1 //===-- DAGCombiner.cpp - Implement a DAG node combiner -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass combines dag nodes to form fewer, simpler DAG nodes. It can be run
11 // both before and after the DAG is legalized.
13 // This pass is not a substitute for the LLVM IR instcombine pass. This pass is
14 // primarily intended to handle simplification opportunities that are implicit
15 // in the LLVM IR and exposed by the various codegen lowering phases.
17 //===----------------------------------------------------------------------===//
19 #include "llvm/CodeGen/SelectionDAG.h"
20 #include "llvm/ADT/SetVector.h"
21 #include "llvm/ADT/SmallBitVector.h"
22 #include "llvm/ADT/SmallPtrSet.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/Analysis/AliasAnalysis.h"
25 #include "llvm/CodeGen/MachineFrameInfo.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/DerivedTypes.h"
29 #include "llvm/IR/Function.h"
30 #include "llvm/IR/LLVMContext.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/ErrorHandling.h"
34 #include "llvm/Support/MathExtras.h"
35 #include "llvm/Support/raw_ostream.h"
36 #include "llvm/Target/TargetLowering.h"
37 #include "llvm/Target/TargetOptions.h"
38 #include "llvm/Target/TargetRegisterInfo.h"
39 #include "llvm/Target/TargetSubtargetInfo.h"
43 #define DEBUG_TYPE "dagcombine"
45 STATISTIC(NodesCombined , "Number of dag nodes combined");
46 STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created");
47 STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created");
48 STATISTIC(OpsNarrowed , "Number of load/op/store narrowed");
49 STATISTIC(LdStFP2Int , "Number of fp load/store pairs transformed to int");
50 STATISTIC(SlicedLoads, "Number of load sliced");
54 CombinerAA("combiner-alias-analysis", cl::Hidden,
55 cl::desc("Enable DAG combiner alias-analysis heuristics"));
58 CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden,
59 cl::desc("Enable DAG combiner's use of IR alias analysis"));
62 UseTBAA("combiner-use-tbaa", cl::Hidden, cl::init(true),
63 cl::desc("Enable DAG combiner's use of TBAA"));
66 static cl::opt<std::string>
67 CombinerAAOnlyFunc("combiner-aa-only-func", cl::Hidden,
68 cl::desc("Only use DAG-combiner alias analysis in this"
72 /// Hidden option to stress test load slicing, i.e., when this option
73 /// is enabled, load slicing bypasses most of its profitability guards.
75 StressLoadSlicing("combiner-stress-load-slicing", cl::Hidden,
76 cl::desc("Bypass the profitability model of load "
81 MaySplitLoadIndex("combiner-split-load-index", cl::Hidden, cl::init(true),
82 cl::desc("DAG combiner may split indexing from loads"));
84 //------------------------------ DAGCombiner ---------------------------------//
88 const TargetLowering &TLI;
90 CodeGenOpt::Level OptLevel;
95 /// \brief Worklist of all of the nodes that need to be simplified.
97 /// This must behave as a stack -- new nodes to process are pushed onto the
98 /// back and when processing we pop off of the back.
100 /// The worklist will not contain duplicates but may contain null entries
101 /// due to nodes being deleted from the underlying DAG.
102 SmallVector<SDNode *, 64> Worklist;
104 /// \brief Mapping from an SDNode to its position on the worklist.
106 /// This is used to find and remove nodes from the worklist (by nulling
107 /// them) when they are deleted from the underlying DAG. It relies on
108 /// stable indices of nodes within the worklist.
109 DenseMap<SDNode *, unsigned> WorklistMap;
111 /// \brief Set of nodes which have been combined (at least once).
113 /// This is used to allow us to reliably add any operands of a DAG node
114 /// which have not yet been combined to the worklist.
115 SmallPtrSet<SDNode *, 64> CombinedNodes;
117 // AA - Used for DAG load/store alias analysis.
120 /// When an instruction is simplified, add all users of the instruction to
121 /// the work lists because they might get more simplified now.
122 void AddUsersToWorklist(SDNode *N) {
123 for (SDNode *Node : N->uses())
127 /// Call the node-specific routine that folds each particular type of node.
128 SDValue visit(SDNode *N);
131 /// Add to the worklist making sure its instance is at the back (next to be
133 void AddToWorklist(SDNode *N) {
134 // Skip handle nodes as they can't usefully be combined and confuse the
135 // zero-use deletion strategy.
136 if (N->getOpcode() == ISD::HANDLENODE)
139 if (WorklistMap.insert(std::make_pair(N, Worklist.size())).second)
140 Worklist.push_back(N);
143 /// Remove all instances of N from the worklist.
144 void removeFromWorklist(SDNode *N) {
145 CombinedNodes.erase(N);
147 auto It = WorklistMap.find(N);
148 if (It == WorklistMap.end())
149 return; // Not in the worklist.
151 // Null out the entry rather than erasing it to avoid a linear operation.
152 Worklist[It->second] = nullptr;
153 WorklistMap.erase(It);
156 void deleteAndRecombine(SDNode *N);
157 bool recursivelyDeleteUnusedNodes(SDNode *N);
159 SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
162 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) {
163 return CombineTo(N, &Res, 1, AddTo);
166 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1,
168 SDValue To[] = { Res0, Res1 };
169 return CombineTo(N, To, 2, AddTo);
172 void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO);
176 /// Check the specified integer node value to see if it can be simplified or
177 /// if things it uses can be simplified by bit propagation.
178 /// If so, return true.
179 bool SimplifyDemandedBits(SDValue Op) {
180 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
181 APInt Demanded = APInt::getAllOnesValue(BitWidth);
182 return SimplifyDemandedBits(Op, Demanded);
185 bool SimplifyDemandedBits(SDValue Op, const APInt &Demanded);
187 bool CombineToPreIndexedLoadStore(SDNode *N);
188 bool CombineToPostIndexedLoadStore(SDNode *N);
189 SDValue SplitIndexingFromLoad(LoadSDNode *LD);
190 bool SliceUpLoad(SDNode *N);
192 /// \brief Replace an ISD::EXTRACT_VECTOR_ELT of a load with a narrowed
195 /// \param EVE ISD::EXTRACT_VECTOR_ELT to be replaced.
196 /// \param InVecVT type of the input vector to EVE with bitcasts resolved.
197 /// \param EltNo index of the vector element to load.
198 /// \param OriginalLoad load that EVE came from to be replaced.
199 /// \returns EVE on success SDValue() on failure.
200 SDValue ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
201 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad);
202 void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad);
203 SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace);
204 SDValue SExtPromoteOperand(SDValue Op, EVT PVT);
205 SDValue ZExtPromoteOperand(SDValue Op, EVT PVT);
206 SDValue PromoteIntBinOp(SDValue Op);
207 SDValue PromoteIntShiftOp(SDValue Op);
208 SDValue PromoteExtend(SDValue Op);
209 bool PromoteLoad(SDValue Op);
211 void ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
212 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
213 ISD::NodeType ExtType);
215 /// Call the node-specific routine that knows how to fold each
216 /// particular type of node. If that doesn't do anything, try the
217 /// target-specific DAG combines.
218 SDValue combine(SDNode *N);
220 // Visitation implementation - Implement dag node combining for different
221 // node types. The semantics are as follows:
223 // SDValue.getNode() == 0 - No change was made
224 // SDValue.getNode() == N - N was replaced, is dead and has been handled.
225 // otherwise - N should be replaced by the returned Operand.
227 SDValue visitTokenFactor(SDNode *N);
228 SDValue visitMERGE_VALUES(SDNode *N);
229 SDValue visitADD(SDNode *N);
230 SDValue visitSUB(SDNode *N);
231 SDValue visitADDC(SDNode *N);
232 SDValue visitSUBC(SDNode *N);
233 SDValue visitADDE(SDNode *N);
234 SDValue visitSUBE(SDNode *N);
235 SDValue visitMUL(SDNode *N);
236 SDValue visitSDIV(SDNode *N);
237 SDValue visitUDIV(SDNode *N);
238 SDValue visitSREM(SDNode *N);
239 SDValue visitUREM(SDNode *N);
240 SDValue visitMULHU(SDNode *N);
241 SDValue visitMULHS(SDNode *N);
242 SDValue visitSMUL_LOHI(SDNode *N);
243 SDValue visitUMUL_LOHI(SDNode *N);
244 SDValue visitSMULO(SDNode *N);
245 SDValue visitUMULO(SDNode *N);
246 SDValue visitSDIVREM(SDNode *N);
247 SDValue visitUDIVREM(SDNode *N);
248 SDValue visitAND(SDNode *N);
249 SDValue 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);
306 SDValue visitMLOAD(SDNode *N);
307 SDValue visitMSTORE(SDNode *N);
309 SDValue XformToShuffleWithZero(SDNode *N);
310 SDValue ReassociateOps(unsigned Opc, SDLoc DL, SDValue LHS, SDValue RHS);
312 SDValue visitShiftByConstant(SDNode *N, ConstantSDNode *Amt);
314 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
315 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
316 SDValue SimplifySelect(SDLoc DL, SDValue N0, SDValue N1, SDValue N2);
317 SDValue SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1, SDValue N2,
318 SDValue N3, ISD::CondCode CC,
319 bool NotExtCompare = false);
320 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
321 SDLoc DL, bool foldBooleans = true);
323 bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
325 bool isOneUseSetCC(SDValue N) const;
327 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
329 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
330 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
331 SDValue BuildSDIV(SDNode *N);
332 SDValue BuildSDIVPow2(SDNode *N);
333 SDValue BuildUDIV(SDNode *N);
334 SDValue BuildReciprocalEstimate(SDValue Op);
335 SDValue BuildRsqrtEstimate(SDValue Op);
336 SDValue BuildRsqrtNROneConst(SDValue Op, SDValue Est, unsigned Iterations);
337 SDValue BuildRsqrtNRTwoConst(SDValue Op, SDValue Est, unsigned Iterations);
338 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
339 bool DemandHighBits = true);
340 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
341 SDNode *MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg,
342 SDValue InnerPos, SDValue InnerNeg,
343 unsigned PosOpcode, unsigned NegOpcode,
345 SDNode *MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL);
346 SDValue ReduceLoadWidth(SDNode *N);
347 SDValue ReduceLoadOpStoreWidth(SDNode *N);
348 SDValue TransformFPLoadStorePair(SDNode *N);
349 SDValue reduceBuildVecExtToExtBuildVec(SDNode *N);
350 SDValue reduceBuildVecConvertToConvertBuildVec(SDNode *N);
352 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
354 /// Walk up chain skipping non-aliasing memory nodes,
355 /// looking for aliasing nodes and adding them to the Aliases vector.
356 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
357 SmallVectorImpl<SDValue> &Aliases);
359 /// Return true if there is any possibility that the two addresses overlap.
360 bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const;
362 /// Walk up chain skipping non-aliasing memory nodes, looking for a better
363 /// chain (aliasing node.)
364 SDValue FindBetterChain(SDNode *N, SDValue Chain);
366 /// Merge consecutive store operations into a wide store.
367 /// This optimization uses wide integers or vectors when possible.
368 /// \return True if some memory operations were changed.
369 bool MergeConsecutiveStores(StoreSDNode *N);
371 /// \brief Try to transform a truncation where C is a constant:
372 /// (trunc (and X, C)) -> (and (trunc X), (trunc C))
374 /// \p N needs to be a truncation and its first operand an AND. Other
375 /// requirements are checked by the function (e.g. that trunc is
376 /// single-use) and if missed an empty SDValue is returned.
377 SDValue distributeTruncateThroughAnd(SDNode *N);
380 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
381 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
382 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {
383 AttributeSet FnAttrs =
384 DAG.getMachineFunction().getFunction()->getAttributes();
386 FnAttrs.hasAttribute(AttributeSet::FunctionIndex,
387 Attribute::OptimizeForSize) ||
388 FnAttrs.hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize);
391 /// Runs the dag combiner on all nodes in the work list
392 void Run(CombineLevel AtLevel);
394 SelectionDAG &getDAG() const { return DAG; }
396 /// Returns a type large enough to hold any valid shift amount - before type
397 /// legalization these can be huge.
398 EVT getShiftAmountTy(EVT LHSTy) {
399 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
400 if (LHSTy.isVector())
402 return LegalTypes ? TLI.getScalarShiftAmountTy(LHSTy)
403 : TLI.getPointerTy();
406 /// This method returns true if we are running before type legalization or
407 /// if the specified VT is legal.
408 bool isTypeLegal(const EVT &VT) {
409 if (!LegalTypes) return true;
410 return TLI.isTypeLegal(VT);
413 /// Convenience wrapper around TargetLowering::getSetCCResultType
414 EVT getSetCCResultType(EVT VT) const {
415 return TLI.getSetCCResultType(*DAG.getContext(), VT);
422 /// This class is a DAGUpdateListener that removes any deleted
423 /// nodes from the worklist.
424 class WorklistRemover : public SelectionDAG::DAGUpdateListener {
427 explicit WorklistRemover(DAGCombiner &dc)
428 : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {}
430 void NodeDeleted(SDNode *N, SDNode *E) override {
431 DC.removeFromWorklist(N);
436 //===----------------------------------------------------------------------===//
437 // TargetLowering::DAGCombinerInfo implementation
438 //===----------------------------------------------------------------------===//
440 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
441 ((DAGCombiner*)DC)->AddToWorklist(N);
444 void TargetLowering::DAGCombinerInfo::RemoveFromWorklist(SDNode *N) {
445 ((DAGCombiner*)DC)->removeFromWorklist(N);
448 SDValue TargetLowering::DAGCombinerInfo::
449 CombineTo(SDNode *N, const std::vector<SDValue> &To, bool AddTo) {
450 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
453 SDValue TargetLowering::DAGCombinerInfo::
454 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
455 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
459 SDValue TargetLowering::DAGCombinerInfo::
460 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
461 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
464 void TargetLowering::DAGCombinerInfo::
465 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
466 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
469 //===----------------------------------------------------------------------===//
471 //===----------------------------------------------------------------------===//
473 void DAGCombiner::deleteAndRecombine(SDNode *N) {
474 removeFromWorklist(N);
476 // If the operands of this node are only used by the node, they will now be
477 // dead. Make sure to re-visit them and recursively delete dead nodes.
478 for (const SDValue &Op : N->ops())
479 // For an operand generating multiple values, one of the values may
480 // become dead allowing further simplification (e.g. split index
481 // arithmetic from an indexed load).
482 if (Op->hasOneUse() || Op->getNumValues() > 1)
483 AddToWorklist(Op.getNode());
488 /// Return 1 if we can compute the negated form of the specified expression for
489 /// the same cost as the expression itself, or 2 if we can compute the negated
490 /// form more cheaply than the expression itself.
491 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
492 const TargetLowering &TLI,
493 const TargetOptions *Options,
494 unsigned Depth = 0) {
495 // fneg is removable even if it has multiple uses.
496 if (Op.getOpcode() == ISD::FNEG) return 2;
498 // Don't allow anything with multiple uses.
499 if (!Op.hasOneUse()) return 0;
501 // Don't recurse exponentially.
502 if (Depth > 6) return 0;
504 switch (Op.getOpcode()) {
505 default: return false;
506 case ISD::ConstantFP:
507 // Don't invert constant FP values after legalize. The negated constant
508 // isn't necessarily legal.
509 return LegalOperations ? 0 : 1;
511 // FIXME: determine better conditions for this xform.
512 if (!Options->UnsafeFPMath) return 0;
514 // After operation legalization, it might not be legal to create new FSUBs.
515 if (LegalOperations &&
516 !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType()))
519 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
520 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
523 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
524 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
527 // We can't turn -(A-B) into B-A when we honor signed zeros.
528 if (!Options->UnsafeFPMath) return 0;
530 // fold (fneg (fsub A, B)) -> (fsub B, A)
535 if (Options->HonorSignDependentRoundingFPMath()) return 0;
537 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
538 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
542 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
548 return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options,
553 /// If isNegatibleForFree returns true, return the newly negated expression.
554 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
555 bool LegalOperations, unsigned Depth = 0) {
556 const TargetOptions &Options = DAG.getTarget().Options;
557 // fneg is removable even if it has multiple uses.
558 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
560 // Don't allow anything with multiple uses.
561 assert(Op.hasOneUse() && "Unknown reuse!");
563 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
564 switch (Op.getOpcode()) {
565 default: llvm_unreachable("Unknown code");
566 case ISD::ConstantFP: {
567 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
569 return DAG.getConstantFP(V, Op.getValueType());
572 // FIXME: determine better conditions for this xform.
573 assert(Options.UnsafeFPMath);
575 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
576 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
577 DAG.getTargetLoweringInfo(), &Options, Depth+1))
578 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
579 GetNegatedExpression(Op.getOperand(0), DAG,
580 LegalOperations, Depth+1),
582 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
583 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
584 GetNegatedExpression(Op.getOperand(1), DAG,
585 LegalOperations, Depth+1),
588 // We can't turn -(A-B) into B-A when we honor signed zeros.
589 assert(Options.UnsafeFPMath);
591 // fold (fneg (fsub 0, B)) -> B
592 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
593 if (N0CFP->getValueAPF().isZero())
594 return Op.getOperand(1);
596 // fold (fneg (fsub A, B)) -> (fsub B, A)
597 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
598 Op.getOperand(1), Op.getOperand(0));
602 assert(!Options.HonorSignDependentRoundingFPMath());
604 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
605 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
606 DAG.getTargetLoweringInfo(), &Options, Depth+1))
607 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
608 GetNegatedExpression(Op.getOperand(0), DAG,
609 LegalOperations, Depth+1),
612 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
613 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
615 GetNegatedExpression(Op.getOperand(1), DAG,
616 LegalOperations, Depth+1));
620 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
621 GetNegatedExpression(Op.getOperand(0), DAG,
622 LegalOperations, Depth+1));
624 return DAG.getNode(ISD::FP_ROUND, SDLoc(Op), Op.getValueType(),
625 GetNegatedExpression(Op.getOperand(0), DAG,
626 LegalOperations, Depth+1),
631 // Return true if this node is a setcc, or is a select_cc
632 // that selects between the target values used for true and false, making it
633 // equivalent to a setcc. Also, set the incoming LHS, RHS, and CC references to
634 // the appropriate nodes based on the type of node we are checking. This
635 // simplifies life a bit for the callers.
636 bool DAGCombiner::isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
638 if (N.getOpcode() == ISD::SETCC) {
639 LHS = N.getOperand(0);
640 RHS = N.getOperand(1);
641 CC = N.getOperand(2);
645 if (N.getOpcode() != ISD::SELECT_CC ||
646 !TLI.isConstTrueVal(N.getOperand(2).getNode()) ||
647 !TLI.isConstFalseVal(N.getOperand(3).getNode()))
650 if (TLI.getBooleanContents(N.getValueType()) ==
651 TargetLowering::UndefinedBooleanContent)
654 LHS = N.getOperand(0);
655 RHS = N.getOperand(1);
656 CC = N.getOperand(4);
660 /// Return true if this is a SetCC-equivalent operation with only one use.
661 /// If this is true, it allows the users to invert the operation for free when
662 /// it is profitable to do so.
663 bool DAGCombiner::isOneUseSetCC(SDValue N) const {
665 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
670 /// Returns true if N is a BUILD_VECTOR node whose
671 /// elements are all the same constant or undefined.
672 static bool isConstantSplatVector(SDNode *N, APInt& SplatValue) {
673 BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(N);
678 unsigned SplatBitSize;
680 EVT EltVT = N->getValueType(0).getVectorElementType();
681 return (C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
683 EltVT.getSizeInBits() >= SplatBitSize);
686 // \brief Returns the SDNode if it is a constant BuildVector or constant.
687 static SDNode *isConstantBuildVectorOrConstantInt(SDValue N) {
688 if (isa<ConstantSDNode>(N))
690 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N);
691 if (BV && BV->isConstant())
696 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
698 static ConstantSDNode *isConstOrConstSplat(SDValue N) {
699 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N))
702 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
703 BitVector UndefElements;
704 ConstantSDNode *CN = BV->getConstantSplatNode(&UndefElements);
706 // BuildVectors can truncate their operands. Ignore that case here.
707 // FIXME: We blindly ignore splats which include undef which is overly
709 if (CN && UndefElements.none() &&
710 CN->getValueType(0) == N.getValueType().getScalarType())
717 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
719 static ConstantFPSDNode *isConstOrConstSplatFP(SDValue N) {
720 if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N))
723 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N)) {
724 BitVector UndefElements;
725 ConstantFPSDNode *CN = BV->getConstantFPSplatNode(&UndefElements);
727 if (CN && UndefElements.none())
734 SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL,
735 SDValue N0, SDValue N1) {
736 EVT VT = N0.getValueType();
737 if (N0.getOpcode() == Opc) {
738 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0.getOperand(1))) {
739 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1)) {
740 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
741 if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, L, R))
742 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
745 if (N0.hasOneUse()) {
746 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one
748 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1);
749 if (!OpNode.getNode())
751 AddToWorklist(OpNode.getNode());
752 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
757 if (N1.getOpcode() == Opc) {
758 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1.getOperand(1))) {
759 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0)) {
760 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
761 if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, R, L))
762 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
765 if (N1.hasOneUse()) {
766 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one
768 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N1.getOperand(0), N0);
769 if (!OpNode.getNode())
771 AddToWorklist(OpNode.getNode());
772 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
780 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
782 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
784 DEBUG(dbgs() << "\nReplacing.1 ";
786 dbgs() << "\nWith: ";
787 To[0].getNode()->dump(&DAG);
788 dbgs() << " and " << NumTo-1 << " other values\n");
789 for (unsigned i = 0, e = NumTo; i != e; ++i)
790 assert((!To[i].getNode() ||
791 N->getValueType(i) == To[i].getValueType()) &&
792 "Cannot combine value to value of different type!");
794 WorklistRemover DeadNodes(*this);
795 DAG.ReplaceAllUsesWith(N, To);
797 // Push the new nodes and any users onto the worklist
798 for (unsigned i = 0, e = NumTo; i != e; ++i) {
799 if (To[i].getNode()) {
800 AddToWorklist(To[i].getNode());
801 AddUsersToWorklist(To[i].getNode());
806 // Finally, if the node is now dead, remove it from the graph. The node
807 // may not be dead if the replacement process recursively simplified to
808 // something else needing this node.
810 deleteAndRecombine(N);
811 return SDValue(N, 0);
815 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
816 // Replace all uses. If any nodes become isomorphic to other nodes and
817 // are deleted, make sure to remove them from our worklist.
818 WorklistRemover DeadNodes(*this);
819 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New);
821 // Push the new node and any (possibly new) users onto the worklist.
822 AddToWorklist(TLO.New.getNode());
823 AddUsersToWorklist(TLO.New.getNode());
825 // Finally, if the node is now dead, remove it from the graph. The node
826 // may not be dead if the replacement process recursively simplified to
827 // something else needing this node.
828 if (TLO.Old.getNode()->use_empty())
829 deleteAndRecombine(TLO.Old.getNode());
832 /// Check the specified integer node value to see if it can be simplified or if
833 /// things it uses can be simplified by bit propagation. If so, return true.
834 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
835 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
836 APInt KnownZero, KnownOne;
837 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
841 AddToWorklist(Op.getNode());
843 // Replace the old value with the new one.
845 DEBUG(dbgs() << "\nReplacing.2 ";
846 TLO.Old.getNode()->dump(&DAG);
847 dbgs() << "\nWith: ";
848 TLO.New.getNode()->dump(&DAG);
851 CommitTargetLoweringOpt(TLO);
855 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
857 EVT VT = Load->getValueType(0);
858 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
860 DEBUG(dbgs() << "\nReplacing.9 ";
862 dbgs() << "\nWith: ";
863 Trunc.getNode()->dump(&DAG);
865 WorklistRemover DeadNodes(*this);
866 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc);
867 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1));
868 deleteAndRecombine(Load);
869 AddToWorklist(Trunc.getNode());
872 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
875 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
876 EVT MemVT = LD->getMemoryVT();
877 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
878 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, PVT, MemVT) ? ISD::ZEXTLOAD
880 : LD->getExtensionType();
882 return DAG.getExtLoad(ExtType, dl, PVT,
883 LD->getChain(), LD->getBasePtr(),
884 MemVT, LD->getMemOperand());
887 unsigned Opc = Op.getOpcode();
890 case ISD::AssertSext:
891 return DAG.getNode(ISD::AssertSext, dl, PVT,
892 SExtPromoteOperand(Op.getOperand(0), PVT),
894 case ISD::AssertZext:
895 return DAG.getNode(ISD::AssertZext, dl, PVT,
896 ZExtPromoteOperand(Op.getOperand(0), PVT),
898 case ISD::Constant: {
900 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
901 return DAG.getNode(ExtOpc, dl, PVT, Op);
905 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
907 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
910 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
911 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
913 EVT OldVT = Op.getValueType();
915 bool Replace = false;
916 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
917 if (!NewOp.getNode())
919 AddToWorklist(NewOp.getNode());
922 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
923 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
924 DAG.getValueType(OldVT));
927 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
928 EVT OldVT = Op.getValueType();
930 bool Replace = false;
931 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
932 if (!NewOp.getNode())
934 AddToWorklist(NewOp.getNode());
937 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
938 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
941 /// Promote the specified integer binary operation if the target indicates it is
942 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
943 /// i32 since i16 instructions are longer.
944 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
945 if (!LegalOperations)
948 EVT VT = Op.getValueType();
949 if (VT.isVector() || !VT.isInteger())
952 // If operation type is 'undesirable', e.g. i16 on x86, consider
954 unsigned Opc = Op.getOpcode();
955 if (TLI.isTypeDesirableForOp(Opc, VT))
959 // Consult target whether it is a good idea to promote this operation and
960 // what's the right type to promote it to.
961 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
962 assert(PVT != VT && "Don't know what type to promote to!");
964 bool Replace0 = false;
965 SDValue N0 = Op.getOperand(0);
966 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
970 bool Replace1 = false;
971 SDValue N1 = Op.getOperand(1);
976 NN1 = PromoteOperand(N1, PVT, Replace1);
981 AddToWorklist(NN0.getNode());
983 AddToWorklist(NN1.getNode());
986 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
988 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
990 DEBUG(dbgs() << "\nPromoting ";
991 Op.getNode()->dump(&DAG));
993 return DAG.getNode(ISD::TRUNCATE, dl, VT,
994 DAG.getNode(Opc, dl, PVT, NN0, NN1));
999 /// Promote the specified integer shift operation if the target indicates it is
1000 /// beneficial. e.g. On x86, it's usually better to promote i16 operations to
1001 /// i32 since i16 instructions are longer.
1002 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
1003 if (!LegalOperations)
1006 EVT VT = Op.getValueType();
1007 if (VT.isVector() || !VT.isInteger())
1010 // If operation type is 'undesirable', e.g. i16 on x86, consider
1012 unsigned Opc = Op.getOpcode();
1013 if (TLI.isTypeDesirableForOp(Opc, VT))
1017 // Consult target whether it is a good idea to promote this operation and
1018 // what's the right type to promote it to.
1019 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1020 assert(PVT != VT && "Don't know what type to promote to!");
1022 bool Replace = false;
1023 SDValue N0 = Op.getOperand(0);
1024 if (Opc == ISD::SRA)
1025 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
1026 else if (Opc == ISD::SRL)
1027 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
1029 N0 = PromoteOperand(N0, PVT, Replace);
1033 AddToWorklist(N0.getNode());
1035 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
1037 DEBUG(dbgs() << "\nPromoting ";
1038 Op.getNode()->dump(&DAG));
1040 return DAG.getNode(ISD::TRUNCATE, dl, VT,
1041 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
1046 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
1047 if (!LegalOperations)
1050 EVT VT = Op.getValueType();
1051 if (VT.isVector() || !VT.isInteger())
1054 // If operation type is 'undesirable', e.g. i16 on x86, consider
1056 unsigned Opc = Op.getOpcode();
1057 if (TLI.isTypeDesirableForOp(Opc, VT))
1061 // Consult target whether it is a good idea to promote this operation and
1062 // what's the right type to promote it to.
1063 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1064 assert(PVT != VT && "Don't know what type to promote to!");
1065 // fold (aext (aext x)) -> (aext x)
1066 // fold (aext (zext x)) -> (zext x)
1067 // fold (aext (sext x)) -> (sext x)
1068 DEBUG(dbgs() << "\nPromoting ";
1069 Op.getNode()->dump(&DAG));
1070 return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0));
1075 bool DAGCombiner::PromoteLoad(SDValue Op) {
1076 if (!LegalOperations)
1079 EVT VT = Op.getValueType();
1080 if (VT.isVector() || !VT.isInteger())
1083 // If operation type is 'undesirable', e.g. i16 on x86, consider
1085 unsigned Opc = Op.getOpcode();
1086 if (TLI.isTypeDesirableForOp(Opc, VT))
1090 // Consult target whether it is a good idea to promote this operation and
1091 // what's the right type to promote it to.
1092 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1093 assert(PVT != VT && "Don't know what type to promote to!");
1096 SDNode *N = Op.getNode();
1097 LoadSDNode *LD = cast<LoadSDNode>(N);
1098 EVT MemVT = LD->getMemoryVT();
1099 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
1100 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, PVT, MemVT) ? ISD::ZEXTLOAD
1102 : LD->getExtensionType();
1103 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
1104 LD->getChain(), LD->getBasePtr(),
1105 MemVT, LD->getMemOperand());
1106 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
1108 DEBUG(dbgs() << "\nPromoting ";
1111 Result.getNode()->dump(&DAG);
1113 WorklistRemover DeadNodes(*this);
1114 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
1115 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1));
1116 deleteAndRecombine(N);
1117 AddToWorklist(Result.getNode());
1123 /// \brief Recursively delete a node which has no uses and any operands for
1124 /// which it is the only use.
1126 /// Note that this both deletes the nodes and removes them from the worklist.
1127 /// It also adds any nodes who have had a user deleted to the worklist as they
1128 /// may now have only one use and subject to other combines.
1129 bool DAGCombiner::recursivelyDeleteUnusedNodes(SDNode *N) {
1130 if (!N->use_empty())
1133 SmallSetVector<SDNode *, 16> Nodes;
1136 N = Nodes.pop_back_val();
1140 if (N->use_empty()) {
1141 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1142 Nodes.insert(N->getOperand(i).getNode());
1144 removeFromWorklist(N);
1149 } while (!Nodes.empty());
1153 //===----------------------------------------------------------------------===//
1154 // Main DAG Combiner implementation
1155 //===----------------------------------------------------------------------===//
1157 void DAGCombiner::Run(CombineLevel AtLevel) {
1158 // set the instance variables, so that the various visit routines may use it.
1160 LegalOperations = Level >= AfterLegalizeVectorOps;
1161 LegalTypes = Level >= AfterLegalizeTypes;
1163 // Add all the dag nodes to the worklist.
1164 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
1165 E = DAG.allnodes_end(); I != E; ++I)
1168 // Create a dummy node (which is not added to allnodes), that adds a reference
1169 // to the root node, preventing it from being deleted, and tracking any
1170 // changes of the root.
1171 HandleSDNode Dummy(DAG.getRoot());
1173 // while the worklist isn't empty, find a node and
1174 // try and combine it.
1175 while (!WorklistMap.empty()) {
1177 // The Worklist holds the SDNodes in order, but it may contain null entries.
1179 N = Worklist.pop_back_val();
1182 bool GoodWorklistEntry = WorklistMap.erase(N);
1183 (void)GoodWorklistEntry;
1184 assert(GoodWorklistEntry &&
1185 "Found a worklist entry without a corresponding map entry!");
1187 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1188 // N is deleted from the DAG, since they too may now be dead or may have a
1189 // reduced number of uses, allowing other xforms.
1190 if (recursivelyDeleteUnusedNodes(N))
1193 WorklistRemover DeadNodes(*this);
1195 // If this combine is running after legalizing the DAG, re-legalize any
1196 // nodes pulled off the worklist.
1197 if (Level == AfterLegalizeDAG) {
1198 SmallSetVector<SDNode *, 16> UpdatedNodes;
1199 bool NIsValid = DAG.LegalizeOp(N, UpdatedNodes);
1201 for (SDNode *LN : UpdatedNodes) {
1203 AddUsersToWorklist(LN);
1209 DEBUG(dbgs() << "\nCombining: "; N->dump(&DAG));
1211 // Add any operands of the new node which have not yet been combined to the
1212 // worklist as well. Because the worklist uniques things already, this
1213 // won't repeatedly process the same operand.
1214 CombinedNodes.insert(N);
1215 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1216 if (!CombinedNodes.count(N->getOperand(i).getNode()))
1217 AddToWorklist(N->getOperand(i).getNode());
1219 SDValue RV = combine(N);
1226 // If we get back the same node we passed in, rather than a new node or
1227 // zero, we know that the node must have defined multiple values and
1228 // CombineTo was used. Since CombineTo takes care of the worklist
1229 // mechanics for us, we have no work to do in this case.
1230 if (RV.getNode() == N)
1233 assert(N->getOpcode() != ISD::DELETED_NODE &&
1234 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1235 "Node was deleted but visit returned new node!");
1237 DEBUG(dbgs() << " ... into: ";
1238 RV.getNode()->dump(&DAG));
1240 // Transfer debug value.
1241 DAG.TransferDbgValues(SDValue(N, 0), RV);
1242 if (N->getNumValues() == RV.getNode()->getNumValues())
1243 DAG.ReplaceAllUsesWith(N, RV.getNode());
1245 assert(N->getValueType(0) == RV.getValueType() &&
1246 N->getNumValues() == 1 && "Type mismatch");
1248 DAG.ReplaceAllUsesWith(N, &OpV);
1251 // Push the new node and any users onto the worklist
1252 AddToWorklist(RV.getNode());
1253 AddUsersToWorklist(RV.getNode());
1255 // Finally, if the node is now dead, remove it from the graph. The node
1256 // may not be dead if the replacement process recursively simplified to
1257 // something else needing this node. This will also take care of adding any
1258 // operands which have lost a user to the worklist.
1259 recursivelyDeleteUnusedNodes(N);
1262 // If the root changed (e.g. it was a dead load, update the root).
1263 DAG.setRoot(Dummy.getValue());
1264 DAG.RemoveDeadNodes();
1267 SDValue DAGCombiner::visit(SDNode *N) {
1268 switch (N->getOpcode()) {
1270 case ISD::TokenFactor: return visitTokenFactor(N);
1271 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1272 case ISD::ADD: return visitADD(N);
1273 case ISD::SUB: return visitSUB(N);
1274 case ISD::ADDC: return visitADDC(N);
1275 case ISD::SUBC: return visitSUBC(N);
1276 case ISD::ADDE: return visitADDE(N);
1277 case ISD::SUBE: return visitSUBE(N);
1278 case ISD::MUL: return visitMUL(N);
1279 case ISD::SDIV: return visitSDIV(N);
1280 case ISD::UDIV: return visitUDIV(N);
1281 case ISD::SREM: return visitSREM(N);
1282 case ISD::UREM: return visitUREM(N);
1283 case ISD::MULHU: return visitMULHU(N);
1284 case ISD::MULHS: return visitMULHS(N);
1285 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1286 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1287 case ISD::SMULO: return visitSMULO(N);
1288 case ISD::UMULO: return visitUMULO(N);
1289 case ISD::SDIVREM: return visitSDIVREM(N);
1290 case ISD::UDIVREM: return visitUDIVREM(N);
1291 case ISD::AND: return visitAND(N);
1292 case ISD::OR: return visitOR(N);
1293 case ISD::XOR: return visitXOR(N);
1294 case ISD::SHL: return visitSHL(N);
1295 case ISD::SRA: return visitSRA(N);
1296 case ISD::SRL: return visitSRL(N);
1298 case ISD::ROTL: return visitRotate(N);
1299 case ISD::CTLZ: return visitCTLZ(N);
1300 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1301 case ISD::CTTZ: return visitCTTZ(N);
1302 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1303 case ISD::CTPOP: return visitCTPOP(N);
1304 case ISD::SELECT: return visitSELECT(N);
1305 case ISD::VSELECT: return visitVSELECT(N);
1306 case ISD::SELECT_CC: return visitSELECT_CC(N);
1307 case ISD::SETCC: return visitSETCC(N);
1308 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1309 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1310 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1311 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1312 case ISD::TRUNCATE: return visitTRUNCATE(N);
1313 case ISD::BITCAST: return visitBITCAST(N);
1314 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1315 case ISD::FADD: return visitFADD(N);
1316 case ISD::FSUB: return visitFSUB(N);
1317 case ISD::FMUL: return visitFMUL(N);
1318 case ISD::FMA: return visitFMA(N);
1319 case ISD::FDIV: return visitFDIV(N);
1320 case ISD::FREM: return visitFREM(N);
1321 case ISD::FSQRT: return visitFSQRT(N);
1322 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1323 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1324 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1325 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1326 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1327 case ISD::FP_ROUND: return visitFP_ROUND(N);
1328 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1329 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1330 case ISD::FNEG: return visitFNEG(N);
1331 case ISD::FABS: return visitFABS(N);
1332 case ISD::FFLOOR: return visitFFLOOR(N);
1333 case ISD::FMINNUM: return visitFMINNUM(N);
1334 case ISD::FMAXNUM: return visitFMAXNUM(N);
1335 case ISD::FCEIL: return visitFCEIL(N);
1336 case ISD::FTRUNC: return visitFTRUNC(N);
1337 case ISD::BRCOND: return visitBRCOND(N);
1338 case ISD::BR_CC: return visitBR_CC(N);
1339 case ISD::LOAD: return visitLOAD(N);
1340 case ISD::STORE: return visitSTORE(N);
1341 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1342 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1343 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1344 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1345 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1346 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1347 case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N);
1348 case ISD::MLOAD: return visitMLOAD(N);
1349 case ISD::MSTORE: return visitMSTORE(N);
1354 SDValue DAGCombiner::combine(SDNode *N) {
1355 SDValue RV = visit(N);
1357 // If nothing happened, try a target-specific DAG combine.
1358 if (!RV.getNode()) {
1359 assert(N->getOpcode() != ISD::DELETED_NODE &&
1360 "Node was deleted but visit returned NULL!");
1362 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1363 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1365 // Expose the DAG combiner to the target combiner impls.
1366 TargetLowering::DAGCombinerInfo
1367 DagCombineInfo(DAG, Level, false, this);
1369 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1373 // If nothing happened still, try promoting the operation.
1374 if (!RV.getNode()) {
1375 switch (N->getOpcode()) {
1383 RV = PromoteIntBinOp(SDValue(N, 0));
1388 RV = PromoteIntShiftOp(SDValue(N, 0));
1390 case ISD::SIGN_EXTEND:
1391 case ISD::ZERO_EXTEND:
1392 case ISD::ANY_EXTEND:
1393 RV = PromoteExtend(SDValue(N, 0));
1396 if (PromoteLoad(SDValue(N, 0)))
1402 // If N is a commutative binary node, try commuting it to enable more
1404 if (!RV.getNode() && SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1405 N->getNumValues() == 1) {
1406 SDValue N0 = N->getOperand(0);
1407 SDValue N1 = N->getOperand(1);
1409 // Constant operands are canonicalized to RHS.
1410 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1411 SDValue Ops[] = {N1, N0};
1413 if (const BinaryWithFlagsSDNode *BinNode =
1414 dyn_cast<BinaryWithFlagsSDNode>(N)) {
1415 CSENode = DAG.getNodeIfExists(
1416 N->getOpcode(), N->getVTList(), Ops, BinNode->hasNoUnsignedWrap(),
1417 BinNode->hasNoSignedWrap(), BinNode->isExact());
1419 CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(), Ops);
1422 return SDValue(CSENode, 0);
1429 /// Given a node, return its input chain if it has one, otherwise return a null
1431 static SDValue getInputChainForNode(SDNode *N) {
1432 if (unsigned NumOps = N->getNumOperands()) {
1433 if (N->getOperand(0).getValueType() == MVT::Other)
1434 return N->getOperand(0);
1435 if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1436 return N->getOperand(NumOps-1);
1437 for (unsigned i = 1; i < NumOps-1; ++i)
1438 if (N->getOperand(i).getValueType() == MVT::Other)
1439 return N->getOperand(i);
1444 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1445 // If N has two operands, where one has an input chain equal to the other,
1446 // the 'other' chain is redundant.
1447 if (N->getNumOperands() == 2) {
1448 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1449 return N->getOperand(0);
1450 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1451 return N->getOperand(1);
1454 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1455 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1456 SmallPtrSet<SDNode*, 16> SeenOps;
1457 bool Changed = false; // If we should replace this token factor.
1459 // Start out with this token factor.
1462 // Iterate through token factors. The TFs grows when new token factors are
1464 for (unsigned i = 0; i < TFs.size(); ++i) {
1465 SDNode *TF = TFs[i];
1467 // Check each of the operands.
1468 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1469 SDValue Op = TF->getOperand(i);
1471 switch (Op.getOpcode()) {
1472 case ISD::EntryToken:
1473 // Entry tokens don't need to be added to the list. They are
1478 case ISD::TokenFactor:
1479 if (Op.hasOneUse() &&
1480 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1481 // Queue up for processing.
1482 TFs.push_back(Op.getNode());
1483 // Clean up in case the token factor is removed.
1484 AddToWorklist(Op.getNode());
1491 // Only add if it isn't already in the list.
1492 if (SeenOps.insert(Op.getNode()).second)
1503 // If we've change things around then replace token factor.
1506 // The entry token is the only possible outcome.
1507 Result = DAG.getEntryNode();
1509 // New and improved token factor.
1510 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Ops);
1513 // Don't add users to work list.
1514 return CombineTo(N, Result, false);
1520 /// MERGE_VALUES can always be eliminated.
1521 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1522 WorklistRemover DeadNodes(*this);
1523 // Replacing results may cause a different MERGE_VALUES to suddenly
1524 // be CSE'd with N, and carry its uses with it. Iterate until no
1525 // uses remain, to ensure that the node can be safely deleted.
1526 // First add the users of this node to the work list so that they
1527 // can be tried again once they have new operands.
1528 AddUsersToWorklist(N);
1530 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1531 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1532 } while (!N->use_empty());
1533 deleteAndRecombine(N);
1534 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1537 SDValue DAGCombiner::visitADD(SDNode *N) {
1538 SDValue N0 = N->getOperand(0);
1539 SDValue N1 = N->getOperand(1);
1540 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1541 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1542 EVT VT = N0.getValueType();
1545 if (VT.isVector()) {
1546 SDValue FoldedVOp = SimplifyVBinOp(N);
1547 if (FoldedVOp.getNode()) return FoldedVOp;
1549 // fold (add x, 0) -> x, vector edition
1550 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1552 if (ISD::isBuildVectorAllZeros(N0.getNode()))
1556 // fold (add x, undef) -> undef
1557 if (N0.getOpcode() == ISD::UNDEF)
1559 if (N1.getOpcode() == ISD::UNDEF)
1561 // fold (add c1, c2) -> c1+c2
1563 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1564 // canonicalize constant to RHS
1566 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, N0);
1567 // fold (add x, 0) -> x
1568 if (N1C && N1C->isNullValue())
1570 // fold (add Sym, c) -> Sym+c
1571 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1572 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1573 GA->getOpcode() == ISD::GlobalAddress)
1574 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1576 (uint64_t)N1C->getSExtValue());
1577 // fold ((c1-A)+c2) -> (c1+c2)-A
1578 if (N1C && N0.getOpcode() == ISD::SUB)
1579 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1580 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1581 DAG.getConstant(N1C->getAPIntValue()+
1582 N0C->getAPIntValue(), VT),
1585 SDValue RADD = ReassociateOps(ISD::ADD, SDLoc(N), N0, N1);
1588 // fold ((0-A) + B) -> B-A
1589 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1590 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1591 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, N0.getOperand(1));
1592 // fold (A + (0-B)) -> A-B
1593 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1594 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1595 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1.getOperand(1));
1596 // fold (A+(B-A)) -> B
1597 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1598 return N1.getOperand(0);
1599 // fold ((B-A)+A) -> B
1600 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1601 return N0.getOperand(0);
1602 // fold (A+(B-(A+C))) to (B-C)
1603 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1604 N0 == N1.getOperand(1).getOperand(0))
1605 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1606 N1.getOperand(1).getOperand(1));
1607 // fold (A+(B-(C+A))) to (B-C)
1608 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1609 N0 == N1.getOperand(1).getOperand(1))
1610 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1611 N1.getOperand(1).getOperand(0));
1612 // fold (A+((B-A)+or-C)) to (B+or-C)
1613 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1614 N1.getOperand(0).getOpcode() == ISD::SUB &&
1615 N0 == N1.getOperand(0).getOperand(1))
1616 return DAG.getNode(N1.getOpcode(), SDLoc(N), VT,
1617 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1619 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1620 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1621 SDValue N00 = N0.getOperand(0);
1622 SDValue N01 = N0.getOperand(1);
1623 SDValue N10 = N1.getOperand(0);
1624 SDValue N11 = N1.getOperand(1);
1626 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1627 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1628 DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10),
1629 DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11));
1632 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1633 return SDValue(N, 0);
1635 // fold (a+b) -> (a|b) iff a and b share no bits.
1636 if (VT.isInteger() && !VT.isVector()) {
1637 APInt LHSZero, LHSOne;
1638 APInt RHSZero, RHSOne;
1639 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1641 if (LHSZero.getBoolValue()) {
1642 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1644 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1645 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1646 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero){
1647 if (!LegalOperations || TLI.isOperationLegal(ISD::OR, VT))
1648 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1);
1653 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1654 if (N1.getOpcode() == ISD::SHL &&
1655 N1.getOperand(0).getOpcode() == ISD::SUB)
1656 if (ConstantSDNode *C =
1657 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1658 if (C->getAPIntValue() == 0)
1659 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0,
1660 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1661 N1.getOperand(0).getOperand(1),
1663 if (N0.getOpcode() == ISD::SHL &&
1664 N0.getOperand(0).getOpcode() == ISD::SUB)
1665 if (ConstantSDNode *C =
1666 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1667 if (C->getAPIntValue() == 0)
1668 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1,
1669 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1670 N0.getOperand(0).getOperand(1),
1673 if (N1.getOpcode() == ISD::AND) {
1674 SDValue AndOp0 = N1.getOperand(0);
1675 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1676 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1677 unsigned DestBits = VT.getScalarType().getSizeInBits();
1679 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1680 // and similar xforms where the inner op is either ~0 or 0.
1681 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1683 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1687 // add (sext i1), X -> sub X, (zext i1)
1688 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1689 N0.getOperand(0).getValueType() == MVT::i1 &&
1690 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1692 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1693 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1696 // add X, (sextinreg Y i1) -> sub X, (and Y 1)
1697 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
1698 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
1699 if (TN->getVT() == MVT::i1) {
1701 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
1702 DAG.getConstant(1, VT));
1703 return DAG.getNode(ISD::SUB, DL, VT, N0, ZExt);
1710 SDValue DAGCombiner::visitADDC(SDNode *N) {
1711 SDValue N0 = N->getOperand(0);
1712 SDValue N1 = N->getOperand(1);
1713 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1714 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1715 EVT VT = N0.getValueType();
1717 // If the flag result is dead, turn this into an ADD.
1718 if (!N->hasAnyUseOfValue(1))
1719 return CombineTo(N, DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N1),
1720 DAG.getNode(ISD::CARRY_FALSE,
1721 SDLoc(N), MVT::Glue));
1723 // canonicalize constant to RHS.
1725 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N1, N0);
1727 // fold (addc x, 0) -> x + no carry out
1728 if (N1C && N1C->isNullValue())
1729 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1730 SDLoc(N), MVT::Glue));
1732 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1733 APInt LHSZero, LHSOne;
1734 APInt RHSZero, RHSOne;
1735 DAG.computeKnownBits(N0, LHSZero, LHSOne);
1737 if (LHSZero.getBoolValue()) {
1738 DAG.computeKnownBits(N1, RHSZero, RHSOne);
1740 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1741 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1742 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1743 return CombineTo(N, DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1),
1744 DAG.getNode(ISD::CARRY_FALSE,
1745 SDLoc(N), MVT::Glue));
1751 SDValue DAGCombiner::visitADDE(SDNode *N) {
1752 SDValue N0 = N->getOperand(0);
1753 SDValue N1 = N->getOperand(1);
1754 SDValue CarryIn = N->getOperand(2);
1755 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1756 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1758 // canonicalize constant to RHS
1760 return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(),
1763 // fold (adde x, y, false) -> (addc x, y)
1764 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1765 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1);
1770 // Since it may not be valid to emit a fold to zero for vector initializers
1771 // check if we can before folding.
1772 static SDValue tryFoldToZero(SDLoc DL, const TargetLowering &TLI, EVT VT,
1774 bool LegalOperations, bool LegalTypes) {
1776 return DAG.getConstant(0, VT);
1777 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
1778 return DAG.getConstant(0, VT);
1782 SDValue DAGCombiner::visitSUB(SDNode *N) {
1783 SDValue N0 = N->getOperand(0);
1784 SDValue N1 = N->getOperand(1);
1785 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1786 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1787 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? nullptr :
1788 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1789 EVT VT = N0.getValueType();
1792 if (VT.isVector()) {
1793 SDValue FoldedVOp = SimplifyVBinOp(N);
1794 if (FoldedVOp.getNode()) return FoldedVOp;
1796 // fold (sub x, 0) -> x, vector edition
1797 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1801 // fold (sub x, x) -> 0
1802 // FIXME: Refactor this and xor and other similar operations together.
1804 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
1805 // fold (sub c1, c2) -> c1-c2
1807 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1808 // fold (sub x, c) -> (add x, -c)
1810 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0,
1811 DAG.getConstant(-N1C->getAPIntValue(), VT));
1812 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1813 if (N0C && N0C->isAllOnesValue())
1814 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
1815 // fold A-(A-B) -> B
1816 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1817 return N1.getOperand(1);
1818 // fold (A+B)-A -> B
1819 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1820 return N0.getOperand(1);
1821 // fold (A+B)-B -> A
1822 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1823 return N0.getOperand(0);
1824 // fold C2-(A+C1) -> (C2-C1)-A
1825 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1826 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(),
1828 return DAG.getNode(ISD::SUB, SDLoc(N), VT, NewC,
1831 // fold ((A+(B+or-C))-B) -> A+or-C
1832 if (N0.getOpcode() == ISD::ADD &&
1833 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1834 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1835 N0.getOperand(1).getOperand(0) == N1)
1836 return DAG.getNode(N0.getOperand(1).getOpcode(), SDLoc(N), VT,
1837 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1838 // fold ((A+(C+B))-B) -> A+C
1839 if (N0.getOpcode() == ISD::ADD &&
1840 N0.getOperand(1).getOpcode() == ISD::ADD &&
1841 N0.getOperand(1).getOperand(1) == N1)
1842 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1843 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1844 // fold ((A-(B-C))-C) -> A-B
1845 if (N0.getOpcode() == ISD::SUB &&
1846 N0.getOperand(1).getOpcode() == ISD::SUB &&
1847 N0.getOperand(1).getOperand(1) == N1)
1848 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1849 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1851 // If either operand of a sub is undef, the result is undef
1852 if (N0.getOpcode() == ISD::UNDEF)
1854 if (N1.getOpcode() == ISD::UNDEF)
1857 // If the relocation model supports it, consider symbol offsets.
1858 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1859 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1860 // fold (sub Sym, c) -> Sym-c
1861 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1862 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1864 (uint64_t)N1C->getSExtValue());
1865 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1866 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1867 if (GA->getGlobal() == GB->getGlobal())
1868 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1872 // sub X, (sextinreg Y i1) -> add X, (and Y 1)
1873 if (N1.getOpcode() == ISD::SIGN_EXTEND_INREG) {
1874 VTSDNode *TN = cast<VTSDNode>(N1.getOperand(1));
1875 if (TN->getVT() == MVT::i1) {
1877 SDValue ZExt = DAG.getNode(ISD::AND, DL, VT, N1.getOperand(0),
1878 DAG.getConstant(1, VT));
1879 return DAG.getNode(ISD::ADD, DL, VT, N0, ZExt);
1886 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1887 SDValue N0 = N->getOperand(0);
1888 SDValue N1 = N->getOperand(1);
1889 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1890 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1891 EVT VT = N0.getValueType();
1893 // If the flag result is dead, turn this into an SUB.
1894 if (!N->hasAnyUseOfValue(1))
1895 return CombineTo(N, DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1),
1896 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1899 // fold (subc x, x) -> 0 + no borrow
1901 return CombineTo(N, DAG.getConstant(0, VT),
1902 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1905 // fold (subc x, 0) -> x + no borrow
1906 if (N1C && N1C->isNullValue())
1907 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1910 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1911 if (N0C && N0C->isAllOnesValue())
1912 return CombineTo(N, DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0),
1913 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1919 SDValue DAGCombiner::visitSUBE(SDNode *N) {
1920 SDValue N0 = N->getOperand(0);
1921 SDValue N1 = N->getOperand(1);
1922 SDValue CarryIn = N->getOperand(2);
1924 // fold (sube x, y, false) -> (subc x, y)
1925 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1926 return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1);
1931 SDValue DAGCombiner::visitMUL(SDNode *N) {
1932 SDValue N0 = N->getOperand(0);
1933 SDValue N1 = N->getOperand(1);
1934 EVT VT = N0.getValueType();
1936 // fold (mul x, undef) -> 0
1937 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1938 return DAG.getConstant(0, VT);
1940 bool N0IsConst = false;
1941 bool N1IsConst = false;
1942 APInt ConstValue0, ConstValue1;
1944 if (VT.isVector()) {
1945 SDValue FoldedVOp = SimplifyVBinOp(N);
1946 if (FoldedVOp.getNode()) return FoldedVOp;
1948 N0IsConst = isConstantSplatVector(N0.getNode(), ConstValue0);
1949 N1IsConst = isConstantSplatVector(N1.getNode(), ConstValue1);
1951 N0IsConst = dyn_cast<ConstantSDNode>(N0) != nullptr;
1952 ConstValue0 = N0IsConst ? (dyn_cast<ConstantSDNode>(N0))->getAPIntValue()
1954 N1IsConst = dyn_cast<ConstantSDNode>(N1) != nullptr;
1955 ConstValue1 = N1IsConst ? (dyn_cast<ConstantSDNode>(N1))->getAPIntValue()
1959 // fold (mul c1, c2) -> c1*c2
1960 if (N0IsConst && N1IsConst)
1961 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0.getNode(), N1.getNode());
1963 // canonicalize constant to RHS
1964 if (N0IsConst && !N1IsConst)
1965 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0);
1966 // fold (mul x, 0) -> 0
1967 if (N1IsConst && ConstValue1 == 0)
1969 // We require a splat of the entire scalar bit width for non-contiguous
1972 ConstValue1.getBitWidth() == VT.getScalarType().getSizeInBits();
1973 // fold (mul x, 1) -> x
1974 if (N1IsConst && ConstValue1 == 1 && IsFullSplat)
1976 // fold (mul x, -1) -> 0-x
1977 if (N1IsConst && ConstValue1.isAllOnesValue())
1978 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1979 DAG.getConstant(0, VT), N0);
1980 // fold (mul x, (1 << c)) -> x << c
1981 if (N1IsConst && ConstValue1.isPowerOf2() && IsFullSplat)
1982 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
1983 DAG.getConstant(ConstValue1.logBase2(),
1984 getShiftAmountTy(N0.getValueType())));
1985 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
1986 if (N1IsConst && (-ConstValue1).isPowerOf2() && IsFullSplat) {
1987 unsigned Log2Val = (-ConstValue1).logBase2();
1988 // FIXME: If the input is something that is easily negated (e.g. a
1989 // single-use add), we should put the negate there.
1990 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1991 DAG.getConstant(0, VT),
1992 DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
1993 DAG.getConstant(Log2Val,
1994 getShiftAmountTy(N0.getValueType()))));
1998 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
1999 if (N1IsConst && N0.getOpcode() == ISD::SHL &&
2000 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2001 isa<ConstantSDNode>(N0.getOperand(1)))) {
2002 SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT,
2003 N1, N0.getOperand(1));
2004 AddToWorklist(C3.getNode());
2005 return DAG.getNode(ISD::MUL, SDLoc(N), VT,
2006 N0.getOperand(0), C3);
2009 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
2012 SDValue Sh(nullptr,0), Y(nullptr,0);
2013 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
2014 if (N0.getOpcode() == ISD::SHL &&
2015 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2016 isa<ConstantSDNode>(N0.getOperand(1))) &&
2017 N0.getNode()->hasOneUse()) {
2019 } else if (N1.getOpcode() == ISD::SHL &&
2020 isa<ConstantSDNode>(N1.getOperand(1)) &&
2021 N1.getNode()->hasOneUse()) {
2026 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2027 Sh.getOperand(0), Y);
2028 return DAG.getNode(ISD::SHL, SDLoc(N), VT,
2029 Mul, Sh.getOperand(1));
2033 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
2034 if (N1IsConst && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
2035 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
2036 isa<ConstantSDNode>(N0.getOperand(1))))
2037 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
2038 DAG.getNode(ISD::MUL, SDLoc(N0), VT,
2039 N0.getOperand(0), N1),
2040 DAG.getNode(ISD::MUL, SDLoc(N1), VT,
2041 N0.getOperand(1), N1));
2044 SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1);
2051 SDValue DAGCombiner::visitSDIV(SDNode *N) {
2052 SDValue N0 = N->getOperand(0);
2053 SDValue N1 = N->getOperand(1);
2054 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2055 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2056 EVT VT = N->getValueType(0);
2059 if (VT.isVector()) {
2060 SDValue FoldedVOp = SimplifyVBinOp(N);
2061 if (FoldedVOp.getNode()) return FoldedVOp;
2064 // fold (sdiv c1, c2) -> c1/c2
2065 if (N0C && N1C && !N1C->isNullValue())
2066 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
2067 // fold (sdiv X, 1) -> X
2068 if (N1C && N1C->getAPIntValue() == 1LL)
2070 // fold (sdiv X, -1) -> 0-X
2071 if (N1C && N1C->isAllOnesValue())
2072 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2073 DAG.getConstant(0, VT), N0);
2074 // If we know the sign bits of both operands are zero, strength reduce to a
2075 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
2076 if (!VT.isVector()) {
2077 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2078 return DAG.getNode(ISD::UDIV, SDLoc(N), N1.getValueType(),
2082 // fold (sdiv X, pow2) -> simple ops after legalize
2083 if (N1C && !N1C->isNullValue() && (N1C->getAPIntValue().isPowerOf2() ||
2084 (-N1C->getAPIntValue()).isPowerOf2())) {
2085 // If dividing by powers of two is cheap, then don't perform the following
2087 if (TLI.isPow2SDivCheap())
2090 // Target-specific implementation of sdiv x, pow2.
2091 SDValue Res = BuildSDIVPow2(N);
2095 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
2097 // Splat the sign bit into the register
2099 DAG.getNode(ISD::SRA, SDLoc(N), VT, N0,
2100 DAG.getConstant(VT.getScalarSizeInBits() - 1,
2101 getShiftAmountTy(N0.getValueType())));
2102 AddToWorklist(SGN.getNode());
2104 // Add (N0 < 0) ? abs2 - 1 : 0;
2106 DAG.getNode(ISD::SRL, SDLoc(N), VT, SGN,
2107 DAG.getConstant(VT.getScalarSizeInBits() - lg2,
2108 getShiftAmountTy(SGN.getValueType())));
2109 SDValue ADD = DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, SRL);
2110 AddToWorklist(SRL.getNode());
2111 AddToWorklist(ADD.getNode()); // Divide by pow2
2112 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), VT, ADD,
2113 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType())));
2115 // If we're dividing by a positive value, we're done. Otherwise, we must
2116 // negate the result.
2117 if (N1C->getAPIntValue().isNonNegative())
2120 AddToWorklist(SRA.getNode());
2121 return DAG.getNode(ISD::SUB, SDLoc(N), VT, DAG.getConstant(0, VT), SRA);
2124 // if integer divide is expensive and we satisfy the requirements, emit an
2125 // alternate sequence.
2126 if (N1C && !TLI.isIntDivCheap()) {
2127 SDValue Op = BuildSDIV(N);
2128 if (Op.getNode()) return Op;
2132 if (N0.getOpcode() == ISD::UNDEF)
2133 return DAG.getConstant(0, VT);
2134 // X / undef -> undef
2135 if (N1.getOpcode() == ISD::UNDEF)
2141 SDValue DAGCombiner::visitUDIV(SDNode *N) {
2142 SDValue N0 = N->getOperand(0);
2143 SDValue N1 = N->getOperand(1);
2144 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2145 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2146 EVT VT = N->getValueType(0);
2149 if (VT.isVector()) {
2150 SDValue FoldedVOp = SimplifyVBinOp(N);
2151 if (FoldedVOp.getNode()) return FoldedVOp;
2154 // fold (udiv c1, c2) -> c1/c2
2155 if (N0C && N1C && !N1C->isNullValue())
2156 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
2157 // fold (udiv x, (1 << c)) -> x >>u c
2158 if (N1C && N1C->getAPIntValue().isPowerOf2())
2159 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0,
2160 DAG.getConstant(N1C->getAPIntValue().logBase2(),
2161 getShiftAmountTy(N0.getValueType())));
2162 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
2163 if (N1.getOpcode() == ISD::SHL) {
2164 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2165 if (SHC->getAPIntValue().isPowerOf2()) {
2166 EVT ADDVT = N1.getOperand(1).getValueType();
2167 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N), ADDVT,
2169 DAG.getConstant(SHC->getAPIntValue()
2172 AddToWorklist(Add.getNode());
2173 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, Add);
2177 // fold (udiv x, c) -> alternate
2178 if (N1C && !TLI.isIntDivCheap()) {
2179 SDValue Op = BuildUDIV(N);
2180 if (Op.getNode()) return Op;
2184 if (N0.getOpcode() == ISD::UNDEF)
2185 return DAG.getConstant(0, VT);
2186 // X / undef -> undef
2187 if (N1.getOpcode() == ISD::UNDEF)
2193 SDValue DAGCombiner::visitSREM(SDNode *N) {
2194 SDValue N0 = N->getOperand(0);
2195 SDValue N1 = N->getOperand(1);
2196 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2197 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2198 EVT VT = N->getValueType(0);
2200 // fold (srem c1, c2) -> c1%c2
2201 if (N0C && N1C && !N1C->isNullValue())
2202 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
2203 // If we know the sign bits of both operands are zero, strength reduce to a
2204 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
2205 if (!VT.isVector()) {
2206 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2207 return DAG.getNode(ISD::UREM, SDLoc(N), VT, N0, N1);
2210 // If X/C can be simplified by the division-by-constant logic, lower
2211 // X%C to the equivalent of X-X/C*C.
2212 if (N1C && !N1C->isNullValue()) {
2213 SDValue Div = DAG.getNode(ISD::SDIV, SDLoc(N), VT, N0, N1);
2214 AddToWorklist(Div.getNode());
2215 SDValue OptimizedDiv = combine(Div.getNode());
2216 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2217 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2219 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2220 AddToWorklist(Mul.getNode());
2226 if (N0.getOpcode() == ISD::UNDEF)
2227 return DAG.getConstant(0, VT);
2228 // X % undef -> undef
2229 if (N1.getOpcode() == ISD::UNDEF)
2235 SDValue DAGCombiner::visitUREM(SDNode *N) {
2236 SDValue N0 = N->getOperand(0);
2237 SDValue N1 = N->getOperand(1);
2238 ConstantSDNode *N0C = isConstOrConstSplat(N0);
2239 ConstantSDNode *N1C = isConstOrConstSplat(N1);
2240 EVT VT = N->getValueType(0);
2242 // fold (urem c1, c2) -> c1%c2
2243 if (N0C && N1C && !N1C->isNullValue())
2244 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
2245 // fold (urem x, pow2) -> (and x, pow2-1)
2246 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
2247 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0,
2248 DAG.getConstant(N1C->getAPIntValue()-1,VT));
2249 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2250 if (N1.getOpcode() == ISD::SHL) {
2251 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2252 if (SHC->getAPIntValue().isPowerOf2()) {
2254 DAG.getNode(ISD::ADD, SDLoc(N), VT, N1,
2255 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
2257 AddToWorklist(Add.getNode());
2258 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, Add);
2263 // If X/C can be simplified by the division-by-constant logic, lower
2264 // X%C to the equivalent of X-X/C*C.
2265 if (N1C && !N1C->isNullValue()) {
2266 SDValue Div = DAG.getNode(ISD::UDIV, SDLoc(N), VT, N0, N1);
2267 AddToWorklist(Div.getNode());
2268 SDValue OptimizedDiv = combine(Div.getNode());
2269 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2270 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2272 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2273 AddToWorklist(Mul.getNode());
2279 if (N0.getOpcode() == ISD::UNDEF)
2280 return DAG.getConstant(0, VT);
2281 // X % undef -> undef
2282 if (N1.getOpcode() == ISD::UNDEF)
2288 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2289 SDValue N0 = N->getOperand(0);
2290 SDValue N1 = N->getOperand(1);
2291 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2292 EVT VT = N->getValueType(0);
2295 // fold (mulhs x, 0) -> 0
2296 if (N1C && N1C->isNullValue())
2298 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2299 if (N1C && N1C->getAPIntValue() == 1)
2300 return DAG.getNode(ISD::SRA, SDLoc(N), N0.getValueType(), N0,
2301 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2302 getShiftAmountTy(N0.getValueType())));
2303 // fold (mulhs x, undef) -> 0
2304 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2305 return DAG.getConstant(0, VT);
2307 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2309 if (VT.isSimple() && !VT.isVector()) {
2310 MVT Simple = VT.getSimpleVT();
2311 unsigned SimpleSize = Simple.getSizeInBits();
2312 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2313 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2314 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2315 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2316 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2317 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2318 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2319 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2326 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2327 SDValue N0 = N->getOperand(0);
2328 SDValue N1 = N->getOperand(1);
2329 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2330 EVT VT = N->getValueType(0);
2333 // fold (mulhu x, 0) -> 0
2334 if (N1C && N1C->isNullValue())
2336 // fold (mulhu x, 1) -> 0
2337 if (N1C && N1C->getAPIntValue() == 1)
2338 return DAG.getConstant(0, N0.getValueType());
2339 // fold (mulhu x, undef) -> 0
2340 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2341 return DAG.getConstant(0, VT);
2343 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2345 if (VT.isSimple() && !VT.isVector()) {
2346 MVT Simple = VT.getSimpleVT();
2347 unsigned SimpleSize = Simple.getSizeInBits();
2348 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2349 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2350 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2351 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2352 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2353 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2354 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2355 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2362 /// Perform optimizations common to nodes that compute two values. LoOp and HiOp
2363 /// give the opcodes for the two computations that are being performed. Return
2364 /// true if a simplification was made.
2365 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2367 // If the high half is not needed, just compute the low half.
2368 bool HiExists = N->hasAnyUseOfValue(1);
2370 (!LegalOperations ||
2371 TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) {
2372 SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2373 return CombineTo(N, Res, Res);
2376 // If the low half is not needed, just compute the high half.
2377 bool LoExists = N->hasAnyUseOfValue(0);
2379 (!LegalOperations ||
2380 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2381 SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2382 return CombineTo(N, Res, Res);
2385 // If both halves are used, return as it is.
2386 if (LoExists && HiExists)
2389 // If the two computed results can be simplified separately, separate them.
2391 SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0), N->ops());
2392 AddToWorklist(Lo.getNode());
2393 SDValue LoOpt = combine(Lo.getNode());
2394 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2395 (!LegalOperations ||
2396 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2397 return CombineTo(N, LoOpt, LoOpt);
2401 SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1), N->ops());
2402 AddToWorklist(Hi.getNode());
2403 SDValue HiOpt = combine(Hi.getNode());
2404 if (HiOpt.getNode() && HiOpt != Hi &&
2405 (!LegalOperations ||
2406 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2407 return CombineTo(N, HiOpt, HiOpt);
2413 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2414 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2415 if (Res.getNode()) return Res;
2417 EVT VT = N->getValueType(0);
2420 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2422 if (VT.isSimple() && !VT.isVector()) {
2423 MVT Simple = VT.getSimpleVT();
2424 unsigned SimpleSize = Simple.getSizeInBits();
2425 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2426 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2427 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2428 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2429 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2430 // Compute the high part as N1.
2431 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2432 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2433 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2434 // Compute the low part as N0.
2435 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2436 return CombineTo(N, Lo, Hi);
2443 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2444 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2445 if (Res.getNode()) return Res;
2447 EVT VT = N->getValueType(0);
2450 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2452 if (VT.isSimple() && !VT.isVector()) {
2453 MVT Simple = VT.getSimpleVT();
2454 unsigned SimpleSize = Simple.getSizeInBits();
2455 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2456 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2457 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2458 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2459 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2460 // Compute the high part as N1.
2461 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2462 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2463 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2464 // Compute the low part as N0.
2465 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2466 return CombineTo(N, Lo, Hi);
2473 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2474 // (smulo x, 2) -> (saddo x, x)
2475 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2476 if (C2->getAPIntValue() == 2)
2477 return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(),
2478 N->getOperand(0), N->getOperand(0));
2483 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2484 // (umulo x, 2) -> (uaddo x, x)
2485 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2486 if (C2->getAPIntValue() == 2)
2487 return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(),
2488 N->getOperand(0), N->getOperand(0));
2493 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2494 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2495 if (Res.getNode()) return Res;
2500 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2501 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2502 if (Res.getNode()) return Res;
2507 /// If this is a binary operator with two operands of the same opcode, try to
2509 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2510 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2511 EVT VT = N0.getValueType();
2512 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2514 // Bail early if none of these transforms apply.
2515 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2517 // For each of OP in AND/OR/XOR:
2518 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2519 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2520 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2521 // fold (OP (bswap x), (bswap y)) -> (bswap (OP x, y))
2522 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2524 // do not sink logical op inside of a vector extend, since it may combine
2526 EVT Op0VT = N0.getOperand(0).getValueType();
2527 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2528 N0.getOpcode() == ISD::SIGN_EXTEND ||
2529 N0.getOpcode() == ISD::BSWAP ||
2530 // Avoid infinite looping with PromoteIntBinOp.
2531 (N0.getOpcode() == ISD::ANY_EXTEND &&
2532 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2533 (N0.getOpcode() == ISD::TRUNCATE &&
2534 (!TLI.isZExtFree(VT, Op0VT) ||
2535 !TLI.isTruncateFree(Op0VT, VT)) &&
2536 TLI.isTypeLegal(Op0VT))) &&
2538 Op0VT == N1.getOperand(0).getValueType() &&
2539 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2540 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2541 N0.getOperand(0).getValueType(),
2542 N0.getOperand(0), N1.getOperand(0));
2543 AddToWorklist(ORNode.getNode());
2544 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode);
2547 // For each of OP in SHL/SRL/SRA/AND...
2548 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2549 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2550 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2551 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2552 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2553 N0.getOperand(1) == N1.getOperand(1)) {
2554 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2555 N0.getOperand(0).getValueType(),
2556 N0.getOperand(0), N1.getOperand(0));
2557 AddToWorklist(ORNode.getNode());
2558 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
2559 ORNode, N0.getOperand(1));
2562 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2563 // Only perform this optimization after type legalization and before
2564 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2565 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2566 // we don't want to undo this promotion.
2567 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2569 if ((N0.getOpcode() == ISD::BITCAST ||
2570 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) &&
2571 Level == AfterLegalizeTypes) {
2572 SDValue In0 = N0.getOperand(0);
2573 SDValue In1 = N1.getOperand(0);
2574 EVT In0Ty = In0.getValueType();
2575 EVT In1Ty = In1.getValueType();
2577 // If both incoming values are integers, and the original types are the
2579 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2580 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1);
2581 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op);
2582 AddToWorklist(Op.getNode());
2587 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2588 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2589 // If both shuffles use the same mask, and both shuffle within a single
2590 // vector, then it is worthwhile to move the swizzle after the operation.
2591 // The type-legalizer generates this pattern when loading illegal
2592 // vector types from memory. In many cases this allows additional shuffle
2594 // There are other cases where moving the shuffle after the xor/and/or
2595 // is profitable even if shuffles don't perform a swizzle.
2596 // If both shuffles use the same mask, and both shuffles have the same first
2597 // or second operand, then it might still be profitable to move the shuffle
2598 // after the xor/and/or operation.
2599 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) {
2600 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2601 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2603 assert(N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() &&
2604 "Inputs to shuffles are not the same type");
2606 // Check that both shuffles use the same mask. The masks are known to be of
2607 // the same length because the result vector type is the same.
2608 // Check also that shuffles have only one use to avoid introducing extra
2610 if (SVN0->hasOneUse() && SVN1->hasOneUse() &&
2611 SVN0->getMask().equals(SVN1->getMask())) {
2612 SDValue ShOp = N0->getOperand(1);
2614 // Don't try to fold this node if it requires introducing a
2615 // build vector of all zeros that might be illegal at this stage.
2616 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2618 ShOp = DAG.getConstant(0, VT);
2623 // (AND (shuf (A, C), shuf (B, C)) -> shuf (AND (A, B), C)
2624 // (OR (shuf (A, C), shuf (B, C)) -> shuf (OR (A, B), C)
2625 // (XOR (shuf (A, C), shuf (B, C)) -> shuf (XOR (A, B), V_0)
2626 if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) {
2627 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2628 N0->getOperand(0), N1->getOperand(0));
2629 AddToWorklist(NewNode.getNode());
2630 return DAG.getVectorShuffle(VT, SDLoc(N), NewNode, ShOp,
2631 &SVN0->getMask()[0]);
2634 // Don't try to fold this node if it requires introducing a
2635 // build vector of all zeros that might be illegal at this stage.
2636 ShOp = N0->getOperand(0);
2637 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2639 ShOp = DAG.getConstant(0, VT);
2644 // (AND (shuf (C, A), shuf (C, B)) -> shuf (C, AND (A, B))
2645 // (OR (shuf (C, A), shuf (C, B)) -> shuf (C, OR (A, B))
2646 // (XOR (shuf (C, A), shuf (C, B)) -> shuf (V_0, XOR (A, B))
2647 if (N0->getOperand(0) == N1->getOperand(0) && ShOp.getNode()) {
2648 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2649 N0->getOperand(1), N1->getOperand(1));
2650 AddToWorklist(NewNode.getNode());
2651 return DAG.getVectorShuffle(VT, SDLoc(N), ShOp, NewNode,
2652 &SVN0->getMask()[0]);
2660 SDValue DAGCombiner::visitAND(SDNode *N) {
2661 SDValue N0 = N->getOperand(0);
2662 SDValue N1 = N->getOperand(1);
2663 SDValue LL, LR, RL, RR, CC0, CC1;
2664 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2665 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2666 EVT VT = N1.getValueType();
2667 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2670 if (VT.isVector()) {
2671 SDValue FoldedVOp = SimplifyVBinOp(N);
2672 if (FoldedVOp.getNode()) return FoldedVOp;
2674 // fold (and x, 0) -> 0, vector edition
2675 if (ISD::isBuildVectorAllZeros(N0.getNode()))
2676 // do not return N0, because undef node may exist in N0
2677 return DAG.getConstant(
2678 APInt::getNullValue(
2679 N0.getValueType().getScalarType().getSizeInBits()),
2681 if (ISD::isBuildVectorAllZeros(N1.getNode()))
2682 // do not return N1, because undef node may exist in N1
2683 return DAG.getConstant(
2684 APInt::getNullValue(
2685 N1.getValueType().getScalarType().getSizeInBits()),
2688 // fold (and x, -1) -> x, vector edition
2689 if (ISD::isBuildVectorAllOnes(N0.getNode()))
2691 if (ISD::isBuildVectorAllOnes(N1.getNode()))
2695 // fold (and x, undef) -> 0
2696 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2697 return DAG.getConstant(0, VT);
2698 // fold (and c1, c2) -> c1&c2
2700 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2701 // canonicalize constant to RHS
2703 return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0);
2704 // fold (and x, -1) -> x
2705 if (N1C && N1C->isAllOnesValue())
2707 // if (and x, c) is known to be zero, return 0
2708 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2709 APInt::getAllOnesValue(BitWidth)))
2710 return DAG.getConstant(0, VT);
2712 SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1);
2715 // fold (and (or x, C), D) -> D if (C & D) == D
2716 if (N1C && N0.getOpcode() == ISD::OR)
2717 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2718 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2720 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2721 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2722 SDValue N0Op0 = N0.getOperand(0);
2723 APInt Mask = ~N1C->getAPIntValue();
2724 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2725 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2726 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
2727 N0.getValueType(), N0Op0);
2729 // Replace uses of the AND with uses of the Zero extend node.
2732 // We actually want to replace all uses of the any_extend with the
2733 // zero_extend, to avoid duplicating things. This will later cause this
2734 // AND to be folded.
2735 CombineTo(N0.getNode(), Zext);
2736 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2739 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2740 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2741 // already be zero by virtue of the width of the base type of the load.
2743 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2745 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2746 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2747 N0.getOpcode() == ISD::LOAD) {
2748 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2749 N0 : N0.getOperand(0) );
2751 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2752 // This can be a pure constant or a vector splat, in which case we treat the
2753 // vector as a scalar and use the splat value.
2754 APInt Constant = APInt::getNullValue(1);
2755 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2756 Constant = C->getAPIntValue();
2757 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2758 APInt SplatValue, SplatUndef;
2759 unsigned SplatBitSize;
2761 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2762 SplatBitSize, HasAnyUndefs);
2764 // Undef bits can contribute to a possible optimisation if set, so
2766 SplatValue |= SplatUndef;
2768 // The splat value may be something like "0x00FFFFFF", which means 0 for
2769 // the first vector value and FF for the rest, repeating. We need a mask
2770 // that will apply equally to all members of the vector, so AND all the
2771 // lanes of the constant together.
2772 EVT VT = Vector->getValueType(0);
2773 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2775 // If the splat value has been compressed to a bitlength lower
2776 // than the size of the vector lane, we need to re-expand it to
2778 if (BitWidth > SplatBitSize)
2779 for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
2780 SplatBitSize < BitWidth;
2781 SplatBitSize = SplatBitSize * 2)
2782 SplatValue |= SplatValue.shl(SplatBitSize);
2784 // Make sure that variable 'Constant' is only set if 'SplatBitSize' is a
2785 // multiple of 'BitWidth'. Otherwise, we could propagate a wrong value.
2786 if (SplatBitSize % BitWidth == 0) {
2787 Constant = APInt::getAllOnesValue(BitWidth);
2788 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
2789 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
2794 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
2795 // actually legal and isn't going to get expanded, else this is a false
2797 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
2798 Load->getValueType(0),
2799 Load->getMemoryVT());
2801 // Resize the constant to the same size as the original memory access before
2802 // extension. If it is still the AllOnesValue then this AND is completely
2805 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
2808 switch (Load->getExtensionType()) {
2809 default: B = false; break;
2810 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
2812 case ISD::NON_EXTLOAD: B = true; break;
2815 if (B && Constant.isAllOnesValue()) {
2816 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
2817 // preserve semantics once we get rid of the AND.
2818 SDValue NewLoad(Load, 0);
2819 if (Load->getExtensionType() == ISD::EXTLOAD) {
2820 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
2821 Load->getValueType(0), SDLoc(Load),
2822 Load->getChain(), Load->getBasePtr(),
2823 Load->getOffset(), Load->getMemoryVT(),
2824 Load->getMemOperand());
2825 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
2826 if (Load->getNumValues() == 3) {
2827 // PRE/POST_INC loads have 3 values.
2828 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
2829 NewLoad.getValue(2) };
2830 CombineTo(Load, To, 3, true);
2832 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
2836 // Fold the AND away, taking care not to fold to the old load node if we
2838 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
2840 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2843 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2844 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2845 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2846 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2848 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2849 LL.getValueType().isInteger()) {
2850 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2851 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
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);
2857 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2858 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2859 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(N0),
2860 LR.getValueType(), LL, RL);
2861 AddToWorklist(ANDNode.getNode());
2862 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
2864 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2865 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2866 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2867 LR.getValueType(), LL, RL);
2868 AddToWorklist(ORNode.getNode());
2869 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
2872 // Simplify (and (setne X, 0), (setne X, -1)) -> (setuge (add X, 1), 2)
2873 if (LL == RL && isa<ConstantSDNode>(LR) && isa<ConstantSDNode>(RR) &&
2874 Op0 == Op1 && LL.getValueType().isInteger() &&
2875 Op0 == ISD::SETNE && ((cast<ConstantSDNode>(LR)->isNullValue() &&
2876 cast<ConstantSDNode>(RR)->isAllOnesValue()) ||
2877 (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
2878 cast<ConstantSDNode>(RR)->isNullValue()))) {
2879 SDValue ADDNode = DAG.getNode(ISD::ADD, SDLoc(N0), LL.getValueType(),
2880 LL, DAG.getConstant(1, LL.getValueType()));
2881 AddToWorklist(ADDNode.getNode());
2882 return DAG.getSetCC(SDLoc(N), VT, ADDNode,
2883 DAG.getConstant(2, LL.getValueType()), ISD::SETUGE);
2885 // canonicalize equivalent to ll == rl
2886 if (LL == RR && LR == RL) {
2887 Op1 = ISD::getSetCCSwappedOperands(Op1);
2890 if (LL == RL && LR == RR) {
2891 bool isInteger = LL.getValueType().isInteger();
2892 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2893 if (Result != ISD::SETCC_INVALID &&
2894 (!LegalOperations ||
2895 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
2896 TLI.isOperationLegal(ISD::SETCC,
2897 getSetCCResultType(N0.getSimpleValueType())))))
2898 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
2903 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
2904 if (N0.getOpcode() == N1.getOpcode()) {
2905 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2906 if (Tmp.getNode()) return Tmp;
2909 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
2910 // fold (and (sra)) -> (and (srl)) when possible.
2911 if (!VT.isVector() &&
2912 SimplifyDemandedBits(SDValue(N, 0)))
2913 return SDValue(N, 0);
2915 // fold (zext_inreg (extload x)) -> (zextload x)
2916 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
2917 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2918 EVT MemVT = LN0->getMemoryVT();
2919 // If we zero all the possible extended bits, then we can turn this into
2920 // a zextload if we are running before legalize or the operation is legal.
2921 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2922 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2923 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2924 ((!LegalOperations && !LN0->isVolatile()) ||
2925 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) {
2926 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2927 LN0->getChain(), LN0->getBasePtr(),
2928 MemVT, LN0->getMemOperand());
2930 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2931 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2934 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
2935 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
2937 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2938 EVT MemVT = LN0->getMemoryVT();
2939 // If we zero all the possible extended bits, then we can turn this into
2940 // a zextload if we are running before legalize or the operation is legal.
2941 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2942 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2943 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2944 ((!LegalOperations && !LN0->isVolatile()) ||
2945 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT))) {
2946 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2947 LN0->getChain(), LN0->getBasePtr(),
2948 MemVT, LN0->getMemOperand());
2950 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2951 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2955 // fold (and (load x), 255) -> (zextload x, i8)
2956 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2957 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2958 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2959 (N0.getOpcode() == ISD::ANY_EXTEND &&
2960 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2961 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2962 LoadSDNode *LN0 = HasAnyExt
2963 ? cast<LoadSDNode>(N0.getOperand(0))
2964 : cast<LoadSDNode>(N0);
2965 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2966 LN0->isUnindexed() && N0.hasOneUse() && SDValue(LN0, 0).hasOneUse()) {
2967 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2968 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2969 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2970 EVT LoadedVT = LN0->getMemoryVT();
2971 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2973 if (ExtVT == LoadedVT &&
2974 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy,
2978 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
2979 LN0->getChain(), LN0->getBasePtr(), ExtVT,
2980 LN0->getMemOperand());
2982 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
2983 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2986 // Do not change the width of a volatile load.
2987 // Do not generate loads of non-round integer types since these can
2988 // be expensive (and would be wrong if the type is not byte sized).
2989 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
2990 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, LoadResultTy,
2992 EVT PtrType = LN0->getOperand(1).getValueType();
2994 unsigned Alignment = LN0->getAlignment();
2995 SDValue NewPtr = LN0->getBasePtr();
2997 // For big endian targets, we need to add an offset to the pointer
2998 // to load the correct bytes. For little endian systems, we merely
2999 // need to read fewer bytes from the same pointer.
3000 if (TLI.isBigEndian()) {
3001 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
3002 unsigned EVTStoreBytes = ExtVT.getStoreSize();
3003 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
3004 NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0), PtrType,
3005 NewPtr, DAG.getConstant(PtrOff, PtrType));
3006 Alignment = MinAlign(Alignment, PtrOff);
3009 AddToWorklist(NewPtr.getNode());
3012 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
3013 LN0->getChain(), NewPtr,
3014 LN0->getPointerInfo(),
3015 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
3016 LN0->isInvariant(), Alignment, LN0->getAAInfo());
3018 CombineTo(LN0, Load, Load.getValue(1));
3019 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3025 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
3026 VT.getSizeInBits() <= 64) {
3027 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3028 APInt ADDC = ADDI->getAPIntValue();
3029 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
3030 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
3031 // immediate for an add, but it is legal if its top c2 bits are set,
3032 // transform the ADD so the immediate doesn't need to be materialized
3034 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
3035 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
3036 SRLI->getZExtValue());
3037 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
3039 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
3041 DAG.getNode(ISD::ADD, SDLoc(N0), VT,
3042 N0.getOperand(0), DAG.getConstant(ADDC, VT));
3043 CombineTo(N0.getNode(), NewAdd);
3044 return SDValue(N, 0); // Return N so it doesn't get rechecked!
3052 // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const)
3053 if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) {
3054 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
3055 N0.getOperand(1), false);
3056 if (BSwap.getNode())
3063 /// Match (a >> 8) | (a << 8) as (bswap a) >> 16.
3064 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
3065 bool DemandHighBits) {
3066 if (!LegalOperations)
3069 EVT VT = N->getValueType(0);
3070 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
3072 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3075 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
3076 bool LookPassAnd0 = false;
3077 bool LookPassAnd1 = false;
3078 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
3080 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
3082 if (N0.getOpcode() == ISD::AND) {
3083 if (!N0.getNode()->hasOneUse())
3085 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3086 if (!N01C || N01C->getZExtValue() != 0xFF00)
3088 N0 = N0.getOperand(0);
3089 LookPassAnd0 = true;
3092 if (N1.getOpcode() == ISD::AND) {
3093 if (!N1.getNode()->hasOneUse())
3095 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3096 if (!N11C || N11C->getZExtValue() != 0xFF)
3098 N1 = N1.getOperand(0);
3099 LookPassAnd1 = true;
3102 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
3104 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
3106 if (!N0.getNode()->hasOneUse() ||
3107 !N1.getNode()->hasOneUse())
3110 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3111 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3114 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
3117 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
3118 SDValue N00 = N0->getOperand(0);
3119 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
3120 if (!N00.getNode()->hasOneUse())
3122 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
3123 if (!N001C || N001C->getZExtValue() != 0xFF)
3125 N00 = N00.getOperand(0);
3126 LookPassAnd0 = true;
3129 SDValue N10 = N1->getOperand(0);
3130 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
3131 if (!N10.getNode()->hasOneUse())
3133 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
3134 if (!N101C || N101C->getZExtValue() != 0xFF00)
3136 N10 = N10.getOperand(0);
3137 LookPassAnd1 = true;
3143 // Make sure everything beyond the low halfword gets set to zero since the SRL
3144 // 16 will clear the top bits.
3145 unsigned OpSizeInBits = VT.getSizeInBits();
3146 if (DemandHighBits && OpSizeInBits > 16) {
3147 // If the left-shift isn't masked out then the only way this is a bswap is
3148 // if all bits beyond the low 8 are 0. In that case the entire pattern
3149 // reduces to a left shift anyway: leave it for other parts of the combiner.
3153 // However, if the right shift isn't masked out then it might be because
3154 // it's not needed. See if we can spot that too.
3155 if (!LookPassAnd1 &&
3156 !DAG.MaskedValueIsZero(
3157 N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16)))
3161 SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00);
3162 if (OpSizeInBits > 16)
3163 Res = DAG.getNode(ISD::SRL, SDLoc(N), VT, Res,
3164 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
3168 /// Return true if the specified node is an element that makes up a 32-bit
3169 /// packed halfword byteswap.
3170 /// ((x & 0x000000ff) << 8) |
3171 /// ((x & 0x0000ff00) >> 8) |
3172 /// ((x & 0x00ff0000) << 8) |
3173 /// ((x & 0xff000000) >> 8)
3174 static bool isBSwapHWordElement(SDValue N, MutableArrayRef<SDNode *> Parts) {
3175 if (!N.getNode()->hasOneUse())
3178 unsigned Opc = N.getOpcode();
3179 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
3182 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3187 switch (N1C->getZExtValue()) {
3190 case 0xFF: Num = 0; break;
3191 case 0xFF00: Num = 1; break;
3192 case 0xFF0000: Num = 2; break;
3193 case 0xFF000000: Num = 3; break;
3196 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
3197 SDValue N0 = N.getOperand(0);
3198 if (Opc == ISD::AND) {
3199 if (Num == 0 || Num == 2) {
3201 // (x >> 8) & 0xff0000
3202 if (N0.getOpcode() != ISD::SRL)
3204 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3205 if (!C || C->getZExtValue() != 8)
3208 // (x << 8) & 0xff00
3209 // (x << 8) & 0xff000000
3210 if (N0.getOpcode() != ISD::SHL)
3212 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3213 if (!C || C->getZExtValue() != 8)
3216 } else if (Opc == ISD::SHL) {
3218 // (x & 0xff0000) << 8
3219 if (Num != 0 && Num != 2)
3221 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3222 if (!C || C->getZExtValue() != 8)
3224 } else { // Opc == ISD::SRL
3225 // (x & 0xff00) >> 8
3226 // (x & 0xff000000) >> 8
3227 if (Num != 1 && Num != 3)
3229 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3230 if (!C || C->getZExtValue() != 8)
3237 Parts[Num] = N0.getOperand(0).getNode();
3241 /// Match a 32-bit packed halfword bswap. That is
3242 /// ((x & 0x000000ff) << 8) |
3243 /// ((x & 0x0000ff00) >> 8) |
3244 /// ((x & 0x00ff0000) << 8) |
3245 /// ((x & 0xff000000) >> 8)
3246 /// => (rotl (bswap x), 16)
3247 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
3248 if (!LegalOperations)
3251 EVT VT = N->getValueType(0);
3254 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3258 // (or (or (and), (and)), (or (and), (and)))
3259 // (or (or (or (and), (and)), (and)), (and))
3260 if (N0.getOpcode() != ISD::OR)
3262 SDValue N00 = N0.getOperand(0);
3263 SDValue N01 = N0.getOperand(1);
3264 SDNode *Parts[4] = {};
3266 if (N1.getOpcode() == ISD::OR &&
3267 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) {
3268 // (or (or (and), (and)), (or (and), (and)))
3269 SDValue N000 = N00.getOperand(0);
3270 if (!isBSwapHWordElement(N000, Parts))
3273 SDValue N001 = N00.getOperand(1);
3274 if (!isBSwapHWordElement(N001, Parts))
3276 SDValue N010 = N01.getOperand(0);
3277 if (!isBSwapHWordElement(N010, Parts))
3279 SDValue N011 = N01.getOperand(1);
3280 if (!isBSwapHWordElement(N011, Parts))
3283 // (or (or (or (and), (and)), (and)), (and))
3284 if (!isBSwapHWordElement(N1, Parts))
3286 if (!isBSwapHWordElement(N01, Parts))
3288 if (N00.getOpcode() != ISD::OR)
3290 SDValue N000 = N00.getOperand(0);
3291 if (!isBSwapHWordElement(N000, Parts))
3293 SDValue N001 = N00.getOperand(1);
3294 if (!isBSwapHWordElement(N001, Parts))
3298 // Make sure the parts are all coming from the same node.
3299 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
3302 SDValue BSwap = DAG.getNode(ISD::BSWAP, SDLoc(N), VT,
3303 SDValue(Parts[0],0));
3305 // Result of the bswap should be rotated by 16. If it's not legal, then
3306 // do (x << 16) | (x >> 16).
3307 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
3308 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3309 return DAG.getNode(ISD::ROTL, SDLoc(N), VT, BSwap, ShAmt);
3310 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3311 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, BSwap, ShAmt);
3312 return DAG.getNode(ISD::OR, SDLoc(N), VT,
3313 DAG.getNode(ISD::SHL, SDLoc(N), VT, BSwap, ShAmt),
3314 DAG.getNode(ISD::SRL, SDLoc(N), VT, BSwap, ShAmt));
3317 SDValue DAGCombiner::visitOR(SDNode *N) {
3318 SDValue N0 = N->getOperand(0);
3319 SDValue N1 = N->getOperand(1);
3320 SDValue LL, LR, RL, RR, CC0, CC1;
3321 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3322 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3323 EVT VT = N1.getValueType();
3326 if (VT.isVector()) {
3327 SDValue FoldedVOp = SimplifyVBinOp(N);
3328 if (FoldedVOp.getNode()) return FoldedVOp;
3330 // fold (or x, 0) -> x, vector edition
3331 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3333 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3336 // fold (or x, -1) -> -1, vector edition
3337 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3338 // do not return N0, because undef node may exist in N0
3339 return DAG.getConstant(
3340 APInt::getAllOnesValue(
3341 N0.getValueType().getScalarType().getSizeInBits()),
3343 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3344 // do not return N1, because undef node may exist in N1
3345 return DAG.getConstant(
3346 APInt::getAllOnesValue(
3347 N1.getValueType().getScalarType().getSizeInBits()),
3350 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask1)
3351 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf B, A, Mask2)
3352 // Do this only if the resulting shuffle is legal.
3353 if (isa<ShuffleVectorSDNode>(N0) &&
3354 isa<ShuffleVectorSDNode>(N1) &&
3355 // Avoid folding a node with illegal type.
3356 TLI.isTypeLegal(VT) &&
3357 N0->getOperand(1) == N1->getOperand(1) &&
3358 ISD::isBuildVectorAllZeros(N0.getOperand(1).getNode())) {
3359 bool CanFold = true;
3360 unsigned NumElts = VT.getVectorNumElements();
3361 const ShuffleVectorSDNode *SV0 = cast<ShuffleVectorSDNode>(N0);
3362 const ShuffleVectorSDNode *SV1 = cast<ShuffleVectorSDNode>(N1);
3363 // We construct two shuffle masks:
3364 // - Mask1 is a shuffle mask for a shuffle with N0 as the first operand
3365 // and N1 as the second operand.
3366 // - Mask2 is a shuffle mask for a shuffle with N1 as the first operand
3367 // and N0 as the second operand.
3368 // We do this because OR is commutable and therefore there might be
3369 // two ways to fold this node into a shuffle.
3370 SmallVector<int,4> Mask1;
3371 SmallVector<int,4> Mask2;
3373 for (unsigned i = 0; i != NumElts && CanFold; ++i) {
3374 int M0 = SV0->getMaskElt(i);
3375 int M1 = SV1->getMaskElt(i);
3377 // Both shuffle indexes are undef. Propagate Undef.
3378 if (M0 < 0 && M1 < 0) {
3379 Mask1.push_back(M0);
3380 Mask2.push_back(M0);
3384 if (M0 < 0 || M1 < 0 ||
3385 (M0 < (int)NumElts && M1 < (int)NumElts) ||
3386 (M0 >= (int)NumElts && M1 >= (int)NumElts)) {
3391 Mask1.push_back(M0 < (int)NumElts ? M0 : M1 + NumElts);
3392 Mask2.push_back(M1 < (int)NumElts ? M1 : M0 + NumElts);
3396 // Fold this sequence only if the resulting shuffle is 'legal'.
3397 if (TLI.isShuffleMaskLegal(Mask1, VT))
3398 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0),
3399 N1->getOperand(0), &Mask1[0]);
3400 if (TLI.isShuffleMaskLegal(Mask2, VT))
3401 return DAG.getVectorShuffle(VT, SDLoc(N), N1->getOperand(0),
3402 N0->getOperand(0), &Mask2[0]);
3407 // fold (or x, undef) -> -1
3408 if (!LegalOperations &&
3409 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3410 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3411 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
3413 // fold (or c1, c2) -> c1|c2
3415 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
3416 // canonicalize constant to RHS
3418 return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0);
3419 // fold (or x, 0) -> x
3420 if (N1C && N1C->isNullValue())
3422 // fold (or x, -1) -> -1
3423 if (N1C && N1C->isAllOnesValue())
3425 // fold (or x, c) -> c iff (x & ~c) == 0
3426 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3429 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3430 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
3431 if (BSwap.getNode())
3433 BSwap = MatchBSwapHWordLow(N, N0, N1);
3434 if (BSwap.getNode())
3438 SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1);
3441 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3442 // iff (c1 & c2) == 0.
3443 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3444 isa<ConstantSDNode>(N0.getOperand(1))) {
3445 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3446 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) {
3447 if (SDValue COR = DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1))
3449 ISD::AND, SDLoc(N), VT,
3450 DAG.getNode(ISD::OR, SDLoc(N0), VT, N0.getOperand(0), N1), COR);
3454 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3455 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3456 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3457 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3459 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3460 LL.getValueType().isInteger()) {
3461 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3462 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3463 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3464 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3465 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(LR),
3466 LR.getValueType(), LL, RL);
3467 AddToWorklist(ORNode.getNode());
3468 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
3470 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3471 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3472 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3473 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3474 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(LR),
3475 LR.getValueType(), LL, RL);
3476 AddToWorklist(ANDNode.getNode());
3477 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
3480 // canonicalize equivalent to ll == rl
3481 if (LL == RR && LR == RL) {
3482 Op1 = ISD::getSetCCSwappedOperands(Op1);
3485 if (LL == RL && LR == RR) {
3486 bool isInteger = LL.getValueType().isInteger();
3487 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3488 if (Result != ISD::SETCC_INVALID &&
3489 (!LegalOperations ||
3490 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
3491 TLI.isOperationLegal(ISD::SETCC,
3492 getSetCCResultType(N0.getValueType())))))
3493 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
3498 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3499 if (N0.getOpcode() == N1.getOpcode()) {
3500 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3501 if (Tmp.getNode()) return Tmp;
3504 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3505 if (N0.getOpcode() == ISD::AND &&
3506 N1.getOpcode() == ISD::AND &&
3507 N0.getOperand(1).getOpcode() == ISD::Constant &&
3508 N1.getOperand(1).getOpcode() == ISD::Constant &&
3509 // Don't increase # computations.
3510 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3511 // We can only do this xform if we know that bits from X that are set in C2
3512 // but not in C1 are already zero. Likewise for Y.
3513 const APInt &LHSMask =
3514 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3515 const APInt &RHSMask =
3516 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3518 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3519 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3520 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3521 N0.getOperand(0), N1.getOperand(0));
3522 return DAG.getNode(ISD::AND, SDLoc(N), VT, X,
3523 DAG.getConstant(LHSMask | RHSMask, VT));
3527 // See if this is some rotate idiom.
3528 if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N)))
3529 return SDValue(Rot, 0);
3531 // Simplify the operands using demanded-bits information.
3532 if (!VT.isVector() &&
3533 SimplifyDemandedBits(SDValue(N, 0)))
3534 return SDValue(N, 0);
3539 /// Match "(X shl/srl V1) & V2" where V2 may not be present.
3540 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3541 if (Op.getOpcode() == ISD::AND) {
3542 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3543 Mask = Op.getOperand(1);
3544 Op = Op.getOperand(0);
3550 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3558 // Return true if we can prove that, whenever Neg and Pos are both in the
3559 // range [0, OpSize), Neg == (Pos == 0 ? 0 : OpSize - Pos). This means that
3560 // for two opposing shifts shift1 and shift2 and a value X with OpBits bits:
3562 // (or (shift1 X, Neg), (shift2 X, Pos))
3564 // reduces to a rotate in direction shift2 by Pos or (equivalently) a rotate
3565 // in direction shift1 by Neg. The range [0, OpSize) means that we only need
3566 // to consider shift amounts with defined behavior.
3567 static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned OpSize) {
3568 // If OpSize is a power of 2 then:
3570 // (a) (Pos == 0 ? 0 : OpSize - Pos) == (OpSize - Pos) & (OpSize - 1)
3571 // (b) Neg == Neg & (OpSize - 1) whenever Neg is in [0, OpSize).
3573 // So if OpSize is a power of 2 and Neg is (and Neg', OpSize-1), we check
3574 // for the stronger condition:
3576 // Neg & (OpSize - 1) == (OpSize - Pos) & (OpSize - 1) [A]
3578 // for all Neg and Pos. Since Neg & (OpSize - 1) == Neg' & (OpSize - 1)
3579 // we can just replace Neg with Neg' for the rest of the function.
3581 // In other cases we check for the even stronger condition:
3583 // Neg == OpSize - Pos [B]
3585 // for all Neg and Pos. Note that the (or ...) then invokes undefined
3586 // behavior if Pos == 0 (and consequently Neg == OpSize).
3588 // We could actually use [A] whenever OpSize is a power of 2, but the
3589 // only extra cases that it would match are those uninteresting ones
3590 // where Neg and Pos are never in range at the same time. E.g. for
3591 // OpSize == 32, using [A] would allow a Neg of the form (sub 64, Pos)
3592 // as well as (sub 32, Pos), but:
3594 // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos))
3596 // always invokes undefined behavior for 32-bit X.
3598 // Below, Mask == OpSize - 1 when using [A] and is all-ones otherwise.
3599 unsigned MaskLoBits = 0;
3600 if (Neg.getOpcode() == ISD::AND &&
3601 isPowerOf2_64(OpSize) &&
3602 Neg.getOperand(1).getOpcode() == ISD::Constant &&
3603 cast<ConstantSDNode>(Neg.getOperand(1))->getAPIntValue() == OpSize - 1) {
3604 Neg = Neg.getOperand(0);
3605 MaskLoBits = Log2_64(OpSize);
3608 // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1.
3609 if (Neg.getOpcode() != ISD::SUB)
3611 ConstantSDNode *NegC = dyn_cast<ConstantSDNode>(Neg.getOperand(0));
3614 SDValue NegOp1 = Neg.getOperand(1);
3616 // On the RHS of [A], if Pos is Pos' & (OpSize - 1), just replace Pos with
3617 // Pos'. The truncation is redundant for the purpose of the equality.
3619 Pos.getOpcode() == ISD::AND &&
3620 Pos.getOperand(1).getOpcode() == ISD::Constant &&
3621 cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() == OpSize - 1)
3622 Pos = Pos.getOperand(0);
3624 // The condition we need is now:
3626 // (NegC - NegOp1) & Mask == (OpSize - Pos) & Mask
3628 // If NegOp1 == Pos then we need:
3630 // OpSize & Mask == NegC & Mask
3632 // (because "x & Mask" is a truncation and distributes through subtraction).
3635 Width = NegC->getAPIntValue();
3636 // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC.
3637 // Then the condition we want to prove becomes:
3639 // (NegC - NegOp1) & Mask == (OpSize - (NegOp1 + PosC)) & Mask
3641 // which, again because "x & Mask" is a truncation, becomes:
3643 // NegC & Mask == (OpSize - PosC) & Mask
3644 // OpSize & Mask == (NegC + PosC) & Mask
3645 else if (Pos.getOpcode() == ISD::ADD &&
3646 Pos.getOperand(0) == NegOp1 &&
3647 Pos.getOperand(1).getOpcode() == ISD::Constant)
3648 Width = (cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() +
3649 NegC->getAPIntValue());
3653 // Now we just need to check that OpSize & Mask == Width & Mask.
3655 // Opsize & Mask is 0 since Mask is Opsize - 1.
3656 return Width.getLoBits(MaskLoBits) == 0;
3657 return Width == OpSize;
3660 // A subroutine of MatchRotate used once we have found an OR of two opposite
3661 // shifts of Shifted. If Neg == <operand size> - Pos then the OR reduces
3662 // to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the
3663 // former being preferred if supported. InnerPos and InnerNeg are Pos and
3664 // Neg with outer conversions stripped away.
3665 SDNode *DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos,
3666 SDValue Neg, SDValue InnerPos,
3667 SDValue InnerNeg, unsigned PosOpcode,
3668 unsigned NegOpcode, SDLoc DL) {
3669 // fold (or (shl x, (*ext y)),
3670 // (srl x, (*ext (sub 32, y)))) ->
3671 // (rotl x, y) or (rotr x, (sub 32, y))
3673 // fold (or (shl x, (*ext (sub 32, y))),
3674 // (srl x, (*ext y))) ->
3675 // (rotr x, y) or (rotl x, (sub 32, y))
3676 EVT VT = Shifted.getValueType();
3677 if (matchRotateSub(InnerPos, InnerNeg, VT.getSizeInBits())) {
3678 bool HasPos = TLI.isOperationLegalOrCustom(PosOpcode, VT);
3679 return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted,
3680 HasPos ? Pos : Neg).getNode();
3686 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3687 // idioms for rotate, and if the target supports rotation instructions, generate
3689 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL) {
3690 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3691 EVT VT = LHS.getValueType();
3692 if (!TLI.isTypeLegal(VT)) return nullptr;
3694 // The target must have at least one rotate flavor.
3695 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3696 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3697 if (!HasROTL && !HasROTR) return nullptr;
3699 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3700 SDValue LHSShift; // The shift.
3701 SDValue LHSMask; // AND value if any.
3702 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3703 return nullptr; // Not part of a rotate.
3705 SDValue RHSShift; // The shift.
3706 SDValue RHSMask; // AND value if any.
3707 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3708 return nullptr; // Not part of a rotate.
3710 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3711 return nullptr; // Not shifting the same value.
3713 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3714 return nullptr; // Shifts must disagree.
3716 // Canonicalize shl to left side in a shl/srl pair.
3717 if (RHSShift.getOpcode() == ISD::SHL) {
3718 std::swap(LHS, RHS);
3719 std::swap(LHSShift, RHSShift);
3720 std::swap(LHSMask , RHSMask );
3723 unsigned OpSizeInBits = VT.getSizeInBits();
3724 SDValue LHSShiftArg = LHSShift.getOperand(0);
3725 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3726 SDValue RHSShiftArg = RHSShift.getOperand(0);
3727 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3729 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3730 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3731 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3732 RHSShiftAmt.getOpcode() == ISD::Constant) {
3733 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3734 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3735 if ((LShVal + RShVal) != OpSizeInBits)
3738 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3739 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
3741 // If there is an AND of either shifted operand, apply it to the result.
3742 if (LHSMask.getNode() || RHSMask.getNode()) {
3743 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3745 if (LHSMask.getNode()) {
3746 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3747 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3749 if (RHSMask.getNode()) {
3750 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3751 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3754 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3757 return Rot.getNode();
3760 // If there is a mask here, and we have a variable shift, we can't be sure
3761 // that we're masking out the right stuff.
3762 if (LHSMask.getNode() || RHSMask.getNode())
3765 // If the shift amount is sign/zext/any-extended just peel it off.
3766 SDValue LExtOp0 = LHSShiftAmt;
3767 SDValue RExtOp0 = RHSShiftAmt;
3768 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3769 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3770 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3771 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3772 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3773 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3774 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3775 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3776 LExtOp0 = LHSShiftAmt.getOperand(0);
3777 RExtOp0 = RHSShiftAmt.getOperand(0);
3780 SDNode *TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt,
3781 LExtOp0, RExtOp0, ISD::ROTL, ISD::ROTR, DL);
3785 SDNode *TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt,
3786 RExtOp0, LExtOp0, ISD::ROTR, ISD::ROTL, DL);
3793 SDValue DAGCombiner::visitXOR(SDNode *N) {
3794 SDValue N0 = N->getOperand(0);
3795 SDValue N1 = N->getOperand(1);
3796 SDValue LHS, RHS, CC;
3797 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3798 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3799 EVT VT = N0.getValueType();
3802 if (VT.isVector()) {
3803 SDValue FoldedVOp = SimplifyVBinOp(N);
3804 if (FoldedVOp.getNode()) return FoldedVOp;
3806 // fold (xor x, 0) -> x, vector edition
3807 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3809 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3813 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3814 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3815 return DAG.getConstant(0, VT);
3816 // fold (xor x, undef) -> undef
3817 if (N0.getOpcode() == ISD::UNDEF)
3819 if (N1.getOpcode() == ISD::UNDEF)
3821 // fold (xor c1, c2) -> c1^c2
3823 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3824 // canonicalize constant to RHS
3826 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
3827 // fold (xor x, 0) -> x
3828 if (N1C && N1C->isNullValue())
3831 SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1);
3835 // fold !(x cc y) -> (x !cc y)
3836 if (TLI.isConstTrueVal(N1.getNode()) && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3837 bool isInt = LHS.getValueType().isInteger();
3838 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3841 if (!LegalOperations ||
3842 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) {
3843 switch (N0.getOpcode()) {
3845 llvm_unreachable("Unhandled SetCC Equivalent!");
3847 return DAG.getSetCC(SDLoc(N), VT, LHS, RHS, NotCC);
3848 case ISD::SELECT_CC:
3849 return DAG.getSelectCC(SDLoc(N), LHS, RHS, N0.getOperand(2),
3850 N0.getOperand(3), NotCC);
3855 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3856 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3857 N0.getNode()->hasOneUse() &&
3858 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3859 SDValue V = N0.getOperand(0);
3860 V = DAG.getNode(ISD::XOR, SDLoc(N0), V.getValueType(), V,
3861 DAG.getConstant(1, V.getValueType()));
3862 AddToWorklist(V.getNode());
3863 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V);
3866 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3867 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3868 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3869 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3870 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3871 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3872 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3873 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3874 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
3875 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3878 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3879 if (N1C && N1C->isAllOnesValue() &&
3880 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3881 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3882 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3883 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3884 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3885 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3886 AddToWorklist(LHS.getNode()); AddToWorklist(RHS.getNode());
3887 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3890 // fold (xor (and x, y), y) -> (and (not x), y)
3891 if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3892 N0->getOperand(1) == N1) {
3893 SDValue X = N0->getOperand(0);
3894 SDValue NotX = DAG.getNOT(SDLoc(X), X, VT);
3895 AddToWorklist(NotX.getNode());
3896 return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1);
3898 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3899 if (N1C && N0.getOpcode() == ISD::XOR) {
3900 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3901 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3903 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(1),
3904 DAG.getConstant(N1C->getAPIntValue() ^
3905 N00C->getAPIntValue(), VT));
3907 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(0),
3908 DAG.getConstant(N1C->getAPIntValue() ^
3909 N01C->getAPIntValue(), VT));
3911 // fold (xor x, x) -> 0
3913 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
3915 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
3916 if (N0.getOpcode() == N1.getOpcode()) {
3917 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3918 if (Tmp.getNode()) return Tmp;
3921 // Simplify the expression using non-local knowledge.
3922 if (!VT.isVector() &&
3923 SimplifyDemandedBits(SDValue(N, 0)))
3924 return SDValue(N, 0);
3929 /// Handle transforms common to the three shifts, when the shift amount is a
3931 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, ConstantSDNode *Amt) {
3932 // We can't and shouldn't fold opaque constants.
3933 if (Amt->isOpaque())
3936 SDNode *LHS = N->getOperand(0).getNode();
3937 if (!LHS->hasOneUse()) return SDValue();
3939 // We want to pull some binops through shifts, so that we have (and (shift))
3940 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
3941 // thing happens with address calculations, so it's important to canonicalize
3943 bool HighBitSet = false; // Can we transform this if the high bit is set?
3945 switch (LHS->getOpcode()) {
3946 default: return SDValue();
3949 HighBitSet = false; // We can only transform sra if the high bit is clear.
3952 HighBitSet = true; // We can only transform sra if the high bit is set.
3955 if (N->getOpcode() != ISD::SHL)
3956 return SDValue(); // only shl(add) not sr[al](add).
3957 HighBitSet = false; // We can only transform sra if the high bit is clear.
3961 // We require the RHS of the binop to be a constant and not opaque as well.
3962 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
3963 if (!BinOpCst || BinOpCst->isOpaque()) return SDValue();
3965 // FIXME: disable this unless the input to the binop is a shift by a constant.
3966 // If it is not a shift, it pessimizes some common cases like:
3968 // void foo(int *X, int i) { X[i & 1235] = 1; }
3969 // int bar(int *X, int i) { return X[i & 255]; }
3970 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
3971 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
3972 BinOpLHSVal->getOpcode() != ISD::SRA &&
3973 BinOpLHSVal->getOpcode() != ISD::SRL) ||
3974 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
3977 EVT VT = N->getValueType(0);
3979 // If this is a signed shift right, and the high bit is modified by the
3980 // logical operation, do not perform the transformation. The highBitSet
3981 // boolean indicates the value of the high bit of the constant which would
3982 // cause it to be modified for this operation.
3983 if (N->getOpcode() == ISD::SRA) {
3984 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
3985 if (BinOpRHSSignSet != HighBitSet)
3989 if (!TLI.isDesirableToCommuteWithShift(LHS))
3992 // Fold the constants, shifting the binop RHS by the shift amount.
3993 SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)),
3995 LHS->getOperand(1), N->getOperand(1));
3996 assert(isa<ConstantSDNode>(NewRHS) && "Folding was not successful!");
3998 // Create the new shift.
3999 SDValue NewShift = DAG.getNode(N->getOpcode(),
4000 SDLoc(LHS->getOperand(0)),
4001 VT, LHS->getOperand(0), N->getOperand(1));
4003 // Create the new binop.
4004 return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS);
4007 SDValue DAGCombiner::distributeTruncateThroughAnd(SDNode *N) {
4008 assert(N->getOpcode() == ISD::TRUNCATE);
4009 assert(N->getOperand(0).getOpcode() == ISD::AND);
4011 // (truncate:TruncVT (and N00, N01C)) -> (and (truncate:TruncVT N00), TruncC)
4012 if (N->hasOneUse() && N->getOperand(0).hasOneUse()) {
4013 SDValue N01 = N->getOperand(0).getOperand(1);
4015 if (ConstantSDNode *N01C = isConstOrConstSplat(N01)) {
4016 EVT TruncVT = N->getValueType(0);
4017 SDValue N00 = N->getOperand(0).getOperand(0);
4018 APInt TruncC = N01C->getAPIntValue();
4019 TruncC = TruncC.trunc(TruncVT.getScalarSizeInBits());
4021 return DAG.getNode(ISD::AND, SDLoc(N), TruncVT,
4022 DAG.getNode(ISD::TRUNCATE, SDLoc(N), TruncVT, N00),
4023 DAG.getConstant(TruncC, TruncVT));
4030 SDValue DAGCombiner::visitRotate(SDNode *N) {
4031 // fold (rot* x, (trunc (and y, c))) -> (rot* x, (and (trunc y), (trunc c))).
4032 if (N->getOperand(1).getOpcode() == ISD::TRUNCATE &&
4033 N->getOperand(1).getOperand(0).getOpcode() == ISD::AND) {
4034 SDValue NewOp1 = distributeTruncateThroughAnd(N->getOperand(1).getNode());
4035 if (NewOp1.getNode())
4036 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
4037 N->getOperand(0), NewOp1);
4042 SDValue DAGCombiner::visitSHL(SDNode *N) {
4043 SDValue N0 = N->getOperand(0);
4044 SDValue N1 = N->getOperand(1);
4045 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4046 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4047 EVT VT = N0.getValueType();
4048 unsigned OpSizeInBits = VT.getScalarSizeInBits();
4051 if (VT.isVector()) {
4052 SDValue FoldedVOp = SimplifyVBinOp(N);
4053 if (FoldedVOp.getNode()) return FoldedVOp;
4055 BuildVectorSDNode *N1CV = dyn_cast<BuildVectorSDNode>(N1);
4056 // If setcc produces all-one true value then:
4057 // (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<<N1CV)
4058 if (N1CV && N1CV->isConstant()) {
4059 if (N0.getOpcode() == ISD::AND) {
4060 SDValue N00 = N0->getOperand(0);
4061 SDValue N01 = N0->getOperand(1);
4062 BuildVectorSDNode *N01CV = dyn_cast<BuildVectorSDNode>(N01);
4064 if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC &&
4065 TLI.getBooleanContents(N00.getOperand(0).getValueType()) ==
4066 TargetLowering::ZeroOrNegativeOneBooleanContent) {
4067 if (SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, VT, N01CV, N1CV))
4068 return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C);
4071 N1C = isConstOrConstSplat(N1);
4076 // fold (shl c1, c2) -> c1<<c2
4078 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
4079 // fold (shl 0, x) -> 0
4080 if (N0C && N0C->isNullValue())
4082 // fold (shl x, c >= size(x)) -> undef
4083 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4084 return DAG.getUNDEF(VT);
4085 // fold (shl x, 0) -> x
4086 if (N1C && N1C->isNullValue())
4088 // fold (shl undef, x) -> 0
4089 if (N0.getOpcode() == ISD::UNDEF)
4090 return DAG.getConstant(0, VT);
4091 // if (shl x, c) is known to be zero, return 0
4092 if (DAG.MaskedValueIsZero(SDValue(N, 0),
4093 APInt::getAllOnesValue(OpSizeInBits)))
4094 return DAG.getConstant(0, VT);
4095 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
4096 if (N1.getOpcode() == ISD::TRUNCATE &&
4097 N1.getOperand(0).getOpcode() == ISD::AND) {
4098 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4099 if (NewOp1.getNode())
4100 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, NewOp1);
4103 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4104 return SDValue(N, 0);
4106 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
4107 if (N1C && N0.getOpcode() == ISD::SHL) {
4108 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4109 uint64_t c1 = N0C1->getZExtValue();
4110 uint64_t c2 = N1C->getZExtValue();
4111 if (c1 + c2 >= OpSizeInBits)
4112 return DAG.getConstant(0, VT);
4113 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4114 DAG.getConstant(c1 + c2, N1.getValueType()));
4118 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
4119 // For this to be valid, the second form must not preserve any of the bits
4120 // that are shifted out by the inner shift in the first form. This means
4121 // the outer shift size must be >= the number of bits added by the ext.
4122 // As a corollary, we don't care what kind of ext it is.
4123 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
4124 N0.getOpcode() == ISD::ANY_EXTEND ||
4125 N0.getOpcode() == ISD::SIGN_EXTEND) &&
4126 N0.getOperand(0).getOpcode() == ISD::SHL) {
4127 SDValue N0Op0 = N0.getOperand(0);
4128 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4129 uint64_t c1 = N0Op0C1->getZExtValue();
4130 uint64_t c2 = N1C->getZExtValue();
4131 EVT InnerShiftVT = N0Op0.getValueType();
4132 uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits();
4133 if (c2 >= OpSizeInBits - InnerShiftSize) {
4134 if (c1 + c2 >= OpSizeInBits)
4135 return DAG.getConstant(0, VT);
4136 return DAG.getNode(ISD::SHL, SDLoc(N0), VT,
4137 DAG.getNode(N0.getOpcode(), SDLoc(N0), VT,
4138 N0Op0->getOperand(0)),
4139 DAG.getConstant(c1 + c2, N1.getValueType()));
4144 // fold (shl (zext (srl x, C)), C) -> (zext (shl (srl x, C), C))
4145 // Only fold this if the inner zext has no other uses to avoid increasing
4146 // the total number of instructions.
4147 if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() &&
4148 N0.getOperand(0).getOpcode() == ISD::SRL) {
4149 SDValue N0Op0 = N0.getOperand(0);
4150 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4151 uint64_t c1 = N0Op0C1->getZExtValue();
4152 if (c1 < VT.getScalarSizeInBits()) {
4153 uint64_t c2 = N1C->getZExtValue();
4155 SDValue NewOp0 = N0.getOperand(0);
4156 EVT CountVT = NewOp0.getOperand(1).getValueType();
4157 SDValue NewSHL = DAG.getNode(ISD::SHL, SDLoc(N), NewOp0.getValueType(),
4158 NewOp0, DAG.getConstant(c2, CountVT));
4159 AddToWorklist(NewSHL.getNode());
4160 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
4166 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
4167 // (and (srl x, (sub c1, c2), MASK)
4168 // Only fold this if the inner shift has no other uses -- if it does, folding
4169 // this will increase the total number of instructions.
4170 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
4171 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4172 uint64_t c1 = N0C1->getZExtValue();
4173 if (c1 < OpSizeInBits) {
4174 uint64_t c2 = N1C->getZExtValue();
4175 APInt Mask = APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - c1);
4178 Mask = Mask.shl(c2 - c1);
4179 Shift = DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4180 DAG.getConstant(c2 - c1, N1.getValueType()));
4182 Mask = Mask.lshr(c1 - c2);
4183 Shift = DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4184 DAG.getConstant(c1 - c2, N1.getValueType()));
4186 return DAG.getNode(ISD::AND, SDLoc(N0), VT, Shift,
4187 DAG.getConstant(Mask, VT));
4191 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
4192 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
4193 unsigned BitSize = VT.getScalarSizeInBits();
4194 SDValue HiBitsMask =
4195 DAG.getConstant(APInt::getHighBitsSet(BitSize,
4196 BitSize - N1C->getZExtValue()), VT);
4197 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4201 // fold (shl (add x, c1), c2) -> (add (shl x, c2), c1 << c2)
4202 // Variant of version done on multiply, except mul by a power of 2 is turned
4205 if (N1C && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
4206 (isa<ConstantSDNode>(N0.getOperand(1)) ||
4207 isConstantSplatVector(N0.getOperand(1).getNode(), Val))) {
4208 SDValue Shl0 = DAG.getNode(ISD::SHL, SDLoc(N0), VT, N0.getOperand(0), N1);
4209 SDValue Shl1 = DAG.getNode(ISD::SHL, SDLoc(N1), VT, N0.getOperand(1), N1);
4210 return DAG.getNode(ISD::ADD, SDLoc(N), VT, Shl0, Shl1);
4214 SDValue NewSHL = visitShiftByConstant(N, N1C);
4215 if (NewSHL.getNode())
4222 SDValue DAGCombiner::visitSRA(SDNode *N) {
4223 SDValue N0 = N->getOperand(0);
4224 SDValue N1 = N->getOperand(1);
4225 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4226 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4227 EVT VT = N0.getValueType();
4228 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4231 if (VT.isVector()) {
4232 SDValue FoldedVOp = SimplifyVBinOp(N);
4233 if (FoldedVOp.getNode()) return FoldedVOp;
4235 N1C = isConstOrConstSplat(N1);
4238 // fold (sra c1, c2) -> (sra c1, c2)
4240 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
4241 // fold (sra 0, x) -> 0
4242 if (N0C && N0C->isNullValue())
4244 // fold (sra -1, x) -> -1
4245 if (N0C && N0C->isAllOnesValue())
4247 // fold (sra x, (setge c, size(x))) -> undef
4248 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4249 return DAG.getUNDEF(VT);
4250 // fold (sra x, 0) -> x
4251 if (N1C && N1C->isNullValue())
4253 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
4255 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
4256 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
4257 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
4259 ExtVT = EVT::getVectorVT(*DAG.getContext(),
4260 ExtVT, VT.getVectorNumElements());
4261 if ((!LegalOperations ||
4262 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
4263 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
4264 N0.getOperand(0), DAG.getValueType(ExtVT));
4267 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
4268 if (N1C && N0.getOpcode() == ISD::SRA) {
4269 if (ConstantSDNode *C1 = isConstOrConstSplat(N0.getOperand(1))) {
4270 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
4271 if (Sum >= OpSizeInBits)
4272 Sum = OpSizeInBits - 1;
4273 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0.getOperand(0),
4274 DAG.getConstant(Sum, N1.getValueType()));
4278 // fold (sra (shl X, m), (sub result_size, n))
4279 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
4280 // result_size - n != m.
4281 // If truncate is free for the target sext(shl) is likely to result in better
4283 if (N0.getOpcode() == ISD::SHL && N1C) {
4284 // Get the two constanst of the shifts, CN0 = m, CN = n.
4285 const ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1));
4287 LLVMContext &Ctx = *DAG.getContext();
4288 // Determine what the truncate's result bitsize and type would be.
4289 EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - N1C->getZExtValue());
4292 TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorNumElements());
4294 // Determine the residual right-shift amount.
4295 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
4297 // If the shift is not a no-op (in which case this should be just a sign
4298 // extend already), the truncated to type is legal, sign_extend is legal
4299 // on that type, and the truncate to that type is both legal and free,
4300 // perform the transform.
4301 if ((ShiftAmt > 0) &&
4302 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
4303 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
4304 TLI.isTruncateFree(VT, TruncVT)) {
4306 SDValue Amt = DAG.getConstant(ShiftAmt,
4307 getShiftAmountTy(N0.getOperand(0).getValueType()));
4308 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0), VT,
4309 N0.getOperand(0), Amt);
4310 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), TruncVT,
4312 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N),
4313 N->getValueType(0), Trunc);
4318 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
4319 if (N1.getOpcode() == ISD::TRUNCATE &&
4320 N1.getOperand(0).getOpcode() == ISD::AND) {
4321 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4322 if (NewOp1.getNode())
4323 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, NewOp1);
4326 // fold (sra (trunc (srl x, c1)), c2) -> (trunc (sra x, c1 + c2))
4327 // if c1 is equal to the number of bits the trunc removes
4328 if (N0.getOpcode() == ISD::TRUNCATE &&
4329 (N0.getOperand(0).getOpcode() == ISD::SRL ||
4330 N0.getOperand(0).getOpcode() == ISD::SRA) &&
4331 N0.getOperand(0).hasOneUse() &&
4332 N0.getOperand(0).getOperand(1).hasOneUse() &&
4334 SDValue N0Op0 = N0.getOperand(0);
4335 if (ConstantSDNode *LargeShift = isConstOrConstSplat(N0Op0.getOperand(1))) {
4336 unsigned LargeShiftVal = LargeShift->getZExtValue();
4337 EVT LargeVT = N0Op0.getValueType();
4339 if (LargeVT.getScalarSizeInBits() - OpSizeInBits == LargeShiftVal) {
4341 DAG.getConstant(LargeShiftVal + N1C->getZExtValue(),
4342 getShiftAmountTy(N0Op0.getOperand(0).getValueType()));
4343 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), LargeVT,
4344 N0Op0.getOperand(0), Amt);
4345 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, SRA);
4350 // Simplify, based on bits shifted out of the LHS.
4351 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4352 return SDValue(N, 0);
4355 // If the sign bit is known to be zero, switch this to a SRL.
4356 if (DAG.SignBitIsZero(N0))
4357 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1);
4360 SDValue NewSRA = visitShiftByConstant(N, N1C);
4361 if (NewSRA.getNode())
4368 SDValue DAGCombiner::visitSRL(SDNode *N) {
4369 SDValue N0 = N->getOperand(0);
4370 SDValue N1 = N->getOperand(1);
4371 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4372 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4373 EVT VT = N0.getValueType();
4374 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4377 if (VT.isVector()) {
4378 SDValue FoldedVOp = SimplifyVBinOp(N);
4379 if (FoldedVOp.getNode()) return FoldedVOp;
4381 N1C = isConstOrConstSplat(N1);
4384 // fold (srl c1, c2) -> c1 >>u c2
4386 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
4387 // fold (srl 0, x) -> 0
4388 if (N0C && N0C->isNullValue())
4390 // fold (srl x, c >= size(x)) -> undef
4391 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4392 return DAG.getUNDEF(VT);
4393 // fold (srl x, 0) -> x
4394 if (N1C && N1C->isNullValue())
4396 // if (srl x, c) is known to be zero, return 0
4397 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
4398 APInt::getAllOnesValue(OpSizeInBits)))
4399 return DAG.getConstant(0, VT);
4401 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
4402 if (N1C && N0.getOpcode() == ISD::SRL) {
4403 if (ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1))) {
4404 uint64_t c1 = N01C->getZExtValue();
4405 uint64_t c2 = N1C->getZExtValue();
4406 if (c1 + c2 >= OpSizeInBits)
4407 return DAG.getConstant(0, VT);
4408 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4409 DAG.getConstant(c1 + c2, N1.getValueType()));
4413 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
4414 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
4415 N0.getOperand(0).getOpcode() == ISD::SRL &&
4416 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
4418 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
4419 uint64_t c2 = N1C->getZExtValue();
4420 EVT InnerShiftVT = N0.getOperand(0).getValueType();
4421 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
4422 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
4423 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
4424 if (c1 + OpSizeInBits == InnerShiftSize) {
4425 if (c1 + c2 >= InnerShiftSize)
4426 return DAG.getConstant(0, VT);
4427 return DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT,
4428 DAG.getNode(ISD::SRL, SDLoc(N0), InnerShiftVT,
4429 N0.getOperand(0)->getOperand(0),
4430 DAG.getConstant(c1 + c2, ShiftCountVT)));
4434 // fold (srl (shl x, c), c) -> (and x, cst2)
4435 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1) {
4436 unsigned BitSize = N0.getScalarValueSizeInBits();
4437 if (BitSize <= 64) {
4438 uint64_t ShAmt = N1C->getZExtValue() + 64 - BitSize;
4439 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4440 DAG.getConstant(~0ULL >> ShAmt, VT));
4444 // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask)
4445 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
4446 // Shifting in all undef bits?
4447 EVT SmallVT = N0.getOperand(0).getValueType();
4448 unsigned BitSize = SmallVT.getScalarSizeInBits();
4449 if (N1C->getZExtValue() >= BitSize)
4450 return DAG.getUNDEF(VT);
4452 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
4453 uint64_t ShiftAmt = N1C->getZExtValue();
4454 SDValue SmallShift = DAG.getNode(ISD::SRL, SDLoc(N0), SmallVT,
4456 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
4457 AddToWorklist(SmallShift.getNode());
4458 APInt Mask = APInt::getAllOnesValue(OpSizeInBits).lshr(ShiftAmt);
4459 return DAG.getNode(ISD::AND, SDLoc(N), VT,
4460 DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, SmallShift),
4461 DAG.getConstant(Mask, VT));
4465 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
4466 // bit, which is unmodified by sra.
4467 if (N1C && N1C->getZExtValue() + 1 == OpSizeInBits) {
4468 if (N0.getOpcode() == ISD::SRA)
4469 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1);
4472 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
4473 if (N1C && N0.getOpcode() == ISD::CTLZ &&
4474 N1C->getAPIntValue() == Log2_32(OpSizeInBits)) {
4475 APInt KnownZero, KnownOne;
4476 DAG.computeKnownBits(N0.getOperand(0), KnownZero, KnownOne);
4478 // If any of the input bits are KnownOne, then the input couldn't be all
4479 // zeros, thus the result of the srl will always be zero.
4480 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
4482 // If all of the bits input the to ctlz node are known to be zero, then
4483 // the result of the ctlz is "32" and the result of the shift is one.
4484 APInt UnknownBits = ~KnownZero;
4485 if (UnknownBits == 0) return DAG.getConstant(1, VT);
4487 // Otherwise, check to see if there is exactly one bit input to the ctlz.
4488 if ((UnknownBits & (UnknownBits - 1)) == 0) {
4489 // Okay, we know that only that the single bit specified by UnknownBits
4490 // could be set on input to the CTLZ node. If this bit is set, the SRL
4491 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
4492 // to an SRL/XOR pair, which is likely to simplify more.
4493 unsigned ShAmt = UnknownBits.countTrailingZeros();
4494 SDValue Op = N0.getOperand(0);
4497 Op = DAG.getNode(ISD::SRL, SDLoc(N0), VT, Op,
4498 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
4499 AddToWorklist(Op.getNode());
4502 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
4503 Op, DAG.getConstant(1, VT));
4507 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
4508 if (N1.getOpcode() == ISD::TRUNCATE &&
4509 N1.getOperand(0).getOpcode() == ISD::AND) {
4510 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4511 if (NewOp1.getNode())
4512 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, NewOp1);
4515 // fold operands of srl based on knowledge that the low bits are not
4517 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4518 return SDValue(N, 0);
4521 SDValue NewSRL = visitShiftByConstant(N, N1C);
4522 if (NewSRL.getNode())
4526 // Attempt to convert a srl of a load into a narrower zero-extending load.
4527 SDValue NarrowLoad = ReduceLoadWidth(N);
4528 if (NarrowLoad.getNode())
4531 // Here is a common situation. We want to optimize:
4534 // %b = and i32 %a, 2
4535 // %c = srl i32 %b, 1
4536 // brcond i32 %c ...
4542 // %c = setcc eq %b, 0
4545 // However when after the source operand of SRL is optimized into AND, the SRL
4546 // itself may not be optimized further. Look for it and add the BRCOND into
4548 if (N->hasOneUse()) {
4549 SDNode *Use = *N->use_begin();
4550 if (Use->getOpcode() == ISD::BRCOND)
4552 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
4553 // Also look pass the truncate.
4554 Use = *Use->use_begin();
4555 if (Use->getOpcode() == ISD::BRCOND)
4563 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
4564 SDValue N0 = N->getOperand(0);
4565 EVT VT = N->getValueType(0);
4567 // fold (ctlz c1) -> c2
4568 if (isa<ConstantSDNode>(N0))
4569 return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0);
4573 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4574 SDValue N0 = N->getOperand(0);
4575 EVT VT = N->getValueType(0);
4577 // fold (ctlz_zero_undef c1) -> c2
4578 if (isa<ConstantSDNode>(N0))
4579 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4583 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4584 SDValue N0 = N->getOperand(0);
4585 EVT VT = N->getValueType(0);
4587 // fold (cttz c1) -> c2
4588 if (isa<ConstantSDNode>(N0))
4589 return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0);
4593 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4594 SDValue N0 = N->getOperand(0);
4595 EVT VT = N->getValueType(0);
4597 // fold (cttz_zero_undef c1) -> c2
4598 if (isa<ConstantSDNode>(N0))
4599 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4603 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4604 SDValue N0 = N->getOperand(0);
4605 EVT VT = N->getValueType(0);
4607 // fold (ctpop c1) -> c2
4608 if (isa<ConstantSDNode>(N0))
4609 return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0);
4614 /// \brief Generate Min/Max node
4615 static SDValue combineMinNumMaxNum(SDLoc DL, EVT VT, SDValue LHS, SDValue RHS,
4616 SDValue True, SDValue False,
4617 ISD::CondCode CC, const TargetLowering &TLI,
4618 SelectionDAG &DAG) {
4619 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
4629 unsigned Opcode = (LHS == True) ? ISD::FMINNUM : ISD::FMAXNUM;
4630 if (TLI.isOperationLegal(Opcode, VT))
4631 return DAG.getNode(Opcode, DL, VT, LHS, RHS);
4640 unsigned Opcode = (LHS == True) ? ISD::FMAXNUM : ISD::FMINNUM;
4641 if (TLI.isOperationLegal(Opcode, VT))
4642 return DAG.getNode(Opcode, DL, VT, LHS, RHS);
4650 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4651 SDValue N0 = N->getOperand(0);
4652 SDValue N1 = N->getOperand(1);
4653 SDValue N2 = N->getOperand(2);
4654 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4655 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4656 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
4657 EVT VT = N->getValueType(0);
4658 EVT VT0 = N0.getValueType();
4660 // fold (select C, X, X) -> X
4663 // fold (select true, X, Y) -> X
4664 if (N0C && !N0C->isNullValue())
4666 // fold (select false, X, Y) -> Y
4667 if (N0C && N0C->isNullValue())
4669 // fold (select C, 1, X) -> (or C, X)
4670 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4671 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4672 // fold (select C, 0, 1) -> (xor C, 1)
4673 // We can't do this reliably if integer based booleans have different contents
4674 // to floating point based booleans. This is because we can't tell whether we
4675 // have an integer-based boolean or a floating-point-based boolean unless we
4676 // can find the SETCC that produced it and inspect its operands. This is
4677 // fairly easy if C is the SETCC node, but it can potentially be
4678 // undiscoverable (or not reasonably discoverable). For example, it could be
4679 // in another basic block or it could require searching a complicated
4681 if (VT.isInteger() &&
4682 (VT0 == MVT::i1 || (VT0.isInteger() &&
4683 TLI.getBooleanContents(false, false) ==
4684 TLI.getBooleanContents(false, true) &&
4685 TLI.getBooleanContents(false, false) ==
4686 TargetLowering::ZeroOrOneBooleanContent)) &&
4687 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4690 return DAG.getNode(ISD::XOR, SDLoc(N), VT0,
4691 N0, DAG.getConstant(1, VT0));
4692 XORNode = DAG.getNode(ISD::XOR, SDLoc(N0), VT0,
4693 N0, DAG.getConstant(1, VT0));
4694 AddToWorklist(XORNode.getNode());
4696 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, XORNode);
4697 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, XORNode);
4699 // fold (select C, 0, X) -> (and (not C), X)
4700 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4701 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4702 AddToWorklist(NOTNode.getNode());
4703 return DAG.getNode(ISD::AND, SDLoc(N), VT, NOTNode, N2);
4705 // fold (select C, X, 1) -> (or (not C), X)
4706 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4707 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4708 AddToWorklist(NOTNode.getNode());
4709 return DAG.getNode(ISD::OR, SDLoc(N), VT, NOTNode, N1);
4711 // fold (select C, X, 0) -> (and C, X)
4712 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4713 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4714 // fold (select X, X, Y) -> (or X, Y)
4715 // fold (select X, 1, Y) -> (or X, Y)
4716 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4717 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4718 // fold (select X, Y, X) -> (and X, Y)
4719 // fold (select X, Y, 0) -> (and X, Y)
4720 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4721 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4723 // If we can fold this based on the true/false value, do so.
4724 if (SimplifySelectOps(N, N1, N2))
4725 return SDValue(N, 0); // Don't revisit N.
4727 // fold selects based on a setcc into other things, such as min/max/abs
4728 if (N0.getOpcode() == ISD::SETCC) {
4729 // select x, y (fcmp lt x, y) -> fminnum x, y
4730 // select x, y (fcmp gt x, y) -> fmaxnum x, y
4732 // This is OK if we don't care about what happens if either operand is a
4736 // FIXME: Instead of testing for UnsafeFPMath, this should be checking for
4737 // no signed zeros as well as no nans.
4738 const TargetOptions &Options = DAG.getTarget().Options;
4739 if (Options.UnsafeFPMath &&
4740 VT.isFloatingPoint() && N0.hasOneUse() &&
4741 DAG.isKnownNeverNaN(N1) && DAG.isKnownNeverNaN(N2)) {
4742 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
4745 combineMinNumMaxNum(SDLoc(N), VT, N0.getOperand(0), N0.getOperand(1),
4746 N1, N2, CC, TLI, DAG);
4751 if ((!LegalOperations &&
4752 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT)) ||
4753 TLI.isOperationLegal(ISD::SELECT_CC, VT))
4754 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT,
4755 N0.getOperand(0), N0.getOperand(1),
4756 N1, N2, N0.getOperand(2));
4757 return SimplifySelect(SDLoc(N), N0, N1, N2);
4764 std::pair<SDValue, SDValue> SplitVSETCC(const SDNode *N, SelectionDAG &DAG) {
4767 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
4769 // Split the inputs.
4770 SDValue Lo, Hi, LL, LH, RL, RH;
4771 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
4772 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
4774 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));
4775 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));
4777 return std::make_pair(Lo, Hi);
4780 // This function assumes all the vselect's arguments are CONCAT_VECTOR
4781 // nodes and that the condition is a BV of ConstantSDNodes (or undefs).
4782 static SDValue ConvertSelectToConcatVector(SDNode *N, SelectionDAG &DAG) {
4784 SDValue Cond = N->getOperand(0);
4785 SDValue LHS = N->getOperand(1);
4786 SDValue RHS = N->getOperand(2);
4787 EVT VT = N->getValueType(0);
4788 int NumElems = VT.getVectorNumElements();
4789 assert(LHS.getOpcode() == ISD::CONCAT_VECTORS &&
4790 RHS.getOpcode() == ISD::CONCAT_VECTORS &&
4791 Cond.getOpcode() == ISD::BUILD_VECTOR);
4793 // CONCAT_VECTOR can take an arbitrary number of arguments. We only care about
4794 // binary ones here.
4795 if (LHS->getNumOperands() != 2 || RHS->getNumOperands() != 2)
4798 // We're sure we have an even number of elements due to the
4799 // concat_vectors we have as arguments to vselect.
4800 // Skip BV elements until we find one that's not an UNDEF
4801 // After we find an UNDEF element, keep looping until we get to half the
4802 // length of the BV and see if all the non-undef nodes are the same.
4803 ConstantSDNode *BottomHalf = nullptr;
4804 for (int i = 0; i < NumElems / 2; ++i) {
4805 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4808 if (BottomHalf == nullptr)
4809 BottomHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4810 else if (Cond->getOperand(i).getNode() != BottomHalf)
4814 // Do the same for the second half of the BuildVector
4815 ConstantSDNode *TopHalf = nullptr;
4816 for (int i = NumElems / 2; i < NumElems; ++i) {
4817 if (Cond->getOperand(i)->getOpcode() == ISD::UNDEF)
4820 if (TopHalf == nullptr)
4821 TopHalf = cast<ConstantSDNode>(Cond.getOperand(i));
4822 else if (Cond->getOperand(i).getNode() != TopHalf)
4826 assert(TopHalf && BottomHalf &&
4827 "One half of the selector was all UNDEFs and the other was all the "
4828 "same value. This should have been addressed before this function.");
4830 ISD::CONCAT_VECTORS, dl, VT,
4831 BottomHalf->isNullValue() ? RHS->getOperand(0) : LHS->getOperand(0),
4832 TopHalf->isNullValue() ? RHS->getOperand(1) : LHS->getOperand(1));
4835 SDValue DAGCombiner::visitMSTORE(SDNode *N) {
4837 if (Level >= AfterLegalizeTypes)
4840 MaskedStoreSDNode *MST = dyn_cast<MaskedStoreSDNode>(N);
4841 SDValue Mask = MST->getMask();
4842 SDValue Data = MST->getValue();
4845 // If the MSTORE data type requires splitting and the mask is provided by a
4846 // SETCC, then split both nodes and its operands before legalization. This
4847 // prevents the type legalizer from unrolling SETCC into scalar comparisons
4848 // and enables future optimizations (e.g. min/max pattern matching on X86).
4849 if (Mask.getOpcode() == ISD::SETCC) {
4851 // Check if any splitting is required.
4852 if (TLI.getTypeAction(*DAG.getContext(), Data.getValueType()) !=
4853 TargetLowering::TypeSplitVector)
4856 SDValue MaskLo, MaskHi, Lo, Hi;
4857 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
4860 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MST->getValueType(0));
4862 SDValue Chain = MST->getChain();
4863 SDValue Ptr = MST->getBasePtr();
4865 EVT MemoryVT = MST->getMemoryVT();
4866 unsigned Alignment = MST->getOriginalAlignment();
4868 // if Alignment is equal to the vector size,
4869 // take the half of it for the second part
4870 unsigned SecondHalfAlignment =
4871 (Alignment == Data->getValueType(0).getSizeInBits()/8) ?
4872 Alignment/2 : Alignment;
4874 EVT LoMemVT, HiMemVT;
4875 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
4877 SDValue DataLo, DataHi;
4878 std::tie(DataLo, DataHi) = DAG.SplitVector(Data, DL);
4880 MachineMemOperand *MMO = DAG.getMachineFunction().
4881 getMachineMemOperand(MST->getPointerInfo(),
4882 MachineMemOperand::MOStore, LoMemVT.getStoreSize(),
4883 Alignment, MST->getAAInfo(), MST->getRanges());
4885 Lo = DAG.getMaskedStore(Chain, DL, DataLo, Ptr, MaskLo, LoMemVT, MMO,
4886 MST->isTruncatingStore());
4888 unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
4889 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
4890 DAG.getConstant(IncrementSize, Ptr.getValueType()));
4892 MMO = DAG.getMachineFunction().
4893 getMachineMemOperand(MST->getPointerInfo(),
4894 MachineMemOperand::MOStore, HiMemVT.getStoreSize(),
4895 SecondHalfAlignment, MST->getAAInfo(),
4898 Hi = DAG.getMaskedStore(Chain, DL, DataHi, Ptr, MaskHi, HiMemVT, MMO,
4899 MST->isTruncatingStore());
4901 AddToWorklist(Lo.getNode());
4902 AddToWorklist(Hi.getNode());
4904 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);
4909 SDValue DAGCombiner::visitMLOAD(SDNode *N) {
4911 if (Level >= AfterLegalizeTypes)
4914 MaskedLoadSDNode *MLD = dyn_cast<MaskedLoadSDNode>(N);
4915 SDValue Mask = MLD->getMask();
4918 // If the MLOAD result requires splitting and the mask is provided by a
4919 // SETCC, then split both nodes and its operands before legalization. This
4920 // prevents the type legalizer from unrolling SETCC into scalar comparisons
4921 // and enables future optimizations (e.g. min/max pattern matching on X86).
4923 if (Mask.getOpcode() == ISD::SETCC) {
4924 EVT VT = N->getValueType(0);
4926 // Check if any splitting is required.
4927 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
4928 TargetLowering::TypeSplitVector)
4931 SDValue MaskLo, MaskHi, Lo, Hi;
4932 std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
4934 SDValue Src0 = MLD->getSrc0();
4935 SDValue Src0Lo, Src0Hi;
4936 std::tie(Src0Lo, Src0Hi) = DAG.SplitVector(Src0, DL);
4939 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MLD->getValueType(0));
4941 SDValue Chain = MLD->getChain();
4942 SDValue Ptr = MLD->getBasePtr();
4943 EVT MemoryVT = MLD->getMemoryVT();
4944 unsigned Alignment = MLD->getOriginalAlignment();
4946 // if Alignment is equal to the vector size,
4947 // take the half of it for the second part
4948 unsigned SecondHalfAlignment =
4949 (Alignment == MLD->getValueType(0).getSizeInBits()/8) ?
4950 Alignment/2 : Alignment;
4952 EVT LoMemVT, HiMemVT;
4953 std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
4955 MachineMemOperand *MMO = DAG.getMachineFunction().
4956 getMachineMemOperand(MLD->getPointerInfo(),
4957 MachineMemOperand::MOLoad, LoMemVT.getStoreSize(),
4958 Alignment, MLD->getAAInfo(), MLD->getRanges());
4960 Lo = DAG.getMaskedLoad(LoVT, DL, Chain, Ptr, MaskLo, Src0Lo, LoMemVT, MMO,
4963 unsigned IncrementSize = LoMemVT.getSizeInBits()/8;
4964 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
4965 DAG.getConstant(IncrementSize, Ptr.getValueType()));
4967 MMO = DAG.getMachineFunction().
4968 getMachineMemOperand(MLD->getPointerInfo(),
4969 MachineMemOperand::MOLoad, HiMemVT.getStoreSize(),
4970 SecondHalfAlignment, MLD->getAAInfo(), MLD->getRanges());
4972 Hi = DAG.getMaskedLoad(HiVT, DL, Chain, Ptr, MaskHi, Src0Hi, HiMemVT, MMO,
4975 AddToWorklist(Lo.getNode());
4976 AddToWorklist(Hi.getNode());
4978 // Build a factor node to remember that this load is independent of the
4980 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo.getValue(1),
4983 // Legalized the chain result - switch anything that used the old chain to
4985 DAG.ReplaceAllUsesOfValueWith(SDValue(MLD, 1), Chain);
4987 SDValue LoadRes = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
4989 SDValue RetOps[] = { LoadRes, Chain };
4990 return DAG.getMergeValues(RetOps, DL);
4995 SDValue DAGCombiner::visitVSELECT(SDNode *N) {
4996 SDValue N0 = N->getOperand(0);
4997 SDValue N1 = N->getOperand(1);
4998 SDValue N2 = N->getOperand(2);
5001 // Canonicalize integer abs.
5002 // vselect (setg[te] X, 0), X, -X ->
5003 // vselect (setgt X, -1), X, -X ->
5004 // vselect (setl[te] X, 0), -X, X ->
5005 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
5006 if (N0.getOpcode() == ISD::SETCC) {
5007 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
5008 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5010 bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
5012 if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
5013 (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) &&
5014 N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1))
5015 isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode());
5016 else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) &&
5017 N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1))
5018 isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode());
5021 EVT VT = LHS.getValueType();
5022 SDValue Shift = DAG.getNode(
5023 ISD::SRA, DL, VT, LHS,
5024 DAG.getConstant(VT.getScalarType().getSizeInBits() - 1, VT));
5025 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift);
5026 AddToWorklist(Shift.getNode());
5027 AddToWorklist(Add.getNode());
5028 return DAG.getNode(ISD::XOR, DL, VT, Add, Shift);
5032 // If the VSELECT result requires splitting and the mask is provided by a
5033 // SETCC, then split both nodes and its operands before legalization. This
5034 // prevents the type legalizer from unrolling SETCC into scalar comparisons
5035 // and enables future optimizations (e.g. min/max pattern matching on X86).
5036 if (N0.getOpcode() == ISD::SETCC) {
5037 EVT VT = N->getValueType(0);
5039 // Check if any splitting is required.
5040 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
5041 TargetLowering::TypeSplitVector)
5044 SDValue Lo, Hi, CCLo, CCHi, LL, LH, RL, RH;
5045 std::tie(CCLo, CCHi) = SplitVSETCC(N0.getNode(), DAG);
5046 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 1);
5047 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 2);
5049 Lo = DAG.getNode(N->getOpcode(), DL, LL.getValueType(), CCLo, LL, RL);
5050 Hi = DAG.getNode(N->getOpcode(), DL, LH.getValueType(), CCHi, LH, RH);
5052 // Add the new VSELECT nodes to the work list in case they need to be split
5054 AddToWorklist(Lo.getNode());
5055 AddToWorklist(Hi.getNode());
5057 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
5060 // Fold (vselect (build_vector all_ones), N1, N2) -> N1
5061 if (ISD::isBuildVectorAllOnes(N0.getNode()))
5063 // Fold (vselect (build_vector all_zeros), N1, N2) -> N2
5064 if (ISD::isBuildVectorAllZeros(N0.getNode()))
5067 // The ConvertSelectToConcatVector function is assuming both the above
5068 // checks for (vselect (build_vector all{ones,zeros) ...) have been made
5070 if (N1.getOpcode() == ISD::CONCAT_VECTORS &&
5071 N2.getOpcode() == ISD::CONCAT_VECTORS &&
5072 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
5073 SDValue CV = ConvertSelectToConcatVector(N, DAG);
5081 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
5082 SDValue N0 = N->getOperand(0);
5083 SDValue N1 = N->getOperand(1);
5084 SDValue N2 = N->getOperand(2);
5085 SDValue N3 = N->getOperand(3);
5086 SDValue N4 = N->getOperand(4);
5087 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
5089 // fold select_cc lhs, rhs, x, x, cc -> x
5093 // Determine if the condition we're dealing with is constant
5094 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
5095 N0, N1, CC, SDLoc(N), false);
5096 if (SCC.getNode()) {
5097 AddToWorklist(SCC.getNode());
5099 if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) {
5100 if (!SCCC->isNullValue())
5101 return N2; // cond always true -> true val
5103 return N3; // cond always false -> false val
5104 } else if (SCC->getOpcode() == ISD::UNDEF) {
5105 // When the condition is UNDEF, just return the first operand. This is
5106 // coherent the DAG creation, no setcc node is created in this case
5108 } else if (SCC.getOpcode() == ISD::SETCC) {
5109 // Fold to a simpler select_cc
5110 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(),
5111 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
5116 // If we can fold this based on the true/false value, do so.
5117 if (SimplifySelectOps(N, N2, N3))
5118 return SDValue(N, 0); // Don't revisit N.
5120 // fold select_cc into other things, such as min/max/abs
5121 return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC);
5124 SDValue DAGCombiner::visitSETCC(SDNode *N) {
5125 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
5126 cast<CondCodeSDNode>(N->getOperand(2))->get(),
5130 // tryToFoldExtendOfConstant - Try to fold a sext/zext/aext
5131 // dag node into a ConstantSDNode or a build_vector of constants.
5132 // This function is called by the DAGCombiner when visiting sext/zext/aext
5133 // dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND).
5134 // Vector extends are not folded if operations are legal; this is to
5135 // avoid introducing illegal build_vector dag nodes.
5136 static SDNode *tryToFoldExtendOfConstant(SDNode *N, const TargetLowering &TLI,
5137 SelectionDAG &DAG, bool LegalTypes,
5138 bool LegalOperations) {
5139 unsigned Opcode = N->getOpcode();
5140 SDValue N0 = N->getOperand(0);
5141 EVT VT = N->getValueType(0);
5143 assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND ||
5144 Opcode == ISD::ANY_EXTEND) && "Expected EXTEND dag node in input!");
5146 // fold (sext c1) -> c1
5147 // fold (zext c1) -> c1
5148 // fold (aext c1) -> c1
5149 if (isa<ConstantSDNode>(N0))
5150 return DAG.getNode(Opcode, SDLoc(N), VT, N0).getNode();
5152 // fold (sext (build_vector AllConstants) -> (build_vector AllConstants)
5153 // fold (zext (build_vector AllConstants) -> (build_vector AllConstants)
5154 // fold (aext (build_vector AllConstants) -> (build_vector AllConstants)
5155 EVT SVT = VT.getScalarType();
5156 if (!(VT.isVector() &&
5157 (!LegalTypes || (!LegalOperations && TLI.isTypeLegal(SVT))) &&
5158 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())))
5161 // We can fold this node into a build_vector.
5162 unsigned VTBits = SVT.getSizeInBits();
5163 unsigned EVTBits = N0->getValueType(0).getScalarType().getSizeInBits();
5164 unsigned ShAmt = VTBits - EVTBits;
5165 SmallVector<SDValue, 8> Elts;
5166 unsigned NumElts = N0->getNumOperands();
5169 for (unsigned i=0; i != NumElts; ++i) {
5170 SDValue Op = N0->getOperand(i);
5171 if (Op->getOpcode() == ISD::UNDEF) {
5172 Elts.push_back(DAG.getUNDEF(SVT));
5176 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
5177 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
5178 if (Opcode == ISD::SIGN_EXTEND)
5179 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
5182 Elts.push_back(DAG.getConstant(C.shl(ShAmt).lshr(ShAmt).getZExtValue(),
5186 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Elts).getNode();
5189 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
5190 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
5191 // transformation. Returns true if extension are possible and the above
5192 // mentioned transformation is profitable.
5193 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
5195 SmallVectorImpl<SDNode *> &ExtendNodes,
5196 const TargetLowering &TLI) {
5197 bool HasCopyToRegUses = false;
5198 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
5199 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
5200 UE = N0.getNode()->use_end();
5205 if (UI.getUse().getResNo() != N0.getResNo())
5207 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
5208 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
5209 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
5210 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
5211 // Sign bits will be lost after a zext.
5214 for (unsigned i = 0; i != 2; ++i) {
5215 SDValue UseOp = User->getOperand(i);
5218 if (!isa<ConstantSDNode>(UseOp))
5223 ExtendNodes.push_back(User);
5226 // If truncates aren't free and there are users we can't
5227 // extend, it isn't worthwhile.
5230 // Remember if this value is live-out.
5231 if (User->getOpcode() == ISD::CopyToReg)
5232 HasCopyToRegUses = true;
5235 if (HasCopyToRegUses) {
5236 bool BothLiveOut = false;
5237 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
5239 SDUse &Use = UI.getUse();
5240 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
5246 // Both unextended and extended values are live out. There had better be
5247 // a good reason for the transformation.
5248 return ExtendNodes.size();
5253 void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
5254 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
5255 ISD::NodeType ExtType) {
5256 // Extend SetCC uses if necessary.
5257 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
5258 SDNode *SetCC = SetCCs[i];
5259 SmallVector<SDValue, 4> Ops;
5261 for (unsigned j = 0; j != 2; ++j) {
5262 SDValue SOp = SetCC->getOperand(j);
5264 Ops.push_back(ExtLoad);
5266 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
5269 Ops.push_back(SetCC->getOperand(2));
5270 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0), Ops));
5274 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
5275 SDValue N0 = N->getOperand(0);
5276 EVT VT = N->getValueType(0);
5278 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5280 return SDValue(Res, 0);
5282 // fold (sext (sext x)) -> (sext x)
5283 // fold (sext (aext x)) -> (sext x)
5284 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5285 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT,
5288 if (N0.getOpcode() == ISD::TRUNCATE) {
5289 // fold (sext (truncate (load x))) -> (sext (smaller load x))
5290 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
5291 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5292 if (NarrowLoad.getNode()) {
5293 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5294 if (NarrowLoad.getNode() != N0.getNode()) {
5295 CombineTo(N0.getNode(), NarrowLoad);
5296 // CombineTo deleted the truncate, if needed, but not what's under it.
5299 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5302 // See if the value being truncated is already sign extended. If so, just
5303 // eliminate the trunc/sext pair.
5304 SDValue Op = N0.getOperand(0);
5305 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
5306 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
5307 unsigned DestBits = VT.getScalarType().getSizeInBits();
5308 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
5310 if (OpBits == DestBits) {
5311 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
5312 // bits, it is already ready.
5313 if (NumSignBits > DestBits-MidBits)
5315 } else if (OpBits < DestBits) {
5316 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
5317 // bits, just sext from i32.
5318 if (NumSignBits > OpBits-MidBits)
5319 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, Op);
5321 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
5322 // bits, just truncate to i32.
5323 if (NumSignBits > OpBits-MidBits)
5324 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5327 // fold (sext (truncate x)) -> (sextinreg x).
5328 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
5329 N0.getValueType())) {
5330 if (OpBits < DestBits)
5331 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op);
5332 else if (OpBits > DestBits)
5333 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op);
5334 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, Op,
5335 DAG.getValueType(N0.getValueType()));
5339 // fold (sext (load x)) -> (sext (truncate (sextload x)))
5340 // None of the supported targets knows how to perform load and sign extend
5341 // on vectors in one instruction. We only perform this transformation on
5343 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5344 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5345 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5346 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, N0.getValueType()))) {
5347 bool DoXform = true;
5348 SmallVector<SDNode*, 4> SetCCs;
5349 if (!N0.hasOneUse())
5350 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
5352 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5353 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5355 LN0->getBasePtr(), N0.getValueType(),
5356 LN0->getMemOperand());
5357 CombineTo(N, ExtLoad);
5358 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5359 N0.getValueType(), ExtLoad);
5360 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5361 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5363 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5367 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
5368 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
5369 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5370 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5371 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5372 EVT MemVT = LN0->getMemoryVT();
5373 if ((!LegalOperations && !LN0->isVolatile()) ||
5374 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, MemVT)) {
5375 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5377 LN0->getBasePtr(), MemVT,
5378 LN0->getMemOperand());
5379 CombineTo(N, ExtLoad);
5380 CombineTo(N0.getNode(),
5381 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5382 N0.getValueType(), ExtLoad),
5383 ExtLoad.getValue(1));
5384 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5388 // fold (sext (and/or/xor (load x), cst)) ->
5389 // (and/or/xor (sextload x), (sext cst))
5390 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5391 N0.getOpcode() == ISD::XOR) &&
5392 isa<LoadSDNode>(N0.getOperand(0)) &&
5393 N0.getOperand(1).getOpcode() == ISD::Constant &&
5394 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, N0.getValueType()) &&
5395 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5396 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5397 if (LN0->getExtensionType() != ISD::ZEXTLOAD && LN0->isUnindexed()) {
5398 bool DoXform = true;
5399 SmallVector<SDNode*, 4> SetCCs;
5400 if (!N0.hasOneUse())
5401 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
5404 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN0), VT,
5405 LN0->getChain(), LN0->getBasePtr(),
5407 LN0->getMemOperand());
5408 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5409 Mask = Mask.sext(VT.getSizeInBits());
5410 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5411 ExtLoad, DAG.getConstant(Mask, VT));
5412 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5413 SDLoc(N0.getOperand(0)),
5414 N0.getOperand(0).getValueType(), ExtLoad);
5416 CombineTo(N0.getOperand(0).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!
5424 if (N0.getOpcode() == ISD::SETCC) {
5425 EVT N0VT = N0.getOperand(0).getValueType();
5426 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
5427 // Only do this before legalize for now.
5428 if (VT.isVector() && !LegalOperations &&
5429 TLI.getBooleanContents(N0VT) ==
5430 TargetLowering::ZeroOrNegativeOneBooleanContent) {
5431 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
5432 // of the same size as the compared operands. Only optimize sext(setcc())
5433 // if this is the case.
5434 EVT SVT = getSetCCResultType(N0VT);
5436 // We know that the # elements of the results is the same as the
5437 // # elements of the compare (and the # elements of the compare result
5438 // for that matter). Check to see that they are the same size. If so,
5439 // we know that the element size of the sext'd result matches the
5440 // element size of the compare operands.
5441 if (VT.getSizeInBits() == SVT.getSizeInBits())
5442 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5444 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5446 // If the desired elements are smaller or larger than the source
5447 // elements we can use a matching integer vector type and then
5448 // truncate/sign extend
5449 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5450 if (SVT == MatchingVectorType) {
5451 SDValue VsetCC = DAG.getSetCC(SDLoc(N), MatchingVectorType,
5452 N0.getOperand(0), N0.getOperand(1),
5453 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5454 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
5458 // sext(setcc x, y, cc) -> (select (setcc x, y, cc), -1, 0)
5459 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
5461 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
5463 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5464 NegOne, DAG.getConstant(0, VT),
5465 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5466 if (SCC.getNode()) return SCC;
5468 if (!VT.isVector()) {
5469 EVT SetCCVT = getSetCCResultType(N0.getOperand(0).getValueType());
5470 if (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, SetCCVT)) {
5472 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5473 SDValue SetCC = DAG.getSetCC(DL, SetCCVT,
5474 N0.getOperand(0), N0.getOperand(1), CC);
5475 return DAG.getSelect(DL, VT, SetCC,
5476 NegOne, DAG.getConstant(0, VT));
5481 // fold (sext x) -> (zext x) if the sign bit is known zero.
5482 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
5483 DAG.SignBitIsZero(N0))
5484 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0);
5489 // isTruncateOf - If N is a truncate of some other value, return true, record
5490 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
5491 // This function computes KnownZero to avoid a duplicated call to
5492 // computeKnownBits in the caller.
5493 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
5496 if (N->getOpcode() == ISD::TRUNCATE) {
5497 Op = N->getOperand(0);
5498 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5502 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
5503 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
5506 SDValue Op0 = N->getOperand(0);
5507 SDValue Op1 = N->getOperand(1);
5508 assert(Op0.getValueType() == Op1.getValueType());
5510 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
5511 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
5512 if (COp0 && COp0->isNullValue())
5514 else if (COp1 && COp1->isNullValue())
5519 DAG.computeKnownBits(Op, KnownZero, KnownOne);
5521 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
5527 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
5528 SDValue N0 = N->getOperand(0);
5529 EVT VT = N->getValueType(0);
5531 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5533 return SDValue(Res, 0);
5535 // fold (zext (zext x)) -> (zext x)
5536 // fold (zext (aext x)) -> (zext x)
5537 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5538 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT,
5541 // fold (zext (truncate x)) -> (zext x) or
5542 // (zext (truncate x)) -> (truncate x)
5543 // This is valid when the truncated bits of x are already zero.
5544 // FIXME: We should extend this to work for vectors too.
5547 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
5548 APInt TruncatedBits =
5549 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
5550 APInt(Op.getValueSizeInBits(), 0) :
5551 APInt::getBitsSet(Op.getValueSizeInBits(),
5552 N0.getValueSizeInBits(),
5553 std::min(Op.getValueSizeInBits(),
5554 VT.getSizeInBits()));
5555 if (TruncatedBits == (KnownZero & TruncatedBits)) {
5556 if (VT.bitsGT(Op.getValueType()))
5557 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, Op);
5558 if (VT.bitsLT(Op.getValueType()))
5559 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5565 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5566 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
5567 if (N0.getOpcode() == ISD::TRUNCATE) {
5568 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5569 if (NarrowLoad.getNode()) {
5570 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5571 if (NarrowLoad.getNode() != N0.getNode()) {
5572 CombineTo(N0.getNode(), NarrowLoad);
5573 // CombineTo deleted the truncate, if needed, but not what's under it.
5576 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5580 // fold (zext (truncate x)) -> (and x, mask)
5581 if (N0.getOpcode() == ISD::TRUNCATE &&
5582 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
5584 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5585 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
5586 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5587 if (NarrowLoad.getNode()) {
5588 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5589 if (NarrowLoad.getNode() != N0.getNode()) {
5590 CombineTo(N0.getNode(), NarrowLoad);
5591 // CombineTo deleted the truncate, if needed, but not what's under it.
5594 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5597 SDValue Op = N0.getOperand(0);
5598 if (Op.getValueType().bitsLT(VT)) {
5599 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, Op);
5600 AddToWorklist(Op.getNode());
5601 } else if (Op.getValueType().bitsGT(VT)) {
5602 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5603 AddToWorklist(Op.getNode());
5605 return DAG.getZeroExtendInReg(Op, SDLoc(N),
5606 N0.getValueType().getScalarType());
5609 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
5610 // if either of the casts is not free.
5611 if (N0.getOpcode() == ISD::AND &&
5612 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5613 N0.getOperand(1).getOpcode() == ISD::Constant &&
5614 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5615 N0.getValueType()) ||
5616 !TLI.isZExtFree(N0.getValueType(), VT))) {
5617 SDValue X = N0.getOperand(0).getOperand(0);
5618 if (X.getValueType().bitsLT(VT)) {
5619 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(X), VT, X);
5620 } else if (X.getValueType().bitsGT(VT)) {
5621 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
5623 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5624 Mask = Mask.zext(VT.getSizeInBits());
5625 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5626 X, DAG.getConstant(Mask, VT));
5629 // fold (zext (load x)) -> (zext (truncate (zextload x)))
5630 // None of the supported targets knows how to perform load and vector_zext
5631 // on vectors in one instruction. We only perform this transformation on
5633 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5634 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5635 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5636 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, N0.getValueType()))) {
5637 bool DoXform = true;
5638 SmallVector<SDNode*, 4> SetCCs;
5639 if (!N0.hasOneUse())
5640 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
5642 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5643 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5645 LN0->getBasePtr(), N0.getValueType(),
5646 LN0->getMemOperand());
5647 CombineTo(N, ExtLoad);
5648 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5649 N0.getValueType(), ExtLoad);
5650 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5652 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5654 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5658 // fold (zext (and/or/xor (load x), cst)) ->
5659 // (and/or/xor (zextload x), (zext cst))
5660 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5661 N0.getOpcode() == ISD::XOR) &&
5662 isa<LoadSDNode>(N0.getOperand(0)) &&
5663 N0.getOperand(1).getOpcode() == ISD::Constant &&
5664 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, N0.getValueType()) &&
5665 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5666 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5667 if (LN0->getExtensionType() != ISD::SEXTLOAD && LN0->isUnindexed()) {
5668 bool DoXform = true;
5669 SmallVector<SDNode*, 4> SetCCs;
5670 if (!N0.hasOneUse())
5671 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
5674 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), VT,
5675 LN0->getChain(), LN0->getBasePtr(),
5677 LN0->getMemOperand());
5678 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5679 Mask = Mask.zext(VT.getSizeInBits());
5680 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5681 ExtLoad, DAG.getConstant(Mask, VT));
5682 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5683 SDLoc(N0.getOperand(0)),
5684 N0.getOperand(0).getValueType(), ExtLoad);
5686 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5687 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5689 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5694 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
5695 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
5696 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5697 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5698 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5699 EVT MemVT = LN0->getMemoryVT();
5700 if ((!LegalOperations && !LN0->isVolatile()) ||
5701 TLI.isLoadExtLegal(ISD::ZEXTLOAD, VT, MemVT)) {
5702 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5704 LN0->getBasePtr(), MemVT,
5705 LN0->getMemOperand());
5706 CombineTo(N, ExtLoad);
5707 CombineTo(N0.getNode(),
5708 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(),
5710 ExtLoad.getValue(1));
5711 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5715 if (N0.getOpcode() == ISD::SETCC) {
5716 if (!LegalOperations && VT.isVector() &&
5717 N0.getValueType().getVectorElementType() == MVT::i1) {
5718 EVT N0VT = N0.getOperand(0).getValueType();
5719 if (getSetCCResultType(N0VT) == N0.getValueType())
5722 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
5723 // Only do this before legalize for now.
5724 EVT EltVT = VT.getVectorElementType();
5725 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
5726 DAG.getConstant(1, EltVT));
5727 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5728 // We know that the # elements of the results is the same as the
5729 // # elements of the compare (and the # elements of the compare result
5730 // for that matter). Check to see that they are the same size. If so,
5731 // we know that the element size of the sext'd result matches the
5732 // element size of the compare operands.
5733 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5734 DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5736 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
5737 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT,
5740 // If the desired elements are smaller or larger than the source
5741 // elements we can use a matching integer vector type and then
5742 // truncate/sign extend
5743 EVT MatchingElementType =
5744 EVT::getIntegerVT(*DAG.getContext(),
5745 N0VT.getScalarType().getSizeInBits());
5746 EVT MatchingVectorType =
5747 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
5748 N0VT.getVectorNumElements());
5750 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5752 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5753 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5754 DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT),
5755 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, OneOps));
5758 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5760 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5761 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5762 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5763 if (SCC.getNode()) return SCC;
5766 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
5767 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
5768 isa<ConstantSDNode>(N0.getOperand(1)) &&
5769 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
5771 SDValue ShAmt = N0.getOperand(1);
5772 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
5773 if (N0.getOpcode() == ISD::SHL) {
5774 SDValue InnerZExt = N0.getOperand(0);
5775 // If the original shl may be shifting out bits, do not perform this
5777 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
5778 InnerZExt.getOperand(0).getValueType().getSizeInBits();
5779 if (ShAmtVal > KnownZeroBits)
5785 // Ensure that the shift amount is wide enough for the shifted value.
5786 if (VT.getSizeInBits() >= 256)
5787 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
5789 return DAG.getNode(N0.getOpcode(), DL, VT,
5790 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
5797 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
5798 SDValue N0 = N->getOperand(0);
5799 EVT VT = N->getValueType(0);
5801 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5803 return SDValue(Res, 0);
5805 // fold (aext (aext x)) -> (aext x)
5806 // fold (aext (zext x)) -> (zext x)
5807 // fold (aext (sext x)) -> (sext x)
5808 if (N0.getOpcode() == ISD::ANY_EXTEND ||
5809 N0.getOpcode() == ISD::ZERO_EXTEND ||
5810 N0.getOpcode() == ISD::SIGN_EXTEND)
5811 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0));
5813 // fold (aext (truncate (load x))) -> (aext (smaller load x))
5814 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
5815 if (N0.getOpcode() == ISD::TRUNCATE) {
5816 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5817 if (NarrowLoad.getNode()) {
5818 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5819 if (NarrowLoad.getNode() != N0.getNode()) {
5820 CombineTo(N0.getNode(), NarrowLoad);
5821 // CombineTo deleted the truncate, if needed, but not what's under it.
5824 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5828 // fold (aext (truncate x))
5829 if (N0.getOpcode() == ISD::TRUNCATE) {
5830 SDValue TruncOp = N0.getOperand(0);
5831 if (TruncOp.getValueType() == VT)
5832 return TruncOp; // x iff x size == zext size.
5833 if (TruncOp.getValueType().bitsGT(VT))
5834 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, TruncOp);
5835 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, TruncOp);
5838 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
5839 // if the trunc is not free.
5840 if (N0.getOpcode() == ISD::AND &&
5841 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5842 N0.getOperand(1).getOpcode() == ISD::Constant &&
5843 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5844 N0.getValueType())) {
5845 SDValue X = N0.getOperand(0).getOperand(0);
5846 if (X.getValueType().bitsLT(VT)) {
5847 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, X);
5848 } else if (X.getValueType().bitsGT(VT)) {
5849 X = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, X);
5851 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5852 Mask = Mask.zext(VT.getSizeInBits());
5853 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5854 X, DAG.getConstant(Mask, VT));
5857 // fold (aext (load x)) -> (aext (truncate (extload x)))
5858 // None of the supported targets knows how to perform load and any_ext
5859 // on vectors in one instruction. We only perform this transformation on
5861 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5862 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5863 TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) {
5864 bool DoXform = true;
5865 SmallVector<SDNode*, 4> SetCCs;
5866 if (!N0.hasOneUse())
5867 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
5869 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5870 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
5872 LN0->getBasePtr(), N0.getValueType(),
5873 LN0->getMemOperand());
5874 CombineTo(N, ExtLoad);
5875 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5876 N0.getValueType(), ExtLoad);
5877 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5878 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5880 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5884 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
5885 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
5886 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
5887 if (N0.getOpcode() == ISD::LOAD &&
5888 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5890 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5891 ISD::LoadExtType ExtType = LN0->getExtensionType();
5892 EVT MemVT = LN0->getMemoryVT();
5893 if (!LegalOperations || TLI.isLoadExtLegal(ExtType, VT, MemVT)) {
5894 SDValue ExtLoad = DAG.getExtLoad(ExtType, SDLoc(N),
5895 VT, LN0->getChain(), LN0->getBasePtr(),
5896 MemVT, LN0->getMemOperand());
5897 CombineTo(N, ExtLoad);
5898 CombineTo(N0.getNode(),
5899 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5900 N0.getValueType(), ExtLoad),
5901 ExtLoad.getValue(1));
5902 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5906 if (N0.getOpcode() == ISD::SETCC) {
5908 // aext(setcc) -> vsetcc
5909 // aext(setcc) -> truncate(vsetcc)
5910 // aext(setcc) -> aext(vsetcc)
5911 // Only do this before legalize for now.
5912 if (VT.isVector() && !LegalOperations) {
5913 EVT N0VT = N0.getOperand(0).getValueType();
5914 // We know that the # elements of the results is the same as the
5915 // # elements of the compare (and the # elements of the compare result
5916 // for that matter). Check to see that they are the same size. If so,
5917 // we know that the element size of the sext'd result matches the
5918 // element size of the compare operands.
5919 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5920 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5922 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5923 // If the desired elements are smaller or larger than the source
5924 // elements we can use a matching integer vector type and then
5925 // truncate/any extend
5927 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5929 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5931 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5932 return DAG.getAnyExtOrTrunc(VsetCC, SDLoc(N), VT);
5936 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5938 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5939 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5940 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5948 /// See if the specified operand can be simplified with the knowledge that only
5949 /// the bits specified by Mask are used. If so, return the simpler operand,
5950 /// otherwise return a null SDValue.
5951 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
5952 switch (V.getOpcode()) {
5954 case ISD::Constant: {
5955 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
5956 assert(CV && "Const value should be ConstSDNode.");
5957 const APInt &CVal = CV->getAPIntValue();
5958 APInt NewVal = CVal & Mask;
5960 return DAG.getConstant(NewVal, V.getValueType());
5965 // If the LHS or RHS don't contribute bits to the or, drop them.
5966 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
5967 return V.getOperand(1);
5968 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
5969 return V.getOperand(0);
5972 // Only look at single-use SRLs.
5973 if (!V.getNode()->hasOneUse())
5975 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
5976 // See if we can recursively simplify the LHS.
5977 unsigned Amt = RHSC->getZExtValue();
5979 // Watch out for shift count overflow though.
5980 if (Amt >= Mask.getBitWidth()) break;
5981 APInt NewMask = Mask << Amt;
5982 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
5983 if (SimplifyLHS.getNode())
5984 return DAG.getNode(ISD::SRL, SDLoc(V), V.getValueType(),
5985 SimplifyLHS, V.getOperand(1));
5991 /// If the result of a wider load is shifted to right of N bits and then
5992 /// truncated to a narrower type and where N is a multiple of number of bits of
5993 /// the narrower type, transform it to a narrower load from address + N / num of
5994 /// bits of new type. If the result is to be extended, also fold the extension
5995 /// to form a extending load.
5996 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
5997 unsigned Opc = N->getOpcode();
5999 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
6000 SDValue N0 = N->getOperand(0);
6001 EVT VT = N->getValueType(0);
6004 // This transformation isn't valid for vector loads.
6008 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
6010 if (Opc == ISD::SIGN_EXTEND_INREG) {
6011 ExtType = ISD::SEXTLOAD;
6012 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
6013 } else if (Opc == ISD::SRL) {
6014 // Another special-case: SRL is basically zero-extending a narrower value.
6015 ExtType = ISD::ZEXTLOAD;
6017 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
6018 if (!N01) return SDValue();
6019 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
6020 VT.getSizeInBits() - N01->getZExtValue());
6022 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, VT, ExtVT))
6025 unsigned EVTBits = ExtVT.getSizeInBits();
6027 // Do not generate loads of non-round integer types since these can
6028 // be expensive (and would be wrong if the type is not byte sized).
6029 if (!ExtVT.isRound())
6033 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
6034 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
6035 ShAmt = N01->getZExtValue();
6036 // Is the shift amount a multiple of size of VT?
6037 if ((ShAmt & (EVTBits-1)) == 0) {
6038 N0 = N0.getOperand(0);
6039 // Is the load width a multiple of size of VT?
6040 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
6044 // At this point, we must have a load or else we can't do the transform.
6045 if (!isa<LoadSDNode>(N0)) return SDValue();
6047 // Because a SRL must be assumed to *need* to zero-extend the high bits
6048 // (as opposed to anyext the high bits), we can't combine the zextload
6049 // lowering of SRL and an sextload.
6050 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
6053 // If the shift amount is larger than the input type then we're not
6054 // accessing any of the loaded bytes. If the load was a zextload/extload
6055 // then the result of the shift+trunc is zero/undef (handled elsewhere).
6056 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
6061 // If the load is shifted left (and the result isn't shifted back right),
6062 // we can fold the truncate through the shift.
6063 unsigned ShLeftAmt = 0;
6064 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
6065 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
6066 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
6067 ShLeftAmt = N01->getZExtValue();
6068 N0 = N0.getOperand(0);
6072 // If we haven't found a load, we can't narrow it. Don't transform one with
6073 // multiple uses, this would require adding a new load.
6074 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse())
6077 // Don't change the width of a volatile load.
6078 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6079 if (LN0->isVolatile())
6082 // Verify that we are actually reducing a load width here.
6083 if (LN0->getMemoryVT().getSizeInBits() < EVTBits)
6086 // For the transform to be legal, the load must produce only two values
6087 // (the value loaded and the chain). Don't transform a pre-increment
6088 // load, for example, which produces an extra value. Otherwise the
6089 // transformation is not equivalent, and the downstream logic to replace
6090 // uses gets things wrong.
6091 if (LN0->getNumValues() > 2)
6094 // If the load that we're shrinking is an extload and we're not just
6095 // discarding the extension we can't simply shrink the load. Bail.
6096 // TODO: It would be possible to merge the extensions in some cases.
6097 if (LN0->getExtensionType() != ISD::NON_EXTLOAD &&
6098 LN0->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits() + ShAmt)
6101 if (!TLI.shouldReduceLoadWidth(LN0, ExtType, ExtVT))
6104 EVT PtrType = N0.getOperand(1).getValueType();
6106 if (PtrType == MVT::Untyped || PtrType.isExtended())
6107 // It's not possible to generate a constant of extended or untyped type.
6110 // For big endian targets, we need to adjust the offset to the pointer to
6111 // load the correct bytes.
6112 if (TLI.isBigEndian()) {
6113 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
6114 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
6115 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
6118 uint64_t PtrOff = ShAmt / 8;
6119 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
6120 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0),
6121 PtrType, LN0->getBasePtr(),
6122 DAG.getConstant(PtrOff, PtrType));
6123 AddToWorklist(NewPtr.getNode());
6126 if (ExtType == ISD::NON_EXTLOAD)
6127 Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr,
6128 LN0->getPointerInfo().getWithOffset(PtrOff),
6129 LN0->isVolatile(), LN0->isNonTemporal(),
6130 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
6132 Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(),NewPtr,
6133 LN0->getPointerInfo().getWithOffset(PtrOff),
6134 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
6135 LN0->isInvariant(), NewAlign, LN0->getAAInfo());
6137 // Replace the old load's chain with the new load's chain.
6138 WorklistRemover DeadNodes(*this);
6139 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
6141 // Shift the result left, if we've swallowed a left shift.
6142 SDValue Result = Load;
6143 if (ShLeftAmt != 0) {
6144 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
6145 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
6147 // If the shift amount is as large as the result size (but, presumably,
6148 // no larger than the source) then the useful bits of the result are
6149 // zero; we can't simply return the shortened shift, because the result
6150 // of that operation is undefined.
6151 if (ShLeftAmt >= VT.getSizeInBits())
6152 Result = DAG.getConstant(0, VT);
6154 Result = DAG.getNode(ISD::SHL, SDLoc(N0), VT,
6155 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
6158 // Return the new loaded value.
6162 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
6163 SDValue N0 = N->getOperand(0);
6164 SDValue N1 = N->getOperand(1);
6165 EVT VT = N->getValueType(0);
6166 EVT EVT = cast<VTSDNode>(N1)->getVT();
6167 unsigned VTBits = VT.getScalarType().getSizeInBits();
6168 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
6170 // fold (sext_in_reg c1) -> c1
6171 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
6172 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1);
6174 // If the input is already sign extended, just drop the extension.
6175 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
6178 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
6179 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
6180 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT()))
6181 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
6182 N0.getOperand(0), N1);
6184 // fold (sext_in_reg (sext x)) -> (sext x)
6185 // fold (sext_in_reg (aext x)) -> (sext x)
6186 // if x is small enough.
6187 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
6188 SDValue N00 = N0.getOperand(0);
6189 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
6190 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
6191 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
6194 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
6195 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
6196 return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT);
6198 // fold operands of sext_in_reg based on knowledge that the top bits are not
6200 if (SimplifyDemandedBits(SDValue(N, 0)))
6201 return SDValue(N, 0);
6203 // fold (sext_in_reg (load x)) -> (smaller sextload x)
6204 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
6205 SDValue NarrowLoad = ReduceLoadWidth(N);
6206 if (NarrowLoad.getNode())
6209 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
6210 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
6211 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
6212 if (N0.getOpcode() == ISD::SRL) {
6213 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
6214 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
6215 // We can turn this into an SRA iff the input to the SRL is already sign
6217 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
6218 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
6219 return DAG.getNode(ISD::SRA, SDLoc(N), VT,
6220 N0.getOperand(0), N0.getOperand(1));
6224 // fold (sext_inreg (extload x)) -> (sextload x)
6225 if (ISD::isEXTLoad(N0.getNode()) &&
6226 ISD::isUNINDEXEDLoad(N0.getNode()) &&
6227 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
6228 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6229 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) {
6230 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6231 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
6233 LN0->getBasePtr(), EVT,
6234 LN0->getMemOperand());
6235 CombineTo(N, ExtLoad);
6236 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
6237 AddToWorklist(ExtLoad.getNode());
6238 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6240 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
6241 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
6243 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
6244 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6245 TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, EVT))) {
6246 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6247 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
6249 LN0->getBasePtr(), EVT,
6250 LN0->getMemOperand());
6251 CombineTo(N, ExtLoad);
6252 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
6253 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6256 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
6257 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
6258 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
6259 N0.getOperand(1), false);
6260 if (BSwap.getNode())
6261 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
6265 // Fold a sext_inreg of a build_vector of ConstantSDNodes or undefs
6266 // into a build_vector.
6267 if (ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
6268 SmallVector<SDValue, 8> Elts;
6269 unsigned NumElts = N0->getNumOperands();
6270 unsigned ShAmt = VTBits - EVTBits;
6272 for (unsigned i = 0; i != NumElts; ++i) {
6273 SDValue Op = N0->getOperand(i);
6274 if (Op->getOpcode() == ISD::UNDEF) {
6279 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
6280 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
6281 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
6282 Op.getValueType()));
6285 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Elts);
6291 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
6292 SDValue N0 = N->getOperand(0);
6293 EVT VT = N->getValueType(0);
6294 bool isLE = TLI.isLittleEndian();
6297 if (N0.getValueType() == N->getValueType(0))
6299 // fold (truncate c1) -> c1
6300 if (isa<ConstantSDNode>(N0))
6301 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0);
6302 // fold (truncate (truncate x)) -> (truncate x)
6303 if (N0.getOpcode() == ISD::TRUNCATE)
6304 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6305 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
6306 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
6307 N0.getOpcode() == ISD::SIGN_EXTEND ||
6308 N0.getOpcode() == ISD::ANY_EXTEND) {
6309 if (N0.getOperand(0).getValueType().bitsLT(VT))
6310 // if the source is smaller than the dest, we still need an extend
6311 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
6313 if (N0.getOperand(0).getValueType().bitsGT(VT))
6314 // if the source is larger than the dest, than we just need the truncate
6315 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
6316 // if the source and dest are the same type, we can drop both the extend
6317 // and the truncate.
6318 return N0.getOperand(0);
6321 // Fold extract-and-trunc into a narrow extract. For example:
6322 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
6323 // i32 y = TRUNCATE(i64 x)
6325 // v16i8 b = BITCAST (v2i64 val)
6326 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
6328 // Note: We only run this optimization after type legalization (which often
6329 // creates this pattern) and before operation legalization after which
6330 // we need to be more careful about the vector instructions that we generate.
6331 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
6332 LegalTypes && !LegalOperations && N0->hasOneUse() && VT != MVT::i1) {
6334 EVT VecTy = N0.getOperand(0).getValueType();
6335 EVT ExTy = N0.getValueType();
6336 EVT TrTy = N->getValueType(0);
6338 unsigned NumElem = VecTy.getVectorNumElements();
6339 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
6341 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
6342 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
6344 SDValue EltNo = N0->getOperand(1);
6345 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
6346 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
6347 EVT IndexTy = TLI.getVectorIdxTy();
6348 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
6350 SDValue V = DAG.getNode(ISD::BITCAST, SDLoc(N),
6351 NVT, N0.getOperand(0));
6353 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
6355 DAG.getConstant(Index, IndexTy));
6359 // trunc (select c, a, b) -> select c, (trunc a), (trunc b)
6360 if (N0.getOpcode() == ISD::SELECT) {
6361 EVT SrcVT = N0.getValueType();
6362 if ((!LegalOperations || TLI.isOperationLegal(ISD::SELECT, SrcVT)) &&
6363 TLI.isTruncateFree(SrcVT, VT)) {
6365 SDValue Cond = N0.getOperand(0);
6366 SDValue TruncOp0 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(1));
6367 SDValue TruncOp1 = DAG.getNode(ISD::TRUNCATE, SL, VT, N0.getOperand(2));
6368 return DAG.getNode(ISD::SELECT, SDLoc(N), VT, Cond, TruncOp0, TruncOp1);
6372 // Fold a series of buildvector, bitcast, and truncate if possible.
6374 // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to
6375 // (2xi32 (buildvector x, y)).
6376 if (Level == AfterLegalizeVectorOps && VT.isVector() &&
6377 N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
6378 N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
6379 N0.getOperand(0).hasOneUse()) {
6381 SDValue BuildVect = N0.getOperand(0);
6382 EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType();
6383 EVT TruncVecEltTy = VT.getVectorElementType();
6385 // Check that the element types match.
6386 if (BuildVectEltTy == TruncVecEltTy) {
6387 // Now we only need to compute the offset of the truncated elements.
6388 unsigned BuildVecNumElts = BuildVect.getNumOperands();
6389 unsigned TruncVecNumElts = VT.getVectorNumElements();
6390 unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts;
6392 assert((BuildVecNumElts % TruncVecNumElts) == 0 &&
6393 "Invalid number of elements");
6395 SmallVector<SDValue, 8> Opnds;
6396 for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset)
6397 Opnds.push_back(BuildVect.getOperand(i));
6399 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
6403 // See if we can simplify the input to this truncate through knowledge that
6404 // only the low bits are being used.
6405 // For example "trunc (or (shl x, 8), y)" // -> trunc y
6406 // Currently we only perform this optimization on scalars because vectors
6407 // may have different active low bits.
6408 if (!VT.isVector()) {
6410 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
6411 VT.getSizeInBits()));
6412 if (Shorter.getNode())
6413 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter);
6415 // fold (truncate (load x)) -> (smaller load x)
6416 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
6417 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
6418 SDValue Reduced = ReduceLoadWidth(N);
6419 if (Reduced.getNode())
6421 // Handle the case where the load remains an extending load even
6422 // after truncation.
6423 if (N0.hasOneUse() && ISD::isUNINDEXEDLoad(N0.getNode())) {
6424 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6425 if (!LN0->isVolatile() &&
6426 LN0->getMemoryVT().getStoreSizeInBits() < VT.getSizeInBits()) {
6427 SDValue NewLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(LN0),
6428 VT, LN0->getChain(), LN0->getBasePtr(),
6430 LN0->getMemOperand());
6431 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLoad.getValue(1));
6436 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
6437 // where ... are all 'undef'.
6438 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
6439 SmallVector<EVT, 8> VTs;
6442 unsigned NumDefs = 0;
6444 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
6445 SDValue X = N0.getOperand(i);
6446 if (X.getOpcode() != ISD::UNDEF) {
6451 // Stop if more than one members are non-undef.
6454 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
6455 VT.getVectorElementType(),
6456 X.getValueType().getVectorNumElements()));
6460 return DAG.getUNDEF(VT);
6463 assert(V.getNode() && "The single defined operand is empty!");
6464 SmallVector<SDValue, 8> Opnds;
6465 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
6467 Opnds.push_back(DAG.getUNDEF(VTs[i]));
6470 SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V);
6471 AddToWorklist(NV.getNode());
6472 Opnds.push_back(NV);
6474 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Opnds);
6478 // Simplify the operands using demanded-bits information.
6479 if (!VT.isVector() &&
6480 SimplifyDemandedBits(SDValue(N, 0)))
6481 return SDValue(N, 0);
6486 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
6487 SDValue Elt = N->getOperand(i);
6488 if (Elt.getOpcode() != ISD::MERGE_VALUES)
6489 return Elt.getNode();
6490 return Elt.getOperand(Elt.getResNo()).getNode();
6493 /// build_pair (load, load) -> load
6494 /// if load locations are consecutive.
6495 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
6496 assert(N->getOpcode() == ISD::BUILD_PAIR);
6498 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
6499 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
6500 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
6501 LD1->getAddressSpace() != LD2->getAddressSpace())
6503 EVT LD1VT = LD1->getValueType(0);
6505 if (ISD::isNON_EXTLoad(LD2) &&
6507 // If both are volatile this would reduce the number of volatile loads.
6508 // If one is volatile it might be ok, but play conservative and bail out.
6509 !LD1->isVolatile() &&
6510 !LD2->isVolatile() &&
6511 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
6512 unsigned Align = LD1->getAlignment();
6513 unsigned NewAlign = TLI.getDataLayout()->
6514 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6516 if (NewAlign <= Align &&
6517 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
6518 return DAG.getLoad(VT, SDLoc(N), LD1->getChain(),
6519 LD1->getBasePtr(), LD1->getPointerInfo(),
6520 false, false, false, Align);
6526 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
6527 SDValue N0 = N->getOperand(0);
6528 EVT VT = N->getValueType(0);
6530 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
6531 // Only do this before legalize, since afterward the target may be depending
6532 // on the bitconvert.
6533 // First check to see if this is all constant.
6535 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
6537 bool isSimple = cast<BuildVectorSDNode>(N0)->isConstant();
6539 EVT DestEltVT = N->getValueType(0).getVectorElementType();
6540 assert(!DestEltVT.isVector() &&
6541 "Element type of vector ValueType must not be vector!");
6543 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
6546 // If the input is a constant, let getNode fold it.
6547 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
6548 // If we can't allow illegal operations, we need to check that this is just
6549 // a fp -> int or int -> conversion and that the resulting operation will
6551 if (!LegalOperations ||
6552 (isa<ConstantSDNode>(N0) && VT.isFloatingPoint() && !VT.isVector() &&
6553 TLI.isOperationLegal(ISD::ConstantFP, VT)) ||
6554 (isa<ConstantFPSDNode>(N0) && VT.isInteger() && !VT.isVector() &&
6555 TLI.isOperationLegal(ISD::Constant, VT)))
6556 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, N0);
6559 // (conv (conv x, t1), t2) -> (conv x, t2)
6560 if (N0.getOpcode() == ISD::BITCAST)
6561 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT,
6564 // fold (conv (load x)) -> (load (conv*)x)
6565 // If the resultant load doesn't need a higher alignment than the original!
6566 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6567 // Do not change the width of a volatile load.
6568 !cast<LoadSDNode>(N0)->isVolatile() &&
6569 // Do not remove the cast if the types differ in endian layout.
6570 TLI.hasBigEndianPartOrdering(N0.getValueType()) ==
6571 TLI.hasBigEndianPartOrdering(VT) &&
6572 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)) &&
6573 TLI.isLoadBitCastBeneficial(N0.getValueType(), VT)) {
6574 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6575 unsigned Align = TLI.getDataLayout()->
6576 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6577 unsigned OrigAlign = LN0->getAlignment();
6579 if (Align <= OrigAlign) {
6580 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(),
6581 LN0->getBasePtr(), LN0->getPointerInfo(),
6582 LN0->isVolatile(), LN0->isNonTemporal(),
6583 LN0->isInvariant(), OrigAlign,
6585 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
6590 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
6591 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
6592 // This often reduces constant pool loads.
6593 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) ||
6594 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) &&
6595 N0.getNode()->hasOneUse() && VT.isInteger() &&
6596 !VT.isVector() && !N0.getValueType().isVector()) {
6597 SDValue NewConv = DAG.getNode(ISD::BITCAST, SDLoc(N0), VT,
6599 AddToWorklist(NewConv.getNode());
6601 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6602 if (N0.getOpcode() == ISD::FNEG)
6603 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
6604 NewConv, DAG.getConstant(SignBit, VT));
6605 assert(N0.getOpcode() == ISD::FABS);
6606 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6607 NewConv, DAG.getConstant(~SignBit, VT));
6610 // fold (bitconvert (fcopysign cst, x)) ->
6611 // (or (and (bitconvert x), sign), (and cst, (not sign)))
6612 // Note that we don't handle (copysign x, cst) because this can always be
6613 // folded to an fneg or fabs.
6614 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
6615 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
6616 VT.isInteger() && !VT.isVector()) {
6617 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
6618 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
6619 if (isTypeLegal(IntXVT)) {
6620 SDValue X = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6621 IntXVT, N0.getOperand(1));
6622 AddToWorklist(X.getNode());
6624 // If X has a different width than the result/lhs, sext it or truncate it.
6625 unsigned VTWidth = VT.getSizeInBits();
6626 if (OrigXWidth < VTWidth) {
6627 X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X);
6628 AddToWorklist(X.getNode());
6629 } else if (OrigXWidth > VTWidth) {
6630 // To get the sign bit in the right place, we have to shift it right
6631 // before truncating.
6632 X = DAG.getNode(ISD::SRL, SDLoc(X),
6633 X.getValueType(), X,
6634 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
6635 AddToWorklist(X.getNode());
6636 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
6637 AddToWorklist(X.getNode());
6640 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6641 X = DAG.getNode(ISD::AND, SDLoc(X), VT,
6642 X, DAG.getConstant(SignBit, VT));
6643 AddToWorklist(X.getNode());
6645 SDValue Cst = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6646 VT, N0.getOperand(0));
6647 Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT,
6648 Cst, DAG.getConstant(~SignBit, VT));
6649 AddToWorklist(Cst.getNode());
6651 return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst);
6655 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
6656 if (N0.getOpcode() == ISD::BUILD_PAIR) {
6657 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
6658 if (CombineLD.getNode())
6665 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
6666 EVT VT = N->getValueType(0);
6667 return CombineConsecutiveLoads(N, VT);
6670 /// We know that BV is a build_vector node with Constant, ConstantFP or Undef
6671 /// operands. DstEltVT indicates the destination element value type.
6672 SDValue DAGCombiner::
6673 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
6674 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
6676 // If this is already the right type, we're done.
6677 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
6679 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
6680 unsigned DstBitSize = DstEltVT.getSizeInBits();
6682 // If this is a conversion of N elements of one type to N elements of another
6683 // type, convert each element. This handles FP<->INT cases.
6684 if (SrcBitSize == DstBitSize) {
6685 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6686 BV->getValueType(0).getVectorNumElements());
6688 // Due to the FP element handling below calling this routine recursively,
6689 // we can end up with a scalar-to-vector node here.
6690 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
6691 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6692 DAG.getNode(ISD::BITCAST, SDLoc(BV),
6693 DstEltVT, BV->getOperand(0)));
6695 SmallVector<SDValue, 8> Ops;
6696 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6697 SDValue Op = BV->getOperand(i);
6698 // If the vector element type is not legal, the BUILD_VECTOR operands
6699 // are promoted and implicitly truncated. Make that explicit here.
6700 if (Op.getValueType() != SrcEltVT)
6701 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op);
6702 Ops.push_back(DAG.getNode(ISD::BITCAST, SDLoc(BV),
6704 AddToWorklist(Ops.back().getNode());
6706 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6709 // Otherwise, we're growing or shrinking the elements. To avoid having to
6710 // handle annoying details of growing/shrinking FP values, we convert them to
6712 if (SrcEltVT.isFloatingPoint()) {
6713 // Convert the input float vector to a int vector where the elements are the
6715 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
6716 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
6720 // Now we know the input is an integer vector. If the output is a FP type,
6721 // convert to integer first, then to FP of the right size.
6722 if (DstEltVT.isFloatingPoint()) {
6723 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
6724 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
6726 // Next, convert to FP elements of the same size.
6727 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
6730 // Okay, we know the src/dst types are both integers of differing types.
6731 // Handling growing first.
6732 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
6733 if (SrcBitSize < DstBitSize) {
6734 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
6736 SmallVector<SDValue, 8> Ops;
6737 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
6738 i += NumInputsPerOutput) {
6739 bool isLE = TLI.isLittleEndian();
6740 APInt NewBits = APInt(DstBitSize, 0);
6741 bool EltIsUndef = true;
6742 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
6743 // Shift the previously computed bits over.
6744 NewBits <<= SrcBitSize;
6745 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
6746 if (Op.getOpcode() == ISD::UNDEF) continue;
6749 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
6750 zextOrTrunc(SrcBitSize).zext(DstBitSize);
6754 Ops.push_back(DAG.getUNDEF(DstEltVT));
6756 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
6759 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
6760 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6763 // Finally, this must be the case where we are shrinking elements: each input
6764 // turns into multiple outputs.
6765 bool isS2V = ISD::isScalarToVector(BV);
6766 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
6767 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6768 NumOutputsPerInput*BV->getNumOperands());
6769 SmallVector<SDValue, 8> Ops;
6771 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6772 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
6773 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
6774 Ops.push_back(DAG.getUNDEF(DstEltVT));
6778 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
6779 getAPIntValue().zextOrTrunc(SrcBitSize);
6781 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
6782 APInt ThisVal = OpVal.trunc(DstBitSize);
6783 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
6784 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
6785 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
6786 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6788 OpVal = OpVal.lshr(DstBitSize);
6791 // For big endian targets, swap the order of the pieces of each element.
6792 if (TLI.isBigEndian())
6793 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
6796 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT, Ops);
6799 SDValue DAGCombiner::visitFADD(SDNode *N) {
6800 SDValue N0 = N->getOperand(0);
6801 SDValue N1 = N->getOperand(1);
6802 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6803 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6804 EVT VT = N->getValueType(0);
6805 const TargetOptions &Options = DAG.getTarget().Options;
6808 if (VT.isVector()) {
6809 SDValue FoldedVOp = SimplifyVBinOp(N);
6810 if (FoldedVOp.getNode()) return FoldedVOp;
6813 // fold (fadd c1, c2) -> c1 + c2
6815 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N1);
6817 // canonicalize constant to RHS
6818 if (N0CFP && !N1CFP)
6819 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N0);
6821 // fold (fadd A, (fneg B)) -> (fsub A, B)
6822 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6823 isNegatibleForFree(N1, LegalOperations, TLI, &Options) == 2)
6824 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0,
6825 GetNegatedExpression(N1, DAG, LegalOperations));
6827 // fold (fadd (fneg A), B) -> (fsub B, A)
6828 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6829 isNegatibleForFree(N0, LegalOperations, TLI, &Options) == 2)
6830 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N1,
6831 GetNegatedExpression(N0, DAG, LegalOperations));
6833 // If 'unsafe math' is enabled, fold lots of things.
6834 if (Options.UnsafeFPMath) {
6835 // No FP constant should be created after legalization as Instruction
6836 // Selection pass has a hard time dealing with FP constants.
6837 bool AllowNewConst = (Level < AfterLegalizeDAG);
6839 // fold (fadd A, 0) -> A
6840 if (N1CFP && N1CFP->getValueAPF().isZero())
6843 // fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
6844 if (N1CFP && N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
6845 isa<ConstantFPSDNode>(N0.getOperand(1)))
6846 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0.getOperand(0),
6847 DAG.getNode(ISD::FADD, SDLoc(N), VT,
6848 N0.getOperand(1), N1));
6850 // If allowed, fold (fadd (fneg x), x) -> 0.0
6851 if (AllowNewConst && N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1)
6852 return DAG.getConstantFP(0.0, VT);
6854 // If allowed, fold (fadd x, (fneg x)) -> 0.0
6855 if (AllowNewConst && N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0)
6856 return DAG.getConstantFP(0.0, VT);
6858 // We can fold chains of FADD's of the same value into multiplications.
6859 // This transform is not safe in general because we are reducing the number
6860 // of rounding steps.
6861 if (TLI.isOperationLegalOrCustom(ISD::FMUL, VT) && !N0CFP && !N1CFP) {
6862 if (N0.getOpcode() == ISD::FMUL) {
6863 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
6864 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
6866 // (fadd (fmul x, c), x) -> (fmul x, c+1)
6867 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
6868 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6870 DAG.getConstantFP(1.0, VT));
6871 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, NewCFP);
6874 // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2)
6875 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
6876 N1.getOperand(0) == N1.getOperand(1) &&
6877 N0.getOperand(0) == N1.getOperand(0)) {
6878 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6880 DAG.getConstantFP(2.0, VT));
6881 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6882 N0.getOperand(0), NewCFP);
6886 if (N1.getOpcode() == ISD::FMUL) {
6887 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
6888 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
6890 // (fadd x, (fmul x, c)) -> (fmul x, c+1)
6891 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
6892 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6894 DAG.getConstantFP(1.0, VT));
6895 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, NewCFP);
6898 // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2)
6899 if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD &&
6900 N0.getOperand(0) == N0.getOperand(1) &&
6901 N1.getOperand(0) == N0.getOperand(0)) {
6902 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6904 DAG.getConstantFP(2.0, VT));
6905 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1.getOperand(0), NewCFP);
6909 if (N0.getOpcode() == ISD::FADD && AllowNewConst) {
6910 ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
6911 // (fadd (fadd x, x), x) -> (fmul x, 3.0)
6912 if (!CFP && N0.getOperand(0) == N0.getOperand(1) &&
6913 (N0.getOperand(0) == N1))
6914 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6915 N1, DAG.getConstantFP(3.0, VT));
6918 if (N1.getOpcode() == ISD::FADD && AllowNewConst) {
6919 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
6920 // (fadd x, (fadd x, x)) -> (fmul x, 3.0)
6921 if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) &&
6922 N1.getOperand(0) == N0)
6923 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6924 N0, DAG.getConstantFP(3.0, VT));
6927 // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0)
6928 if (AllowNewConst &&
6929 N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
6930 N0.getOperand(0) == N0.getOperand(1) &&
6931 N1.getOperand(0) == N1.getOperand(1) &&
6932 N0.getOperand(0) == N1.getOperand(0))
6933 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6934 N0.getOperand(0), DAG.getConstantFP(4.0, VT));
6936 } // enable-unsafe-fp-math
6938 // FADD -> FMA combines:
6939 if ((Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath) &&
6940 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
6941 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
6943 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
6944 if (N0.getOpcode() == ISD::FMUL &&
6945 (N0->hasOneUse() || TLI.enableAggressiveFMAFusion(VT)))
6946 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6947 N0.getOperand(0), N0.getOperand(1), N1);
6949 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
6950 // Note: Commutes FADD operands.
6951 if (N1.getOpcode() == ISD::FMUL &&
6952 (N1->hasOneUse() || TLI.enableAggressiveFMAFusion(VT)))
6953 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6954 N1.getOperand(0), N1.getOperand(1), N0);
6956 // When FP_EXTEND nodes are free on the target, and there is an opportunity
6957 // to combine into FMA, arrange such nodes accordingly.
6958 if (TLI.isFPExtFree(VT)) {
6960 // fold (fadd (fpext (fmul x, y)), z) -> (fma (fpext x), (fpext y), z)
6961 if (N0.getOpcode() == ISD::FP_EXTEND) {
6962 SDValue N00 = N0.getOperand(0);
6963 if (N00.getOpcode() == ISD::FMUL)
6964 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6965 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
6967 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
6968 N00.getOperand(1)), N1);
6971 // fold (fadd x, (fpext (fmul y, z)), z) -> (fma (fpext y), (fpext z), x)
6972 // Note: Commutes FADD operands.
6973 if (N1.getOpcode() == ISD::FP_EXTEND) {
6974 SDValue N10 = N1.getOperand(0);
6975 if (N10.getOpcode() == ISD::FMUL)
6976 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6977 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
6979 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
6980 N10.getOperand(1)), N0);
6984 // More folding opportunities when target permits.
6985 if (TLI.enableAggressiveFMAFusion(VT)) {
6987 // fold (fadd (fma x, y, (fmul u, v)), z) -> (fma x, y (fma u, v, z))
6988 if (N0.getOpcode() == ISD::FMA &&
6989 N0.getOperand(2).getOpcode() == ISD::FMUL)
6990 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6991 N0.getOperand(0), N0.getOperand(1),
6992 DAG.getNode(ISD::FMA, SDLoc(N), VT,
6993 N0.getOperand(2).getOperand(0),
6994 N0.getOperand(2).getOperand(1),
6997 // fold (fadd x, (fma y, z, (fmul u, v)) -> (fma y, z (fma u, v, x))
6998 if (N1->getOpcode() == ISD::FMA &&
6999 N1.getOperand(2).getOpcode() == ISD::FMUL)
7000 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7001 N1.getOperand(0), N1.getOperand(1),
7002 DAG.getNode(ISD::FMA, SDLoc(N), VT,
7003 N1.getOperand(2).getOperand(0),
7004 N1.getOperand(2).getOperand(1),
7012 SDValue DAGCombiner::visitFSUB(SDNode *N) {
7013 SDValue N0 = N->getOperand(0);
7014 SDValue N1 = N->getOperand(1);
7015 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
7016 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
7017 EVT VT = N->getValueType(0);
7019 const TargetOptions &Options = DAG.getTarget().Options;
7022 if (VT.isVector()) {
7023 SDValue FoldedVOp = SimplifyVBinOp(N);
7024 if (FoldedVOp.getNode()) return FoldedVOp;
7027 // fold (fsub c1, c2) -> c1-c2
7029 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0, N1);
7031 // fold (fsub A, (fneg B)) -> (fadd A, B)
7032 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
7033 return DAG.getNode(ISD::FADD, dl, VT, N0,
7034 GetNegatedExpression(N1, DAG, LegalOperations));
7036 // If 'unsafe math' is enabled, fold lots of things.
7037 if (Options.UnsafeFPMath) {
7039 if (N1CFP && N1CFP->getValueAPF().isZero())
7042 // (fsub 0, B) -> -B
7043 if (N0CFP && N0CFP->getValueAPF().isZero()) {
7044 if (isNegatibleForFree(N1, LegalOperations, TLI, &Options))
7045 return GetNegatedExpression(N1, DAG, LegalOperations);
7046 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7047 return DAG.getNode(ISD::FNEG, dl, VT, N1);
7050 // (fsub x, x) -> 0.0
7052 return DAG.getConstantFP(0.0f, VT);
7054 // (fsub x, (fadd x, y)) -> (fneg y)
7055 // (fsub x, (fadd y, x)) -> (fneg y)
7056 if (N1.getOpcode() == ISD::FADD) {
7057 SDValue N10 = N1->getOperand(0);
7058 SDValue N11 = N1->getOperand(1);
7060 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI, &Options))
7061 return GetNegatedExpression(N11, DAG, LegalOperations);
7063 if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI, &Options))
7064 return GetNegatedExpression(N10, DAG, LegalOperations);
7068 // FSUB -> FMA combines:
7069 if ((Options.AllowFPOpFusion == FPOpFusion::Fast || Options.UnsafeFPMath) &&
7070 TLI.isFMAFasterThanFMulAndFAdd(VT) &&
7071 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
7073 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
7074 if (N0.getOpcode() == ISD::FMUL &&
7075 (N0->hasOneUse() || TLI.enableAggressiveFMAFusion(VT)))
7076 return DAG.getNode(ISD::FMA, dl, VT,
7077 N0.getOperand(0), N0.getOperand(1),
7078 DAG.getNode(ISD::FNEG, dl, VT, N1));
7080 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
7081 // Note: Commutes FSUB operands.
7082 if (N1.getOpcode() == ISD::FMUL &&
7083 (N1->hasOneUse() || TLI.enableAggressiveFMAFusion(VT)))
7084 return DAG.getNode(ISD::FMA, dl, VT,
7085 DAG.getNode(ISD::FNEG, dl, VT,
7087 N1.getOperand(1), N0);
7089 // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
7090 if (N0.getOpcode() == ISD::FNEG &&
7091 N0.getOperand(0).getOpcode() == ISD::FMUL &&
7092 ((N0->hasOneUse() && N0.getOperand(0).hasOneUse()) ||
7093 TLI.enableAggressiveFMAFusion(VT))) {
7094 SDValue N00 = N0.getOperand(0).getOperand(0);
7095 SDValue N01 = N0.getOperand(0).getOperand(1);
7096 return DAG.getNode(ISD::FMA, dl, VT,
7097 DAG.getNode(ISD::FNEG, dl, VT, N00), N01,
7098 DAG.getNode(ISD::FNEG, dl, VT, N1));
7101 // When FP_EXTEND nodes are free on the target, and there is an opportunity
7102 // to combine into FMA, arrange such nodes accordingly.
7103 if (TLI.isFPExtFree(VT)) {
7105 // fold (fsub (fpext (fmul x, y)), z)
7106 // -> (fma (fpext x), (fpext y), (fneg z))
7107 if (N0.getOpcode() == ISD::FP_EXTEND) {
7108 SDValue N00 = N0.getOperand(0);
7109 if (N00.getOpcode() == ISD::FMUL)
7110 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7111 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7113 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7115 DAG.getNode(ISD::FNEG, SDLoc(N), VT, N1));
7118 // fold (fsub x, (fpext (fmul y, z)))
7119 // -> (fma (fneg (fpext y)), (fpext z), x)
7120 // Note: Commutes FSUB operands.
7121 if (N1.getOpcode() == ISD::FP_EXTEND) {
7122 SDValue N10 = N1.getOperand(0);
7123 if (N10.getOpcode() == ISD::FMUL)
7124 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7125 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7126 DAG.getNode(ISD::FP_EXTEND, SDLoc(N),
7127 VT, N10.getOperand(0))),
7128 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7133 // fold (fsub (fpext (fneg (fmul, x, y))), z)
7134 // -> (fma (fneg (fpext x)), (fpext y), (fneg z))
7135 if (N0.getOpcode() == ISD::FP_EXTEND) {
7136 SDValue N00 = N0.getOperand(0);
7137 if (N00.getOpcode() == ISD::FNEG) {
7138 SDValue N000 = N00.getOperand(0);
7139 if (N000.getOpcode() == ISD::FMUL) {
7140 return DAG.getNode(ISD::FMA, dl, VT,
7141 DAG.getNode(ISD::FNEG, dl, VT,
7142 DAG.getNode(ISD::FP_EXTEND, SDLoc(N),
7143 VT, N000.getOperand(0))),
7144 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7145 N000.getOperand(1)),
7146 DAG.getNode(ISD::FNEG, dl, VT, N1));
7151 // fold (fsub (fneg (fpext (fmul, x, y))), z)
7152 // -> (fma (fneg (fpext x)), (fpext y), (fneg z))
7153 if (N0.getOpcode() == ISD::FNEG) {
7154 SDValue N00 = N0.getOperand(0);
7155 if (N00.getOpcode() == ISD::FP_EXTEND) {
7156 SDValue N000 = N00.getOperand(0);
7157 if (N000.getOpcode() == ISD::FMUL) {
7158 return DAG.getNode(ISD::FMA, dl, VT,
7159 DAG.getNode(ISD::FNEG, dl, VT,
7160 DAG.getNode(ISD::FP_EXTEND, SDLoc(N),
7161 VT, N000.getOperand(0))),
7162 DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT,
7163 N000.getOperand(1)),
7164 DAG.getNode(ISD::FNEG, dl, VT, N1));
7170 // More folding opportunities when target permits.
7171 if (TLI.enableAggressiveFMAFusion(VT)) {
7173 // fold (fsub (fma x, y, (fmul u, v)), z)
7174 // -> (fma x, y (fma u, v, (fneg z)))
7175 if (N0.getOpcode() == ISD::FMA &&
7176 N0.getOperand(2).getOpcode() == ISD::FMUL)
7177 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7178 N0.getOperand(0), N0.getOperand(1),
7179 DAG.getNode(ISD::FMA, SDLoc(N), VT,
7180 N0.getOperand(2).getOperand(0),
7181 N0.getOperand(2).getOperand(1),
7182 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7185 // fold (fsub x, (fma y, z, (fmul u, v)))
7186 // -> (fma (fneg y), z, (fma (fneg u), v, x))
7187 if (N1.getOpcode() == ISD::FMA &&
7188 N1.getOperand(2).getOpcode() == ISD::FMUL) {
7189 SDValue N20 = N1.getOperand(2).getOperand(0);
7190 SDValue N21 = N1.getOperand(2).getOperand(1);
7191 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
7192 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7195 DAG.getNode(ISD::FMA, SDLoc(N), VT,
7196 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7206 SDValue DAGCombiner::visitFMUL(SDNode *N) {
7207 SDValue N0 = N->getOperand(0);
7208 SDValue N1 = N->getOperand(1);
7209 ConstantFPSDNode *N0CFP = isConstOrConstSplatFP(N0);
7210 ConstantFPSDNode *N1CFP = isConstOrConstSplatFP(N1);
7211 EVT VT = N->getValueType(0);
7212 const TargetOptions &Options = DAG.getTarget().Options;
7215 if (VT.isVector()) {
7216 // This just handles C1 * C2 for vectors. Other vector folds are below.
7217 SDValue FoldedVOp = SimplifyVBinOp(N);
7218 if (FoldedVOp.getNode())
7220 // Canonicalize vector constant to RHS.
7221 if (N0.getOpcode() == ISD::BUILD_VECTOR &&
7222 N1.getOpcode() != ISD::BUILD_VECTOR)
7223 if (auto *BV0 = dyn_cast<BuildVectorSDNode>(N0))
7224 if (BV0->isConstant())
7225 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, N1, N0);
7228 // fold (fmul c1, c2) -> c1*c2
7230 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, N1);
7232 // canonicalize constant to RHS
7233 if (N0CFP && !N1CFP)
7234 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, N0);
7236 // fold (fmul A, 1.0) -> A
7237 if (N1CFP && N1CFP->isExactlyValue(1.0))
7240 if (Options.UnsafeFPMath) {
7241 // fold (fmul A, 0) -> 0
7242 if (N1CFP && N1CFP->getValueAPF().isZero())
7245 // fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
7246 if (N0.getOpcode() == ISD::FMUL) {
7247 // Fold scalars or any vector constants (not just splats).
7248 // This fold is done in general by InstCombine, but extra fmul insts
7249 // may have been generated during lowering.
7250 SDValue N01 = N0.getOperand(1);
7251 auto *BV1 = dyn_cast<BuildVectorSDNode>(N1);
7252 auto *BV01 = dyn_cast<BuildVectorSDNode>(N01);
7253 if ((N1CFP && isConstOrConstSplatFP(N01)) ||
7254 (BV1 && BV01 && BV1->isConstant() && BV01->isConstant())) {
7256 SDValue MulConsts = DAG.getNode(ISD::FMUL, SL, VT, N01, N1);
7257 return DAG.getNode(ISD::FMUL, SL, VT, N0.getOperand(0), MulConsts);
7261 // fold (fmul (fadd x, x), c) -> (fmul x, (fmul 2.0, c))
7262 // Undo the fmul 2.0, x -> fadd x, x transformation, since if it occurs
7263 // during an early run of DAGCombiner can prevent folding with fmuls
7264 // inserted during lowering.
7265 if (N0.getOpcode() == ISD::FADD && N0.getOperand(0) == N0.getOperand(1)) {
7267 const SDValue Two = DAG.getConstantFP(2.0, VT);
7268 SDValue MulConsts = DAG.getNode(ISD::FMUL, SL, VT, Two, N1);
7269 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0.getOperand(0), MulConsts);
7273 // fold (fmul X, 2.0) -> (fadd X, X)
7274 if (N1CFP && N1CFP->isExactlyValue(+2.0))
7275 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N0);
7277 // fold (fmul X, -1.0) -> (fneg X)
7278 if (N1CFP && N1CFP->isExactlyValue(-1.0))
7279 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7280 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
7282 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
7283 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
7284 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
7285 // Both can be negated for free, check to see if at least one is cheaper
7287 if (LHSNeg == 2 || RHSNeg == 2)
7288 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7289 GetNegatedExpression(N0, DAG, LegalOperations),
7290 GetNegatedExpression(N1, DAG, LegalOperations));
7297 SDValue DAGCombiner::visitFMA(SDNode *N) {
7298 SDValue N0 = N->getOperand(0);
7299 SDValue N1 = N->getOperand(1);
7300 SDValue N2 = N->getOperand(2);
7301 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7302 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7303 EVT VT = N->getValueType(0);
7305 const TargetOptions &Options = DAG.getTarget().Options;
7307 // Constant fold FMA.
7308 if (isa<ConstantFPSDNode>(N0) &&
7309 isa<ConstantFPSDNode>(N1) &&
7310 isa<ConstantFPSDNode>(N2)) {
7311 return DAG.getNode(ISD::FMA, dl, VT, N0, N1, N2);
7314 if (Options.UnsafeFPMath) {
7315 if (N0CFP && N0CFP->isZero())
7317 if (N1CFP && N1CFP->isZero())
7320 if (N0CFP && N0CFP->isExactlyValue(1.0))
7321 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2);
7322 if (N1CFP && N1CFP->isExactlyValue(1.0))
7323 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2);
7325 // Canonicalize (fma c, x, y) -> (fma x, c, y)
7326 if (N0CFP && !N1CFP)
7327 return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2);
7329 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
7330 if (Options.UnsafeFPMath && N1CFP &&
7331 N2.getOpcode() == ISD::FMUL &&
7332 N0 == N2.getOperand(0) &&
7333 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
7334 return DAG.getNode(ISD::FMUL, dl, VT, N0,
7335 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
7339 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
7340 if (Options.UnsafeFPMath &&
7341 N0.getOpcode() == ISD::FMUL && N1CFP &&
7342 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
7343 return DAG.getNode(ISD::FMA, dl, VT,
7345 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
7349 // (fma x, 1, y) -> (fadd x, y)
7350 // (fma x, -1, y) -> (fadd (fneg x), y)
7352 if (N1CFP->isExactlyValue(1.0))
7353 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
7355 if (N1CFP->isExactlyValue(-1.0) &&
7356 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
7357 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
7358 AddToWorklist(RHSNeg.getNode());
7359 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
7363 // (fma x, c, x) -> (fmul x, (c+1))
7364 if (Options.UnsafeFPMath && N1CFP && N0 == N2)
7365 return DAG.getNode(ISD::FMUL, dl, VT, N0,
7366 DAG.getNode(ISD::FADD, dl, VT,
7367 N1, DAG.getConstantFP(1.0, VT)));
7369 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
7370 if (Options.UnsafeFPMath && N1CFP &&
7371 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0)
7372 return DAG.getNode(ISD::FMUL, dl, VT, N0,
7373 DAG.getNode(ISD::FADD, dl, VT,
7374 N1, DAG.getConstantFP(-1.0, VT)));
7380 SDValue DAGCombiner::visitFDIV(SDNode *N) {
7381 SDValue N0 = N->getOperand(0);
7382 SDValue N1 = N->getOperand(1);
7383 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7384 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7385 EVT VT = N->getValueType(0);
7387 const TargetOptions &Options = DAG.getTarget().Options;
7390 if (VT.isVector()) {
7391 SDValue FoldedVOp = SimplifyVBinOp(N);
7392 if (FoldedVOp.getNode()) return FoldedVOp;
7395 // fold (fdiv c1, c2) -> c1/c2
7397 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1);
7399 if (Options.UnsafeFPMath) {
7400 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
7402 // Compute the reciprocal 1.0 / c2.
7403 APFloat N1APF = N1CFP->getValueAPF();
7404 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
7405 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
7406 // Only do the transform if the reciprocal is a legal fp immediate that
7407 // isn't too nasty (eg NaN, denormal, ...).
7408 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
7409 (!LegalOperations ||
7410 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
7411 // backend)... we should handle this gracefully after Legalize.
7412 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
7413 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
7414 TLI.isFPImmLegal(Recip, VT)))
7415 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0,
7416 DAG.getConstantFP(Recip, VT));
7419 // If this FDIV is part of a reciprocal square root, it may be folded
7420 // into a target-specific square root estimate instruction.
7421 if (N1.getOpcode() == ISD::FSQRT) {
7422 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0))) {
7423 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7425 } else if (N1.getOpcode() == ISD::FP_EXTEND &&
7426 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
7427 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0).getOperand(0))) {
7428 RV = DAG.getNode(ISD::FP_EXTEND, SDLoc(N1), VT, RV);
7429 AddToWorklist(RV.getNode());
7430 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7432 } else if (N1.getOpcode() == ISD::FP_ROUND &&
7433 N1.getOperand(0).getOpcode() == ISD::FSQRT) {
7434 if (SDValue RV = BuildRsqrtEstimate(N1.getOperand(0).getOperand(0))) {
7435 RV = DAG.getNode(ISD::FP_ROUND, SDLoc(N1), VT, RV, N1.getOperand(1));
7436 AddToWorklist(RV.getNode());
7437 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7439 } else if (N1.getOpcode() == ISD::FMUL) {
7440 // Look through an FMUL. Even though this won't remove the FDIV directly,
7441 // it's still worthwhile to get rid of the FSQRT if possible.
7444 if (N1.getOperand(0).getOpcode() == ISD::FSQRT) {
7445 SqrtOp = N1.getOperand(0);
7446 OtherOp = N1.getOperand(1);
7447 } else if (N1.getOperand(1).getOpcode() == ISD::FSQRT) {
7448 SqrtOp = N1.getOperand(1);
7449 OtherOp = N1.getOperand(0);
7451 if (SqrtOp.getNode()) {
7452 // We found a FSQRT, so try to make this fold:
7453 // x / (y * sqrt(z)) -> x * (rsqrt(z) / y)
7454 if (SDValue RV = BuildRsqrtEstimate(SqrtOp.getOperand(0))) {
7455 RV = DAG.getNode(ISD::FDIV, SDLoc(N1), VT, RV, OtherOp);
7456 AddToWorklist(RV.getNode());
7457 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7462 // Fold into a reciprocal estimate and multiply instead of a real divide.
7463 if (SDValue RV = BuildReciprocalEstimate(N1)) {
7464 AddToWorklist(RV.getNode());
7465 return DAG.getNode(ISD::FMUL, DL, VT, N0, RV);
7469 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
7470 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI, &Options)) {
7471 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI, &Options)) {
7472 // Both can be negated for free, check to see if at least one is cheaper
7474 if (LHSNeg == 2 || RHSNeg == 2)
7475 return DAG.getNode(ISD::FDIV, SDLoc(N), VT,
7476 GetNegatedExpression(N0, DAG, LegalOperations),
7477 GetNegatedExpression(N1, DAG, LegalOperations));
7481 // Combine multiple FDIVs with the same divisor into multiple FMULs by the
7483 // E.g., (a / D; b / D;) -> (recip = 1.0 / D; a * recip; b * recip)
7484 // Notice that this is not always beneficial. One reason is different target
7485 // may have different costs for FDIV and FMUL, so sometimes the cost of two
7486 // FDIVs may be lower than the cost of one FDIV and two FMULs. Another reason
7487 // is the critical path is increased from "one FDIV" to "one FDIV + one FMUL".
7488 if (Options.UnsafeFPMath) {
7489 // Skip if current node is a reciprocal.
7490 if (N0CFP && N0CFP->isExactlyValue(1.0))
7493 SmallVector<SDNode *, 4> Users;
7494 // Find all FDIV users of the same divisor.
7495 for (SDNode::use_iterator UI = N1.getNode()->use_begin(),
7496 UE = N1.getNode()->use_end();
7498 SDNode *User = UI.getUse().getUser();
7499 if (User->getOpcode() == ISD::FDIV && User->getOperand(1) == N1)
7500 Users.push_back(User);
7503 if (TLI.combineRepeatedFPDivisors(Users.size())) {
7504 SDValue FPOne = DAG.getConstantFP(1.0, VT); // floating point 1.0
7505 SDValue Reciprocal = DAG.getNode(ISD::FDIV, SDLoc(N), VT, FPOne, N1);
7507 // Dividend / Divisor -> Dividend * Reciprocal
7508 for (auto I = Users.begin(), E = Users.end(); I != E; ++I) {
7509 if ((*I)->getOperand(0) != FPOne) {
7510 SDValue NewNode = DAG.getNode(ISD::FMUL, SDLoc(*I), VT,
7511 (*I)->getOperand(0), Reciprocal);
7512 DAG.ReplaceAllUsesWith(*I, NewNode.getNode());
7522 SDValue DAGCombiner::visitFREM(SDNode *N) {
7523 SDValue N0 = N->getOperand(0);
7524 SDValue N1 = N->getOperand(1);
7525 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7526 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7527 EVT VT = N->getValueType(0);
7529 // fold (frem c1, c2) -> fmod(c1,c2)
7531 return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1);
7536 SDValue DAGCombiner::visitFSQRT(SDNode *N) {
7537 if (DAG.getTarget().Options.UnsafeFPMath &&
7538 !TLI.isFsqrtCheap()) {
7539 // Compute this as X * (1/sqrt(X)) = X * (X ** -0.5)
7540 if (SDValue RV = BuildRsqrtEstimate(N->getOperand(0))) {
7541 EVT VT = RV.getValueType();
7542 RV = DAG.getNode(ISD::FMUL, SDLoc(N), VT, N->getOperand(0), RV);
7543 AddToWorklist(RV.getNode());
7545 // Unfortunately, RV is now NaN if the input was exactly 0.
7546 // Select out this case and force the answer to 0.
7547 SDValue Zero = DAG.getConstantFP(0.0, VT);
7549 DAG.getSetCC(SDLoc(N), TLI.getSetCCResultType(*DAG.getContext(), VT),
7550 N->getOperand(0), Zero, ISD::SETEQ);
7551 AddToWorklist(ZeroCmp.getNode());
7552 AddToWorklist(RV.getNode());
7554 RV = DAG.getNode(VT.isVector() ? ISD::VSELECT : ISD::SELECT,
7555 SDLoc(N), VT, ZeroCmp, Zero, RV);
7562 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
7563 SDValue N0 = N->getOperand(0);
7564 SDValue N1 = N->getOperand(1);
7565 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7566 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7567 EVT VT = N->getValueType(0);
7569 if (N0CFP && N1CFP) // Constant fold
7570 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1);
7573 const APFloat& V = N1CFP->getValueAPF();
7574 // copysign(x, c1) -> fabs(x) iff ispos(c1)
7575 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
7576 if (!V.isNegative()) {
7577 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
7578 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7580 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
7581 return DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7582 DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0));
7586 // copysign(fabs(x), y) -> copysign(x, y)
7587 // copysign(fneg(x), y) -> copysign(x, y)
7588 // copysign(copysign(x,z), y) -> copysign(x, y)
7589 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
7590 N0.getOpcode() == ISD::FCOPYSIGN)
7591 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7592 N0.getOperand(0), N1);
7594 // copysign(x, abs(y)) -> abs(x)
7595 if (N1.getOpcode() == ISD::FABS)
7596 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7598 // copysign(x, copysign(y,z)) -> copysign(x, z)
7599 if (N1.getOpcode() == ISD::FCOPYSIGN)
7600 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7601 N0, N1.getOperand(1));
7603 // copysign(x, fp_extend(y)) -> copysign(x, y)
7604 // copysign(x, fp_round(y)) -> copysign(x, y)
7605 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
7606 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7607 N0, N1.getOperand(0));
7612 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
7613 SDValue N0 = N->getOperand(0);
7614 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
7615 EVT VT = N->getValueType(0);
7616 EVT OpVT = N0.getValueType();
7618 // fold (sint_to_fp c1) -> c1fp
7620 // ...but only if the target supports immediate floating-point values
7621 (!LegalOperations ||
7622 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
7623 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
7625 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
7626 // but UINT_TO_FP is legal on this target, try to convert.
7627 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
7628 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
7629 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
7630 if (DAG.SignBitIsZero(N0))
7631 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
7634 // The next optimizations are desirable only if SELECT_CC can be lowered.
7635 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
7636 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7637 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
7639 (!LegalOperations ||
7640 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7642 { N0.getOperand(0), N0.getOperand(1),
7643 DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT),
7645 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7648 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
7649 // (select_cc x, y, 1.0, 0.0,, cc)
7650 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
7651 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
7652 (!LegalOperations ||
7653 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7655 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
7656 DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT),
7657 N0.getOperand(0).getOperand(2) };
7658 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7665 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
7666 SDValue N0 = N->getOperand(0);
7667 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
7668 EVT VT = N->getValueType(0);
7669 EVT OpVT = N0.getValueType();
7671 // fold (uint_to_fp c1) -> c1fp
7673 // ...but only if the target supports immediate floating-point values
7674 (!LegalOperations ||
7675 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
7676 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
7678 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
7679 // but SINT_TO_FP is legal on this target, try to convert.
7680 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
7681 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
7682 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
7683 if (DAG.SignBitIsZero(N0))
7684 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
7687 // The next optimizations are desirable only if SELECT_CC can be lowered.
7688 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT) || !LegalOperations) {
7689 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7691 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
7692 (!LegalOperations ||
7693 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7695 { N0.getOperand(0), N0.getOperand(1),
7696 DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT),
7698 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops);
7705 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
7706 SDValue N0 = N->getOperand(0);
7707 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7708 EVT VT = N->getValueType(0);
7710 // fold (fp_to_sint c1fp) -> c1
7712 return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0);
7717 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
7718 SDValue N0 = N->getOperand(0);
7719 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7720 EVT VT = N->getValueType(0);
7722 // fold (fp_to_uint c1fp) -> c1
7724 return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0);
7729 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
7730 SDValue N0 = N->getOperand(0);
7731 SDValue N1 = N->getOperand(1);
7732 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7733 EVT VT = N->getValueType(0);
7735 // fold (fp_round c1fp) -> c1fp
7737 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1);
7739 // fold (fp_round (fp_extend x)) -> x
7740 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
7741 return N0.getOperand(0);
7743 // fold (fp_round (fp_round x)) -> (fp_round x)
7744 if (N0.getOpcode() == ISD::FP_ROUND) {
7745 // This is a value preserving truncation if both round's are.
7746 bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
7747 N0.getNode()->getConstantOperandVal(1) == 1;
7748 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0.getOperand(0),
7749 DAG.getIntPtrConstant(IsTrunc));
7752 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
7753 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
7754 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT,
7755 N0.getOperand(0), N1);
7756 AddToWorklist(Tmp.getNode());
7757 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7758 Tmp, N0.getOperand(1));
7764 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
7765 SDValue N0 = N->getOperand(0);
7766 EVT VT = N->getValueType(0);
7767 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
7768 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7770 // fold (fp_round_inreg c1fp) -> c1fp
7771 if (N0CFP && isTypeLegal(EVT)) {
7772 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
7773 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, Round);
7779 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
7780 SDValue N0 = N->getOperand(0);
7781 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7782 EVT VT = N->getValueType(0);
7784 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
7785 if (N->hasOneUse() &&
7786 N->use_begin()->getOpcode() == ISD::FP_ROUND)
7789 // fold (fp_extend c1fp) -> c1fp
7791 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0);
7793 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
7795 if (N0.getOpcode() == ISD::FP_ROUND
7796 && N0.getNode()->getConstantOperandVal(1) == 1) {
7797 SDValue In = N0.getOperand(0);
7798 if (In.getValueType() == VT) return In;
7799 if (VT.bitsLT(In.getValueType()))
7800 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT,
7801 In, N0.getOperand(1));
7802 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In);
7805 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
7806 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7807 TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) {
7808 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7809 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
7811 LN0->getBasePtr(), N0.getValueType(),
7812 LN0->getMemOperand());
7813 CombineTo(N, ExtLoad);
7814 CombineTo(N0.getNode(),
7815 DAG.getNode(ISD::FP_ROUND, SDLoc(N0),
7816 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
7817 ExtLoad.getValue(1));
7818 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7824 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
7825 SDValue N0 = N->getOperand(0);
7826 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7827 EVT VT = N->getValueType(0);
7829 // fold (fceil c1) -> fceil(c1)
7831 return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0);
7836 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
7837 SDValue N0 = N->getOperand(0);
7838 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7839 EVT VT = N->getValueType(0);
7841 // fold (ftrunc c1) -> ftrunc(c1)
7843 return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0);
7848 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
7849 SDValue N0 = N->getOperand(0);
7850 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7851 EVT VT = N->getValueType(0);
7853 // fold (ffloor c1) -> ffloor(c1)
7855 return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0);
7860 // FIXME: FNEG and FABS have a lot in common; refactor.
7861 SDValue DAGCombiner::visitFNEG(SDNode *N) {
7862 SDValue N0 = N->getOperand(0);
7863 EVT VT = N->getValueType(0);
7865 if (VT.isVector()) {
7866 SDValue FoldedVOp = SimplifyVUnaryOp(N);
7867 if (FoldedVOp.getNode()) return FoldedVOp;
7870 // Constant fold FNEG.
7871 if (isa<ConstantFPSDNode>(N0))
7872 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N->getOperand(0));
7874 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
7875 &DAG.getTarget().Options))
7876 return GetNegatedExpression(N0, DAG, LegalOperations);
7878 // Transform fneg(bitconvert(x)) -> bitconvert(x ^ sign) to avoid loading
7879 // constant pool values.
7880 if (!TLI.isFNegFree(VT) &&
7881 N0.getOpcode() == ISD::BITCAST &&
7882 N0.getNode()->hasOneUse()) {
7883 SDValue Int = N0.getOperand(0);
7884 EVT IntVT = Int.getValueType();
7885 if (IntVT.isInteger() && !IntVT.isVector()) {
7887 if (N0.getValueType().isVector()) {
7888 // For a vector, get a mask such as 0x80... per scalar element
7890 SignMask = APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
7891 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
7893 // For a scalar, just generate 0x80...
7894 SignMask = APInt::getSignBit(IntVT.getSizeInBits());
7896 Int = DAG.getNode(ISD::XOR, SDLoc(N0), IntVT, Int,
7897 DAG.getConstant(SignMask, IntVT));
7898 AddToWorklist(Int.getNode());
7899 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Int);
7903 // (fneg (fmul c, x)) -> (fmul -c, x)
7904 if (N0.getOpcode() == ISD::FMUL) {
7905 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
7907 APFloat CVal = CFP1->getValueAPF();
7909 if (Level >= AfterLegalizeDAG &&
7910 (TLI.isFPImmLegal(CVal, N->getValueType(0)) ||
7911 TLI.isOperationLegal(ISD::ConstantFP, N->getValueType(0))))
7913 ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
7914 DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0.getOperand(1)));
7921 SDValue DAGCombiner::visitFMINNUM(SDNode *N) {
7922 SDValue N0 = N->getOperand(0);
7923 SDValue N1 = N->getOperand(1);
7924 const ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7925 const ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7927 if (N0CFP && N1CFP) {
7928 const APFloat &C0 = N0CFP->getValueAPF();
7929 const APFloat &C1 = N1CFP->getValueAPF();
7930 return DAG.getConstantFP(minnum(C0, C1), N->getValueType(0));
7934 EVT VT = N->getValueType(0);
7935 // Canonicalize to constant on RHS.
7936 return DAG.getNode(ISD::FMINNUM, SDLoc(N), VT, N1, N0);
7942 SDValue DAGCombiner::visitFMAXNUM(SDNode *N) {
7943 SDValue N0 = N->getOperand(0);
7944 SDValue N1 = N->getOperand(1);
7945 const ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7946 const ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
7948 if (N0CFP && N1CFP) {
7949 const APFloat &C0 = N0CFP->getValueAPF();
7950 const APFloat &C1 = N1CFP->getValueAPF();
7951 return DAG.getConstantFP(maxnum(C0, C1), N->getValueType(0));
7955 EVT VT = N->getValueType(0);
7956 // Canonicalize to constant on RHS.
7957 return DAG.getNode(ISD::FMAXNUM, SDLoc(N), VT, N1, N0);
7963 SDValue DAGCombiner::visitFABS(SDNode *N) {
7964 SDValue N0 = N->getOperand(0);
7965 EVT VT = N->getValueType(0);
7967 if (VT.isVector()) {
7968 SDValue FoldedVOp = SimplifyVUnaryOp(N);
7969 if (FoldedVOp.getNode()) return FoldedVOp;
7972 // fold (fabs c1) -> fabs(c1)
7973 if (isa<ConstantFPSDNode>(N0))
7974 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7976 // fold (fabs (fabs x)) -> (fabs x)
7977 if (N0.getOpcode() == ISD::FABS)
7978 return N->getOperand(0);
7980 // fold (fabs (fneg x)) -> (fabs x)
7981 // fold (fabs (fcopysign x, y)) -> (fabs x)
7982 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
7983 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0));
7985 // Transform fabs(bitconvert(x)) -> bitconvert(x & ~sign) to avoid loading
7986 // constant pool values.
7987 if (!TLI.isFAbsFree(VT) &&
7988 N0.getOpcode() == ISD::BITCAST &&
7989 N0.getNode()->hasOneUse()) {
7990 SDValue Int = N0.getOperand(0);
7991 EVT IntVT = Int.getValueType();
7992 if (IntVT.isInteger() && !IntVT.isVector()) {
7994 if (N0.getValueType().isVector()) {
7995 // For a vector, get a mask such as 0x7f... per scalar element
7997 SignMask = ~APInt::getSignBit(N0.getValueType().getScalarSizeInBits());
7998 SignMask = APInt::getSplat(IntVT.getSizeInBits(), SignMask);
8000 // For a scalar, just generate 0x7f...
8001 SignMask = ~APInt::getSignBit(IntVT.getSizeInBits());
8003 Int = DAG.getNode(ISD::AND, SDLoc(N0), IntVT, Int,
8004 DAG.getConstant(SignMask, IntVT));
8005 AddToWorklist(Int.getNode());
8006 return DAG.getNode(ISD::BITCAST, SDLoc(N), N->getValueType(0), Int);
8013 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
8014 SDValue Chain = N->getOperand(0);
8015 SDValue N1 = N->getOperand(1);
8016 SDValue N2 = N->getOperand(2);
8018 // If N is a constant we could fold this into a fallthrough or unconditional
8019 // branch. However that doesn't happen very often in normal code, because
8020 // Instcombine/SimplifyCFG should have handled the available opportunities.
8021 // If we did this folding here, it would be necessary to update the
8022 // MachineBasicBlock CFG, which is awkward.
8024 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
8026 if (N1.getOpcode() == ISD::SETCC &&
8027 TLI.isOperationLegalOrCustom(ISD::BR_CC,
8028 N1.getOperand(0).getValueType())) {
8029 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
8030 Chain, N1.getOperand(2),
8031 N1.getOperand(0), N1.getOperand(1), N2);
8034 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
8035 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
8036 (N1.getOperand(0).hasOneUse() &&
8037 N1.getOperand(0).getOpcode() == ISD::SRL))) {
8038 SDNode *Trunc = nullptr;
8039 if (N1.getOpcode() == ISD::TRUNCATE) {
8040 // Look pass the truncate.
8041 Trunc = N1.getNode();
8042 N1 = N1.getOperand(0);
8045 // Match this pattern so that we can generate simpler code:
8048 // %b = and i32 %a, 2
8049 // %c = srl i32 %b, 1
8050 // brcond i32 %c ...
8055 // %b = and i32 %a, 2
8056 // %c = setcc eq %b, 0
8059 // This applies only when the AND constant value has one bit set and the
8060 // SRL constant is equal to the log2 of the AND constant. The back-end is
8061 // smart enough to convert the result into a TEST/JMP sequence.
8062 SDValue Op0 = N1.getOperand(0);
8063 SDValue Op1 = N1.getOperand(1);
8065 if (Op0.getOpcode() == ISD::AND &&
8066 Op1.getOpcode() == ISD::Constant) {
8067 SDValue AndOp1 = Op0.getOperand(1);
8069 if (AndOp1.getOpcode() == ISD::Constant) {
8070 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
8072 if (AndConst.isPowerOf2() &&
8073 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
8075 DAG.getSetCC(SDLoc(N),
8076 getSetCCResultType(Op0.getValueType()),
8077 Op0, DAG.getConstant(0, Op0.getValueType()),
8080 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, SDLoc(N),
8081 MVT::Other, Chain, SetCC, N2);
8082 // Don't add the new BRCond into the worklist or else SimplifySelectCC
8083 // will convert it back to (X & C1) >> C2.
8084 CombineTo(N, NewBRCond, false);
8085 // Truncate is dead.
8087 deleteAndRecombine(Trunc);
8088 // Replace the uses of SRL with SETCC
8089 WorklistRemover DeadNodes(*this);
8090 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
8091 deleteAndRecombine(N1.getNode());
8092 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8098 // Restore N1 if the above transformation doesn't match.
8099 N1 = N->getOperand(1);
8102 // Transform br(xor(x, y)) -> br(x != y)
8103 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
8104 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
8105 SDNode *TheXor = N1.getNode();
8106 SDValue Op0 = TheXor->getOperand(0);
8107 SDValue Op1 = TheXor->getOperand(1);
8108 if (Op0.getOpcode() == Op1.getOpcode()) {
8109 // Avoid missing important xor optimizations.
8110 SDValue Tmp = visitXOR(TheXor);
8111 if (Tmp.getNode()) {
8112 if (Tmp.getNode() != TheXor) {
8113 DEBUG(dbgs() << "\nReplacing.8 ";
8115 dbgs() << "\nWith: ";
8116 Tmp.getNode()->dump(&DAG);
8118 WorklistRemover DeadNodes(*this);
8119 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
8120 deleteAndRecombine(TheXor);
8121 return DAG.getNode(ISD::BRCOND, SDLoc(N),
8122 MVT::Other, Chain, Tmp, N2);
8125 // visitXOR has changed XOR's operands or replaced the XOR completely,
8127 return SDValue(N, 0);
8131 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
8133 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
8134 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
8135 Op0.getOpcode() == ISD::XOR) {
8136 TheXor = Op0.getNode();
8140 EVT SetCCVT = N1.getValueType();
8142 SetCCVT = getSetCCResultType(SetCCVT);
8143 SDValue SetCC = DAG.getSetCC(SDLoc(TheXor),
8146 Equal ? ISD::SETEQ : ISD::SETNE);
8147 // Replace the uses of XOR with SETCC
8148 WorklistRemover DeadNodes(*this);
8149 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
8150 deleteAndRecombine(N1.getNode());
8151 return DAG.getNode(ISD::BRCOND, SDLoc(N),
8152 MVT::Other, Chain, SetCC, N2);
8159 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
8161 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
8162 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
8163 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
8165 // If N is a constant we could fold this into a fallthrough or unconditional
8166 // branch. However that doesn't happen very often in normal code, because
8167 // Instcombine/SimplifyCFG should have handled the available opportunities.
8168 // If we did this folding here, it would be necessary to update the
8169 // MachineBasicBlock CFG, which is awkward.
8171 // Use SimplifySetCC to simplify SETCC's.
8172 SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()),
8173 CondLHS, CondRHS, CC->get(), SDLoc(N),
8175 if (Simp.getNode()) AddToWorklist(Simp.getNode());
8177 // fold to a simpler setcc
8178 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
8179 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
8180 N->getOperand(0), Simp.getOperand(2),
8181 Simp.getOperand(0), Simp.getOperand(1),
8187 /// Return true if 'Use' is a load or a store that uses N as its base pointer
8188 /// and that N may be folded in the load / store addressing mode.
8189 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
8191 const TargetLowering &TLI) {
8193 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
8194 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
8196 VT = Use->getValueType(0);
8197 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
8198 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
8200 VT = ST->getValue().getValueType();
8204 TargetLowering::AddrMode AM;
8205 if (N->getOpcode() == ISD::ADD) {
8206 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
8209 AM.BaseOffs = Offset->getSExtValue();
8213 } else if (N->getOpcode() == ISD::SUB) {
8214 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
8217 AM.BaseOffs = -Offset->getSExtValue();
8224 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
8227 /// Try turning a load/store into a pre-indexed load/store when the base
8228 /// pointer is an add or subtract and it has other uses besides the load/store.
8229 /// After the transformation, the new indexed load/store has effectively folded
8230 /// the add/subtract in and all of its other uses are redirected to the
8232 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
8233 if (Level < AfterLegalizeDAG)
8239 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8240 if (LD->isIndexed())
8242 VT = LD->getMemoryVT();
8243 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
8244 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
8246 Ptr = LD->getBasePtr();
8247 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8248 if (ST->isIndexed())
8250 VT = ST->getMemoryVT();
8251 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
8252 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
8254 Ptr = ST->getBasePtr();
8260 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
8261 // out. There is no reason to make this a preinc/predec.
8262 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
8263 Ptr.getNode()->hasOneUse())
8266 // Ask the target to do addressing mode selection.
8269 ISD::MemIndexedMode AM = ISD::UNINDEXED;
8270 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
8273 // Backends without true r+i pre-indexed forms may need to pass a
8274 // constant base with a variable offset so that constant coercion
8275 // will work with the patterns in canonical form.
8276 bool Swapped = false;
8277 if (isa<ConstantSDNode>(BasePtr)) {
8278 std::swap(BasePtr, Offset);
8282 // Don't create a indexed load / store with zero offset.
8283 if (isa<ConstantSDNode>(Offset) &&
8284 cast<ConstantSDNode>(Offset)->isNullValue())
8287 // Try turning it into a pre-indexed load / store except when:
8288 // 1) The new base ptr is a frame index.
8289 // 2) If N is a store and the new base ptr is either the same as or is a
8290 // predecessor of the value being stored.
8291 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
8292 // that would create a cycle.
8293 // 4) All uses are load / store ops that use it as old base ptr.
8295 // Check #1. Preinc'ing a frame index would require copying the stack pointer
8296 // (plus the implicit offset) to a register to preinc anyway.
8297 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
8302 SDValue Val = cast<StoreSDNode>(N)->getValue();
8303 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
8307 // If the offset is a constant, there may be other adds of constants that
8308 // can be folded with this one. We should do this to avoid having to keep
8309 // a copy of the original base pointer.
8310 SmallVector<SDNode *, 16> OtherUses;
8311 if (isa<ConstantSDNode>(Offset))
8312 for (SDNode *Use : BasePtr.getNode()->uses()) {
8313 if (Use == Ptr.getNode())
8316 if (Use->isPredecessorOf(N))
8319 if (Use->getOpcode() != ISD::ADD && Use->getOpcode() != ISD::SUB) {
8324 SDValue Op0 = Use->getOperand(0), Op1 = Use->getOperand(1);
8325 if (Op1.getNode() == BasePtr.getNode())
8326 std::swap(Op0, Op1);
8327 assert(Op0.getNode() == BasePtr.getNode() &&
8328 "Use of ADD/SUB but not an operand");
8330 if (!isa<ConstantSDNode>(Op1)) {
8335 // FIXME: In some cases, we can be smarter about this.
8336 if (Op1.getValueType() != Offset.getValueType()) {
8341 OtherUses.push_back(Use);
8345 std::swap(BasePtr, Offset);
8347 // Now check for #3 and #4.
8348 bool RealUse = false;
8350 // Caches for hasPredecessorHelper
8351 SmallPtrSet<const SDNode *, 32> Visited;
8352 SmallVector<const SDNode *, 16> Worklist;
8354 for (SDNode *Use : Ptr.getNode()->uses()) {
8357 if (N->hasPredecessorHelper(Use, Visited, Worklist))
8360 // If Ptr may be folded in addressing mode of other use, then it's
8361 // not profitable to do this transformation.
8362 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
8371 Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
8372 BasePtr, Offset, AM);
8374 Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
8375 BasePtr, Offset, AM);
8378 DEBUG(dbgs() << "\nReplacing.4 ";
8380 dbgs() << "\nWith: ";
8381 Result.getNode()->dump(&DAG);
8383 WorklistRemover DeadNodes(*this);
8385 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
8386 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
8388 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
8391 // Finally, since the node is now dead, remove it from the graph.
8392 deleteAndRecombine(N);
8395 std::swap(BasePtr, Offset);
8397 // Replace other uses of BasePtr that can be updated to use Ptr
8398 for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) {
8399 unsigned OffsetIdx = 1;
8400 if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode())
8402 assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() ==
8403 BasePtr.getNode() && "Expected BasePtr operand");
8405 // We need to replace ptr0 in the following expression:
8406 // x0 * offset0 + y0 * ptr0 = t0
8408 // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store)
8410 // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the
8411 // indexed load/store and the expresion that needs to be re-written.
8413 // Therefore, we have:
8414 // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1
8416 ConstantSDNode *CN =
8417 cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx));
8419 APInt Offset0 = CN->getAPIntValue();
8420 APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue();
8422 X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1;
8423 Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1;
8424 X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1;
8425 Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1;
8427 unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD;
8429 APInt CNV = Offset0;
8430 if (X0 < 0) CNV = -CNV;
8431 if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1;
8432 else CNV = CNV - Offset1;
8434 // We can now generate the new expression.
8435 SDValue NewOp1 = DAG.getConstant(CNV, CN->getValueType(0));
8436 SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0);
8438 SDValue NewUse = DAG.getNode(Opcode,
8439 SDLoc(OtherUses[i]),
8440 OtherUses[i]->getValueType(0), NewOp1, NewOp2);
8441 DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse);
8442 deleteAndRecombine(OtherUses[i]);
8445 // Replace the uses of Ptr with uses of the updated base value.
8446 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
8447 deleteAndRecombine(Ptr.getNode());
8452 /// Try to combine a load/store with a add/sub of the base pointer node into a
8453 /// post-indexed load/store. The transformation folded the add/subtract into the
8454 /// new indexed load/store effectively and all of its uses are redirected to the
8456 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
8457 if (Level < AfterLegalizeDAG)
8463 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
8464 if (LD->isIndexed())
8466 VT = LD->getMemoryVT();
8467 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
8468 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
8470 Ptr = LD->getBasePtr();
8471 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
8472 if (ST->isIndexed())
8474 VT = ST->getMemoryVT();
8475 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
8476 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
8478 Ptr = ST->getBasePtr();
8484 if (Ptr.getNode()->hasOneUse())
8487 for (SDNode *Op : Ptr.getNode()->uses()) {
8489 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
8494 ISD::MemIndexedMode AM = ISD::UNINDEXED;
8495 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
8496 // Don't create a indexed load / store with zero offset.
8497 if (isa<ConstantSDNode>(Offset) &&
8498 cast<ConstantSDNode>(Offset)->isNullValue())
8501 // Try turning it into a post-indexed load / store except when
8502 // 1) All uses are load / store ops that use it as base ptr (and
8503 // it may be folded as addressing mmode).
8504 // 2) Op must be independent of N, i.e. Op is neither a predecessor
8505 // nor a successor of N. Otherwise, if Op is folded that would
8508 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
8512 bool TryNext = false;
8513 for (SDNode *Use : BasePtr.getNode()->uses()) {
8514 if (Use == Ptr.getNode())
8517 // If all the uses are load / store addresses, then don't do the
8519 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
8520 bool RealUse = false;
8521 for (SDNode *UseUse : Use->uses()) {
8522 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
8537 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
8538 SDValue Result = isLoad
8539 ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
8540 BasePtr, Offset, AM)
8541 : DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
8542 BasePtr, Offset, AM);
8545 DEBUG(dbgs() << "\nReplacing.5 ";
8547 dbgs() << "\nWith: ";
8548 Result.getNode()->dump(&DAG);
8550 WorklistRemover DeadNodes(*this);
8552 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
8553 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
8555 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
8558 // Finally, since the node is now dead, remove it from the graph.
8559 deleteAndRecombine(N);
8561 // Replace the uses of Use with uses of the updated base value.
8562 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
8563 Result.getValue(isLoad ? 1 : 0));
8564 deleteAndRecombine(Op);
8573 /// \brief Return the base-pointer arithmetic from an indexed \p LD.
8574 SDValue DAGCombiner::SplitIndexingFromLoad(LoadSDNode *LD) {
8575 ISD::MemIndexedMode AM = LD->getAddressingMode();
8576 assert(AM != ISD::UNINDEXED);
8577 SDValue BP = LD->getOperand(1);
8578 SDValue Inc = LD->getOperand(2);
8580 // Some backends use TargetConstants for load offsets, but don't expect
8581 // TargetConstants in general ADD nodes. We can convert these constants into
8582 // regular Constants (if the constant is not opaque).
8583 assert((Inc.getOpcode() != ISD::TargetConstant ||
8584 !cast<ConstantSDNode>(Inc)->isOpaque()) &&
8585 "Cannot split out indexing using opaque target constants");
8586 if (Inc.getOpcode() == ISD::TargetConstant) {
8587 ConstantSDNode *ConstInc = cast<ConstantSDNode>(Inc);
8588 Inc = DAG.getConstant(*ConstInc->getConstantIntValue(),
8589 ConstInc->getValueType(0));
8593 (AM == ISD::PRE_INC || AM == ISD::POST_INC ? ISD::ADD : ISD::SUB);
8594 return DAG.getNode(Opc, SDLoc(LD), BP.getSimpleValueType(), BP, Inc);
8597 SDValue DAGCombiner::visitLOAD(SDNode *N) {
8598 LoadSDNode *LD = cast<LoadSDNode>(N);
8599 SDValue Chain = LD->getChain();
8600 SDValue Ptr = LD->getBasePtr();
8602 // If load is not volatile and there are no uses of the loaded value (and
8603 // the updated indexed value in case of indexed loads), change uses of the
8604 // chain value into uses of the chain input (i.e. delete the dead load).
8605 if (!LD->isVolatile()) {
8606 if (N->getValueType(1) == MVT::Other) {
8608 if (!N->hasAnyUseOfValue(0)) {
8609 // It's not safe to use the two value CombineTo variant here. e.g.
8610 // v1, chain2 = load chain1, loc
8611 // v2, chain3 = load chain2, loc
8613 // Now we replace use of chain2 with chain1. This makes the second load
8614 // isomorphic to the one we are deleting, and thus makes this load live.
8615 DEBUG(dbgs() << "\nReplacing.6 ";
8617 dbgs() << "\nWith chain: ";
8618 Chain.getNode()->dump(&DAG);
8620 WorklistRemover DeadNodes(*this);
8621 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
8624 deleteAndRecombine(N);
8626 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8630 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
8632 // If this load has an opaque TargetConstant offset, then we cannot split
8633 // the indexing into an add/sub directly (that TargetConstant may not be
8634 // valid for a different type of node, and we cannot convert an opaque
8635 // target constant into a regular constant).
8636 bool HasOTCInc = LD->getOperand(2).getOpcode() == ISD::TargetConstant &&
8637 cast<ConstantSDNode>(LD->getOperand(2))->isOpaque();
8639 if (!N->hasAnyUseOfValue(0) &&
8640 ((MaySplitLoadIndex && !HasOTCInc) || !N->hasAnyUseOfValue(1))) {
8641 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
8643 if (N->hasAnyUseOfValue(1) && MaySplitLoadIndex && !HasOTCInc) {
8644 Index = SplitIndexingFromLoad(LD);
8645 // Try to fold the base pointer arithmetic into subsequent loads and
8647 AddUsersToWorklist(N);
8649 Index = DAG.getUNDEF(N->getValueType(1));
8650 DEBUG(dbgs() << "\nReplacing.7 ";
8652 dbgs() << "\nWith: ";
8653 Undef.getNode()->dump(&DAG);
8654 dbgs() << " and 2 other values\n");
8655 WorklistRemover DeadNodes(*this);
8656 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
8657 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Index);
8658 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
8659 deleteAndRecombine(N);
8660 return SDValue(N, 0); // Return N so it doesn't get rechecked!
8665 // If this load is directly stored, replace the load value with the stored
8667 // TODO: Handle store large -> read small portion.
8668 // TODO: Handle TRUNCSTORE/LOADEXT
8669 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
8670 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
8671 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
8672 if (PrevST->getBasePtr() == Ptr &&
8673 PrevST->getValue().getValueType() == N->getValueType(0))
8674 return CombineTo(N, Chain.getOperand(1), Chain);
8678 // Try to infer better alignment information than the load already has.
8679 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
8680 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
8681 if (Align > LD->getMemOperand()->getBaseAlignment()) {
8683 DAG.getExtLoad(LD->getExtensionType(), SDLoc(N),
8684 LD->getValueType(0),
8685 Chain, Ptr, LD->getPointerInfo(),
8687 LD->isVolatile(), LD->isNonTemporal(),
8688 LD->isInvariant(), Align, LD->getAAInfo());
8689 return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true);
8694 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
8695 : DAG.getSubtarget().useAA();
8697 if (CombinerAAOnlyFunc.getNumOccurrences() &&
8698 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
8701 if (UseAA && LD->isUnindexed()) {
8702 // Walk up chain skipping non-aliasing memory nodes.
8703 SDValue BetterChain = FindBetterChain(N, Chain);
8705 // If there is a better chain.
8706 if (Chain != BetterChain) {
8709 // Replace the chain to void dependency.
8710 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
8711 ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD),
8712 BetterChain, Ptr, LD->getMemOperand());
8714 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD),
8715 LD->getValueType(0),
8716 BetterChain, Ptr, LD->getMemoryVT(),
8717 LD->getMemOperand());
8720 // Create token factor to keep old chain connected.
8721 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
8722 MVT::Other, Chain, ReplLoad.getValue(1));
8724 // Make sure the new and old chains are cleaned up.
8725 AddToWorklist(Token.getNode());
8727 // Replace uses with load result and token factor. Don't add users
8729 return CombineTo(N, ReplLoad.getValue(0), Token, false);
8733 // Try transforming N to an indexed load.
8734 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
8735 return SDValue(N, 0);
8737 // Try to slice up N to more direct loads if the slices are mapped to
8738 // different register banks or pairing can take place.
8740 return SDValue(N, 0);
8746 /// \brief Helper structure used to slice a load in smaller loads.
8747 /// Basically a slice is obtained from the following sequence:
8748 /// Origin = load Ty1, Base
8749 /// Shift = srl Ty1 Origin, CstTy Amount
8750 /// Inst = trunc Shift to Ty2
8752 /// Then, it will be rewriten into:
8753 /// Slice = load SliceTy, Base + SliceOffset
8754 /// [Inst = zext Slice to Ty2], only if SliceTy <> Ty2
8756 /// SliceTy is deduced from the number of bits that are actually used to
8758 struct LoadedSlice {
8759 /// \brief Helper structure used to compute the cost of a slice.
8761 /// Are we optimizing for code size.
8766 unsigned CrossRegisterBanksCopies;
8770 Cost(bool ForCodeSize = false)
8771 : ForCodeSize(ForCodeSize), Loads(0), Truncates(0),
8772 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {}
8774 /// \brief Get the cost of one isolated slice.
8775 Cost(const LoadedSlice &LS, bool ForCodeSize = false)
8776 : ForCodeSize(ForCodeSize), Loads(1), Truncates(0),
8777 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {
8778 EVT TruncType = LS.Inst->getValueType(0);
8779 EVT LoadedType = LS.getLoadedType();
8780 if (TruncType != LoadedType &&
8781 !LS.DAG->getTargetLoweringInfo().isZExtFree(LoadedType, TruncType))
8785 /// \brief Account for slicing gain in the current cost.
8786 /// Slicing provide a few gains like removing a shift or a
8787 /// truncate. This method allows to grow the cost of the original
8788 /// load with the gain from this slice.
8789 void addSliceGain(const LoadedSlice &LS) {
8790 // Each slice saves a truncate.
8791 const TargetLowering &TLI = LS.DAG->getTargetLoweringInfo();
8792 if (!TLI.isTruncateFree(LS.Inst->getValueType(0),
8793 LS.Inst->getOperand(0).getValueType()))
8795 // If there is a shift amount, this slice gets rid of it.
8798 // If this slice can merge a cross register bank copy, account for it.
8799 if (LS.canMergeExpensiveCrossRegisterBankCopy())
8800 ++CrossRegisterBanksCopies;
8803 Cost &operator+=(const Cost &RHS) {
8805 Truncates += RHS.Truncates;
8806 CrossRegisterBanksCopies += RHS.CrossRegisterBanksCopies;
8812 bool operator==(const Cost &RHS) const {
8813 return Loads == RHS.Loads && Truncates == RHS.Truncates &&
8814 CrossRegisterBanksCopies == RHS.CrossRegisterBanksCopies &&
8815 ZExts == RHS.ZExts && Shift == RHS.Shift;
8818 bool operator!=(const Cost &RHS) const { return !(*this == RHS); }
8820 bool operator<(const Cost &RHS) const {
8821 // Assume cross register banks copies are as expensive as loads.
8822 // FIXME: Do we want some more target hooks?
8823 unsigned ExpensiveOpsLHS = Loads + CrossRegisterBanksCopies;
8824 unsigned ExpensiveOpsRHS = RHS.Loads + RHS.CrossRegisterBanksCopies;
8825 // Unless we are optimizing for code size, consider the
8826 // expensive operation first.
8827 if (!ForCodeSize && ExpensiveOpsLHS != ExpensiveOpsRHS)
8828 return ExpensiveOpsLHS < ExpensiveOpsRHS;
8829 return (Truncates + ZExts + Shift + ExpensiveOpsLHS) <
8830 (RHS.Truncates + RHS.ZExts + RHS.Shift + ExpensiveOpsRHS);
8833 bool operator>(const Cost &RHS) const { return RHS < *this; }
8835 bool operator<=(const Cost &RHS) const { return !(RHS < *this); }
8837 bool operator>=(const Cost &RHS) const { return !(*this < RHS); }
8839 // The last instruction that represent the slice. This should be a
8840 // truncate instruction.
8842 // The original load instruction.
8844 // The right shift amount in bits from the original load.
8846 // The DAG from which Origin came from.
8847 // This is used to get some contextual information about legal types, etc.
8850 LoadedSlice(SDNode *Inst = nullptr, LoadSDNode *Origin = nullptr,
8851 unsigned Shift = 0, SelectionDAG *DAG = nullptr)
8852 : Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {}
8854 LoadedSlice(const LoadedSlice &LS)
8855 : Inst(LS.Inst), Origin(LS.Origin), Shift(LS.Shift), DAG(LS.DAG) {}
8857 /// \brief Get the bits used in a chunk of bits \p BitWidth large.
8858 /// \return Result is \p BitWidth and has used bits set to 1 and
8859 /// not used bits set to 0.
8860 APInt getUsedBits() const {
8861 // Reproduce the trunc(lshr) sequence:
8862 // - Start from the truncated value.
8863 // - Zero extend to the desired bit width.
8865 assert(Origin && "No original load to compare against.");
8866 unsigned BitWidth = Origin->getValueSizeInBits(0);
8867 assert(Inst && "This slice is not bound to an instruction");
8868 assert(Inst->getValueSizeInBits(0) <= BitWidth &&
8869 "Extracted slice is bigger than the whole type!");
8870 APInt UsedBits(Inst->getValueSizeInBits(0), 0);
8871 UsedBits.setAllBits();
8872 UsedBits = UsedBits.zext(BitWidth);
8877 /// \brief Get the size of the slice to be loaded in bytes.
8878 unsigned getLoadedSize() const {
8879 unsigned SliceSize = getUsedBits().countPopulation();
8880 assert(!(SliceSize & 0x7) && "Size is not a multiple of a byte.");
8881 return SliceSize / 8;
8884 /// \brief Get the type that will be loaded for this slice.
8885 /// Note: This may not be the final type for the slice.
8886 EVT getLoadedType() const {
8887 assert(DAG && "Missing context");
8888 LLVMContext &Ctxt = *DAG->getContext();
8889 return EVT::getIntegerVT(Ctxt, getLoadedSize() * 8);
8892 /// \brief Get the alignment of the load used for this slice.
8893 unsigned getAlignment() const {
8894 unsigned Alignment = Origin->getAlignment();
8895 unsigned Offset = getOffsetFromBase();
8897 Alignment = MinAlign(Alignment, Alignment + Offset);
8901 /// \brief Check if this slice can be rewritten with legal operations.
8902 bool isLegal() const {
8903 // An invalid slice is not legal.
8904 if (!Origin || !Inst || !DAG)
8907 // Offsets are for indexed load only, we do not handle that.
8908 if (Origin->getOffset().getOpcode() != ISD::UNDEF)
8911 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
8913 // Check that the type is legal.
8914 EVT SliceType = getLoadedType();
8915 if (!TLI.isTypeLegal(SliceType))
8918 // Check that the load is legal for this type.
8919 if (!TLI.isOperationLegal(ISD::LOAD, SliceType))
8922 // Check that the offset can be computed.
8923 // 1. Check its type.
8924 EVT PtrType = Origin->getBasePtr().getValueType();
8925 if (PtrType == MVT::Untyped || PtrType.isExtended())
8928 // 2. Check that it fits in the immediate.
8929 if (!TLI.isLegalAddImmediate(getOffsetFromBase()))
8932 // 3. Check that the computation is legal.
8933 if (!TLI.isOperationLegal(ISD::ADD, PtrType))
8936 // Check that the zext is legal if it needs one.
8937 EVT TruncateType = Inst->getValueType(0);
8938 if (TruncateType != SliceType &&
8939 !TLI.isOperationLegal(ISD::ZERO_EXTEND, TruncateType))
8945 /// \brief Get the offset in bytes of this slice in the original chunk of
8947 /// \pre DAG != nullptr.
8948 uint64_t getOffsetFromBase() const {
8949 assert(DAG && "Missing context.");
8951 DAG->getTargetLoweringInfo().getDataLayout()->isBigEndian();
8952 assert(!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported.");
8953 uint64_t Offset = Shift / 8;
8954 unsigned TySizeInBytes = Origin->getValueSizeInBits(0) / 8;
8955 assert(!(Origin->getValueSizeInBits(0) & 0x7) &&
8956 "The size of the original loaded type is not a multiple of a"
8958 // If Offset is bigger than TySizeInBytes, it means we are loading all
8959 // zeros. This should have been optimized before in the process.
8960 assert(TySizeInBytes > Offset &&
8961 "Invalid shift amount for given loaded size");
8963 Offset = TySizeInBytes - Offset - getLoadedSize();
8967 /// \brief Generate the sequence of instructions to load the slice
8968 /// represented by this object and redirect the uses of this slice to
8969 /// this new sequence of instructions.
8970 /// \pre this->Inst && this->Origin are valid Instructions and this
8971 /// object passed the legal check: LoadedSlice::isLegal returned true.
8972 /// \return The last instruction of the sequence used to load the slice.
8973 SDValue loadSlice() const {
8974 assert(Inst && Origin && "Unable to replace a non-existing slice.");
8975 const SDValue &OldBaseAddr = Origin->getBasePtr();
8976 SDValue BaseAddr = OldBaseAddr;
8977 // Get the offset in that chunk of bytes w.r.t. the endianess.
8978 int64_t Offset = static_cast<int64_t>(getOffsetFromBase());
8979 assert(Offset >= 0 && "Offset too big to fit in int64_t!");
8981 // BaseAddr = BaseAddr + Offset.
8982 EVT ArithType = BaseAddr.getValueType();
8983 BaseAddr = DAG->getNode(ISD::ADD, SDLoc(Origin), ArithType, BaseAddr,
8984 DAG->getConstant(Offset, ArithType));
8987 // Create the type of the loaded slice according to its size.
8988 EVT SliceType = getLoadedType();
8990 // Create the load for the slice.
8991 SDValue LastInst = DAG->getLoad(
8992 SliceType, SDLoc(Origin), Origin->getChain(), BaseAddr,
8993 Origin->getPointerInfo().getWithOffset(Offset), Origin->isVolatile(),
8994 Origin->isNonTemporal(), Origin->isInvariant(), getAlignment());
8995 // If the final type is not the same as the loaded type, this means that
8996 // we have to pad with zero. Create a zero extend for that.
8997 EVT FinalType = Inst->getValueType(0);
8998 if (SliceType != FinalType)
9000 DAG->getNode(ISD::ZERO_EXTEND, SDLoc(LastInst), FinalType, LastInst);
9004 /// \brief Check if this slice can be merged with an expensive cross register
9005 /// bank copy. E.g.,
9007 /// f = bitcast i32 i to float
9008 bool canMergeExpensiveCrossRegisterBankCopy() const {
9009 if (!Inst || !Inst->hasOneUse())
9011 SDNode *Use = *Inst->use_begin();
9012 if (Use->getOpcode() != ISD::BITCAST)
9014 assert(DAG && "Missing context");
9015 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
9016 EVT ResVT = Use->getValueType(0);
9017 const TargetRegisterClass *ResRC = TLI.getRegClassFor(ResVT.getSimpleVT());
9018 const TargetRegisterClass *ArgRC =
9019 TLI.getRegClassFor(Use->getOperand(0).getValueType().getSimpleVT());
9020 if (ArgRC == ResRC || !TLI.isOperationLegal(ISD::LOAD, ResVT))
9023 // At this point, we know that we perform a cross-register-bank copy.
9024 // Check if it is expensive.
9025 const TargetRegisterInfo *TRI = DAG->getSubtarget().getRegisterInfo();
9026 // Assume bitcasts are cheap, unless both register classes do not
9027 // explicitly share a common sub class.
9028 if (!TRI || TRI->getCommonSubClass(ArgRC, ResRC))
9031 // Check if it will be merged with the load.
9032 // 1. Check the alignment constraint.
9033 unsigned RequiredAlignment = TLI.getDataLayout()->getABITypeAlignment(
9034 ResVT.getTypeForEVT(*DAG->getContext()));
9036 if (RequiredAlignment > getAlignment())
9039 // 2. Check that the load is a legal operation for that type.
9040 if (!TLI.isOperationLegal(ISD::LOAD, ResVT))
9043 // 3. Check that we do not have a zext in the way.
9044 if (Inst->getValueType(0) != getLoadedType())
9052 /// \brief Check that all bits set in \p UsedBits form a dense region, i.e.,
9053 /// \p UsedBits looks like 0..0 1..1 0..0.
9054 static bool areUsedBitsDense(const APInt &UsedBits) {
9055 // If all the bits are one, this is dense!
9056 if (UsedBits.isAllOnesValue())
9059 // Get rid of the unused bits on the right.
9060 APInt NarrowedUsedBits = UsedBits.lshr(UsedBits.countTrailingZeros());
9061 // Get rid of the unused bits on the left.
9062 if (NarrowedUsedBits.countLeadingZeros())
9063 NarrowedUsedBits = NarrowedUsedBits.trunc(NarrowedUsedBits.getActiveBits());
9064 // Check that the chunk of bits is completely used.
9065 return NarrowedUsedBits.isAllOnesValue();
9068 /// \brief Check whether or not \p First and \p Second are next to each other
9069 /// in memory. This means that there is no hole between the bits loaded
9070 /// by \p First and the bits loaded by \p Second.
9071 static bool areSlicesNextToEachOther(const LoadedSlice &First,
9072 const LoadedSlice &Second) {
9073 assert(First.Origin == Second.Origin && First.Origin &&
9074 "Unable to match different memory origins.");
9075 APInt UsedBits = First.getUsedBits();
9076 assert((UsedBits & Second.getUsedBits()) == 0 &&
9077 "Slices are not supposed to overlap.");
9078 UsedBits |= Second.getUsedBits();
9079 return areUsedBitsDense(UsedBits);
9082 /// \brief Adjust the \p GlobalLSCost according to the target
9083 /// paring capabilities and the layout of the slices.
9084 /// \pre \p GlobalLSCost should account for at least as many loads as
9085 /// there is in the slices in \p LoadedSlices.
9086 static void adjustCostForPairing(SmallVectorImpl<LoadedSlice> &LoadedSlices,
9087 LoadedSlice::Cost &GlobalLSCost) {
9088 unsigned NumberOfSlices = LoadedSlices.size();
9089 // If there is less than 2 elements, no pairing is possible.
9090 if (NumberOfSlices < 2)
9093 // Sort the slices so that elements that are likely to be next to each
9094 // other in memory are next to each other in the list.
9095 std::sort(LoadedSlices.begin(), LoadedSlices.end(),
9096 [](const LoadedSlice &LHS, const LoadedSlice &RHS) {
9097 assert(LHS.Origin == RHS.Origin && "Different bases not implemented.");
9098 return LHS.getOffsetFromBase() < RHS.getOffsetFromBase();
9100 const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo();
9101 // First (resp. Second) is the first (resp. Second) potentially candidate
9102 // to be placed in a paired load.
9103 const LoadedSlice *First = nullptr;
9104 const LoadedSlice *Second = nullptr;
9105 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice,
9106 // Set the beginning of the pair.
9109 Second = &LoadedSlices[CurrSlice];
9111 // If First is NULL, it means we start a new pair.
9112 // Get to the next slice.
9116 EVT LoadedType = First->getLoadedType();
9118 // If the types of the slices are different, we cannot pair them.
9119 if (LoadedType != Second->getLoadedType())
9122 // Check if the target supplies paired loads for this type.
9123 unsigned RequiredAlignment = 0;
9124 if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) {
9125 // move to the next pair, this type is hopeless.
9129 // Check if we meet the alignment requirement.
9130 if (RequiredAlignment > First->getAlignment())
9133 // Check that both loads are next to each other in memory.
9134 if (!areSlicesNextToEachOther(*First, *Second))
9137 assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!");
9138 --GlobalLSCost.Loads;
9139 // Move to the next pair.
9144 /// \brief Check the profitability of all involved LoadedSlice.
9145 /// Currently, it is considered profitable if there is exactly two
9146 /// involved slices (1) which are (2) next to each other in memory, and
9147 /// whose cost (\see LoadedSlice::Cost) is smaller than the original load (3).
9149 /// Note: The order of the elements in \p LoadedSlices may be modified, but not
9150 /// the elements themselves.
9152 /// FIXME: When the cost model will be mature enough, we can relax
9153 /// constraints (1) and (2).
9154 static bool isSlicingProfitable(SmallVectorImpl<LoadedSlice> &LoadedSlices,
9155 const APInt &UsedBits, bool ForCodeSize) {
9156 unsigned NumberOfSlices = LoadedSlices.size();
9157 if (StressLoadSlicing)
9158 return NumberOfSlices > 1;
9161 if (NumberOfSlices != 2)
9165 if (!areUsedBitsDense(UsedBits))
9169 LoadedSlice::Cost OrigCost(ForCodeSize), GlobalSlicingCost(ForCodeSize);
9170 // The original code has one big load.
9172 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice) {
9173 const LoadedSlice &LS = LoadedSlices[CurrSlice];
9174 // Accumulate the cost of all the slices.
9175 LoadedSlice::Cost SliceCost(LS, ForCodeSize);
9176 GlobalSlicingCost += SliceCost;
9178 // Account as cost in the original configuration the gain obtained
9179 // with the current slices.
9180 OrigCost.addSliceGain(LS);
9183 // If the target supports paired load, adjust the cost accordingly.
9184 adjustCostForPairing(LoadedSlices, GlobalSlicingCost);
9185 return OrigCost > GlobalSlicingCost;
9188 /// \brief If the given load, \p LI, is used only by trunc or trunc(lshr)
9189 /// operations, split it in the various pieces being extracted.
9191 /// This sort of thing is introduced by SROA.
9192 /// This slicing takes care not to insert overlapping loads.
9193 /// \pre LI is a simple load (i.e., not an atomic or volatile load).
9194 bool DAGCombiner::SliceUpLoad(SDNode *N) {
9195 if (Level < AfterLegalizeDAG)
9198 LoadSDNode *LD = cast<LoadSDNode>(N);
9199 if (LD->isVolatile() || !ISD::isNormalLoad(LD) ||
9200 !LD->getValueType(0).isInteger())
9203 // Keep track of already used bits to detect overlapping values.
9204 // In that case, we will just abort the transformation.
9205 APInt UsedBits(LD->getValueSizeInBits(0), 0);
9207 SmallVector<LoadedSlice, 4> LoadedSlices;
9209 // Check if this load is used as several smaller chunks of bits.
9210 // Basically, look for uses in trunc or trunc(lshr) and record a new chain
9211 // of computation for each trunc.
9212 for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end();
9213 UI != UIEnd; ++UI) {
9214 // Skip the uses of the chain.
9215 if (UI.getUse().getResNo() != 0)
9221 // Check if this is a trunc(lshr).
9222 if (User->getOpcode() == ISD::SRL && User->hasOneUse() &&
9223 isa<ConstantSDNode>(User->getOperand(1))) {
9224 Shift = cast<ConstantSDNode>(User->getOperand(1))->getZExtValue();
9225 User = *User->use_begin();
9228 // At this point, User is a Truncate, iff we encountered, trunc or
9230 if (User->getOpcode() != ISD::TRUNCATE)
9233 // The width of the type must be a power of 2 and greater than 8-bits.
9234 // Otherwise the load cannot be represented in LLVM IR.
9235 // Moreover, if we shifted with a non-8-bits multiple, the slice
9236 // will be across several bytes. We do not support that.
9237 unsigned Width = User->getValueSizeInBits(0);
9238 if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7))
9241 // Build the slice for this chain of computations.
9242 LoadedSlice LS(User, LD, Shift, &DAG);
9243 APInt CurrentUsedBits = LS.getUsedBits();
9245 // Check if this slice overlaps with another.
9246 if ((CurrentUsedBits & UsedBits) != 0)
9248 // Update the bits used globally.
9249 UsedBits |= CurrentUsedBits;
9251 // Check if the new slice would be legal.
9255 // Record the slice.
9256 LoadedSlices.push_back(LS);
9259 // Abort slicing if it does not seem to be profitable.
9260 if (!isSlicingProfitable(LoadedSlices, UsedBits, ForCodeSize))
9265 // Rewrite each chain to use an independent load.
9266 // By construction, each chain can be represented by a unique load.
9268 // Prepare the argument for the new token factor for all the slices.
9269 SmallVector<SDValue, 8> ArgChains;
9270 for (SmallVectorImpl<LoadedSlice>::const_iterator
9271 LSIt = LoadedSlices.begin(),
9272 LSItEnd = LoadedSlices.end();
9273 LSIt != LSItEnd; ++LSIt) {
9274 SDValue SliceInst = LSIt->loadSlice();
9275 CombineTo(LSIt->Inst, SliceInst, true);
9276 if (SliceInst.getNode()->getOpcode() != ISD::LOAD)
9277 SliceInst = SliceInst.getOperand(0);
9278 assert(SliceInst->getOpcode() == ISD::LOAD &&
9279 "It takes more than a zext to get to the loaded slice!!");
9280 ArgChains.push_back(SliceInst.getValue(1));
9283 SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other,
9285 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
9289 /// Check to see if V is (and load (ptr), imm), where the load is having
9290 /// specific bytes cleared out. If so, return the byte size being masked out
9291 /// and the shift amount.
9292 static std::pair<unsigned, unsigned>
9293 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
9294 std::pair<unsigned, unsigned> Result(0, 0);
9296 // Check for the structure we're looking for.
9297 if (V->getOpcode() != ISD::AND ||
9298 !isa<ConstantSDNode>(V->getOperand(1)) ||
9299 !ISD::isNormalLoad(V->getOperand(0).getNode()))
9302 // Check the chain and pointer.
9303 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
9304 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
9306 // The store should be chained directly to the load or be an operand of a
9308 if (LD == Chain.getNode())
9310 else if (Chain->getOpcode() != ISD::TokenFactor)
9311 return Result; // Fail.
9314 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
9315 if (Chain->getOperand(i).getNode() == LD) {
9319 if (!isOk) return Result;
9322 // This only handles simple types.
9323 if (V.getValueType() != MVT::i16 &&
9324 V.getValueType() != MVT::i32 &&
9325 V.getValueType() != MVT::i64)
9328 // Check the constant mask. Invert it so that the bits being masked out are
9329 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
9330 // follow the sign bit for uniformity.
9331 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
9332 unsigned NotMaskLZ = countLeadingZeros(NotMask);
9333 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
9334 unsigned NotMaskTZ = countTrailingZeros(NotMask);
9335 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
9336 if (NotMaskLZ == 64) return Result; // All zero mask.
9338 // See if we have a continuous run of bits. If so, we have 0*1+0*
9339 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
9342 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
9343 if (V.getValueType() != MVT::i64 && NotMaskLZ)
9344 NotMaskLZ -= 64-V.getValueSizeInBits();
9346 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
9347 switch (MaskedBytes) {
9351 default: return Result; // All one mask, or 5-byte mask.
9354 // Verify that the first bit starts at a multiple of mask so that the access
9355 // is aligned the same as the access width.
9356 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
9358 Result.first = MaskedBytes;
9359 Result.second = NotMaskTZ/8;
9364 /// Check to see if IVal is something that provides a value as specified by
9365 /// MaskInfo. If so, replace the specified store with a narrower store of
9368 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
9369 SDValue IVal, StoreSDNode *St,
9371 unsigned NumBytes = MaskInfo.first;
9372 unsigned ByteShift = MaskInfo.second;
9373 SelectionDAG &DAG = DC->getDAG();
9375 // Check to see if IVal is all zeros in the part being masked in by the 'or'
9376 // that uses this. If not, this is not a replacement.
9377 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
9378 ByteShift*8, (ByteShift+NumBytes)*8);
9379 if (!DAG.MaskedValueIsZero(IVal, Mask)) return nullptr;
9381 // Check that it is legal on the target to do this. It is legal if the new
9382 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
9384 MVT VT = MVT::getIntegerVT(NumBytes*8);
9385 if (!DC->isTypeLegal(VT))
9388 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
9389 // shifted by ByteShift and truncated down to NumBytes.
9391 IVal = DAG.getNode(ISD::SRL, SDLoc(IVal), IVal.getValueType(), IVal,
9392 DAG.getConstant(ByteShift*8,
9393 DC->getShiftAmountTy(IVal.getValueType())));
9395 // Figure out the offset for the store and the alignment of the access.
9397 unsigned NewAlign = St->getAlignment();
9399 if (DAG.getTargetLoweringInfo().isLittleEndian())
9400 StOffset = ByteShift;
9402 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
9404 SDValue Ptr = St->getBasePtr();
9406 Ptr = DAG.getNode(ISD::ADD, SDLoc(IVal), Ptr.getValueType(),
9407 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
9408 NewAlign = MinAlign(NewAlign, StOffset);
9411 // Truncate down to the new size.
9412 IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal);
9415 return DAG.getStore(St->getChain(), SDLoc(St), IVal, Ptr,
9416 St->getPointerInfo().getWithOffset(StOffset),
9417 false, false, NewAlign).getNode();
9421 /// Look for sequence of load / op / store where op is one of 'or', 'xor', and
9422 /// 'and' of immediates. If 'op' is only touching some of the loaded bits, try
9423 /// narrowing the load and store if it would end up being a win for performance
9425 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
9426 StoreSDNode *ST = cast<StoreSDNode>(N);
9427 if (ST->isVolatile())
9430 SDValue Chain = ST->getChain();
9431 SDValue Value = ST->getValue();
9432 SDValue Ptr = ST->getBasePtr();
9433 EVT VT = Value.getValueType();
9435 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
9438 unsigned Opc = Value.getOpcode();
9440 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
9441 // is a byte mask indicating a consecutive number of bytes, check to see if
9442 // Y is known to provide just those bytes. If so, we try to replace the
9443 // load + replace + store sequence with a single (narrower) store, which makes
9445 if (Opc == ISD::OR) {
9446 std::pair<unsigned, unsigned> MaskedLoad;
9447 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
9448 if (MaskedLoad.first)
9449 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
9450 Value.getOperand(1), ST,this))
9451 return SDValue(NewST, 0);
9453 // Or is commutative, so try swapping X and Y.
9454 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
9455 if (MaskedLoad.first)
9456 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
9457 Value.getOperand(0), ST,this))
9458 return SDValue(NewST, 0);
9461 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
9462 Value.getOperand(1).getOpcode() != ISD::Constant)
9465 SDValue N0 = Value.getOperand(0);
9466 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
9467 Chain == SDValue(N0.getNode(), 1)) {
9468 LoadSDNode *LD = cast<LoadSDNode>(N0);
9469 if (LD->getBasePtr() != Ptr ||
9470 LD->getPointerInfo().getAddrSpace() !=
9471 ST->getPointerInfo().getAddrSpace())
9474 // Find the type to narrow it the load / op / store to.
9475 SDValue N1 = Value.getOperand(1);
9476 unsigned BitWidth = N1.getValueSizeInBits();
9477 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
9478 if (Opc == ISD::AND)
9479 Imm ^= APInt::getAllOnesValue(BitWidth);
9480 if (Imm == 0 || Imm.isAllOnesValue())
9482 unsigned ShAmt = Imm.countTrailingZeros();
9483 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
9484 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
9485 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
9486 // The narrowing should be profitable, the load/store operation should be
9487 // legal (or custom) and the store size should be equal to the NewVT width.
9488 while (NewBW < BitWidth &&
9489 !(TLI.isOperationLegalOrCustom(Opc, NewVT) &&
9490 TLI.isNarrowingProfitable(VT, NewVT))) {
9491 NewBW = NextPowerOf2(NewBW);
9492 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
9494 if (NewBW >= BitWidth)
9497 // If the lsb changed does not start at the type bitwidth boundary,
9498 // start at the previous one.
9500 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
9501 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt,
9502 std::min(BitWidth, ShAmt + NewBW));
9503 if ((Imm & Mask) == Imm) {
9504 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
9505 if (Opc == ISD::AND)
9506 NewImm ^= APInt::getAllOnesValue(NewBW);
9507 uint64_t PtrOff = ShAmt / 8;
9508 // For big endian targets, we need to adjust the offset to the pointer to
9509 // load the correct bytes.
9510 if (TLI.isBigEndian())
9511 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
9513 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
9514 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
9515 if (NewAlign < TLI.getDataLayout()->getABITypeAlignment(NewVTTy))
9518 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD),
9519 Ptr.getValueType(), Ptr,
9520 DAG.getConstant(PtrOff, Ptr.getValueType()));
9521 SDValue NewLD = DAG.getLoad(NewVT, SDLoc(N0),
9522 LD->getChain(), NewPtr,
9523 LD->getPointerInfo().getWithOffset(PtrOff),
9524 LD->isVolatile(), LD->isNonTemporal(),
9525 LD->isInvariant(), NewAlign,
9527 SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD,
9528 DAG.getConstant(NewImm, NewVT));
9529 SDValue NewST = DAG.getStore(Chain, SDLoc(N),
9531 ST->getPointerInfo().getWithOffset(PtrOff),
9532 false, false, NewAlign);
9534 AddToWorklist(NewPtr.getNode());
9535 AddToWorklist(NewLD.getNode());
9536 AddToWorklist(NewVal.getNode());
9537 WorklistRemover DeadNodes(*this);
9538 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
9547 /// For a given floating point load / store pair, if the load value isn't used
9548 /// by any other operations, then consider transforming the pair to integer
9549 /// load / store operations if the target deems the transformation profitable.
9550 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
9551 StoreSDNode *ST = cast<StoreSDNode>(N);
9552 SDValue Chain = ST->getChain();
9553 SDValue Value = ST->getValue();
9554 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
9555 Value.hasOneUse() &&
9556 Chain == SDValue(Value.getNode(), 1)) {
9557 LoadSDNode *LD = cast<LoadSDNode>(Value);
9558 EVT VT = LD->getMemoryVT();
9559 if (!VT.isFloatingPoint() ||
9560 VT != ST->getMemoryVT() ||
9561 LD->isNonTemporal() ||
9562 ST->isNonTemporal() ||
9563 LD->getPointerInfo().getAddrSpace() != 0 ||
9564 ST->getPointerInfo().getAddrSpace() != 0)
9567 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
9568 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
9569 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
9570 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
9571 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
9574 unsigned LDAlign = LD->getAlignment();
9575 unsigned STAlign = ST->getAlignment();
9576 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
9577 unsigned ABIAlign = TLI.getDataLayout()->getABITypeAlignment(IntVTTy);
9578 if (LDAlign < ABIAlign || STAlign < ABIAlign)
9581 SDValue NewLD = DAG.getLoad(IntVT, SDLoc(Value),
9582 LD->getChain(), LD->getBasePtr(),
9583 LD->getPointerInfo(),
9584 false, false, false, LDAlign);
9586 SDValue NewST = DAG.getStore(NewLD.getValue(1), SDLoc(N),
9587 NewLD, ST->getBasePtr(),
9588 ST->getPointerInfo(),
9589 false, false, STAlign);
9591 AddToWorklist(NewLD.getNode());
9592 AddToWorklist(NewST.getNode());
9593 WorklistRemover DeadNodes(*this);
9594 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
9602 /// Helper struct to parse and store a memory address as base + index + offset.
9603 /// We ignore sign extensions when it is safe to do so.
9604 /// The following two expressions are not equivalent. To differentiate we need
9605 /// to store whether there was a sign extension involved in the index
9607 /// (load (i64 add (i64 copyfromreg %c)
9608 /// (i64 signextend (add (i8 load %index)
9612 /// (load (i64 add (i64 copyfromreg %c)
9613 /// (i64 signextend (i32 add (i32 signextend (i8 load %index))
9615 struct BaseIndexOffset {
9619 bool IsIndexSignExt;
9621 BaseIndexOffset() : Offset(0), IsIndexSignExt(false) {}
9623 BaseIndexOffset(SDValue Base, SDValue Index, int64_t Offset,
9624 bool IsIndexSignExt) :
9625 Base(Base), Index(Index), Offset(Offset), IsIndexSignExt(IsIndexSignExt) {}
9627 bool equalBaseIndex(const BaseIndexOffset &Other) {
9628 return Other.Base == Base && Other.Index == Index &&
9629 Other.IsIndexSignExt == IsIndexSignExt;
9632 /// Parses tree in Ptr for base, index, offset addresses.
9633 static BaseIndexOffset match(SDValue Ptr) {
9634 bool IsIndexSignExt = false;
9636 // We only can pattern match BASE + INDEX + OFFSET. If Ptr is not an ADD
9637 // instruction, then it could be just the BASE or everything else we don't
9638 // know how to handle. Just use Ptr as BASE and give up.
9639 if (Ptr->getOpcode() != ISD::ADD)
9640 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
9642 // We know that we have at least an ADD instruction. Try to pattern match
9643 // the simple case of BASE + OFFSET.
9644 if (isa<ConstantSDNode>(Ptr->getOperand(1))) {
9645 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
9646 return BaseIndexOffset(Ptr->getOperand(0), SDValue(), Offset,
9650 // Inside a loop the current BASE pointer is calculated using an ADD and a
9651 // MUL instruction. In this case Ptr is the actual BASE pointer.
9652 // (i64 add (i64 %array_ptr)
9653 // (i64 mul (i64 %induction_var)
9654 // (i64 %element_size)))
9655 if (Ptr->getOperand(1)->getOpcode() == ISD::MUL)
9656 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
9658 // Look at Base + Index + Offset cases.
9659 SDValue Base = Ptr->getOperand(0);
9660 SDValue IndexOffset = Ptr->getOperand(1);
9662 // Skip signextends.
9663 if (IndexOffset->getOpcode() == ISD::SIGN_EXTEND) {
9664 IndexOffset = IndexOffset->getOperand(0);
9665 IsIndexSignExt = true;
9668 // Either the case of Base + Index (no offset) or something else.
9669 if (IndexOffset->getOpcode() != ISD::ADD)
9670 return BaseIndexOffset(Base, IndexOffset, 0, IsIndexSignExt);
9672 // Now we have the case of Base + Index + offset.
9673 SDValue Index = IndexOffset->getOperand(0);
9674 SDValue Offset = IndexOffset->getOperand(1);
9676 if (!isa<ConstantSDNode>(Offset))
9677 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
9679 // Ignore signextends.
9680 if (Index->getOpcode() == ISD::SIGN_EXTEND) {
9681 Index = Index->getOperand(0);
9682 IsIndexSignExt = true;
9683 } else IsIndexSignExt = false;
9685 int64_t Off = cast<ConstantSDNode>(Offset)->getSExtValue();
9686 return BaseIndexOffset(Base, Index, Off, IsIndexSignExt);
9690 /// Holds a pointer to an LSBaseSDNode as well as information on where it
9691 /// is located in a sequence of memory operations connected by a chain.
9693 MemOpLink (LSBaseSDNode *N, int64_t Offset, unsigned Seq):
9694 MemNode(N), OffsetFromBase(Offset), SequenceNum(Seq) { }
9695 // Ptr to the mem node.
9696 LSBaseSDNode *MemNode;
9697 // Offset from the base ptr.
9698 int64_t OffsetFromBase;
9699 // What is the sequence number of this mem node.
9700 // Lowest mem operand in the DAG starts at zero.
9701 unsigned SequenceNum;
9704 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
9705 EVT MemVT = St->getMemoryVT();
9706 int64_t ElementSizeBytes = MemVT.getSizeInBits()/8;
9707 bool NoVectors = DAG.getMachineFunction().getFunction()->getAttributes().
9708 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
9710 // Don't merge vectors into wider inputs.
9711 if (MemVT.isVector() || !MemVT.isSimple())
9714 // Perform an early exit check. Do not bother looking at stored values that
9715 // are not constants or loads.
9716 SDValue StoredVal = St->getValue();
9717 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
9718 if (!isa<ConstantSDNode>(StoredVal) && !isa<ConstantFPSDNode>(StoredVal) &&
9722 // Only look at ends of store sequences.
9723 SDValue Chain = SDValue(St, 0);
9724 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
9727 // This holds the base pointer, index, and the offset in bytes from the base
9729 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr());
9731 // We must have a base and an offset.
9732 if (!BasePtr.Base.getNode())
9735 // Do not handle stores to undef base pointers.
9736 if (BasePtr.Base.getOpcode() == ISD::UNDEF)
9739 // Save the LoadSDNodes that we find in the chain.
9740 // We need to make sure that these nodes do not interfere with
9741 // any of the store nodes.
9742 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
9744 // Save the StoreSDNodes that we find in the chain.
9745 SmallVector<MemOpLink, 8> StoreNodes;
9747 // Walk up the chain and look for nodes with offsets from the same
9748 // base pointer. Stop when reaching an instruction with a different kind
9749 // or instruction which has a different base pointer.
9751 StoreSDNode *Index = St;
9753 // If the chain has more than one use, then we can't reorder the mem ops.
9754 if (Index != St && !SDValue(Index, 0)->hasOneUse())
9757 // Find the base pointer and offset for this memory node.
9758 BaseIndexOffset Ptr = BaseIndexOffset::match(Index->getBasePtr());
9760 // Check that the base pointer is the same as the original one.
9761 if (!Ptr.equalBaseIndex(BasePtr))
9764 // Check that the alignment is the same.
9765 if (Index->getAlignment() != St->getAlignment())
9768 // The memory operands must not be volatile.
9769 if (Index->isVolatile() || Index->isIndexed())
9773 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
9774 if (St->isTruncatingStore())
9777 // The stored memory type must be the same.
9778 if (Index->getMemoryVT() != MemVT)
9781 // We do not allow unaligned stores because we want to prevent overriding
9783 if (Index->getAlignment()*8 != MemVT.getSizeInBits())
9786 // We found a potential memory operand to merge.
9787 StoreNodes.push_back(MemOpLink(Index, Ptr.Offset, Seq++));
9789 // Find the next memory operand in the chain. If the next operand in the
9790 // chain is a store then move up and continue the scan with the next
9791 // memory operand. If the next operand is a load save it and use alias
9792 // information to check if it interferes with anything.
9793 SDNode *NextInChain = Index->getChain().getNode();
9795 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
9796 // We found a store node. Use it for the next iteration.
9799 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
9800 if (Ldn->isVolatile()) {
9805 // Save the load node for later. Continue the scan.
9806 AliasLoadNodes.push_back(Ldn);
9807 NextInChain = Ldn->getChain().getNode();
9816 // Check if there is anything to merge.
9817 if (StoreNodes.size() < 2)
9820 // Sort the memory operands according to their distance from the base pointer.
9821 std::sort(StoreNodes.begin(), StoreNodes.end(),
9822 [](MemOpLink LHS, MemOpLink RHS) {
9823 return LHS.OffsetFromBase < RHS.OffsetFromBase ||
9824 (LHS.OffsetFromBase == RHS.OffsetFromBase &&
9825 LHS.SequenceNum > RHS.SequenceNum);
9828 // Scan the memory operations on the chain and find the first non-consecutive
9829 // store memory address.
9830 unsigned LastConsecutiveStore = 0;
9831 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
9832 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) {
9834 // Check that the addresses are consecutive starting from the second
9835 // element in the list of stores.
9837 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
9838 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
9843 // Check if this store interferes with any of the loads that we found.
9844 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld)
9845 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) {
9849 // We found a load that alias with this store. Stop the sequence.
9853 // Mark this node as useful.
9854 LastConsecutiveStore = i;
9857 // The node with the lowest store address.
9858 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
9860 // Store the constants into memory as one consecutive store.
9862 unsigned LastLegalType = 0;
9863 unsigned LastLegalVectorType = 0;
9864 bool NonZero = false;
9865 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
9866 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9867 SDValue StoredVal = St->getValue();
9869 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
9870 NonZero |= !C->isNullValue();
9871 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) {
9872 NonZero |= !C->getConstantFPValue()->isNullValue();
9878 // Find a legal type for the constant store.
9879 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
9880 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9881 if (TLI.isTypeLegal(StoreTy))
9882 LastLegalType = i+1;
9883 // Or check whether a truncstore is legal.
9884 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
9885 TargetLowering::TypePromoteInteger) {
9886 EVT LegalizedStoredValueTy =
9887 TLI.getTypeToTransformTo(*DAG.getContext(), StoredVal.getValueType());
9888 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy))
9889 LastLegalType = i+1;
9892 // Find a legal type for the vector store.
9893 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
9894 if (TLI.isTypeLegal(Ty))
9895 LastLegalVectorType = i + 1;
9898 // We only use vectors if the constant is known to be zero and the
9899 // function is not marked with the noimplicitfloat attribute.
9900 if (NonZero || NoVectors)
9901 LastLegalVectorType = 0;
9903 // Check if we found a legal integer type to store.
9904 if (LastLegalType == 0 && LastLegalVectorType == 0)
9907 bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors;
9908 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType;
9910 // Make sure we have something to merge.
9914 unsigned EarliestNodeUsed = 0;
9915 for (unsigned i=0; i < NumElem; ++i) {
9916 // Find a chain for the new wide-store operand. Notice that some
9917 // of the store nodes that we found may not be selected for inclusion
9918 // in the wide store. The chain we use needs to be the chain of the
9919 // earliest store node which is *used* and replaced by the wide store.
9920 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
9921 EarliestNodeUsed = i;
9924 // The earliest Node in the DAG.
9925 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
9926 SDLoc DL(StoreNodes[0].MemNode);
9930 // Find a legal type for the vector store.
9931 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
9932 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
9933 StoredVal = DAG.getConstant(0, Ty);
9935 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
9936 APInt StoreInt(StoreBW, 0);
9938 // Construct a single integer constant which is made of the smaller
9940 bool IsLE = TLI.isLittleEndian();
9941 for (unsigned i = 0; i < NumElem ; ++i) {
9942 unsigned Idx = IsLE ?(NumElem - 1 - i) : i;
9943 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
9944 SDValue Val = St->getValue();
9945 StoreInt<<=ElementSizeBytes*8;
9946 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
9947 StoreInt|=C->getAPIntValue().zext(StoreBW);
9948 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
9949 StoreInt|= C->getValueAPF().bitcastToAPInt().zext(StoreBW);
9951 llvm_unreachable("Invalid constant element type");
9955 // Create the new Load and Store operations.
9956 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9957 StoredVal = DAG.getConstant(StoreInt, StoreTy);
9960 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), DL, StoredVal,
9961 FirstInChain->getBasePtr(),
9962 FirstInChain->getPointerInfo(),
9964 FirstInChain->getAlignment());
9966 // Replace the first store with the new store
9967 CombineTo(EarliestOp, NewStore);
9968 // Erase all other stores.
9969 for (unsigned i = 0; i < NumElem ; ++i) {
9970 if (StoreNodes[i].MemNode == EarliestOp)
9972 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9973 // ReplaceAllUsesWith will replace all uses that existed when it was
9974 // called, but graph optimizations may cause new ones to appear. For
9975 // example, the case in pr14333 looks like
9977 // St's chain -> St -> another store -> X
9979 // And the only difference from St to the other store is the chain.
9980 // When we change it's chain to be St's chain they become identical,
9981 // get CSEed and the net result is that X is now a use of St.
9982 // Since we know that St is redundant, just iterate.
9983 while (!St->use_empty())
9984 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain());
9985 deleteAndRecombine(St);
9991 // Below we handle the case of multiple consecutive stores that
9992 // come from multiple consecutive loads. We merge them into a single
9993 // wide load and a single wide store.
9995 // Look for load nodes which are used by the stored values.
9996 SmallVector<MemOpLink, 8> LoadNodes;
9998 // Find acceptable loads. Loads need to have the same chain (token factor),
9999 // must not be zext, volatile, indexed, and they must be consecutive.
10000 BaseIndexOffset LdBasePtr;
10001 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
10002 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10003 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
10006 // Loads must only have one use.
10007 if (!Ld->hasNUsesOfValue(1, 0))
10010 // Check that the alignment is the same as the stores.
10011 if (Ld->getAlignment() != St->getAlignment())
10014 // The memory operands must not be volatile.
10015 if (Ld->isVolatile() || Ld->isIndexed())
10018 // We do not accept ext loads.
10019 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
10022 // The stored memory type must be the same.
10023 if (Ld->getMemoryVT() != MemVT)
10026 BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld->getBasePtr());
10027 // If this is not the first ptr that we check.
10028 if (LdBasePtr.Base.getNode()) {
10029 // The base ptr must be the same.
10030 if (!LdPtr.equalBaseIndex(LdBasePtr))
10033 // Check that all other base pointers are the same as this one.
10037 // We found a potential memory operand to merge.
10038 LoadNodes.push_back(MemOpLink(Ld, LdPtr.Offset, 0));
10041 if (LoadNodes.size() < 2)
10044 // If we have load/store pair instructions and we only have two values,
10046 unsigned RequiredAlignment;
10047 if (LoadNodes.size() == 2 && TLI.hasPairedLoad(MemVT, RequiredAlignment) &&
10048 St->getAlignment() >= RequiredAlignment)
10051 // Scan the memory operations on the chain and find the first non-consecutive
10052 // load memory address. These variables hold the index in the store node
10054 unsigned LastConsecutiveLoad = 0;
10055 // This variable refers to the size and not index in the array.
10056 unsigned LastLegalVectorType = 0;
10057 unsigned LastLegalIntegerType = 0;
10058 StartAddress = LoadNodes[0].OffsetFromBase;
10059 SDValue FirstChain = LoadNodes[0].MemNode->getChain();
10060 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
10061 // All loads much share the same chain.
10062 if (LoadNodes[i].MemNode->getChain() != FirstChain)
10065 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
10066 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
10068 LastConsecutiveLoad = i;
10070 // Find a legal type for the vector store.
10071 EVT StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
10072 if (TLI.isTypeLegal(StoreTy))
10073 LastLegalVectorType = i + 1;
10075 // Find a legal type for the integer store.
10076 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
10077 StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
10078 if (TLI.isTypeLegal(StoreTy))
10079 LastLegalIntegerType = i + 1;
10080 // Or check whether a truncstore and extload is legal.
10081 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
10082 TargetLowering::TypePromoteInteger) {
10083 EVT LegalizedStoredValueTy =
10084 TLI.getTypeToTransformTo(*DAG.getContext(), StoreTy);
10085 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
10086 TLI.isLoadExtLegal(ISD::ZEXTLOAD, LegalizedStoredValueTy, StoreTy) &&
10087 TLI.isLoadExtLegal(ISD::SEXTLOAD, LegalizedStoredValueTy, StoreTy) &&
10088 TLI.isLoadExtLegal(ISD::EXTLOAD, LegalizedStoredValueTy, StoreTy))
10089 LastLegalIntegerType = i+1;
10093 // Only use vector types if the vector type is larger than the integer type.
10094 // If they are the same, use integers.
10095 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors;
10096 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType);
10098 // We add +1 here because the LastXXX variables refer to location while
10099 // the NumElem refers to array/index size.
10100 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1;
10101 NumElem = std::min(LastLegalType, NumElem);
10106 // The earliest Node in the DAG.
10107 unsigned EarliestNodeUsed = 0;
10108 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
10109 for (unsigned i=1; i<NumElem; ++i) {
10110 // Find a chain for the new wide-store operand. Notice that some
10111 // of the store nodes that we found may not be selected for inclusion
10112 // in the wide store. The chain we use needs to be the chain of the
10113 // earliest store node which is *used* and replaced by the wide store.
10114 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
10115 EarliestNodeUsed = i;
10118 // Find if it is better to use vectors or integers to load and store
10122 JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
10124 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
10125 JointMemOpVT = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
10128 SDLoc LoadDL(LoadNodes[0].MemNode);
10129 SDLoc StoreDL(StoreNodes[0].MemNode);
10131 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
10132 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL,
10133 FirstLoad->getChain(),
10134 FirstLoad->getBasePtr(),
10135 FirstLoad->getPointerInfo(),
10136 false, false, false,
10137 FirstLoad->getAlignment());
10139 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), StoreDL, NewLoad,
10140 FirstInChain->getBasePtr(),
10141 FirstInChain->getPointerInfo(), false, false,
10142 FirstInChain->getAlignment());
10144 // Replace one of the loads with the new load.
10145 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode);
10146 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
10147 SDValue(NewLoad.getNode(), 1));
10149 // Remove the rest of the load chains.
10150 for (unsigned i = 1; i < NumElem ; ++i) {
10151 // Replace all chain users of the old load nodes with the chain of the new
10153 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
10154 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain());
10157 // Replace the first store with the new store.
10158 CombineTo(EarliestOp, NewStore);
10159 // Erase all other stores.
10160 for (unsigned i = 0; i < NumElem ; ++i) {
10161 // Remove all Store nodes.
10162 if (StoreNodes[i].MemNode == EarliestOp)
10164 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
10165 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain());
10166 deleteAndRecombine(St);
10172 SDValue DAGCombiner::visitSTORE(SDNode *N) {
10173 StoreSDNode *ST = cast<StoreSDNode>(N);
10174 SDValue Chain = ST->getChain();
10175 SDValue Value = ST->getValue();
10176 SDValue Ptr = ST->getBasePtr();
10178 // If this is a store of a bit convert, store the input value if the
10179 // resultant store does not need a higher alignment than the original.
10180 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
10181 ST->isUnindexed()) {
10182 unsigned OrigAlign = ST->getAlignment();
10183 EVT SVT = Value.getOperand(0).getValueType();
10184 unsigned Align = TLI.getDataLayout()->
10185 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
10186 if (Align <= OrigAlign &&
10187 ((!LegalOperations && !ST->isVolatile()) ||
10188 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
10189 return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0),
10190 Ptr, ST->getPointerInfo(), ST->isVolatile(),
10191 ST->isNonTemporal(), OrigAlign,
10195 // Turn 'store undef, Ptr' -> nothing.
10196 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
10199 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
10200 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
10201 // NOTE: If the original store is volatile, this transform must not increase
10202 // the number of stores. For example, on x86-32 an f64 can be stored in one
10203 // processor operation but an i64 (which is not legal) requires two. So the
10204 // transform should not be done in this case.
10205 if (Value.getOpcode() != ISD::TargetConstantFP) {
10207 switch (CFP->getSimpleValueType(0).SimpleTy) {
10208 default: llvm_unreachable("Unknown FP type");
10209 case MVT::f16: // We don't do this for these yet.
10215 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
10216 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
10217 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
10218 bitcastToAPInt().getZExtValue(), MVT::i32);
10219 return DAG.getStore(Chain, SDLoc(N), Tmp,
10220 Ptr, ST->getMemOperand());
10224 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
10225 !ST->isVolatile()) ||
10226 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
10227 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
10228 getZExtValue(), MVT::i64);
10229 return DAG.getStore(Chain, SDLoc(N), Tmp,
10230 Ptr, ST->getMemOperand());
10233 if (!ST->isVolatile() &&
10234 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
10235 // Many FP stores are not made apparent until after legalize, e.g. for
10236 // argument passing. Since this is so common, custom legalize the
10237 // 64-bit integer store into two 32-bit stores.
10238 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
10239 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
10240 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
10241 if (TLI.isBigEndian()) std::swap(Lo, Hi);
10243 unsigned Alignment = ST->getAlignment();
10244 bool isVolatile = ST->isVolatile();
10245 bool isNonTemporal = ST->isNonTemporal();
10246 AAMDNodes AAInfo = ST->getAAInfo();
10248 SDValue St0 = DAG.getStore(Chain, SDLoc(ST), Lo,
10249 Ptr, ST->getPointerInfo(),
10250 isVolatile, isNonTemporal,
10251 ST->getAlignment(), AAInfo);
10252 Ptr = DAG.getNode(ISD::ADD, SDLoc(N), Ptr.getValueType(), Ptr,
10253 DAG.getConstant(4, Ptr.getValueType()));
10254 Alignment = MinAlign(Alignment, 4U);
10255 SDValue St1 = DAG.getStore(Chain, SDLoc(ST), Hi,
10256 Ptr, ST->getPointerInfo().getWithOffset(4),
10257 isVolatile, isNonTemporal,
10258 Alignment, AAInfo);
10259 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other,
10268 // Try to infer better alignment information than the store already has.
10269 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
10270 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
10271 if (Align > ST->getAlignment())
10272 return DAG.getTruncStore(Chain, SDLoc(N), Value,
10273 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
10274 ST->isVolatile(), ST->isNonTemporal(), Align,
10279 // Try transforming a pair floating point load / store ops to integer
10280 // load / store ops.
10281 SDValue NewST = TransformFPLoadStorePair(N);
10282 if (NewST.getNode())
10285 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA
10286 : DAG.getSubtarget().useAA();
10288 if (CombinerAAOnlyFunc.getNumOccurrences() &&
10289 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
10292 if (UseAA && ST->isUnindexed()) {
10293 // Walk up chain skipping non-aliasing memory nodes.
10294 SDValue BetterChain = FindBetterChain(N, Chain);
10296 // If there is a better chain.
10297 if (Chain != BetterChain) {
10300 // Replace the chain to avoid dependency.
10301 if (ST->isTruncatingStore()) {
10302 ReplStore = DAG.getTruncStore(BetterChain, SDLoc(N), Value, Ptr,
10303 ST->getMemoryVT(), ST->getMemOperand());
10305 ReplStore = DAG.getStore(BetterChain, SDLoc(N), Value, Ptr,
10306 ST->getMemOperand());
10309 // Create token to keep both nodes around.
10310 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
10311 MVT::Other, Chain, ReplStore);
10313 // Make sure the new and old chains are cleaned up.
10314 AddToWorklist(Token.getNode());
10316 // Don't add users to work list.
10317 return CombineTo(N, Token, false);
10321 // Try transforming N to an indexed store.
10322 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
10323 return SDValue(N, 0);
10325 // FIXME: is there such a thing as a truncating indexed store?
10326 if (ST->isTruncatingStore() && ST->isUnindexed() &&
10327 Value.getValueType().isInteger()) {
10328 // See if we can simplify the input to this truncstore with knowledge that
10329 // only the low bits are being used. For example:
10330 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
10332 GetDemandedBits(Value,
10333 APInt::getLowBitsSet(
10334 Value.getValueType().getScalarType().getSizeInBits(),
10335 ST->getMemoryVT().getScalarType().getSizeInBits()));
10336 AddToWorklist(Value.getNode());
10337 if (Shorter.getNode())
10338 return DAG.getTruncStore(Chain, SDLoc(N), Shorter,
10339 Ptr, ST->getMemoryVT(), ST->getMemOperand());
10341 // Otherwise, see if we can simplify the operation with
10342 // SimplifyDemandedBits, which only works if the value has a single use.
10343 if (SimplifyDemandedBits(Value,
10344 APInt::getLowBitsSet(
10345 Value.getValueType().getScalarType().getSizeInBits(),
10346 ST->getMemoryVT().getScalarType().getSizeInBits())))
10347 return SDValue(N, 0);
10350 // If this is a load followed by a store to the same location, then the store
10352 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
10353 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
10354 ST->isUnindexed() && !ST->isVolatile() &&
10355 // There can't be any side effects between the load and store, such as
10356 // a call or store.
10357 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
10358 // The store is dead, remove it.
10363 // If this is a store followed by a store with the same value to the same
10364 // location, then the store is dead/noop.
10365 if (StoreSDNode *ST1 = dyn_cast<StoreSDNode>(Chain)) {
10366 if (ST1->getBasePtr() == Ptr && ST->getMemoryVT() == ST1->getMemoryVT() &&
10367 ST1->getValue() == Value && ST->isUnindexed() && !ST->isVolatile() &&
10368 ST1->isUnindexed() && !ST1->isVolatile()) {
10369 // The store is dead, remove it.
10374 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
10375 // truncating store. We can do this even if this is already a truncstore.
10376 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
10377 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
10378 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
10379 ST->getMemoryVT())) {
10380 return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0),
10381 Ptr, ST->getMemoryVT(), ST->getMemOperand());
10384 // Only perform this optimization before the types are legal, because we
10385 // don't want to perform this optimization on every DAGCombine invocation.
10387 bool EverChanged = false;
10390 // There can be multiple store sequences on the same chain.
10391 // Keep trying to merge store sequences until we are unable to do so
10392 // or until we merge the last store on the chain.
10393 bool Changed = MergeConsecutiveStores(ST);
10394 EverChanged |= Changed;
10395 if (!Changed) break;
10396 } while (ST->getOpcode() != ISD::DELETED_NODE);
10399 return SDValue(N, 0);
10402 return ReduceLoadOpStoreWidth(N);
10405 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
10406 SDValue InVec = N->getOperand(0);
10407 SDValue InVal = N->getOperand(1);
10408 SDValue EltNo = N->getOperand(2);
10411 // If the inserted element is an UNDEF, just use the input vector.
10412 if (InVal.getOpcode() == ISD::UNDEF)
10415 EVT VT = InVec.getValueType();
10417 // If we can't generate a legal BUILD_VECTOR, exit
10418 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
10421 // Check that we know which element is being inserted
10422 if (!isa<ConstantSDNode>(EltNo))
10424 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
10426 // Canonicalize insert_vector_elt dag nodes.
10428 // (insert_vector_elt (insert_vector_elt A, Idx0), Idx1)
10429 // -> (insert_vector_elt (insert_vector_elt A, Idx1), Idx0)
10431 // Do this only if the child insert_vector node has one use; also
10432 // do this only if indices are both constants and Idx1 < Idx0.
10433 if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT && InVec.hasOneUse()
10434 && isa<ConstantSDNode>(InVec.getOperand(2))) {
10435 unsigned OtherElt =
10436 cast<ConstantSDNode>(InVec.getOperand(2))->getZExtValue();
10437 if (Elt < OtherElt) {
10439 SDValue NewOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(N), VT,
10440 InVec.getOperand(0), InVal, EltNo);
10441 AddToWorklist(NewOp.getNode());
10442 return DAG.getNode(ISD::INSERT_VECTOR_ELT, SDLoc(InVec.getNode()),
10443 VT, NewOp, InVec.getOperand(1), InVec.getOperand(2));
10447 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
10448 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
10449 // vector elements.
10450 SmallVector<SDValue, 8> Ops;
10451 // Do not combine these two vectors if the output vector will not replace
10452 // the input vector.
10453 if (InVec.getOpcode() == ISD::BUILD_VECTOR && InVec.hasOneUse()) {
10454 Ops.append(InVec.getNode()->op_begin(),
10455 InVec.getNode()->op_end());
10456 } else if (InVec.getOpcode() == ISD::UNDEF) {
10457 unsigned NElts = VT.getVectorNumElements();
10458 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
10463 // Insert the element
10464 if (Elt < Ops.size()) {
10465 // All the operands of BUILD_VECTOR must have the same type;
10466 // we enforce that here.
10467 EVT OpVT = Ops[0].getValueType();
10468 if (InVal.getValueType() != OpVT)
10469 InVal = OpVT.bitsGT(InVal.getValueType()) ?
10470 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
10471 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
10475 // Return the new vector
10476 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
10479 SDValue DAGCombiner::ReplaceExtractVectorEltOfLoadWithNarrowedLoad(
10480 SDNode *EVE, EVT InVecVT, SDValue EltNo, LoadSDNode *OriginalLoad) {
10481 EVT ResultVT = EVE->getValueType(0);
10482 EVT VecEltVT = InVecVT.getVectorElementType();
10483 unsigned Align = OriginalLoad->getAlignment();
10484 unsigned NewAlign = TLI.getDataLayout()->getABITypeAlignment(
10485 VecEltVT.getTypeForEVT(*DAG.getContext()));
10487 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VecEltVT))
10492 SDValue NewPtr = OriginalLoad->getBasePtr();
10494 EVT PtrType = NewPtr.getValueType();
10495 MachinePointerInfo MPI;
10496 if (auto *ConstEltNo = dyn_cast<ConstantSDNode>(EltNo)) {
10497 int Elt = ConstEltNo->getZExtValue();
10498 unsigned PtrOff = VecEltVT.getSizeInBits() * Elt / 8;
10499 if (TLI.isBigEndian())
10500 PtrOff = InVecVT.getSizeInBits() / 8 - PtrOff;
10501 Offset = DAG.getConstant(PtrOff, PtrType);
10502 MPI = OriginalLoad->getPointerInfo().getWithOffset(PtrOff);
10504 Offset = DAG.getNode(
10505 ISD::MUL, SDLoc(EVE), EltNo.getValueType(), EltNo,
10506 DAG.getConstant(VecEltVT.getStoreSize(), EltNo.getValueType()));
10507 if (TLI.isBigEndian())
10508 Offset = DAG.getNode(
10509 ISD::SUB, SDLoc(EVE), EltNo.getValueType(),
10510 DAG.getConstant(InVecVT.getStoreSize(), EltNo.getValueType()), Offset);
10511 MPI = OriginalLoad->getPointerInfo();
10513 NewPtr = DAG.getNode(ISD::ADD, SDLoc(EVE), PtrType, NewPtr, Offset);
10515 // The replacement we need to do here is a little tricky: we need to
10516 // replace an extractelement of a load with a load.
10517 // Use ReplaceAllUsesOfValuesWith to do the replacement.
10518 // Note that this replacement assumes that the extractvalue is the only
10519 // use of the load; that's okay because we don't want to perform this
10520 // transformation in other cases anyway.
10523 if (ResultVT.bitsGT(VecEltVT)) {
10524 // If the result type of vextract is wider than the load, then issue an
10525 // extending load instead.
10526 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, ResultVT,
10530 Load = DAG.getExtLoad(
10531 ExtType, SDLoc(EVE), ResultVT, OriginalLoad->getChain(), NewPtr, MPI,
10532 VecEltVT, OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
10533 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
10534 Chain = Load.getValue(1);
10536 Load = DAG.getLoad(
10537 VecEltVT, SDLoc(EVE), OriginalLoad->getChain(), NewPtr, MPI,
10538 OriginalLoad->isVolatile(), OriginalLoad->isNonTemporal(),
10539 OriginalLoad->isInvariant(), Align, OriginalLoad->getAAInfo());
10540 Chain = Load.getValue(1);
10541 if (ResultVT.bitsLT(VecEltVT))
10542 Load = DAG.getNode(ISD::TRUNCATE, SDLoc(EVE), ResultVT, Load);
10544 Load = DAG.getNode(ISD::BITCAST, SDLoc(EVE), ResultVT, Load);
10546 WorklistRemover DeadNodes(*this);
10547 SDValue From[] = { SDValue(EVE, 0), SDValue(OriginalLoad, 1) };
10548 SDValue To[] = { Load, Chain };
10549 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
10550 // Since we're explicitly calling ReplaceAllUses, add the new node to the
10551 // worklist explicitly as well.
10552 AddToWorklist(Load.getNode());
10553 AddUsersToWorklist(Load.getNode()); // Add users too
10554 // Make sure to revisit this node to clean it up; it will usually be dead.
10555 AddToWorklist(EVE);
10557 return SDValue(EVE, 0);
10560 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
10561 // (vextract (scalar_to_vector val, 0) -> val
10562 SDValue InVec = N->getOperand(0);
10563 EVT VT = InVec.getValueType();
10564 EVT NVT = N->getValueType(0);
10566 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
10567 // Check if the result type doesn't match the inserted element type. A
10568 // SCALAR_TO_VECTOR may truncate the inserted element and the
10569 // EXTRACT_VECTOR_ELT may widen the extracted vector.
10570 SDValue InOp = InVec.getOperand(0);
10571 if (InOp.getValueType() != NVT) {
10572 assert(InOp.getValueType().isInteger() && NVT.isInteger());
10573 return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT);
10578 SDValue EltNo = N->getOperand(1);
10579 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
10581 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
10582 // We only perform this optimization before the op legalization phase because
10583 // we may introduce new vector instructions which are not backed by TD
10584 // patterns. For example on AVX, extracting elements from a wide vector
10585 // without using extract_subvector. However, if we can find an underlying
10586 // scalar value, then we can always use that.
10587 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
10589 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
10590 int NumElem = VT.getVectorNumElements();
10591 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
10592 // Find the new index to extract from.
10593 int OrigElt = SVOp->getMaskElt(Elt);
10595 // Extracting an undef index is undef.
10597 return DAG.getUNDEF(NVT);
10599 // Select the right vector half to extract from.
10601 if (OrigElt < NumElem) {
10602 SVInVec = InVec->getOperand(0);
10604 SVInVec = InVec->getOperand(1);
10605 OrigElt -= NumElem;
10608 if (SVInVec.getOpcode() == ISD::BUILD_VECTOR) {
10609 SDValue InOp = SVInVec.getOperand(OrigElt);
10610 if (InOp.getValueType() != NVT) {
10611 assert(InOp.getValueType().isInteger() && NVT.isInteger());
10612 InOp = DAG.getSExtOrTrunc(InOp, SDLoc(SVInVec), NVT);
10618 // FIXME: We should handle recursing on other vector shuffles and
10619 // scalar_to_vector here as well.
10621 if (!LegalOperations) {
10622 EVT IndexTy = TLI.getVectorIdxTy();
10623 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT,
10624 SVInVec, DAG.getConstant(OrigElt, IndexTy));
10628 bool BCNumEltsChanged = false;
10629 EVT ExtVT = VT.getVectorElementType();
10632 // If the result of load has to be truncated, then it's not necessarily
10634 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
10637 if (InVec.getOpcode() == ISD::BITCAST) {
10638 // Don't duplicate a load with other uses.
10639 if (!InVec.hasOneUse())
10642 EVT BCVT = InVec.getOperand(0).getValueType();
10643 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
10645 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
10646 BCNumEltsChanged = true;
10647 InVec = InVec.getOperand(0);
10648 ExtVT = BCVT.getVectorElementType();
10651 // (vextract (vN[if]M load $addr), i) -> ([if]M load $addr + i * size)
10652 if (!LegalOperations && !ConstEltNo && InVec.hasOneUse() &&
10653 ISD::isNormalLoad(InVec.getNode()) &&
10654 !N->getOperand(1)->hasPredecessor(InVec.getNode())) {
10655 SDValue Index = N->getOperand(1);
10656 if (LoadSDNode *OrigLoad = dyn_cast<LoadSDNode>(InVec))
10657 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, Index,
10661 // Perform only after legalization to ensure build_vector / vector_shuffle
10662 // optimizations have already been done.
10663 if (!LegalOperations) return SDValue();
10665 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
10666 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
10667 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
10670 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
10672 LoadSDNode *LN0 = nullptr;
10673 const ShuffleVectorSDNode *SVN = nullptr;
10674 if (ISD::isNormalLoad(InVec.getNode())) {
10675 LN0 = cast<LoadSDNode>(InVec);
10676 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
10677 InVec.getOperand(0).getValueType() == ExtVT &&
10678 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
10679 // Don't duplicate a load with other uses.
10680 if (!InVec.hasOneUse())
10683 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
10684 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
10685 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
10687 // (load $addr+1*size)
10689 // Don't duplicate a load with other uses.
10690 if (!InVec.hasOneUse())
10693 // If the bit convert changed the number of elements, it is unsafe
10694 // to examine the mask.
10695 if (BCNumEltsChanged)
10698 // Select the input vector, guarding against out of range extract vector.
10699 unsigned NumElems = VT.getVectorNumElements();
10700 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
10701 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
10703 if (InVec.getOpcode() == ISD::BITCAST) {
10704 // Don't duplicate a load with other uses.
10705 if (!InVec.hasOneUse())
10708 InVec = InVec.getOperand(0);
10710 if (ISD::isNormalLoad(InVec.getNode())) {
10711 LN0 = cast<LoadSDNode>(InVec);
10712 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
10713 EltNo = DAG.getConstant(Elt, EltNo.getValueType());
10717 // Make sure we found a non-volatile load and the extractelement is
10719 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
10722 // If Idx was -1 above, Elt is going to be -1, so just return undef.
10724 return DAG.getUNDEF(LVT);
10726 return ReplaceExtractVectorEltOfLoadWithNarrowedLoad(N, VT, EltNo, LN0);
10732 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
10733 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
10734 // We perform this optimization post type-legalization because
10735 // the type-legalizer often scalarizes integer-promoted vectors.
10736 // Performing this optimization before may create bit-casts which
10737 // will be type-legalized to complex code sequences.
10738 // We perform this optimization only before the operation legalizer because we
10739 // may introduce illegal operations.
10740 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
10743 unsigned NumInScalars = N->getNumOperands();
10745 EVT VT = N->getValueType(0);
10747 // Check to see if this is a BUILD_VECTOR of a bunch of values
10748 // which come from any_extend or zero_extend nodes. If so, we can create
10749 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
10750 // optimizations. We do not handle sign-extend because we can't fill the sign
10752 EVT SourceType = MVT::Other;
10753 bool AllAnyExt = true;
10755 for (unsigned i = 0; i != NumInScalars; ++i) {
10756 SDValue In = N->getOperand(i);
10757 // Ignore undef inputs.
10758 if (In.getOpcode() == ISD::UNDEF) continue;
10760 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
10761 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
10763 // Abort if the element is not an extension.
10764 if (!ZeroExt && !AnyExt) {
10765 SourceType = MVT::Other;
10769 // The input is a ZeroExt or AnyExt. Check the original type.
10770 EVT InTy = In.getOperand(0).getValueType();
10772 // Check that all of the widened source types are the same.
10773 if (SourceType == MVT::Other)
10776 else if (InTy != SourceType) {
10777 // Multiple income types. Abort.
10778 SourceType = MVT::Other;
10782 // Check if all of the extends are ANY_EXTENDs.
10783 AllAnyExt &= AnyExt;
10786 // In order to have valid types, all of the inputs must be extended from the
10787 // same source type and all of the inputs must be any or zero extend.
10788 // Scalar sizes must be a power of two.
10789 EVT OutScalarTy = VT.getScalarType();
10790 bool ValidTypes = SourceType != MVT::Other &&
10791 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
10792 isPowerOf2_32(SourceType.getSizeInBits());
10794 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
10795 // turn into a single shuffle instruction.
10799 bool isLE = TLI.isLittleEndian();
10800 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
10801 assert(ElemRatio > 1 && "Invalid element size ratio");
10802 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
10803 DAG.getConstant(0, SourceType);
10805 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
10806 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
10808 // Populate the new build_vector
10809 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
10810 SDValue Cast = N->getOperand(i);
10811 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
10812 Cast.getOpcode() == ISD::ZERO_EXTEND ||
10813 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
10815 if (Cast.getOpcode() == ISD::UNDEF)
10816 In = DAG.getUNDEF(SourceType);
10818 In = Cast->getOperand(0);
10819 unsigned Index = isLE ? (i * ElemRatio) :
10820 (i * ElemRatio + (ElemRatio - 1));
10822 assert(Index < Ops.size() && "Invalid index");
10826 // The type of the new BUILD_VECTOR node.
10827 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
10828 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
10829 "Invalid vector size");
10830 // Check if the new vector type is legal.
10831 if (!isTypeLegal(VecVT)) return SDValue();
10833 // Make the new BUILD_VECTOR.
10834 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, Ops);
10836 // The new BUILD_VECTOR node has the potential to be further optimized.
10837 AddToWorklist(BV.getNode());
10838 // Bitcast to the desired type.
10839 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
10842 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
10843 EVT VT = N->getValueType(0);
10845 unsigned NumInScalars = N->getNumOperands();
10848 EVT SrcVT = MVT::Other;
10849 unsigned Opcode = ISD::DELETED_NODE;
10850 unsigned NumDefs = 0;
10852 for (unsigned i = 0; i != NumInScalars; ++i) {
10853 SDValue In = N->getOperand(i);
10854 unsigned Opc = In.getOpcode();
10856 if (Opc == ISD::UNDEF)
10859 // If all scalar values are floats and converted from integers.
10860 if (Opcode == ISD::DELETED_NODE &&
10861 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
10868 EVT InVT = In.getOperand(0).getValueType();
10870 // If all scalar values are typed differently, bail out. It's chosen to
10871 // simplify BUILD_VECTOR of integer types.
10872 if (SrcVT == MVT::Other)
10879 // If the vector has just one element defined, it's not worth to fold it into
10880 // a vectorized one.
10884 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
10885 && "Should only handle conversion from integer to float.");
10886 assert(SrcVT != MVT::Other && "Cannot determine source type!");
10888 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
10890 if (!TLI.isOperationLegalOrCustom(Opcode, NVT))
10893 SmallVector<SDValue, 8> Opnds;
10894 for (unsigned i = 0; i != NumInScalars; ++i) {
10895 SDValue In = N->getOperand(i);
10897 if (In.getOpcode() == ISD::UNDEF)
10898 Opnds.push_back(DAG.getUNDEF(SrcVT));
10900 Opnds.push_back(In.getOperand(0));
10902 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT, Opnds);
10903 AddToWorklist(BV.getNode());
10905 return DAG.getNode(Opcode, dl, VT, BV);
10908 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
10909 unsigned NumInScalars = N->getNumOperands();
10911 EVT VT = N->getValueType(0);
10913 // A vector built entirely of undefs is undef.
10914 if (ISD::allOperandsUndef(N))
10915 return DAG.getUNDEF(VT);
10917 SDValue V = reduceBuildVecExtToExtBuildVec(N);
10921 V = reduceBuildVecConvertToConvertBuildVec(N);
10925 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
10926 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
10927 // at most two distinct vectors, turn this into a shuffle node.
10929 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
10930 if (!isTypeLegal(VT))
10933 // May only combine to shuffle after legalize if shuffle is legal.
10934 if (LegalOperations && !TLI.isOperationLegal(ISD::VECTOR_SHUFFLE, VT))
10937 SDValue VecIn1, VecIn2;
10938 bool UsesZeroVector = false;
10939 for (unsigned i = 0; i != NumInScalars; ++i) {
10940 SDValue Op = N->getOperand(i);
10941 // Ignore undef inputs.
10942 if (Op.getOpcode() == ISD::UNDEF) continue;
10944 // See if we can combine this build_vector into a blend with a zero vector.
10945 if (!VecIn2.getNode() && ((Op.getOpcode() == ISD::Constant &&
10946 cast<ConstantSDNode>(Op.getNode())->isNullValue()) ||
10947 (Op.getOpcode() == ISD::ConstantFP &&
10948 cast<ConstantFPSDNode>(Op.getNode())->getValueAPF().isZero()))) {
10949 UsesZeroVector = true;
10953 // If this input is something other than a EXTRACT_VECTOR_ELT with a
10954 // constant index, bail out.
10955 if (Op.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
10956 !isa<ConstantSDNode>(Op.getOperand(1))) {
10957 VecIn1 = VecIn2 = SDValue(nullptr, 0);
10961 // We allow up to two distinct input vectors.
10962 SDValue ExtractedFromVec = Op.getOperand(0);
10963 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
10966 if (!VecIn1.getNode()) {
10967 VecIn1 = ExtractedFromVec;
10968 } else if (!VecIn2.getNode() && !UsesZeroVector) {
10969 VecIn2 = ExtractedFromVec;
10971 // Too many inputs.
10972 VecIn1 = VecIn2 = SDValue(nullptr, 0);
10977 // If everything is good, we can make a shuffle operation.
10978 if (VecIn1.getNode()) {
10979 unsigned InNumElements = VecIn1.getValueType().getVectorNumElements();
10980 SmallVector<int, 8> Mask;
10981 for (unsigned i = 0; i != NumInScalars; ++i) {
10982 unsigned Opcode = N->getOperand(i).getOpcode();
10983 if (Opcode == ISD::UNDEF) {
10984 Mask.push_back(-1);
10988 // Operands can also be zero.
10989 if (Opcode != ISD::EXTRACT_VECTOR_ELT) {
10990 assert(UsesZeroVector &&
10991 (Opcode == ISD::Constant || Opcode == ISD::ConstantFP) &&
10992 "Unexpected node found!");
10993 Mask.push_back(NumInScalars+i);
10997 // If extracting from the first vector, just use the index directly.
10998 SDValue Extract = N->getOperand(i);
10999 SDValue ExtVal = Extract.getOperand(1);
11000 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
11001 if (Extract.getOperand(0) == VecIn1) {
11002 Mask.push_back(ExtIndex);
11006 // Otherwise, use InIdx + InputVecSize
11007 Mask.push_back(InNumElements + ExtIndex);
11010 // Avoid introducing illegal shuffles with zero.
11011 if (UsesZeroVector && !TLI.isVectorClearMaskLegal(Mask, VT))
11014 // We can't generate a shuffle node with mismatched input and output types.
11015 // Attempt to transform a single input vector to the correct type.
11016 if ((VT != VecIn1.getValueType())) {
11017 // If the input vector type has a different base type to the output
11018 // vector type, bail out.
11019 EVT VTElemType = VT.getVectorElementType();
11020 if ((VecIn1.getValueType().getVectorElementType() != VTElemType) ||
11021 (VecIn2.getNode() &&
11022 (VecIn2.getValueType().getVectorElementType() != VTElemType)))
11025 // If the input vector is too small, widen it.
11026 // We only support widening of vectors which are half the size of the
11027 // output registers. For example XMM->YMM widening on X86 with AVX.
11028 EVT VecInT = VecIn1.getValueType();
11029 if (VecInT.getSizeInBits() * 2 == VT.getSizeInBits()) {
11030 // If we only have one small input, widen it by adding undef values.
11031 if (!VecIn2.getNode())
11032 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, VecIn1,
11033 DAG.getUNDEF(VecIn1.getValueType()));
11034 else if (VecIn1.getValueType() == VecIn2.getValueType()) {
11035 // If we have two small inputs of the same type, try to concat them.
11036 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, VecIn1, VecIn2);
11037 VecIn2 = SDValue(nullptr, 0);
11040 } else if (VecInT.getSizeInBits() == VT.getSizeInBits() * 2) {
11041 // If the input vector is too large, try to split it.
11042 // We don't support having two input vectors that are too large.
11043 // If the zero vector was used, we can not split the vector,
11044 // since we'd need 3 inputs.
11045 if (UsesZeroVector || VecIn2.getNode())
11048 if (!TLI.isExtractSubvectorCheap(VT, VT.getVectorNumElements()))
11051 // Try to replace VecIn1 with two extract_subvectors
11052 // No need to update the masks, they should still be correct.
11053 VecIn2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, VecIn1,
11054 DAG.getConstant(VT.getVectorNumElements(), TLI.getVectorIdxTy()));
11055 VecIn1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, VecIn1,
11056 DAG.getConstant(0, TLI.getVectorIdxTy()));
11061 if (UsesZeroVector)
11062 VecIn2 = VT.isInteger() ? DAG.getConstant(0, VT) :
11063 DAG.getConstantFP(0.0, VT);
11065 // If VecIn2 is unused then change it to undef.
11066 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
11068 // Check that we were able to transform all incoming values to the same
11070 if (VecIn2.getValueType() != VecIn1.getValueType() ||
11071 VecIn1.getValueType() != VT)
11074 // Return the new VECTOR_SHUFFLE node.
11078 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]);
11084 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
11085 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
11086 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
11087 // inputs come from at most two distinct vectors, turn this into a shuffle
11090 // If we only have one input vector, we don't need to do any concatenation.
11091 if (N->getNumOperands() == 1)
11092 return N->getOperand(0);
11094 // Check if all of the operands are undefs.
11095 EVT VT = N->getValueType(0);
11096 if (ISD::allOperandsUndef(N))
11097 return DAG.getUNDEF(VT);
11099 // Optimize concat_vectors where one of the vectors is undef.
11100 if (N->getNumOperands() == 2 &&
11101 N->getOperand(1)->getOpcode() == ISD::UNDEF) {
11102 SDValue In = N->getOperand(0);
11103 assert(In.getValueType().isVector() && "Must concat vectors");
11105 // Transform: concat_vectors(scalar, undef) -> scalar_to_vector(sclr).
11106 if (In->getOpcode() == ISD::BITCAST &&
11107 !In->getOperand(0)->getValueType(0).isVector()) {
11108 SDValue Scalar = In->getOperand(0);
11109 EVT SclTy = Scalar->getValueType(0);
11111 if (!SclTy.isFloatingPoint() && !SclTy.isInteger())
11114 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SclTy,
11115 VT.getSizeInBits() / SclTy.getSizeInBits());
11116 if (!TLI.isTypeLegal(NVT) || !TLI.isTypeLegal(Scalar.getValueType()))
11119 SDLoc dl = SDLoc(N);
11120 SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NVT, Scalar);
11121 return DAG.getNode(ISD::BITCAST, dl, VT, Res);
11125 // fold (concat_vectors (BUILD_VECTOR A, B, ...), (BUILD_VECTOR C, D, ...))
11126 // -> (BUILD_VECTOR A, B, ..., C, D, ...)
11127 if (N->getNumOperands() == 2 &&
11128 N->getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
11129 N->getOperand(1).getOpcode() == ISD::BUILD_VECTOR) {
11130 EVT VT = N->getValueType(0);
11131 SDValue N0 = N->getOperand(0);
11132 SDValue N1 = N->getOperand(1);
11133 SmallVector<SDValue, 8> Opnds;
11134 unsigned BuildVecNumElts = N0.getNumOperands();
11136 EVT SclTy0 = N0.getOperand(0)->getValueType(0);
11137 EVT SclTy1 = N1.getOperand(0)->getValueType(0);
11138 if (SclTy0.isFloatingPoint()) {
11139 for (unsigned i = 0; i != BuildVecNumElts; ++i)
11140 Opnds.push_back(N0.getOperand(i));
11141 for (unsigned i = 0; i != BuildVecNumElts; ++i)
11142 Opnds.push_back(N1.getOperand(i));
11144 // If BUILD_VECTOR are from built from integer, they may have different
11145 // operand types. Get the smaller type and truncate all operands to it.
11146 EVT MinTy = SclTy0.bitsLE(SclTy1) ? SclTy0 : SclTy1;
11147 for (unsigned i = 0; i != BuildVecNumElts; ++i)
11148 Opnds.push_back(DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinTy,
11149 N0.getOperand(i)));
11150 for (unsigned i = 0; i != BuildVecNumElts; ++i)
11151 Opnds.push_back(DAG.getNode(ISD::TRUNCATE, SDLoc(N), MinTy,
11152 N1.getOperand(i)));
11155 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, Opnds);
11158 // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
11159 // nodes often generate nop CONCAT_VECTOR nodes.
11160 // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that
11161 // place the incoming vectors at the exact same location.
11162 SDValue SingleSource = SDValue();
11163 unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements();
11165 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
11166 SDValue Op = N->getOperand(i);
11168 if (Op.getOpcode() == ISD::UNDEF)
11171 // Check if this is the identity extract:
11172 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
11175 // Find the single incoming vector for the extract_subvector.
11176 if (SingleSource.getNode()) {
11177 if (Op.getOperand(0) != SingleSource)
11180 SingleSource = Op.getOperand(0);
11182 // Check the source type is the same as the type of the result.
11183 // If not, this concat may extend the vector, so we can not
11184 // optimize it away.
11185 if (SingleSource.getValueType() != N->getValueType(0))
11189 unsigned IdentityIndex = i * PartNumElem;
11190 ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1));
11191 // The extract index must be constant.
11195 // Check that we are reading from the identity index.
11196 if (CS->getZExtValue() != IdentityIndex)
11200 if (SingleSource.getNode())
11201 return SingleSource;
11206 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
11207 EVT NVT = N->getValueType(0);
11208 SDValue V = N->getOperand(0);
11210 if (V->getOpcode() == ISD::CONCAT_VECTORS) {
11212 // (extract_subvec (concat V1, V2, ...), i)
11215 // Only operand 0 is checked as 'concat' assumes all inputs of the same
11217 if (V->getOperand(0).getValueType() != NVT)
11219 unsigned Idx = dyn_cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
11220 unsigned NumElems = NVT.getVectorNumElements();
11221 assert((Idx % NumElems) == 0 &&
11222 "IDX in concat is not a multiple of the result vector length.");
11223 return V->getOperand(Idx / NumElems);
11227 if (V->getOpcode() == ISD::BITCAST)
11228 V = V.getOperand(0);
11230 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
11232 // Handle only simple case where vector being inserted and vector
11233 // being extracted are of same type, and are half size of larger vectors.
11234 EVT BigVT = V->getOperand(0).getValueType();
11235 EVT SmallVT = V->getOperand(1).getValueType();
11236 if (!NVT.bitsEq(SmallVT) || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
11239 // Only handle cases where both indexes are constants with the same type.
11240 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
11241 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
11243 if (InsIdx && ExtIdx &&
11244 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
11245 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
11247 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
11249 // indices are equal or bit offsets are equal => V1
11250 // otherwise => (extract_subvec V1, ExtIdx)
11251 if (InsIdx->getZExtValue() * SmallVT.getScalarType().getSizeInBits() ==
11252 ExtIdx->getZExtValue() * NVT.getScalarType().getSizeInBits())
11253 return DAG.getNode(ISD::BITCAST, dl, NVT, V->getOperand(1));
11254 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT,
11255 DAG.getNode(ISD::BITCAST, dl,
11256 N->getOperand(0).getValueType(),
11257 V->getOperand(0)), N->getOperand(1));
11264 static SDValue simplifyShuffleOperandRecursively(SmallBitVector &UsedElements,
11265 SDValue V, SelectionDAG &DAG) {
11267 EVT VT = V.getValueType();
11269 switch (V.getOpcode()) {
11273 case ISD::CONCAT_VECTORS: {
11274 EVT OpVT = V->getOperand(0).getValueType();
11275 int OpSize = OpVT.getVectorNumElements();
11276 SmallBitVector OpUsedElements(OpSize, false);
11277 bool FoundSimplification = false;
11278 SmallVector<SDValue, 4> NewOps;
11279 NewOps.reserve(V->getNumOperands());
11280 for (int i = 0, NumOps = V->getNumOperands(); i < NumOps; ++i) {
11281 SDValue Op = V->getOperand(i);
11282 bool OpUsed = false;
11283 for (int j = 0; j < OpSize; ++j)
11284 if (UsedElements[i * OpSize + j]) {
11285 OpUsedElements[j] = true;
11289 OpUsed ? simplifyShuffleOperandRecursively(OpUsedElements, Op, DAG)
11290 : DAG.getUNDEF(OpVT));
11291 FoundSimplification |= Op == NewOps.back();
11292 OpUsedElements.reset();
11294 if (FoundSimplification)
11295 V = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, NewOps);
11299 case ISD::INSERT_SUBVECTOR: {
11300 SDValue BaseV = V->getOperand(0);
11301 SDValue SubV = V->getOperand(1);
11302 auto *IdxN = dyn_cast<ConstantSDNode>(V->getOperand(2));
11306 int SubSize = SubV.getValueType().getVectorNumElements();
11307 int Idx = IdxN->getZExtValue();
11308 bool SubVectorUsed = false;
11309 SmallBitVector SubUsedElements(SubSize, false);
11310 for (int i = 0; i < SubSize; ++i)
11311 if (UsedElements[i + Idx]) {
11312 SubVectorUsed = true;
11313 SubUsedElements[i] = true;
11314 UsedElements[i + Idx] = false;
11317 // Now recurse on both the base and sub vectors.
11318 SDValue SimplifiedSubV =
11320 ? simplifyShuffleOperandRecursively(SubUsedElements, SubV, DAG)
11321 : DAG.getUNDEF(SubV.getValueType());
11322 SDValue SimplifiedBaseV = simplifyShuffleOperandRecursively(UsedElements, BaseV, DAG);
11323 if (SimplifiedSubV != SubV || SimplifiedBaseV != BaseV)
11324 V = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT,
11325 SimplifiedBaseV, SimplifiedSubV, V->getOperand(2));
11331 static SDValue simplifyShuffleOperands(ShuffleVectorSDNode *SVN, SDValue N0,
11332 SDValue N1, SelectionDAG &DAG) {
11333 EVT VT = SVN->getValueType(0);
11334 int NumElts = VT.getVectorNumElements();
11335 SmallBitVector N0UsedElements(NumElts, false), N1UsedElements(NumElts, false);
11336 for (int M : SVN->getMask())
11337 if (M >= 0 && M < NumElts)
11338 N0UsedElements[M] = true;
11339 else if (M >= NumElts)
11340 N1UsedElements[M - NumElts] = true;
11342 SDValue S0 = simplifyShuffleOperandRecursively(N0UsedElements, N0, DAG);
11343 SDValue S1 = simplifyShuffleOperandRecursively(N1UsedElements, N1, DAG);
11344 if (S0 == N0 && S1 == N1)
11347 return DAG.getVectorShuffle(VT, SDLoc(SVN), S0, S1, SVN->getMask());
11350 // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat.
11351 static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
11352 EVT VT = N->getValueType(0);
11353 unsigned NumElts = VT.getVectorNumElements();
11355 SDValue N0 = N->getOperand(0);
11356 SDValue N1 = N->getOperand(1);
11357 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
11359 SmallVector<SDValue, 4> Ops;
11360 EVT ConcatVT = N0.getOperand(0).getValueType();
11361 unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements();
11362 unsigned NumConcats = NumElts / NumElemsPerConcat;
11364 // Look at every vector that's inserted. We're looking for exact
11365 // subvector-sized copies from a concatenated vector
11366 for (unsigned I = 0; I != NumConcats; ++I) {
11367 // Make sure we're dealing with a copy.
11368 unsigned Begin = I * NumElemsPerConcat;
11369 bool AllUndef = true, NoUndef = true;
11370 for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) {
11371 if (SVN->getMaskElt(J) >= 0)
11378 if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0)
11381 for (unsigned J = 1; J != NumElemsPerConcat; ++J)
11382 if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J))
11385 unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat;
11386 if (FirstElt < N0.getNumOperands())
11387 Ops.push_back(N0.getOperand(FirstElt));
11389 Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands()));
11391 } else if (AllUndef) {
11392 Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType()));
11393 } else { // Mixed with general masks and undefs, can't do optimization.
11398 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops);
11401 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
11402 EVT VT = N->getValueType(0);
11403 unsigned NumElts = VT.getVectorNumElements();
11405 SDValue N0 = N->getOperand(0);
11406 SDValue N1 = N->getOperand(1);
11408 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
11410 // Canonicalize shuffle undef, undef -> undef
11411 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
11412 return DAG.getUNDEF(VT);
11414 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
11416 // Canonicalize shuffle v, v -> v, undef
11418 SmallVector<int, 8> NewMask;
11419 for (unsigned i = 0; i != NumElts; ++i) {
11420 int Idx = SVN->getMaskElt(i);
11421 if (Idx >= (int)NumElts) Idx -= NumElts;
11422 NewMask.push_back(Idx);
11424 return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT),
11428 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
11429 if (N0.getOpcode() == ISD::UNDEF) {
11430 SmallVector<int, 8> NewMask;
11431 for (unsigned i = 0; i != NumElts; ++i) {
11432 int Idx = SVN->getMaskElt(i);
11434 if (Idx >= (int)NumElts)
11437 Idx = -1; // remove reference to lhs
11439 NewMask.push_back(Idx);
11441 return DAG.getVectorShuffle(VT, SDLoc(N), N1, DAG.getUNDEF(VT),
11445 // Remove references to rhs if it is undef
11446 if (N1.getOpcode() == ISD::UNDEF) {
11447 bool Changed = false;
11448 SmallVector<int, 8> NewMask;
11449 for (unsigned i = 0; i != NumElts; ++i) {
11450 int Idx = SVN->getMaskElt(i);
11451 if (Idx >= (int)NumElts) {
11455 NewMask.push_back(Idx);
11458 return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, &NewMask[0]);
11461 // If it is a splat, check if the argument vector is another splat or a
11462 // build_vector with all scalar elements the same.
11463 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
11464 SDNode *V = N0.getNode();
11466 // If this is a bit convert that changes the element type of the vector but
11467 // not the number of vector elements, look through it. Be careful not to
11468 // look though conversions that change things like v4f32 to v2f64.
11469 if (V->getOpcode() == ISD::BITCAST) {
11470 SDValue ConvInput = V->getOperand(0);
11471 if (ConvInput.getValueType().isVector() &&
11472 ConvInput.getValueType().getVectorNumElements() == NumElts)
11473 V = ConvInput.getNode();
11476 if (V->getOpcode() == ISD::BUILD_VECTOR) {
11477 assert(V->getNumOperands() == NumElts &&
11478 "BUILD_VECTOR has wrong number of operands");
11480 bool AllSame = true;
11481 for (unsigned i = 0; i != NumElts; ++i) {
11482 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
11483 Base = V->getOperand(i);
11487 // Splat of <u, u, u, u>, return <u, u, u, u>
11488 if (!Base.getNode())
11490 for (unsigned i = 0; i != NumElts; ++i) {
11491 if (V->getOperand(i) != Base) {
11496 // Splat of <x, x, x, x>, return <x, x, x, x>
11502 // There are various patterns used to build up a vector from smaller vectors,
11503 // subvectors, or elements. Scan chains of these and replace unused insertions
11504 // or components with undef.
11505 if (SDValue S = simplifyShuffleOperands(SVN, N0, N1, DAG))
11508 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
11509 Level < AfterLegalizeVectorOps &&
11510 (N1.getOpcode() == ISD::UNDEF ||
11511 (N1.getOpcode() == ISD::CONCAT_VECTORS &&
11512 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) {
11513 SDValue V = partitionShuffleOfConcats(N, DAG);
11519 // Canonicalize shuffles according to rules:
11520 // shuffle(A, shuffle(A, B)) -> shuffle(shuffle(A,B), A)
11521 // shuffle(B, shuffle(A, B)) -> shuffle(shuffle(A,B), B)
11522 // shuffle(B, shuffle(A, Undef)) -> shuffle(shuffle(A, Undef), B)
11523 if (N1.getOpcode() == ISD::VECTOR_SHUFFLE &&
11524 N0.getOpcode() != ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
11525 TLI.isTypeLegal(VT)) {
11526 // The incoming shuffle must be of the same type as the result of the
11527 // current shuffle.
11528 assert(N1->getOperand(0).getValueType() == VT &&
11529 "Shuffle types don't match");
11531 SDValue SV0 = N1->getOperand(0);
11532 SDValue SV1 = N1->getOperand(1);
11533 bool HasSameOp0 = N0 == SV0;
11534 bool IsSV1Undef = SV1.getOpcode() == ISD::UNDEF;
11535 if (HasSameOp0 || IsSV1Undef || N0 == SV1)
11536 // Commute the operands of this shuffle so that next rule
11538 return DAG.getCommutedVectorShuffle(*SVN);
11541 // Try to fold according to rules:
11542 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2)
11543 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2)
11544 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2)
11545 // Don't try to fold shuffles with illegal type.
11546 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
11547 TLI.isTypeLegal(VT)) {
11548 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
11550 // The incoming shuffle must be of the same type as the result of the
11551 // current shuffle.
11552 assert(OtherSV->getOperand(0).getValueType() == VT &&
11553 "Shuffle types don't match");
11556 SmallVector<int, 4> Mask;
11557 // Compute the combined shuffle mask for a shuffle with SV0 as the first
11558 // operand, and SV1 as the second operand.
11559 for (unsigned i = 0; i != NumElts; ++i) {
11560 int Idx = SVN->getMaskElt(i);
11562 // Propagate Undef.
11563 Mask.push_back(Idx);
11567 SDValue CurrentVec;
11568 if (Idx < (int)NumElts) {
11569 // This shuffle index refers to the inner shuffle N0. Lookup the inner
11570 // shuffle mask to identify which vector is actually referenced.
11571 Idx = OtherSV->getMaskElt(Idx);
11573 // Propagate Undef.
11574 Mask.push_back(Idx);
11578 CurrentVec = (Idx < (int) NumElts) ? OtherSV->getOperand(0)
11579 : OtherSV->getOperand(1);
11581 // This shuffle index references an element within N1.
11585 // Simple case where 'CurrentVec' is UNDEF.
11586 if (CurrentVec.getOpcode() == ISD::UNDEF) {
11587 Mask.push_back(-1);
11591 // Canonicalize the shuffle index. We don't know yet if CurrentVec
11592 // will be the first or second operand of the combined shuffle.
11593 Idx = Idx % NumElts;
11594 if (!SV0.getNode() || SV0 == CurrentVec) {
11595 // Ok. CurrentVec is the left hand side.
11596 // Update the mask accordingly.
11598 Mask.push_back(Idx);
11602 // Bail out if we cannot convert the shuffle pair into a single shuffle.
11603 if (SV1.getNode() && SV1 != CurrentVec)
11606 // Ok. CurrentVec is the right hand side.
11607 // Update the mask accordingly.
11609 Mask.push_back(Idx + NumElts);
11612 // Check if all indices in Mask are Undef. In case, propagate Undef.
11613 bool isUndefMask = true;
11614 for (unsigned i = 0; i != NumElts && isUndefMask; ++i)
11615 isUndefMask &= Mask[i] < 0;
11618 return DAG.getUNDEF(VT);
11620 if (!SV0.getNode())
11621 SV0 = DAG.getUNDEF(VT);
11622 if (!SV1.getNode())
11623 SV1 = DAG.getUNDEF(VT);
11625 // Avoid introducing shuffles with illegal mask.
11626 if (!TLI.isShuffleMaskLegal(Mask, VT)) {
11627 // Compute the commuted shuffle mask and test again.
11628 for (unsigned i = 0; i != NumElts; ++i) {
11632 else if (idx < (int)NumElts)
11633 Mask[i] = idx + NumElts;
11635 Mask[i] = idx - NumElts;
11638 if (!TLI.isShuffleMaskLegal(Mask, VT))
11641 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, A, M2)
11642 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, A, M2)
11643 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(C, B, M2)
11644 std::swap(SV0, SV1);
11647 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, B, M2)
11648 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(A, C, M2)
11649 // shuffle(shuffle(A, B, M0), C, M1) -> shuffle(B, C, M2)
11650 return DAG.getVectorShuffle(VT, SDLoc(N), SV0, SV1, &Mask[0]);
11656 SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
11657 SDValue N0 = N->getOperand(0);
11658 SDValue N2 = N->getOperand(2);
11660 // If the input vector is a concatenation, and the insert replaces
11661 // one of the halves, we can optimize into a single concat_vectors.
11662 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
11663 N0->getNumOperands() == 2 && N2.getOpcode() == ISD::Constant) {
11664 APInt InsIdx = cast<ConstantSDNode>(N2)->getAPIntValue();
11665 EVT VT = N->getValueType(0);
11667 // Lower half: fold (insert_subvector (concat_vectors X, Y), Z) ->
11668 // (concat_vectors Z, Y)
11670 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
11671 N->getOperand(1), N0.getOperand(1));
11673 // Upper half: fold (insert_subvector (concat_vectors X, Y), Z) ->
11674 // (concat_vectors X, Z)
11675 if (InsIdx == VT.getVectorNumElements()/2)
11676 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
11677 N0.getOperand(0), N->getOperand(1));
11683 /// Returns a vector_shuffle if it able to transform an AND to a vector_shuffle
11684 /// with the destination vector and a zero vector.
11685 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
11686 /// vector_shuffle V, Zero, <0, 4, 2, 4>
11687 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
11688 EVT VT = N->getValueType(0);
11690 SDValue LHS = N->getOperand(0);
11691 SDValue RHS = N->getOperand(1);
11692 if (N->getOpcode() == ISD::AND) {
11693 if (RHS.getOpcode() == ISD::BITCAST)
11694 RHS = RHS.getOperand(0);
11695 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
11696 SmallVector<int, 8> Indices;
11697 unsigned NumElts = RHS.getNumOperands();
11698 for (unsigned i = 0; i != NumElts; ++i) {
11699 SDValue Elt = RHS.getOperand(i);
11700 if (!isa<ConstantSDNode>(Elt))
11703 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
11704 Indices.push_back(i);
11705 else if (cast<ConstantSDNode>(Elt)->isNullValue())
11706 Indices.push_back(NumElts+i);
11711 // Let's see if the target supports this vector_shuffle.
11712 EVT RVT = RHS.getValueType();
11713 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
11716 // Return the new VECTOR_SHUFFLE node.
11717 EVT EltVT = RVT.getVectorElementType();
11718 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
11719 DAG.getConstant(0, EltVT));
11720 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), RVT, ZeroOps);
11721 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
11722 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
11723 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
11730 /// Visit a binary vector operation, like ADD.
11731 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
11732 assert(N->getValueType(0).isVector() &&
11733 "SimplifyVBinOp only works on vectors!");
11735 SDValue LHS = N->getOperand(0);
11736 SDValue RHS = N->getOperand(1);
11737 SDValue Shuffle = XformToShuffleWithZero(N);
11738 if (Shuffle.getNode()) return Shuffle;
11740 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
11742 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
11743 RHS.getOpcode() == ISD::BUILD_VECTOR) {
11744 // Check if both vectors are constants. If not bail out.
11745 if (!(cast<BuildVectorSDNode>(LHS)->isConstant() &&
11746 cast<BuildVectorSDNode>(RHS)->isConstant()))
11749 SmallVector<SDValue, 8> Ops;
11750 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
11751 SDValue LHSOp = LHS.getOperand(i);
11752 SDValue RHSOp = RHS.getOperand(i);
11754 // Can't fold divide by zero.
11755 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
11756 N->getOpcode() == ISD::FDIV) {
11757 if ((RHSOp.getOpcode() == ISD::Constant &&
11758 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
11759 (RHSOp.getOpcode() == ISD::ConstantFP &&
11760 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
11764 EVT VT = LHSOp.getValueType();
11765 EVT RVT = RHSOp.getValueType();
11767 // Integer BUILD_VECTOR operands may have types larger than the element
11768 // size (e.g., when the element type is not legal). Prior to type
11769 // legalization, the types may not match between the two BUILD_VECTORS.
11770 // Truncate one of the operands to make them match.
11771 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
11772 RHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, RHSOp);
11774 LHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), RVT, LHSOp);
11778 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(LHS), VT,
11780 if (FoldOp.getOpcode() != ISD::UNDEF &&
11781 FoldOp.getOpcode() != ISD::Constant &&
11782 FoldOp.getOpcode() != ISD::ConstantFP)
11784 Ops.push_back(FoldOp);
11785 AddToWorklist(FoldOp.getNode());
11788 if (Ops.size() == LHS.getNumOperands())
11789 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), LHS.getValueType(), Ops);
11792 // Type legalization might introduce new shuffles in the DAG.
11793 // Fold (VBinOp (shuffle (A, Undef, Mask)), (shuffle (B, Undef, Mask)))
11794 // -> (shuffle (VBinOp (A, B)), Undef, Mask).
11795 if (LegalTypes && isa<ShuffleVectorSDNode>(LHS) &&
11796 isa<ShuffleVectorSDNode>(RHS) && LHS.hasOneUse() && RHS.hasOneUse() &&
11797 LHS.getOperand(1).getOpcode() == ISD::UNDEF &&
11798 RHS.getOperand(1).getOpcode() == ISD::UNDEF) {
11799 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(LHS);
11800 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(RHS);
11802 if (SVN0->getMask().equals(SVN1->getMask())) {
11803 EVT VT = N->getValueType(0);
11804 SDValue UndefVector = LHS.getOperand(1);
11805 SDValue NewBinOp = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
11806 LHS.getOperand(0), RHS.getOperand(0));
11807 AddUsersToWorklist(N);
11808 return DAG.getVectorShuffle(VT, SDLoc(N), NewBinOp, UndefVector,
11809 &SVN0->getMask()[0]);
11816 /// Visit a binary vector operation, like FABS/FNEG.
11817 SDValue DAGCombiner::SimplifyVUnaryOp(SDNode *N) {
11818 assert(N->getValueType(0).isVector() &&
11819 "SimplifyVUnaryOp only works on vectors!");
11821 SDValue N0 = N->getOperand(0);
11823 if (N0.getOpcode() != ISD::BUILD_VECTOR)
11826 // Operand is a BUILD_VECTOR node, see if we can constant fold it.
11827 SmallVector<SDValue, 8> Ops;
11828 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
11829 SDValue Op = N0.getOperand(i);
11830 if (Op.getOpcode() != ISD::UNDEF &&
11831 Op.getOpcode() != ISD::ConstantFP)
11833 EVT EltVT = Op.getValueType();
11834 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(N0), EltVT, Op);
11835 if (FoldOp.getOpcode() != ISD::UNDEF &&
11836 FoldOp.getOpcode() != ISD::ConstantFP)
11838 Ops.push_back(FoldOp);
11839 AddToWorklist(FoldOp.getNode());
11842 if (Ops.size() != N0.getNumOperands())
11845 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), N0.getValueType(), Ops);
11848 SDValue DAGCombiner::SimplifySelect(SDLoc DL, SDValue N0,
11849 SDValue N1, SDValue N2){
11850 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
11852 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
11853 cast<CondCodeSDNode>(N0.getOperand(2))->get());
11855 // If we got a simplified select_cc node back from SimplifySelectCC, then
11856 // break it down into a new SETCC node, and a new SELECT node, and then return
11857 // the SELECT node, since we were called with a SELECT node.
11858 if (SCC.getNode()) {
11859 // Check to see if we got a select_cc back (to turn into setcc/select).
11860 // Otherwise, just return whatever node we got back, like fabs.
11861 if (SCC.getOpcode() == ISD::SELECT_CC) {
11862 SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0),
11864 SCC.getOperand(0), SCC.getOperand(1),
11865 SCC.getOperand(4));
11866 AddToWorklist(SETCC.getNode());
11867 return DAG.getSelect(SDLoc(SCC), SCC.getValueType(), SETCC,
11868 SCC.getOperand(2), SCC.getOperand(3));
11876 /// Given a SELECT or a SELECT_CC node, where LHS and RHS are the two values
11877 /// being selected between, see if we can simplify the select. Callers of this
11878 /// should assume that TheSelect is deleted if this returns true. As such, they
11879 /// should return the appropriate thing (e.g. the node) back to the top-level of
11880 /// the DAG combiner loop to avoid it being looked at.
11881 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
11884 // Cannot simplify select with vector condition
11885 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
11887 // If this is a select from two identical things, try to pull the operation
11888 // through the select.
11889 if (LHS.getOpcode() != RHS.getOpcode() ||
11890 !LHS.hasOneUse() || !RHS.hasOneUse())
11893 // If this is a load and the token chain is identical, replace the select
11894 // of two loads with a load through a select of the address to load from.
11895 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
11896 // constants have been dropped into the constant pool.
11897 if (LHS.getOpcode() == ISD::LOAD) {
11898 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
11899 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
11901 // Token chains must be identical.
11902 if (LHS.getOperand(0) != RHS.getOperand(0) ||
11903 // Do not let this transformation reduce the number of volatile loads.
11904 LLD->isVolatile() || RLD->isVolatile() ||
11905 // If this is an EXTLOAD, the VT's must match.
11906 LLD->getMemoryVT() != RLD->getMemoryVT() ||
11907 // If this is an EXTLOAD, the kind of extension must match.
11908 (LLD->getExtensionType() != RLD->getExtensionType() &&
11909 // The only exception is if one of the extensions is anyext.
11910 LLD->getExtensionType() != ISD::EXTLOAD &&
11911 RLD->getExtensionType() != ISD::EXTLOAD) ||
11912 // FIXME: this discards src value information. This is
11913 // over-conservative. It would be beneficial to be able to remember
11914 // both potential memory locations. Since we are discarding
11915 // src value info, don't do the transformation if the memory
11916 // locations are not in the default address space.
11917 LLD->getPointerInfo().getAddrSpace() != 0 ||
11918 RLD->getPointerInfo().getAddrSpace() != 0 ||
11919 !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(),
11920 LLD->getBasePtr().getValueType()))
11923 // Check that the select condition doesn't reach either load. If so,
11924 // folding this will induce a cycle into the DAG. If not, this is safe to
11925 // xform, so create a select of the addresses.
11927 if (TheSelect->getOpcode() == ISD::SELECT) {
11928 SDNode *CondNode = TheSelect->getOperand(0).getNode();
11929 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
11930 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
11932 // The loads must not depend on one another.
11933 if (LLD->isPredecessorOf(RLD) ||
11934 RLD->isPredecessorOf(LLD))
11936 Addr = DAG.getSelect(SDLoc(TheSelect),
11937 LLD->getBasePtr().getValueType(),
11938 TheSelect->getOperand(0), LLD->getBasePtr(),
11939 RLD->getBasePtr());
11940 } else { // Otherwise SELECT_CC
11941 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
11942 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
11944 if ((LLD->hasAnyUseOfValue(1) &&
11945 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
11946 (RLD->hasAnyUseOfValue(1) &&
11947 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
11950 Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect),
11951 LLD->getBasePtr().getValueType(),
11952 TheSelect->getOperand(0),
11953 TheSelect->getOperand(1),
11954 LLD->getBasePtr(), RLD->getBasePtr(),
11955 TheSelect->getOperand(4));
11959 // It is safe to replace the two loads if they have different alignments,
11960 // but the new load must be the minimum (most restrictive) alignment of the
11962 bool isInvariant = LLD->isInvariant() & RLD->isInvariant();
11963 unsigned Alignment = std::min(LLD->getAlignment(), RLD->getAlignment());
11964 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
11965 Load = DAG.getLoad(TheSelect->getValueType(0),
11967 // FIXME: Discards pointer and AA info.
11968 LLD->getChain(), Addr, MachinePointerInfo(),
11969 LLD->isVolatile(), LLD->isNonTemporal(),
11970 isInvariant, Alignment);
11972 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
11973 RLD->getExtensionType() : LLD->getExtensionType(),
11975 TheSelect->getValueType(0),
11976 // FIXME: Discards pointer and AA info.
11977 LLD->getChain(), Addr, MachinePointerInfo(),
11978 LLD->getMemoryVT(), LLD->isVolatile(),
11979 LLD->isNonTemporal(), isInvariant, Alignment);
11982 // Users of the select now use the result of the load.
11983 CombineTo(TheSelect, Load);
11985 // Users of the old loads now use the new load's chain. We know the
11986 // old-load value is dead now.
11987 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
11988 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
11995 /// Simplify an expression of the form (N0 cond N1) ? N2 : N3
11996 /// where 'cond' is the comparison specified by CC.
11997 SDValue DAGCombiner::SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1,
11998 SDValue N2, SDValue N3,
11999 ISD::CondCode CC, bool NotExtCompare) {
12000 // (x ? y : y) -> y.
12001 if (N2 == N3) return N2;
12003 EVT VT = N2.getValueType();
12004 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
12005 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
12006 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
12008 // Determine if the condition we're dealing with is constant
12009 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
12010 N0, N1, CC, DL, false);
12011 if (SCC.getNode()) AddToWorklist(SCC.getNode());
12012 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
12014 // fold select_cc true, x, y -> x
12015 if (SCCC && !SCCC->isNullValue())
12017 // fold select_cc false, x, y -> y
12018 if (SCCC && SCCC->isNullValue())
12021 // Check to see if we can simplify the select into an fabs node
12022 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
12023 // Allow either -0.0 or 0.0
12024 if (CFP->getValueAPF().isZero()) {
12025 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
12026 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
12027 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
12028 N2 == N3.getOperand(0))
12029 return DAG.getNode(ISD::FABS, DL, VT, N0);
12031 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
12032 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
12033 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
12034 N2.getOperand(0) == N3)
12035 return DAG.getNode(ISD::FABS, DL, VT, N3);
12039 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
12040 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
12041 // in it. This is a win when the constant is not otherwise available because
12042 // it replaces two constant pool loads with one. We only do this if the FP
12043 // type is known to be legal, because if it isn't, then we are before legalize
12044 // types an we want the other legalization to happen first (e.g. to avoid
12045 // messing with soft float) and if the ConstantFP is not legal, because if
12046 // it is legal, we may not need to store the FP constant in a constant pool.
12047 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
12048 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
12049 if (TLI.isTypeLegal(N2.getValueType()) &&
12050 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
12051 TargetLowering::Legal &&
12052 !TLI.isFPImmLegal(TV->getValueAPF(), TV->getValueType(0)) &&
12053 !TLI.isFPImmLegal(FV->getValueAPF(), FV->getValueType(0))) &&
12054 // If both constants have multiple uses, then we won't need to do an
12055 // extra load, they are likely around in registers for other users.
12056 (TV->hasOneUse() || FV->hasOneUse())) {
12057 Constant *Elts[] = {
12058 const_cast<ConstantFP*>(FV->getConstantFPValue()),
12059 const_cast<ConstantFP*>(TV->getConstantFPValue())
12061 Type *FPTy = Elts[0]->getType();
12062 const DataLayout &TD = *TLI.getDataLayout();
12064 // Create a ConstantArray of the two constants.
12065 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
12066 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
12067 TD.getPrefTypeAlignment(FPTy));
12068 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
12070 // Get the offsets to the 0 and 1 element of the array so that we can
12071 // select between them.
12072 SDValue Zero = DAG.getIntPtrConstant(0);
12073 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
12074 SDValue One = DAG.getIntPtrConstant(EltSize);
12076 SDValue Cond = DAG.getSetCC(DL,
12077 getSetCCResultType(N0.getValueType()),
12079 AddToWorklist(Cond.getNode());
12080 SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(),
12082 AddToWorklist(CstOffset.getNode());
12083 CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx,
12085 AddToWorklist(CPIdx.getNode());
12086 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
12087 MachinePointerInfo::getConstantPool(), false,
12088 false, false, Alignment);
12093 // Check to see if we can perform the "gzip trick", transforming
12094 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
12095 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
12096 (N1C->isNullValue() || // (a < 0) ? b : 0
12097 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
12098 EVT XType = N0.getValueType();
12099 EVT AType = N2.getValueType();
12100 if (XType.bitsGE(AType)) {
12101 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
12102 // single-bit constant.
12103 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
12104 unsigned ShCtV = N2C->getAPIntValue().logBase2();
12105 ShCtV = XType.getSizeInBits()-ShCtV-1;
12106 SDValue ShCt = DAG.getConstant(ShCtV,
12107 getShiftAmountTy(N0.getValueType()));
12108 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0),
12110 AddToWorklist(Shift.getNode());
12112 if (XType.bitsGT(AType)) {
12113 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
12114 AddToWorklist(Shift.getNode());
12117 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
12120 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0),
12122 DAG.getConstant(XType.getSizeInBits()-1,
12123 getShiftAmountTy(N0.getValueType())));
12124 AddToWorklist(Shift.getNode());
12126 if (XType.bitsGT(AType)) {
12127 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
12128 AddToWorklist(Shift.getNode());
12131 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
12135 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
12136 // where y is has a single bit set.
12137 // A plaintext description would be, we can turn the SELECT_CC into an AND
12138 // when the condition can be materialized as an all-ones register. Any
12139 // single bit-test can be materialized as an all-ones register with
12140 // shift-left and shift-right-arith.
12141 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
12142 N0->getValueType(0) == VT &&
12143 N1C && N1C->isNullValue() &&
12144 N2C && N2C->isNullValue()) {
12145 SDValue AndLHS = N0->getOperand(0);
12146 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
12147 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
12148 // Shift the tested bit over the sign bit.
12149 APInt AndMask = ConstAndRHS->getAPIntValue();
12151 DAG.getConstant(AndMask.countLeadingZeros(),
12152 getShiftAmountTy(AndLHS.getValueType()));
12153 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt);
12155 // Now arithmetic right shift it all the way over, so the result is either
12156 // all-ones, or zero.
12158 DAG.getConstant(AndMask.getBitWidth()-1,
12159 getShiftAmountTy(Shl.getValueType()));
12160 SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt);
12162 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
12166 // fold select C, 16, 0 -> shl C, 4
12167 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
12168 TLI.getBooleanContents(N0.getValueType()) ==
12169 TargetLowering::ZeroOrOneBooleanContent) {
12171 // If the caller doesn't want us to simplify this into a zext of a compare,
12173 if (NotExtCompare && N2C->getAPIntValue() == 1)
12176 // Get a SetCC of the condition
12177 // NOTE: Don't create a SETCC if it's not legal on this target.
12178 if (!LegalOperations ||
12179 TLI.isOperationLegal(ISD::SETCC,
12180 LegalTypes ? getSetCCResultType(N0.getValueType()) : MVT::i1)) {
12182 // cast from setcc result type to select result type
12184 SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()),
12186 if (N2.getValueType().bitsLT(SCC.getValueType()))
12187 Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2),
12188 N2.getValueType());
12190 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
12191 N2.getValueType(), SCC);
12193 SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC);
12194 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
12195 N2.getValueType(), SCC);
12198 AddToWorklist(SCC.getNode());
12199 AddToWorklist(Temp.getNode());
12201 if (N2C->getAPIntValue() == 1)
12204 // shl setcc result by log2 n2c
12205 return DAG.getNode(
12206 ISD::SHL, DL, N2.getValueType(), Temp,
12207 DAG.getConstant(N2C->getAPIntValue().logBase2(),
12208 getShiftAmountTy(Temp.getValueType())));
12212 // Check to see if this is the equivalent of setcc
12213 // FIXME: Turn all of these into setcc if setcc if setcc is legal
12214 // otherwise, go ahead with the folds.
12215 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
12216 EVT XType = N0.getValueType();
12217 if (!LegalOperations ||
12218 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(XType))) {
12219 SDValue Res = DAG.getSetCC(DL, getSetCCResultType(XType), N0, N1, CC);
12220 if (Res.getValueType() != VT)
12221 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
12225 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
12226 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
12227 (!LegalOperations ||
12228 TLI.isOperationLegal(ISD::CTLZ, XType))) {
12229 SDValue Ctlz = DAG.getNode(ISD::CTLZ, SDLoc(N0), XType, N0);
12230 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
12231 DAG.getConstant(Log2_32(XType.getSizeInBits()),
12232 getShiftAmountTy(Ctlz.getValueType())));
12234 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
12235 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
12236 SDValue NegN0 = DAG.getNode(ISD::SUB, SDLoc(N0),
12237 XType, DAG.getConstant(0, XType), N0);
12238 SDValue NotN0 = DAG.getNOT(SDLoc(N0), N0, XType);
12239 return DAG.getNode(ISD::SRL, DL, XType,
12240 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
12241 DAG.getConstant(XType.getSizeInBits()-1,
12242 getShiftAmountTy(XType)));
12244 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
12245 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
12246 SDValue Sign = DAG.getNode(ISD::SRL, SDLoc(N0), XType, N0,
12247 DAG.getConstant(XType.getSizeInBits()-1,
12248 getShiftAmountTy(N0.getValueType())));
12249 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
12253 // Check to see if this is an integer abs.
12254 // select_cc setg[te] X, 0, X, -X ->
12255 // select_cc setgt X, -1, X, -X ->
12256 // select_cc setl[te] X, 0, -X, X ->
12257 // select_cc setlt X, 1, -X, X ->
12258 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
12260 ConstantSDNode *SubC = nullptr;
12261 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
12262 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
12263 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
12264 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
12265 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
12266 (N1C->isOne() && CC == ISD::SETLT)) &&
12267 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
12268 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
12270 EVT XType = N0.getValueType();
12271 if (SubC && SubC->isNullValue() && XType.isInteger()) {
12272 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0), XType,
12274 DAG.getConstant(XType.getSizeInBits()-1,
12275 getShiftAmountTy(N0.getValueType())));
12276 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0),
12278 AddToWorklist(Shift.getNode());
12279 AddToWorklist(Add.getNode());
12280 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
12287 /// This is a stub for TargetLowering::SimplifySetCC.
12288 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
12289 SDValue N1, ISD::CondCode Cond,
12290 SDLoc DL, bool foldBooleans) {
12291 TargetLowering::DAGCombinerInfo
12292 DagCombineInfo(DAG, Level, false, this);
12293 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
12296 /// Given an ISD::SDIV node expressing a divide by constant, return
12297 /// a DAG expression to select that will generate the same value by multiplying
12298 /// by a magic number.
12299 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
12300 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
12301 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
12305 // Avoid division by zero.
12306 if (!C->getAPIntValue())
12309 std::vector<SDNode*> Built;
12311 TLI.BuildSDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
12313 for (SDNode *N : Built)
12318 /// Given an ISD::SDIV node expressing a divide by constant power of 2, return a
12319 /// DAG expression that will generate the same value by right shifting.
12320 SDValue DAGCombiner::BuildSDIVPow2(SDNode *N) {
12321 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
12325 // Avoid division by zero.
12326 if (!C->getAPIntValue())
12329 std::vector<SDNode *> Built;
12330 SDValue S = TLI.BuildSDIVPow2(N, C->getAPIntValue(), DAG, &Built);
12332 for (SDNode *N : Built)
12337 /// Given an ISD::UDIV node expressing a divide by constant, return a DAG
12338 /// expression that will generate the same value by multiplying by a magic
12340 /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide".
12341 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
12342 ConstantSDNode *C = isConstOrConstSplat(N->getOperand(1));
12346 // Avoid division by zero.
12347 if (!C->getAPIntValue())
12350 std::vector<SDNode*> Built;
12352 TLI.BuildUDIV(N, C->getAPIntValue(), DAG, LegalOperations, &Built);
12354 for (SDNode *N : Built)
12359 SDValue DAGCombiner::BuildReciprocalEstimate(SDValue Op) {
12360 if (Level >= AfterLegalizeDAG)
12363 // Expose the DAG combiner to the target combiner implementations.
12364 TargetLowering::DAGCombinerInfo DCI(DAG, Level, false, this);
12366 unsigned Iterations = 0;
12367 if (SDValue Est = TLI.getRecipEstimate(Op, DCI, Iterations)) {
12369 // Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
12370 // For the reciprocal, we need to find the zero of the function:
12371 // F(X) = A X - 1 [which has a zero at X = 1/A]
12373 // X_{i+1} = X_i (2 - A X_i) = X_i + X_i (1 - A X_i) [this second form
12374 // does not require additional intermediate precision]
12375 EVT VT = Op.getValueType();
12377 SDValue FPOne = DAG.getConstantFP(1.0, VT);
12379 AddToWorklist(Est.getNode());
12381 // Newton iterations: Est = Est + Est (1 - Arg * Est)
12382 for (unsigned i = 0; i < Iterations; ++i) {
12383 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Op, Est);
12384 AddToWorklist(NewEst.getNode());
12386 NewEst = DAG.getNode(ISD::FSUB, DL, VT, FPOne, NewEst);
12387 AddToWorklist(NewEst.getNode());
12389 NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst);
12390 AddToWorklist(NewEst.getNode());
12392 Est = DAG.getNode(ISD::FADD, DL, VT, Est, NewEst);
12393 AddToWorklist(Est.getNode());
12402 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
12403 /// For the reciprocal sqrt, we need to find the zero of the function:
12404 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
12406 /// X_{i+1} = X_i (1.5 - A X_i^2 / 2)
12407 /// As a result, we precompute A/2 prior to the iteration loop.
12408 SDValue DAGCombiner::BuildRsqrtNROneConst(SDValue Arg, SDValue Est,
12409 unsigned Iterations) {
12410 EVT VT = Arg.getValueType();
12412 SDValue ThreeHalves = DAG.getConstantFP(1.5, VT);
12414 // We now need 0.5 * Arg which we can write as (1.5 * Arg - Arg) so that
12415 // this entire sequence requires only one FP constant.
12416 SDValue HalfArg = DAG.getNode(ISD::FMUL, DL, VT, ThreeHalves, Arg);
12417 AddToWorklist(HalfArg.getNode());
12419 HalfArg = DAG.getNode(ISD::FSUB, DL, VT, HalfArg, Arg);
12420 AddToWorklist(HalfArg.getNode());
12422 // Newton iterations: Est = Est * (1.5 - HalfArg * Est * Est)
12423 for (unsigned i = 0; i < Iterations; ++i) {
12424 SDValue NewEst = DAG.getNode(ISD::FMUL, DL, VT, Est, Est);
12425 AddToWorklist(NewEst.getNode());
12427 NewEst = DAG.getNode(ISD::FMUL, DL, VT, HalfArg, NewEst);
12428 AddToWorklist(NewEst.getNode());
12430 NewEst = DAG.getNode(ISD::FSUB, DL, VT, ThreeHalves, NewEst);
12431 AddToWorklist(NewEst.getNode());
12433 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, NewEst);
12434 AddToWorklist(Est.getNode());
12439 /// Newton iteration for a function: F(X) is X_{i+1} = X_i - F(X_i)/F'(X_i)
12440 /// For the reciprocal sqrt, we need to find the zero of the function:
12441 /// F(X) = 1/X^2 - A [which has a zero at X = 1/sqrt(A)]
12443 /// X_{i+1} = (-0.5 * X_i) * (A * X_i * X_i + (-3.0))
12444 SDValue DAGCombiner::BuildRsqrtNRTwoConst(SDValue Arg, SDValue Est,
12445 unsigned Iterations) {
12446 EVT VT = Arg.getValueType();
12448 SDValue MinusThree = DAG.getConstantFP(-3.0, VT);
12449 SDValue MinusHalf = DAG.getConstantFP(-0.5, VT);
12451 // Newton iterations: Est = -0.5 * Est * (-3.0 + Arg * Est * Est)
12452 for (unsigned i = 0; i < Iterations; ++i) {
12453 SDValue HalfEst = DAG.getNode(ISD::FMUL, DL, VT, Est, MinusHalf);
12454 AddToWorklist(HalfEst.getNode());
12456 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Est);
12457 AddToWorklist(Est.getNode());
12459 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, Arg);
12460 AddToWorklist(Est.getNode());
12462 Est = DAG.getNode(ISD::FADD, DL, VT, Est, MinusThree);
12463 AddToWorklist(Est.getNode());
12465 Est = DAG.getNode(ISD::FMUL, DL, VT, Est, HalfEst);
12466 AddToWorklist(Est.getNode());
12471 SDValue DAGCombiner::BuildRsqrtEstimate(SDValue Op) {
12472 if (Level >= AfterLegalizeDAG)
12475 // Expose the DAG combiner to the target combiner implementations.
12476 TargetLowering::DAGCombinerInfo DCI(DAG, Level, false, this);
12477 unsigned Iterations = 0;
12478 bool UseOneConstNR = false;
12479 if (SDValue Est = TLI.getRsqrtEstimate(Op, DCI, Iterations, UseOneConstNR)) {
12480 AddToWorklist(Est.getNode());
12482 Est = UseOneConstNR ?
12483 BuildRsqrtNROneConst(Op, Est, Iterations) :
12484 BuildRsqrtNRTwoConst(Op, Est, Iterations);
12492 /// Return true if base is a frame index, which is known not to alias with
12493 /// anything but itself. Provides base object and offset as results.
12494 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
12495 const GlobalValue *&GV, const void *&CV) {
12496 // Assume it is a primitive operation.
12497 Base = Ptr; Offset = 0; GV = nullptr; CV = nullptr;
12499 // If it's an adding a simple constant then integrate the offset.
12500 if (Base.getOpcode() == ISD::ADD) {
12501 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
12502 Base = Base.getOperand(0);
12503 Offset += C->getZExtValue();
12507 // Return the underlying GlobalValue, and update the Offset. Return false
12508 // for GlobalAddressSDNode since the same GlobalAddress may be represented
12509 // by multiple nodes with different offsets.
12510 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
12511 GV = G->getGlobal();
12512 Offset += G->getOffset();
12516 // Return the underlying Constant value, and update the Offset. Return false
12517 // for ConstantSDNodes since the same constant pool entry may be represented
12518 // by multiple nodes with different offsets.
12519 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
12520 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
12521 : (const void *)C->getConstVal();
12522 Offset += C->getOffset();
12525 // If it's any of the following then it can't alias with anything but itself.
12526 return isa<FrameIndexSDNode>(Base);
12529 /// Return true if there is any possibility that the two addresses overlap.
12530 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) const {
12531 // If they are the same then they must be aliases.
12532 if (Op0->getBasePtr() == Op1->getBasePtr()) return true;
12534 // If they are both volatile then they cannot be reordered.
12535 if (Op0->isVolatile() && Op1->isVolatile()) return true;
12537 // Gather base node and offset information.
12538 SDValue Base1, Base2;
12539 int64_t Offset1, Offset2;
12540 const GlobalValue *GV1, *GV2;
12541 const void *CV1, *CV2;
12542 bool isFrameIndex1 = FindBaseOffset(Op0->getBasePtr(),
12543 Base1, Offset1, GV1, CV1);
12544 bool isFrameIndex2 = FindBaseOffset(Op1->getBasePtr(),
12545 Base2, Offset2, GV2, CV2);
12547 // If they have a same base address then check to see if they overlap.
12548 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
12549 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
12550 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
12552 // It is possible for different frame indices to alias each other, mostly
12553 // when tail call optimization reuses return address slots for arguments.
12554 // To catch this case, look up the actual index of frame indices to compute
12555 // the real alias relationship.
12556 if (isFrameIndex1 && isFrameIndex2) {
12557 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12558 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
12559 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
12560 return !((Offset1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= Offset2 ||
12561 (Offset2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= Offset1);
12564 // Otherwise, if we know what the bases are, and they aren't identical, then
12565 // we know they cannot alias.
12566 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
12569 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
12570 // compared to the size and offset of the access, we may be able to prove they
12571 // do not alias. This check is conservative for now to catch cases created by
12572 // splitting vector types.
12573 if ((Op0->getOriginalAlignment() == Op1->getOriginalAlignment()) &&
12574 (Op0->getSrcValueOffset() != Op1->getSrcValueOffset()) &&
12575 (Op0->getMemoryVT().getSizeInBits() >> 3 ==
12576 Op1->getMemoryVT().getSizeInBits() >> 3) &&
12577 (Op0->getOriginalAlignment() > Op0->getMemoryVT().getSizeInBits()) >> 3) {
12578 int64_t OffAlign1 = Op0->getSrcValueOffset() % Op0->getOriginalAlignment();
12579 int64_t OffAlign2 = Op1->getSrcValueOffset() % Op1->getOriginalAlignment();
12581 // There is no overlap between these relatively aligned accesses of similar
12582 // size, return no alias.
12583 if ((OffAlign1 + (Op0->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign2 ||
12584 (OffAlign2 + (Op1->getMemoryVT().getSizeInBits() >> 3)) <= OffAlign1)
12588 bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0
12590 : DAG.getSubtarget().useAA();
12592 if (CombinerAAOnlyFunc.getNumOccurrences() &&
12593 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
12597 Op0->getMemOperand()->getValue() && Op1->getMemOperand()->getValue()) {
12598 // Use alias analysis information.
12599 int64_t MinOffset = std::min(Op0->getSrcValueOffset(),
12600 Op1->getSrcValueOffset());
12601 int64_t Overlap1 = (Op0->getMemoryVT().getSizeInBits() >> 3) +
12602 Op0->getSrcValueOffset() - MinOffset;
12603 int64_t Overlap2 = (Op1->getMemoryVT().getSizeInBits() >> 3) +
12604 Op1->getSrcValueOffset() - MinOffset;
12605 AliasAnalysis::AliasResult AAResult =
12606 AA.alias(AliasAnalysis::Location(Op0->getMemOperand()->getValue(),
12608 UseTBAA ? Op0->getAAInfo() : AAMDNodes()),
12609 AliasAnalysis::Location(Op1->getMemOperand()->getValue(),
12611 UseTBAA ? Op1->getAAInfo() : AAMDNodes()));
12612 if (AAResult == AliasAnalysis::NoAlias)
12616 // Otherwise we have to assume they alias.
12620 /// Walk up chain skipping non-aliasing memory nodes,
12621 /// looking for aliasing nodes and adding them to the Aliases vector.
12622 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
12623 SmallVectorImpl<SDValue> &Aliases) {
12624 SmallVector<SDValue, 8> Chains; // List of chains to visit.
12625 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
12627 // Get alias information for node.
12628 bool IsLoad = isa<LoadSDNode>(N) && !cast<LSBaseSDNode>(N)->isVolatile();
12631 Chains.push_back(OriginalChain);
12632 unsigned Depth = 0;
12634 // Look at each chain and determine if it is an alias. If so, add it to the
12635 // aliases list. If not, then continue up the chain looking for the next
12637 while (!Chains.empty()) {
12638 SDValue Chain = Chains.back();
12641 // For TokenFactor nodes, look at each operand and only continue up the
12642 // chain until we find two aliases. If we've seen two aliases, assume we'll
12643 // find more and revert to original chain since the xform is unlikely to be
12646 // FIXME: The depth check could be made to return the last non-aliasing
12647 // chain we found before we hit a tokenfactor rather than the original
12649 if (Depth > 6 || Aliases.size() == 2) {
12651 Aliases.push_back(OriginalChain);
12655 // Don't bother if we've been before.
12656 if (!Visited.insert(Chain.getNode()).second)
12659 switch (Chain.getOpcode()) {
12660 case ISD::EntryToken:
12661 // Entry token is ideal chain operand, but handled in FindBetterChain.
12666 // Get alias information for Chain.
12667 bool IsOpLoad = isa<LoadSDNode>(Chain.getNode()) &&
12668 !cast<LSBaseSDNode>(Chain.getNode())->isVolatile();
12670 // If chain is alias then stop here.
12671 if (!(IsLoad && IsOpLoad) &&
12672 isAlias(cast<LSBaseSDNode>(N), cast<LSBaseSDNode>(Chain.getNode()))) {
12673 Aliases.push_back(Chain);
12675 // Look further up the chain.
12676 Chains.push_back(Chain.getOperand(0));
12682 case ISD::TokenFactor:
12683 // We have to check each of the operands of the token factor for "small"
12684 // token factors, so we queue them up. Adding the operands to the queue
12685 // (stack) in reverse order maintains the original order and increases the
12686 // likelihood that getNode will find a matching token factor (CSE.)
12687 if (Chain.getNumOperands() > 16) {
12688 Aliases.push_back(Chain);
12691 for (unsigned n = Chain.getNumOperands(); n;)
12692 Chains.push_back(Chain.getOperand(--n));
12697 // For all other instructions we will just have to take what we can get.
12698 Aliases.push_back(Chain);
12703 // We need to be careful here to also search for aliases through the
12704 // value operand of a store, etc. Consider the following situation:
12706 // L1 = load Token1, %52
12707 // S1 = store Token1, L1, %51
12708 // L2 = load Token1, %52+8
12709 // S2 = store Token1, L2, %51+8
12710 // Token2 = Token(S1, S2)
12711 // L3 = load Token2, %53
12712 // S3 = store Token2, L3, %52
12713 // L4 = load Token2, %53+8
12714 // S4 = store Token2, L4, %52+8
12715 // If we search for aliases of S3 (which loads address %52), and we look
12716 // only through the chain, then we'll miss the trivial dependence on L1
12717 // (which also loads from %52). We then might change all loads and
12718 // stores to use Token1 as their chain operand, which could result in
12719 // copying %53 into %52 before copying %52 into %51 (which should
12722 // The problem is, however, that searching for such data dependencies
12723 // can become expensive, and the cost is not directly related to the
12724 // chain depth. Instead, we'll rule out such configurations here by
12725 // insisting that we've visited all chain users (except for users
12726 // of the original chain, which is not necessary). When doing this,
12727 // we need to look through nodes we don't care about (otherwise, things
12728 // like register copies will interfere with trivial cases).
12730 SmallVector<const SDNode *, 16> Worklist;
12731 for (const SDNode *N : Visited)
12732 if (N != OriginalChain.getNode())
12733 Worklist.push_back(N);
12735 while (!Worklist.empty()) {
12736 const SDNode *M = Worklist.pop_back_val();
12738 // We have already visited M, and want to make sure we've visited any uses
12739 // of M that we care about. For uses that we've not visisted, and don't
12740 // care about, queue them to the worklist.
12742 for (SDNode::use_iterator UI = M->use_begin(),
12743 UIE = M->use_end(); UI != UIE; ++UI)
12744 if (UI.getUse().getValueType() == MVT::Other &&
12745 Visited.insert(*UI).second) {
12746 if (isa<MemIntrinsicSDNode>(*UI) || isa<MemSDNode>(*UI)) {
12747 // We've not visited this use, and we care about it (it could have an
12748 // ordering dependency with the original node).
12750 Aliases.push_back(OriginalChain);
12754 // We've not visited this use, but we don't care about it. Mark it as
12755 // visited and enqueue it to the worklist.
12756 Worklist.push_back(*UI);
12761 /// Walk up chain skipping non-aliasing memory nodes, looking for a better chain
12762 /// (aliasing node.)
12763 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
12764 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
12766 // Accumulate all the aliases to this node.
12767 GatherAllAliases(N, OldChain, Aliases);
12769 // If no operands then chain to entry token.
12770 if (Aliases.size() == 0)
12771 return DAG.getEntryNode();
12773 // If a single operand then chain to it. We don't need to revisit it.
12774 if (Aliases.size() == 1)
12777 // Construct a custom tailored token factor.
12778 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other, Aliases);
12781 /// This is the entry point for the file.
12782 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
12783 CodeGenOpt::Level OptLevel) {
12784 /// This is the main entry point to this class.
12785 DAGCombiner(*this, AA, OptLevel).Run(Level);