)+A) -> B
1593 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1594 return N0.getOperand(0);
1595 // fold (A+(B-(A+C))) to (B-C)
1596 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1597 N0 == N1.getOperand(1).getOperand(0))
1598 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1599 N1.getOperand(1).getOperand(1));
1600 // fold (A+(B-(C+A))) to (B-C)
1601 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1602 N0 == N1.getOperand(1).getOperand(1))
1603 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1604 N1.getOperand(1).getOperand(0));
1605 // fold (A+((B-A)+or-C)) to (B+or-C)
1606 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1607 N1.getOperand(0).getOpcode() == ISD::SUB &&
1608 N0 == N1.getOperand(0).getOperand(1))
1609 return DAG.getNode(N1.getOpcode(), SDLoc(N), VT,
1610 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1612 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1613 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1614 SDValue N00 = N0.getOperand(0);
1615 SDValue N01 = N0.getOperand(1);
1616 SDValue N10 = N1.getOperand(0);
1617 SDValue N11 = N1.getOperand(1);
1619 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1620 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1621 DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10),
1622 DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11));
1625 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1626 return SDValue(N, 0);
1628 // fold (a+b) -> (a|b) iff a and b share no bits.
1629 if (VT.isInteger() && !VT.isVector()) {
1630 APInt LHSZero, LHSOne;
1631 APInt RHSZero, RHSOne;
1632 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1634 if (LHSZero.getBoolValue()) {
1635 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1637 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1638 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1639 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero){
1640 if (!LegalOperations || TLI.isOperationLegal(ISD::OR, VT))
1641 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1);
1646 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1647 if (N1.getOpcode() == ISD::SHL &&
1648 N1.getOperand(0).getOpcode() == ISD::SUB)
1649 if (ConstantSDNode *C =
1650 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1651 if (C->getAPIntValue() == 0)
1652 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0,
1653 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1654 N1.getOperand(0).getOperand(1),
1656 if (N0.getOpcode() == ISD::SHL &&
1657 N0.getOperand(0).getOpcode() == ISD::SUB)
1658 if (ConstantSDNode *C =
1659 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1660 if (C->getAPIntValue() == 0)
1661 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1,
1662 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1663 N0.getOperand(0).getOperand(1),
1666 if (N1.getOpcode() == ISD::AND) {
1667 SDValue AndOp0 = N1.getOperand(0);
1668 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1669 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1670 unsigned DestBits = VT.getScalarType().getSizeInBits();
1672 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1673 // and similar xforms where the inner op is either ~0 or 0.
1674 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1676 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1680 // add (sext i1), X -> sub X, (zext i1)
1681 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1682 N0.getOperand(0).getValueType() == MVT::i1 &&
1683 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1685 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1686 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1689 // add X, (sextinreg Y i1) -> sub X, (and Y 1)
1690 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
1691 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
1692 if (TN->getVT() == MVT::i1) {
1694 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
1695 DAG.getConstant(1, VT));
1696 return DAG.getNode(ISD::SUB, DL, VT, N0, ZExt);
1703 SDValue DAGCombiner::visitADDC(SDNode *N) {
1704 SDValue N0 = N->getOperand(0);
1705 SDValue N1 = N->getOperand(1);
1706 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1707 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1708 EVT VT = N0.getValueType();
1710 // If the flag result is dead, turn this into an ADD.
1711 if (!N->hasAnyUseOfValue(1))
1712 return CombineTo(N, DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N1),
1713 DAG.getNode(ISD::CARRY_FALSE,
1714 SDLoc(N), MVT::Glue));
1716 // canonicalize constant to RHS.
1718 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N1, N0);
1720 // fold (addc x, 0) -> x + no carry out
1721 if (N1C && N1C->isNullValue())
1722 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1723 SDLoc(N), MVT::Glue));
1725 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1726 APInt LHSZero, LHSOne;
1727 APInt RHSZero, RHSOne;
1728 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1730 if (LHSZero.getBoolValue()) {
1731 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1733 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1734 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1735 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1736 return CombineTo(N, DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1),
1737 DAG.getNode(ISD::CARRY_FALSE,
1738 SDLoc(N), MVT::Glue));
1744 SDValue DAGCombiner::visitADDE(SDNode *N) {
1745 SDValue N0 = N->getOperand(0);
1746 SDValue N1 = N->getOperand(1);
1747 SDValue CarryIn = N->getOperand(2);
1748 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1749 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1751 // canonicalize constant to RHS
1753 return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(),
1756 // fold (adde x, y, false) -> (addc x, y)
1757 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1758 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1);
1763 // Since it may not be valid to emit a fold to zero for vector initializers
1764 // check if we can before folding.
1765 static SDValue tryFoldToZero(SDLoc DL, const TargetLowering &TLI, EVT VT,
1767 bool LegalOperations, bool LegalTypes) {
1769 return DAG.getConstant(0, VT);
1770 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
1771 return DAG.getConstant(0, VT);
1775 SDValue DAGCombiner::visitSUB(SDNode *N) {
1776 SDValue N0 = N->getOperand(0);
1777 SDValue N1 = N->getOperand(1);
1778 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1779 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1780 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? nullptr :
1781 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1782 EVT VT = N0.getValueType();
1785 if (VT.isVector()) {
1786 SDValue FoldedVOp = SimplifyVBinOp(N);
1787 if (FoldedVOp.getNode()) return FoldedVOp;
1789 // fold (sub x, 0) -> x, vector edition
1790 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1794 // fold (sub x, x) -> 0
1795 // FIXME: Refactor this and xor and other similar operations together.
1797 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
1798 // fold (sub c1, c2) -> c1-c2
1800 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1801 // fold (sub x, c) -> (add x, -c)
1803 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0,
1804 DAG.getConstant(-N1C->getAPIntValue(), VT));
1805 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1806 if (N0C && N0C->isAllOnesValue())
1807 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
1808 // fold A-(A-B) -> B
1809 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1810 return N1.getOperand(1);
1811 // fold (A+B)-A -> B
1812 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1813 return N0.getOperand(1);
1814 // fold (A+B)-B -> A
1815 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1816 return N0.getOperand(0);
1817 // fold C2-(A+C1) -> (C2-C1)-A
1818 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1819 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(),
1821 return DAG.getNode(ISD::SUB, SDLoc(N), VT, NewC,
1824 // fold ((A+(B+or-C))-B) -> A+or-C
1825 if (N0.getOpcode() == ISD::ADD &&
1826 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1827 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1828 N0.getOperand(1).getOperand(0) == N1)
1829 return DAG.getNode(N0.getOperand(1).getOpcode(), SDLoc(N), VT,
1830 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1831 // fold ((A+(C+B))-B) -> A+C
1832 if (N0.getOpcode() == ISD::ADD &&
1833 N0.getOperand(1).getOpcode() == ISD::ADD &&
1834 N0.getOperand(1).getOperand(1) == N1)
1835 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1836 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1837 // fold ((A-(B-C))-C) -> A-B
1838 if (N0.getOpcode() == ISD::SUB &&
1839 N0.getOperand(1).getOpcode() == ISD::SUB &&
1840 N0.getOperand(1).getOperand(1) == N1)
1841 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1842 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1844 // If either operand of a sub is undef, the result is undef
1845 if (N0.getOpcode() == ISD::UNDEF)
1847 if (N1.getOpcode() == ISD::UNDEF)
1850 // If the relocation model supports it, consider symbol offsets.
1851 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1852 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1853 // fold (sub Sym, c) -> Sym-c
1854 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1855 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1857 (uint64_t)N1C->getSExtValue());
1858 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1859 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1860 if (GA->getGlobal() == GB->getGlobal())
1861 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1865 // sub X, (sextinreg Y i1) -> add X, (and Y 1)
1866 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
1867 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
1868 if (TN->getVT() == MVT::i1) {
1870 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
1871 DAG.getConstant(1, VT));
1872 return DAG.getNode(ISD::ADD, DL, VT, N0, ZExt);
1879 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1880 SDValue N0 = N->getOperand(0);
1881 SDValue N1 = N->getOperand(1);
1882 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1883 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1884 EVT VT = N0.getValueType();
1886 // If the flag result is dead, turn this into an SUB.
1887 if (!N->hasAnyUseOfValue(1))
1888 return CombineTo(N, DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1),
1889 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1892 // fold (subc x, x) -> 0 + no borrow
1894 return CombineTo(N, DAG.getConstant(0, VT),
1895 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1898 // fold (subc x, 0) -> x + no borrow
1899 if (N1C && N1C->isNullValue())
1900 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1903 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1904 if (N0C && N0C->isAllOnesValue())
1905 return CombineTo(N, DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0),
1906 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1912 SDValue DAGCombiner::visitSUBE(SDNode *N) {
1913 SDValue N0 = N->getOperand(0);
1914 SDValue N1 = N->getOperand(1);
1915 SDValue CarryIn = N->getOperand(2);
1917 // fold (sube x, y, false) -> (subc x, y)
1918 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1919 return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1);
1924 SDValue DAGCombiner::visitMUL(SDNode *N) {
1925 SDValue N0 = N->getOperand(0);
1926 SDValue N1 = N->getOperand(1);
1927 EVT VT = N0.getValueType();
1929 // fold (mul x, undef) -> 0
1930 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1931 return DAG.getConstant(0, VT);
1933 bool N0IsConst = false;
1934 bool N1IsConst = false;
1935 APInt ConstValue0, ConstValue1;
1937 if (VT.isVector()) {
1938 SDValue FoldedVOp = SimplifyVBinOp(N);
1939 if (FoldedVOp.getNode()) return FoldedVOp;
1941 N0IsConst = isConstantSplatVector(N0.getNode(), ConstValue0);
1942 N1IsConst = isConstantSplatVector(N1.getNode(), ConstValue1);
1944 N0IsConst = dyn_cast<ConstantSDNode>(N0) != nullptr;
1945 ConstValue0 = N0IsConst ? (dyn_cast<ConstantSDNode>(N0))->getAPIntValue()
1947 N1IsConst = dyn_cast<ConstantSDNode>(N1) != nullptr;
1948 ConstValue1 = N1IsConst ? (dyn_cast<ConstantSDNode>(N1))->getAPIntValue()
1952 // fold (mul c1, c2) -> c1*c2
1953 if (N0IsConst && N1IsConst)
1954 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0.getNode(), N1.getNode());
1956 // canonicalize constant to RHS
1957 if (N0IsConst && !N1IsConst)
1958 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0);
1959 // fold (mul x, 0) -> 0
1960 if (N1IsConst && ConstValue1 == 0)
1962 // We require a splat of the entire scalar bit width for non-contiguous
1965 ConstValue1.getBitWidth() == VT.getScalarType().getSizeInBits();
1966 // fold (mul x, 1) -> x
1967 if (N1IsConst && ConstValue1 == 1 && IsFullSplat)
1969 // fold (mul x, -1) -> 0-x
1970 if (N1IsConst && ConstValue1.isAllOnesValue())
1971 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1972 DAG.getConstant(0, VT), N0);
1973 // fold (mul x, (1 << c)) -> x << c
1974 if (N1IsConst && ConstValue1.isPowerOf2() && IsFullSplat)
1975 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
1976 DAG.getConstant(ConstValue1.logBase2(),
1977 getShiftAmountTy(N0.getValueType())));
1978 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
1979 if (N1IsConst && (-ConstValue1).isPowerOf2() && IsFullSplat) {
1980 unsigned Log2Val = (-ConstValue1).logBase2();
1981 // FIXME: If the input is something that is easily negated (e.g. a
1982 // single-use add), we should put the negate there.
1983 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1984 DAG.getConstant(0, VT),
1985 DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
1986 DAG.getConstant(Log2Val,
1987 getShiftAmountTy(N0.getValueType()))));
1991 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
1992 if (N1IsConst && N0.getOpcode() == ISD::SHL &&
1993 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1994 isa<ConstantSDNode>(N0.getOperand(1)))) {
1995 SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT,
1996 N1, N0.getOperand(1));
1997 AddToWorklist(C3.getNode());
1998 return DAG.getNode(ISD::MUL, SDLoc(N), VT,
1999 N0.getOperand(0), C3);
2002 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
2005 SDValue Sh(nullptr,0), Y(nullptr,0);
2006 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
2007 if (N0.getOpcode() == ISD::SHL &&
2008 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2009 isa<ConstantSDNode>(N0.getOperand(1))) &&
2010 N0.getNode()->hasOneUse()) {
2012 } else if (N1.getOpcode() == ISD::SHL &&
2013 isa<ConstantSDNode>(N1.getOperand(1)) &&
2014 N1.getNode()->hasOneUse()) {
2019 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2020 Sh.getOperand(0), Y);
2021 return DAG.getNode(ISD::SHL, SDLoc(N), VT,
2022 Mul, Sh.getOperand(1));
2026 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
2027 if (N1IsConst && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
2028 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2029 isa<ConstantSDNode>(N0.getOperand(1))))
2030 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
2031 DAG.getNode(ISD::MUL, SDLoc(N0), VT,
2032 N0.getOperand(0), N1),
2033 DAG.getNode(ISD::MUL, SDLoc(N1), VT,
2034 N0.getOperand(1), N1));
2037 SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1);
2044 SDValue DAGCombiner::visitSDIV(SDNode *N) {
2045 SDValue N0 = N->getOperand(0);
2046 SDValue N1 = N->getOperand(1);
2047 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2048 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2049 EVT VT = N->getValueType(0);
2052 if (VT.isVector()) {
2053 SDValue FoldedVOp = SimplifyVBinOp(N);
2054 if (FoldedVOp.getNode()) return FoldedVOp;
2057 // fold (sdiv c1, c2) -> c1/c2
2058 if (N0C && N1C && !N1C->isNullValue())
2059 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
2060 // fold (sdiv X, 1) -> X
2061 if (N1C && N1C->getAPIntValue() == 1LL)
2063 // fold (sdiv X, -1) -> 0-X
2064 if (N1C && N1C->isAllOnesValue())
2065 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2066 DAG.getConstant(0, VT), N0);
2067 // If we know the sign bits of both operands are zero, strength reduce to a
2068 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
2069 if (!VT.isVector()) {
2070 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2071 return DAG.getNode(ISD::UDIV, SDLoc(N), N1.getValueType(),
2075 // fold (sdiv X, pow2) -> simple ops after legalize
2076 if (N1C && !N1C->isNullValue() && (N1C->getAPIntValue().isPowerOf2() ||
2077 (-N1C->getAPIntValue()).isPowerOf2())) {
2078 // If dividing by powers of two is cheap, then don't perform the following
2080 if (TLI.isPow2SDivCheap())
2083 // Target-specific implementation of sdiv x, pow2.
2084 SDValue Res = BuildSDIVPow2(N);
2088 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
2090 // Splat the sign bit into the register
2092 DAG.getNode(ISD::SRA, SDLoc(N), VT, N0,
2093 DAG.getConstant(VT.getScalarSizeInBits() - 1,
2094 getShiftAmountTy(N0.getValueType())));
2095 AddToWorklist(SGN.getNode());
2097 // Add (N0 < 0) ? abs2 - 1 : 0;
2099 DAG.getNode(ISD::SRL, SDLoc(N), VT, SGN,
2100 DAG.getConstant(VT.getScalarSizeInBits() - lg2,
2101 getShiftAmountTy(SGN.getValueType())));
2102 SDValue ADD = DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, SRL);
2103 AddToWorklist(SRL.getNode());
2104 AddToWorklist(ADD.getNode()); // Divide by pow2
2105 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), VT, ADD,
2106 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType())));
2108 // If we're dividing by a positive value, we're done. Otherwise, we must
2109 // negate the result.
2110 if (N1C->getAPIntValue().isNonNegative())
2113 AddToWorklist(SRA.getNode());
2114 return DAG.getNode(ISD::SUB, SDLoc(N), VT, DAG.getConstant(0, VT), SRA);
2117 // if integer divide is expensive and we satisfy the requirements, emit an
2118 // alternate sequence.
2119 if (N1C && !TLI.isIntDivCheap()) {
2120 SDValue Op = BuildSDIV(N);
2121 if (Op.getNode()) return Op;
2125 if (N0.getOpcode() == ISD::UNDEF)
2126 return DAG.getConstant(0, VT);
2127 // X / undef -> undef
2128 if (N1.getOpcode() == ISD::UNDEF)
2134 SDValue DAGCombiner::visitUDIV(SDNode *N) {
2135 SDValue N0 = N->getOperand(0);
2136 SDValue N1 = N->getOperand(1);
2137 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2138 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2139 EVT VT = N->getValueType(0);
2142 if (VT.isVector()) {
2143 SDValue FoldedVOp = SimplifyVBinOp(N);
2144 if (FoldedVOp.getNode()) return FoldedVOp;
2147 // fold (udiv c1, c2) -> c1/c2
2148 if (N0C && N1C && !N1C->isNullValue())
2149 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
2150 // fold (udiv x, (1 << c)) -> x >>u c
2151 if (N1C && N1C->getAPIntValue().isPowerOf2())
2152 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0,
2153 DAG.getConstant(N1C->getAPIntValue().logBase2(),
2154 getShiftAmountTy(N0.getValueType())));
2155 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
2156 if (N1.getOpcode() == ISD::SHL) {
2157 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2158 if (SHC->getAPIntValue().isPowerOf2()) {
2159 EVT ADDVT = N1.getOperand(1).getValueType();
2160 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N), ADDVT,
2162 DAG.getConstant(SHC->getAPIntValue()
2165 AddToWorklist(Add.getNode());
2166 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, Add);
2170 // fold (udiv x, c) -> alternate
2171 if (N1C && !TLI.isIntDivCheap()) {
2172 SDValue Op = BuildUDIV(N);
2173 if (Op.getNode()) return Op;
2177 if (N0.getOpcode() == ISD::UNDEF)
2178 return DAG.getConstant(0, VT);
2179 // X / undef -> undef
2180 if (N1.getOpcode() == ISD::UNDEF)
2186 SDValue DAGCombiner::visitSREM(SDNode *N) {
2187 SDValue N0 = N->getOperand(0);
2188 SDValue N1 = N->getOperand(1);
2189 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2190 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2191 EVT VT = N->getValueType(0);
2193 // fold (srem c1, c2) -> c1%c2
2194 if (N0C && N1C && !N1C->isNullValue())
2195 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
2196 // If we know the sign bits of both operands are zero, strength reduce to a
2197 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
2198 if (!VT.isVector()) {
2199 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2200 return DAG.getNode(ISD::UREM, SDLoc(N), VT, N0, N1);
2203 // If X/C can be simplified by the division-by-constant logic, lower
2204 // X%C to the equivalent of X-X/C*C.
2205 if (N1C && !N1C->isNullValue()) {
2206 SDValue Div = DAG.getNode(ISD::SDIV, SDLoc(N), VT, N0, N1);
2207 AddToWorklist(Div.getNode());
2208 SDValue OptimizedDiv = combine(Div.getNode());
2209 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2210 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2212 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2213 AddToWorklist(Mul.getNode());
2219 if (N0.getOpcode() == ISD::UNDEF)
2220 return DAG.getConstant(0, VT);
2221 // X % undef -> undef
2222 if (N1.getOpcode() == ISD::UNDEF)
2228 SDValue DAGCombiner::visitUREM(SDNode *N) {
2229 SDValue N0 = N->getOperand(0);
2230 SDValue N1 = N->getOperand(1);
2231 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2232 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2233 EVT VT = N->getValueType(0);
2235 // fold (urem c1, c2) -> c1%c2
2236 if (N0C && N1C && !N1C->isNullValue())
2237 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
2238 // fold (urem x, pow2) -> (and x, pow2-1)
2239 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
2240 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0,
2241 DAG.getConstant(N1C->getAPIntValue()-1,VT));
2242 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2243 if (N1.getOpcode() == ISD::SHL) {
2244 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2245 if (SHC->getAPIntValue().isPowerOf2()) {
2247 DAG.getNode(ISD::ADD, SDLoc(N), VT, N1,
2248 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
2250 AddToWorklist(Add.getNode());
2251 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, Add);
2256 // If X/C can be simplified by the division-by-constant logic, lower
2257 // X%C to the equivalent of X-X/C*C.
2258 if (N1C && !N1C->isNullValue()) {
2259 SDValue Div = DAG.getNode(ISD::UDIV, SDLoc(N), VT, N0, N1);
2260 AddToWorklist(Div.getNode());
2261 SDValue OptimizedDiv = combine(Div.getNode());
2262 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2263 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2265 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2266 AddToWorklist(Mul.getNode());
2272 if (N0.getOpcode() == ISD::UNDEF)
2273 return DAG.getConstant(0, VT);
2274 // X % undef -> undef
2275 if (N1.getOpcode() == ISD::UNDEF)
2281 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2282 SDValue N0 = N->getOperand(0);
2283 SDValue N1 = N->getOperand(1);
2284 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2285 EVT VT = N->getValueType(0);
2288 // fold (mulhs x, 0) -> 0
2289 if (N1C && N1C->isNullValue())
2291 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2292 if (N1C && N1C->getAPIntValue() == 1)
2293 return DAG.getNode(ISD::SRA, SDLoc(N), N0.getValueType(), N0,
2294 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2295 getShiftAmountTy(N0.getValueType())));
2296 // fold (mulhs x, undef) -> 0
2297 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2298 return DAG.getConstant(0, VT);
2300 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2302 if (VT.isSimple() && !VT.isVector()) {
2303 MVT Simple = VT.getSimpleVT();
2304 unsigned SimpleSize = Simple.getSizeInBits();
2305 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2306 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2307 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2308 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2309 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2310 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2311 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2312 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2319 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2320 SDValue N0 = N->getOperand(0);
2321 SDValue N1 = N->getOperand(1);
2322 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2323 EVT VT = N->getValueType(0);
2326 // fold (mulhu x, 0) -> 0
2327 if (N1C && N1C->isNullValue())
2329 // fold (mulhu x, 1) -> 0
2330 if (N1C && N1C->getAPIntValue() == 1)
2331 return DAG.getConstant(0, N0.getValueType());
2332 // fold (mulhu x, undef) -> 0
2333 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2334 return DAG.getConstant(0, VT);
2336 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2338 if (VT.isSimple() && !VT.isVector()) {
2339 MVT Simple = VT.getSimpleVT();
2340 unsigned SimpleSize = Simple.getSizeInBits();
2341 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2342 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2343 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2344 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2345 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2346 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2347 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2348 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2355 /// Perform optimizations common to nodes that compute two values. LoOp and HiOp
2356 /// give the opcodes for the two computations that are being performed. Return
2357 /// true if a simplification was made.
2358 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2360 // If the high half is not needed, just compute the low half.
2361 bool HiExists = N->hasAnyUseOfValue(1);
2363 (!LegalOperations ||
2364 TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) {
2365 SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2366 return CombineTo(N, Res, Res);
2369 // If the low half is not needed, just compute the high half.
2370 bool LoExists = N->hasAnyUseOfValue(0);
2372 (!LegalOperations ||
2373 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2374 SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2375 return CombineTo(N, Res, Res);
2378 // If both halves are used, return as it is.
2379 if (LoExists && HiExists)
2382 // If the two computed results can be simplified separately, separate them.
2384 SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2385 AddToWorklist(Lo.getNode());
2386 SDValue LoOpt = combine(Lo.getNode());
2387 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2388 (!LegalOperations ||
2389 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2390 return CombineTo(N, LoOpt, LoOpt);
2394 SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2395 AddToWorklist(Hi.getNode());
2396 SDValue HiOpt = combine(Hi.getNode());
2397 if (HiOpt.getNode() && HiOpt != Hi &&
2398 (!LegalOperations ||
2399 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2400 return CombineTo(N, HiOpt, HiOpt);
2406 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2407 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2408 if (Res.getNode()) return Res;
2410 EVT VT = N->getValueType(0);
2413 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2415 if (VT.isSimple() && !VT.isVector()) {
2416 MVT Simple = VT.getSimpleVT();
2417 unsigned SimpleSize = Simple.getSizeInBits();
2418 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2419 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2420 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2421 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2422 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2423 // Compute the high part as N1.
2424 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2425 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2426 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2427 // Compute the low part as N0.
2428 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2429 return CombineTo(N, Lo, Hi);
2436 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2437 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2438 if (Res.getNode()) return Res;
2440 EVT VT = N->getValueType(0);
2443 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2445 if (VT.isSimple() && !VT.isVector()) {
2446 MVT Simple = VT.getSimpleVT();
2447 unsigned SimpleSize = Simple.getSizeInBits();
2448 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2449 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2450 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2451 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2452 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2453 // Compute the high part as N1.
2454 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2455 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2456 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2457 // Compute the low part as N0.
2458 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2459 return CombineTo(N, Lo, Hi);
2466 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2467 // (smulo x, 2) -> (saddo x, x)
2468 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2469 if (C2->getAPIntValue() == 2)
2470 return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(),
2471 N->getOperand(0), N->getOperand(0));
2476 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2477 // (umulo x, 2) -> (uaddo x, x)
2478 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2479 if (C2->getAPIntValue() == 2)
2480 return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(),
2481 N->getOperand(0), N->getOperand(0));
2486 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2487 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2488 if (Res.getNode()) return Res;
2493 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2494 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2495 if (Res.getNode()) return Res;
2500 /// If this is a binary operator with two operands of the same opcode, try to
2502 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2503 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2504 EVT VT = N0.getValueType();
2505 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2507 // Bail early if none of these transforms apply.
2508 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2510 // For each of OP in AND/OR/XOR:
2511 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2512 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2513 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2514 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2516 // do not sink logical op inside of a vector extend, since it may combine
2518 EVT Op0VT = N0.getOperand(0).getValueType();
2519 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2520 N0.getOpcode() == ISD::SIGN_EXTEND ||
2521 // Avoid infinite looping with PromoteIntBinOp.
2522 (N0.getOpcode() == ISD::ANY_EXTEND &&
2523 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2524 (N0.getOpcode() == ISD::TRUNCATE &&
2525 (!TLI.isZExtFree(VT, Op0VT) ||
2526 !TLI.isTruncateFree(Op0VT, VT)) &&
2527 TLI.isTypeLegal(Op0VT))) &&
2529 Op0VT == N1.getOperand(0).getValueType() &&
2530 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2531 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2532 N0.getOperand(0).getValueType(),
2533 N0.getOperand(0), N1.getOperand(0));
2534 AddToWorklist(ORNode.getNode());
2535 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode);
2538 // For each of OP in SHL/SRL/SRA/AND...
2539 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2540 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2541 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2542 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2543 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2544 N0.getOperand(1) == N1.getOperand(1)) {
2545 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2546 N0.getOperand(0).getValueType(),
2547 N0.getOperand(0), N1.getOperand(0));
2548 AddToWorklist(ORNode.getNode());
2549 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
2550 ORNode, N0.getOperand(1));
2553 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2554 // Only perform this optimization after type legalization and before
2555 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2556 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2557 // we don't want to undo this promotion.
2558 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2560 if ((N0.getOpcode() == ISD::BITCAST ||
2561 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) &&
2562 Level == AfterLegalizeTypes) {
2563 SDValue In0 = N0.getOperand(0);
2564 SDValue In1 = N1.getOperand(0);
2565 EVT In0Ty = In0.getValueType();
2566 EVT In1Ty = In1.getValueType();
2568 // If both incoming values are integers, and the original types are the
2570 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2571 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1);
2572 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op);
2573 AddToWorklist(Op.getNode());
2578 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2579 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2580 // If both shuffles use the same mask, and both shuffle within a single
2581 // vector, then it is worthwhile to move the swizzle after the operation.
2582 // The type-legalizer generates this pattern when loading illegal
2583 // vector types from memory. In many cases this allows additional shuffle
2585 // There are other cases where moving the shuffle after the xor/and/or
2586 // is profitable even if shuffles don't perform a swizzle.
2587 // If both shuffles use the same mask, and both shuffles have the same first
2588 // or second operand, then it might still be profitable to move the shuffle
2589 // after the xor/and/or operation.
2590 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) {
2591 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2592 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2594 assert(N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() &&
2595 "Inputs to shuffles are not the same type");
2597 // Check that both shuffles use the same mask. The masks are known to be of
2598 // the same length because the result vector type is the same.
2599 // Check also that shuffles have only one use to avoid introducing extra
2601 if (SVN0->hasOneUse() && SVN1->hasOneUse() &&
2602 SVN0->getMask().equals(SVN1->getMask())) {
2603 SDValue ShOp = N0->getOperand(1);
2605 // Don't try to fold this node if it requires introducing a
2606 // build vector of all zeros that might be illegal at this stage.
2607 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2609 ShOp = DAG.getConstant(0, VT);
2614 // (AND (shuf (A, C), shuf (B, C)) -> shuf (AND (A, B), C)
2615 // (OR (shuf (A, C), shuf (B, C)) -> shuf (OR (A, B), C)
2616 // (XOR (shuf (A, C), shuf (B, C)) -> shuf (XOR (A, B), V_0)
2617 if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) {
2618 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2619 N0->getOperand(0), N1->getOperand(0));
2620 AddToWorklist(NewNode.getNode());
2621 return DAG.getVectorShuffle(VT, SDLoc(N), NewNode, ShOp,
2622 &SVN0->getMask()[0]);
2625 // Don't try to fold this node if it requires introducing a
2626 // build vector of all zeros that might be illegal at this stage.
2627 ShOp = N0->getOperand(0);
2628 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2630 ShOp = DAG.getConstant(0, VT);
2635 // (AND (shuf (C, A), shuf (C, B)) -> shuf (C, AND (A, B))
2636 // (OR (shuf (C, A), shuf (C, B)) -> shuf (C, OR (A, B))
2637 // (XOR (shuf (C, A), shuf (C, B)) -> shuf (V_0, XOR (A, B))
2638 if (N0->getOperand(0) == N1->getOperand(0) && ShOp.getNode()) {
2639 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2640 N0->getOperand(1), N1->getOperand(1));
2641 AddToWorklist(NewNode.getNode());
2642 return DAG.getVectorShuffle(VT, SDLoc(N), ShOp, NewNode,
2643 &SVN0->getMask()[0]);
2651 SDValue DAGCombiner::visitAND(SDNode *N) {
2652 SDValue N0 = N->getOperand(0);
2653 SDValue N1 = N->getOperand(1);
2654 SDValue LL, LR, RL, RR, CC0, CC1;
2655 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2656 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2657 EVT VT = N1.getValueType();
2658 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2661 if (VT.isVector()) {
2662 SDValue FoldedVOp = SimplifyVBinOp(N);
2663 if (FoldedVOp.getNode()) return FoldedVOp;
2665 // fold (and x, 0) -> 0, vector edition
2666 if (ISD::isBuildVectorAllZeros(N0.getNode()))
2667 // do not return N0, because undef node may exist in N0
2668 return DAG.getConstant(
2669 APInt::getNullValue(
2670 N0.getValueType().getScalarType().getSizeInBits()),
2672 if (ISD::isBuildVectorAllZeros(N1.getNode()))
2673 // do not return N1, because undef node may exist in N1
2674 return DAG.getConstant(
2675 APInt::getNullValue(
2676 N1.getValueType().getScalarType().getSizeInBits()),
2679 // fold (and x, -1) -> x, vector edition
2680 if (ISD::isBuildVectorAllOnes(N0.getNode()))
2682 if (ISD::isBuildVectorAllOnes(N1.getNode()))
2686 // fold (and x, undef) -> 0
2687 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2688 return DAG.getConstant(0, VT);
2689 // fold (and c1, c2) -> c1&c2
2691 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2692 // canonicalize constant to RHS
2694 return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0);
2695 // fold (and x, -1) -> x
2696 if (N1C && N1C->isAllOnesValue())
2698 // if (and x, c) is known to be zero, return 0
2699 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2700 APInt::getAllOnesValue(BitWidth)))
2701 return DAG.getConstant(0, VT);
2703 SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1);
2706 // fold (and (or x, C), D) -> D if (C & D) == D
2707 if (N1C && N0.getOpcode() == ISD::OR)
2708 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2709 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2711 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2712 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2713 SDValue N0Op0 = N0.getOperand(0);
2714 APInt Mask = ~N1C->getAPIntValue();
2715 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2716 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2717 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
2718 N0.getValueType(), N0Op0);
2720 // Replace uses of the AND with uses of the Zero extend node.
2723 // We actually want to replace all uses of the any_extend with the
2724 // zero_extend, to avoid duplicating things. This will later cause this
2725 // AND to be folded.
2726 CombineTo(N0.getNode(), Zext);
2727 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2730 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2731 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2732 // already be zero by virtue of the width of the base type of the load.
2734 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2736 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2737 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2738 N0.getOpcode() == ISD::LOAD) {
2739 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2740 N0 : N0.getOperand(0) );
2742 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2743 // This can be a pure constant or a vector splat, in which case we treat the
2744 // vector as a scalar and use the splat value.
2745 APInt Constant = APInt::getNullValue(1);
2746 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2747 Constant = C->getAPIntValue();
2748 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2749 APInt SplatValue, SplatUndef;
2750 unsigned SplatBitSize;
2752 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2753 SplatBitSize, HasAnyUndefs);
2755 // Undef bits can contribute to a possible optimisation if set, so
2757 SplatValue |= SplatUndef;
2759 // The splat value may be something like "0x00FFFFFF", which means 0 for
2760 // the first vector value and FF for the rest, repeating. We need a mask
2761 // that will apply equally to all members of the vector, so AND all the
2762 // lanes of the constant together.
2763 EVT VT = Vector->getValueType(0);
2764 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2766 // If the splat value has been compressed to a bitlength lower
2767 // than the size of the vector lane, we need to re-expand it to
2769 if (BitWidth > SplatBitSize)
2770 for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
2771 SplatBitSize < BitWidth;
2772 SplatBitSize = SplatBitSize * 2)
2773 SplatValue |= SplatValue.shl(SplatBitSize);
2775 Constant = APInt::getAllOnesValue(BitWidth);
2776 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
2777 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
2781 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
2782 // actually legal and isn't going to get expanded, else this is a false
2784 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
2785 Load->getMemoryVT());
2787 // Resize the constant to the same size as the original memory access before
2788 // extension. If it is still the AllOnesValue then this AND is completely
2791 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
2794 switch (Load->getExtensionType()) {
2795 default: B = false; break;
2796 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
2798 case ISD::NON_EXTLOAD: B = true; break;
2801 if (B && Constant.isAllOnesValue()) {
2802 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
2803 // preserve semantics once we get rid of the AND.
2804 SDValue NewLoad(Load, 0);
2805 if (Load->getExtensionType() == ISD::EXTLOAD) {
2806 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
2807 Load->getValueType(0), SDLoc(Load),
2808 Load->getChain(), Load->getBasePtr(),
2809 Load->getOffset(), Load->getMemoryVT(),
2810 Load->getMemOperand());
2811 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
2812 if (Load->getNumValues() == 3) {
2813 // PRE/POST_INC loads have 3 values.
2814 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
2815 NewLoad.getValue(2) };
2816 CombineTo(Load, To, 3, true);
2818 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
2822 // Fold the AND away, taking care not to fold to the old load node if we
2824 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
2826 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2829 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2830 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2831 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2832 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2834 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2835 LL.getValueType().isInteger()) {
2836 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2837 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
2838 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2839 LR.getValueType(), LL, RL);
2840 AddToWorklist(ORNode.getNode());
2841 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
2843 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2844 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2845 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(N0),
2846 LR.getValueType(), LL, RL);
2847 AddToWorklist(ANDNode.getNode());
2848 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
2850 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2851 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2852 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2853 LR.getValueType(), LL, RL);
2854 AddToWorklist(ORNode.getNode());
2855 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
2858 // Simplify (and (setne X, 0), (setne X, -1)) -> (setuge (add X, 1), 2)
2859 if (LL == RL && isa<ConstantSDNode>(LR) && isa<ConstantSDNode>(RR) &&
2860 Op0 == Op1 && LL.getValueType().isInteger() &&
2861 Op0 == ISD::SETNE && ((cast<ConstantSDNode>(LR)->isNullValue() &&
2862 cast<ConstantSDNode>(RR)->isAllOnesValue()) ||
2863 (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
2864 cast<ConstantSDNode>(RR)->isNullValue()))) {
2865 SDValue ADDNode = DAG.getNode(ISD::ADD, SDLoc(N0), LL.getValueType(),
2866 LL, DAG.getConstant(1, LL.getValueType()));
2867 AddToWorklist(ADDNode.getNode());
2868 return DAG.getSetCC(SDLoc(N), VT, ADDNode,
2869 DAG.getConstant(2, LL.getValueType()), ISD::SETUGE);
2871 // canonicalize equivalent to ll == rl
2872 if (LL == RR && LR == RL) {
2873 Op1 = ISD::getSetCCSwappedOperands(Op1);
2876 if (LL == RL && LR == RR) {
2877 bool isInteger = LL.getValueType().isInteger();
2878 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2879 if (Result != ISD::SETCC_INVALID &&
2880 (!LegalOperations ||
2881 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
2882 TLI.isOperationLegal(ISD::SETCC,
2883 getSetCCResultType(N0.getSimpleValueType())))))
2884 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
2889 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
2890 if (N0.getOpcode() == N1.getOpcode()) {
2891 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2892 if (Tmp.getNode()) return Tmp;
2895 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
2896 // fold (and (sra)) -> (and (srl)) when possible.
2897 if (!VT.isVector() &&
2898 SimplifyDemandedBits(SDValue(N, 0)))
2899 return SDValue(N, 0);
2901 // fold (zext_inreg (extload x)) -> (zextload x)
2902 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
2903 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2904 EVT MemVT = LN0->getMemoryVT();
2905 // If we zero all the possible extended bits, then we can turn this into
2906 // a zextload if we are running before legalize or the operation is legal.
2907 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2908 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2909 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2910 ((!LegalOperations && !LN0->isVolatile()) ||
2911 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2912 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2913 LN0->getChain(), LN0->getBasePtr(),
2914 MemVT, LN0->getMemOperand());
2916 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2917 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2920 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
2921 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
2923 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2924 EVT MemVT = LN0->getMemoryVT();
2925 // If we zero all the possible extended bits, then we can turn this into
2926 // a zextload if we are running before legalize or the operation is legal.
2927 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2928 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2929 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2930 ((!LegalOperations && !LN0->isVolatile()) ||
2931 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2932 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2933 LN0->getChain(), LN0->getBasePtr(),
2934 MemVT, LN0->getMemOperand());
2936 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2937 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2941 // fold (and (load x), 255) -> (zextload x, i8)
2942 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2943 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2944 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2945 (N0.getOpcode() == ISD::ANY_EXTEND &&
2946 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2947 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2948 LoadSDNode *LN0 = HasAnyExt
2949 ? cast<LoadSDNode>(N0.getOperand(0))
2950 : cast<LoadSDNode>(N0);
2951 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2952 LN0->isUnindexed() && N0.hasOneUse() && SDValue(LN0, 0).hasOneUse()) {
2953 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2954 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2955 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2956 EVT LoadedVT = LN0->getMemoryVT();
2958 if (ExtVT == LoadedVT &&
2959 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2960 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2963 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
2964 LN0->getChain(), LN0->getBasePtr(), ExtVT,
2965 LN0->getMemOperand());
2967 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
2968 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2971 // Do not change the width of a volatile load.
2972 // Do not generate loads of non-round integer types since these can
2973 // be expensive (and would be wrong if the type is not byte sized).
2974 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
2975 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2976 EVT PtrType = LN0->getOperand(1).getValueType();
2978 unsigned Alignment = LN0->getAlignment();
2979 SDValue NewPtr = LN0->getBasePtr();
2981 // For big endian targets, we need to add an offset to the pointer
2982 // to load the correct bytes. For little endian systems, we merely
2983 // need to read fewer bytes from the same pointer.
2984 if (TLI.isBigEndian()) {
2985 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
2986 unsigned EVTStoreBytes = ExtVT.getStoreSize();
2987 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
2988 NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0), PtrType,
2989 NewPtr, DAG.getConstant(PtrOff, PtrType));
2990 Alignment = MinAlign(Alignment, PtrOff);
2993 AddToWorklist(NewPtr.getNode());
2995 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2997 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
2998 LN0->getChain(), NewPtr,
2999 LN0->getPointerInfo(),
3000 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
3001 LN0->isInvariant(), Alignment, LN0->getAAInfo());
3003 CombineTo(LN0, Load, Load.getValue(1));
3004 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3010 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
3011 VT.getSizeInBits() <= 64) {
3012 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3013 APInt ADDC = ADDI->getAPIntValue();
3014 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
3015 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
3016 // immediate for an add, but it is legal if its top c2 bits are set,
3017 // transform the ADD so the immediate doesn't need to be materialized
3019 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
3020 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
3021 SRLI->getZExtValue());
3022 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
3024 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
3026 DAG.getNode(ISD::ADD, SDLoc(N0), VT,
3027 N0.getOperand(0), DAG.getConstant(ADDC, VT));
3028 CombineTo(N0.getNode(), NewAdd);
3029 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3037 // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const)
3038 if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) {
3039 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
3040 N0.getOperand(1), false);
3041 if (BSwap.getNode())
3048 /// Match (a >> 8) | (a << 8) as (bswap a) >> 16.
3049 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
3050 bool DemandHighBits) {
3051 if (!LegalOperations)
3054 EVT VT = N->getValueType(0);
3055 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
3057 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3060 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
3061 bool LookPassAnd0 = false;
3062 bool LookPassAnd1 = false;
3063 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
3065 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
3067 if (N0.getOpcode() == ISD::AND) {
3068 if (!N0.getNode()->hasOneUse())
3070 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3071 if (!N01C || N01C->getZExtValue() != 0xFF00)
3073 N0 = N0.getOperand(0);
3074 LookPassAnd0 = true;
3077 if (N1.getOpcode() == ISD::AND) {
3078 if (!N1.getNode()->hasOneUse())
3080 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3081 if (!N11C || N11C->getZExtValue() != 0xFF)
3083 N1 = N1.getOperand(0);
3084 LookPassAnd1 = true;
3087 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
3089 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
3091 if (!N0.getNode()->hasOneUse() ||
3092 !N1.getNode()->hasOneUse())
3095 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3096 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3099 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
3102 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
3103 SDValue N00 = N0->getOperand(0);
3104 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
3105 if (!N00.getNode()->hasOneUse())
3107 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
3108 if (!N001C || N001C->getZExtValue() != 0xFF)
3110 N00 = N00.getOperand(0);
3111 LookPassAnd0 = true;
3114 SDValue N10 = N1->getOperand(0);
3115 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
3116 if (!N10.getNode()->hasOneUse())
3118 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
3119 if (!N101C || N101C->getZExtValue() != 0xFF00)
3121 N10 = N10.getOperand(0);
3122 LookPassAnd1 = true;
3128 // Make sure everything beyond the low halfword gets set to zero since the SRL
3129 // 16 will clear the top bits.
3130 unsigned OpSizeInBits = VT.getSizeInBits();
3131 if (DemandHighBits && OpSizeInBits > 16) {
3132 // If the left-shift isn't masked out then the only way this is a bswap is
3133 // if all bits beyond the low 8 are 0. In that case the entire pattern
3134 // reduces to a left shift anyway: leave it for other parts of the combiner.
3138 // However, if the right shift isn't masked out then it might be because
3139 // it's not needed. See if we can spot that too.
3140 if (!LookPassAnd1 &&
3141 !DAG.MaskedValueIsZero(
3142 N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16)))
3146 SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00);
3147 if (OpSizeInBits > 16)
3148 Res = DAG.getNode(ISD::SRL, SDLoc(N), VT, Res,
3149 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
3153 /// Return true if the specified node is an element that makes up a 32-bit
3154 /// packed halfword byteswap.
3155 /// ((x & 0x000000ff) << 8) |
3156 /// ((x & 0x0000ff00) >> 8) |
3157 /// ((x & 0x00ff0000) << 8) |
3158 /// ((x & 0xff000000) >> 8)
3159 static bool isBSwapHWordElement(SDValue N, MutableArrayRef<SDNode *> Parts) {
3160 if (!N.getNode()->hasOneUse())
3163 unsigned Opc = N.getOpcode();
3164 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
3167 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3172 switch (N1C->getZExtValue()) {
3175 case 0xFF: Num = 0; break;
3176 case 0xFF00: Num = 1; break;
3177 case 0xFF0000: Num = 2; break;
3178 case 0xFF000000: Num = 3; break;
3181 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
3182 SDValue N0 = N.getOperand(0);
3183 if (Opc == ISD::AND) {
3184 if (Num == 0 || Num == 2) {
3186 // (x >> 8) & 0xff0000
3187 if (N0.getOpcode() != ISD::SRL)
3189 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3190 if (!C || C->getZExtValue() != 8)
3193 // (x << 8) & 0xff00
3194 // (x << 8) & 0xff000000
3195 if (N0.getOpcode() != ISD::SHL)
3197 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3198 if (!C || C->getZExtValue() != 8)
3201 } else if (Opc == ISD::SHL) {
3203 // (x & 0xff0000) << 8
3204 if (Num != 0 && Num != 2)
3206 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3207 if (!C || C->getZExtValue() != 8)
3209 } else { // Opc == ISD::SRL
3210 // (x & 0xff00) >> 8
3211 // (x & 0xff000000) >> 8
3212 if (Num != 1 && Num != 3)
3214 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3215 if (!C || C->getZExtValue() != 8)
3222 Parts[Num] = N0.getOperand(0).getNode();
3226 /// Match a 32-bit packed halfword bswap. That is
3227 /// ((x & 0x000000ff) << 8) |
3228 /// ((x & 0x0000ff00) >> 8) |
3229 /// ((x & 0x00ff0000) << 8) |
3230 /// ((x & 0xff000000) >> 8)
3231 /// => (rotl (bswap x), 16)
3232 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
3233 if (!LegalOperations)
3236 EVT VT = N->getValueType(0);
3239 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3243 // (or (or (and), (and)), (or (and), (and)))
3244 // (or (or (or (and), (and)), (and)), (and))
3245 if (N0.getOpcode() != ISD::OR)
3247 SDValue N00 = N0.getOperand(0);
3248 SDValue N01 = N0.getOperand(1);
3249 SDNode *Parts[4] = {};
3251 if (N1.getOpcode() == ISD::OR &&
3252 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) {
3253 // (or (or (and), (and)), (or (and), (and)))
3254 SDValue N000 = N00.getOperand(0);
3255 if (!isBSwapHWordElement(N000, Parts))
3258 SDValue N001 = N00.getOperand(1);
3259 if (!isBSwapHWordElement(N001, Parts))
3261 SDValue N010 = N01.getOperand(0);
3262 if (!isBSwapHWordElement(N010, Parts))
3264 SDValue N011 = N01.getOperand(1);
3265 if (!isBSwapHWordElement(N011, Parts))
3268 // (or (or (or (and), (and)), (and)), (and))
3269 if (!isBSwapHWordElement(N1, Parts))
3271 if (!isBSwapHWordElement(N01, Parts))
3273 if (N00.getOpcode() != ISD::OR)
3275 SDValue N000 = N00.getOperand(0);
3276 if (!isBSwapHWordElement(N000, Parts))
3278 SDValue N001 = N00.getOperand(1);
3279 if (!isBSwapHWordElement(N001, Parts))
3283 // Make sure the parts are all coming from the same node.
3284 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
3287 SDValue BSwap = DAG.getNode(ISD::BSWAP, SDLoc(N), VT,
3288 SDValue(Parts[0],0));
3290 // Result of the bswap should be rotated by 16. If it's not legal, then
3291 // do (x << 16) | (x >> 16).
3292 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
3293 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3294 return DAG.getNode(ISD::ROTL, SDLoc(N), VT, BSwap, ShAmt);
3295 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3296 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, BSwap, ShAmt);
3297 return DAG.getNode(ISD::OR, SDLoc(N), VT,
3298 DAG.getNode(ISD::SHL, SDLoc(N), VT, BSwap, ShAmt),
3299 DAG.getNode(ISD::SRL, SDLoc(N), VT, BSwap, ShAmt));
3302 SDValue DAGCombiner::visitOR(SDNode *N) {
3303 SDValue N0 = N->getOperand(0);
3304 SDValue N1 = N->getOperand(1);
3305 SDValue LL, LR, RL, RR, CC0, CC1;
3306 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3307 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3308 EVT VT = N1.getValueType();
3311 if (VT.isVector()) {
3312 SDValue FoldedVOp = SimplifyVBinOp(N);
3313 if (FoldedVOp.getNode()) return FoldedVOp;
3315 // fold (or x, 0) -> x, vector edition
3316 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3318 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3321 // fold (or x, -1) -> -1, vector edition
3322 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3323 // do not return N0, because undef node may exist in N0
3324 return DAG.getConstant(
3325 APInt::getAllOnesValue(
3326 N0.getValueType().getScalarType().getSizeInBits()),
3328 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3329 // do not return N1, because undef node may exist in N1
3330 return DAG.getConstant(
3331 APInt::getAllOnesValue(
3332 N1.getValueType().getScalarType().getSizeInBits()),
3335 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask1)
3336 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf B, A, Mask2)
3337 // Do this only if the resulting shuffle is legal.
3338 if (isa<ShuffleVectorSDNode>(N0) &&
3339 isa<ShuffleVectorSDNode>(N1) &&
3340 // Avoid folding a node with illegal type.
3341 TLI.isTypeLegal(VT) &&
3342 N0->getOperand(1) == N1->getOperand(1) &&
3343 ISD::isBuildVectorAllZeros(N0.getOperand(1).getNode())) {
3344 bool CanFold = true;
3345 unsigned NumElts = VT.getVectorNumElements();
3346 const ShuffleVectorSDNode *SV0 = cast<ShuffleVectorSDNode>(N0);
3347 const ShuffleVectorSDNode *SV1 = cast<ShuffleVectorSDNode>(N1);
3348 // We construct two shuffle masks:
3349 // - Mask1 is a shuffle mask for a shuffle with N0 as the first operand
3350 // and N1 as the second operand.
3351 // - Mask2 is a shuffle mask for a shuffle with N1 as the first operand
3352 // and N0 as the second operand.
3353 // We do this because OR is commutable and therefore there might be
3354 // two ways to fold this node into a shuffle.
3355 SmallVector<int,4> Mask1;
3356 SmallVector<int,4> Mask2;
3358 for (unsigned i = 0; i != NumElts && CanFold; ++i) {
3359 int M0 = SV0->getMaskElt(i);
3360 int M1 = SV1->getMaskElt(i);
3362 // Both shuffle indexes are undef. Propagate Undef.
3363 if (M0 < 0 && M1 < 0) {
3364 Mask1.push_back(M0);
3365 Mask2.push_back(M0);
3369 if (M0 < 0 || M1 < 0 ||
3370 (M0 < (int)NumElts && M1 < (int)NumElts) ||
3371 (M0 >= (int)NumElts && M1 >= (int)NumElts)) {
3376 Mask1.push_back(M0 < (int)NumElts ? M0 : M1 + NumElts);
3377 Mask2.push_back(M1 < (int)NumElts ? M1 : M0 + NumElts);
3381 // Fold this sequence only if the resulting shuffle is 'legal'.
3382 if (TLI.isShuffleMaskLegal(Mask1, VT))
3383 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0),
3384 N1->getOperand(0), &Mask1[0]);
3385 if (TLI.isShuffleMaskLegal(Mask2, VT))
3386 return DAG.getVectorShuffle(VT, SDLoc(N), N1->getOperand(0),
3387 N0->getOperand(0), &Mask2[0]);
3392 // fold (or x, undef) -> -1
3393 if (!LegalOperations &&
3394 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3395 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3396 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
3398 // fold (or c1, c2) -> c1|c2
3400 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
3401 // canonicalize constant to RHS
3403 return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0);
3404 // fold (or x, 0) -> x
3405 if (N1C && N1C->isNullValue())
3407 // fold (or x, -1) -> -1
3408 if (N1C && N1C->isAllOnesValue())
3410 // fold (or x, c) -> c iff (x & ~c) == 0
3411 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3414 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3415 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
3416 if (BSwap.getNode())
3418 BSwap = MatchBSwapHWordLow(N, N0, N1);
3419 if (BSwap.getNode())
3423 SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1);
3426 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3427 // iff (c1 & c2) == 0.
3428 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3429 isa<ConstantSDNode>(N0.getOperand(1))) {
3430 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3431 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) {
3432 SDValue COR = DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1);
3435 return DAG.getNode(ISD::AND, SDLoc(N), VT,
3436 DAG.getNode(ISD::OR, SDLoc(N0), VT,
3437 N0.getOperand(0), N1), COR);
3440 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3441 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3442 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3443 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3445 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3446 LL.getValueType().isInteger()) {
3447 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3448 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3449 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3450 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3451 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(LR),
3452 LR.getValueType(), LL, RL);
3453 AddToWorklist(ORNode.getNode());
3454 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
3456 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3457 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3458 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3459 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3460 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(LR),
3461 LR.getValueType(), LL, RL);
3462 AddToWorklist(ANDNode.getNode());
3463 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
3466 // canonicalize equivalent to ll == rl
3467 if (LL == RR && LR == RL) {
3468 Op1 = ISD::getSetCCSwappedOperands(Op1);
3471 if (LL == RL && LR == RR) {
3472 bool isInteger = LL.getValueType().isInteger();
3473 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3474 if (Result != ISD::SETCC_INVALID &&
3475 (!LegalOperations ||
3476 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
3477 TLI.isOperationLegal(ISD::SETCC,
3478 getSetCCResultType(N0.getValueType())))))
3479 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
3484 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3485 if (N0.getOpcode() == N1.getOpcode()) {
3486 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3487 if (Tmp.getNode()) return Tmp;
3490 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3491 if (N0.getOpcode() == ISD::AND &&
3492 N1.getOpcode() == ISD::AND &&
3493 N0.getOperand(1).getOpcode() == ISD::Constant &&
3494 N1.getOperand(1).getOpcode() == ISD::Constant &&
3495 // Don't increase # computations.
3496 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3497 // We can only do this xform if we know that bits from X that are set in C2
3498 // but not in C1 are already zero. Likewise for Y.
3499 const APInt &LHSMask =
3500 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3501 const APInt &RHSMask =
3502 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3504 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3505 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3506 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3507 N0.getOperand(0), N1.getOperand(0));
3508 return DAG.getNode(ISD::AND, SDLoc(N), VT, X,
3509 DAG.getConstant(LHSMask | RHSMask, VT));
3513 // See if this is some rotate idiom.
3514 if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N)))
3515 return SDValue(Rot, 0);
3517 // Simplify the operands using demanded-bits information.
3518 if (!VT.isVector() &&
3519 SimplifyDemandedBits(SDValue(N, 0)))
3520 return SDValue(N, 0);
3525 /// Match "(X shl/srl V1) & V2" where V2 may not be present.
3526 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3527 if (Op.getOpcode() == ISD::AND) {
3528 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3529 Mask = Op.getOperand(1);
3530 Op = Op.getOperand(0);
3536 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3544 // Return true if we can prove that, whenever Neg and Pos are both in the
3545 // range [0, OpSize), Neg == (Pos == 0 ? 0 : OpSize - Pos). This means that
3546 // for two opposing shifts shift1 and shift2 and a value X with OpBits bits:
3548 // (or (shift1 X, Neg), (shift2 X, Pos))
3550 // reduces to a rotate in direction shift2 by Pos or (equivalently) a rotate
3551 // in direction shift1 by Neg. The range [0, OpSize) means that we only need
3552 // to consider shift amounts with defined behavior.
3553 static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned OpSize) {
3554 // If OpSize is a power of 2 then:
3556 // (a) (Pos == 0 ? 0 : OpSize - Pos) == (OpSize - Pos) & (OpSize - 1)
3557 // (b) Neg == Neg & (OpSize - 1) whenever Neg is in [0, OpSize).
3559 // So if OpSize is a power of 2 and Neg is (and Neg', OpSize-1), we check
3560 // for the stronger condition:
3562 // Neg & (OpSize - 1) == (OpSize - Pos) & (OpSize - 1) [A]
3564 // for all Neg and Pos. Since Neg & (OpSize - 1) == Neg' & (OpSize - 1)
3565 // we can just replace Neg with Neg' for the rest of the function.
3567 // In other cases we check for the even stronger condition:
3569 // Neg == OpSize - Pos [B]
3571 // for all Neg and Pos. Note that the (or ...) then invokes undefined
3572 // behavior if Pos == 0 (and consequently Neg == OpSize).
3574 // We could actually use [A] whenever OpSize is a power of 2, but the
3575 // only extra cases that it would match are those uninteresting ones
3576 // where Neg and Pos are never in range at the same time. E.g. for
3577 // OpSize == 32, using [A] would allow a Neg of the form (sub 64, Pos)
3578 // as well as (sub 32, Pos), but:
3580 // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos))
3582 // always invokes undefined behavior for 32-bit X.
3584 // Below, Mask == OpSize - 1 when using [A] and is all-ones otherwise.
3585 unsigned MaskLoBits = 0;
3586 if (Neg.getOpcode() == ISD::AND &&
3587 isPowerOf2_64(OpSize) &&
3588 Neg.getOperand(1).getOpcode() == ISD::Constant &&
3589 cast<ConstantSDNode>(Neg.getOperand(1))->getAPIntValue() == OpSize - 1) {
3590 Neg = Neg.getOperand(0);
3591 MaskLoBits = Log2_64(OpSize);
3594 // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1.
3595 if (Neg.getOpcode() != ISD::SUB)
3597 ConstantSDNode *NegC = dyn_cast<ConstantSDNode>(Neg.getOperand(0));
3600 SDValue NegOp1 = Neg.getOperand(1);
3602 // On the RHS of [A], if Pos is Pos' & (OpSize - 1), just replace Pos with
3603 // Pos'. The truncation is redundant for the purpose of the equality.
3605 Pos.getOpcode() == ISD::AND &&
3606 Pos.getOperand(1).getOpcode() == ISD::Constant &&
3607 cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() == OpSize - 1)
3608 Pos = Pos.getOperand(0);
3610 // The condition we need is now:
3612 // (NegC - NegOp1) & Mask == (OpSize - Pos) & Mask
3614 // If NegOp1 == Pos then we need:
3616 // OpSize & Mask == NegC & Mask
3618 // (because "x & Mask" is a truncation and distributes through subtraction).
3621 Width = NegC->getAPIntValue();
3622 // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC.
3623 // Then the condition we want to prove becomes:
3625 // (NegC - NegOp1) & Mask == (OpSize - (NegOp1 + PosC)) & Mask
3627 // which, again because "x & Mask" is a truncation, becomes:
3629 // NegC & Mask == (OpSize - PosC) & Mask
3630 // OpSize & Mask == (NegC + PosC) & Mask
3631 else if (Pos.getOpcode() == ISD::ADD &&
3632 Pos.getOperand(0) == NegOp1 &&
3633 Pos.getOperand(1).getOpcode() == ISD::Constant)
3634 Width = (cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() +
3635 NegC->getAPIntValue());
3639 // Now we just need to check that OpSize & Mask == Width & Mask.
3641 // Opsize & Mask is 0 since Mask is Opsize - 1.
3642 return Width.getLoBits(MaskLoBits) == 0;
3643 return Width == OpSize;
3646 // A subroutine of MatchRotate used once we have found an OR of two opposite
3647 // shifts of Shifted. If Neg == <operand size> - Pos then the OR reduces
3648 // to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the
3649 // former being preferred if supported. InnerPos and InnerNeg are Pos and
3650 // Neg with outer conversions stripped away.
3651 SDNode *DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos,
3652 SDValue Neg, SDValue InnerPos,
3653 SDValue InnerNeg, unsigned PosOpcode,
3654 unsigned NegOpcode, SDLoc DL) {
3655 // fold (or (shl x, (*ext y)),
3656 // (srl x, (*ext (sub 32, y)))) ->
3657 // (rotl x, y) or (rotr x, (sub 32, y))
3659 // fold (or (shl x, (*ext (sub 32, y))),
3660 // (srl x, (*ext y))) ->
3661 // (rotr x, y) or (rotl x, (sub 32, y))
3662 EVT VT = Shifted.getValueType();
3663 if (matchRotateSub(InnerPos, InnerNeg, VT.getSizeInBits())) {
3664 bool HasPos = TLI.isOperationLegalOrCustom(PosOpcode, VT);
3665 return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted,
3666 HasPos ? Pos : Neg).getNode();
3672 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3673 // idioms for rotate, and if the target supports rotation instructions, generate
3675 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL) {
3676 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3677 EVT VT = LHS.getValueType();
3678 if (!TLI.isTypeLegal(VT)) return nullptr;
3680 // The target must have at least one rotate flavor.
3681 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3682 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3683 if (!HasROTL && !HasROTR) return nullptr;
3685 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3686 SDValue LHSShift; // The shift.
3687 SDValue LHSMask; // AND value if any.
3688 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3689 return nullptr; // Not part of a rotate.
3691 SDValue RHSShift; // The shift.
3692 SDValue RHSMask; // AND value if any.
3693 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3694 return nullptr; // Not part of a rotate.
3696 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3697 return nullptr; // Not shifting the same value.
3699 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3700 return nullptr; // Shifts must disagree.
3702 // Canonicalize shl to left side in a shl/srl pair.
3703 if (RHSShift.getOpcode() == ISD::SHL) {
3704 std::swap(LHS, RHS);
3705 std::swap(LHSShift, RHSShift);
3706 std::swap(LHSMask , RHSMask );
3709 unsigned OpSizeInBits = VT.getSizeInBits();
3710 SDValue LHSShiftArg = LHSShift.getOperand(0);
3711 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3712 SDValue RHSShiftArg = RHSShift.getOperand(0);
3713 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3715 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3716 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3717 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3718 RHSShiftAmt.getOpcode() == ISD::Constant) {
3719 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3720 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3721 if ((LShVal + RShVal) != OpSizeInBits)
3724 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3725 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
3727 // If there is an AND of either shifted operand, apply it to the result.
3728 if (LHSMask.getNode() || RHSMask.getNode()) {
3729 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3731 if (LHSMask.getNode()) {
3732 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3733 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3735 if (RHSMask.getNode()) {
3736 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3737 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3740 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3743 return Rot.getNode();
3746 // If there is a mask here, and we have a variable shift, we can't be sure
3747 // that we're masking out the right stuff.
3748 if (LHSMask.getNode() || RHSMask.getNode())
3751 // If the shift amount is sign/zext/any-extended just peel it off.
3752 SDValue LExtOp0 = LHSShiftAmt;
3753 SDValue RExtOp0 = RHSShiftAmt;
3754 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3755 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3756 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3757 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3758 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3759 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3760 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3761 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3762 LExtOp0 = LHSShiftAmt.getOperand(0);
3763 RExtOp0 = RHSShiftAmt.getOperand(0);
3766 SDNode *TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt,
3767 LExtOp0, RExtOp0, ISD::ROTL, ISD::ROTR, DL);
3771 SDNode *TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt,
3772 RExtOp0, LExtOp0, ISD::ROTR, ISD::ROTL, DL);
3779 SDValue DAGCombiner::visitXOR(SDNode *N) {
3780 SDValue N0 = N->getOperand(0);
3781 SDValue N1 = N->getOperand(1);
3782 SDValue LHS, RHS, CC;
3783 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3784 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3785 EVT VT = N0.getValueType();
3788 if (VT.isVector()) {
3789 SDValue FoldedVOp = SimplifyVBinOp(N);
3790 if (FoldedVOp.getNode()) return FoldedVOp;
3792 // fold (xor x, 0) -> x, vector edition
3793 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3795 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3799 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3800 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3801 return DAG.getConstant(0, VT);
3802 // fold (xor x, undef) -> undef
3803 if (N0.getOpcode() == ISD::UNDEF)
3805 if (N1.getOpcode() == ISD::UNDEF)
3807 // fold (xor c1, c2) -> c1^c2
3809 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3810 // canonicalize constant to RHS
3812 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
3813 // fold (xor x, 0) -> x
3814 if (N1C && N1C->isNullValue())
3817 SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1);
3821 // fold !(x cc y) -> (x !cc y)
3822 if (N1C && N1C->getAPIntValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3823 bool isInt = LHS.getValueType().isInteger();
3824 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3827 if (!LegalOperations ||
3828 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) {
3829 switch (N0.getOpcode()) {
3831 llvm_unreachable("Unhandled SetCC Equivalent!");
3833 return DAG.getSetCC(SDLoc(N), VT, LHS, RHS, NotCC);
3834 case ISD::SELECT_CC:
3835 return DAG.getSelectCC(SDLoc(N), LHS, RHS, N0.getOperand(2),
3836 N0.getOperand(3), NotCC);
3841 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3842 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3843 N0.getNode()->hasOneUse() &&
3844 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3845 SDValue V = N0.getOperand(0);
3846 V = DAG.getNode(ISD::XOR, SDLoc(N0), V.getValueType(), V,
3847 DAG.getConstant(1, V.getValueType()));
3848 AddToWorklist(V.getNode());
3849 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V);
3852 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3853 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3854 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3855 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3856 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3857 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3858 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3859 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3860 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
3861 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3864 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3865 if (N1C && N1C->isAllOnesValue() &&
3866 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3867 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3868 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3869 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3870 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3871 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3872 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
3873 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3876 // fold (xor (and x, y), y) -> (and (not x), y)
3877 if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3878 N0->getOperand(1) == N1) {
3879 SDValue X = N0->getOperand(0);
3880 SDValue NotX = DAG.getNOT(SDLoc(X), X, VT);
3881 AddToWorklist(NotX.getNode());
3882 return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1);
3884 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3885 if (N1C && N0.getOpcode() == ISD::XOR) {
3886 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3887 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3889 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(1),
3890 DAG.getConstant(N1C->getAPIntValue() ^
3891 N00C->getAPIntValue(), VT));
3893 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(0),
3894 DAG.getConstant(N1C->getAPIntValue() ^
3895 N01C->getAPIntValue(), VT));
3897 // fold (xor x, x) -> 0
3899 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
3901 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
3902 if (N0.getOpcode() == N1.getOpcode()) {
3903 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3904 if (Tmp.getNode()) return Tmp;
3907 // Simplify the expression using non-local knowledge.
3908 if (!VT.isVector() &&
3909 SimplifyDemandedBits(SDValue(N, 0)))
3910 return SDValue(N, 0);
3915 /// Handle transforms common to the three shifts, when the shift amount is a
3917 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, ConstantSDNode *Amt) {
3918 // We can't and shouldn't fold opaque constants.
3919 if (Amt->isOpaque())
3922 SDNode *LHS = N->getOperand(0).getNode();
3923 if (!LHS->hasOneUse()) return SDValue();
3925 // We want to pull some binops through shifts, so that we have (and (shift))
3926 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
3927 // thing happens with address calculations, so it's important to canonicalize
3929 bool HighBitSet = false; // Can we transform this if the high bit is set?
3931 switch (LHS->getOpcode()) {
3932 default: return SDValue();
3935 HighBitSet = false; // We can only transform sra if the high bit is clear.
3938 HighBitSet = true; // We can only transform sra if the high bit is set.
3941 if (N->getOpcode() != ISD::SHL)
3942 return SDValue(); // only shl(add) not sr[al](add).
3943 HighBitSet = false; // We can only transform sra if the high bit is clear.
3947 // We require the RHS of the binop to be a constant and not opaque as well.
3948 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
3949 if (!BinOpCst || BinOpCst->isOpaque()) return SDValue();
3951 // FIXME: disable this unless the input to the binop is a shift by a constant.
3952 // If it is not a shift, it pessimizes some common cases like:
3954 // void foo(int *X, int i) { X[i & 1235] = 1; }
3955 // int bar(int *X, int i) { return X[i & 255]; }
3956 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
3957 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
3958 BinOpLHSVal->getOpcode() != ISD::SRA &&
3959 BinOpLHSVal->getOpcode() != ISD::SRL) ||
3960 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
3963 EVT VT = N->getValueType(0);
3965 // If this is a signed shift right, and the high bit is modified by the
3966 // logical operation, do not perform the transformation. The highBitSet
3967 // boolean indicates the value of the high bit of the constant which would
3968 // cause it to be modified for this operation.
3969 if (N->getOpcode() == ISD::SRA) {
3970 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
3971 if (BinOpRHSSignSet != HighBitSet)
3975 if (!TLI.isDesirableToCommuteWithShift(LHS))
3978 // Fold the constants, shifting the binop RHS by the shift amount.
3979 SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)),
3981 LHS->getOperand(1), N->getOperand(1));
3982 assert(isa<ConstantSDNode>(NewRHS) && "Folding was not successful!");
3984 // Create the new shift.
3985 SDValue NewShift = DAG.getNode(N->getOpcode(),
3986 SDLoc(LHS->getOperand(0)),
3987 VT, LHS->getOperand(0), N->getOperand(1));
3989 // Create the new binop.
3990 return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS);
3993 SDValue DAGCombiner::distributeTruncateThroughAnd(SDNode *N) {
3994 assert(N->getOpcode() == ISD::TRUNCATE);
3995 assert(N->getOperand(0).getOpcode() == ISD::AND);
3997 // (truncate:TruncVT (and N00, N01C)) -> (and (truncate:TruncVT N00), TruncC)
3998 if (N->hasOneUse() && N->getOperand(0).hasOneUse()) {
3999 SDValue N01 = N->getOperand(0).getOperand(1);
4001 if (ConstantSDNode *N01C = isConstOrConstSplat(N01)) {
4002 EVT TruncVT = N->getValueType(0);
4003 SDValue N00 = N->getOperand(0).getOperand(0);
4004 APInt TruncC = N01C->getAPIntValue();
4005 TruncC = TruncC.trunc(TruncVT.getScalarSizeInBits());
4007 return DAG.getNode(ISD::AND, SDLoc(N), TruncVT,
4008 DAG.getNode(ISD::TRUNCATE, SDLoc(N), TruncVT, N00),
4009 DAG.getConstant(TruncC, TruncVT));
4016 SDValue DAGCombiner::visitRotate(SDNode *N) {
4017 // fold (rot* x, (trunc (and y, c))) -> (rot* x, (and (trunc y), (trunc c))).
4018 if (N->getOperand(1).getOpcode() == ISD::TRUNCATE &&
4019 N->getOperand(1).getOperand(0).getOpcode() == ISD::AND) {
4020 SDValue NewOp1 = distributeTruncateThroughAnd(N->getOperand(1).getNode());
4021 if (NewOp1.getNode())
4022 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
4023 N->getOperand(0), NewOp1);
4028 SDValue DAGCombiner::visitSHL(SDNode *N) {
4029 SDValue N0 = N->getOperand(0);
4030 SDValue N1 = N->getOperand(1);
4031 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4032 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4033 EVT VT = N0.getValueType();
4034 unsigned OpSizeInBits = VT.getScalarSizeInBits();
4037 if (VT.isVector()) {
4038 SDValue FoldedVOp = SimplifyVBinOp(N);
4039 if (FoldedVOp.getNode()) return FoldedVOp;
4041 BuildVectorSDNode *N1CV = dyn_cast<BuildVectorSDNode>(N1);
4042 // If setcc produces all-one true value then:
4043 // (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<<N1CV)
4044 if (N1CV && N1CV->isConstant()) {
4045 if (N0.getOpcode() == ISD::AND) {
4046 SDValue N00 = N0->getOperand(0);
4047 SDValue N01 = N0->getOperand(1);
4048 BuildVectorSDNode *N01CV = dyn_cast<BuildVectorSDNode>(N01);
4050 if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC &&
4051 TLI.getBooleanContents(N00.getOperand(0).getValueType()) ==
4052 TargetLowering::ZeroOrNegativeOneBooleanContent) {
4053 SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, VT, N01CV, N1CV);
4055 return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C);
4058 N1C = isConstOrConstSplat(N1);
4063 // fold (shl c1, c2) -> c1<<c2
4065 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
4066 // fold (shl 0, x) -> 0
4067 if (N0C && N0C->isNullValue())
4069 // fold (shl x, c >= size(x)) -> undef
4070 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4071 return DAG.getUNDEF(VT);
4072 // fold (shl x, 0) -> x
4073 if (N1C && N1C->isNullValue())
4075 // fold (shl undef, x) -> 0
4076 if (N0.getOpcode() == ISD::UNDEF)
4077 return DAG.getConstant(0, VT);
4078 // if (shl x, c) is known to be zero, return 0
4079 if (DAG.MaskedValueIsZero(SDValue(N, 0),
4080 APInt::getAllOnesValue(OpSizeInBits)))
4081 return DAG.getConstant(0, VT);
4082 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
4083 if (N1.getOpcode() == ISD::TRUNCATE &&
4084 N1.getOperand(0).getOpcode() == ISD::AND) {
4085 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4086 if (NewOp1.getNode())
4087 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, NewOp1);
4090 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4091 return SDValue(N, 0);
4093 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
4094 if (N1C && N0.getOpcode() == ISD::SHL) {
4095 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4096 uint64_t c1 = N0C1->getZExtValue();
4097 uint64_t c2 = N1C->getZExtValue();
4098 if (c1 + c2 >= OpSizeInBits)
4099 return DAG.getConstant(0, VT);
4100 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4101 DAG.getConstant(c1 + c2, N1.getValueType()));
4105 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
4106 // For this to be valid, the second form must not preserve any of the bits
4107 // that are shifted out by the inner shift in the first form. This means
4108 // the outer shift size must be >= the number of bits added by the ext.
4109 // As a corollary, we don't care what kind of ext it is.
4110 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
4111 N0.getOpcode() == ISD::ANY_EXTEND ||
4112 N0.getOpcode() == ISD::SIGN_EXTEND) &&
4113 N0.getOperand(0).getOpcode() == ISD::SHL) {
4114 SDValue N0Op0 = N0.getOperand(0);
4115 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4116 uint64_t c1 = N0Op0C1->getZExtValue();
4117 uint64_t c2 = N1C->getZExtValue();
4118 EVT InnerShiftVT = N0Op0.getValueType();
4119 uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits();
4120 if (c2 >= OpSizeInBits - InnerShiftSize) {
4121 if (c1 + c2 >= OpSizeInBits)
4122 return DAG.getConstant(0, VT);
4123 return DAG.getNode(ISD::SHL, SDLoc(N0), VT,
4124 DAG.getNode(N0.getOpcode(), SDLoc(N0), VT,
4125 N0Op0->getOperand(0)),
4126 DAG.getConstant(c1 + c2, N1.getValueType()));
4131 // fold (shl (zext (srl x, C)), C) -> (zext (shl (srl x, C), C))
4132 // Only fold this if the inner zext has no other uses to avoid increasing
4133 // the total number of instructions.
4134 if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() &&
4135 N0.getOperand(0).getOpcode() == ISD::SRL) {
4136 SDValue N0Op0 = N0.getOperand(0);
4137 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4138 uint64_t c1 = N0Op0C1->getZExtValue();
4139 if (c1 < VT.getScalarSizeInBits()) {
4140 uint64_t c2 = N1C->getZExtValue();
4142 SDValue NewOp0 = N0.getOperand(0);
4143 EVT CountVT = NewOp0.getOperand(1).getValueType();
4144 SDValue NewSHL = DAG.getNode(ISD::SHL, SDLoc(N), NewOp0.getValueType(),
4145 NewOp0, DAG.getConstant(c2, CountVT));
4146 AddToWorklist(NewSHL.getNode());
4147 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
4153 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
4154 // (and (srl x, (sub c1, c2), MASK)
4155 // Only fold this if the inner shift has no other uses -- if it does, folding
4156 // this will increase the total number of instructions.
4157 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
4158 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4159 uint64_t c1 = N0C1->getZExtValue();
4160 if (c1 < OpSizeInBits) {
4161 uint64_t c2 = N1C->getZExtValue();
4162 APInt Mask = APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - c1);
4165 Mask = Mask.shl(c2 - c1);
4166 Shift = DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4167 DAG.getConstant(c2 - c1, N1.getValueType()));
4169 Mask = Mask.lshr(c1 - c2);
4170 Shift = DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4171 DAG.getConstant(c1 - c2, N1.getValueType()));
4173 return DAG.getNode(ISD::AND, SDLoc(N0), VT, Shift,
4174 DAG.getConstant(Mask, VT));
4178 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
4179 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
4180 unsigned BitSize = VT.getScalarSizeInBits();
4181 SDValue HiBitsMask =
4182 DAG.getConstant(APInt::getHighBitsSet(BitSize,
4183 BitSize - N1C->getZExtValue()), VT);
4184 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4188 // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
4189 // Variant of version done on multiply, except mul by a power of 2 is turned
4192 if (N1C && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
4193 (isa<ConstantSDNode>(N0.getOperand(1)) ||
4194 isConstantSplatVector(N0.getOperand(1).getNode(), Val))) {
4195 SDValue Shl0 = DAG.getNode(ISD::SHL, SDLoc(N0), VT, N0.getOperand(0), N1);
4196 SDValue Shl1 = DAG.getNode(ISD::SHL, SDLoc(N1), VT, N0.getOperand(1), N1);
4197 return DAG.getNode(ISD::ADD, SDLoc(N), VT, Shl0, Shl1);
4201 SDValue NewSHL = visitShiftByConstant(N, N1C);
4202 if (NewSHL.getNode())
4209 SDValue DAGCombiner::visitSRA(SDNode *N) {
4210 SDValue N0 = N->getOperand(0);
4211 SDValue N1 = N->getOperand(1);
4212 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4213 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4214 EVT VT = N0.getValueType();
4215 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4218 if (VT.isVector()) {
4219 SDValue FoldedVOp = SimplifyVBinOp(N);
4220 if (FoldedVOp.getNode()) return FoldedVOp;
4222 N1C = isConstOrConstSplat(N1);
4225 // fold (sra c1, c2) -> (sra c1, c2)
4227 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
4228 // fold (sra 0, x) -> 0
4229 if (N0C && N0C->isNullValue())
4231 // fold (sra -1, x) -> -1
4232 if (N0C && N0C->isAllOnesValue())
4234 // fold (sra x, (setge c, size(x))) -> undef
4235 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4236 return DAG.getUNDEF(VT);
4237 // fold (sra x, 0) -> x
4238 if (N1C && N1C->isNullValue())
4240 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
4242 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
4243 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
4244 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
4246 ExtVT = EVT::getVectorVT(*DAG.getContext(),
4247 ExtVT, VT.getVectorNumElements());
4248 if ((!LegalOperations ||
4249 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
4250 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
4251 N0.getOperand(0), DAG.getValueType(ExtVT));
4254 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
4255 if (N1C && N0.getOpcode() == ISD::SRA) {
4256 if (ConstantSDNode *C1 = isConstOrConstSplat(N0.getOperand(1))) {
4257 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
4258 if (Sum >= OpSizeInBits)
4259 Sum = OpSizeInBits - 1;
4260 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0.getOperand(0),
4261 DAG.getConstant(Sum, N1.getValueType()));
4265 // fold (sra (shl X, m), (sub result_size, n))
4266 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
4267 // result_size - n != m.
4268 // If truncate is free for the target sext(shl) is likely to result in better
4270 if (N0.getOpcode() == ISD::SHL && N1C) {
4271 // Get the two constanst of the shifts, CN0 = m, CN = n.
4272 const ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1));
4274 LLVMContext &Ctx = *DAG.getContext();
4275 // Determine what the truncate's result bitsize and type would be.
4276 EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - N1C->getZExtValue());
4279 TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorNumElements());
4281 // Determine the residual right-shift amount.
4282 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
4284 // If the shift is not a no-op (in which case this should be just a sign
4285 // extend already), the truncated to type is legal, sign_extend is legal
4286 // on that type, and the truncate to that type is both legal and free,
4287 // perform the transform.
4288 if ((ShiftAmt > 0) &&
4289 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
4290 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
4291 TLI.isTruncateFree(VT, TruncVT)) {
4293 SDValue Amt = DAG.getConstant(ShiftAmt,
4294 getShiftAmountTy(N0.getOperand(0).getValueType()));
4295 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0), VT,
4296 N0.getOperand(0), Amt);
4297 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), TruncVT,
4299 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N),
4300 N->getValueType(0), Trunc);
4305 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
4306 if (N1.getOpcode() == ISD::TRUNCATE &&
4307 N1.getOperand(0).getOpcode() == ISD::AND) {
4308 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4309 if (NewOp1.getNode())
4310 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, NewOp1);
4313 // fold (sra (trunc (srl x, c1)), c2) -> (trunc (sra x, c1 + c2))
4314 // if c1 is equal to the number of bits the trunc removes
4315 if (N0.getOpcode() == ISD::TRUNCATE &&
4316 (N0.getOperand(0).getOpcode() == ISD::SRL ||
4317 N0.getOperand(0).getOpcode() == ISD::SRA) &&
4318 N0.getOperand(0).hasOneUse() &&
4319 N0.getOperand(0).getOperand(1).hasOneUse() &&
4321 SDValue N0Op0 = N0.getOperand(0);
4322 if (ConstantSDNode *LargeShift = isConstOrConstSplat(N0Op0.getOperand(1))) {
4323 unsigned LargeShiftVal = LargeShift->getZExtValue();
4324 EVT LargeVT = N0Op0.getValueType();
4326 if (LargeVT.getScalarSizeInBits() - OpSizeInBits == LargeShiftVal) {
4328 DAG.getConstant(LargeShiftVal + N1C->getZExtValue(),
4329 getShiftAmountTy(N0Op0.getOperand(0).getValueType()));
4330 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), LargeVT,
4331 N0Op0.getOperand(0), Amt);
4332 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, SRA);
4337 // Simplify, based on bits shifted out of the LHS.
4338 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4339 return SDValue(N, 0);
4342 // If the sign bit is known to be zero, switch this to a SRL.
4343 if (DAG.SignBitIsZero(N0))
4344 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1);
4347 SDValue NewSRA = visitShiftByConstant(N, N1C);
4348 if (NewSRA.getNode())
4355 SDValue DAGCombiner::visitSRL(SDNode *N) {
4356 SDValue N0 = N->getOperand(0);
4357 SDValue N1 = N->getOperand(1);
4358 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4359 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4360 EVT VT = N0.getValueType();
4361 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4364 if (VT.isVector()) {
4365 SDValue FoldedVOp = SimplifyVBinOp(N);
4366 if (FoldedVOp.getNode()) return FoldedVOp;
4368 N1C = isConstOrConstSplat(N1);
4371 // fold (srl c1, c2) -> c1 >>u c2
4373 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
4374 // fold (srl 0, x) -> 0
4375 if (N0C && N0C->isNullValue())
4377 // fold (srl x, c >= size(x)) -> undef
4378 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4379 return DAG.getUNDEF(VT);
4380 // fold (srl x, 0) -> x
4381 if (N1C && N1C->isNullValue())
4383 // if (srl x, c) is known to be zero, return 0
4384 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
4385 APInt::getAllOnesValue(OpSizeInBits)))
4386 return DAG.getConstant(0, VT);
4388 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
4389 if (N1C && N0.getOpcode() == ISD::SRL) {
4390 if (ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1))) {
4391 uint64_t c1 = N01C->getZExtValue();
4392 uint64_t c2 = N1C->getZExtValue();
4393 if (c1 + c2 >= OpSizeInBits)
4394 return DAG.getConstant(0, VT);
4395 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4396 DAG.getConstant(c1 + c2, N1.getValueType()));
4400 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
4401 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
4402 N0.getOperand(0).getOpcode() == ISD::SRL &&
4403 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
4405 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
4406 uint64_t c2 = N1C->getZExtValue();
4407 EVT InnerShiftVT = N0.getOperand(0).getValueType();
4408 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
4409 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
4410 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
4411 if (c1 + OpSizeInBits == InnerShiftSize) {
4412 if (c1 + c2 >= InnerShiftSize)
4413 return DAG.getConstant(0, VT);
4414 return DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT,
4415 DAG.getNode(ISD::SRL, SDLoc(N0), InnerShiftVT,
4416 N0.getOperand(0)->getOperand(0),
4417 DAG.getConstant(c1 + c2, ShiftCountVT)));
4421 // fold (srl (shl x, c), c) -> (and x, cst2)
4422 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1) {
4423 unsigned BitSize = N0.getScalarValueSizeInBits();
4424 if (BitSize <= 64) {
4425 uint64_t ShAmt = N1C->getZExtValue() + 64 - BitSize;
4426 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4427 DAG.getConstant(~0ULL >> ShAmt, VT));
4431 // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask)
4432 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
4433 // Shifting in all undef bits?
4434 EVT SmallVT = N0.getOperand(0).getValueType();
4435 unsigned BitSize = SmallVT.getScalarSizeInBits();
4436 if (N1C->getZExtValue() >= BitSize)
4437 return DAG.getUNDEF(VT);
4439 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
4440 uint64_t ShiftAmt = N1C->getZExtValue();
4441 SDValue SmallShift = DAG.getNode(ISD::SRL, SDLoc(N0), SmallVT,
4443 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
4444 AddToWorklist(SmallShift.getNode());
4445 APInt Mask = APInt::getAllOnesValue(OpSizeInBits).lshr(ShiftAmt);
4446 return DAG.getNode(ISD::AND, SDLoc(N), VT,
4447 DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, SmallShift),
4448 DAG.getConstant(Mask, VT));
4452 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
4453 // bit, which is unmodified by sra.
4454 if (N1C && N1C->getZExtValue() + 1 == OpSizeInBits) {
4455 if (N0.getOpcode() == ISD::SRA)
4456 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1);
4459 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
4460 if (N1C && N0.getOpcode() == ISD::CTLZ &&
4461 N1C->getAPIntValue() == Log2_32(OpSizeInBits)) {
4462 APInt KnownZero, KnownOne;
4463 DAG.computeKnownBits(N0.getOperand(0), KnownZero, KnownOne);
4465 // If any of the input bits are KnownOne, then the input couldn't be all
4466 // zeros, thus the result of the srl will always be zero.
4467 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
4469 // If all of the bits input the to ctlz node are known to be zero, then
4470 // the result of the ctlz is "32" and the result of the shift is one.
4471 APInt UnknownBits = ~KnownZero;
4472 if (UnknownBits == 0) return DAG.getConstant(1, VT);
4474 // Otherwise, check to see if there is exactly one bit input to the ctlz.
4475 if ((UnknownBits & (UnknownBits - 1)) == 0) {
4476 // Okay, we know that only that the single bit specified by UnknownBits
4477 // could be set on input to the CTLZ node. If this bit is set, the SRL
4478 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
4479 // to an SRL/XOR pair, which is likely to simplify more.
4480 unsigned ShAmt = UnknownBits.countTrailingZeros();
4481 SDValue Op = N0.getOperand(0);
4484 Op = DAG.getNode(ISD::SRL, SDLoc(N0), VT, Op,
4485 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
4486 AddToWorklist(Op.getNode());
4489 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
4490 Op, DAG.getConstant(1, VT));
4494 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
4495 if (N1.getOpcode() == ISD::TRUNCATE &&
4496 N1.getOperand(0).getOpcode() == ISD::AND) {
4497 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4498 if (NewOp1.getNode())
4499 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, NewOp1);
4502 // fold operands of srl based on knowledge that the low bits are not
4504 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4505 return SDValue(N, 0);
4508 SDValue NewSRL = visitShiftByConstant(N, N1C);
4509 if (NewSRL.getNode())
4513 // Attempt to convert a srl of a load into a narrower zero-extending load.
4514 SDValue NarrowLoad = ReduceLoadWidth(N);
4515 if (NarrowLoad.getNode())
4518 // Here is a common situation. We want to optimize:
4521 // %b = and i32 %a, 2
4522 // %c = srl i32 %b, 1
4523 // brcond i32 %c ...
4529 // %c = setcc eq %b, 0
4532 // However when after the source operand of SRL is optimized into AND, the SRL
4533 // itself may not be optimized further. Look for it and add the BRCOND into
4535 if (N->hasOneUse()) {
4536 SDNode *Use = *N->use_begin();
4537 if (Use->getOpcode() == ISD::BRCOND)
4539 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
4540 // Also look pass the truncate.
4541 Use = *Use->use_begin();
4542 if (Use->getOpcode() == ISD::BRCOND)
4550 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
4551 SDValue N0 = N->getOperand(0);
4552 EVT VT = N->getValueType(0);
4554 // fold (ctlz c1) -> c2
4555 if (isa<ConstantSDNode>(N0))
4556 return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0);
4560 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4561 SDValue N0 = N->getOperand(0);
4562 EVT VT = N->getValueType(0);
4564 // fold (ctlz_zero_undef c1) -> c2
4565 if (isa<ConstantSDNode>(N0))
4566 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4570 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4571 SDValue N0 = N->getOperand(0);
4572 EVT VT = N->getValueType(0);
4574 // fold (cttz c1) -> c2
4575 if (isa<ConstantSDNode>(N0))
4576 return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0);
4580 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4581 SDValue N0 = N->getOperand(0);
4582 EVT VT = N->getValueType(0);
4584 // fold (cttz_zero_undef c1) -> c2
4585 if (isa<ConstantSDNode>(N0))
4586 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4590 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4591 SDValue N0 = N->getOperand(0);
4592 EVT VT = N->getValueType(0);
4594 // fold (ctpop c1) -> c2
4595 if (isa<ConstantSDNode>(N0))
4596 return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0);
4600 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4601 SDValue N0 = N->getOperand(0);
4602 SDValue N1 = N->getOperand(1);
4603 SDValue N2 = N->getOperand(2);
4604 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4605 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4606 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
4607 EVT VT = N->getValueType(0);
4608 EVT VT0 = N0.getValueType();
4610 // fold (select C, X, X) -> X
4613 // fold (select true, X, Y) -> X
4614 if (N0C && !N0C->isNullValue())
4616 // fold (select false, X, Y) -> Y
4617 if (N0C && N0C->isNullValue())
4619 // fold (select C, 1, X) -> (or C, X)
4620 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4621 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4622 // fold (select C, 0, 1) -> (xor C, 1)
4623 // We can't do this reliably if integer based booleans have different contents
4624 // to floating point based booleans. This is because we can't tell whether we
4625 // have an integer-based boolean or a floating-point-based boolean unless we
4626 // can find the SETCC that produced it and inspect its operands. This is
4627 // fairly easy if C is the SETCC node, but it can potentially be
4628 // undiscoverable (or not reasonably discoverable). For example, it could be
4629 // in another basic block or it could require searching a complicated
4631 if (VT.isInteger() &&
4632 (VT0 == MVT::i1 || (VT0.isInteger() &&
4633 TLI.getBooleanContents(false, false) ==
4634 TLI.getBooleanContents(false, true) &&
4635 TLI.getBooleanContents(false, false) ==
4636 TargetLowering::ZeroOrOneBooleanContent)) &&
4637 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4640 return DAG.getNode(ISD::XOR, SDLoc(N), VT0,
4641 N0, DAG.getConstant(1, VT0));
4642 XORNode = DAG.getNode(ISD::XOR, SDLoc(N0), VT0,
4643 N0, DAG.getConstant(1, VT0));
4644 AddToWorklist(XORNode.getNode());
4646 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, XORNode);
4647 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, XORNode);
4649 // fold (select C, 0, X) -> (and (not C), X)
4650 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4651 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4652 AddToWorklist(NOTNode.getNode());
4653 return DAG.getNode(ISD::AND, SDLoc(N), VT, NOTNode, N2);
4655 // fold (select C, X, 1) -> (or (not C), X)
4656 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4657 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4658 AddToWorklist(NOTNode.getNode());
4659 return DAG.getNode(ISD::OR, SDLoc(N), VT, NOTNode, N1);
4661 // fold (select C, X, 0) -> (and C, X)
4662 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4663 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4664 // fold (select X, X, Y) -> (or X, Y)
4665 // fold (select X, 1, Y) -> (or X, Y)
4666 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4667 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4668 // fold (select X, Y, X) -> (and X, Y)
4669 // fold (select X, Y, 0) -> (and X, Y)
4670 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4671 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4673 // If we can fold this based on the true/false value, do so.
4674 if (SimplifySelectOps(N, N1, N2))
4675 return SDValue(N, 0); // Don't revisit N.
4677 // fold selects based on a setcc into other things, such as min/max/abs
4678 if (N0.getOpcode() == ISD::SETCC) {
4679 if ((!LegalOperations &&
4680 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT)) ||
4681 TLI.isOperationLegal(ISD::SELECT_CC, VT))
4682 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT,
4683 N0.getOperand(0), N0.getOperand(1),
4684 N1, N2, N0.getOperand(2));
4685 return SimplifySelect(SDLoc(N), N0, N1, N2);
4692 std::pair<SDValue, SDValue> SplitVSETCC(const SDNode *N, SelectionDAG &DAG) {
4695 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
4697 // Split the inputs.
4698 SDValue Lo, Hi, LL, LH, RL, RH;
4699 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
4700 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
4702 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));
4703 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));
4705 return std::make_pair(Lo, Hi);
4708 // This function assumes all the vselect's arguments are CONCAT_VECTOR
4709 // nodes and that the condition is a BV of ConstantSDNodes (or undefs).
4710 static SDValue ConvertSelectToConcatVector(SDNode *N, SelectionDAG &DAG) {
4712 SDValue Cond = N->getOperand(0);
4713 SDValue LHS = N->getOperand(1);
4714 SDValue RHS = N->getOperand(2);
4715 EVT VT = N->getValueType(0);
4716 int NumElems = VT.getVectorNumElements();
4717 assert(LHS.getOpcode() == ISD::CONCAT_VECTORS &&
4718 RHS.getOpcode() == ISD::CONCAT_VECTORS &&
4719 Cond.getOpcode() == ISD::BUILD_VECTOR);
4721 // CONCAT_VECTOR can take an arbitrary number of arguments. We only care about
4722 // binary ones here.
4723 if (LHS->getNumOperands() != 2 || RHS->getNumOperands() != 2)
4726 // We're sure we have an even number of elements due to the
4727 // concat_vectors we have as arguments to vselect.
4728 // Skip BV elements until we find one that's not an UNDEF
4729 // After we find an UNDEF element, keep looping until we get to half the
4730 // length of the BV and see if all the non-undef nodes are the same.
4731 ConstantSDNode *BottomHalf = nullptr;
4732 for (int i = 0; i < NumElems / 2; ++i) {
4733 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4736 if (BottomHalf == nullptr)
4737 BottomHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4738 else if (Cond->getOperand(i).getNode() != BottomHalf)
4742 // Do the same for the second half of the BuildVector
4743 ConstantSDNode *TopHalf = nullptr;
4744 for (int i = NumElems / 2; i < NumElems; ++i) {
4745 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4748 if (TopHalf == nullptr)
4749 TopHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4750 else if (Cond->getOperand(i).getNode() != TopHalf)
4754 assert(TopHalf && BottomHalf &&
4755 "One half of the selector was all UNDEFs and the other was all the "
4756 "same value. This should have been addressed before this function.");
4758 ISD::CONCAT_VECTORS, dl, VT,
4759 BottomHalf->isNullValue() ? RHS->getOperand(0) : LHS->getOperand(0),
4760 TopHalf->isNullValue() ? RHS->getOperand(1) : LHS->getOperand(1));
4763 SDValue DAGCombiner::visitVSELECT(SDNode *N) {
4764 SDValue N0 = N->getOperand(0);
4765 SDValue N1 = N->getOperand(1);
4766 SDValue N2 = N->getOperand(2);
4769 // Canonicalize integer abs.
4770 // vselect (setg[te] X, 0), X, -X ->
4771 // vselect (setgt X, -1), X, -X ->
4772 // vselect (setl[te] X, 0), -X, X ->
4773 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
4774 if (N0.getOpcode() == ISD::SETCC) {
4775 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
4776 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
4778 bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
4780 if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
4781 (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) &&
4782 N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1))
4783 isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode());
4784 else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) &&
4785 N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1))
4786 isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode());
4789 EVT VT = LHS.getValueType();
4790 SDValue Shift = DAG.getNode(
4791 ISD::SRA, DL, VT, LHS,
4792 DAG.getConstant(VT.getScalarType().getSizeInBits() - 1, VT));
4793 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift);
4794 AddToWorklist(Shift.getNode());
4795 AddToWorklist(Add.getNode());
4796 return DAG.getNode(ISD::XOR, DL, VT, Add, Shift);
4800 // If the VSELECT result requires splitting and the mask is provided by a
4801 // SETCC, then split both nodes and its operands before legalization. This
4802 // prevents the type legalizer from unrolling SETCC into scalar comparisons
4803 // and enables future optimizations (e.g. min/max pattern matching on X86).
4804 if (N0.getOpcode() == ISD::SETCC) {
4805 EVT VT = N->getValueType(0);
4807 // Check if any splitting is required.
4808 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
4809 TargetLowering::TypeSplitVector)
4812 SDValue Lo, Hi, CCLo, CCHi, LL, LH, RL, RH;
4813 std::tie(CCLo, CCHi) = SplitVSETCC(N0.getNode(), DAG);
4814 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 1);
4815 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 2);
4817 Lo = DAG.getNode(N->getOpcode(), DL, LL.getValueType(), CCLo, LL, RL);
4818 Hi = DAG.getNode(N->getOpcode(), DL, LH.getValueType(), CCHi, LH, RH);
4820 // Add the new VSELECT nodes to the work list in case they need to be split
4822 AddToWorklist(Lo.getNode());
4823 AddToWorklist(Hi.getNode());
4825 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
4828 // Fold (vselect (build_vector all_ones), N1, N2) -> N1
4829 if (ISD::isBuildVectorAllOnes(N0.getNode()))
4831 // Fold (vselect (build_vector all_zeros), N1, N2) -> N2
4832 if (ISD::isBuildVectorAllZeros(N0.getNode()))
4835 // The ConvertSelectToConcatVector function is assuming both the above
4836 // checks for (vselect (build_vector all{ones,zeros) ...) have been made
4838 if (N1.getOpcode() == ISD::CONCAT_VECTORS &&
4839 N2.getOpcode() == ISD::CONCAT_VECTORS &&
4840 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
4841 SDValue CV = ConvertSelectToConcatVector(N, DAG);
4849 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
4850 SDValue N0 = N->getOperand(0);
4851 SDValue N1 = N->getOperand(1);
4852 SDValue N2 = N->getOperand(2);
4853 SDValue N3 = N->getOperand(3);
4854 SDValue N4 = N->getOperand(4);
4855 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
4857 // fold select_cc lhs, rhs, x, x, cc -> x
4861 // Determine if the condition we're dealing with is constant
4862 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
4863 N0, N1, CC, SDLoc(N), false);
4864 if (SCC.getNode()) {
4865 AddToWorklist(SCC.getNode());
4867 if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) {
4868 if (!SCCC->isNullValue())
4869 return N2; // cond always true -> true val
4871 return N3; // cond always false -> false val
4874 // Fold to a simpler select_cc
4875 if (SCC.getOpcode() == ISD::SETCC)
4876 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(),
4877 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
4881 // If we can fold this based on the true/false value, do so.
4882 if (SimplifySelectOps(N, N2, N3))
4883 return SDValue(N, 0); // Don't revisit N.
4885 // fold select_cc into other things, such as min/max/abs
4886 return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC);
4889 SDValue DAGCombiner::visitSETCC(SDNode *N) {
4890 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
4891 cast<CondCodeSDNode>(N->getOperand(2))->get(),
4895 // tryToFoldExtendOfConstant - Try to fold a sext/zext/aext
4896 // dag node into a ConstantSDNode or a build_vector of constants.
4897 // This function is called by the DAGCombiner when visiting sext/zext/aext
4898 // dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND).
4899 // Vector extends are not folded if operations are legal; this is to
4900 // avoid introducing illegal build_vector dag nodes.
4901 static SDNode *tryToFoldExtendOfConstant(SDNode *N, const TargetLowering &TLI,
4902 SelectionDAG &DAG, bool LegalTypes,
4903 bool LegalOperations) {
4904 unsigned Opcode = N->getOpcode();
4905 SDValue N0 = N->getOperand(0);
4906 EVT VT = N->getValueType(0);
4908 assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND ||
4909 Opcode == ISD::ANY_EXTEND) && "Expected EXTEND dag node in input!");
4911 // fold (sext c1) -> c1
4912 // fold (zext c1) -> c1
4913 // fold (aext c1) -> c1
4914 if (isa<ConstantSDNode>(N0))
4915 return DAG.getNode(Opcode, SDLoc(N), VT, N0).getNode();
4917 // fold (sext (build_vector AllConstants) -> (build_vector AllConstants)
4918 // fold (zext (build_vector AllConstants) -> (build_vector AllConstants)
4919 // fold (aext (build_vector AllConstants) -> (build_vector AllConstants)
4920 EVT SVT = VT.getScalarType();
4921 if (!(VT.isVector() &&
4922 (!LegalTypes || (!LegalOperations && TLI.isTypeLegal(SVT))) &&
4923 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())))
4926 // We can fold this node into a build_vector.
4927 unsigned VTBits = SVT.getSizeInBits();
4928 unsigned EVTBits = N0->getValueType(0).getScalarType().getSizeInBits();
4929 unsigned ShAmt = VTBits - EVTBits;
4930 SmallVector<SDValue, 8> Elts;
4931 unsigned NumElts = N0->getNumOperands();
4934 for (unsigned i=0; i != NumElts; ++i) {
4935 SDValue Op = N0->getOperand(i);
4936 if (Op->getOpcode() == ISD::UNDEF) {
4937 Elts.push_back(DAG.getUNDEF(SVT));
4941 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
4942 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
4943 if (Opcode == ISD::SIGN_EXTEND)
4944 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
4947 Elts.push_back(DAG.getConstant(C.shl(ShAmt).lshr(ShAmt).getZExtValue(),
4951 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Elts).getNode();
4954 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
4955 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
4956 // transformation. Returns true if extension are possible and the above
4957 // mentioned transformation is profitable.
4958 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
4960 SmallVectorImpl<SDNode *> &ExtendNodes,
4961 const TargetLowering &TLI) {
4962 bool HasCopyToRegUses = false;
4963 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
4964 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
4965 UE = N0.getNode()->use_end();
4970 if (UI.getUse().getResNo() != N0.getResNo())
4972 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
4973 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
4974 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
4975 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
4976 // Sign bits will be lost after a zext.
4979 for (unsigned i = 0; i != 2; ++i) {
4980 SDValue UseOp = User->getOperand(i);
4983 if (!isa<ConstantSDNode>(UseOp))
4988 ExtendNodes.push_back(User);
4991 // If truncates aren't free and there are users we can't
4992 // extend, it isn't worthwhile.
4995 // Remember if this value is live-out.
4996 if (User->getOpcode() == ISD::CopyToReg)
4997 HasCopyToRegUses = true;
5000 if (HasCopyToRegUses) {
5001 bool BothLiveOut = false;
5002 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
5004 SDUse &Use = UI.getUse();
5005 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
5011 // Both unextended and extended values are live out. There had better be
5012 // a good reason for the transformation.
5013 return ExtendNodes.size();
5018 void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
5019 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
5020 ISD::NodeType ExtType) {
5021 // Extend SetCC uses if necessary.
5022 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
5023 SDNode *SetCC = SetCCs[i];
5024 SmallVector<SDValue, 4> Ops;
5026 for (unsigned j = 0; j != 2; ++j) {
5027 SDValue SOp = SetCC->getOperand(j);
5029 Ops.push_back(ExtLoad);
5031 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
5034 Ops.push_back(SetCC->getOperand(2));
5035 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), Ops));
5039 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
5040 SDValue N0 = N->getOperand(0);
5041 EVT VT = N->getValueType(0);
5043 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5045 return SDValue(Res, 0);
5047 // fold (sext (sext x)) -> (sext x)
5048 // fold (sext (aext x)) -> (sext x)
5049 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5050 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT,
5053 if (N0.getOpcode() == ISD::TRUNCATE) {
5054 // fold (sext (truncate (load x))) -> (sext (smaller load x))
5055 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
5056 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5057 if (NarrowLoad.getNode()) {
5058 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5059 if (NarrowLoad.getNode() != N0.getNode()) {
5060 CombineTo(N0.getNode(), NarrowLoad);
5061 // CombineTo deleted the truncate, if needed, but not what's under it.
5064 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5067 // See if the value being truncated is already sign extended. If so, just
5068 // eliminate the trunc/sext pair.
5069 SDValue Op = N0.getOperand(0);
5070 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
5071 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
5072 unsigned DestBits = VT.getScalarType().getSizeInBits();
5073 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
5075 if (OpBits == DestBits) {
5076 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
5077 // bits, it is already ready.
5078 if (NumSignBits > DestBits-MidBits)
5080 } else if (OpBits < DestBits) {
5081 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
5082 // bits, just sext from i32.
5083 if (NumSignBits > OpBits-MidBits)
5084 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, Op);
5086 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
5087 // bits, just truncate to i32.
5088 if (NumSignBits > OpBits-MidBits)
5089 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5092 // fold (sext (truncate x)) -> (sextinreg x).
5093 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
5094 N0.getValueType())) {
5095 if (OpBits < DestBits)
5096 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op);
5097 else if (OpBits > DestBits)
5098 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op);
5099 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, Op,
5100 DAG.getValueType(N0.getValueType()));
5104 // fold (sext (load x)) -> (sext (truncate (sextload x)))
5105 // None of the supported targets knows how to perform load and sign extend
5106 // on vectors in one instruction. We only perform this transformation on
5108 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5109 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5110 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5111 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) {
5112 bool DoXform = true;
5113 SmallVector<SDNode*, 4> SetCCs;
5114 if (!N0.hasOneUse())
5115 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
5117 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5118 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5120 LN0->getBasePtr(), N0.getValueType(),
5121 LN0->getMemOperand());
5122 CombineTo(N, ExtLoad);
5123 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5124 N0.getValueType(), ExtLoad);
5125 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5126 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5128 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5132 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
5133 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
5134 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5135 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5136 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5137 EVT MemVT = LN0->getMemoryVT();
5138 if ((!LegalOperations && !LN0->isVolatile()) ||
5139 TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) {
5140 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5142 LN0->getBasePtr(), MemVT,
5143 LN0->getMemOperand());
5144 CombineTo(N, ExtLoad);
5145 CombineTo(N0.getNode(),
5146 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5147 N0.getValueType(), ExtLoad),
5148 ExtLoad.getValue(1));
5149 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5153 // fold (sext (and/or/xor (load x), cst)) ->
5154 // (and/or/xor (sextload x), (sext cst))
5155 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5156 N0.getOpcode() == ISD::XOR) &&
5157 isa<LoadSDNode>(N0.getOperand(0)) &&
5158 N0.getOperand(1).getOpcode() == ISD::Constant &&
5159 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()) &&
5160 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5161 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5162 if (LN0->getExtensionType() != ISD::ZEXTLOAD && LN0->isUnindexed()) {
5163 bool DoXform = true;
5164 SmallVector<SDNode*, 4> SetCCs;
5165 if (!N0.hasOneUse())
5166 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
5169 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN0), VT,
5170 LN0->getChain(), LN0->getBasePtr(),
5172 LN0->getMemOperand());
5173 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5174 Mask = Mask.sext(VT.getSizeInBits());
5175 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5176 ExtLoad, DAG.getConstant(Mask, VT));
5177 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5178 SDLoc(N0.getOperand(0)),
5179 N0.getOperand(0).getValueType(), ExtLoad);
5181 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5182 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5184 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5189 if (N0.getOpcode() == ISD::SETCC) {
5190 EVT N0VT = N0.getOperand(0).getValueType();
5191 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
5192 // Only do this before legalize for now.
5193 if (VT.isVector() && !LegalOperations &&
5194 TLI.getBooleanContents(N0VT) ==
5195 TargetLowering::ZeroOrNegativeOneBooleanContent) {
5196 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
5197 // of the same size as the compared operands. Only optimize sext(setcc())
5198 // if this is the case.
5199 EVT SVT = getSetCCResultType(N0VT);
5201 // We know that the # elements of the results is the same as the
5202 // # elements of the compare (and the # elements of the compare result
5203 // for that matter). Check to see that they are the same size. If so,
5204 // we know that the element size of the sext'd result matches the
5205 // element size of the compare operands.
5206 if (VT.getSizeInBits() == SVT.getSizeInBits())
5207 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5209 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5211 // If the desired elements are smaller or larger than the source
5212 // elements we can use a matching integer vector type and then
5213 // truncate/sign extend
5214 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5215 if (SVT == MatchingVectorType) {
5216 SDValue VsetCC = DAG.getSetCC(SDLoc(N), MatchingVectorType,
5217 N0.getOperand(0), N0.getOperand(1),
5218 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5219 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
5223 // sext(setcc x, y, cc) -> (select (setcc x, y, cc), -1, 0)
5224 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
5226 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
5228 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5229 NegOne, DAG.getConstant(0, VT),
5230 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5231 if (SCC.getNode()) return SCC;
5233 if (!VT.isVector()) {
5234 EVT SetCCVT = getSetCCResultType(N0.getOperand(0).getValueType());
5235 if (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, SetCCVT)) {
5237 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5238 SDValue SetCC = DAG.getSetCC(DL, SetCCVT,
5239 N0.getOperand(0), N0.getOperand(1), CC);
5240 return DAG.getSelect(DL, VT, SetCC,
5241 NegOne, DAG.getConstant(0, VT));
5246 // fold (sext x) -> (zext x) if the sign bit is known zero.
5247 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
5248 DAG.SignBitIsZero(N0))
5249 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0);
5254 // isTruncateOf - If N is a truncate of some other value, return true, record
5255 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
5256 // This function computes KnownZero to avoid a duplicated call to
5257 // computeKnownBits in the caller.
5258 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
5261 if (N->getOpcode() == ISD::TRUNCATE) {
5262 Op = N->getOperand(0);
5263 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5267 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
5268 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
5271 SDValue Op0 = N->getOperand(0);
5272 SDValue Op1 = N->getOperand(1);
5273 assert(Op0.getValueType() == Op1.getValueType());
5275 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
5276 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
5277 if (COp0 && COp0->isNullValue())
5279 else if (COp1 && COp1->isNullValue())
5284 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5286 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
5292 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
5293 SDValue N0 = N->getOperand(0);
5294 EVT VT = N->getValueType(0);
5296 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5298 return SDValue(Res, 0);
5300 // fold (zext (zext x)) -> (zext x)
5301 // fold (zext (aext x)) -> (zext x)
5302 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5303 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT,
5306 // fold (zext (truncate x)) -> (zext x) or
5307 // (zext (truncate x)) -> (truncate x)
5308 // This is valid when the truncated bits of x are already zero.
5309 // FIXME: We should extend this to work for vectors too.
5312 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
5313 APInt TruncatedBits =
5314 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
5315 APInt(Op.getValueSizeInBits(), 0) :
5316 APInt::getBitsSet(Op.getValueSizeInBits(),
5317 N0.getValueSizeInBits(),
5318 std::min(Op.getValueSizeInBits(),
5319 VT.getSizeInBits()));
5320 if (TruncatedBits == (KnownZero & TruncatedBits)) {
5321 if (VT.bitsGT(Op.getValueType()))
5322 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, Op);
5323 if (VT.bitsLT(Op.getValueType()))
5324 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5330 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5331 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
5332 if (N0.getOpcode() == ISD::TRUNCATE) {
5333 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5334 if (NarrowLoad.getNode()) {
5335 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5336 if (NarrowLoad.getNode() != N0.getNode()) {
5337 CombineTo(N0.getNode(), NarrowLoad);
5338 // CombineTo deleted the truncate, if needed, but not what's under it.
5341 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5345 // fold (zext (truncate x)) -> (and x, mask)
5346 if (N0.getOpcode() == ISD::TRUNCATE &&
5347 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
5349 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5350 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
5351 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5352 if (NarrowLoad.getNode()) {
5353 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5354 if (NarrowLoad.getNode() != N0.getNode()) {
5355 CombineTo(N0.getNode(), NarrowLoad);
5356 // CombineTo deleted the truncate, if needed, but not what's under it.
5359 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5362 SDValue Op = N0.getOperand(0);
5363 if (Op.getValueType().bitsLT(VT)) {
5364 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, Op);
5365 AddToWorklist(Op.getNode());
5366 } else if (Op.getValueType().bitsGT(VT)) {
5367 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5368 AddToWorklist(Op.getNode());
5370 return DAG.getZeroExtendInReg(Op, SDLoc(N),
5371 N0.getValueType().getScalarType());
5374 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
5375 // if either of the casts is not free.
5376 if (N0.getOpcode() == ISD::AND &&
5377 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5378 N0.getOperand(1).getOpcode() == ISD::Constant &&
5379 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5380 N0.getValueType()) ||
5381 !TLI.isZExtFree(N0.getValueType(), VT))) {
5382 SDValue X = N0.getOperand(0).getOperand(0);
5383 if (X.getValueType().bitsLT(VT)) {
5384 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(X), VT, X);
5385 } else if (X.getValueType().bitsGT(VT)) {
5386 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
5388 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5389 Mask = Mask.zext(VT.getSizeInBits());
5390 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5391 X, DAG.getConstant(Mask, VT));
5394 // fold (zext (load x)) -> (zext (truncate (zextload x)))
5395 // None of the supported targets knows how to perform load and vector_zext
5396 // on vectors in one instruction. We only perform this transformation on
5398 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5399 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5400 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5401 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
5402 bool DoXform = true;
5403 SmallVector<SDNode*, 4> SetCCs;
5404 if (!N0.hasOneUse())
5405 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
5407 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5408 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5410 LN0->getBasePtr(), N0.getValueType(),
5411 LN0->getMemOperand());
5412 CombineTo(N, ExtLoad);
5413 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5414 N0.getValueType(), ExtLoad);
5415 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5417 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5419 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5423 // fold (zext (and/or/xor (load x), cst)) ->
5424 // (and/or/xor (zextload x), (zext cst))
5425 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5426 N0.getOpcode() == ISD::XOR) &&
5427 isa<LoadSDNode>(N0.getOperand(0)) &&
5428 N0.getOperand(1).getOpcode() == ISD::Constant &&
5429 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()) &&
5430 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5431 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5432 if (LN0->getExtensionType() != ISD::SEXTLOAD && LN0->isUnindexed()) {
5433 bool DoXform = true;
5434 SmallVector<SDNode*, 4> SetCCs;
5435 if (!N0.hasOneUse())
5436 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
5439 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), VT,
5440 LN0->getChain(), LN0->getBasePtr(),
5442 LN0->getMemOperand());
5443 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5444 Mask = Mask.zext(VT.getSizeInBits());
5445 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5446 ExtLoad, DAG.getConstant(Mask, VT));
5447 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5448 SDLoc(N0.getOperand(0)),
5449 N0.getOperand(0).getValueType(), ExtLoad);
5451 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5452 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5454 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5459 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
5460 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
5461 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5462 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5463 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5464 EVT MemVT = LN0->getMemoryVT();
5465 if ((!LegalOperations && !LN0->isVolatile()) ||
5466 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) {
5467 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5469 LN0->getBasePtr(), MemVT,
5470 LN0->getMemOperand());
5471 CombineTo(N, ExtLoad);
5472 CombineTo(N0.getNode(),
5473 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(),
5475 ExtLoad.getValue(1));
5476 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5480 if (N0.getOpcode() == ISD::SETCC) {
5481 if (!LegalOperations && VT.isVector() &&
5482 N0.getValueType().getVectorElementType() == MVT::i1) {
5483 EVT N0VT = N0.getOperand(0).getValueType();
5484 if (getSetCCResultType(N0VT) == N0.getValueType())
5487 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
5488 // Only do this before legalize for now.
5489 EVT EltVT = VT.getVectorElementType();
5490 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
5491 DAG.getConstant(1, EltVT));
5492 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5493 // We know that the # elements of the results is the same as the
5494 // # elements of the compare (and the # elements of the compare result
5495 // for that matter). Check to see that they are the same size. If so,
5496 // we know that the element size of the sext'd result matches the
5497 // element size of the compare operands.
5498 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5499 DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5501 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
5502 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT,
5505 // If the desired elements are smaller or larger than the source
5506 // elements we can use a matching integer vector type and then
5507 // truncate/sign extend
5508 EVT MatchingElementType =
5509 EVT::getIntegerVT(*DAG.getContext(),
5510 N0VT.getScalarType().getSizeInBits());
5511 EVT MatchingVectorType =
5512 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
5513 N0VT.getVectorNumElements());
5515 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5517 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5518 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5519 DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT),
5520 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, OneOps));
5523 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5525 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5526 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5527 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5528 if (SCC.getNode()) return SCC;
5531 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
5532 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
5533 isa<ConstantSDNode>(N0.getOperand(1)) &&
5534 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
5536 SDValue ShAmt = N0.getOperand(1);
5537 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
5538 if (N0.getOpcode() == ISD::SHL) {
5539 SDValue InnerZExt = N0.getOperand(0);
5540 // If the original shl may be shifting out bits, do not perform this
5542 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
5543 InnerZExt.getOperand(0).getValueType().getSizeInBits();
5544 if (ShAmtVal > KnownZeroBits)
5550 // Ensure that the shift amount is wide enough for the shifted value.
5551 if (VT.getSizeInBits() >= 256)
5552 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
5554 return DAG.getNode(N0.getOpcode(), DL, VT,
5555 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
5562 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
5563 SDValue N0 = N->getOperand(0);
5564 EVT VT = N->getValueType(0);
5566 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5568 return SDValue(Res, 0);
5570 // fold (aext (aext x)) -> (aext x)
5571 // fold (aext (zext x)) -> (zext x)
5572 // fold (aext (sext x)) -> (sext x)
5573 if (N0.getOpcode() == ISD::ANY_EXTEND ||
5574 N0.getOpcode() == ISD::ZERO_EXTEND ||
5575 N0.getOpcode() == ISD::SIGN_EXTEND)
5576 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0));
5578 // fold (aext (truncate (load x))) -> (aext (smaller load x))
5579 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
5580 if (N0.getOpcode() == ISD::TRUNCATE) {
5581 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5582 if (NarrowLoad.getNode()) {
5583 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5584 if (NarrowLoad.getNode() != N0.getNode()) {
5585 CombineTo(N0.getNode(), NarrowLoad);
5586 // CombineTo deleted the truncate, if needed, but not what's under it.
5589 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5593 // fold (aext (truncate x))
5594 if (N0.getOpcode() == ISD::TRUNCATE) {
5595 SDValue TruncOp = N0.getOperand(0);
5596 if (TruncOp.getValueType() == VT)
5597 return TruncOp; // x iff x size == zext size.
5598 if (TruncOp.getValueType().bitsGT(VT))
5599 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, TruncOp);
5600 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, TruncOp);
5603 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
5604 // if the trunc is not free.
5605 if (N0.getOpcode() == ISD::AND &&
5606 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5607 N0.getOperand(1).getOpcode() == ISD::Constant &&
5608 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5609 N0.getValueType())) {
5610 SDValue X = N0.getOperand(0).getOperand(0);
5611 if (X.getValueType().bitsLT(VT)) {
5612 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, X);
5613 } else if (X.getValueType().bitsGT(VT)) {
5614 X = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, X);
5616 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5617 Mask = Mask.zext(VT.getSizeInBits());
5618 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5619 X, DAG.getConstant(Mask, VT));
5622 // fold (aext (load x)) -> (aext (truncate (extload x)))
5623 // None of the supported targets knows how to perform load and any_ext
5624 // on vectors in one instruction. We only perform this transformation on
5626 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5627 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5628 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType())) {
5629 bool DoXform = true;
5630 SmallVector<SDNode*, 4> SetCCs;
5631 if (!N0.hasOneUse())
5632 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
5634 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5635 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
5637 LN0->getBasePtr(), N0.getValueType(),
5638 LN0->getMemOperand());
5639 CombineTo(N, ExtLoad);
5640 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5641 N0.getValueType(), ExtLoad);
5642 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5643 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5645 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5649 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
5650 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
5651 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
5652 if (N0.getOpcode() == ISD::LOAD &&
5653 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5655 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5656 ISD::LoadExtType ExtType = LN0->getExtensionType();
5657 EVT MemVT = LN0->getMemoryVT();
5658 if (!LegalOperations || TLI.isLoadExtLegal(ExtType, MemVT)) {
5659 SDValue ExtLoad = DAG.getExtLoad(ExtType, SDLoc(N),
5660 VT, LN0->getChain(), LN0->getBasePtr(),
5661 MemVT, LN0->getMemOperand());
5662 CombineTo(N, ExtLoad);
5663 CombineTo(N0.getNode(),
5664 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5665 N0.getValueType(), ExtLoad),
5666 ExtLoad.getValue(1));
5667 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5671 if (N0.getOpcode() == ISD::SETCC) {
5673 // aext(setcc) -> vsetcc
5674 // aext(setcc) -> truncate(vsetcc)
5675 // aext(setcc) -> aext(vsetcc)
5676 // Only do this before legalize for now.
5677 if (VT.isVector() && !LegalOperations) {
5678 EVT N0VT = N0.getOperand(0).getValueType();
5679 // We know that the # elements of the results is the same as the
5680 // # elements of the compare (and the # elements of the compare result
5681 // for that matter). Check to see that they are the same size. If so,
5682 // we know that the element size of the sext'd result matches the
5683 // element size of the compare operands.
5684 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5685 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5687 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5688 // If the desired elements are smaller or larger than the source
5689 // elements we can use a matching integer vector type and then
5690 // truncate/any extend
5692 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5694 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5696 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5697 return DAG.getAnyExtOrTrunc(VsetCC, SDLoc(N), VT);
5701 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5703 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5704 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5705 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5713 /// See if the specified operand can be simplified with the knowledge that only
5714 /// the bits specified by Mask are used. If so, return the simpler operand,
5715 /// otherwise return a null SDValue.
5716 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
5717 switch (V.getOpcode()) {
5719 case ISD::Constant: {
5720 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
5721 assert(CV && "Const value should be ConstSDNode.");
5722 const APInt &CVal = CV->getAPIntValue();
5723 APInt NewVal = CVal & Mask;
5725 return DAG.getConstant(NewVal, V.getValueType());
5730 // If the LHS or RHS don't contribute bits to the or, drop them.
5731 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
5732 return V.getOperand(1);
5733 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
5734 return V.getOperand(0);
5737 // Only look at single-use SRLs.
5738 if (!V.getNode()->hasOneUse())
5740 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
5741 // See if we can recursively simplify the LHS.
5742 unsigned Amt = RHSC->getZExtValue();
5744 // Watch out for shift count overflow though.
5745 if (Amt >= Mask.getBitWidth()) break;
5746 APInt NewMask = Mask << Amt;
5747 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
5748 if (SimplifyLHS.getNode())
5749 return DAG.getNode(ISD::SRL, SDLoc(V), V.getValueType(),
5750 SimplifyLHS, V.getOperand(1));
5756 /// If the result of a wider load is shifted to right of N bits and then
5757 /// truncated to a narrower type and where N is a multiple of number of bits of
5758 /// the narrower type, transform it to a narrower load from address + N / num of
5759 /// bits of new type. If the result is to be extended, also fold the extension
5760 /// to form a extending load.
5761 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
5762 unsigned Opc = N->getOpcode();
5764 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
5765 SDValue N0 = N->getOperand(0);
5766 EVT VT = N->getValueType(0);
5769 // This transformation isn't valid for vector loads.
5773 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
5775 if (Opc == ISD::SIGN_EXTEND_INREG) {
5776 ExtType = ISD::SEXTLOAD;
5777 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
5778 } else if (Opc == ISD::SRL) {
5779 // Another special-case: SRL is basically zero-extending a narrower value.
5780 ExtType = ISD::ZEXTLOAD;
5782 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
5783 if (!N01) return SDValue();
5784 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
5785 VT.getSizeInBits() - N01->getZExtValue());
5787 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, ExtVT))
5790 unsigned EVTBits = ExtVT.getSizeInBits();
5792 // Do not generate loads of non-round integer types since these can
5793 // be expensive (and would be wrong if the type is not byte sized).
5794 if (!ExtVT.isRound())
5798 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
5799 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5800 ShAmt = N01->getZExtValue();
5801 // Is the shift amount a multiple of size of VT?
5802 if ((ShAmt & (EVTBits-1)) == 0) {
5803 N0 = N0.getOperand(0);
5804 // Is the load width a multiple of size of VT?
5805 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
5809 // At this point, we must have a load or else we can't do the transform.
5810 if (!isa<LoadSDNode>(N0)) return SDValue();
5812 // Because a SRL must be assumed to *need* to zero-extend the high bits
5813 // (as opposed to anyext the high bits), we can't combine the zextload
5814 // lowering of SRL and an sextload.
5815 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
5818 // If the shift amount is larger than the input type then we're not
5819 // accessing any of the loaded bytes. If the load was a zextload/extload
5820 // then the result of the shift+trunc is zero/undef (handled elsewhere).
5821 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
5826 // If the load is shifted left (and the result isn't shifted back right),
5827 // we can fold the truncate through the shift.
5828 unsigned ShLeftAmt = 0;
5829 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
5830 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
5831 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5832 ShLeftAmt = N01->getZExtValue();
5833 N0 = N0.getOperand(0);
5837 // If we haven't found a load, we can't narrow it. Don't transform one with
5838 // multiple uses, this would require adding a new load.
5839 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse())
5842 // Don't change the width of a volatile load.
5843 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5844 if (LN0->isVolatile())
5847 // Verify that we are actually reducing a load width here.
5848 if (LN0->getMemoryVT().getSizeInBits() < EVTBits)
5851 // For the transform to be legal, the load must produce only two values
5852 // (the value loaded and the chain). Don't transform a pre-increment
5853 // load, for example, which produces an extra value. Otherwise the
5854 // transformation is not equivalent, and the downstream logic to replace
5855 // uses gets things wrong.
5856 if (LN0->getNumValues() > 2)
5859 // If the load that we're shrinking is an extload and we're not just
5860 // discarding the extension we can't simply shrink the load. Bail.
5861 // TODO: It would be possible to merge the extensions in some cases.
5862 if (LN0->getExtensionType() != ISD::NON_EXTLOAD &&
5863 LN0->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits() + ShAmt)
5866 EVT PtrType = N0.getOperand(1).getValueType();
5868 if (PtrType == MVT::Untyped || PtrType.isExtended())
5869 // It's not possible to generate a constant of extended or untyped type.
5872 // For big endian targets, we need to adjust the offset to the pointer to
5873 // load the correct bytes.
5874 if (TLI.isBigEndian()) {
5875 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
5876 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
5877 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
5880 uint64_t PtrOff = ShAmt / 8;
5881 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
5882 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0),
5883 PtrType, LN0->getBasePtr(),
5884 DAG.getConstant(PtrOff, PtrType));
5885 AddToWorklist(NewPtr.getNode());
5888 if (ExtType == ISD::NON_EXTLOAD)
5889 Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr,
5890 LN0->getPointerInfo().getWithOffset(PtrOff),
5891 LN0->isVolatile(), LN0->isNonTemporal(),
5892 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
5894 Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(),NewPtr,
5895 LN0->getPointerInfo().getWithOffset(PtrOff),
5896 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
5897 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
5899 // Replace the old load's chain with the new load's chain.
5900 WorklistRemover DeadNodes(*this);
5901 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
5903 // Shift the result left, if we've swallowed a left shift.
5904 SDValue Result = Load;
5905 if (ShLeftAmt != 0) {
5906 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
5907 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
5909 // If the shift amount is as large as the result size (but, presumably,
5910 // no larger than the source) then the useful bits of the result are
5911 // zero; we can't simply return the shortened shift, because the result
5912 // of that operation is undefined.
5913 if (ShLeftAmt >= VT.getSizeInBits())
5914 Result = DAG.getConstant(0, VT);
5916 Result = DAG.getNode(ISD::SHL, SDLoc(N0), VT,
5917 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
5920 // Return the new loaded value.
5924 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
5925 SDValue N0 = N->getOperand(0);
5926 SDValue N1 = N->getOperand(1);
5927 EVT VT = N->getValueType(0);
5928 EVT EVT = cast<VTSDNode>(N1)->getVT();
5929 unsigned VTBits = VT.getScalarType().getSizeInBits();
5930 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
5932 // fold (sext_in_reg c1) -> c1
5933 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
5934 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1);
5936 // If the input is already sign extended, just drop the extension.
5937 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
5940 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
5941 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
5942 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT()))
5943 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
5944 N0.getOperand(0), N1);
5946 // fold (sext_in_reg (sext x)) -> (sext x)
5947 // fold (sext_in_reg (aext x)) -> (sext x)
5948 // if x is small enough.
5949 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
5950 SDValue N00 = N0.getOperand(0);
5951 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
5952 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
5953 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
5956 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
5957 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
5958 return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT);
5960 // fold operands of sext_in_reg based on knowledge that the top bits are not
5962 if (SimplifyDemandedBits(SDValue(N, 0)))
5963 return SDValue(N, 0);
5965 // fold (sext_in_reg (load x)) -> (smaller sextload x)
5966 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
5967 SDValue NarrowLoad = ReduceLoadWidth(N);
5968 if (NarrowLoad.getNode())
5971 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
5972 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
5973 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
5974 if (N0.getOpcode() == ISD::SRL) {
5975 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
5976 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
5977 // We can turn this into an SRA iff the input to the SRL is already sign
5979 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
5980 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
5981 return DAG.getNode(ISD::SRA, SDLoc(N), VT,
5982 N0.getOperand(0), N0.getOperand(1));
5986 // fold (sext_inreg (extload x)) -> (sextload x)
5987 if (ISD::isEXTLoad(N0.getNode()) &&
5988 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5989 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5990 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5991 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5992 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5993 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5995 LN0->getBasePtr(), EVT,
5996 LN0->getMemOperand());
5997 CombineTo(N, ExtLoad);
5998 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5999 AddToWorklist(ExtLoad.getNode());
6000 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6002 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
6003 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
6005 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
6006 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6007 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
6008 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6009 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
6011 LN0->getBasePtr(), EVT,
6012 LN0->getMemOperand());
6013 CombineTo(N, ExtLoad);
6014 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
6015 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6018 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
6019 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
6020 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
6021 N0.getOperand(1), false);
6022 if (BSwap.getNode())
6023 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
6027 // Fold a sext_inreg of a build_vector of ConstantSDNodes or undefs
6028 // into a build_vector.
6029 if (ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
6030 SmallVector<SDValue, 8> Elts;
6031 unsigned NumElts = N0->getNumOperands();
6032 unsigned ShAmt = VTBits - EVTBits;
6034 for (unsigned i = 0; i != NumElts; ++i) {
6035 SDValue Op = N0->getOperand(i);
6036 if (Op->getOpcode() == ISD::UNDEF) {
6041 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
6042 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
6043 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
6044 Op.getValueType()));
6047 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Elts);
6053 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
6054 SDValue N0 = N->getOperand(0);
6055 EVT VT = N->getValueType(0);
6056 bool isLE = TLI.isLittleEndian();
6059 if (N0.getValueType() == N->getValueType(0))
6061 // fold (truncate c1) -> c1
6062 if (isa<ConstantSDNode>(N0))
6063 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0);
6064 // fold (truncate (truncate x)) -> (truncate x)
6065 if (N0.getOpcode() == ISD::TRUNCATE)
6066 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6067 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
6068 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
6069 N0.getOpcode() == ISD::SIGN_EXTEND ||
6070 N0.getOpcode() == ISD::ANY_EXTEND) {
6071 if (N0.getOperand(0).getValueType().bitsLT(VT))
6072 // if the source is smaller than the dest, we still need an extend
6073 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
6075 if (N0.getOperand(0).getValueType().bitsGT(VT))
6076 // if the source is larger than the dest, than we just need the truncate
6077 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6078 // if the source and dest are the same type, we can drop both the extend
6079 // and the truncate.
6080 return N0.getOperand(0);
6083 // Fold extract-and-trunc into a narrow extract. For example:
6084 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
6085 // i32 y = TRUNCATE(i64 x)
6087 // v16i8 b = BITCAST (v2i64 val)
6088 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
6090 // Note: We only run this optimization after type legalization (which often
6091 // creates this pattern) and before operation legalization after which
6092 // we need to be more careful about the vector instructions that we generate.
6093 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6094 LegalTypes && !LegalOperations && N0->hasOneUse() && VT != MVT::i1) {
6096 EVT VecTy = N0.getOperand(0).getValueType();
6097 EVT ExTy = N0.getValueType();
6098 EVT TrTy = N->getValueType(0);
6100 unsigned NumElem = VecTy.getVectorNumElements();
6101 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
6103 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
6104 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
6106 SDValue EltNo = N0->getOperand(1);
6107 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
6108 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
6109 EVT IndexTy = TLI.getVectorIdxTy();
6110 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
6112 SDValue V = DAG.getNode(ISD::BITCAST, SDLoc(N),
6113 NVT, N0.getOperand(0));
6115 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
6117 DAG.getConstant(Index, IndexTy));
6121 // trunc (select c, a, b) -> select c, (trunc a), (trunc b)
6122 if (N0.getOpcode() == ISD::SELECT) {
6123 EVT SrcVT = N0.getValueType();
6124 if ((!LegalOperations || TLI.isOperationLegal(ISD::SELECT, SrcVT)) &&
6125 TLI.isTruncateFree(SrcVT, VT)) {
6127 SDValue Cond = N0.getOperand(0);
6128 SDValue TruncOp0 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(1));
6129 SDValue TruncOp1 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(2));
6130 return DAG.getNode(ISD::SELECT, SDLoc(N), VT, Cond, TruncOp0, TruncOp1);
6134 // Fold a series of buildvector, bitcast, and truncate if possible.
6136 // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to
6137 // (2xi32 (buildvector x, y)).
6138 if (Level == AfterLegalizeVectorOps && VT.isVector() &&
6139 N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
6140 N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
6141 N0.getOperand(0).hasOneUse()) {
6143 SDValue BuildVect = N0.getOperand(0);
6144 EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType();
6145 EVT TruncVecEltTy = VT.getVectorElementType();
6147 // Check that the element types match.
6148 if (BuildVectEltTy == TruncVecEltTy) {
6149 // Now we only need to compute the offset of the truncated elements.
6150 unsigned BuildVecNumElts = BuildVect.getNumOperands();
6151 unsigned TruncVecNumElts = VT.getVectorNumElements();
6152 unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts;
6154 assert((BuildVecNumElts % TruncVecNumElts) == 0 &&
6155 "Invalid number of elements");
6157 SmallVector<SDValue, 8> Opnds;
6158 for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset)
6159 Opnds.push_back(BuildVect.getOperand(i));
6161 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
6165 // See if we can simplify the input to this truncate through knowledge that
6166 // only the low bits are being used.
6167 // For example "trunc (or (shl x, 8), y)" // -> trunc y
6168 // Currently we only perform this optimization on scalars because vectors
6169 // may have different active low bits.
6170 if (!VT.isVector()) {
6172 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
6173 VT.getSizeInBits()));
6174 if (Shorter.getNode())
6175 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter);
6177 // fold (truncate (load x)) -> (smaller load x)
6178 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
6179 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
6180 SDValue Reduced = ReduceLoadWidth(N);
6181 if (Reduced.getNode())
6183 // Handle the case where the load remains an extending load even
6184 // after truncation.
6185 if (N0.hasOneUse() && ISD::isUNINDEXEDLoad(N0.getNode())) {
6186 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6187 if (!LN0->isVolatile() &&
6188 LN0->getMemoryVT().getStoreSizeInBits() < VT.getSizeInBits()) {
6189 SDValue NewLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(LN0),
6190 VT, LN0->getChain(), LN0->getBasePtr(),
6192 LN0->getMemOperand());
6193 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLoad.getValue(1));
6198 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
6199 // where ... are all 'undef'.
6200 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
6201 SmallVector<EVT, 8> VTs;
6204 unsigned NumDefs = 0;
6206 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
6207 SDValue X = N0.getOperand(i);
6208 if (X.getOpcode() != ISD::UNDEF) {
6213 // Stop if more than one members are non-undef.
6216 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
6217 VT.getVectorElementType(),
6218 X.getValueType().getVectorNumElements()));
6222 return DAG.getUNDEF(VT);
6225 assert(V.getNode() && "The single defined operand is empty!");
6226 SmallVector<SDValue, 8> Opnds;
6227 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
6229 Opnds.push_back(DAG.getUNDEF(VTs[i]));
6232 SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V);
6233 AddToWorklist(NV.getNode());
6234 Opnds.push_back(NV);
6236 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Opnds);
6240 // Simplify the operands using demanded-bits information.
6241 if (!VT.isVector() &&
6242 SimplifyDemandedBits(SDValue(N, 0)))
6243 return SDValue(N, 0);
6248 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
6249 SDValue Elt = N->getOperand(i);
6250 if (Elt.getOpcode() != ISD::MERGE_VALUES)
6251 return Elt.getNode();
6252 return Elt.getOperand(Elt.getResNo()).getNode();
6255 /// build_pair (load, load) -> load
6256 /// if load locations are consecutive.
6257 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
6258 assert(N->getOpcode() == ISD::BUILD_PAIR);
6260 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
6261 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
6262 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
6263 LD1->getAddressSpace() != LD2->getAddressSpace())
6265 EVT LD1VT = LD1->getValueType(0);
6267 if (ISD::isNON_EXTLoad(LD2) &&
6269 // If both are volatile this would reduce the number of volatile loads.
6270 // If one is volatile it might be ok, but play conservative and bail out.
6271 !LD1->isVolatile() &&
6272 !LD2->isVolatile() &&
6273 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
6274 unsigned Align = LD1->getAlignment();
6275 unsigned NewAlign = TLI.getDataLayout()->
6276 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6278 if (NewAlign <= Align &&
6279 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
6280 return DAG.getLoad(VT, SDLoc(N), LD1->getChain(),
6281 LD1->getBasePtr(), LD1->getPointerInfo(),
6282 false, false, false, Align);
6288 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
6289 SDValue N0 = N->getOperand(0);
6290 EVT VT = N->getValueType(0);
6292 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
6293 // Only do this before legalize, since afterward the target may be depending
6294 // on the bitconvert.
6295 // First check to see if this is all constant.
6297 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
6299 bool isSimple = cast<BuildVectorSDNode>(N0)->isConstant();
6301 EVT DestEltVT = N->getValueType(0).getVectorElementType();
6302 assert(!DestEltVT.isVector() &&
6303 "Element type of vector ValueType must not be vector!");
6305 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
6308 // If the input is a constant, let getNode fold it.
6309 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
6310 SDValue Res = DAG.getNode(ISD::BITCAST, SDLoc(N), VT, N0);
6311 if (Res.getNode() != N) {
6312 if (!LegalOperations ||
6313 TLI.isOperationLegal(Res.getNode()->getOpcode(), VT))
6316 // Folding it resulted in an illegal node, and it's too late to
6317 // do that. Clean up the old node and forego the transformation.
6318 // Ideally this won't happen very often, because instcombine
6319 // and the earlier dagcombine runs (where illegal nodes are
6320 // permitted) should have folded most of them already.
6321 deleteAndRecombine(Res.getNode());
6325 // (conv (conv x, t1), t2) -> (conv x, t2)
6326 if (N0.getOpcode() == ISD::BITCAST)
6327 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT,
6330 // fold (conv (load x)) -> (load (conv*)x)
6331 // If the resultant load doesn't need a higher alignment than the original!
6332 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6333 // Do not change the width of a volatile load.
6334 !cast<LoadSDNode>(N0)->isVolatile() &&
6335 // Do not remove the cast if the types differ in endian layout.
6336 TLI.hasBigEndianPartOrdering(N0.getValueType()) ==
6337 TLI.hasBigEndianPartOrdering(VT) &&
6338 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)) &&
6339 TLI.isLoadBitCastBeneficial(N0.getValueType(), VT)) {
6340 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6341 unsigned Align = TLI.getDataLayout()->
6342 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6343 unsigned OrigAlign = LN0->getAlignment();
6345 if (Align <= OrigAlign) {
6346 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(),
6347 LN0->getBasePtr(), LN0->getPointerInfo(),
6348 LN0->isVolatile(), LN0->isNonTemporal(),
6349 LN0->isInvariant(), OrigAlign,
6351 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
6356 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
6357 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
6358 // This often reduces constant pool loads.
6359 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) ||
6360 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) &&
6361 N0.getNode()->hasOneUse() && VT.isInteger() &&
6362 !VT.isVector() && !N0.getValueType().isVector()) {
6363 SDValue NewConv = DAG.getNode(ISD::BITCAST, SDLoc(N0), VT,
6365 AddToWorklist(NewConv.getNode());
6367 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6368 if (N0.getOpcode() == ISD::FNEG)
6369 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
6370 NewConv, DAG.getConstant(SignBit, VT));
6371 assert(N0.getOpcode() == ISD::FABS);
6372 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6373 NewConv, DAG.getConstant(~SignBit, VT));
6376 // fold (bitconvert (fcopysign cst, x)) ->
6377 // (or (and (bitconvert x), sign), (and cst, (not sign)))
6378 // Note that we don't handle (copysign x, cst) because this can always be
6379 // folded to an fneg or fabs.
6380 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
6381 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
6382 VT.isInteger() && !VT.isVector()) {
6383 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
6384 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
6385 if (isTypeLegal(IntXVT)) {
6386 SDValue X = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6387 IntXVT, N0.getOperand(1));
6388 AddToWorklist(X.getNode());
6390 // If X has a different width than the result/lhs, sext it or truncate it.
6391 unsigned VTWidth = VT.getSizeInBits();
6392 if (OrigXWidth < VTWidth) {
6393 X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X);
6394 AddToWorklist(X.getNode());
6395 } else if (OrigXWidth > VTWidth) {
6396 // To get the sign bit in the right place, we have to shift it right
6397 // before truncating.
6398 X = DAG.getNode(ISD::SRL, SDLoc(X),
6399 X.getValueType(), X,
6400 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
6401 AddToWorklist(X.getNode());
6402 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
6403 AddToWorklist(X.getNode());
6406 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6407 X = DAG.getNode(ISD::AND, SDLoc(X), VT,
6408 X, DAG.getConstant(SignBit, VT));
6409 AddToWorklist(X.getNode());
6411 SDValue Cst = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6412 VT, N0.getOperand(0));
6413 Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT,
6414 Cst, DAG.getConstant(~SignBit, VT));
6415 AddToWorklist(Cst.getNode());
6417 return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst);
6421 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
6422 if (N0.getOpcode() == ISD::BUILD_PAIR) {
6423 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
6424 if (CombineLD.getNode())
6431 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
6432 EVT VT = N->getValueType(0);
6433 return CombineConsecutiveLoads(N, VT);
6436 /// We know that BV is a build_vector node with Constant, ConstantFP or Undef
6437 /// operands. DstEltVT indicates the destination element value type.
6438 SDValue DAGCombiner::
6439 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
6440 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
6442 // If this is already the right type, we're done.
6443 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
6445 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
6446 unsigned DstBitSize = DstEltVT.getSizeInBits();
6448 // If this is a conversion of N elements of one type to N elements of another
6449 // type, convert each element. This handles FP<->INT cases.
6450 if (SrcBitSize == DstBitSize) {
6451 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6452 BV->getValueType(0).getVectorNumElements());
6454 // Due to the FP element handling below calling this routine recursively,
6455 // we can end up with a scalar-to-vector node here.
6456 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
6457 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6458 DAG.getNode(ISD::BITCAST, SDLoc(BV),
6459 DstEltVT, BV->getOperand(0)));
6461 SmallVector<SDValue, 8> Ops;
6462 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6463 SDValue Op = BV->getOperand(i);
6464 // If the vector element type is not legal, the BUILD_VECTOR operands
6465 // are promoted and implicitly truncated. Make that explicit here.
6466 if (Op.getValueType() != SrcEltVT)
6467 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op);
6468 Ops.push_back(DAG.getNode(ISD::BITCAST, SDLoc(BV),
6470 AddToWorklist(Ops.back().getNode());
6472 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6475 // Otherwise, we're growing or shrinking the elements. To avoid having to
6476 // handle annoying details of growing/shrinking FP values, we convert them to
6478 if (SrcEltVT.isFloatingPoint()) {
6479 // Convert the input float vector to a int vector where the elements are the
6481 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
6482 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
6483 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
6487 // Now we know the input is an integer vector. If the output is a FP type,
6488 // convert to integer first, then to FP of the right size.
6489 if (DstEltVT.isFloatingPoint()) {
6490 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
6491 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
6492 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
6494 // Next, convert to FP elements of the same size.
6495 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
6498 // Okay, we know the src/dst types are both integers of differing types.
6499 // Handling growing first.
6500 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
6501 if (SrcBitSize < DstBitSize) {
6502 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
6504 SmallVector<SDValue, 8> Ops;
6505 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
6506 i += NumInputsPerOutput) {
6507 bool isLE = TLI.isLittleEndian();
6508 APInt NewBits = APInt(DstBitSize, 0);
6509 bool EltIsUndef = true;
6510 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
6511 // Shift the previously computed bits over.
6512 NewBits <<= SrcBitSize;
6513 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
6514 if (Op.getOpcode() == ISD::UNDEF) continue;
6517 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
6518 zextOrTrunc(SrcBitSize).zext(DstBitSize);
6522 Ops.push_back(DAG.getUNDEF(DstEltVT));
6524 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
6527 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
6528 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6531 // Finally, this must be the case where we are shrinking elements: each input
6532 // turns into multiple outputs.
6533 bool isS2V = ISD::isScalarToVector(BV);
6534 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
6535 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6536 NumOutputsPerInput*BV->getNumOperands());
6537 SmallVector<SDValue, 8> Ops;
6539 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6540 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
6541 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
6542 Ops.push_back(DAG.getUNDEF(DstEltVT));
6546 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
6547 getAPIntValue().zextOrTrunc(SrcBitSize);
6549 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
6550 APInt ThisVal = OpVal.trunc(DstBitSize);
6551 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
6552 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
6553 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
6554 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6556 OpVal = OpVal.lshr(DstBitSize);
6559 // For big endian targets, swap the order of the pieces of each element.
6560 if (TLI.isBigEndian())
6561 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
6564 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6567 SDValue DAGCombiner::visitFADD(SDNode *N) {
6568 SDValue N0 = N->getOperand(0);
6569 SDValue N1 = N->getOperand(1);
6570 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6571 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6572 EVT VT = N->getValueType(0);
6573 const TargetOptions &Options = DAG.getTarget().Options;
6576 if (VT.isVector()) {
6577 SDValue FoldedVOp = SimplifyVBinOp(N);
6578 if (FoldedVOp.getNode()) return FoldedVOp;
6581 // fold (fadd c1, c2) -> c1 + c2
6583 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N1);
6585 // canonicalize constant to RHS
6586 if (N0CFP && !N1CFP)
6587 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N0);
6589 // fold (fadd A, (fneg B)) -> (fsub A, B)
6590 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6591 isNegatibleForFree(N1, LegalOperations, TLI, &Options) == 2)
6592 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0,
6593 GetNegatedExpression(N1, DAG, LegalOperations));
6595 // fold (fadd (fneg A), B) -> (fsub B, A)
6596 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6597 isNegatibleForFree(N0, LegalOperations, TLI, &Options) == 2)
6598 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N1,
6599 GetNegatedExpression(N0, DAG, LegalOperations));
6601 // If 'unsafe math' is enabled, fold lots of things.
6602 if (Options.UnsafeFPMath) {
6603 // No FP constant should be created after legalization as Instruction
6604 // Selection pass has a hard time dealing with FP constants.
6605 bool AllowNewConst = (Level < AfterLegalizeDAG);
6607 // fold (fadd A, 0) -> A
6608 if (N1CFP && N1CFP->getValueAPF().isZero())
6611 // fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
6612 if (N1CFP && N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
6613 isa<ConstantFPSDNode>(N0.getOperand(1)))
6614 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0.getOperand(0),
6615 DAG.getNode(ISD::FADD, SDLoc(N), VT,
6616 N0.getOperand(1), N1));
6618 // If allowed, fold (fadd (fneg x), x) -> 0.0
6619 if (AllowNewConst && N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1)
6620 return DAG.getConstantFP(0.0, VT);
6622 // If allowed, fold (fadd x, (fneg x)) -> 0.0
6623 if (AllowNewConst && N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0)
6624 return DAG.getConstantFP(0.0, VT);
6626 // We can fold chains of FADD's of the same value into multiplications.
6627 // This transform is not safe in general because we are reducing the number
6628 // of rounding steps.
6629 if (TLI.isOperationLegalOrCustom(ISD::FMUL, VT) && !N0CFP && !N1CFP) {
6630 if (N0.getOpcode() == ISD::FMUL) {
6631 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
6632 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
6634 // (fadd (fmul x, c), x) -> (fmul x, c+1)
6635 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
6636 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6638 DAG.getConstantFP(1.0, VT));
6639 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, NewCFP);
6642 // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2)
6643 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
6644 N1.getOperand(0) == N1.getOperand(1) &&
6645 N0.getOperand(0) == N1.getOperand(0)) {
6646 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6648 DAG.getConstantFP(2.0, VT));
6649 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6650 N0.getOperand(0), NewCFP);
6654 if (N1.getOpcode() == ISD::FMUL) {
6655 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
6656 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
6658 // (fadd x, (fmul x, c)) -> (fmul x, c+1)
6659 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
6660 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6662 DAG.getConstantFP(1.0, VT));
6663 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, NewCFP);
6666 // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2)
6667 if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD &&
6668 N0.getOperand(0) == N0.getOperand(1) &&
6669 N1.getOperand(0) == N0.getOperand(0)) {
6670 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6672 DAG.getConstantFP(2.0, VT));
6673 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1.getOperand(0), NewCFP);
6677 if (N0.getOpcode() == ISD::FADD && AllowNewConst) {
6678 ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
6679 // (fadd (fadd x, x), x) -> (fmul x, 3.0)
6680 if (!CFP && N0.getOperand(0) == N0.getOperand(1) &&
6681 (N0.getOperand(0) == N1))
6682 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6683 N1, DAG.getConstantFP(3.0, VT));
6686 if (N1.getOpcode() == ISD::FADD && AllowNewConst) {
6687 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
6688 // (fadd x, (fadd x, x)) -> (fmul x, 3.0)
6689 if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) &&
6690 N1.getOperand(0) == N0)
6691 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6692 N0, DAG.getConstantFP(3.0, VT));
6695 // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0)
6696 if (AllowNewConst &&
6697 N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
6698 N0.getOperand(0) == N0.getOperand(1) &&
6699 N1.getOperand(0) == N1.getOperand(1) &&
6700 N0.getOperand(0) == N1.getOperand(0))
6701 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6702 N0.getOperand(0), DAG.getConstantFP(4.0, VT));
6704 } // enable-unsafe-fp-math
6706 // FADD -> FMA combines:
6707 if ((Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath) &&
6708 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
6709 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
6711 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
6712 if (N0.getOpcode() == ISD::FMUL &&
6713 (N0->hasOneUse() || TLI.enableAggressiveFMAFusion(VT)))
6714 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6715 N0.getOperand(0), N0.getOperand(1), N1);
6717 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
6718 // Note: Commutes FADD operands.
6719 if (N1.getOpcode() == ISD::FMUL &&
6720 (N1->hasOneUse() || TLI.enableAggressiveFMAFusion(VT)))
6721 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6722 N1.getOperand(0), N1.getOperand(1), N0);
6728 SDValue DAGCombiner::visitFSUB(SDNode *N) {
6729 SDValue N0 = N->getOperand(0);
6730 SDValue N1 = N->getOperand(1);
6731 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
6732 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
6733 EVT VT = N->getValueType(0);
6735 const TargetOptions &Options = DAG.getTarget().Options;
6738 if (VT.isVector()) {
6739 SDValue FoldedVOp = SimplifyVBinOp(N);
6740 if (FoldedVOp.getNode()) return FoldedVOp;
6743 // fold (fsub c1, c2) -> c1-c2
6745 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0, N1);
6747 // fold (fsub A, (fneg B)) -> (fadd A, B)
6748 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
6749 return DAG.getNode(ISD::FADD, dl, VT, N0,
6750 GetNegatedExpression(N1, DAG, LegalOperations));
6752 // If 'unsafe math' is enabled, fold lots of things.
6753 if (Options.UnsafeFPMath) {
6755 if (N1CFP && N1CFP->getValueAPF().isZero())
6758 // (fsub 0, B) -> -B
6759 if (N0CFP && N0CFP->getValueAPF().isZero()) {
6760 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
6761 return GetNegatedExpression(N1, DAG, LegalOperations);
6762 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6763 return DAG.getNode(ISD::FNEG, dl, VT, N1);
6766 // (fsub x, x) -> 0.0
6768 return DAG.getConstantFP(0.0f, VT);
6770 // (fsub x, (fadd x, y)) -> (fneg y)
6771 // (fsub x, (fadd y, x)) -> (fneg y)
6772 if (N1.getOpcode() == ISD::FADD) {
6773 SDValue N10 = N1->getOperand(0);
6774 SDValue N11 = N1->getOperand(1);
6776 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI, &Options))
6777 return GetNegatedExpression(N11, DAG, LegalOperations);
6779 if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI, &Options))
6780 return GetNegatedExpression(N10, DAG, LegalOperations);
6784 // FSUB -> FMA combines:
6785 if ((Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath) &&
6786 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
6787 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
6789 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
6790 if (N0.getOpcode() == ISD::FMUL &&
6791 (N0->hasOneUse() || TLI.enableAggressiveFMAFusion(VT)))
6792 return DAG.getNode(ISD::FMA, dl, VT,
6793 N0.getOperand(0), N0.getOperand(1),
6794 DAG.getNode(ISD::FNEG, dl, VT, N1));
6796 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
6797 // Note: Commutes FSUB operands.
6798 if (N1.getOpcode() == ISD::FMUL &&
6799 (N1->hasOneUse() || TLI.enableAggressiveFMAFusion(VT)))
6800 return DAG.getNode(ISD::FMA, dl, VT,
6801 DAG.getNode(ISD::FNEG, dl, VT,
6803 N1.getOperand(1), N0);
6805 // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
6806 if (N0.getOpcode() == ISD::FNEG &&
6807 N0.getOperand(0).getOpcode() == ISD::FMUL &&
6808 ((N0->hasOneUse() && N0.getOperand(0).hasOneUse()) ||
6809 TLI.enableAggressiveFMAFusion(VT))) {
6810 SDValue N00 = N0.getOperand(0).getOperand(0);
6811 SDValue N01 = N0.getOperand(0).getOperand(1);
6812 return DAG.getNode(ISD::FMA, dl, VT,
6813 DAG.getNode(ISD::FNEG, dl, VT, N00), N01,
6814 DAG.getNode(ISD::FNEG, dl, VT, N1));
6821 SDValue DAGCombiner::visitFMUL(SDNode *N) {
6822 SDValue N0 = N->getOperand(0);
6823 SDValue N1 = N->getOperand(1);
6824 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
6825 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
6826 EVT VT = N->getValueType(0);
6827 const TargetOptions &Options = DAG.getTarget().Options;
6830 if (VT.isVector()) {
6831 // This just handles C1 * C2 for vectors. Other vector folds are below.
6832 SDValue FoldedVOp = SimplifyVBinOp(N);
6833 if (FoldedVOp.getNode())
6835 // Canonicalize vector constant to RHS.
6836 if (N0.getOpcode() == ISD::BUILD_VECTOR &&
6837 N1.getOpcode() != ISD::BUILD_VECTOR)
6838 if (auto *BV0 = dyn_cast<BuildVectorSDNode>(N0))
6839 if (BV0->isConstant())
6840 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, N1, N0);
6843 // fold (fmul c1, c2) -> c1*c2
6845 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, N1);
6847 // canonicalize constant to RHS
6848 if (N0CFP && !N1CFP)
6849 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, N0);
6851 // fold (fmul A, 1.0) -> A
6852 if (N1CFP && N1CFP->isExactlyValue(1.0))
6855 if (Options.UnsafeFPMath) {
6856 // fold (fmul A, 0) -> 0
6857 if (N1CFP && N1CFP->getValueAPF().isZero())
6860 // fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
6861 if (N0.getOpcode() == ISD::FMUL) {
6862 // Fold scalars or any vector constants (not just splats).
6863 // This fold is done in general by InstCombine, but extra fmul insts
6864 // may have been generated during lowering.
6865 SDValue N01 = N0.getOperand(1);
6866 auto *BV1 = dyn_cast<BuildVectorSDNode>(N1);
6867 auto *BV01 = dyn_cast<BuildVectorSDNode>(N01);
6868 if ((N1CFP && isConstOrConstSplatFP(N01)) ||
6869 (BV1 && BV01 && BV1->isConstant() && BV01->isConstant())) {
6871 SDValue MulConsts = DAG.getNode(ISD::FMUL, SL, VT, N01, N1);
6872 return DAG.getNode(ISD::FMUL, SL, VT, N0.getOperand(0), MulConsts);
6876 // fold (fmul (fadd x, x), c) -> (fmul x, (fmul 2.0, c))
6877 // Undo the fmul 2.0, x -> fadd x, x transformation, since if it occurs
6878 // during an early run of DAGCombiner can prevent folding with fmuls
6879 // inserted during lowering.
6880 if (N0.getOpcode() == ISD::FADD && N0.getOperand(0) == N0.getOperand(1)) {
6882 const SDValue Two = DAG.getConstantFP(2.0, VT);
6883 SDValue MulConsts = DAG.getNode(ISD::FMUL, SL, VT, Two, N1);
6884 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0.getOperand(0), MulConsts);
6888 // fold (fmul X, 2.0) -> (fadd X, X)
6889 if (N1CFP && N1CFP->isExactlyValue(+2.0))
6890 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N0);
6892 // fold (fmul X, -1.0) -> (fneg X)
6893 if (N1CFP && N1CFP->isExactlyValue(-1.0))
6894 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6895 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
6897 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
6898 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
6899 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
6900 // Both can be negated for free, check to see if at least one is cheaper
6902 if (LHSNeg == 2 || RHSNeg == 2)
6903 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6904 GetNegatedExpression(N0, DAG, LegalOperations),
6905 GetNegatedExpression(N1, DAG, LegalOperations));
6912 SDValue DAGCombiner::visitFMA(SDNode *N) {
6913 SDValue N0 = N->getOperand(0);
6914 SDValue N1 = N->getOperand(1);
6915 SDValue N2 = N->getOperand(2);
6916 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6917 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6918 EVT VT = N->getValueType(0);
6920 const TargetOptions &Options = DAG.getTarget().Options;
6922 // Constant fold FMA.
6923 if (isa<ConstantFPSDNode>(N0) &&
6924 isa<ConstantFPSDNode>(N1) &&
6925 isa<ConstantFPSDNode>(N2)) {
6926 return DAG.getNode(ISD::FMA, dl, VT, N0, N1, N2);
6929 if (Options.UnsafeFPMath) {
6930 if (N0CFP && N0CFP->isZero())
6932 if (N1CFP && N1CFP->isZero())
6935 if (N0CFP && N0CFP->isExactlyValue(1.0))
6936 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2);
6937 if (N1CFP && N1CFP->isExactlyValue(1.0))
6938 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2);
6940 // Canonicalize (fma c, x, y) -> (fma x, c, y)
6941 if (N0CFP && !N1CFP)
6942 return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2);
6944 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
6945 if (Options.UnsafeFPMath && N1CFP &&
6946 N2.getOpcode() == ISD::FMUL &&
6947 N0 == N2.getOperand(0) &&
6948 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
6949 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6950 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
6954 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
6955 if (Options.UnsafeFPMath &&
6956 N0.getOpcode() == ISD::FMUL && N1CFP &&
6957 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
6958 return DAG.getNode(ISD::FMA, dl, VT,
6960 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
6964 // (fma x, 1, y) -> (fadd x, y)
6965 // (fma x, -1, y) -> (fadd (fneg x), y)
6967 if (N1CFP->isExactlyValue(1.0))
6968 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
6970 if (N1CFP->isExactlyValue(-1.0) &&
6971 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
6972 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
6973 AddToWorklist(RHSNeg.getNode());
6974 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
6978 // (fma x, c, x) -> (fmul x, (c+1))
6979 if (Options.UnsafeFPMath && N1CFP && N0 == N2)
6980 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6981 DAG.getNode(ISD::FADD, dl, VT,
6982 N1, DAG.getConstantFP(1.0, VT)));
6984 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
6985 if (Options.UnsafeFPMath && N1CFP &&
6986 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0)
6987 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6988 DAG.getNode(ISD::FADD, dl, VT,
6989 N1, DAG.getConstantFP(-1.0, VT)));
6995 SDValue DAGCombiner::visitFDIV(SDNode *N) {
6996 SDValue N0 = N->getOperand(0);
6997 SDValue N1 = N->getOperand(1);
6998 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6999 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7000 EVT VT = N->getValueType(0);
7002 const TargetOptions &Options = DAG.getTarget().Options;
7005 if (VT.isVector()) {
7006 SDValue FoldedVOp = SimplifyVBinOp(N);
7007 if (FoldedVOp.getNode()) return FoldedVOp;
7010 // fold (fdiv c1, c2) -> c1/c2
7012 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1);
7014 if (Options.UnsafeFPMath) {
7015 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
7017 // Compute the reciprocal 1.0 / c2.
7018 APFloat N1APF = N1CFP->getValueAPF();
7019 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
7020 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
7021 // Only do the transform if the reciprocal is a legal fp immediate that
7022 // isn't too nasty (eg NaN, denormal, ...).
7023 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
7024 (!LegalOperations ||
7025 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
7026 // backend)... we should handle this gracefully after Legalize.
7027 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
7028 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
7029 TLI.isFPImmLegal(Recip, VT)))
7030 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0,
7031 DAG.getConstantFP(Recip, VT));
7034 // If this FDIV is part of a reciprocal square root, it may be folded
7035 // into a target-specific square root estimate instruction.
7036 if (N1.getOpcode() == ISD::FSQRT) {
7037 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0))) {
7038 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7040 } else if (N1.getOpcode() == ISD::FP_EXTEND &&
7041 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
7042 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0).getOperand(0))) {
7043 RV = DAG.getNode(ISD::FP_EXTEND, SDLoc(N1), VT, RV);
7044 AddToWorklist(RV.getNode());
7045 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7047 } else if (N1.getOpcode() == ISD::FP_ROUND &&
7048 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
7049 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0).getOperand(0))) {
7050 RV = DAG.getNode(ISD::FP_ROUND, SDLoc(N1), VT, RV, N1.getOperand(1));
7051 AddToWorklist(RV.getNode());
7052 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7054 } else if (N1.getOpcode() == ISD::FMUL) {
7055 // Look through an FMUL. Even though this won't remove the FDIV directly,
7056 // it's still worthwhile to get rid of the FSQRT if possible.
7059 if (N1.getOperand(0).getOpcode() == ISD::FSQRT) {
7060 SqrtOp = N1.getOperand(0);
7061 OtherOp = N1.getOperand(1);
7062 } else if (N1.getOperand(1).getOpcode() == ISD::FSQRT) {
7063 SqrtOp = N1.getOperand(1);
7064 OtherOp = N1.getOperand(0);
7066 if (SqrtOp.getNode()) {
7067 // We found a FSQRT, so try to make this fold:
7068 // x / (y * sqrt(z)) -> x * (rsqrt(z) / y)
7069 if (SDValue RV = BuildRsqrtEstimate(SqrtOp.getOperand(0))) {
7070 RV = DAG.getNode(ISD::FDIV, SDLoc(N1), VT, RV, OtherOp);
7071 AddToWorklist(RV.getNode());
7072 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7077 // Fold into a reciprocal estimate and multiply instead of a real divide.
7078 if (SDValue RV = BuildReciprocalEstimate(N1)) {
7079 AddToWorklist(RV.getNode());
7080 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7084 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
7085 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
7086 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
7087 // Both can be negated for free, check to see if at least one is cheaper
7089 if (LHSNeg == 2 || RHSNeg == 2)
7090 return DAG.getNode(ISD::FDIV, SDLoc(N), VT,
7091 GetNegatedExpression(N0, DAG, LegalOperations),
7092 GetNegatedExpression(N1, DAG, LegalOperations));
7099 SDValue DAGCombiner::visitFREM(SDNode *N) {
7100 SDValue N0 = N->getOperand(0);
7101 SDValue N1 = N->getOperand(1);
7102 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7103 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7104 EVT VT = N->getValueType(0);
7106 // fold (frem c1, c2) -> fmod(c1,c2)
7108 return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1);
7113 SDValue DAGCombiner::visitFSQRT(SDNode *N) {
7114 if (DAG.getTarget().Options.UnsafeFPMath) {
7115 // Compute this as X * (1/sqrt(X)) = X * (X ** -0.5)
7116 if (SDValue RV = BuildRsqrtEstimate(N->getOperand(0))) {
7117 EVT VT = RV.getValueType();
7118 RV = DAG.getNode(ISD::FMUL, SDLoc(N), VT, N->getOperand(0), RV);
7119 AddToWorklist(RV.getNode());
7121 // Unfortunately, RV is now NaN if the input was exactly 0.
7122 // Select out this case and force the answer to 0.
7123 SDValue Zero = DAG.getConstantFP(0.0, VT);
7125 DAG.getSetCC(SDLoc(N), TLI.getSetCCResultType(*DAG.getContext(), VT),
7126 N->getOperand(0), Zero, ISD::SETEQ);
7127 AddToWorklist(ZeroCmp.getNode());
7128 AddToWorklist(RV.getNode());
7130 RV = DAG.getNode(VT.isVector() ? ISD::VSELECT : ISD::SELECT,
7131 SDLoc(N), VT, ZeroCmp, Zero, RV);
7138 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
7139 SDValue N0 = N->getOperand(0);
7140 SDValue N1 = N->getOperand(1);
7141 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7142 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7143 EVT VT = N->getValueType(0);
7145 if (N0CFP && N1CFP) // Constant fold
7146 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1);
7149 const APFloat& V = N1CFP->getValueAPF();
7150 // copysign(x, c1) -> fabs(x) iff ispos(c1)
7151 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
7152 if (!V.isNegative()) {
7153 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
7154 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7156 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7157 return DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7158 DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0));
7162 // copysign(fabs(x), y) -> copysign(x, y)
7163 // copysign(fneg(x), y) -> copysign(x, y)
7164 // copysign(copysign(x,z), y) -> copysign(x, y)
7165 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
7166 N0.getOpcode() == ISD::FCOPYSIGN)
7167 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7168 N0.getOperand(0), N1);
7170 // copysign(x, abs(y)) -> abs(x)
7171 if (N1.getOpcode() == ISD::FABS)
7172 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7174 // copysign(x, copysign(y,z)) -> copysign(x, z)
7175 if (N1.getOpcode() == ISD::FCOPYSIGN)
7176 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7177 N0, N1.getOperand(1));
7179 // copysign(x, fp_extend(y)) -> copysign(x, y)
7180 // copysign(x, fp_round(y)) -> copysign(x, y)
7181 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
7182 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7183 N0, N1.getOperand(0));
7188 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
7189 SDValue N0 = N->getOperand(0);
7190 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
7191 EVT VT = N->getValueType(0);
7192 EVT OpVT = N0.getValueType();
7194 // fold (sint_to_fp c1) -> c1fp
7196 // ...but only if the target supports immediate floating-point values
7197 (!LegalOperations ||
7198 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
7199 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
7201 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
7202 // but UINT_TO_FP is legal on this target, try to convert.
7203 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
7204 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
7205 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
7206 if (DAG.SignBitIsZero(N0))
7207 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
7210 // The next optimizations are desirable only if SELECT_CC can be lowered.
7211 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
7212 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7213 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
7215 (!LegalOperations ||
7216 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7218 { N0.getOperand(0), N0.getOperand(1),
7219 DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT),
7221 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7224 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
7225 // (select_cc x, y, 1.0, 0.0,, cc)
7226 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
7227 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
7228 (!LegalOperations ||
7229 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7231 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
7232 DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT),
7233 N0.getOperand(0).getOperand(2) };
7234 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7241 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
7242 SDValue N0 = N->getOperand(0);
7243 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
7244 EVT VT = N->getValueType(0);
7245 EVT OpVT = N0.getValueType();
7247 // fold (uint_to_fp c1) -> c1fp
7249 // ...but only if the target supports immediate floating-point values
7250 (!LegalOperations ||
7251 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
7252 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
7254 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
7255 // but SINT_TO_FP is legal on this target, try to convert.
7256 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
7257 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
7258 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
7259 if (DAG.SignBitIsZero(N0))
7260 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
7263 // The next optimizations are desirable only if SELECT_CC can be lowered.
7264 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
7265 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7267 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
7268 (!LegalOperations ||
7269 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7271 { N0.getOperand(0), N0.getOperand(1),
7272 DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT),
7274 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7281 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
7282 SDValue N0 = N->getOperand(0);
7283 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7284 EVT VT = N->getValueType(0);
7286 // fold (fp_to_sint c1fp) -> c1
7288 return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0);
7293 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
7294 SDValue N0 = N->getOperand(0);
7295 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7296 EVT VT = N->getValueType(0);
7298 // fold (fp_to_uint c1fp) -> c1
7300 return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0);
7305 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
7306 SDValue N0 = N->getOperand(0);
7307 SDValue N1 = N->getOperand(1);
7308 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7309 EVT VT = N->getValueType(0);
7311 // fold (fp_round c1fp) -> c1fp
7313 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1);
7315 // fold (fp_round (fp_extend x)) -> x
7316 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
7317 return N0.getOperand(0);
7319 // fold (fp_round (fp_round x)) -> (fp_round x)
7320 if (N0.getOpcode() == ISD::FP_ROUND) {
7321 // This is a value preserving truncation if both round's are.
7322 bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
7323 N0.getNode()->getConstantOperandVal(1) == 1;
7324 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0.getOperand(0),
7325 DAG.getIntPtrConstant(IsTrunc));
7328 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
7329 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
7330 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT,
7331 N0.getOperand(0), N1);
7332 AddToWorklist(Tmp.getNode());
7333 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7334 Tmp, N0.getOperand(1));
7340 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
7341 SDValue N0 = N->getOperand(0);
7342 EVT VT = N->getValueType(0);
7343 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
7344 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7346 // fold (fp_round_inreg c1fp) -> c1fp
7347 if (N0CFP && isTypeLegal(EVT)) {
7348 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
7349 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, Round);
7355 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
7356 SDValue N0 = N->getOperand(0);
7357 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7358 EVT VT = N->getValueType(0);
7360 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
7361 if (N->hasOneUse() &&
7362 N->use_begin()->getOpcode() == ISD::FP_ROUND)
7365 // fold (fp_extend c1fp) -> c1fp
7367 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0);
7369 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
7371 if (N0.getOpcode() == ISD::FP_ROUND
7372 && N0.getNode()->getConstantOperandVal(1) == 1) {
7373 SDValue In = N0.getOperand(0);
7374 if (In.getValueType() == VT) return In;
7375 if (VT.bitsLT(In.getValueType()))
7376 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT,
7377 In, N0.getOperand(1));
7378 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In);
7381 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
7382 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7383 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType())) {
7384 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7385 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
7387 LN0->getBasePtr(), N0.getValueType(),
7388 LN0->getMemOperand());
7389 CombineTo(N, ExtLoad);
7390 CombineTo(N0.getNode(),
7391 DAG.getNode(ISD::FP_ROUND, SDLoc(N0),
7392 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
7393 ExtLoad.getValue(1));
7394 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7400 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
7401 SDValue N0 = N->getOperand(0);
7402 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7403 EVT VT = N->getValueType(0);
7405 // fold (fceil c1) -> fceil(c1)
7407 return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0);
7412 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
7413 SDValue N0 = N->getOperand(0);
7414 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7415 EVT VT = N->getValueType(0);
7417 // fold (ftrunc c1) -> ftrunc(c1)
7419 return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0);
7424 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
7425 SDValue N0 = N->getOperand(0);
7426 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7427 EVT VT = N->getValueType(0);
7429 // fold (ffloor c1) -> ffloor(c1)
7431 return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0);
7436 // FIXME: FNEG and FABS have a lot in common; refactor.
7437 SDValue DAGCombiner::visitFNEG(SDNode *N) {
7438 SDValue N0 = N->getOperand(0);
7439 EVT VT = N->getValueType(0);
7441 if (VT.isVector()) {
7442 SDValue FoldedVOp = SimplifyVUnaryOp(N);
7443 if (FoldedVOp.getNode()) return FoldedVOp;
7446 // Constant fold FNEG.
7447 if (isa<ConstantFPSDNode>(N0))
7448 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N->getOperand(0));
7450 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
7451 &DAG.getTarget().Options))
7452 return GetNegatedExpression(N0, DAG, LegalOperations);
7454 // Transform fneg(bitconvert(x)) -> bitconvert(x ^ sign) to avoid loading
7455 // constant pool values.
7456 if (!TLI.isFNegFree(VT) &&
7457 N0.getOpcode() == ISD::BITCAST &&
7458 N0.getNode()->hasOneUse()) {
7459 SDValue Int = N0.getOperand(0);
7460 EVT IntVT = Int.getValueType();
7461 if (IntVT.isInteger() && !IntVT.isVector()) {
7463 if (N0.getValueType().isVector()) {
7464 // For a vector, get a mask such as 0x80... per scalar element
7466 SignMask = APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
7467 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
7469 // For a scalar, just generate 0x80...
7470 SignMask = APInt::getSignBit(IntVT.getSizeInBits());
7472 Int = DAG.getNode(ISD::XOR, SDLoc(N0), IntVT, Int,
7473 DAG.getConstant(SignMask, IntVT));
7474 AddToWorklist(Int.getNode());
7475 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Int);
7479 // (fneg (fmul c, x)) -> (fmul -c, x)
7480 if (N0.getOpcode() == ISD::FMUL) {
7481 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
7483 APFloat CVal = CFP1->getValueAPF();
7485 if (Level >= AfterLegalizeDAG &&
7486 (TLI.isFPImmLegal(CVal, N->getValueType(0)) ||
7487 TLI.isOperationLegal(ISD::ConstantFP, N->getValueType(0))))
7489 ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
7490 DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0.getOperand(1)));
7497 SDValue DAGCombiner::visitFMINNUM(SDNode *N) {
7498 SDValue N0 = N->getOperand(0);
7499 SDValue N1 = N->getOperand(1);
7500 const ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7501 const ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7503 if (N0CFP && N1CFP) {
7504 const APFloat &C0 = N0CFP->getValueAPF();
7505 const APFloat &C1 = N1CFP->getValueAPF();
7506 return DAG.getConstantFP(minnum(C0, C1), N->getValueType(0));
7510 EVT VT = N->getValueType(0);
7511 // Canonicalize to constant on RHS.
7512 return DAG.getNode(ISD::FMINNUM, SDLoc(N), VT, N1, N0);
7518 SDValue DAGCombiner::visitFMAXNUM(SDNode *N) {
7519 SDValue N0 = N->getOperand(0);
7520 SDValue N1 = N->getOperand(1);
7521 const ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7522 const ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7524 if (N0CFP && N1CFP) {
7525 const APFloat &C0 = N0CFP->getValueAPF();
7526 const APFloat &C1 = N1CFP->getValueAPF();
7527 return DAG.getConstantFP(maxnum(C0, C1), N->getValueType(0));
7531 EVT VT = N->getValueType(0);
7532 // Canonicalize to constant on RHS.
7533 return DAG.getNode(ISD::FMAXNUM, SDLoc(N), VT, N1, N0);
7539 SDValue DAGCombiner::visitFABS(SDNode *N) {
7540 SDValue N0 = N->getOperand(0);
7541 EVT VT = N->getValueType(0);
7543 if (VT.isVector()) {
7544 SDValue FoldedVOp = SimplifyVUnaryOp(N);
7545 if (FoldedVOp.getNode()) return FoldedVOp;
7548 // fold (fabs c1) -> fabs(c1)
7549 if (isa<ConstantFPSDNode>(N0))
7550 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7552 // fold (fabs (fabs x)) -> (fabs x)
7553 if (N0.getOpcode() == ISD::FABS)
7554 return N->getOperand(0);
7556 // fold (fabs (fneg x)) -> (fabs x)
7557 // fold (fabs (fcopysign x, y)) -> (fabs x)
7558 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
7559 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0));
7561 // Transform fabs(bitconvert(x)) -> bitconvert(x & ~sign) to avoid loading
7562 // constant pool values.
7563 if (!TLI.isFAbsFree(VT) &&
7564 N0.getOpcode() == ISD::BITCAST &&
7565 N0.getNode()->hasOneUse()) {
7566 SDValue Int = N0.getOperand(0);
7567 EVT IntVT = Int.getValueType();
7568 if (IntVT.isInteger() && !IntVT.isVector()) {
7570 if (N0.getValueType().isVector()) {
7571 // For a vector, get a mask such as 0x7f... per scalar element
7573 SignMask = ~APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
7574 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
7576 // For a scalar, just generate 0x7f...
7577 SignMask = ~APInt::getSignBit(IntVT.getSizeInBits());
7579 Int = DAG.getNode(ISD::AND, SDLoc(N0), IntVT, Int,
7580 DAG.getConstant(SignMask, IntVT));
7581 AddToWorklist(Int.getNode());
7582 return DAG.getNode(ISD::BITCAST, SDLoc(N), N->getValueType(0), Int);
7589 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
7590 SDValue Chain = N->getOperand(0);
7591 SDValue N1 = N->getOperand(1);
7592 SDValue N2 = N->getOperand(2);
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 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
7602 if (N1.getOpcode() == ISD::SETCC &&
7603 TLI.isOperationLegalOrCustom(ISD::BR_CC,
7604 N1.getOperand(0).getValueType())) {
7605 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
7606 Chain, N1.getOperand(2),
7607 N1.getOperand(0), N1.getOperand(1), N2);
7610 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
7611 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
7612 (N1.getOperand(0).hasOneUse() &&
7613 N1.getOperand(0).getOpcode() == ISD::SRL))) {
7614 SDNode *Trunc = nullptr;
7615 if (N1.getOpcode() == ISD::TRUNCATE) {
7616 // Look pass the truncate.
7617 Trunc = N1.getNode();
7618 N1 = N1.getOperand(0);
7621 // Match this pattern so that we can generate simpler code:
7624 // %b = and i32 %a, 2
7625 // %c = srl i32 %b, 1
7626 // brcond i32 %c ...
7631 // %b = and i32 %a, 2
7632 // %c = setcc eq %b, 0
7635 // This applies only when the AND constant value has one bit set and the
7636 // SRL constant is equal to the log2 of the AND constant. The back-end is
7637 // smart enough to convert the result into a TEST/JMP sequence.
7638 SDValue Op0 = N1.getOperand(0);
7639 SDValue Op1 = N1.getOperand(1);
7641 if (Op0.getOpcode() == ISD::AND &&
7642 Op1.getOpcode() == ISD::Constant) {
7643 SDValue AndOp1 = Op0.getOperand(1);
7645 if (AndOp1.getOpcode() == ISD::Constant) {
7646 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
7648 if (AndConst.isPowerOf2() &&
7649 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
7651 DAG.getSetCC(SDLoc(N),
7652 getSetCCResultType(Op0.getValueType()),
7653 Op0, DAG.getConstant(0, Op0.getValueType()),
7656 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, SDLoc(N),
7657 MVT::Other, Chain, SetCC, N2);
7658 // Don't add the new BRCond into the worklist or else SimplifySelectCC
7659 // will convert it back to (X & C1) >> C2.
7660 CombineTo(N, NewBRCond, false);
7661 // Truncate is dead.
7663 deleteAndRecombine(Trunc);
7664 // Replace the uses of SRL with SETCC
7665 WorklistRemover DeadNodes(*this);
7666 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
7667 deleteAndRecombine(N1.getNode());
7668 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7674 // Restore N1 if the above transformation doesn't match.
7675 N1 = N->getOperand(1);
7678 // Transform br(xor(x, y)) -> br(x != y)
7679 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
7680 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
7681 SDNode *TheXor = N1.getNode();
7682 SDValue Op0 = TheXor->getOperand(0);
7683 SDValue Op1 = TheXor->getOperand(1);
7684 if (Op0.getOpcode() == Op1.getOpcode()) {
7685 // Avoid missing important xor optimizations.
7686 SDValue Tmp = visitXOR(TheXor);
7687 if (Tmp.getNode()) {
7688 if (Tmp.getNode() != TheXor) {
7689 DEBUG(dbgs() << "\nReplacing.8 ";
7691 dbgs() << "\nWith: ";
7692 Tmp.getNode()->dump(&DAG);
7694 WorklistRemover DeadNodes(*this);
7695 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
7696 deleteAndRecombine(TheXor);
7697 return DAG.getNode(ISD::BRCOND, SDLoc(N),
7698 MVT::Other, Chain, Tmp, N2);
7701 // visitXOR has changed XOR's operands or replaced the XOR completely,
7703 return SDValue(N, 0);
7707 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
7709 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
7710 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
7711 Op0.getOpcode() == ISD::XOR) {
7712 TheXor = Op0.getNode();
7716 EVT SetCCVT = N1.getValueType();
7718 SetCCVT = getSetCCResultType(SetCCVT);
7719 SDValue SetCC = DAG.getSetCC(SDLoc(TheXor),
7722 Equal ? ISD::SETEQ : ISD::SETNE);
7723 // Replace the uses of XOR with SETCC
7724 WorklistRemover DeadNodes(*this);
7725 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
7726 deleteAndRecombine(N1.getNode());
7727 return DAG.getNode(ISD::BRCOND, SDLoc(N),
7728 MVT::Other, Chain, SetCC, N2);
7735 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
7737 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
7738 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
7739 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
7741 // If N is a constant we could fold this into a fallthrough or unconditional
7742 // branch. However that doesn't happen very often in normal code, because
7743 // Instcombine/SimplifyCFG should have handled the available opportunities.
7744 // If we did this folding here, it would be necessary to update the
7745 // MachineBasicBlock CFG, which is awkward.
7747 // Use SimplifySetCC to simplify SETCC's.
7748 SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()),
7749 CondLHS, CondRHS, CC->get(), SDLoc(N),
7751 if (Simp.getNode()) AddToWorklist(Simp.getNode());
7753 // fold to a simpler setcc
7754 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
7755 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
7756 N->getOperand(0), Simp.getOperand(2),
7757 Simp.getOperand(0), Simp.getOperand(1),
7763 /// Return true if 'Use' is a load or a store that uses N as its base pointer
7764 /// and that N may be folded in the load / store addressing mode.
7765 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
7767 const TargetLowering &TLI) {
7769 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
7770 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
7772 VT = Use->getValueType(0);
7773 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
7774 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
7776 VT = ST->getValue().getValueType();
7780 TargetLowering::AddrMode AM;
7781 if (N->getOpcode() == ISD::ADD) {
7782 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
7785 AM.BaseOffs = Offset->getSExtValue();
7789 } else if (N->getOpcode() == ISD::SUB) {
7790 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
7793 AM.BaseOffs = -Offset->getSExtValue();
7800 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
7803 /// Try turning a load/store into a pre-indexed load/store when the base
7804 /// pointer is an add or subtract and it has other uses besides the load/store.
7805 /// After the transformation, the new indexed load/store has effectively folded
7806 /// the add/subtract in and all of its other uses are redirected to the
7808 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
7809 if (Level < AfterLegalizeDAG)
7815 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7816 if (LD->isIndexed())
7818 VT = LD->getMemoryVT();
7819 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
7820 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
7822 Ptr = LD->getBasePtr();
7823 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7824 if (ST->isIndexed())
7826 VT = ST->getMemoryVT();
7827 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
7828 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
7830 Ptr = ST->getBasePtr();
7836 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
7837 // out. There is no reason to make this a preinc/predec.
7838 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
7839 Ptr.getNode()->hasOneUse())
7842 // Ask the target to do addressing mode selection.
7845 ISD::MemIndexedMode AM = ISD::UNINDEXED;
7846 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
7849 // Backends without true r+i pre-indexed forms may need to pass a
7850 // constant base with a variable offset so that constant coercion
7851 // will work with the patterns in canonical form.
7852 bool Swapped = false;
7853 if (isa<ConstantSDNode>(BasePtr)) {
7854 std::swap(BasePtr, Offset);
7858 // Don't create a indexed load / store with zero offset.
7859 if (isa<ConstantSDNode>(Offset) &&
7860 cast<ConstantSDNode>(Offset)->isNullValue())
7863 // Try turning it into a pre-indexed load / store except when:
7864 // 1) The new base ptr is a frame index.
7865 // 2) If N is a store and the new base ptr is either the same as or is a
7866 // predecessor of the value being stored.
7867 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
7868 // that would create a cycle.
7869 // 4) All uses are load / store ops that use it as old base ptr.
7871 // Check #1. Preinc'ing a frame index would require copying the stack pointer
7872 // (plus the implicit offset) to a register to preinc anyway.
7873 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
7878 SDValue Val = cast<StoreSDNode>(N)->getValue();
7879 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
7883 // If the offset is a constant, there may be other adds of constants that
7884 // can be folded with this one. We should do this to avoid having to keep
7885 // a copy of the original base pointer.
7886 SmallVector<SDNode *, 16> OtherUses;
7887 if (isa<ConstantSDNode>(Offset))
7888 for (SDNode *Use : BasePtr.getNode()->uses()) {
7889 if (Use == Ptr.getNode())
7892 if (Use->isPredecessorOf(N))
7895 if (Use->getOpcode() != ISD::ADD && Use->getOpcode() != ISD::SUB) {
7900 SDValue Op0 = Use->getOperand(0), Op1 = Use->getOperand(1);
7901 if (Op1.getNode() == BasePtr.getNode())
7902 std::swap(Op0, Op1);
7903 assert(Op0.getNode() == BasePtr.getNode() &&
7904 "Use of ADD/SUB but not an operand");
7906 if (!isa<ConstantSDNode>(Op1)) {
7911 // FIXME: In some cases, we can be smarter about this.
7912 if (Op1.getValueType() != Offset.getValueType()) {
7917 OtherUses.push_back(Use);
7921 std::swap(BasePtr, Offset);
7923 // Now check for #3 and #4.
7924 bool RealUse = false;
7926 // Caches for hasPredecessorHelper
7927 SmallPtrSet<const SDNode *, 32> Visited;
7928 SmallVector<const SDNode *, 16> Worklist;
7930 for (SDNode *Use : Ptr.getNode()->uses()) {
7933 if (N->hasPredecessorHelper(Use, Visited, Worklist))
7936 // If Ptr may be folded in addressing mode of other use, then it's
7937 // not profitable to do this transformation.
7938 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
7947 Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
7948 BasePtr, Offset, AM);
7950 Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
7951 BasePtr, Offset, AM);
7954 DEBUG(dbgs() << "\nReplacing.4 ";
7956 dbgs() << "\nWith: ";
7957 Result.getNode()->dump(&DAG);
7959 WorklistRemover DeadNodes(*this);
7961 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
7962 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
7964 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
7967 // Finally, since the node is now dead, remove it from the graph.
7968 deleteAndRecombine(N);
7971 std::swap(BasePtr, Offset);
7973 // Replace other uses of BasePtr that can be updated to use Ptr
7974 for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) {
7975 unsigned OffsetIdx = 1;
7976 if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode())
7978 assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() ==
7979 BasePtr.getNode() && "Expected BasePtr operand");
7981 // We need to replace ptr0 in the following expression:
7982 // x0 * offset0 + y0 * ptr0 = t0
7984 // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store)
7986 // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the
7987 // indexed load/store and the expresion that needs to be re-written.
7989 // Therefore, we have:
7990 // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1
7992 ConstantSDNode *CN =
7993 cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx));
7995 APInt Offset0 = CN->getAPIntValue();
7996 APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue();
7998 X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1;
7999 Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1;
8000 X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1;
8001 Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1;
8003 unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD;
8005 APInt CNV = Offset0;
8006 if (X0 < 0) CNV = -CNV;
8007 if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1;
8008 else CNV = CNV - Offset1;
8010 // We can now generate the new expression.
8011 SDValue NewOp1 = DAG.getConstant(CNV, CN->getValueType(0));
8012 SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0);
8014 SDValue NewUse = DAG.getNode(Opcode,
8015 SDLoc(OtherUses[i]),
8016 OtherUses[i]->getValueType(0), NewOp1, NewOp2);
8017 DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse);
8018 deleteAndRecombine(OtherUses[i]);
8021 // Replace the uses of Ptr with uses of the updated base value.
8022 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
8023 deleteAndRecombine(Ptr.getNode());
8028 /// Try to combine a load/store with a add/sub of the base pointer node into a
8029 /// post-indexed load/store. The transformation folded the add/subtract into the
8030 /// new indexed load/store effectively and all of its uses are redirected to the
8032 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
8033 if (Level < AfterLegalizeDAG)
8039 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8040 if (LD->isIndexed())
8042 VT = LD->getMemoryVT();
8043 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
8044 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
8046 Ptr = LD->getBasePtr();
8047 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8048 if (ST->isIndexed())
8050 VT = ST->getMemoryVT();
8051 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
8052 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
8054 Ptr = ST->getBasePtr();
8060 if (Ptr.getNode()->hasOneUse())
8063 for (SDNode *Op : Ptr.getNode()->uses()) {
8065 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
8070 ISD::MemIndexedMode AM = ISD::UNINDEXED;
8071 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
8072 // Don't create a indexed load / store with zero offset.
8073 if (isa<ConstantSDNode>(Offset) &&
8074 cast<ConstantSDNode>(Offset)->isNullValue())
8077 // Try turning it into a post-indexed load / store except when
8078 // 1) All uses are load / store ops that use it as base ptr (and
8079 // it may be folded as addressing mmode).
8080 // 2) Op must be independent of N, i.e. Op is neither a predecessor
8081 // nor a successor of N. Otherwise, if Op is folded that would
8084 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
8088 bool TryNext = false;
8089 for (SDNode *Use : BasePtr.getNode()->uses()) {
8090 if (Use == Ptr.getNode())
8093 // If all the uses are load / store addresses, then don't do the
8095 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
8096 bool RealUse = false;
8097 for (SDNode *UseUse : Use->uses()) {
8098 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
8113 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
8114 SDValue Result = isLoad
8115 ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
8116 BasePtr, Offset, AM)
8117 : DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
8118 BasePtr, Offset, AM);
8121 DEBUG(dbgs() << "\nReplacing.5 ";
8123 dbgs() << "\nWith: ";
8124 Result.getNode()->dump(&DAG);
8126 WorklistRemover DeadNodes(*this);
8128 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
8129 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
8131 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
8134 // Finally, since the node is now dead, remove it from the graph.
8135 deleteAndRecombine(N);
8137 // Replace the uses of Use with uses of the updated base value.
8138 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
8139 Result.getValue(isLoad ? 1 : 0));
8140 deleteAndRecombine(Op);
8149 /// \brief Return the base-pointer arithmetic from an indexed \p LD.
8150 SDValue DAGCombiner::SplitIndexingFromLoad(LoadSDNode *LD) {
8151 ISD::MemIndexedMode AM = LD->getAddressingMode();
8152 assert(AM != ISD::UNINDEXED);
8153 SDValue BP = LD->getOperand(1);
8154 SDValue Inc = LD->getOperand(2);
8156 // Some backends use TargetConstants for load offsets, but don't expect
8157 // TargetConstants in general ADD nodes. We can convert these constants into
8158 // regular Constants (if the constant is not opaque).
8159 assert((Inc.getOpcode() != ISD::TargetConstant ||
8160 !cast<ConstantSDNode>(Inc)->isOpaque()) &&
8161 "Cannot split out indexing using opaque target constants");
8162 if (Inc.getOpcode() == ISD::TargetConstant) {
8163 ConstantSDNode *ConstInc = cast<ConstantSDNode>(Inc);
8164 Inc = DAG.getConstant(*ConstInc->getConstantIntValue(),
8165 ConstInc->getValueType(0));
8169 (AM == ISD::PRE_INC || AM == ISD::POST_INC ? ISD::ADD : ISD::SUB);
8170 return DAG.getNode(Opc, SDLoc(LD), BP.getSimpleValueType(), BP, Inc);
8173 SDValue DAGCombiner::visitLOAD(SDNode *N) {
8174 LoadSDNode *LD = cast<LoadSDNode>(N);
8175 SDValue Chain = LD->getChain();
8176 SDValue Ptr = LD->getBasePtr();
8178 // If load is not volatile and there are no uses of the loaded value (and
8179 // the updated indexed value in case of indexed loads), change uses of the
8180 // chain value into uses of the chain input (i.e. delete the dead load).
8181 if (!LD->isVolatile()) {
8182 if (N->getValueType(1) == MVT::Other) {
8184 if (!N->hasAnyUseOfValue(0)) {
8185 // It's not safe to use the two value CombineTo variant here. e.g.
8186 // v1, chain2 = load chain1, loc
8187 // v2, chain3 = load chain2, loc
8189 // Now we replace use of chain2 with chain1. This makes the second load
8190 // isomorphic to the one we are deleting, and thus makes this load live.
8191 DEBUG(dbgs() << "\nReplacing.6 ";
8193 dbgs() << "\nWith chain: ";
8194 Chain.getNode()->dump(&DAG);
8196 WorklistRemover DeadNodes(*this);
8197 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
8200 deleteAndRecombine(N);
8202 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8206 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
8208 // If this load has an opaque TargetConstant offset, then we cannot split
8209 // the indexing into an add/sub directly (that TargetConstant may not be
8210 // valid for a different type of node, and we cannot convert an opaque
8211 // target constant into a regular constant).
8212 bool HasOTCInc = LD->getOperand(2).getOpcode() == ISD::TargetConstant &&
8213 cast<ConstantSDNode>(LD->getOperand(2))->isOpaque();
8215 if (!N->hasAnyUseOfValue(0) &&
8216 ((MaySplitLoadIndex && !HasOTCInc) || !N->hasAnyUseOfValue(1))) {
8217 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
8219 if (N->hasAnyUseOfValue(1) && MaySplitLoadIndex && !HasOTCInc) {
8220 Index = SplitIndexingFromLoad(LD);
8221 // Try to fold the base pointer arithmetic into subsequent loads and
8223 AddUsersToWorklist(N);
8225 Index = DAG.getUNDEF(N->getValueType(1));
8226 DEBUG(dbgs() << "\nReplacing.7 ";
8228 dbgs() << "\nWith: ";
8229 Undef.getNode()->dump(&DAG);
8230 dbgs() << " and 2 other values\n");
8231 WorklistRemover DeadNodes(*this);
8232 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
8233 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Index);
8234 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
8235 deleteAndRecombine(N);
8236 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8241 // If this load is directly stored, replace the load value with the stored
8243 // TODO: Handle store large -> read small portion.
8244 // TODO: Handle TRUNCSTORE/LOADEXT
8245 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
8246 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
8247 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
8248 if (PrevST->getBasePtr() == Ptr &&
8249 PrevST->getValue().getValueType() == N->getValueType(0))
8250 return CombineTo(N, Chain.getOperand(1), Chain);
8254 // Try to infer better alignment information than the load already has.
8255 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
8256 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
8257 if (Align > LD->getMemOperand()->getBaseAlignment()) {
8259 DAG.getExtLoad(LD->getExtensionType(), SDLoc(N),
8260 LD->getValueType(0),
8261 Chain, Ptr, LD->getPointerInfo(),
8263 LD->isVolatile(), LD->isNonTemporal(),
8264 LD->isInvariant(), Align, LD->getAAInfo());
8265 return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true);
8270 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
8271 : DAG.getSubtarget().useAA();
8273 if (CombinerAAOnlyFunc.getNumOccurrences() &&
8274 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
8277 if (UseAA && LD->isUnindexed()) {
8278 // Walk up chain skipping non-aliasing memory nodes.
8279 SDValue BetterChain = FindBetterChain(N, Chain);
8281 // If there is a better chain.
8282 if (Chain != BetterChain) {
8285 // Replace the chain to void dependency.
8286 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
8287 ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD),
8288 BetterChain, Ptr, LD->getMemOperand());
8290 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD),
8291 LD->getValueType(0),
8292 BetterChain, Ptr, LD->getMemoryVT(),
8293 LD->getMemOperand());
8296 // Create token factor to keep old chain connected.
8297 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
8298 MVT::Other, Chain, ReplLoad.getValue(1));
8300 // Make sure the new and old chains are cleaned up.
8301 AddToWorklist(Token.getNode());
8303 // Replace uses with load result and token factor. Don't add users
8305 return CombineTo(N, ReplLoad.getValue(0), Token, false);
8309 // Try transforming N to an indexed load.
8310 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
8311 return SDValue(N, 0);
8313 // Try to slice up N to more direct loads if the slices are mapped to
8314 // different register banks or pairing can take place.
8316 return SDValue(N, 0);
8322 /// \brief Helper structure used to slice a load in smaller loads.
8323 /// Basically a slice is obtained from the following sequence:
8324 /// Origin = load Ty1, Base
8325 /// Shift = srl Ty1 Origin, CstTy Amount
8326 /// Inst = trunc Shift to Ty2
8328 /// Then, it will be rewriten into:
8329 /// Slice = load SliceTy, Base + SliceOffset
8330 /// [Inst = zext Slice to Ty2], only if SliceTy <> Ty2
8332 /// SliceTy is deduced from the number of bits that are actually used to
8334 struct LoadedSlice {
8335 /// \brief Helper structure used to compute the cost of a slice.
8337 /// Are we optimizing for code size.
8342 unsigned CrossRegisterBanksCopies;
8346 Cost(bool ForCodeSize = false)
8347 : ForCodeSize(ForCodeSize), Loads(0), Truncates(0),
8348 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {}
8350 /// \brief Get the cost of one isolated slice.
8351 Cost(const LoadedSlice &LS, bool ForCodeSize = false)
8352 : ForCodeSize(ForCodeSize), Loads(1), Truncates(0),
8353 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {
8354 EVT TruncType = LS.Inst->getValueType(0);
8355 EVT LoadedType = LS.getLoadedType();
8356 if (TruncType != LoadedType &&
8357 !LS.DAG->getTargetLoweringInfo().isZExtFree(LoadedType, TruncType))
8361 /// \brief Account for slicing gain in the current cost.
8362 /// Slicing provide a few gains like removing a shift or a
8363 /// truncate. This method allows to grow the cost of the original
8364 /// load with the gain from this slice.
8365 void addSliceGain(const LoadedSlice &LS) {
8366 // Each slice saves a truncate.
8367 const TargetLowering &TLI = LS.DAG->getTargetLoweringInfo();
8368 if (!TLI.isTruncateFree(LS.Inst->getValueType(0),
8369 LS.Inst->getOperand(0).getValueType()))
8371 // If there is a shift amount, this slice gets rid of it.
8374 // If this slice can merge a cross register bank copy, account for it.
8375 if (LS.canMergeExpensiveCrossRegisterBankCopy())
8376 ++CrossRegisterBanksCopies;
8379 Cost &operator+=(const Cost &RHS) {
8381 Truncates += RHS.Truncates;
8382 CrossRegisterBanksCopies += RHS.CrossRegisterBanksCopies;
8388 bool operator==(const Cost &RHS) const {
8389 return Loads == RHS.Loads && Truncates == RHS.Truncates &&
8390 CrossRegisterBanksCopies == RHS.CrossRegisterBanksCopies &&
8391 ZExts == RHS.ZExts && Shift == RHS.Shift;
8394 bool operator!=(const Cost &RHS) const { return !(*this == RHS); }
8396 bool operator<(const Cost &RHS) const {
8397 // Assume cross register banks copies are as expensive as loads.
8398 // FIXME: Do we want some more target hooks?
8399 unsigned ExpensiveOpsLHS = Loads + CrossRegisterBanksCopies;
8400 unsigned ExpensiveOpsRHS = RHS.Loads + RHS.CrossRegisterBanksCopies;
8401 // Unless we are optimizing for code size, consider the
8402 // expensive operation first.
8403 if (!ForCodeSize && ExpensiveOpsLHS != ExpensiveOpsRHS)
8404 return ExpensiveOpsLHS < ExpensiveOpsRHS;
8405 return (Truncates + ZExts + Shift + ExpensiveOpsLHS) <
8406 (RHS.Truncates + RHS.ZExts + RHS.Shift + ExpensiveOpsRHS);
8409 bool operator>(const Cost &RHS) const { return RHS < *this; }
8411 bool operator<=(const Cost &RHS) const { return !(RHS < *this); }
8413 bool operator>=(const Cost &RHS) const { return !(*this < RHS); }
8415 // The last instruction that represent the slice. This should be a
8416 // truncate instruction.
8418 // The original load instruction.
8420 // The right shift amount in bits from the original load.
8422 // The DAG from which Origin came from.
8423 // This is used to get some contextual information about legal types, etc.
8426 LoadedSlice(SDNode *Inst = nullptr, LoadSDNode *Origin = nullptr,
8427 unsigned Shift = 0, SelectionDAG *DAG = nullptr)
8428 : Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {}
8430 LoadedSlice(const LoadedSlice &LS)
8431 : Inst(LS.Inst), Origin(LS.Origin), Shift(LS.Shift), DAG(LS.DAG) {}
8433 /// \brief Get the bits used in a chunk of bits \p BitWidth large.
8434 /// \return Result is \p BitWidth and has used bits set to 1 and
8435 /// not used bits set to 0.
8436 APInt getUsedBits() const {
8437 // Reproduce the trunc(lshr) sequence:
8438 // - Start from the truncated value.
8439 // - Zero extend to the desired bit width.
8441 assert(Origin && "No original load to compare against.");
8442 unsigned BitWidth = Origin->getValueSizeInBits(0);
8443 assert(Inst && "This slice is not bound to an instruction");
8444 assert(Inst->getValueSizeInBits(0) <= BitWidth &&
8445 "Extracted slice is bigger than the whole type!");
8446 APInt UsedBits(Inst->getValueSizeInBits(0), 0);
8447 UsedBits.setAllBits();
8448 UsedBits = UsedBits.zext(BitWidth);
8453 /// \brief Get the size of the slice to be loaded in bytes.
8454 unsigned getLoadedSize() const {
8455 unsigned SliceSize = getUsedBits().countPopulation();
8456 assert(!(SliceSize & 0x7) && "Size is not a multiple of a byte.");
8457 return SliceSize / 8;
8460 /// \brief Get the type that will be loaded for this slice.
8461 /// Note: This may not be the final type for the slice.
8462 EVT getLoadedType() const {
8463 assert(DAG && "Missing context");
8464 LLVMContext &Ctxt = *DAG->getContext();
8465 return EVT::getIntegerVT(Ctxt, getLoadedSize() * 8);
8468 /// \brief Get the alignment of the load used for this slice.
8469 unsigned getAlignment() const {
8470 unsigned Alignment = Origin->getAlignment();
8471 unsigned Offset = getOffsetFromBase();
8473 Alignment = MinAlign(Alignment, Alignment + Offset);
8477 /// \brief Check if this slice can be rewritten with legal operations.
8478 bool isLegal() const {
8479 // An invalid slice is not legal.
8480 if (!Origin || !Inst || !DAG)
8483 // Offsets are for indexed load only, we do not handle that.
8484 if (Origin->getOffset().getOpcode() != ISD::UNDEF)
8487 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
8489 // Check that the type is legal.
8490 EVT SliceType = getLoadedType();
8491 if (!TLI.isTypeLegal(SliceType))
8494 // Check that the load is legal for this type.
8495 if (!TLI.isOperationLegal(ISD::LOAD, SliceType))
8498 // Check that the offset can be computed.
8499 // 1. Check its type.
8500 EVT PtrType = Origin->getBasePtr().getValueType();
8501 if (PtrType == MVT::Untyped || PtrType.isExtended())
8504 // 2. Check that it fits in the immediate.
8505 if (!TLI.isLegalAddImmediate(getOffsetFromBase()))
8508 // 3. Check that the computation is legal.
8509 if (!TLI.isOperationLegal(ISD::ADD, PtrType))
8512 // Check that the zext is legal if it needs one.
8513 EVT TruncateType = Inst->getValueType(0);
8514 if (TruncateType != SliceType &&
8515 !TLI.isOperationLegal(ISD::ZERO_EXTEND, TruncateType))
8521 /// \brief Get the offset in bytes of this slice in the original chunk of
8523 /// \pre DAG != nullptr.
8524 uint64_t getOffsetFromBase() const {
8525 assert(DAG && "Missing context.");
8527 DAG->getTargetLoweringInfo().getDataLayout()->isBigEndian();
8528 assert(!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported.");
8529 uint64_t Offset = Shift / 8;
8530 unsigned TySizeInBytes = Origin->getValueSizeInBits(0) / 8;
8531 assert(!(Origin->getValueSizeInBits(0) & 0x7) &&
8532 "The size of the original loaded type is not a multiple of a"
8534 // If Offset is bigger than TySizeInBytes, it means we are loading all
8535 // zeros. This should have been optimized before in the process.
8536 assert(TySizeInBytes > Offset &&
8537 "Invalid shift amount for given loaded size");
8539 Offset = TySizeInBytes - Offset - getLoadedSize();
8543 /// \brief Generate the sequence of instructions to load the slice
8544 /// represented by this object and redirect the uses of this slice to
8545 /// this new sequence of instructions.
8546 /// \pre this->Inst && this->Origin are valid Instructions and this
8547 /// object passed the legal check: LoadedSlice::isLegal returned true.
8548 /// \return The last instruction of the sequence used to load the slice.
8549 SDValue loadSlice() const {
8550 assert(Inst && Origin && "Unable to replace a non-existing slice.");
8551 const SDValue &OldBaseAddr = Origin->getBasePtr();
8552 SDValue BaseAddr = OldBaseAddr;
8553 // Get the offset in that chunk of bytes w.r.t. the endianess.
8554 int64_t Offset = static_cast<int64_t>(getOffsetFromBase());
8555 assert(Offset >= 0 && "Offset too big to fit in int64_t!");
8557 // BaseAddr = BaseAddr + Offset.
8558 EVT ArithType = BaseAddr.getValueType();
8559 BaseAddr = DAG->getNode(ISD::ADD, SDLoc(Origin), ArithType, BaseAddr,
8560 DAG->getConstant(Offset, ArithType));
8563 // Create the type of the loaded slice according to its size.
8564 EVT SliceType = getLoadedType();
8566 // Create the load for the slice.
8567 SDValue LastInst = DAG->getLoad(
8568 SliceType, SDLoc(Origin), Origin->getChain(), BaseAddr,
8569 Origin->getPointerInfo().getWithOffset(Offset), Origin->isVolatile(),
8570 Origin->isNonTemporal(), Origin->isInvariant(), getAlignment());
8571 // If the final type is not the same as the loaded type, this means that
8572 // we have to pad with zero. Create a zero extend for that.
8573 EVT FinalType = Inst->getValueType(0);
8574 if (SliceType != FinalType)
8576 DAG->getNode(ISD::ZERO_EXTEND, SDLoc(LastInst), FinalType, LastInst);
8580 /// \brief Check if this slice can be merged with an expensive cross register
8581 /// bank copy. E.g.,
8583 /// f = bitcast i32 i to float
8584 bool canMergeExpensiveCrossRegisterBankCopy() const {
8585 if (!Inst || !Inst->hasOneUse())
8587 SDNode *Use = *Inst->use_begin();
8588 if (Use->getOpcode() != ISD::BITCAST)
8590 assert(DAG && "Missing context");
8591 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
8592 EVT ResVT = Use->getValueType(0);
8593 const TargetRegisterClass *ResRC = TLI.getRegClassFor(ResVT.getSimpleVT());
8594 const TargetRegisterClass *ArgRC =
8595 TLI.getRegClassFor(Use->getOperand(0).getValueType().getSimpleVT());
8596 if (ArgRC == ResRC || !TLI.isOperationLegal(ISD::LOAD, ResVT))
8599 // At this point, we know that we perform a cross-register-bank copy.
8600 // Check if it is expensive.
8601 const TargetRegisterInfo *TRI = DAG->getSubtarget().getRegisterInfo();
8602 // Assume bitcasts are cheap, unless both register classes do not
8603 // explicitly share a common sub class.
8604 if (!TRI || TRI->getCommonSubClass(ArgRC, ResRC))
8607 // Check if it will be merged with the load.
8608 // 1. Check the alignment constraint.
8609 unsigned RequiredAlignment = TLI.getDataLayout()->getABITypeAlignment(
8610 ResVT.getTypeForEVT(*DAG->getContext()));
8612 if (RequiredAlignment > getAlignment())
8615 // 2. Check that the load is a legal operation for that type.
8616 if (!TLI.isOperationLegal(ISD::LOAD, ResVT))
8619 // 3. Check that we do not have a zext in the way.
8620 if (Inst->getValueType(0) != getLoadedType())
8628 /// \brief Check that all bits set in \p UsedBits form a dense region, i.e.,
8629 /// \p UsedBits looks like 0..0 1..1 0..0.
8630 static bool areUsedBitsDense(const APInt &UsedBits) {
8631 // If all the bits are one, this is dense!
8632 if (UsedBits.isAllOnesValue())
8635 // Get rid of the unused bits on the right.
8636 APInt NarrowedUsedBits = UsedBits.lshr(UsedBits.countTrailingZeros());
8637 // Get rid of the unused bits on the left.
8638 if (NarrowedUsedBits.countLeadingZeros())
8639 NarrowedUsedBits = NarrowedUsedBits.trunc(NarrowedUsedBits.getActiveBits());
8640 // Check that the chunk of bits is completely used.
8641 return NarrowedUsedBits.isAllOnesValue();
8644 /// \brief Check whether or not \p First and \p Second are next to each other
8645 /// in memory. This means that there is no hole between the bits loaded
8646 /// by \p First and the bits loaded by \p Second.
8647 static bool areSlicesNextToEachOther(const LoadedSlice &First,
8648 const LoadedSlice &Second) {
8649 assert(First.Origin == Second.Origin && First.Origin &&
8650 "Unable to match different memory origins.");
8651 APInt UsedBits = First.getUsedBits();
8652 assert((UsedBits & Second.getUsedBits()) == 0 &&
8653 "Slices are not supposed to overlap.");
8654 UsedBits |= Second.getUsedBits();
8655 return areUsedBitsDense(UsedBits);
8658 /// \brief Adjust the \p GlobalLSCost according to the target
8659 /// paring capabilities and the layout of the slices.
8660 /// \pre \p GlobalLSCost should account for at least as many loads as
8661 /// there is in the slices in \p LoadedSlices.
8662 static void adjustCostForPairing(SmallVectorImpl<LoadedSlice> &LoadedSlices,
8663 LoadedSlice::Cost &GlobalLSCost) {
8664 unsigned NumberOfSlices = LoadedSlices.size();
8665 // If there is less than 2 elements, no pairing is possible.
8666 if (NumberOfSlices < 2)
8669 // Sort the slices so that elements that are likely to be next to each
8670 // other in memory are next to each other in the list.
8671 std::sort(LoadedSlices.begin(), LoadedSlices.end(),
8672 [](const LoadedSlice &LHS, const LoadedSlice &RHS) {
8673 assert(LHS.Origin == RHS.Origin && "Different bases not implemented.");
8674 return LHS.getOffsetFromBase() < RHS.getOffsetFromBase();
8676 const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo();
8677 // First (resp. Second) is the first (resp. Second) potentially candidate
8678 // to be placed in a paired load.
8679 const LoadedSlice *First = nullptr;
8680 const LoadedSlice *Second = nullptr;
8681 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice,
8682 // Set the beginning of the pair.
8685 Second = &LoadedSlices[CurrSlice];
8687 // If First is NULL, it means we start a new pair.
8688 // Get to the next slice.
8692 EVT LoadedType = First->getLoadedType();
8694 // If the types of the slices are different, we cannot pair them.
8695 if (LoadedType != Second->getLoadedType())
8698 // Check if the target supplies paired loads for this type.
8699 unsigned RequiredAlignment = 0;
8700 if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) {
8701 // move to the next pair, this type is hopeless.
8705 // Check if we meet the alignment requirement.
8706 if (RequiredAlignment > First->getAlignment())
8709 // Check that both loads are next to each other in memory.
8710 if (!areSlicesNextToEachOther(*First, *Second))
8713 assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!");
8714 --GlobalLSCost.Loads;
8715 // Move to the next pair.
8720 /// \brief Check the profitability of all involved LoadedSlice.
8721 /// Currently, it is considered profitable if there is exactly two
8722 /// involved slices (1) which are (2) next to each other in memory, and
8723 /// whose cost (\see LoadedSlice::Cost) is smaller than the original load (3).
8725 /// Note: The order of the elements in \p LoadedSlices may be modified, but not
8726 /// the elements themselves.
8728 /// FIXME: When the cost model will be mature enough, we can relax
8729 /// constraints (1) and (2).
8730 static bool isSlicingProfitable(SmallVectorImpl<LoadedSlice> &LoadedSlices,
8731 const APInt &UsedBits, bool ForCodeSize) {
8732 unsigned NumberOfSlices = LoadedSlices.size();
8733 if (StressLoadSlicing)
8734 return NumberOfSlices > 1;
8737 if (NumberOfSlices != 2)
8741 if (!areUsedBitsDense(UsedBits))
8745 LoadedSlice::Cost OrigCost(ForCodeSize), GlobalSlicingCost(ForCodeSize);
8746 // The original code has one big load.
8748 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice) {
8749 const LoadedSlice &LS = LoadedSlices[CurrSlice];
8750 // Accumulate the cost of all the slices.
8751 LoadedSlice::Cost SliceCost(LS, ForCodeSize);
8752 GlobalSlicingCost += SliceCost;
8754 // Account as cost in the original configuration the gain obtained
8755 // with the current slices.
8756 OrigCost.addSliceGain(LS);
8759 // If the target supports paired load, adjust the cost accordingly.
8760 adjustCostForPairing(LoadedSlices, GlobalSlicingCost);
8761 return OrigCost > GlobalSlicingCost;
8764 /// \brief If the given load, \p LI, is used only by trunc or trunc(lshr)
8765 /// operations, split it in the various pieces being extracted.
8767 /// This sort of thing is introduced by SROA.
8768 /// This slicing takes care not to insert overlapping loads.
8769 /// \pre LI is a simple load (i.e., not an atomic or volatile load).
8770 bool DAGCombiner::SliceUpLoad(SDNode *N) {
8771 if (Level < AfterLegalizeDAG)
8774 LoadSDNode *LD = cast<LoadSDNode>(N);
8775 if (LD->isVolatile() || !ISD::isNormalLoad(LD) ||
8776 !LD->getValueType(0).isInteger())
8779 // Keep track of already used bits to detect overlapping values.
8780 // In that case, we will just abort the transformation.
8781 APInt UsedBits(LD->getValueSizeInBits(0), 0);
8783 SmallVector<LoadedSlice, 4> LoadedSlices;
8785 // Check if this load is used as several smaller chunks of bits.
8786 // Basically, look for uses in trunc or trunc(lshr) and record a new chain
8787 // of computation for each trunc.
8788 for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end();
8789 UI != UIEnd; ++UI) {
8790 // Skip the uses of the chain.
8791 if (UI.getUse().getResNo() != 0)
8797 // Check if this is a trunc(lshr).
8798 if (User->getOpcode() == ISD::SRL && User->hasOneUse() &&
8799 isa<ConstantSDNode>(User->getOperand(1))) {
8800 Shift = cast<ConstantSDNode>(User->getOperand(1))->getZExtValue();
8801 User = *User->use_begin();
8804 // At this point, User is a Truncate, iff we encountered, trunc or
8806 if (User->getOpcode() != ISD::TRUNCATE)
8809 // The width of the type must be a power of 2 and greater than 8-bits.
8810 // Otherwise the load cannot be represented in LLVM IR.
8811 // Moreover, if we shifted with a non-8-bits multiple, the slice
8812 // will be across several bytes. We do not support that.
8813 unsigned Width = User->getValueSizeInBits(0);
8814 if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7))
8817 // Build the slice for this chain of computations.
8818 LoadedSlice LS(User, LD, Shift, &DAG);
8819 APInt CurrentUsedBits = LS.getUsedBits();
8821 // Check if this slice overlaps with another.
8822 if ((CurrentUsedBits & UsedBits) != 0)
8824 // Update the bits used globally.
8825 UsedBits |= CurrentUsedBits;
8827 // Check if the new slice would be legal.
8831 // Record the slice.
8832 LoadedSlices.push_back(LS);
8835 // Abort slicing if it does not seem to be profitable.
8836 if (!isSlicingProfitable(LoadedSlices, UsedBits, ForCodeSize))
8841 // Rewrite each chain to use an independent load.
8842 // By construction, each chain can be represented by a unique load.
8844 // Prepare the argument for the new token factor for all the slices.
8845 SmallVector<SDValue, 8> ArgChains;
8846 for (SmallVectorImpl<LoadedSlice>::const_iterator
8847 LSIt = LoadedSlices.begin(),
8848 LSItEnd = LoadedSlices.end();
8849 LSIt != LSItEnd; ++LSIt) {
8850 SDValue SliceInst = LSIt->loadSlice();
8851 CombineTo(LSIt->Inst, SliceInst, true);
8852 if (SliceInst.getNode()->getOpcode() != ISD::LOAD)
8853 SliceInst = SliceInst.getOperand(0);
8854 assert(SliceInst->getOpcode() == ISD::LOAD &&
8855 "It takes more than a zext to get to the loaded slice!!");
8856 ArgChains.push_back(SliceInst.getValue(1));
8859 SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other,
8861 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
8865 /// Check to see if V is (and load (ptr), imm), where the load is having
8866 /// specific bytes cleared out. If so, return the byte size being masked out
8867 /// and the shift amount.
8868 static std::pair<unsigned, unsigned>
8869 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
8870 std::pair<unsigned, unsigned> Result(0, 0);
8872 // Check for the structure we're looking for.
8873 if (V->getOpcode() != ISD::AND ||
8874 !isa<ConstantSDNode>(V->getOperand(1)) ||
8875 !ISD::isNormalLoad(V->getOperand(0).getNode()))
8878 // Check the chain and pointer.
8879 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
8880 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
8882 // The store should be chained directly to the load or be an operand of a
8884 if (LD == Chain.getNode())
8886 else if (Chain->getOpcode() != ISD::TokenFactor)
8887 return Result; // Fail.
8890 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
8891 if (Chain->getOperand(i).getNode() == LD) {
8895 if (!isOk) return Result;
8898 // This only handles simple types.
8899 if (V.getValueType() != MVT::i16 &&
8900 V.getValueType() != MVT::i32 &&
8901 V.getValueType() != MVT::i64)
8904 // Check the constant mask. Invert it so that the bits being masked out are
8905 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
8906 // follow the sign bit for uniformity.
8907 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
8908 unsigned NotMaskLZ = countLeadingZeros(NotMask);
8909 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
8910 unsigned NotMaskTZ = countTrailingZeros(NotMask);
8911 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
8912 if (NotMaskLZ == 64) return Result; // All zero mask.
8914 // See if we have a continuous run of bits. If so, we have 0*1+0*
8915 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
8918 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
8919 if (V.getValueType() != MVT::i64 && NotMaskLZ)
8920 NotMaskLZ -= 64-V.getValueSizeInBits();
8922 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
8923 switch (MaskedBytes) {
8927 default: return Result; // All one mask, or 5-byte mask.
8930 // Verify that the first bit starts at a multiple of mask so that the access
8931 // is aligned the same as the access width.
8932 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
8934 Result.first = MaskedBytes;
8935 Result.second = NotMaskTZ/8;
8940 /// Check to see if IVal is something that provides a value as specified by
8941 /// MaskInfo. If so, replace the specified store with a narrower store of
8944 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
8945 SDValue IVal, StoreSDNode *St,
8947 unsigned NumBytes = MaskInfo.first;
8948 unsigned ByteShift = MaskInfo.second;
8949 SelectionDAG &DAG = DC->getDAG();
8951 // Check to see if IVal is all zeros in the part being masked in by the 'or'
8952 // that uses this. If not, this is not a replacement.
8953 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
8954 ByteShift*8, (ByteShift+NumBytes)*8);
8955 if (!DAG.MaskedValueIsZero(IVal, Mask)) return nullptr;
8957 // Check that it is legal on the target to do this. It is legal if the new
8958 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
8960 MVT VT = MVT::getIntegerVT(NumBytes*8);
8961 if (!DC->isTypeLegal(VT))
8964 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
8965 // shifted by ByteShift and truncated down to NumBytes.
8967 IVal = DAG.getNode(ISD::SRL, SDLoc(IVal), IVal.getValueType(), IVal,
8968 DAG.getConstant(ByteShift*8,
8969 DC->getShiftAmountTy(IVal.getValueType())));
8971 // Figure out the offset for the store and the alignment of the access.
8973 unsigned NewAlign = St->getAlignment();
8975 if (DAG.getTargetLoweringInfo().isLittleEndian())
8976 StOffset = ByteShift;
8978 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
8980 SDValue Ptr = St->getBasePtr();
8982 Ptr = DAG.getNode(ISD::ADD, SDLoc(IVal), Ptr.getValueType(),
8983 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
8984 NewAlign = MinAlign(NewAlign, StOffset);
8987 // Truncate down to the new size.
8988 IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal);
8991 return DAG.getStore(St->getChain(), SDLoc(St), IVal, Ptr,
8992 St->getPointerInfo().getWithOffset(StOffset),
8993 false, false, NewAlign).getNode();
8997 /// Look for sequence of load / op / store where op is one of 'or', 'xor', and
8998 /// 'and' of immediates. If 'op' is only touching some of the loaded bits, try
8999 /// narrowing the load and store if it would end up being a win for performance
9001 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
9002 StoreSDNode *ST = cast<StoreSDNode>(N);
9003 if (ST->isVolatile())
9006 SDValue Chain = ST->getChain();
9007 SDValue Value = ST->getValue();
9008 SDValue Ptr = ST->getBasePtr();
9009 EVT VT = Value.getValueType();
9011 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
9014 unsigned Opc = Value.getOpcode();
9016 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
9017 // is a byte mask indicating a consecutive number of bytes, check to see if
9018 // Y is known to provide just those bytes. If so, we try to replace the
9019 // load + replace + store sequence with a single (narrower) store, which makes
9021 if (Opc == ISD::OR) {
9022 std::pair<unsigned, unsigned> MaskedLoad;
9023 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
9024 if (MaskedLoad.first)
9025 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
9026 Value.getOperand(1), ST,this))
9027 return SDValue(NewST, 0);
9029 // Or is commutative, so try swapping X and Y.
9030 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
9031 if (MaskedLoad.first)
9032 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
9033 Value.getOperand(0), ST,this))
9034 return SDValue(NewST, 0);
9037 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
9038 Value.getOperand(1).getOpcode() != ISD::Constant)
9041 SDValue N0 = Value.getOperand(0);
9042 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
9043 Chain == SDValue(N0.getNode(), 1)) {
9044 LoadSDNode *LD = cast<LoadSDNode>(N0);
9045 if (LD->getBasePtr() != Ptr ||
9046 LD->getPointerInfo().getAddrSpace() !=
9047 ST->getPointerInfo().getAddrSpace())
9050 // Find the type to narrow it the load / op / store to.
9051 SDValue N1 = Value.getOperand(1);
9052 unsigned BitWidth = N1.getValueSizeInBits();
9053 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
9054 if (Opc == ISD::AND)
9055 Imm ^= APInt::getAllOnesValue(BitWidth);
9056 if (Imm == 0 || Imm.isAllOnesValue())
9058 unsigned ShAmt = Imm.countTrailingZeros();
9059 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
9060 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
9061 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
9062 while (NewBW < BitWidth &&
9063 !(TLI.isOperationLegalOrCustom(Opc, NewVT) &&
9064 TLI.isNarrowingProfitable(VT, NewVT))) {
9065 NewBW = NextPowerOf2(NewBW);
9066 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
9068 if (NewBW >= BitWidth)
9071 // If the lsb changed does not start at the type bitwidth boundary,
9072 // start at the previous one.
9074 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
9075 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt,
9076 std::min(BitWidth, ShAmt + NewBW));
9077 if ((Imm & Mask) == Imm) {
9078 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
9079 if (Opc == ISD::AND)
9080 NewImm ^= APInt::getAllOnesValue(NewBW);
9081 uint64_t PtrOff = ShAmt / 8;
9082 // For big endian targets, we need to adjust the offset to the pointer to
9083 // load the correct bytes.
9084 if (TLI.isBigEndian())
9085 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
9087 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
9088 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
9089 if (NewAlign < TLI.getDataLayout()->getABITypeAlignment(NewVTTy))
9092 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD),
9093 Ptr.getValueType(), Ptr,
9094 DAG.getConstant(PtrOff, Ptr.getValueType()));
9095 SDValue NewLD = DAG.getLoad(NewVT, SDLoc(N0),
9096 LD->getChain(), NewPtr,
9097 LD->getPointerInfo().getWithOffset(PtrOff),
9098 LD->isVolatile(), LD->isNonTemporal(),
9099 LD->isInvariant(), NewAlign,
9101 SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD,
9102 DAG.getConstant(NewImm, NewVT));
9103 SDValue NewST = DAG.getStore(Chain, SDLoc(N),
9105 ST->getPointerInfo().getWithOffset(PtrOff),
9106 false, false, NewAlign);
9108 AddToWorklist(NewPtr.getNode());
9109 AddToWorklist(NewLD.getNode());
9110 AddToWorklist(NewVal.getNode());
9111 WorklistRemover DeadNodes(*this);
9112 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
9121 /// For a given floating point load / store pair, if the load value isn't used
9122 /// by any other operations, then consider transforming the pair to integer
9123 /// load / store operations if the target deems the transformation profitable.
9124 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
9125 StoreSDNode *ST = cast<StoreSDNode>(N);
9126 SDValue Chain = ST->getChain();
9127 SDValue Value = ST->getValue();
9128 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
9129 Value.hasOneUse() &&
9130 Chain == SDValue(Value.getNode(), 1)) {
9131 LoadSDNode *LD = cast<LoadSDNode>(Value);
9132 EVT VT = LD->getMemoryVT();
9133 if (!VT.isFloatingPoint() ||
9134 VT != ST->getMemoryVT() ||
9135 LD->isNonTemporal() ||
9136 ST->isNonTemporal() ||
9137 LD->getPointerInfo().getAddrSpace() != 0 ||
9138 ST->getPointerInfo().getAddrSpace() != 0)
9141 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
9142 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
9143 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
9144 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
9145 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
9148 unsigned LDAlign = LD->getAlignment();
9149 unsigned STAlign = ST->getAlignment();
9150 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
9151 unsigned ABIAlign = TLI.getDataLayout()->getABITypeAlignment(IntVTTy);
9152 if (LDAlign < ABIAlign || STAlign < ABIAlign)
9155 SDValue NewLD = DAG.getLoad(IntVT, SDLoc(Value),
9156 LD->getChain(), LD->getBasePtr(),
9157 LD->getPointerInfo(),
9158 false, false, false, LDAlign);
9160 SDValue NewST = DAG.getStore(NewLD.getValue(1), SDLoc(N),
9161 NewLD, ST->getBasePtr(),
9162 ST->getPointerInfo(),
9163 false, false, STAlign);
9165 AddToWorklist(NewLD.getNode());
9166 AddToWorklist(NewST.getNode());
9167 WorklistRemover DeadNodes(*this);
9168 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
9176 /// Helper struct to parse and store a memory address as base + index + offset.
9177 /// We ignore sign extensions when it is safe to do so.
9178 /// The following two expressions are not equivalent. To differentiate we need
9179 /// to store whether there was a sign extension involved in the index
9181 /// (load (i64 add (i64 copyfromreg %c)
9182 /// (i64 signextend (add (i8 load %index)
9186 /// (load (i64 add (i64 copyfromreg %c)
9187 /// (i64 signextend (i32 add (i32 signextend (i8 load %index))
9189 struct BaseIndexOffset {
9193 bool IsIndexSignExt;
9195 BaseIndexOffset() : Offset(0), IsIndexSignExt(false) {}
9197 BaseIndexOffset(SDValue Base, SDValue Index, int64_t Offset,
9198 bool IsIndexSignExt) :
9199 Base(Base), Index(Index), Offset(Offset), IsIndexSignExt(IsIndexSignExt) {}
9201 bool equalBaseIndex(const BaseIndexOffset &Other) {
9202 return Other.Base == Base && Other.Index == Index &&
9203 Other.IsIndexSignExt == IsIndexSignExt;
9206 /// Parses tree in Ptr for base, index, offset addresses.
9207 static BaseIndexOffset match(SDValue Ptr) {
9208 bool IsIndexSignExt = false;
9210 // We only can pattern match BASE + INDEX + OFFSET. If Ptr is not an ADD
9211 // instruction, then it could be just the BASE or everything else we don't
9212 // know how to handle. Just use Ptr as BASE and give up.
9213 if (Ptr->getOpcode() != ISD::ADD)
9214 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
9216 // We know that we have at least an ADD instruction. Try to pattern match
9217 // the simple case of BASE + OFFSET.
9218 if (isa<ConstantSDNode>(Ptr->getOperand(1))) {
9219 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
9220 return BaseIndexOffset(Ptr->getOperand(0), SDValue(), Offset,
9224 // Inside a loop the current BASE pointer is calculated using an ADD and a
9225 // MUL instruction. In this case Ptr is the actual BASE pointer.
9226 // (i64 add (i64 %array_ptr)
9227 // (i64 mul (i64 %induction_var)
9228 // (i64 %element_size)))
9229 if (Ptr->getOperand(1)->getOpcode() == ISD::MUL)
9230 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
9232 // Look at Base + Index + Offset cases.
9233 SDValue Base = Ptr->getOperand(0);
9234 SDValue IndexOffset = Ptr->getOperand(1);
9236 // Skip signextends.
9237 if (IndexOffset->getOpcode() == ISD::SIGN_EXTEND) {
9238 IndexOffset = IndexOffset->getOperand(0);
9239 IsIndexSignExt = true;
9242 // Either the case of Base + Index (no offset) or something else.
9243 if (IndexOffset->getOpcode() != ISD::ADD)
9244 return BaseIndexOffset(Base, IndexOffset, 0, IsIndexSignExt);
9246 // Now we have the case of Base + Index + offset.
9247 SDValue Index = IndexOffset->getOperand(0);
9248 SDValue Offset = IndexOffset->getOperand(1);
9250 if (!isa<ConstantSDNode>(Offset))
9251 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
9253 // Ignore signextends.
9254 if (Index->getOpcode() == ISD::SIGN_EXTEND) {
9255 Index = Index->getOperand(0);
9256 IsIndexSignExt = true;
9257 } else IsIndexSignExt = false;
9259 int64_t Off = cast<ConstantSDNode>(Offset)->getSExtValue();
9260 return BaseIndexOffset(Base, Index, Off, IsIndexSignExt);
9264 /// Holds a pointer to an LSBaseSDNode as well as information on where it
9265 /// is located in a sequence of memory operations connected by a chain.
9267 MemOpLink (LSBaseSDNode *N, int64_t Offset, unsigned Seq):
9268 MemNode(N), OffsetFromBase(Offset), SequenceNum(Seq) { }
9269 // Ptr to the mem node.
9270 LSBaseSDNode *MemNode;
9271 // Offset from the base ptr.
9272 int64_t OffsetFromBase;
9273 // What is the sequence number of this mem node.
9274 // Lowest mem operand in the DAG starts at zero.
9275 unsigned SequenceNum;
9278 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
9279 EVT MemVT = St->getMemoryVT();
9280 int64_t ElementSizeBytes = MemVT.getSizeInBits()/8;
9281 bool NoVectors = DAG.getMachineFunction().getFunction()->getAttributes().
9282 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
9284 // Don't merge vectors into wider inputs.
9285 if (MemVT.isVector() || !MemVT.isSimple())
9288 // Perform an early exit check. Do not bother looking at stored values that
9289 // are not constants or loads.
9290 SDValue StoredVal = St->getValue();
9291 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
9292 if (!isa<ConstantSDNode>(StoredVal) && !isa<ConstantFPSDNode>(StoredVal) &&
9296 // Only look at ends of store sequences.
9297 SDValue Chain = SDValue(St, 0);
9298 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
9301 // This holds the base pointer, index, and the offset in bytes from the base
9303 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr());
9305 // We must have a base and an offset.
9306 if (!BasePtr.Base.getNode())
9309 // Do not handle stores to undef base pointers.
9310 if (BasePtr.Base.getOpcode() == ISD::UNDEF)
9313 // Save the LoadSDNodes that we find in the chain.
9314 // We need to make sure that these nodes do not interfere with
9315 // any of the store nodes.
9316 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
9318 // Save the StoreSDNodes that we find in the chain.
9319 SmallVector<MemOpLink, 8> StoreNodes;
9321 // Walk up the chain and look for nodes with offsets from the same
9322 // base pointer. Stop when reaching an instruction with a different kind
9323 // or instruction which has a different base pointer.
9325 StoreSDNode *Index = St;
9327 // If the chain has more than one use, then we can't reorder the mem ops.
9328 if (Index != St && !SDValue(Index, 0)->hasOneUse())
9331 // Find the base pointer and offset for this memory node.
9332 BaseIndexOffset Ptr = BaseIndexOffset::match(Index->getBasePtr());
9334 // Check that the base pointer is the same as the original one.
9335 if (!Ptr.equalBaseIndex(BasePtr))
9338 // Check that the alignment is the same.
9339 if (Index->getAlignment() != St->getAlignment())
9342 // The memory operands must not be volatile.
9343 if (Index->isVolatile() || Index->isIndexed())
9347 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
9348 if (St->isTruncatingStore())
9351 // The stored memory type must be the same.
9352 if (Index->getMemoryVT() != MemVT)
9355 // We do not allow unaligned stores because we want to prevent overriding
9357 if (Index->getAlignment()*8 != MemVT.getSizeInBits())
9360 // We found a potential memory operand to merge.
9361 StoreNodes.push_back(MemOpLink(Index, Ptr.Offset, Seq++));
9363 // Find the next memory operand in the chain. If the next operand in the
9364 // chain is a store then move up and continue the scan with the next
9365 // memory operand. If the next operand is a load save it and use alias
9366 // information to check if it interferes with anything.
9367 SDNode *NextInChain = Index->getChain().getNode();
9369 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
9370 // We found a store node. Use it for the next iteration.
9373 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
9374 if (Ldn->isVolatile()) {
9379 // Save the load node for later. Continue the scan.
9380 AliasLoadNodes.push_back(Ldn);
9381 NextInChain = Ldn->getChain().getNode();
9390 // Check if there is anything to merge.
9391 if (StoreNodes.size() < 2)
9394 // Sort the memory operands according to their distance from the base pointer.
9395 std::sort(StoreNodes.begin(), StoreNodes.end(),
9396 [](MemOpLink LHS, MemOpLink RHS) {
9397 return LHS.OffsetFromBase < RHS.OffsetFromBase ||
9398 (LHS.OffsetFromBase == RHS.OffsetFromBase &&
9399 LHS.SequenceNum > RHS.SequenceNum);
9402 // Scan the memory operations on the chain and find the first non-consecutive
9403 // store memory address.
9404 unsigned LastConsecutiveStore = 0;
9405 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
9406 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) {
9408 // Check that the addresses are consecutive starting from the second
9409 // element in the list of stores.
9411 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
9412 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
9417 // Check if this store interferes with any of the loads that we found.
9418 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld)
9419 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) {
9423 // We found a load that alias with this store. Stop the sequence.
9427 // Mark this node as useful.
9428 LastConsecutiveStore = i;
9431 // The node with the lowest store address.
9432 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
9434 // Store the constants into memory as one consecutive store.
9436 unsigned LastLegalType = 0;
9437 unsigned LastLegalVectorType = 0;
9438 bool NonZero = false;
9439 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
9440 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9441 SDValue StoredVal = St->getValue();
9443 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
9444 NonZero |= !C->isNullValue();
9445 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) {
9446 NonZero |= !C->getConstantFPValue()->isNullValue();
9452 // Find a legal type for the constant store.
9453 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
9454 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9455 if (TLI.isTypeLegal(StoreTy))
9456 LastLegalType = i+1;
9457 // Or check whether a truncstore is legal.
9458 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
9459 TargetLowering::TypePromoteInteger) {
9460 EVT LegalizedStoredValueTy =
9461 TLI.getTypeToTransformTo(*DAG.getContext(), StoredVal.getValueType());
9462 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy))
9463 LastLegalType = i+1;
9466 // Find a legal type for the vector store.
9467 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
9468 if (TLI.isTypeLegal(Ty))
9469 LastLegalVectorType = i + 1;
9472 // We only use vectors if the constant is known to be zero and the
9473 // function is not marked with the noimplicitfloat attribute.
9474 if (NonZero || NoVectors)
9475 LastLegalVectorType = 0;
9477 // Check if we found a legal integer type to store.
9478 if (LastLegalType == 0 && LastLegalVectorType == 0)
9481 bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors;
9482 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType;
9484 // Make sure we have something to merge.
9488 unsigned EarliestNodeUsed = 0;
9489 for (unsigned i=0; i < NumElem; ++i) {
9490 // Find a chain for the new wide-store operand. Notice that some
9491 // of the store nodes that we found may not be selected for inclusion
9492 // in the wide store. The chain we use needs to be the chain of the
9493 // earliest store node which is *used* and replaced by the wide store.
9494 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
9495 EarliestNodeUsed = i;
9498 // The earliest Node in the DAG.
9499 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
9500 SDLoc DL(StoreNodes[0].MemNode);
9504 // Find a legal type for the vector store.
9505 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
9506 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
9507 StoredVal = DAG.getConstant(0, Ty);
9509 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
9510 APInt StoreInt(StoreBW, 0);
9512 // Construct a single integer constant which is made of the smaller
9514 bool IsLE = TLI.isLittleEndian();
9515 for (unsigned i = 0; i < NumElem ; ++i) {
9516 unsigned Idx = IsLE ?(NumElem - 1 - i) : i;
9517 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
9518 SDValue Val = St->getValue();
9519 StoreInt<<=ElementSizeBytes*8;
9520 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
9521 StoreInt|=C->getAPIntValue().zext(StoreBW);
9522 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
9523 StoreInt|= C->getValueAPF().bitcastToAPInt().zext(StoreBW);
9525 assert(false && "Invalid constant element type");
9529 // Create the new Load and Store operations.
9530 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9531 StoredVal = DAG.getConstant(StoreInt, StoreTy);
9534 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), DL, StoredVal,
9535 FirstInChain->getBasePtr(),
9536 FirstInChain->getPointerInfo(),
9538 FirstInChain->getAlignment());
9540 // Replace the first store with the new store
9541 CombineTo(EarliestOp, NewStore);
9542 // Erase all other stores.
9543 for (unsigned i = 0; i < NumElem ; ++i) {
9544 if (StoreNodes[i].MemNode == EarliestOp)
9546 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9547 // ReplaceAllUsesWith will replace all uses that existed when it was
9548 // called, but graph optimizations may cause new ones to appear. For
9549 // example, the case in pr14333 looks like
9551 // St's chain -> St -> another store -> X
9553 // And the only difference from St to the other store is the chain.
9554 // When we change it's chain to be St's chain they become identical,
9555 // get CSEed and the net result is that X is now a use of St.
9556 // Since we know that St is redundant, just iterate.
9557 while (!St->use_empty())
9558 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain());
9559 deleteAndRecombine(St);
9565 // Below we handle the case of multiple consecutive stores that
9566 // come from multiple consecutive loads. We merge them into a single
9567 // wide load and a single wide store.
9569 // Look for load nodes which are used by the stored values.
9570 SmallVector<MemOpLink, 8> LoadNodes;
9572 // Find acceptable loads. Loads need to have the same chain (token factor),
9573 // must not be zext, volatile, indexed, and they must be consecutive.
9574 BaseIndexOffset LdBasePtr;
9575 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
9576 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9577 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
9580 // Loads must only have one use.
9581 if (!Ld->hasNUsesOfValue(1, 0))
9584 // Check that the alignment is the same as the stores.
9585 if (Ld->getAlignment() != St->getAlignment())
9588 // The memory operands must not be volatile.
9589 if (Ld->isVolatile() || Ld->isIndexed())
9592 // We do not accept ext loads.
9593 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
9596 // The stored memory type must be the same.
9597 if (Ld->getMemoryVT() != MemVT)
9600 BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld->getBasePtr());
9601 // If this is not the first ptr that we check.
9602 if (LdBasePtr.Base.getNode()) {
9603 // The base ptr must be the same.
9604 if (!LdPtr.equalBaseIndex(LdBasePtr))
9607 // Check that all other base pointers are the same as this one.
9611 // We found a potential memory operand to merge.
9612 LoadNodes.push_back(MemOpLink(Ld, LdPtr.Offset, 0));
9615 if (LoadNodes.size() < 2)
9618 // If we have load/store pair instructions and we only have two values,
9620 unsigned RequiredAlignment;
9621 if (LoadNodes.size() == 2 && TLI.hasPairedLoad(MemVT, RequiredAlignment) &&
9622 St->getAlignment() >= RequiredAlignment)
9625 // Scan the memory operations on the chain and find the first non-consecutive
9626 // load memory address. These variables hold the index in the store node
9628 unsigned LastConsecutiveLoad = 0;
9629 // This variable refers to the size and not index in the array.
9630 unsigned LastLegalVectorType = 0;
9631 unsigned LastLegalIntegerType = 0;
9632 StartAddress = LoadNodes[0].OffsetFromBase;
9633 SDValue FirstChain = LoadNodes[0].MemNode->getChain();
9634 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
9635 // All loads much share the same chain.
9636 if (LoadNodes[i].MemNode->getChain() != FirstChain)
9639 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
9640 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
9642 LastConsecutiveLoad = i;
9644 // Find a legal type for the vector store.
9645 EVT StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
9646 if (TLI.isTypeLegal(StoreTy))
9647 LastLegalVectorType = i + 1;
9649 // Find a legal type for the integer store.
9650 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
9651 StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9652 if (TLI.isTypeLegal(StoreTy))
9653 LastLegalIntegerType = i + 1;
9654 // Or check whether a truncstore and extload is legal.
9655 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
9656 TargetLowering::TypePromoteInteger) {
9657 EVT LegalizedStoredValueTy =
9658 TLI.getTypeToTransformTo(*DAG.getContext(), StoreTy);
9659 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
9660 TLI.isLoadExtLegal(ISD::ZEXTLOAD, StoreTy) &&
9661 TLI.isLoadExtLegal(ISD::SEXTLOAD, StoreTy) &&
9662 TLI.isLoadExtLegal(ISD::EXTLOAD, StoreTy))
9663 LastLegalIntegerType = i+1;
9667 // Only use vector types if the vector type is larger than the integer type.
9668 // If they are the same, use integers.
9669 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors;
9670 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType);
9672 // We add +1 here because the LastXXX variables refer to location while
9673 // the NumElem refers to array/index size.
9674 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1;
9675 NumElem = std::min(LastLegalType, NumElem);
9680 // The earliest Node in the DAG.
9681 unsigned EarliestNodeUsed = 0;
9682 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
9683 for (unsigned i=1; i<NumElem; ++i) {
9684 // Find a chain for the new wide-store operand. Notice that some
9685 // of the store nodes that we found may not be selected for inclusion
9686 // in the wide store. The chain we use needs to be the chain of the
9687 // earliest store node which is *used* and replaced by the wide store.
9688 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
9689 EarliestNodeUsed = i;
9692 // Find if it is better to use vectors or integers to load and store
9696 JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
9698 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
9699 JointMemOpVT = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9702 SDLoc LoadDL(LoadNodes[0].MemNode);
9703 SDLoc StoreDL(StoreNodes[0].MemNode);
9705 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
9706 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL,
9707 FirstLoad->getChain(),
9708 FirstLoad->getBasePtr(),
9709 FirstLoad->getPointerInfo(),
9710 false, false, false,
9711 FirstLoad->getAlignment());
9713 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), StoreDL, NewLoad,
9714 FirstInChain->getBasePtr(),
9715 FirstInChain->getPointerInfo(), false, false,
9716 FirstInChain->getAlignment());
9718 // Replace one of the loads with the new load.
9719 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode);
9720 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
9721 SDValue(NewLoad.getNode(), 1));
9723 // Remove the rest of the load chains.
9724 for (unsigned i = 1; i < NumElem ; ++i) {
9725 // Replace all chain users of the old load nodes with the chain of the new
9727 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
9728 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain());
9731 // Replace the first store with the new store.
9732 CombineTo(EarliestOp, NewStore);
9733 // Erase all other stores.
9734 for (unsigned i = 0; i < NumElem ; ++i) {
9735 // Remove all Store nodes.
9736 if (StoreNodes[i].MemNode == EarliestOp)
9738 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9739 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain());
9740 deleteAndRecombine(St);
9746 SDValue DAGCombiner::visitSTORE(SDNode *N) {
9747 StoreSDNode *ST = cast<StoreSDNode>(N);
9748 SDValue Chain = ST->getChain();
9749 SDValue Value = ST->getValue();
9750 SDValue Ptr = ST->getBasePtr();
9752 // If this is a store of a bit convert, store the input value if the
9753 // resultant store does not need a higher alignment than the original.
9754 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
9755 ST->isUnindexed()) {
9756 unsigned OrigAlign = ST->getAlignment();
9757 EVT SVT = Value.getOperand(0).getValueType();
9758 unsigned Align = TLI.getDataLayout()->
9759 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
9760 if (Align <= OrigAlign &&
9761 ((!LegalOperations && !ST->isVolatile()) ||
9762 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
9763 return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0),
9764 Ptr, ST->getPointerInfo(), ST->isVolatile(),
9765 ST->isNonTemporal(), OrigAlign,
9769 // Turn 'store undef, Ptr' -> nothing.
9770 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
9773 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
9774 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
9775 // NOTE: If the original store is volatile, this transform must not increase
9776 // the number of stores. For example, on x86-32 an f64 can be stored in one
9777 // processor operation but an i64 (which is not legal) requires two. So the
9778 // transform should not be done in this case.
9779 if (Value.getOpcode() != ISD::TargetConstantFP) {
9781 switch (CFP->getSimpleValueType(0).SimpleTy) {
9782 default: llvm_unreachable("Unknown FP type");
9783 case MVT::f16: // We don't do this for these yet.
9789 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
9790 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
9791 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
9792 bitcastToAPInt().getZExtValue(), MVT::i32);
9793 return DAG.getStore(Chain, SDLoc(N), Tmp,
9794 Ptr, ST->getMemOperand());
9798 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
9799 !ST->isVolatile()) ||
9800 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
9801 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
9802 getZExtValue(), MVT::i64);
9803 return DAG.getStore(Chain, SDLoc(N), Tmp,
9804 Ptr, ST->getMemOperand());
9807 if (!ST->isVolatile() &&
9808 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
9809 // Many FP stores are not made apparent until after legalize, e.g. for
9810 // argument passing. Since this is so common, custom legalize the
9811 // 64-bit integer store into two 32-bit stores.
9812 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
9813 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
9814 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
9815 if (TLI.isBigEndian()) std::swap(Lo, Hi);
9817 unsigned Alignment = ST->getAlignment();
9818 bool isVolatile = ST->isVolatile();
9819 bool isNonTemporal = ST->isNonTemporal();
9820 AAMDNodes AAInfo = ST->getAAInfo();
9822 SDValue St0 = DAG.getStore(Chain, SDLoc(ST), Lo,
9823 Ptr, ST->getPointerInfo(),
9824 isVolatile, isNonTemporal,
9825 ST->getAlignment(), AAInfo);
9826 Ptr = DAG.getNode(ISD::ADD, SDLoc(N), Ptr.getValueType(), Ptr,
9827 DAG.getConstant(4, Ptr.getValueType()));
9828 Alignment = MinAlign(Alignment, 4U);
9829 SDValue St1 = DAG.getStore(Chain, SDLoc(ST), Hi,
9830 Ptr, ST->getPointerInfo().getWithOffset(4),
9831 isVolatile, isNonTemporal,
9833 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other,
9842 // Try to infer better alignment information than the store already has.
9843 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
9844 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
9845 if (Align > ST->getAlignment())
9846 return DAG.getTruncStore(Chain, SDLoc(N), Value,
9847 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
9848 ST->isVolatile(), ST->isNonTemporal(), Align,
9853 // Try transforming a pair floating point load / store ops to integer
9854 // load / store ops.
9855 SDValue NewST = TransformFPLoadStorePair(N);
9856 if (NewST.getNode())
9859 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
9860 : DAG.getSubtarget().useAA();
9862 if (CombinerAAOnlyFunc.getNumOccurrences() &&
9863 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
9866 if (UseAA && ST->isUnindexed()) {
9867 // Walk up chain skipping non-aliasing memory nodes.
9868 SDValue BetterChain = FindBetterChain(N, Chain);
9870 // If there is a better chain.
9871 if (Chain != BetterChain) {
9874 // Replace the chain to avoid dependency.
9875 if (ST->isTruncatingStore()) {
9876 ReplStore = DAG.getTruncStore(BetterChain, SDLoc(N), Value, Ptr,
9877 ST->getMemoryVT(), ST->getMemOperand());
9879 ReplStore = DAG.getStore(BetterChain, SDLoc(N), Value, Ptr,
9880 ST->getMemOperand());
9883 // Create token to keep both nodes around.
9884 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
9885 MVT::Other, Chain, ReplStore);
9887 // Make sure the new and old chains are cleaned up.
9888 AddToWorklist(Token.getNode());
9890 // Don't add users to work list.
9891 return CombineTo(N, Token, false);
9895 // Try transforming N to an indexed store.
9896 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
9897 return SDValue(N, 0);
9899 // FIXME: is there such a thing as a truncating indexed store?
9900 if (ST->isTruncatingStore() && ST->isUnindexed() &&
9901 Value.getValueType().isInteger()) {
9902 // See if we can simplify the input to this truncstore with knowledge that
9903 // only the low bits are being used. For example:
9904 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
9906 GetDemandedBits(Value,
9907 APInt::getLowBitsSet(
9908 Value.getValueType().getScalarType().getSizeInBits(),
9909 ST->getMemoryVT().getScalarType().getSizeInBits()));
9910 AddToWorklist(Value.getNode());
9911 if (Shorter.getNode())
9912 return DAG.getTruncStore(Chain, SDLoc(N), Shorter,
9913 Ptr, ST->getMemoryVT(), ST->getMemOperand());
9915 // Otherwise, see if we can simplify the operation with
9916 // SimplifyDemandedBits, which only works if the value has a single use.
9917 if (SimplifyDemandedBits(Value,
9918 APInt::getLowBitsSet(
9919 Value.getValueType().getScalarType().getSizeInBits(),
9920 ST->getMemoryVT().getScalarType().getSizeInBits())))
9921 return SDValue(N, 0);
9924 // If this is a load followed by a store to the same location, then the store
9926 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
9927 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
9928 ST->isUnindexed() && !ST->isVolatile() &&
9929 // There can't be any side effects between the load and store, such as
9931 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
9932 // The store is dead, remove it.
9937 // If this is a store followed by a store with the same value to the same
9938 // location, then the store is dead/noop.
9939 if (StoreSDNode *ST1 = dyn_cast<StoreSDNode>(Chain)) {
9940 if (ST1->getBasePtr() == Ptr && ST->getMemoryVT() == ST1->getMemoryVT() &&
9941 ST1->getValue() == Value && ST->isUnindexed() && !ST->isVolatile() &&
9942 ST1->isUnindexed() && !ST1->isVolatile()) {
9943 // The store is dead, remove it.
9948 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
9949 // truncating store. We can do this even if this is already a truncstore.
9950 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
9951 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
9952 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
9953 ST->getMemoryVT())) {
9954 return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0),
9955 Ptr, ST->getMemoryVT(), ST->getMemOperand());
9958 // Only perform this optimization before the types are legal, because we
9959 // don't want to perform this optimization on every DAGCombine invocation.
9961 bool EverChanged = false;
9964 // There can be multiple store sequences on the same chain.
9965 // Keep trying to merge store sequences until we are unable to do so
9966 // or until we merge the last store on the chain.
9967 bool Changed = MergeConsecutiveStores(ST);
9968 EverChanged |= Changed;
9969 if (!Changed) break;
9970 } while (ST->getOpcode() != ISD::DELETED_NODE);
9973 return SDValue(N, 0);
9976 return ReduceLoadOpStoreWidth(N);
9979 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
9980 SDValue InVec = N->getOperand(0);
9981 SDValue InVal = N->getOperand(1);
9982 SDValue EltNo = N->getOperand(2);
9985 // If the inserted element is an UNDEF, just use the input vector.
9986 if (InVal.getOpcode() == ISD::UNDEF)
9989 EVT VT = InVec.getValueType();
9991 // If we can't generate a legal BUILD_VECTOR, exit
9992 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
9995 // Check that we know which element is being inserted
9996 if (!isa<ConstantSDNode>(EltNo))
9998 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
10000 // Canonicalize insert_vector_elt dag nodes.
10002 // (insert_vector_elt (insert_vector_elt A, Idx0), Idx1)
10003 // -> (insert_vector_elt (insert_vector_elt A, Idx1), Idx0)
10005 // Do this only if the child insert_vector node has one use; also
10006 // do this only if indices are both constants and Idx1 < Idx0.
10007 if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT && InVec.hasOneUse()
10008 && isa<ConstantSDNode>(InVec.getOperand(2))) {
10009 unsigned OtherElt =
10010 cast<ConstantSDNode>(InVec.getOperand(2))->getZExtValue();
10011 if (Elt < OtherElt) {
10013 SDValue NewOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(N), VT,
10014 InVec.getOperand(0), InVal, EltNo);
10015 AddToWorklist(NewOp.getNode());
10016 return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(InVec.getNode()),
10017 VT, NewOp, InVec.getOperand(1), InVec.getOperand(2));
10021 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
10022 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
10023 // vector elements.
10024 SmallVector<SDValue, 8> Ops;
10025 // Do not combine these two vectors if the output vector will not replace
10026 // the input vector.
10027 if (InVec.getOpcode() == ISD::BUILD_VECTOR && InVec.hasOneUse()) {
10028 Ops.append(InVec.getNode()->op_begin(),
10029 InVec.getNode()->op_end());
10030 } else if (InVec.getOpcode() == ISD::UNDEF) {
10031 unsigned NElts = VT.getVectorNumElements();
10032 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
10037 // Insert the element
10038 if (Elt < Ops.size()) {
10039 // All the operands of BUILD_VECTOR must have the same type;
10040 // we enforce that here.
10041 EVT OpVT = Ops[0].getValueType();
10042 if (InVal.getValueType() != OpVT)
10043 InVal = OpVT.bitsGT(InVal.getValueType()) ?
10044 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
10045 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
10049 // Return the new vector
10050 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
10053 SDValue DAGCombiner::ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
10054 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad) {
10055 EVT ResultVT = EVE->getValueType(0);
10056 EVT VecEltVT = InVecVT.getVectorElementType();
10057 unsigned Align = OriginalLoad->getAlignment();
10058 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
10059 VecEltVT.getTypeForEVT(*DAG.getContext()));
10061 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VecEltVT))
10066 SDValue NewPtr = OriginalLoad->getBasePtr();
10068 EVT PtrType = NewPtr.getValueType();
10069 MachinePointerInfo MPI;
10070 if (auto *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo)) {
10071 int Elt = ConstEltNo->getZExtValue();
10072 unsigned PtrOff = VecEltVT.getSizeInBits() * Elt / 8;
10073 if (TLI.isBigEndian())
10074 PtrOff = InVecVT.getSizeInBits() / 8 - PtrOff;
10075 Offset = DAG.getConstant(PtrOff, PtrType);
10076 MPI = OriginalLoad->getPointerInfo().getWithOffset(PtrOff);
10078 Offset = DAG.getNode(
10079 ISD::MUL, SDLoc(EVE), EltNo.getValueType(), EltNo,
10080 DAG.getConstant(VecEltVT.getStoreSize(), EltNo.getValueType()));
10081 if (TLI.isBigEndian())
10082 Offset = DAG.getNode(
10083 ISD::SUB, SDLoc(EVE), EltNo.getValueType(),
10084 DAG.getConstant(InVecVT.getStoreSize(), EltNo.getValueType()), Offset);
10085 MPI = OriginalLoad->getPointerInfo();
10087 NewPtr = DAG.getNode(ISD::ADD, SDLoc(EVE), PtrType, NewPtr, Offset);
10089 // The replacement we need to do here is a little tricky: we need to
10090 // replace an extractelement of a load with a load.
10091 // Use ReplaceAllUsesOfValuesWith to do the replacement.
10092 // Note that this replacement assumes that the extractvalue is the only
10093 // use of the load; that's okay because we don't want to perform this
10094 // transformation in other cases anyway.
10097 if (ResultVT.bitsGT(VecEltVT)) {
10098 // If the result type of vextract is wider than the load, then issue an
10099 // extending load instead.
10100 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, VecEltVT)
10103 Load = DAG.getExtLoad(
10104 ExtType, SDLoc(EVE), ResultVT, OriginalLoad->getChain(), NewPtr, MPI,
10105 VecEltVT, OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
10106 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
10107 Chain = Load.getValue(1);
10109 Load = DAG.getLoad(
10110 VecEltVT, SDLoc(EVE), OriginalLoad->getChain(), NewPtr, MPI,
10111 OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
10112 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
10113 Chain = Load.getValue(1);
10114 if (ResultVT.bitsLT(VecEltVT))
10115 Load = DAG.getNode(ISD::TRUNCATE, SDLoc(EVE), ResultVT, Load);
10117 Load = DAG.getNode(ISD::BITCAST, SDLoc(EVE), ResultVT, Load);
10119 WorklistRemover DeadNodes(*this);
10120 SDValue From[] = { SDValue(EVE, 0), SDValue(OriginalLoad, 1) };
10121 SDValue To[] = { Load, Chain };
10122 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
10123 // Since we're explicitly calling ReplaceAllUses, add the new node to the
10124 // worklist explicitly as well.
10125 AddToWorklist(Load.getNode());
10126 AddUsersToWorklist(Load.getNode()); // Add users too
10127 // Make sure to revisit this node to clean it up; it will usually be dead.
10128 AddToWorklist(EVE);
10130 return SDValue(EVE, 0);
10133 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
10134 // (vextract (scalar_to_vector val, 0) -> val
10135 SDValue InVec = N->getOperand(0);
10136 EVT VT = InVec.getValueType();
10137 EVT NVT = N->getValueType(0);
10139 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
10140 // Check if the result type doesn't match the inserted element type. A
10141 // SCALAR_TO_VECTOR may truncate the inserted element and the
10142 // EXTRACT_VECTOR_ELT may widen the extracted vector.
10143 SDValue InOp = InVec.getOperand(0);
10144 if (InOp.getValueType() != NVT) {
10145 assert(InOp.getValueType().isInteger() && NVT.isInteger());
10146 return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT);
10151 SDValue EltNo = N->getOperand(1);
10152 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
10154 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
10155 // We only perform this optimization before the op legalization phase because
10156 // we may introduce new vector instructions which are not backed by TD
10157 // patterns. For example on AVX, extracting elements from a wide vector
10158 // without using extract_subvector. However, if we can find an underlying
10159 // scalar value, then we can always use that.
10160 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
10162 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
10163 int NumElem = VT.getVectorNumElements();
10164 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
10165 // Find the new index to extract from.
10166 int OrigElt = SVOp->getMaskElt(Elt);
10168 // Extracting an undef index is undef.
10170 return DAG.getUNDEF(NVT);
10172 // Select the right vector half to extract from.
10174 if (OrigElt < NumElem) {
10175 SVInVec = InVec->getOperand(0);
10177 SVInVec = InVec->getOperand(1);
10178 OrigElt -= NumElem;
10181 if (SVInVec.getOpcode() == ISD::BUILD_VECTOR) {
10182 SDValue InOp = SVInVec.getOperand(OrigElt);
10183 if (InOp.getValueType() != NVT) {
10184 assert(InOp.getValueType().isInteger() && NVT.isInteger());
10185 InOp = DAG.getSExtOrTrunc(InOp, SDLoc(SVInVec), NVT);
10191 // FIXME: We should handle recursing on other vector shuffles and
10192 // scalar_to_vector here as well.
10194 if (!LegalOperations) {
10195 EVT IndexTy = TLI.getVectorIdxTy();
10196 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT,
10197 SVInVec, DAG.getConstant(OrigElt, IndexTy));
10201 bool BCNumEltsChanged = false;
10202 EVT ExtVT = VT.getVectorElementType();
10205 // If the result of load has to be truncated, then it's not necessarily
10207 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
10210 if (InVec.getOpcode() == ISD::BITCAST) {
10211 // Don't duplicate a load with other uses.
10212 if (!InVec.hasOneUse())
10215 EVT BCVT = InVec.getOperand(0).getValueType();
10216 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
10218 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
10219 BCNumEltsChanged = true;
10220 InVec = InVec.getOperand(0);
10221 ExtVT = BCVT.getVectorElementType();
10224 // (vextract (vN[if]M load $addr), i) -> ([if]M load $addr + i * size)
10225 if (!LegalOperations && !ConstEltNo && InVec.hasOneUse() &&
10226 ISD::isNormalLoad(InVec.getNode()) &&
10227 !N->getOperand(1)->hasPredecessor(InVec.getNode())) {
10228 SDValue Index = N->getOperand(1);
10229 if (LoadSDNode *OrigLoad = dyn_cast<LoadSDNode>(InVec))
10230 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, Index,
10234 // Perform only after legalization to ensure build_vector / vector_shuffle
10235 // optimizations have already been done.
10236 if (!LegalOperations) return SDValue();
10238 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
10239 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
10240 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
10243 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
10245 LoadSDNode *LN0 = nullptr;
10246 const ShuffleVectorSDNode *SVN = nullptr;
10247 if (ISD::isNormalLoad(InVec.getNode())) {
10248 LN0 = cast<LoadSDNode>(InVec);
10249 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
10250 InVec.getOperand(0).getValueType() == ExtVT &&
10251 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
10252 // Don't duplicate a load with other uses.
10253 if (!InVec.hasOneUse())
10256 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
10257 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
10258 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
10260 // (load $addr+1*size)
10262 // Don't duplicate a load with other uses.
10263 if (!InVec.hasOneUse())
10266 // If the bit convert changed the number of elements, it is unsafe
10267 // to examine the mask.
10268 if (BCNumEltsChanged)
10271 // Select the input vector, guarding against out of range extract vector.
10272 unsigned NumElems = VT.getVectorNumElements();
10273 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
10274 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
10276 if (InVec.getOpcode() == ISD::BITCAST) {
10277 // Don't duplicate a load with other uses.
10278 if (!InVec.hasOneUse())
10281 InVec = InVec.getOperand(0);
10283 if (ISD::isNormalLoad(InVec.getNode())) {
10284 LN0 = cast<LoadSDNode>(InVec);
10285 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
10286 EltNo = DAG.getConstant(Elt, EltNo.getValueType());
10290 // Make sure we found a non-volatile load and the extractelement is
10292 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
10295 // If Idx was -1 above, Elt is going to be -1, so just return undef.
10297 return DAG.getUNDEF(LVT);
10299 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, EltNo, LN0);
10305 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
10306 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
10307 // We perform this optimization post type-legalization because
10308 // the type-legalizer often scalarizes integer-promoted vectors.
10309 // Performing this optimization before may create bit-casts which
10310 // will be type-legalized to complex code sequences.
10311 // We perform this optimization only before the operation legalizer because we
10312 // may introduce illegal operations.
10313 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
10316 unsigned NumInScalars = N->getNumOperands();
10318 EVT VT = N->getValueType(0);
10320 // Check to see if this is a BUILD_VECTOR of a bunch of values
10321 // which come from any_extend or zero_extend nodes. If so, we can create
10322 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
10323 // optimizations. We do not handle sign-extend because we can't fill the sign
10325 EVT SourceType = MVT::Other;
10326 bool AllAnyExt = true;
10328 for (unsigned i = 0; i != NumInScalars; ++i) {
10329 SDValue In = N->getOperand(i);
10330 // Ignore undef inputs.
10331 if (In.getOpcode() == ISD::UNDEF) continue;
10333 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
10334 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
10336 // Abort if the element is not an extension.
10337 if (!ZeroExt && !AnyExt) {
10338 SourceType = MVT::Other;
10342 // The input is a ZeroExt or AnyExt. Check the original type.
10343 EVT InTy = In.getOperand(0).getValueType();
10345 // Check that all of the widened source types are the same.
10346 if (SourceType == MVT::Other)
10349 else if (InTy != SourceType) {
10350 // Multiple income types. Abort.
10351 SourceType = MVT::Other;
10355 // Check if all of the extends are ANY_EXTENDs.
10356 AllAnyExt &= AnyExt;
10359 // In order to have valid types, all of the inputs must be extended from the
10360 // same source type and all of the inputs must be any or zero extend.
10361 // Scalar sizes must be a power of two.
10362 EVT OutScalarTy = VT.getScalarType();
10363 bool ValidTypes = SourceType != MVT::Other &&
10364 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
10365 isPowerOf2_32(SourceType.getSizeInBits());
10367 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
10368 // turn into a single shuffle instruction.
10372 bool isLE = TLI.isLittleEndian();
10373 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
10374 assert(ElemRatio > 1 && "Invalid element size ratio");
10375 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
10376 DAG.getConstant(0, SourceType);
10378 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
10379 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
10381 // Populate the new build_vector
10382 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
10383 SDValue Cast = N->getOperand(i);
10384 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
10385 Cast.getOpcode() == ISD::ZERO_EXTEND ||
10386 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
10388 if (Cast.getOpcode() == ISD::UNDEF)
10389 In = DAG.getUNDEF(SourceType);
10391 In = Cast->getOperand(0);
10392 unsigned Index = isLE ? (i * ElemRatio) :
10393 (i * ElemRatio + (ElemRatio - 1));
10395 assert(Index < Ops.size() && "Invalid index");
10399 // The type of the new BUILD_VECTOR node.
10400 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
10401 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
10402 "Invalid vector size");
10403 // Check if the new vector type is legal.
10404 if (!isTypeLegal(VecVT)) return SDValue();
10406 // Make the new BUILD_VECTOR.
10407 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
10409 // The new BUILD_VECTOR node has the potential to be further optimized.
10410 AddToWorklist(BV.getNode());
10411 // Bitcast to the desired type.
10412 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
10415 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
10416 EVT VT = N->getValueType(0);
10418 unsigned NumInScalars = N->getNumOperands();
10421 EVT SrcVT = MVT::Other;
10422 unsigned Opcode = ISD::DELETED_NODE;
10423 unsigned NumDefs = 0;
10425 for (unsigned i = 0; i != NumInScalars; ++i) {
10426 SDValue In = N->getOperand(i);
10427 unsigned Opc = In.getOpcode();
10429 if (Opc == ISD::UNDEF)
10432 // If all scalar values are floats and converted from integers.
10433 if (Opcode == ISD::DELETED_NODE &&
10434 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
10441 EVT InVT = In.getOperand(0).getValueType();
10443 // If all scalar values are typed differently, bail out. It's chosen to
10444 // simplify BUILD_VECTOR of integer types.
10445 if (SrcVT == MVT::Other)
10452 // If the vector has just one element defined, it's not worth to fold it into
10453 // a vectorized one.
10457 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
10458 && "Should only handle conversion from integer to float.");
10459 assert(SrcVT != MVT::Other && "Cannot determine source type!");
10461 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
10463 if (!TLI.isOperationLegalOrCustom(Opcode, NVT))
10466 SmallVector<SDValue, 8> Opnds;
10467 for (unsigned i = 0; i != NumInScalars; ++i) {
10468 SDValue In = N->getOperand(i);
10470 if (In.getOpcode() == ISD::UNDEF)
10471 Opnds.push_back(DAG.getUNDEF(SrcVT));
10473 Opnds.push_back(In.getOperand(0));
10475 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Opnds);
10476 AddToWorklist(BV.getNode());
10478 return DAG.getNode(Opcode, dl, VT, BV);
10481 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
10482 unsigned NumInScalars = N->getNumOperands();
10484 EVT VT = N->getValueType(0);
10486 // A vector built entirely of undefs is undef.
10487 if (ISD::allOperandsUndef(N))
10488 return DAG.getUNDEF(VT);
10490 SDValue V = reduceBuildVecExtToExtBuildVec(N);
10494 V = reduceBuildVecConvertToConvertBuildVec(N);
10498 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
10499 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
10500 // at most two distinct vectors, turn this into a shuffle node.
10502 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
10503 if (!isTypeLegal(VT))
10506 // May only combine to shuffle after legalize if shuffle is legal.
10507 if (LegalOperations && !TLI.isOperationLegal(ISD::VECTOR_SHUFFLE, VT))
10510 SDValue VecIn1, VecIn2;
10511 for (unsigned i = 0; i != NumInScalars; ++i) {
10512 // Ignore undef inputs.
10513 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
10515 // If this input is something other than a EXTRACT_VECTOR_ELT with a
10516 // constant index, bail out.
10517 if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
10518 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
10519 VecIn1 = VecIn2 = SDValue(nullptr, 0);
10523 // We allow up to two distinct input vectors.
10524 SDValue ExtractedFromVec = N->getOperand(i).getOperand(0);
10525 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
10528 if (!VecIn1.getNode()) {
10529 VecIn1 = ExtractedFromVec;
10530 } else if (!VecIn2.getNode()) {
10531 VecIn2 = ExtractedFromVec;
10533 // Too many inputs.
10534 VecIn1 = VecIn2 = SDValue(nullptr, 0);
10539 // If everything is good, we can make a shuffle operation.
10540 if (VecIn1.getNode()) {
10541 SmallVector<int, 8> Mask;
10542 for (unsigned i = 0; i != NumInScalars; ++i) {
10543 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
10544 Mask.push_back(-1);
10548 // If extracting from the first vector, just use the index directly.
10549 SDValue Extract = N->getOperand(i);
10550 SDValue ExtVal = Extract.getOperand(1);
10551 if (Extract.getOperand(0) == VecIn1) {
10552 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
10553 if (ExtIndex > VT.getVectorNumElements())
10556 Mask.push_back(ExtIndex);
10560 // Otherwise, use InIdx + VecSize
10561 unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue();
10562 Mask.push_back(Idx+NumInScalars);
10565 // We can't generate a shuffle node with mismatched input and output types.
10566 // Attempt to transform a single input vector to the correct type.
10567 if ((VT != VecIn1.getValueType())) {
10568 // We don't support shuffeling between TWO values of different types.
10569 if (VecIn2.getNode())
10572 // We only support widening of vectors which are half the size of the
10573 // output registers. For example XMM->YMM widening on X86 with AVX.
10574 if (VecIn1.getValueType().getSizeInBits()*2 != VT.getSizeInBits())
10577 // If the input vector type has a different base type to the output
10578 // vector type, bail out.
10579 if (VecIn1.getValueType().getVectorElementType() !=
10580 VT.getVectorElementType())
10583 // Widen the input vector by adding undef values.
10584 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
10585 VecIn1, DAG.getUNDEF(VecIn1.getValueType()));
10588 // If VecIn2 is unused then change it to undef.
10589 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
10591 // Check that we were able to transform all incoming values to the same
10593 if (VecIn2.getValueType() != VecIn1.getValueType() ||
10594 VecIn1.getValueType() != VT)
10597 // Return the new VECTOR_SHUFFLE node.
10601 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]);
10607 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
10608 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
10609 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
10610 // inputs come from at most two distinct vectors, turn this into a shuffle
10613 // If we only have one input vector, we don't need to do any concatenation.
10614 if (N->getNumOperands() == 1)
10615 return N->getOperand(0);
10617 // Check if all of the operands are undefs.
10618 EVT VT = N->getValueType(0);
10619 if (ISD::allOperandsUndef(N))
10620 return DAG.getUNDEF(VT);
10622 // Optimize concat_vectors where one of the vectors is undef.
10623 if (N->getNumOperands() == 2 &&
10624 N->getOperand(1)->getOpcode() == ISD::UNDEF) {
10625 SDValue In = N->getOperand(0);
10626 assert(In.getValueType().isVector() && "Must concat vectors");
10628 // Transform: concat_vectors(scalar, undef) -> scalar_to_vector(sclr).
10629 if (In->getOpcode() == ISD::BITCAST &&
10630 !In->getOperand(0)->getValueType(0).isVector()) {
10631 SDValue Scalar = In->getOperand(0);
10632 EVT SclTy = Scalar->getValueType(0);
10634 if (!SclTy.isFloatingPoint() && !SclTy.isInteger())
10637 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SclTy,
10638 VT.getSizeInBits() / SclTy.getSizeInBits());
10639 if (!TLI.isTypeLegal(NVT) || !TLI.isTypeLegal(Scalar.getValueType()))
10642 SDLoc dl = SDLoc(N);
10643 SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NVT, Scalar);
10644 return DAG.getNode(ISD::BITCAST, dl, VT, Res);
10648 // fold (concat_vectors (BUILD_VECTOR A, B, ...), (BUILD_VECTOR C, D, ...))
10649 // -> (BUILD_VECTOR A, B, ..., C, D, ...)
10650 if (N->getNumOperands() == 2 &&
10651 N->getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
10652 N->getOperand(1).getOpcode() == ISD::BUILD_VECTOR) {
10653 EVT VT = N->getValueType(0);
10654 SDValue N0 = N->getOperand(0);
10655 SDValue N1 = N->getOperand(1);
10656 SmallVector<SDValue, 8> Opnds;
10657 unsigned BuildVecNumElts = N0.getNumOperands();
10659 EVT SclTy0 = N0.getOperand(0)->getValueType(0);
10660 EVT SclTy1 = N1.getOperand(0)->getValueType(0);
10661 if (SclTy0.isFloatingPoint()) {
10662 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10663 Opnds.push_back(N0.getOperand(i));
10664 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10665 Opnds.push_back(N1.getOperand(i));
10667 // If BUILD_VECTOR are from built from integer, they may have different
10668 // operand types. Get the smaller type and truncate all operands to it.
10669 EVT MinTy = SclTy0.bitsLE(SclTy1) ? SclTy0 : SclTy1;
10670 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10671 Opnds.push_back(DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinTy,
10672 N0.getOperand(i)));
10673 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10674 Opnds.push_back(DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinTy,
10675 N1.getOperand(i)));
10678 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
10681 // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
10682 // nodes often generate nop CONCAT_VECTOR nodes.
10683 // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that
10684 // place the incoming vectors at the exact same location.
10685 SDValue SingleSource = SDValue();
10686 unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements();
10688 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
10689 SDValue Op = N->getOperand(i);
10691 if (Op.getOpcode() == ISD::UNDEF)
10694 // Check if this is the identity extract:
10695 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
10698 // Find the single incoming vector for the extract_subvector.
10699 if (SingleSource.getNode()) {
10700 if (Op.getOperand(0) != SingleSource)
10703 SingleSource = Op.getOperand(0);
10705 // Check the source type is the same as the type of the result.
10706 // If not, this concat may extend the vector, so we can not
10707 // optimize it away.
10708 if (SingleSource.getValueType() != N->getValueType(0))
10712 unsigned IdentityIndex = i * PartNumElem;
10713 ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1));
10714 // The extract index must be constant.
10718 // Check that we are reading from the identity index.
10719 if (CS->getZExtValue() != IdentityIndex)
10723 if (SingleSource.getNode())
10724 return SingleSource;
10729 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
10730 EVT NVT = N->getValueType(0);
10731 SDValue V = N->getOperand(0);
10733 if (V->getOpcode() == ISD::CONCAT_VECTORS) {
10735 // (extract_subvec (concat V1, V2, ...), i)
10738 // Only operand 0 is checked as 'concat' assumes all inputs of the same
10740 if (V->getOperand(0).getValueType() != NVT)
10742 unsigned Idx = dyn_cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
10743 unsigned NumElems = NVT.getVectorNumElements();
10744 assert((Idx % NumElems) == 0 &&
10745 "IDX in concat is not a multiple of the result vector length.");
10746 return V->getOperand(Idx / NumElems);
10750 if (V->getOpcode() == ISD::BITCAST)
10751 V = V.getOperand(0);
10753 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
10755 // Handle only simple case where vector being inserted and vector
10756 // being extracted are of same type, and are half size of larger vectors.
10757 EVT BigVT = V->getOperand(0).getValueType();
10758 EVT SmallVT = V->getOperand(1).getValueType();
10759 if (!NVT.bitsEq(SmallVT) || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
10762 // Only handle cases where both indexes are constants with the same type.
10763 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
10764 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
10766 if (InsIdx && ExtIdx &&
10767 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
10768 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
10770 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
10772 // indices are equal or bit offsets are equal => V1
10773 // otherwise => (extract_subvec V1, ExtIdx)
10774 if (InsIdx->getZExtValue() * SmallVT.getScalarType().getSizeInBits() ==
10775 ExtIdx->getZExtValue() * NVT.getScalarType().getSizeInBits())
10776 return DAG.getNode(ISD::BITCAST, dl, NVT, V->getOperand(1));
10777 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT,
10778 DAG.getNode(ISD::BITCAST, dl,
10779 N->getOperand(0).getValueType(),
10780 V->getOperand(0)), N->getOperand(1));
10787 static SDValue simplifyShuffleOperandRecursively(SmallBitVector &UsedElements,
10788 SDValue V, SelectionDAG &DAG) {
10790 EVT VT = V.getValueType();
10792 switch (V.getOpcode()) {
10796 case ISD::CONCAT_VECTORS: {
10797 EVT OpVT = V->getOperand(0).getValueType();
10798 int OpSize = OpVT.getVectorNumElements();
10799 SmallBitVector OpUsedElements(OpSize, false);
10800 bool FoundSimplification = false;
10801 SmallVector<SDValue, 4> NewOps;
10802 NewOps.reserve(V->getNumOperands());
10803 for (int i = 0, NumOps = V->getNumOperands(); i < NumOps; ++i) {
10804 SDValue Op = V->getOperand(i);
10805 bool OpUsed = false;
10806 for (int j = 0; j < OpSize; ++j)
10807 if (UsedElements[i * OpSize + j]) {
10808 OpUsedElements[j] = true;
10812 OpUsed ? simplifyShuffleOperandRecursively(OpUsedElements, Op, DAG)
10813 : DAG.getUNDEF(OpVT));
10814 FoundSimplification |= Op == NewOps.back();
10815 OpUsedElements.reset();
10817 if (FoundSimplification)
10818 V = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, NewOps);
10822 case ISD::INSERT_SUBVECTOR: {
10823 SDValue BaseV = V->getOperand(0);
10824 SDValue SubV = V->getOperand(1);
10825 auto *IdxN = dyn_cast<ConstantSDNode>(V->getOperand(2));
10829 int SubSize = SubV.getValueType().getVectorNumElements();
10830 int Idx = IdxN->getZExtValue();
10831 bool SubVectorUsed = false;
10832 SmallBitVector SubUsedElements(SubSize, false);
10833 for (int i = 0; i < SubSize; ++i)
10834 if (UsedElements[i + Idx]) {
10835 SubVectorUsed = true;
10836 SubUsedElements[i] = true;
10837 UsedElements[i + Idx] = false;
10840 // Now recurse on both the base and sub vectors.
10841 SDValue SimplifiedSubV =
10843 ? simplifyShuffleOperandRecursively(SubUsedElements, SubV, DAG)
10844 : DAG.getUNDEF(SubV.getValueType());
10845 SDValue SimplifiedBaseV = simplifyShuffleOperandRecursively(UsedElements, BaseV, DAG);
10846 if (SimplifiedSubV != SubV || SimplifiedBaseV != BaseV)
10847 V = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT,
10848 SimplifiedBaseV, SimplifiedSubV, V->getOperand(2));
10854 static SDValue simplifyShuffleOperands(ShuffleVectorSDNode *SVN, SDValue N0,
10855 SDValue N1, SelectionDAG &DAG) {
10856 EVT VT = SVN->getValueType(0);
10857 int NumElts = VT.getVectorNumElements();
10858 SmallBitVector N0UsedElements(NumElts, false), N1UsedElements(NumElts, false);
10859 for (int M : SVN->getMask())
10860 if (M >= 0 && M < NumElts)
10861 N0UsedElements[M] = true;
10862 else if (M >= NumElts)
10863 N1UsedElements[M - NumElts] = true;
10865 SDValue S0 = simplifyShuffleOperandRecursively(N0UsedElements, N0, DAG);
10866 SDValue S1 = simplifyShuffleOperandRecursively(N1UsedElements, N1, DAG);
10867 if (S0 == N0 && S1 == N1)
10870 return DAG.getVectorShuffle(VT, SDLoc(SVN), S0, S1, SVN->getMask());
10873 // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat.
10874 static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
10875 EVT VT = N->getValueType(0);
10876 unsigned NumElts = VT.getVectorNumElements();
10878 SDValue N0 = N->getOperand(0);
10879 SDValue N1 = N->getOperand(1);
10880 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
10882 SmallVector<SDValue, 4> Ops;
10883 EVT ConcatVT = N0.getOperand(0).getValueType();
10884 unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements();
10885 unsigned NumConcats = NumElts / NumElemsPerConcat;
10887 // Look at every vector that's inserted. We're looking for exact
10888 // subvector-sized copies from a concatenated vector
10889 for (unsigned I = 0; I != NumConcats; ++I) {
10890 // Make sure we're dealing with a copy.
10891 unsigned Begin = I * NumElemsPerConcat;
10892 bool AllUndef = true, NoUndef = true;
10893 for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) {
10894 if (SVN->getMaskElt(J) >= 0)
10901 if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0)
10904 for (unsigned J = 1; J != NumElemsPerConcat; ++J)
10905 if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J))
10908 unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat;
10909 if (FirstElt < N0.getNumOperands())
10910 Ops.push_back(N0.getOperand(FirstElt));
10912 Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands()));
10914 } else if (AllUndef) {
10915 Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType()));
10916 } else { // Mixed with general masks and undefs, can't do optimization.
10921 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops);
10924 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
10925 EVT VT = N->getValueType(0);
10926 unsigned NumElts = VT.getVectorNumElements();
10928 SDValue N0 = N->getOperand(0);
10929 SDValue N1 = N->getOperand(1);
10931 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
10933 // Canonicalize shuffle undef, undef -> undef
10934 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
10935 return DAG.getUNDEF(VT);
10937 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
10939 // Canonicalize shuffle v, v -> v, undef
10941 SmallVector<int, 8> NewMask;
10942 for (unsigned i = 0; i != NumElts; ++i) {
10943 int Idx = SVN->getMaskElt(i);
10944 if (Idx >= (int)NumElts) Idx -= NumElts;
10945 NewMask.push_back(Idx);
10947 return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT),
10951 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
10952 if (N0.getOpcode() == ISD::UNDEF) {
10953 SmallVector<int, 8> NewMask;
10954 for (unsigned i = 0; i != NumElts; ++i) {
10955 int Idx = SVN->getMaskElt(i);
10957 if (Idx >= (int)NumElts)
10960 Idx = -1; // remove reference to lhs
10962 NewMask.push_back(Idx);
10964 return DAG.getVectorShuffle(VT, SDLoc(N), N1, DAG.getUNDEF(VT),
10968 // Remove references to rhs if it is undef
10969 if (N1.getOpcode() == ISD::UNDEF) {
10970 bool Changed = false;
10971 SmallVector<int, 8> NewMask;
10972 for (unsigned i = 0; i != NumElts; ++i) {
10973 int Idx = SVN->getMaskElt(i);
10974 if (Idx >= (int)NumElts) {
10978 NewMask.push_back(Idx);
10981 return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, &NewMask[0]);
10984 // If it is a splat, check if the argument vector is another splat or a
10985 // build_vector with all scalar elements the same.
10986 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
10987 SDNode *V = N0.getNode();
10989 // If this is a bit convert that changes the element type of the vector but
10990 // not the number of vector elements, look through it. Be careful not to
10991 // look though conversions that change things like v4f32 to v2f64.
10992 if (V->getOpcode() == ISD::BITCAST) {
10993 SDValue ConvInput = V->getOperand(0);
10994 if (ConvInput.getValueType().isVector() &&
10995 ConvInput.getValueType().getVectorNumElements() == NumElts)
10996 V = ConvInput.getNode();
10999 if (V->getOpcode() == ISD::BUILD_VECTOR) {
11000 assert(V->getNumOperands() == NumElts &&
11001 "BUILD_VECTOR has wrong number of operands");
11003 bool AllSame = true;
11004 for (unsigned i = 0; i != NumElts; ++i) {
11005 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
11006 Base = V->getOperand(i);
11010 // Splat of <u, u, u, u>, return <u, u, u, u>
11011 if (!Base.getNode())
11013 for (unsigned i = 0; i != NumElts; ++i) {
11014 if (V->getOperand(i) != Base) {
11019 // Splat of <x, x, x, x>, return <x, x, x, x>
11025 // There are various patterns used to build up a vector from smaller vectors,
11026 // subvectors, or elements. Scan chains of these and replace unused insertions
11027 // or components with undef.
11028 if (SDValue S = simplifyShuffleOperands(SVN, N0, N1, DAG))
11031 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
11032 Level < AfterLegalizeVectorOps &&
11033 (N1.getOpcode() == ISD::UNDEF ||
11034 (N1.getOpcode() == ISD::CONCAT_VECTORS &&
11035 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) {
11036 SDValue V = partitionShuffleOfConcats(N, DAG);
11042 // If this shuffle node is simply a swizzle of another shuffle node,
11043 // then try to simplify it.
11044 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
11045 N1.getOpcode() == ISD::UNDEF) {
11047 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
11049 // The incoming shuffle must be of the same type as the result of the
11050 // current shuffle.
11051 assert(OtherSV->getOperand(0).getValueType() == VT &&
11052 "Shuffle types don't match");
11054 SmallVector<int, 4> Mask;
11055 // Compute the combined shuffle mask.
11056 for (unsigned i = 0; i != NumElts; ++i) {
11057 int Idx = SVN->getMaskElt(i);
11058 assert(Idx < (int)NumElts && "Index references undef operand");
11059 // Next, this index comes from the first value, which is the incoming
11060 // shuffle. Adopt the incoming index.
11062 Idx = OtherSV->getMaskElt(Idx);
11063 Mask.push_back(Idx);
11066 // Check if all indices in Mask are Undef. In case, propagate Undef.
11067 bool isUndefMask = true;
11068 for (unsigned i = 0; i != NumElts && isUndefMask; ++i)
11069 isUndefMask &= Mask[i] < 0;
11072 return DAG.getUNDEF(VT);
11074 bool CommuteOperands = false;
11075 if (N0.getOperand(1).getOpcode() != ISD::UNDEF) {
11076 // To be valid, the combine shuffle mask should only reference elements
11077 // from one of the two vectors in input to the inner shufflevector.
11078 bool IsValidMask = true;
11079 for (unsigned i = 0; i != NumElts && IsValidMask; ++i)
11080 // See if the combined mask only reference undefs or elements coming
11081 // from the first shufflevector operand.
11082 IsValidMask = Mask[i] < 0 || (unsigned)Mask[i] < NumElts;
11084 if (!IsValidMask) {
11085 IsValidMask = true;
11086 for (unsigned i = 0; i != NumElts && IsValidMask; ++i)
11087 // Check that all the elements come from the second shuffle operand.
11088 IsValidMask = Mask[i] < 0 || (unsigned)Mask[i] >= NumElts;
11089 CommuteOperands = IsValidMask;
11092 // Early exit if the combined shuffle mask is not valid.
11097 // See if this pair of shuffles can be safely folded according to either
11098 // of the following rules:
11099 // shuffle(shuffle(x, y), undef) -> x
11100 // shuffle(shuffle(x, undef), undef) -> x
11101 // shuffle(shuffle(x, y), undef) -> y
11102 bool IsIdentityMask = true;
11103 unsigned BaseMaskIndex = CommuteOperands ? NumElts : 0;
11104 for (unsigned i = 0; i != NumElts && IsIdentityMask; ++i) {
11109 // The combined shuffle must map each index to itself.
11110 IsIdentityMask = (unsigned)Mask[i] == i + BaseMaskIndex;
11113 if (IsIdentityMask) {
11114 if (CommuteOperands)
11115 // optimize shuffle(shuffle(x, y), undef) -> y.
11116 return OtherSV->getOperand(1);
11118 // optimize shuffle(shuffle(x, undef), undef) -> x
11119 // optimize shuffle(shuffle(x, y), undef) -> x
11120 return OtherSV->getOperand(0);
11123 // It may still be beneficial to combine the two shuffles if the
11124 // resulting shuffle is legal.
11125 if (TLI.isTypeLegal(VT)) {
11126 if (!CommuteOperands) {
11127 if (TLI.isShuffleMaskLegal(Mask, VT))
11128 // shuffle(shuffle(x, undef, M1), undef, M2) -> shuffle(x, undef, M3).
11129 // shuffle(shuffle(x, y, M1), undef, M2) -> shuffle(x, undef, M3)
11130 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0), N1,
11133 // Compute the commuted shuffle mask.
11134 for (unsigned i = 0; i != NumElts; ++i) {
11138 else if (idx < (int)NumElts)
11139 Mask[i] = idx + NumElts;
11141 Mask[i] = idx - NumElts;
11144 if (TLI.isShuffleMaskLegal(Mask, VT))
11145 // shuffle(shuffle(x, y, M1), undef, M2) -> shuffle(y, undef, M3)
11146 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(1), N1,
11152 // Canonicalize shuffles according to rules:
11153 // shuffle(A, shuffle(A, B)) -> shuffle(shuffle(A,B), A)
11154 // shuffle(B, shuffle(A, B)) -> shuffle(shuffle(A,B), B)
11155 // shuffle(B, shuffle(A, Undef)) -> shuffle(shuffle(A, Undef), B)
11156 if (N1.getOpcode() == ISD::VECTOR_SHUFFLE && N0.getOpcode() != ISD::UNDEF &&
11157 N0.getOpcode() != ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
11158 TLI.isTypeLegal(VT)) {
11159 // The incoming shuffle must be of the same type as the result of the
11160 // current shuffle.
11161 assert(N1->getOperand(0).getValueType() == VT &&
11162 "Shuffle types don't match");
11164 SDValue SV0 = N1->getOperand(0);
11165 SDValue SV1 = N1->getOperand(1);
11166 bool HasSameOp0 = N0 == SV0;
11167 bool IsSV1Undef = SV1.getOpcode() == ISD::UNDEF;
11168 if (HasSameOp0 || IsSV1Undef || N0 == SV1)
11169 // Commute the operands of this shuffle so that next rule
11171 return DAG.getCommutedVectorShuffle(*SVN);
11174 // Try to fold according to rules:
11175 // shuffle(shuffle(A, B, M0), B, M1) -> shuffle(A, B, M2)
11176 // shuffle(shuffle(A, B, M0), A, M1) -> shuffle(A, B, M2)
11177 // shuffle(shuffle(A, Undef, M0), B, M1) -> shuffle(A, B, M2)
11178 // shuffle(shuffle(A, Undef, M0), A, M1) -> shuffle(A, Undef, M2)
11179 // Don't try to fold shuffles with illegal type.
11180 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
11181 N1.getOpcode() != ISD::UNDEF && TLI.isTypeLegal(VT)) {
11182 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
11184 // The incoming shuffle must be of the same type as the result of the
11185 // current shuffle.
11186 assert(OtherSV->getOperand(0).getValueType() == VT &&
11187 "Shuffle types don't match");
11189 SDValue SV0 = OtherSV->getOperand(0);
11190 SDValue SV1 = OtherSV->getOperand(1);
11191 bool HasSameOp0 = N1 == SV0;
11192 bool IsSV1Undef = SV1.getOpcode() == ISD::UNDEF;
11193 if (!HasSameOp0 && !IsSV1Undef && N1 != SV1)
11197 SmallVector<int, 4> Mask;
11198 // Compute the combined shuffle mask for a shuffle with SV0 as the first
11199 // operand, and SV1 as the second operand.
11200 for (unsigned i = 0; i != NumElts; ++i) {
11201 int Idx = SVN->getMaskElt(i);
11203 // Propagate Undef.
11204 Mask.push_back(Idx);
11208 if (Idx < (int)NumElts) {
11209 Idx = OtherSV->getMaskElt(Idx);
11210 if (IsSV1Undef && Idx >= (int) NumElts)
11211 Idx = -1; // Propagate Undef.
11213 Idx = HasSameOp0 ? Idx - NumElts : Idx;
11215 Mask.push_back(Idx);
11218 // Check if all indices in Mask are Undef. In case, propagate Undef.
11219 bool isUndefMask = true;
11220 for (unsigned i = 0; i != NumElts && isUndefMask; ++i)
11221 isUndefMask &= Mask[i] < 0;
11224 return DAG.getUNDEF(VT);
11226 // Avoid introducing shuffles with illegal mask.
11227 if (TLI.isShuffleMaskLegal(Mask, VT)) {
11229 // shuffle(shuffle(A, Undef, M0), B, M1) -> shuffle(A, B, M2)
11230 // shuffle(shuffle(A, Undef, M0), A, M1) -> shuffle(A, Undef, M2)
11231 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, N1, &Mask[0]);
11232 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, &Mask[0]);
11239 SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
11240 SDValue N0 = N->getOperand(0);
11241 SDValue N2 = N->getOperand(2);
11243 // If the input vector is a concatenation, and the insert replaces
11244 // one of the halves, we can optimize into a single concat_vectors.
11245 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
11246 N0->getNumOperands() == 2 && N2.getOpcode() == ISD::Constant) {
11247 APInt InsIdx = cast<ConstantSDNode>(N2)->getAPIntValue();
11248 EVT VT = N->getValueType(0);
11250 // Lower half: fold (insert_subvector (concat_vectors X, Y), Z) ->
11251 // (concat_vectors Z, Y)
11253 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
11254 N->getOperand(1), N0.getOperand(1));
11256 // Upper half: fold (insert_subvector (concat_vectors X, Y), Z) ->
11257 // (concat_vectors X, Z)
11258 if (InsIdx == VT.getVectorNumElements()/2)
11259 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
11260 N0.getOperand(0), N->getOperand(1));
11266 /// Returns a vector_shuffle if it able to transform an AND to a vector_shuffle
11267 /// with the destination vector and a zero vector.
11268 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
11269 /// vector_shuffle V, Zero, <0, 4, 2, 4>
11270 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
11271 EVT VT = N->getValueType(0);
11273 SDValue LHS = N->getOperand(0);
11274 SDValue RHS = N->getOperand(1);
11275 if (N->getOpcode() == ISD::AND) {
11276 if (RHS.getOpcode() == ISD::BITCAST)
11277 RHS = RHS.getOperand(0);
11278 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
11279 SmallVector<int, 8> Indices;
11280 unsigned NumElts = RHS.getNumOperands();
11281 for (unsigned i = 0; i != NumElts; ++i) {
11282 SDValue Elt = RHS.getOperand(i);
11283 if (!isa<ConstantSDNode>(Elt))
11286 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
11287 Indices.push_back(i);
11288 else if (cast<ConstantSDNode>(Elt)->isNullValue())
11289 Indices.push_back(NumElts);
11294 // Let's see if the target supports this vector_shuffle.
11295 EVT RVT = RHS.getValueType();
11296 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
11299 // Return the new VECTOR_SHUFFLE node.
11300 EVT EltVT = RVT.getVectorElementType();
11301 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
11302 DAG.getConstant(0, EltVT));
11303 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), RVT, ZeroOps);
11304 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
11305 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
11306 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
11313 /// Visit a binary vector operation, like ADD.
11314 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
11315 assert(N->getValueType(0).isVector() &&
11316 "SimplifyVBinOp only works on vectors!");
11318 SDValue LHS = N->getOperand(0);
11319 SDValue RHS = N->getOperand(1);
11320 SDValue Shuffle = XformToShuffleWithZero(N);
11321 if (Shuffle.getNode()) return Shuffle;
11323 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
11325 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
11326 RHS.getOpcode() == ISD::BUILD_VECTOR) {
11327 // Check if both vectors are constants. If not bail out.
11328 if (!(cast<BuildVectorSDNode>(LHS)->isConstant() &&
11329 cast<BuildVectorSDNode>(RHS)->isConstant()))
11332 SmallVector<SDValue, 8> Ops;
11333 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
11334 SDValue LHSOp = LHS.getOperand(i);
11335 SDValue RHSOp = RHS.getOperand(i);
11337 // Can't fold divide by zero.
11338 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
11339 N->getOpcode() == ISD::FDIV) {
11340 if ((RHSOp.getOpcode() == ISD::Constant &&
11341 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
11342 (RHSOp.getOpcode() == ISD::ConstantFP &&
11343 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
11347 EVT VT = LHSOp.getValueType();
11348 EVT RVT = RHSOp.getValueType();
11350 // Integer BUILD_VECTOR operands may have types larger than the element
11351 // size (e.g., when the element type is not legal). Prior to type
11352 // legalization, the types may not match between the two BUILD_VECTORS.
11353 // Truncate one of the operands to make them match.
11354 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
11355 RHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, RHSOp);
11357 LHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), RVT, LHSOp);
11361 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(LHS), VT,
11363 if (FoldOp.getOpcode() != ISD::UNDEF &&
11364 FoldOp.getOpcode() != ISD::Constant &&
11365 FoldOp.getOpcode() != ISD::ConstantFP)
11367 Ops.push_back(FoldOp);
11368 AddToWorklist(FoldOp.getNode());
11371 if (Ops.size() == LHS.getNumOperands())
11372 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), LHS.getValueType(), Ops);
11375 // Type legalization might introduce new shuffles in the DAG.
11376 // Fold (VBinOp (shuffle (A, Undef, Mask)), (shuffle (B, Undef, Mask)))
11377 // -> (shuffle (VBinOp (A, B)), Undef, Mask).
11378 if (LegalTypes && isa<ShuffleVectorSDNode>(LHS) &&
11379 isa<ShuffleVectorSDNode>(RHS) && LHS.hasOneUse() && RHS.hasOneUse() &&
11380 LHS.getOperand(1).getOpcode() == ISD::UNDEF &&
11381 RHS.getOperand(1).getOpcode() == ISD::UNDEF) {
11382 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(LHS);
11383 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(RHS);
11385 if (SVN0->getMask().equals(SVN1->getMask())) {
11386 EVT VT = N->getValueType(0);
11387 SDValue UndefVector = LHS.getOperand(1);
11388 SDValue NewBinOp = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
11389 LHS.getOperand(0), RHS.getOperand(0));
11390 AddUsersToWorklist(N);
11391 return DAG.getVectorShuffle(VT, SDLoc(N), NewBinOp, UndefVector,
11392 &SVN0->getMask()[0]);
11399 /// Visit a binary vector operation, like FABS/FNEG.
11400 SDValue DAGCombiner::SimplifyVUnaryOp(SDNode *N) {
11401 assert(N->getValueType(0).isVector() &&
11402 "SimplifyVUnaryOp only works on vectors!");
11404 SDValue N0 = N->getOperand(0);
11406 if (N0.getOpcode() != ISD::BUILD_VECTOR)
11409 // Operand is a BUILD_VECTOR node, see if we can constant fold it.
11410 SmallVector<SDValue, 8> Ops;
11411 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
11412 SDValue Op = N0.getOperand(i);
11413 if (Op.getOpcode() != ISD::UNDEF &&
11414 Op.getOpcode() != ISD::ConstantFP)
11416 EVT EltVT = Op.getValueType();
11417 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(N0), EltVT, Op);
11418 if (FoldOp.getOpcode() != ISD::UNDEF &&
11419 FoldOp.getOpcode() != ISD::ConstantFP)
11421 Ops.push_back(FoldOp);
11422 AddToWorklist(FoldOp.getNode());
11425 if (Ops.size() != N0.getNumOperands())
11428 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), N0.getValueType(), Ops);
11431 SDValue DAGCombiner::SimplifySelect(SDLoc DL, SDValue N0,
11432 SDValue N1, SDValue N2){
11433 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
11435 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
11436 cast<CondCodeSDNode>(N0.getOperand(2))->get());
11438 // If we got a simplified select_cc node back from SimplifySelectCC, then
11439 // break it down into a new SETCC node, and a new SELECT node, and then return
11440 // the SELECT node, since we were called with a SELECT node.
11441 if (SCC.getNode()) {
11442 // Check to see if we got a select_cc back (to turn into setcc/select).
11443 // Otherwise, just return whatever node we got back, like fabs.
11444 if (SCC.getOpcode() == ISD::SELECT_CC) {
11445 SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0),
11447 SCC.getOperand(0), SCC.getOperand(1),
11448 SCC.getOperand(4));
11449 AddToWorklist(SETCC.getNode());
11450 return DAG.getSelect(SDLoc(SCC), SCC.getValueType(), SETCC,
11451 SCC.getOperand(2), SCC.getOperand(3));
11459 /// Given a SELECT or a SELECT_CC node, where LHS and RHS are the two values
11460 /// being selected between, see if we can simplify the select. Callers of this
11461 /// should assume that TheSelect is deleted if this returns true. As such, they
11462 /// should return the appropriate thing (e.g. the node) back to the top-level of
11463 /// the DAG combiner loop to avoid it being looked at.
11464 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
11467 // Cannot simplify select with vector condition
11468 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
11470 // If this is a select from two identical things, try to pull the operation
11471 // through the select.
11472 if (LHS.getOpcode() != RHS.getOpcode() ||
11473 !LHS.hasOneUse() || !RHS.hasOneUse())
11476 // If this is a load and the token chain is identical, replace the select
11477 // of two loads with a load through a select of the address to load from.
11478 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
11479 // constants have been dropped into the constant pool.
11480 if (LHS.getOpcode() == ISD::LOAD) {
11481 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
11482 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
11484 // Token chains must be identical.
11485 if (LHS.getOperand(0) != RHS.getOperand(0) ||
11486 // Do not let this transformation reduce the number of volatile loads.
11487 LLD->isVolatile() || RLD->isVolatile() ||
11488 // If this is an EXTLOAD, the VT's must match.
11489 LLD->getMemoryVT() != RLD->getMemoryVT() ||
11490 // If this is an EXTLOAD, the kind of extension must match.
11491 (LLD->getExtensionType() != RLD->getExtensionType() &&
11492 // The only exception is if one of the extensions is anyext.
11493 LLD->getExtensionType() != ISD::EXTLOAD &&
11494 RLD->getExtensionType() != ISD::EXTLOAD) ||
11495 // FIXME: this discards src value information. This is
11496 // over-conservative. It would be beneficial to be able to remember
11497 // both potential memory locations. Since we are discarding
11498 // src value info, don't do the transformation if the memory
11499 // locations are not in the default address space.
11500 LLD->getPointerInfo().getAddrSpace() != 0 ||
11501 RLD->getPointerInfo().getAddrSpace() != 0 ||
11502 !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(),
11503 LLD->getBasePtr().getValueType()))
11506 // Check that the select condition doesn't reach either load. If so,
11507 // folding this will induce a cycle into the DAG. If not, this is safe to
11508 // xform, so create a select of the addresses.
11510 if (TheSelect->getOpcode() == ISD::SELECT) {
11511 SDNode *CondNode = TheSelect->getOperand(0).getNode();
11512 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
11513 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
11515 // The loads must not depend on one another.
11516 if (LLD->isPredecessorOf(RLD) ||
11517 RLD->isPredecessorOf(LLD))
11519 Addr = DAG.getSelect(SDLoc(TheSelect),
11520 LLD->getBasePtr().getValueType(),
11521 TheSelect->getOperand(0), LLD->getBasePtr(),
11522 RLD->getBasePtr());
11523 } else { // Otherwise SELECT_CC
11524 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
11525 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
11527 if ((LLD->hasAnyUseOfValue(1) &&
11528 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
11529 (RLD->hasAnyUseOfValue(1) &&
11530 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
11533 Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect),
11534 LLD->getBasePtr().getValueType(),
11535 TheSelect->getOperand(0),
11536 TheSelect->getOperand(1),
11537 LLD->getBasePtr(), RLD->getBasePtr(),
11538 TheSelect->getOperand(4));
11542 // It is safe to replace the two loads if they have different alignments,
11543 // but the new load must be the minimum (most restrictive) alignment of the
11545 bool isInvariant = LLD->isInvariant() & RLD->isInvariant();
11546 unsigned Alignment = std::min(LLD->getAlignment(), RLD->getAlignment());
11547 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
11548 Load = DAG.getLoad(TheSelect->getValueType(0),
11550 // FIXME: Discards pointer and AA info.
11551 LLD->getChain(), Addr, MachinePointerInfo(),
11552 LLD->isVolatile(), LLD->isNonTemporal(),
11553 isInvariant, Alignment);
11555 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
11556 RLD->getExtensionType() : LLD->getExtensionType(),
11558 TheSelect->getValueType(0),
11559 // FIXME: Discards pointer and AA info.
11560 LLD->getChain(), Addr, MachinePointerInfo(),
11561 LLD->getMemoryVT(), LLD->isVolatile(),
11562 LLD->isNonTemporal(), isInvariant, Alignment);
11565 // Users of the select now use the result of the load.
11566 CombineTo(TheSelect, Load);
11568 // Users of the old loads now use the new load's chain. We know the
11569 // old-load value is dead now.
11570 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
11571 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
11578 /// Simplify an expression of the form (N0 cond N1) ? N2 : N3
11579 /// where 'cond' is the comparison specified by CC.
11580 SDValue DAGCombiner::SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1,
11581 SDValue N2, SDValue N3,
11582 ISD::CondCode CC, bool NotExtCompare) {
11583 // (x ? y : y) -> y.
11584 if (N2 == N3) return N2;
11586 EVT VT = N2.getValueType();
11587 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
11588 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
11589 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
11591 // Determine if the condition we're dealing with is constant
11592 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
11593 N0, N1, CC, DL, false);
11594 if (SCC.getNode()) AddToWorklist(SCC.getNode());
11595 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
11597 // fold select_cc true, x, y -> x
11598 if (SCCC && !SCCC->isNullValue())
11600 // fold select_cc false, x, y -> y
11601 if (SCCC && SCCC->isNullValue())
11604 // Check to see if we can simplify the select into an fabs node
11605 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
11606 // Allow either -0.0 or 0.0
11607 if (CFP->getValueAPF().isZero()) {
11608 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
11609 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
11610 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
11611 N2 == N3.getOperand(0))
11612 return DAG.getNode(ISD::FABS, DL, VT, N0);
11614 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
11615 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
11616 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
11617 N2.getOperand(0) == N3)
11618 return DAG.getNode(ISD::FABS, DL, VT, N3);
11622 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
11623 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
11624 // in it. This is a win when the constant is not otherwise available because
11625 // it replaces two constant pool loads with one. We only do this if the FP
11626 // type is known to be legal, because if it isn't, then we are before legalize
11627 // types an we want the other legalization to happen first (e.g. to avoid
11628 // messing with soft float) and if the ConstantFP is not legal, because if
11629 // it is legal, we may not need to store the FP constant in a constant pool.
11630 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
11631 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
11632 if (TLI.isTypeLegal(N2.getValueType()) &&
11633 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
11634 TargetLowering::Legal &&
11635 !TLI.isFPImmLegal(TV->getValueAPF(), TV->getValueType(0)) &&
11636 !TLI.isFPImmLegal(FV->getValueAPF(), FV->getValueType(0))) &&
11637 // If both constants have multiple uses, then we won't need to do an
11638 // extra load, they are likely around in registers for other users.
11639 (TV->hasOneUse() || FV->hasOneUse())) {
11640 Constant *Elts[] = {
11641 const_cast<ConstantFP*>(FV->getConstantFPValue()),
11642 const_cast<ConstantFP*>(TV->getConstantFPValue())
11644 Type *FPTy = Elts[0]->getType();
11645 const DataLayout &TD = *TLI.getDataLayout();
11647 // Create a ConstantArray of the two constants.
11648 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
11649 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
11650 TD.getPrefTypeAlignment(FPTy));
11651 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
11653 // Get the offsets to the 0 and 1 element of the array so that we can
11654 // select between them.
11655 SDValue Zero = DAG.getIntPtrConstant(0);
11656 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
11657 SDValue One = DAG.getIntPtrConstant(EltSize);
11659 SDValue Cond = DAG.getSetCC(DL,
11660 getSetCCResultType(N0.getValueType()),
11662 AddToWorklist(Cond.getNode());
11663 SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(),
11665 AddToWorklist(CstOffset.getNode());
11666 CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx,
11668 AddToWorklist(CPIdx.getNode());
11669 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
11670 MachinePointerInfo::getConstantPool(), false,
11671 false, false, Alignment);
11676 // Check to see if we can perform the "gzip trick", transforming
11677 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
11678 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
11679 (N1C->isNullValue() || // (a < 0) ? b : 0
11680 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
11681 EVT XType = N0.getValueType();
11682 EVT AType = N2.getValueType();
11683 if (XType.bitsGE(AType)) {
11684 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
11685 // single-bit constant.
11686 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
11687 unsigned ShCtV = N2C->getAPIntValue().logBase2();
11688 ShCtV = XType.getSizeInBits()-ShCtV-1;
11689 SDValue ShCt = DAG.getConstant(ShCtV,
11690 getShiftAmountTy(N0.getValueType()));
11691 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0),
11693 AddToWorklist(Shift.getNode());
11695 if (XType.bitsGT(AType)) {
11696 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
11697 AddToWorklist(Shift.getNode());
11700 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
11703 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0),
11705 DAG.getConstant(XType.getSizeInBits()-1,
11706 getShiftAmountTy(N0.getValueType())));
11707 AddToWorklist(Shift.getNode());
11709 if (XType.bitsGT(AType)) {
11710 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
11711 AddToWorklist(Shift.getNode());
11714 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
11718 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
11719 // where y is has a single bit set.
11720 // A plaintext description would be, we can turn the SELECT_CC into an AND
11721 // when the condition can be materialized as an all-ones register. Any
11722 // single bit-test can be materialized as an all-ones register with
11723 // shift-left and shift-right-arith.
11724 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
11725 N0->getValueType(0) == VT &&
11726 N1C && N1C->isNullValue() &&
11727 N2C && N2C->isNullValue()) {
11728 SDValue AndLHS = N0->getOperand(0);
11729 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
11730 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
11731 // Shift the tested bit over the sign bit.
11732 APInt AndMask = ConstAndRHS->getAPIntValue();
11734 DAG.getConstant(AndMask.countLeadingZeros(),
11735 getShiftAmountTy(AndLHS.getValueType()));
11736 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt);
11738 // Now arithmetic right shift it all the way over, so the result is either
11739 // all-ones, or zero.
11741 DAG.getConstant(AndMask.getBitWidth()-1,
11742 getShiftAmountTy(Shl.getValueType()));
11743 SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt);
11745 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
11749 // fold select C, 16, 0 -> shl C, 4
11750 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
11751 TLI.getBooleanContents(N0.getValueType()) ==
11752 TargetLowering::ZeroOrOneBooleanContent) {
11754 // If the caller doesn't want us to simplify this into a zext of a compare,
11756 if (NotExtCompare && N2C->getAPIntValue() == 1)
11759 // Get a SetCC of the condition
11760 // NOTE: Don't create a SETCC if it's not legal on this target.
11761 if (!LegalOperations ||
11762 TLI.isOperationLegal(ISD::SETCC,
11763 LegalTypes ? getSetCCResultType(N0.getValueType()) : MVT::i1)) {
11765 // cast from setcc result type to select result type
11767 SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()),
11769 if (N2.getValueType().bitsLT(SCC.getValueType()))
11770 Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2),
11771 N2.getValueType());
11773 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
11774 N2.getValueType(), SCC);
11776 SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC);
11777 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
11778 N2.getValueType(), SCC);
11781 AddToWorklist(SCC.getNode());
11782 AddToWorklist(Temp.getNode());
11784 if (N2C->getAPIntValue() == 1)
11787 // shl setcc result by log2 n2c
11788 return DAG.getNode(
11789 ISD::SHL, DL, N2.getValueType(), Temp,
11790 DAG.getConstant(N2C->getAPIntValue().logBase2(),
11791 getShiftAmountTy(Temp.getValueType())));
11795 // Check to see if this is the equivalent of setcc
11796 // FIXME: Turn all of these into setcc if setcc if setcc is legal
11797 // otherwise, go ahead with the folds.
11798 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
11799 EVT XType = N0.getValueType();
11800 if (!LegalOperations ||
11801 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(XType))) {
11802 SDValue Res = DAG.getSetCC(DL, getSetCCResultType(XType), N0, N1, CC);
11803 if (Res.getValueType() != VT)
11804 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
11808 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
11809 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
11810 (!LegalOperations ||
11811 TLI.isOperationLegal(ISD::CTLZ, XType))) {
11812 SDValue Ctlz = DAG.getNode(ISD::CTLZ, SDLoc(N0), XType, N0);
11813 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
11814 DAG.getConstant(Log2_32(XType.getSizeInBits()),
11815 getShiftAmountTy(Ctlz.getValueType())));
11817 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
11818 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
11819 SDValue NegN0 = DAG.getNode(ISD::SUB, SDLoc(N0),
11820 XType, DAG.getConstant(0, XType), N0);
11821 SDValue NotN0 = DAG.getNOT(SDLoc(N0), N0, XType);
11822 return DAG.getNode(ISD::SRL, DL, XType,
11823 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
11824 DAG.getConstant(XType.getSizeInBits()-1,
11825 getShiftAmountTy(XType)));
11827 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
11828 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
11829 SDValue Sign = DAG.getNode(ISD::SRL, SDLoc(N0), XType, N0,
11830 DAG.getConstant(XType.getSizeInBits()-1,
11831 getShiftAmountTy(N0.getValueType())));
11832 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
11836 // Check to see if this is an integer abs.
11837 // select_cc setg[te] X, 0, X, -X ->
11838 // select_cc setgt X, -1, X, -X ->
11839 // select_cc setl[te] X, 0, -X, X ->
11840 // select_cc setlt X, 1, -X, X ->
11841 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
11843 ConstantSDNode *SubC = nullptr;
11844 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
11845 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
11846 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
11847 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
11848 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
11849 (N1C->isOne() && CC == ISD::SETLT)) &&
11850 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
11851 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
11853 EVT XType = N0.getValueType();
11854 if (SubC && SubC->isNullValue() && XType.isInteger()) {
11855 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0), XType,
11857 DAG.getConstant(XType.getSizeInBits()-1,
11858 getShiftAmountTy(N0.getValueType())));
11859 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0),
11861 AddToWorklist(Shift.getNode());
11862 AddToWorklist(Add.getNode());
11863 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
11870 /// This is a stub for TargetLowering::SimplifySetCC.
11871 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
11872 SDValue N1, ISD::CondCode Cond,
11873 SDLoc DL, bool foldBooleans) {
11874 TargetLowering::DAGCombinerInfo
11875 DagCombineInfo(DAG, Level, false, this);
11876 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
11879 /// Given an ISD::SDIV node expressing a divide by constant, return
11880 /// a DAG expression to select that will generate the same value by multiplying
11881 /// by a magic number.
11882 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
11883 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
11884 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
11888 // Avoid division by zero.
11889 if (!C->getAPIntValue())
11892 std::vector<SDNode*> Built;
11894 TLI.BuildSDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
11896 for (SDNode *N : Built)
11901 /// Given an ISD::SDIV node expressing a divide by constant power of 2, return a
11902 /// DAG expression that will generate the same value by right shifting.
11903 SDValue DAGCombiner::BuildSDIVPow2(SDNode *N) {
11904 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
11908 // Avoid division by zero.
11909 if (!C->getAPIntValue())
11912 std::vector<SDNode *> Built;
11913 SDValue S = TLI.BuildSDIVPow2(N, C->getAPIntValue(), DAG, &Built);
11915 for (SDNode *N : Built)
11920 /// Given an ISD::UDIV node expressing a divide by constant, return a DAG
11921 /// expression that will generate the same value by multiplying by a magic
11923 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
11924 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
11925 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
11929 // Avoid division by zero.
11930 if (!C->getAPIntValue())
11933 std::vector<SDNode*> Built;
11935 TLI.BuildUDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
11937 for (SDNode *N : Built)
11942 SDValue DAGCombiner::BuildReciprocalEstimate(SDValue Op) {
11943 if (Level >= AfterLegalizeDAG)
11946 // Expose the DAG combiner to the target combiner implementations.
11947 TargetLowering::DAGCombinerInfo DCI(DAG, Level, false, this);
11949 unsigned Iterations = 0;
11950 if (SDValue Est = TLI.getRecipEstimate(Op, DCI, Iterations)) {
11952 // Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
11953 // For the reciprocal, we need to find the zero of the function:
11954 // F(X) = A X - 1 [which has a zero at X = 1/A]
11956 // X_{i+1} = X_i (2 - A X_i) = X_i + X_i (1 - A X_i) [this second form
11957 // does not require additional intermediate precision]
11958 EVT VT = Op.getValueType();
11960 SDValue FPOne = DAG.getConstantFP(1.0, VT);
11962 AddToWorklist(Est.getNode());
11964 // Newton iterations: Est = Est + Est (1 - Arg * Est)
11965 for (unsigned i = 0; i < Iterations; ++i) {
11966 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Op, Est);
11967 AddToWorklist(NewEst.getNode());
11969 NewEst = DAG.getNode(ISD::FSUB, DL, VT, FPOne, NewEst);
11970 AddToWorklist(NewEst.getNode());
11972 NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst);
11973 AddToWorklist(NewEst.getNode());
11975 Est = DAG.getNode(ISD::FADD, DL, VT, Est, NewEst);
11976 AddToWorklist(Est.getNode());
11985 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
11986 /// For the reciprocal sqrt, we need to find the zero of the function:
11987 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
11989 /// X_{i+1} = X_i (1.5 - A X_i^2 / 2)
11990 /// As a result, we precompute A/2 prior to the iteration loop.
11991 SDValue DAGCombiner::BuildRsqrtNROneConst(SDValue Arg, SDValue Est,
11992 unsigned Iterations) {
11993 EVT VT = Arg.getValueType();
11995 SDValue ThreeHalves = DAG.getConstantFP(1.5, VT);
11997 // We now need 0.5 * Arg which we can write as (1.5 * Arg - Arg) so that
11998 // this entire sequence requires only one FP constant.
11999 SDValue HalfArg = DAG.getNode(ISD::FMUL, DL, VT, ThreeHalves, Arg);
12000 AddToWorklist(HalfArg.getNode());
12002 HalfArg = DAG.getNode(ISD::FSUB, DL, VT, HalfArg, Arg);
12003 AddToWorklist(HalfArg.getNode());
12005 // Newton iterations: Est = Est * (1.5 - HalfArg * Est * Est)
12006 for (unsigned i = 0; i < Iterations; ++i) {
12007 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, Est);
12008 AddToWorklist(NewEst.getNode());
12010 NewEst = DAG.getNode(ISD::FMUL, DL, VT, HalfArg, NewEst);
12011 AddToWorklist(NewEst.getNode());
12013 NewEst = DAG.getNode(ISD::FSUB, DL, VT, ThreeHalves, NewEst);
12014 AddToWorklist(NewEst.getNode());
12016 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst);
12017 AddToWorklist(Est.getNode());
12022 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
12023 /// For the reciprocal sqrt, we need to find the zero of the function:
12024 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
12026 /// X_{i+1} = (-0.5 * X_i) * (A * X_i * X_i + (-3.0))
12027 SDValue DAGCombiner::BuildRsqrtNRTwoConst(SDValue Arg, SDValue Est,
12028 unsigned Iterations) {
12029 EVT VT = Arg.getValueType();
12031 SDValue MinusThree = DAG.getConstantFP(-3.0, VT);
12032 SDValue MinusHalf = DAG.getConstantFP(-0.5, VT);
12034 // Newton iterations: Est = -0.5 * Est * (-3.0 + Arg * Est * Est)
12035 for (unsigned i = 0; i < Iterations; ++i) {
12036 SDValue HalfEst = DAG.getNode(ISD::FMUL, DL, VT, Est, MinusHalf);
12037 AddToWorklist(HalfEst.getNode());
12039 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Est);
12040 AddToWorklist(Est.getNode());
12042 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Arg);
12043 AddToWorklist(Est.getNode());
12045 Est = DAG.getNode(ISD::FADD, DL, VT, Est, MinusThree);
12046 AddToWorklist(Est.getNode());
12048 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, HalfEst);
12049 AddToWorklist(Est.getNode());
12054 SDValue DAGCombiner::BuildRsqrtEstimate(SDValue Op) {
12055 if (Level >= AfterLegalizeDAG)
12058 // Expose the DAG combiner to the target combiner implementations.
12059 TargetLowering::DAGCombinerInfo DCI(DAG, Level, false, this);
12060 unsigned Iterations = 0;
12061 bool UseOneConstNR = false;
12062 if (SDValue Est = TLI.getRsqrtEstimate(Op, DCI, Iterations, UseOneConstNR)) {
12063 AddToWorklist(Est.getNode());
12065 Est = UseOneConstNR ?
12066 BuildRsqrtNROneConst(Op, Est, Iterations) :
12067 BuildRsqrtNRTwoConst(Op, Est, Iterations);
12075 /// Return true if base is a frame index, which is known not to alias with
12076 /// anything but itself. Provides base object and offset as results.
12077 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
12078 const GlobalValue *&GV, const void *&CV) {
12079 // Assume it is a primitive operation.
12080 Base = Ptr; Offset = 0; GV = nullptr; CV = nullptr;
12082 // If it's an adding a simple constant then integrate the offset.
12083 if (Base.getOpcode() == ISD::ADD) {
12084 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
12085 Base = Base.getOperand(0);
12086 Offset += C->getZExtValue();
12090 // Return the underlying GlobalValue, and update the Offset. Return false
12091 // for GlobalAddressSDNode since the same GlobalAddress may be represented
12092 // by multiple nodes with different offsets.
12093 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
12094 GV = G->getGlobal();
12095 Offset += G->getOffset();
12099 // Return the underlying Constant value, and update the Offset. Return false
12100 // for ConstantSDNodes since the same constant pool entry may be represented
12101 // by multiple nodes with different offsets.
12102 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
12103 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
12104 : (const void *)C->getConstVal();
12105 Offset += C->getOffset();
12108 // If it's any of the following then it can't alias with anything but itself.
12109 return isa<FrameIndexSDNode>(Base);
12112 /// Return true if there is any possibility that the two addresses overlap.
12113 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const {
12114 // If they are the same then they must be aliases.
12115 if (Op0->getBasePtr() == Op1->getBasePtr()) return true;
12117 // If they are both volatile then they cannot be reordered.
12118 if (Op0->isVolatile() && Op1->isVolatile()) return true;
12120 // Gather base node and offset information.
12121 SDValue Base1, Base2;
12122 int64_t Offset1, Offset2;
12123 const GlobalValue *GV1, *GV2;
12124 const void *CV1, *CV2;
12125 bool isFrameIndex1 = FindBaseOffset(Op0->getBasePtr(),
12126 Base1, Offset1, GV1, CV1);
12127 bool isFrameIndex2 = FindBaseOffset(Op1->getBasePtr(),
12128 Base2, Offset2, GV2, CV2);
12130 // If they have a same base address then check to see if they overlap.
12131 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
12132 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
12133 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
12135 // It is possible for different frame indices to alias each other, mostly
12136 // when tail call optimization reuses return address slots for arguments.
12137 // To catch this case, look up the actual index of frame indices to compute
12138 // the real alias relationship.
12139 if (isFrameIndex1 && isFrameIndex2) {
12140 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12141 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
12142 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
12143 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
12144 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
12147 // Otherwise, if we know what the bases are, and they aren't identical, then
12148 // we know they cannot alias.
12149 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
12152 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
12153 // compared to the size and offset of the access, we may be able to prove they
12154 // do not alias. This check is conservative for now to catch cases created by
12155 // splitting vector types.
12156 if ((Op0->getOriginalAlignment() == Op1->getOriginalAlignment()) &&
12157 (Op0->getSrcValueOffset() != Op1->getSrcValueOffset()) &&
12158 (Op0->getMemoryVT().getSizeInBits() >> 3 ==
12159 Op1->getMemoryVT().getSizeInBits() >> 3) &&
12160 (Op0->getOriginalAlignment() > Op0->getMemoryVT().getSizeInBits()) >> 3) {
12161 int64_t OffAlign1 = Op0->getSrcValueOffset() % Op0->getOriginalAlignment();
12162 int64_t OffAlign2 = Op1->getSrcValueOffset() % Op1->getOriginalAlignment();
12164 // There is no overlap between these relatively aligned accesses of similar
12165 // size, return no alias.
12166 if ((OffAlign1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign2 ||
12167 (OffAlign2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign1)
12171 bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0
12173 : DAG.getSubtarget().useAA();
12175 if (CombinerAAOnlyFunc.getNumOccurrences() &&
12176 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
12180 Op0->getMemOperand()->getValue() && Op1->getMemOperand()->getValue()) {
12181 // Use alias analysis information.
12182 int64_t MinOffset = std::min(Op0->getSrcValueOffset(),
12183 Op1->getSrcValueOffset());
12184 int64_t Overlap1 = (Op0->getMemoryVT().getSizeInBits() >> 3) +
12185 Op0->getSrcValueOffset() - MinOffset;
12186 int64_t Overlap2 = (Op1->getMemoryVT().getSizeInBits() >> 3) +
12187 Op1->getSrcValueOffset() - MinOffset;
12188 AliasAnalysis::AliasResult AAResult =
12189 AA.alias(AliasAnalysis::Location(Op0->getMemOperand()->getValue(),
12191 UseTBAA ? Op0->getAAInfo() : AAMDNodes()),
12192 AliasAnalysis::Location(Op1->getMemOperand()->getValue(),
12194 UseTBAA ? Op1->getAAInfo() : AAMDNodes()));
12195 if (AAResult == AliasAnalysis::NoAlias)
12199 // Otherwise we have to assume they alias.
12203 /// Walk up chain skipping non-aliasing memory nodes,
12204 /// looking for aliasing nodes and adding them to the Aliases vector.
12205 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
12206 SmallVectorImpl<SDValue> &Aliases) {
12207 SmallVector<SDValue, 8> Chains; // List of chains to visit.
12208 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
12210 // Get alias information for node.
12211 bool IsLoad = isa<LoadSDNode>(N) && !cast<LSBaseSDNode>(N)->isVolatile();
12214 Chains.push_back(OriginalChain);
12215 unsigned Depth = 0;
12217 // Look at each chain and determine if it is an alias. If so, add it to the
12218 // aliases list. If not, then continue up the chain looking for the next
12220 while (!Chains.empty()) {
12221 SDValue Chain = Chains.back();
12224 // For TokenFactor nodes, look at each operand and only continue up the
12225 // chain until we find two aliases. If we've seen two aliases, assume we'll
12226 // find more and revert to original chain since the xform is unlikely to be
12229 // FIXME: The depth check could be made to return the last non-aliasing
12230 // chain we found before we hit a tokenfactor rather than the original
12232 if (Depth > 6 || Aliases.size() == 2) {
12234 Aliases.push_back(OriginalChain);
12238 // Don't bother if we've been before.
12239 if (!Visited.insert(Chain.getNode()))
12242 switch (Chain.getOpcode()) {
12243 case ISD::EntryToken:
12244 // Entry token is ideal chain operand, but handled in FindBetterChain.
12249 // Get alias information for Chain.
12250 bool IsOpLoad = isa<LoadSDNode>(Chain.getNode()) &&
12251 !cast<LSBaseSDNode>(Chain.getNode())->isVolatile();
12253 // If chain is alias then stop here.
12254 if (!(IsLoad && IsOpLoad) &&
12255 isAlias(cast<LSBaseSDNode>(N), cast<LSBaseSDNode>(Chain.getNode()))) {
12256 Aliases.push_back(Chain);
12258 // Look further up the chain.
12259 Chains.push_back(Chain.getOperand(0));
12265 case ISD::TokenFactor:
12266 // We have to check each of the operands of the token factor for "small"
12267 // token factors, so we queue them up. Adding the operands to the queue
12268 // (stack) in reverse order maintains the original order and increases the
12269 // likelihood that getNode will find a matching token factor (CSE.)
12270 if (Chain.getNumOperands() > 16) {
12271 Aliases.push_back(Chain);
12274 for (unsigned n = Chain.getNumOperands(); n;)
12275 Chains.push_back(Chain.getOperand(--n));
12280 // For all other instructions we will just have to take what we can get.
12281 Aliases.push_back(Chain);
12286 // We need to be careful here to also search for aliases through the
12287 // value operand of a store, etc. Consider the following situation:
12289 // L1 = load Token1, %52
12290 // S1 = store Token1, L1, %51
12291 // L2 = load Token1, %52+8
12292 // S2 = store Token1, L2, %51+8
12293 // Token2 = Token(S1, S2)
12294 // L3 = load Token2, %53
12295 // S3 = store Token2, L3, %52
12296 // L4 = load Token2, %53+8
12297 // S4 = store Token2, L4, %52+8
12298 // If we search for aliases of S3 (which loads address %52), and we look
12299 // only through the chain, then we'll miss the trivial dependence on L1
12300 // (which also loads from %52). We then might change all loads and
12301 // stores to use Token1 as their chain operand, which could result in
12302 // copying %53 into %52 before copying %52 into %51 (which should
12305 // The problem is, however, that searching for such data dependencies
12306 // can become expensive, and the cost is not directly related to the
12307 // chain depth. Instead, we'll rule out such configurations here by
12308 // insisting that we've visited all chain users (except for users
12309 // of the original chain, which is not necessary). When doing this,
12310 // we need to look through nodes we don't care about (otherwise, things
12311 // like register copies will interfere with trivial cases).
12313 SmallVector<const SDNode *, 16> Worklist;
12314 for (const SDNode *N : Visited)
12315 if (N != OriginalChain.getNode())
12316 Worklist.push_back(N);
12318 while (!Worklist.empty()) {
12319 const SDNode *M = Worklist.pop_back_val();
12321 // We have already visited M, and want to make sure we've visited any uses
12322 // of M that we care about. For uses that we've not visisted, and don't
12323 // care about, queue them to the worklist.
12325 for (SDNode::use_iterator UI = M->use_begin(),
12326 UIE = M->use_end(); UI != UIE; ++UI)
12327 if (UI.getUse().getValueType() == MVT::Other && Visited.insert(*UI)) {
12328 if (isa<MemIntrinsicSDNode>(*UI) || isa<MemSDNode>(*UI)) {
12329 // We've not visited this use, and we care about it (it could have an
12330 // ordering dependency with the original node).
12332 Aliases.push_back(OriginalChain);
12336 // We've not visited this use, but we don't care about it. Mark it as
12337 // visited and enqueue it to the worklist.
12338 Worklist.push_back(*UI);
12343 /// Walk up chain skipping non-aliasing memory nodes, looking for a better chain
12344 /// (aliasing node.)
12345 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
12346 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
12348 // Accumulate all the aliases to this node.
12349 GatherAllAliases(N, OldChain, Aliases);
12351 // If no operands then chain to entry token.
12352 if (Aliases.size() == 0)
12353 return DAG.getEntryNode();
12355 // If a single operand then chain to it. We don't need to revisit it.
12356 if (Aliases.size() == 1)
12359 // Construct a custom tailored token factor.
12360 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Aliases);
12363 /// This is the entry point for the file.
12364 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
12365 CodeGenOpt::Level OptLevel) {
12366 /// This is the main entry point to this class.
12367 DAGCombiner(*this, AA, OptLevel).Run(Level);