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 #define DEBUG_TYPE "dagcombine"
20 #include "llvm/CodeGen/SelectionDAG.h"
21 #include "llvm/ADT/SmallPtrSet.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/DerivedTypes.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/LLVMContext.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Target/TargetLowering.h"
36 #include "llvm/Target/TargetMachine.h"
37 #include "llvm/Target/TargetOptions.h"
38 #include "llvm/Target/TargetRegisterInfo.h"
39 #include "llvm/Target/TargetSubtargetInfo.h"
43 STATISTIC(NodesCombined , "Number of dag nodes combined");
44 STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created");
45 STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created");
46 STATISTIC(OpsNarrowed , "Number of load/op/store narrowed");
47 STATISTIC(LdStFP2Int , "Number of fp load/store pairs transformed to int");
48 STATISTIC(SlicedLoads, "Number of load sliced");
52 CombinerAA("combiner-alias-analysis", cl::Hidden,
53 cl::desc("Enable DAG combiner alias-analysis heuristics"));
56 CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden,
57 cl::desc("Enable DAG combiner's use of IR alias analysis"));
60 UseTBAA("combiner-use-tbaa", cl::Hidden, cl::init(true),
61 cl::desc("Enable DAG combiner's use of TBAA"));
64 static cl::opt<std::string>
65 CombinerAAOnlyFunc("combiner-aa-only-func", cl::Hidden,
66 cl::desc("Only use DAG-combiner alias analysis in this"
70 /// Hidden option to stress test load slicing, i.e., when this option
71 /// is enabled, load slicing bypasses most of its profitability guards.
73 StressLoadSlicing("combiner-stress-load-slicing", cl::Hidden,
74 cl::desc("Bypass the profitability model of load "
78 //------------------------------ DAGCombiner ---------------------------------//
82 const TargetLowering &TLI;
84 CodeGenOpt::Level OptLevel;
89 // Worklist of all of the nodes that need to be simplified.
91 // This has the semantics that when adding to the worklist,
92 // the item added must be next to be processed. It should
93 // also only appear once. The naive approach to this takes
96 // To reduce the insert/remove time to logarithmic, we use
97 // a set and a vector to maintain our worklist.
99 // The set contains the items on the worklist, but does not
100 // maintain the order they should be visited.
102 // The vector maintains the order nodes should be visited, but may
103 // contain duplicate or removed nodes. When choosing a node to
104 // visit, we pop off the order stack until we find an item that is
105 // also in the contents set. All operations are O(log N).
106 SmallPtrSet<SDNode*, 64> WorkListContents;
107 SmallVector<SDNode*, 64> WorkListOrder;
109 // AA - Used for DAG load/store alias analysis.
112 /// AddUsersToWorkList - When an instruction is simplified, add all users of
113 /// the instruction to the work lists because they might get more simplified
116 void AddUsersToWorkList(SDNode *N) {
117 for (SDNode *Node : N->uses())
121 /// visit - call the node-specific routine that knows how to fold each
122 /// particular type of node.
123 SDValue visit(SDNode *N);
126 /// AddToWorkList - Add to the work list making sure its instance is at the
127 /// back (next to be processed.)
128 void AddToWorkList(SDNode *N) {
129 WorkListContents.insert(N);
130 WorkListOrder.push_back(N);
133 /// removeFromWorkList - remove all instances of N from the worklist.
135 void removeFromWorkList(SDNode *N) {
136 WorkListContents.erase(N);
139 SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
142 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) {
143 return CombineTo(N, &Res, 1, AddTo);
146 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1,
148 SDValue To[] = { Res0, Res1 };
149 return CombineTo(N, To, 2, AddTo);
152 void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO);
156 /// SimplifyDemandedBits - Check the specified integer node value to see if
157 /// it can be simplified or if things it uses can be simplified by bit
158 /// propagation. If so, return true.
159 bool SimplifyDemandedBits(SDValue Op) {
160 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
161 APInt Demanded = APInt::getAllOnesValue(BitWidth);
162 return SimplifyDemandedBits(Op, Demanded);
165 bool SimplifyDemandedBits(SDValue Op, const APInt &Demanded);
167 bool CombineToPreIndexedLoadStore(SDNode *N);
168 bool CombineToPostIndexedLoadStore(SDNode *N);
169 bool SliceUpLoad(SDNode *N);
171 void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad);
172 SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace);
173 SDValue SExtPromoteOperand(SDValue Op, EVT PVT);
174 SDValue ZExtPromoteOperand(SDValue Op, EVT PVT);
175 SDValue PromoteIntBinOp(SDValue Op);
176 SDValue PromoteIntShiftOp(SDValue Op);
177 SDValue PromoteExtend(SDValue Op);
178 bool PromoteLoad(SDValue Op);
180 void ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
181 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
182 ISD::NodeType ExtType);
184 /// combine - call the node-specific routine that knows how to fold each
185 /// particular type of node. If that doesn't do anything, try the
186 /// target-specific DAG combines.
187 SDValue combine(SDNode *N);
189 // Visitation implementation - Implement dag node combining for different
190 // node types. The semantics are as follows:
192 // SDValue.getNode() == 0 - No change was made
193 // SDValue.getNode() == N - N was replaced, is dead and has been handled.
194 // otherwise - N should be replaced by the returned Operand.
196 SDValue visitTokenFactor(SDNode *N);
197 SDValue visitMERGE_VALUES(SDNode *N);
198 SDValue visitADD(SDNode *N);
199 SDValue visitSUB(SDNode *N);
200 SDValue visitADDC(SDNode *N);
201 SDValue visitSUBC(SDNode *N);
202 SDValue visitADDE(SDNode *N);
203 SDValue visitSUBE(SDNode *N);
204 SDValue visitMUL(SDNode *N);
205 SDValue visitSDIV(SDNode *N);
206 SDValue visitUDIV(SDNode *N);
207 SDValue visitSREM(SDNode *N);
208 SDValue visitUREM(SDNode *N);
209 SDValue visitMULHU(SDNode *N);
210 SDValue visitMULHS(SDNode *N);
211 SDValue visitSMUL_LOHI(SDNode *N);
212 SDValue visitUMUL_LOHI(SDNode *N);
213 SDValue visitSMULO(SDNode *N);
214 SDValue visitUMULO(SDNode *N);
215 SDValue visitSDIVREM(SDNode *N);
216 SDValue visitUDIVREM(SDNode *N);
217 SDValue visitAND(SDNode *N);
218 SDValue visitOR(SDNode *N);
219 SDValue visitXOR(SDNode *N);
220 SDValue SimplifyVBinOp(SDNode *N);
221 SDValue SimplifyVUnaryOp(SDNode *N);
222 SDValue visitSHL(SDNode *N);
223 SDValue visitSRA(SDNode *N);
224 SDValue visitSRL(SDNode *N);
225 SDValue visitRotate(SDNode *N);
226 SDValue visitCTLZ(SDNode *N);
227 SDValue visitCTLZ_ZERO_UNDEF(SDNode *N);
228 SDValue visitCTTZ(SDNode *N);
229 SDValue visitCTTZ_ZERO_UNDEF(SDNode *N);
230 SDValue visitCTPOP(SDNode *N);
231 SDValue visitSELECT(SDNode *N);
232 SDValue visitVSELECT(SDNode *N);
233 SDValue visitSELECT_CC(SDNode *N);
234 SDValue visitSETCC(SDNode *N);
235 SDValue visitSIGN_EXTEND(SDNode *N);
236 SDValue visitZERO_EXTEND(SDNode *N);
237 SDValue visitANY_EXTEND(SDNode *N);
238 SDValue visitSIGN_EXTEND_INREG(SDNode *N);
239 SDValue visitTRUNCATE(SDNode *N);
240 SDValue visitBITCAST(SDNode *N);
241 SDValue visitBUILD_PAIR(SDNode *N);
242 SDValue visitFADD(SDNode *N);
243 SDValue visitFSUB(SDNode *N);
244 SDValue visitFMUL(SDNode *N);
245 SDValue visitFMA(SDNode *N);
246 SDValue visitFDIV(SDNode *N);
247 SDValue visitFREM(SDNode *N);
248 SDValue visitFCOPYSIGN(SDNode *N);
249 SDValue visitSINT_TO_FP(SDNode *N);
250 SDValue visitUINT_TO_FP(SDNode *N);
251 SDValue visitFP_TO_SINT(SDNode *N);
252 SDValue visitFP_TO_UINT(SDNode *N);
253 SDValue visitFP_ROUND(SDNode *N);
254 SDValue visitFP_ROUND_INREG(SDNode *N);
255 SDValue visitFP_EXTEND(SDNode *N);
256 SDValue visitFNEG(SDNode *N);
257 SDValue visitFABS(SDNode *N);
258 SDValue visitFCEIL(SDNode *N);
259 SDValue visitFTRUNC(SDNode *N);
260 SDValue visitFFLOOR(SDNode *N);
261 SDValue visitBRCOND(SDNode *N);
262 SDValue visitBR_CC(SDNode *N);
263 SDValue visitLOAD(SDNode *N);
264 SDValue visitSTORE(SDNode *N);
265 SDValue visitINSERT_VECTOR_ELT(SDNode *N);
266 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
267 SDValue visitBUILD_VECTOR(SDNode *N);
268 SDValue visitCONCAT_VECTORS(SDNode *N);
269 SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
270 SDValue visitVECTOR_SHUFFLE(SDNode *N);
271 SDValue visitINSERT_SUBVECTOR(SDNode *N);
273 SDValue XformToShuffleWithZero(SDNode *N);
274 SDValue ReassociateOps(unsigned Opc, SDLoc DL, SDValue LHS, SDValue RHS);
276 SDValue visitShiftByConstant(SDNode *N, ConstantSDNode *Amt);
278 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
279 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
280 SDValue SimplifySelect(SDLoc DL, SDValue N0, SDValue N1, SDValue N2);
281 SDValue SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1, SDValue N2,
282 SDValue N3, ISD::CondCode CC,
283 bool NotExtCompare = false);
284 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
285 SDLoc DL, bool foldBooleans = true);
287 bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
289 bool isOneUseSetCC(SDValue N) const;
291 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
293 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
294 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
295 SDValue BuildSDIV(SDNode *N);
296 SDValue BuildUDIV(SDNode *N);
297 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
298 bool DemandHighBits = true);
299 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
300 SDNode *MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg,
301 SDValue InnerPos, SDValue InnerNeg,
302 unsigned PosOpcode, unsigned NegOpcode,
304 SDNode *MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL);
305 SDValue ReduceLoadWidth(SDNode *N);
306 SDValue ReduceLoadOpStoreWidth(SDNode *N);
307 SDValue TransformFPLoadStorePair(SDNode *N);
308 SDValue reduceBuildVecExtToExtBuildVec(SDNode *N);
309 SDValue reduceBuildVecConvertToConvertBuildVec(SDNode *N);
311 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
313 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
314 /// looking for aliasing nodes and adding them to the Aliases vector.
315 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
316 SmallVectorImpl<SDValue> &Aliases);
318 /// isAlias - Return true if there is any possibility that the two addresses
320 bool isAlias(SDValue Ptr1, int64_t Size1, bool IsVolatile1,
321 const Value *SrcValue1, int SrcValueOffset1,
322 unsigned SrcValueAlign1,
323 const MDNode *TBAAInfo1,
324 SDValue Ptr2, int64_t Size2, bool IsVolatile2,
325 const Value *SrcValue2, int SrcValueOffset2,
326 unsigned SrcValueAlign2,
327 const MDNode *TBAAInfo2) const;
329 /// isAlias - Return true if there is any possibility that the two addresses
331 bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1);
333 /// FindAliasInfo - Extracts the relevant alias information from the memory
334 /// node. Returns true if the operand was a load.
335 bool FindAliasInfo(SDNode *N,
336 SDValue &Ptr, int64_t &Size, bool &IsVolatile,
337 const Value *&SrcValue, int &SrcValueOffset,
338 unsigned &SrcValueAlignment,
339 const MDNode *&TBAAInfo) const;
341 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes,
342 /// looking for a better chain (aliasing node.)
343 SDValue FindBetterChain(SDNode *N, SDValue Chain);
345 /// Merge consecutive store operations into a wide store.
346 /// This optimization uses wide integers or vectors when possible.
347 /// \return True if some memory operations were changed.
348 bool MergeConsecutiveStores(StoreSDNode *N);
350 /// \brief Try to transform a truncation where C is a constant:
351 /// (trunc (and X, C)) -> (and (trunc X), (trunc C))
353 /// \p N needs to be a truncation and its first operand an AND. Other
354 /// requirements are checked by the function (e.g. that trunc is
355 /// single-use) and if missed an empty SDValue is returned.
356 SDValue distributeTruncateThroughAnd(SDNode *N);
359 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
360 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
361 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {
362 AttributeSet FnAttrs =
363 DAG.getMachineFunction().getFunction()->getAttributes();
365 FnAttrs.hasAttribute(AttributeSet::FunctionIndex,
366 Attribute::OptimizeForSize) ||
367 FnAttrs.hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize);
370 /// Run - runs the dag combiner on all nodes in the work list
371 void Run(CombineLevel AtLevel);
373 SelectionDAG &getDAG() const { return DAG; }
375 /// getShiftAmountTy - Returns a type large enough to hold any valid
376 /// shift amount - before type legalization these can be huge.
377 EVT getShiftAmountTy(EVT LHSTy) {
378 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
379 if (LHSTy.isVector())
381 return LegalTypes ? TLI.getScalarShiftAmountTy(LHSTy)
382 : TLI.getPointerTy();
385 /// isTypeLegal - This method returns true if we are running before type
386 /// legalization or if the specified VT is legal.
387 bool isTypeLegal(const EVT &VT) {
388 if (!LegalTypes) return true;
389 return TLI.isTypeLegal(VT);
392 /// getSetCCResultType - Convenience wrapper around
393 /// TargetLowering::getSetCCResultType
394 EVT getSetCCResultType(EVT VT) const {
395 return TLI.getSetCCResultType(*DAG.getContext(), VT);
402 /// WorkListRemover - This class is a DAGUpdateListener that removes any deleted
403 /// nodes from the worklist.
404 class WorkListRemover : public SelectionDAG::DAGUpdateListener {
407 explicit WorkListRemover(DAGCombiner &dc)
408 : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {}
410 void NodeDeleted(SDNode *N, SDNode *E) override {
411 DC.removeFromWorkList(N);
416 //===----------------------------------------------------------------------===//
417 // TargetLowering::DAGCombinerInfo implementation
418 //===----------------------------------------------------------------------===//
420 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
421 ((DAGCombiner*)DC)->AddToWorkList(N);
424 void TargetLowering::DAGCombinerInfo::RemoveFromWorklist(SDNode *N) {
425 ((DAGCombiner*)DC)->removeFromWorkList(N);
428 SDValue TargetLowering::DAGCombinerInfo::
429 CombineTo(SDNode *N, const std::vector<SDValue> &To, bool AddTo) {
430 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
433 SDValue TargetLowering::DAGCombinerInfo::
434 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
435 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
439 SDValue TargetLowering::DAGCombinerInfo::
440 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
441 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
444 void TargetLowering::DAGCombinerInfo::
445 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
446 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
449 //===----------------------------------------------------------------------===//
451 //===----------------------------------------------------------------------===//
453 /// isNegatibleForFree - Return 1 if we can compute the negated form of the
454 /// specified expression for the same cost as the expression itself, or 2 if we
455 /// can compute the negated form more cheaply than the expression itself.
456 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
457 const TargetLowering &TLI,
458 const TargetOptions *Options,
459 unsigned Depth = 0) {
460 // fneg is removable even if it has multiple uses.
461 if (Op.getOpcode() == ISD::FNEG) return 2;
463 // Don't allow anything with multiple uses.
464 if (!Op.hasOneUse()) return 0;
466 // Don't recurse exponentially.
467 if (Depth > 6) return 0;
469 switch (Op.getOpcode()) {
470 default: return false;
471 case ISD::ConstantFP:
472 // Don't invert constant FP values after legalize. The negated constant
473 // isn't necessarily legal.
474 return LegalOperations ? 0 : 1;
476 // FIXME: determine better conditions for this xform.
477 if (!Options->UnsafeFPMath) return 0;
479 // After operation legalization, it might not be legal to create new FSUBs.
480 if (LegalOperations &&
481 !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType()))
484 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
485 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
488 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
489 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
492 // We can't turn -(A-B) into B-A when we honor signed zeros.
493 if (!Options->UnsafeFPMath) return 0;
495 // fold (fneg (fsub A, B)) -> (fsub B, A)
500 if (Options->HonorSignDependentRoundingFPMath()) return 0;
502 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
503 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
507 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
513 return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options,
518 /// GetNegatedExpression - If isNegatibleForFree returns true, this function
519 /// returns the newly negated expression.
520 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
521 bool LegalOperations, unsigned Depth = 0) {
522 // fneg is removable even if it has multiple uses.
523 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
525 // Don't allow anything with multiple uses.
526 assert(Op.hasOneUse() && "Unknown reuse!");
528 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
529 switch (Op.getOpcode()) {
530 default: llvm_unreachable("Unknown code");
531 case ISD::ConstantFP: {
532 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
534 return DAG.getConstantFP(V, Op.getValueType());
537 // FIXME: determine better conditions for this xform.
538 assert(DAG.getTarget().Options.UnsafeFPMath);
540 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
541 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
542 DAG.getTargetLoweringInfo(),
543 &DAG.getTarget().Options, Depth+1))
544 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
545 GetNegatedExpression(Op.getOperand(0), DAG,
546 LegalOperations, Depth+1),
548 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
549 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
550 GetNegatedExpression(Op.getOperand(1), DAG,
551 LegalOperations, Depth+1),
554 // We can't turn -(A-B) into B-A when we honor signed zeros.
555 assert(DAG.getTarget().Options.UnsafeFPMath);
557 // fold (fneg (fsub 0, B)) -> B
558 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
559 if (N0CFP->getValueAPF().isZero())
560 return Op.getOperand(1);
562 // fold (fneg (fsub A, B)) -> (fsub B, A)
563 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
564 Op.getOperand(1), Op.getOperand(0));
568 assert(!DAG.getTarget().Options.HonorSignDependentRoundingFPMath());
570 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
571 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
572 DAG.getTargetLoweringInfo(),
573 &DAG.getTarget().Options, Depth+1))
574 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
575 GetNegatedExpression(Op.getOperand(0), DAG,
576 LegalOperations, Depth+1),
579 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
580 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
582 GetNegatedExpression(Op.getOperand(1), DAG,
583 LegalOperations, Depth+1));
587 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
588 GetNegatedExpression(Op.getOperand(0), DAG,
589 LegalOperations, Depth+1));
591 return DAG.getNode(ISD::FP_ROUND, SDLoc(Op), Op.getValueType(),
592 GetNegatedExpression(Op.getOperand(0), DAG,
593 LegalOperations, Depth+1),
598 // isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
599 // that selects between the target values used for true and false, making it
600 // equivalent to a setcc. Also, set the incoming LHS, RHS, and CC references to
601 // the appropriate nodes based on the type of node we are checking. This
602 // simplifies life a bit for the callers.
603 bool DAGCombiner::isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
605 if (N.getOpcode() == ISD::SETCC) {
606 LHS = N.getOperand(0);
607 RHS = N.getOperand(1);
608 CC = N.getOperand(2);
612 if (N.getOpcode() != ISD::SELECT_CC ||
613 !TLI.isConstTrueVal(N.getOperand(2).getNode()) ||
614 !TLI.isConstFalseVal(N.getOperand(3).getNode()))
617 LHS = N.getOperand(0);
618 RHS = N.getOperand(1);
619 CC = N.getOperand(4);
623 // isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only
624 // one use. If this is true, it allows the users to invert the operation for
625 // free when it is profitable to do so.
626 bool DAGCombiner::isOneUseSetCC(SDValue N) const {
628 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
633 /// isConstantSplatVector - Returns true if N is a BUILD_VECTOR node whose
634 /// elements are all the same constant or undefined.
635 static bool isConstantSplatVector(SDNode *N, APInt& SplatValue) {
636 BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(N);
641 unsigned SplatBitSize;
643 EVT EltVT = N->getValueType(0).getVectorElementType();
644 return (C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
646 EltVT.getSizeInBits() >= SplatBitSize);
649 // \brief Returns the SDNode if it is a constant BuildVector or constant.
650 static SDNode *isConstantBuildVectorOrConstantInt(SDValue N) {
651 if (isa<ConstantSDNode>(N))
653 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N);
654 if(BV && BV->isConstant())
659 // \brief Returns the SDNode if it is a constant splat BuildVector or constant
661 static ConstantSDNode *isConstOrConstSplat(SDValue N) {
662 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N))
665 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N))
666 return BV->getConstantSplatValue();
671 SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL,
672 SDValue N0, SDValue N1) {
673 EVT VT = N0.getValueType();
674 if (N0.getOpcode() == Opc) {
675 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0.getOperand(1))) {
676 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1)) {
677 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
678 SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, L, R);
679 if (!OpNode.getNode())
681 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
683 if (N0.hasOneUse()) {
684 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one
686 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1);
687 if (!OpNode.getNode())
689 AddToWorkList(OpNode.getNode());
690 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
695 if (N1.getOpcode() == Opc) {
696 if (SDNode *R = isConstantBuildVectorOrConstantInt(N1.getOperand(1))) {
697 if (SDNode *L = isConstantBuildVectorOrConstantInt(N0)) {
698 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
699 SDValue OpNode = DAG.FoldConstantArithmetic(Opc, VT, R, L);
700 if (!OpNode.getNode())
702 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
704 if (N1.hasOneUse()) {
705 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one
707 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N1.getOperand(0), N0);
708 if (!OpNode.getNode())
710 AddToWorkList(OpNode.getNode());
711 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
719 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
721 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
723 DEBUG(dbgs() << "\nReplacing.1 ";
725 dbgs() << "\nWith: ";
726 To[0].getNode()->dump(&DAG);
727 dbgs() << " and " << NumTo-1 << " other values\n";
728 for (unsigned i = 0, e = NumTo; i != e; ++i)
729 assert((!To[i].getNode() ||
730 N->getValueType(i) == To[i].getValueType()) &&
731 "Cannot combine value to value of different type!"));
732 WorkListRemover DeadNodes(*this);
733 DAG.ReplaceAllUsesWith(N, To);
735 // Push the new nodes and any users onto the worklist
736 for (unsigned i = 0, e = NumTo; i != e; ++i) {
737 if (To[i].getNode()) {
738 AddToWorkList(To[i].getNode());
739 AddUsersToWorkList(To[i].getNode());
744 // Finally, if the node is now dead, remove it from the graph. The node
745 // may not be dead if the replacement process recursively simplified to
746 // something else needing this node.
747 if (N->use_empty()) {
748 // Nodes can be reintroduced into the worklist. Make sure we do not
749 // process a node that has been replaced.
750 removeFromWorkList(N);
752 // Finally, since the node is now dead, remove it from the graph.
755 return SDValue(N, 0);
759 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
760 // Replace all uses. If any nodes become isomorphic to other nodes and
761 // are deleted, make sure to remove them from our worklist.
762 WorkListRemover DeadNodes(*this);
763 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New);
765 // Push the new node and any (possibly new) users onto the worklist.
766 AddToWorkList(TLO.New.getNode());
767 AddUsersToWorkList(TLO.New.getNode());
769 // Finally, if the node is now dead, remove it from the graph. The node
770 // may not be dead if the replacement process recursively simplified to
771 // something else needing this node.
772 if (TLO.Old.getNode()->use_empty()) {
773 removeFromWorkList(TLO.Old.getNode());
775 // If the operands of this node are only used by the node, they will now
776 // be dead. Make sure to visit them first to delete dead nodes early.
777 for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands(); i != e; ++i)
778 if (TLO.Old.getNode()->getOperand(i).getNode()->hasOneUse())
779 AddToWorkList(TLO.Old.getNode()->getOperand(i).getNode());
781 DAG.DeleteNode(TLO.Old.getNode());
785 /// SimplifyDemandedBits - Check the specified integer node value to see if
786 /// it can be simplified or if things it uses can be simplified by bit
787 /// propagation. If so, return true.
788 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
789 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
790 APInt KnownZero, KnownOne;
791 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
795 AddToWorkList(Op.getNode());
797 // Replace the old value with the new one.
799 DEBUG(dbgs() << "\nReplacing.2 ";
800 TLO.Old.getNode()->dump(&DAG);
801 dbgs() << "\nWith: ";
802 TLO.New.getNode()->dump(&DAG);
805 CommitTargetLoweringOpt(TLO);
809 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
811 EVT VT = Load->getValueType(0);
812 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
814 DEBUG(dbgs() << "\nReplacing.9 ";
816 dbgs() << "\nWith: ";
817 Trunc.getNode()->dump(&DAG);
819 WorkListRemover DeadNodes(*this);
820 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc);
821 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1));
822 removeFromWorkList(Load);
823 DAG.DeleteNode(Load);
824 AddToWorkList(Trunc.getNode());
827 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
830 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
831 EVT MemVT = LD->getMemoryVT();
832 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
833 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
835 : LD->getExtensionType();
837 return DAG.getExtLoad(ExtType, dl, PVT,
838 LD->getChain(), LD->getBasePtr(),
839 MemVT, LD->getMemOperand());
842 unsigned Opc = Op.getOpcode();
845 case ISD::AssertSext:
846 return DAG.getNode(ISD::AssertSext, dl, PVT,
847 SExtPromoteOperand(Op.getOperand(0), PVT),
849 case ISD::AssertZext:
850 return DAG.getNode(ISD::AssertZext, dl, PVT,
851 ZExtPromoteOperand(Op.getOperand(0), PVT),
853 case ISD::Constant: {
855 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
856 return DAG.getNode(ExtOpc, dl, PVT, Op);
860 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
862 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
865 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
866 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
868 EVT OldVT = Op.getValueType();
870 bool Replace = false;
871 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
872 if (NewOp.getNode() == 0)
874 AddToWorkList(NewOp.getNode());
877 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
878 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
879 DAG.getValueType(OldVT));
882 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
883 EVT OldVT = Op.getValueType();
885 bool Replace = false;
886 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
887 if (NewOp.getNode() == 0)
889 AddToWorkList(NewOp.getNode());
892 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
893 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
896 /// PromoteIntBinOp - Promote the specified integer binary operation if the
897 /// target indicates it is beneficial. e.g. On x86, it's usually better to
898 /// promote i16 operations to i32 since i16 instructions are longer.
899 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
900 if (!LegalOperations)
903 EVT VT = Op.getValueType();
904 if (VT.isVector() || !VT.isInteger())
907 // If operation type is 'undesirable', e.g. i16 on x86, consider
909 unsigned Opc = Op.getOpcode();
910 if (TLI.isTypeDesirableForOp(Opc, VT))
914 // Consult target whether it is a good idea to promote this operation and
915 // what's the right type to promote it to.
916 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
917 assert(PVT != VT && "Don't know what type to promote to!");
919 bool Replace0 = false;
920 SDValue N0 = Op.getOperand(0);
921 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
922 if (NN0.getNode() == 0)
925 bool Replace1 = false;
926 SDValue N1 = Op.getOperand(1);
931 NN1 = PromoteOperand(N1, PVT, Replace1);
932 if (NN1.getNode() == 0)
936 AddToWorkList(NN0.getNode());
938 AddToWorkList(NN1.getNode());
941 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
943 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
945 DEBUG(dbgs() << "\nPromoting ";
946 Op.getNode()->dump(&DAG));
948 return DAG.getNode(ISD::TRUNCATE, dl, VT,
949 DAG.getNode(Opc, dl, PVT, NN0, NN1));
954 /// PromoteIntShiftOp - Promote the specified integer shift operation if the
955 /// target indicates it is beneficial. e.g. On x86, it's usually better to
956 /// promote i16 operations to i32 since i16 instructions are longer.
957 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
958 if (!LegalOperations)
961 EVT VT = Op.getValueType();
962 if (VT.isVector() || !VT.isInteger())
965 // If operation type is 'undesirable', e.g. i16 on x86, consider
967 unsigned Opc = Op.getOpcode();
968 if (TLI.isTypeDesirableForOp(Opc, VT))
972 // Consult target whether it is a good idea to promote this operation and
973 // what's the right type to promote it to.
974 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
975 assert(PVT != VT && "Don't know what type to promote to!");
977 bool Replace = false;
978 SDValue N0 = Op.getOperand(0);
980 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
981 else if (Opc == ISD::SRL)
982 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
984 N0 = PromoteOperand(N0, PVT, Replace);
985 if (N0.getNode() == 0)
988 AddToWorkList(N0.getNode());
990 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
992 DEBUG(dbgs() << "\nPromoting ";
993 Op.getNode()->dump(&DAG));
995 return DAG.getNode(ISD::TRUNCATE, dl, VT,
996 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
1001 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
1002 if (!LegalOperations)
1005 EVT VT = Op.getValueType();
1006 if (VT.isVector() || !VT.isInteger())
1009 // If operation type is 'undesirable', e.g. i16 on x86, consider
1011 unsigned Opc = Op.getOpcode();
1012 if (TLI.isTypeDesirableForOp(Opc, VT))
1016 // Consult target whether it is a good idea to promote this operation and
1017 // what's the right type to promote it to.
1018 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1019 assert(PVT != VT && "Don't know what type to promote to!");
1020 // fold (aext (aext x)) -> (aext x)
1021 // fold (aext (zext x)) -> (zext x)
1022 // fold (aext (sext x)) -> (sext x)
1023 DEBUG(dbgs() << "\nPromoting ";
1024 Op.getNode()->dump(&DAG));
1025 return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0));
1030 bool DAGCombiner::PromoteLoad(SDValue Op) {
1031 if (!LegalOperations)
1034 EVT VT = Op.getValueType();
1035 if (VT.isVector() || !VT.isInteger())
1038 // If operation type is 'undesirable', e.g. i16 on x86, consider
1040 unsigned Opc = Op.getOpcode();
1041 if (TLI.isTypeDesirableForOp(Opc, VT))
1045 // Consult target whether it is a good idea to promote this operation and
1046 // what's the right type to promote it to.
1047 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
1048 assert(PVT != VT && "Don't know what type to promote to!");
1051 SDNode *N = Op.getNode();
1052 LoadSDNode *LD = cast<LoadSDNode>(N);
1053 EVT MemVT = LD->getMemoryVT();
1054 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
1055 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
1057 : LD->getExtensionType();
1058 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
1059 LD->getChain(), LD->getBasePtr(),
1060 MemVT, LD->getMemOperand());
1061 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
1063 DEBUG(dbgs() << "\nPromoting ";
1066 Result.getNode()->dump(&DAG);
1068 WorkListRemover DeadNodes(*this);
1069 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
1070 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1));
1071 removeFromWorkList(N);
1073 AddToWorkList(Result.getNode());
1080 //===----------------------------------------------------------------------===//
1081 // Main DAG Combiner implementation
1082 //===----------------------------------------------------------------------===//
1084 void DAGCombiner::Run(CombineLevel AtLevel) {
1085 // set the instance variables, so that the various visit routines may use it.
1087 LegalOperations = Level >= AfterLegalizeVectorOps;
1088 LegalTypes = Level >= AfterLegalizeTypes;
1090 // Add all the dag nodes to the worklist.
1091 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
1092 E = DAG.allnodes_end(); I != E; ++I)
1095 // Create a dummy node (which is not added to allnodes), that adds a reference
1096 // to the root node, preventing it from being deleted, and tracking any
1097 // changes of the root.
1098 HandleSDNode Dummy(DAG.getRoot());
1100 // The root of the dag may dangle to deleted nodes until the dag combiner is
1101 // done. Set it to null to avoid confusion.
1102 DAG.setRoot(SDValue());
1104 // while the worklist isn't empty, find a node and
1105 // try and combine it.
1106 while (!WorkListContents.empty()) {
1108 // The WorkListOrder holds the SDNodes in order, but it may contain
1110 // In order to avoid a linear scan, we use a set (O(log N)) to hold what the
1111 // worklist *should* contain, and check the node we want to visit is should
1112 // actually be visited.
1114 N = WorkListOrder.pop_back_val();
1115 } while (!WorkListContents.erase(N));
1117 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1118 // N is deleted from the DAG, since they too may now be dead or may have a
1119 // reduced number of uses, allowing other xforms.
1120 if (N->use_empty() && N != &Dummy) {
1121 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1122 AddToWorkList(N->getOperand(i).getNode());
1128 SDValue RV = combine(N);
1130 if (RV.getNode() == 0)
1135 // If we get back the same node we passed in, rather than a new node or
1136 // zero, we know that the node must have defined multiple values and
1137 // CombineTo was used. Since CombineTo takes care of the worklist
1138 // mechanics for us, we have no work to do in this case.
1139 if (RV.getNode() == N)
1142 assert(N->getOpcode() != ISD::DELETED_NODE &&
1143 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1144 "Node was deleted but visit returned new node!");
1146 DEBUG(dbgs() << "\nReplacing.3 ";
1148 dbgs() << "\nWith: ";
1149 RV.getNode()->dump(&DAG);
1152 // Transfer debug value.
1153 DAG.TransferDbgValues(SDValue(N, 0), RV);
1154 WorkListRemover DeadNodes(*this);
1155 if (N->getNumValues() == RV.getNode()->getNumValues())
1156 DAG.ReplaceAllUsesWith(N, RV.getNode());
1158 assert(N->getValueType(0) == RV.getValueType() &&
1159 N->getNumValues() == 1 && "Type mismatch");
1161 DAG.ReplaceAllUsesWith(N, &OpV);
1164 // Push the new node and any users onto the worklist
1165 AddToWorkList(RV.getNode());
1166 AddUsersToWorkList(RV.getNode());
1168 // Add any uses of the old node to the worklist in case this node is the
1169 // last one that uses them. They may become dead after this node is
1171 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1172 AddToWorkList(N->getOperand(i).getNode());
1174 // Finally, if the node is now dead, remove it from the graph. The node
1175 // may not be dead if the replacement process recursively simplified to
1176 // something else needing this node.
1177 if (N->use_empty()) {
1178 // Nodes can be reintroduced into the worklist. Make sure we do not
1179 // process a node that has been replaced.
1180 removeFromWorkList(N);
1182 // Finally, since the node is now dead, remove it from the graph.
1187 // If the root changed (e.g. it was a dead load, update the root).
1188 DAG.setRoot(Dummy.getValue());
1189 DAG.RemoveDeadNodes();
1192 SDValue DAGCombiner::visit(SDNode *N) {
1193 switch (N->getOpcode()) {
1195 case ISD::TokenFactor: return visitTokenFactor(N);
1196 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1197 case ISD::ADD: return visitADD(N);
1198 case ISD::SUB: return visitSUB(N);
1199 case ISD::ADDC: return visitADDC(N);
1200 case ISD::SUBC: return visitSUBC(N);
1201 case ISD::ADDE: return visitADDE(N);
1202 case ISD::SUBE: return visitSUBE(N);
1203 case ISD::MUL: return visitMUL(N);
1204 case ISD::SDIV: return visitSDIV(N);
1205 case ISD::UDIV: return visitUDIV(N);
1206 case ISD::SREM: return visitSREM(N);
1207 case ISD::UREM: return visitUREM(N);
1208 case ISD::MULHU: return visitMULHU(N);
1209 case ISD::MULHS: return visitMULHS(N);
1210 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1211 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1212 case ISD::SMULO: return visitSMULO(N);
1213 case ISD::UMULO: return visitUMULO(N);
1214 case ISD::SDIVREM: return visitSDIVREM(N);
1215 case ISD::UDIVREM: return visitUDIVREM(N);
1216 case ISD::AND: return visitAND(N);
1217 case ISD::OR: return visitOR(N);
1218 case ISD::XOR: return visitXOR(N);
1219 case ISD::SHL: return visitSHL(N);
1220 case ISD::SRA: return visitSRA(N);
1221 case ISD::SRL: return visitSRL(N);
1223 case ISD::ROTL: return visitRotate(N);
1224 case ISD::CTLZ: return visitCTLZ(N);
1225 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1226 case ISD::CTTZ: return visitCTTZ(N);
1227 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1228 case ISD::CTPOP: return visitCTPOP(N);
1229 case ISD::SELECT: return visitSELECT(N);
1230 case ISD::VSELECT: return visitVSELECT(N);
1231 case ISD::SELECT_CC: return visitSELECT_CC(N);
1232 case ISD::SETCC: return visitSETCC(N);
1233 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1234 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1235 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1236 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1237 case ISD::TRUNCATE: return visitTRUNCATE(N);
1238 case ISD::BITCAST: return visitBITCAST(N);
1239 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1240 case ISD::FADD: return visitFADD(N);
1241 case ISD::FSUB: return visitFSUB(N);
1242 case ISD::FMUL: return visitFMUL(N);
1243 case ISD::FMA: return visitFMA(N);
1244 case ISD::FDIV: return visitFDIV(N);
1245 case ISD::FREM: return visitFREM(N);
1246 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1247 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1248 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1249 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1250 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1251 case ISD::FP_ROUND: return visitFP_ROUND(N);
1252 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1253 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1254 case ISD::FNEG: return visitFNEG(N);
1255 case ISD::FABS: return visitFABS(N);
1256 case ISD::FFLOOR: return visitFFLOOR(N);
1257 case ISD::FCEIL: return visitFCEIL(N);
1258 case ISD::FTRUNC: return visitFTRUNC(N);
1259 case ISD::BRCOND: return visitBRCOND(N);
1260 case ISD::BR_CC: return visitBR_CC(N);
1261 case ISD::LOAD: return visitLOAD(N);
1262 case ISD::STORE: return visitSTORE(N);
1263 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1264 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1265 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1266 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1267 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1268 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1269 case ISD::INSERT_SUBVECTOR: return visitINSERT_SUBVECTOR(N);
1274 SDValue DAGCombiner::combine(SDNode *N) {
1275 SDValue RV = visit(N);
1277 // If nothing happened, try a target-specific DAG combine.
1278 if (RV.getNode() == 0) {
1279 assert(N->getOpcode() != ISD::DELETED_NODE &&
1280 "Node was deleted but visit returned NULL!");
1282 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1283 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1285 // Expose the DAG combiner to the target combiner impls.
1286 TargetLowering::DAGCombinerInfo
1287 DagCombineInfo(DAG, Level, false, this);
1289 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1293 // If nothing happened still, try promoting the operation.
1294 if (RV.getNode() == 0) {
1295 switch (N->getOpcode()) {
1303 RV = PromoteIntBinOp(SDValue(N, 0));
1308 RV = PromoteIntShiftOp(SDValue(N, 0));
1310 case ISD::SIGN_EXTEND:
1311 case ISD::ZERO_EXTEND:
1312 case ISD::ANY_EXTEND:
1313 RV = PromoteExtend(SDValue(N, 0));
1316 if (PromoteLoad(SDValue(N, 0)))
1322 // If N is a commutative binary node, try commuting it to enable more
1324 if (RV.getNode() == 0 &&
1325 SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1326 N->getNumValues() == 1) {
1327 SDValue N0 = N->getOperand(0);
1328 SDValue N1 = N->getOperand(1);
1330 // Constant operands are canonicalized to RHS.
1331 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1332 SDValue Ops[] = { N1, N0 };
1333 SDNode *CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(),
1336 return SDValue(CSENode, 0);
1343 /// getInputChainForNode - Given a node, return its input chain if it has one,
1344 /// otherwise return a null sd operand.
1345 static SDValue getInputChainForNode(SDNode *N) {
1346 if (unsigned NumOps = N->getNumOperands()) {
1347 if (N->getOperand(0).getValueType() == MVT::Other)
1348 return N->getOperand(0);
1349 if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1350 return N->getOperand(NumOps-1);
1351 for (unsigned i = 1; i < NumOps-1; ++i)
1352 if (N->getOperand(i).getValueType() == MVT::Other)
1353 return N->getOperand(i);
1358 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1359 // If N has two operands, where one has an input chain equal to the other,
1360 // the 'other' chain is redundant.
1361 if (N->getNumOperands() == 2) {
1362 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1363 return N->getOperand(0);
1364 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1365 return N->getOperand(1);
1368 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1369 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1370 SmallPtrSet<SDNode*, 16> SeenOps;
1371 bool Changed = false; // If we should replace this token factor.
1373 // Start out with this token factor.
1376 // Iterate through token factors. The TFs grows when new token factors are
1378 for (unsigned i = 0; i < TFs.size(); ++i) {
1379 SDNode *TF = TFs[i];
1381 // Check each of the operands.
1382 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1383 SDValue Op = TF->getOperand(i);
1385 switch (Op.getOpcode()) {
1386 case ISD::EntryToken:
1387 // Entry tokens don't need to be added to the list. They are
1392 case ISD::TokenFactor:
1393 if (Op.hasOneUse() &&
1394 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1395 // Queue up for processing.
1396 TFs.push_back(Op.getNode());
1397 // Clean up in case the token factor is removed.
1398 AddToWorkList(Op.getNode());
1405 // Only add if it isn't already in the list.
1406 if (SeenOps.insert(Op.getNode()))
1417 // If we've change things around then replace token factor.
1420 // The entry token is the only possible outcome.
1421 Result = DAG.getEntryNode();
1423 // New and improved token factor.
1424 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N),
1425 MVT::Other, &Ops[0], Ops.size());
1428 // Don't add users to work list.
1429 return CombineTo(N, Result, false);
1435 /// MERGE_VALUES can always be eliminated.
1436 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1437 WorkListRemover DeadNodes(*this);
1438 // Replacing results may cause a different MERGE_VALUES to suddenly
1439 // be CSE'd with N, and carry its uses with it. Iterate until no
1440 // uses remain, to ensure that the node can be safely deleted.
1441 // First add the users of this node to the work list so that they
1442 // can be tried again once they have new operands.
1443 AddUsersToWorkList(N);
1445 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1446 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1447 } while (!N->use_empty());
1448 removeFromWorkList(N);
1450 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1454 SDValue combineShlAddConstant(SDLoc DL, SDValue N0, SDValue N1,
1455 SelectionDAG &DAG) {
1456 EVT VT = N0.getValueType();
1457 SDValue N00 = N0.getOperand(0);
1458 SDValue N01 = N0.getOperand(1);
1459 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N01);
1461 if (N01C && N00.getOpcode() == ISD::ADD && N00.getNode()->hasOneUse() &&
1462 isa<ConstantSDNode>(N00.getOperand(1))) {
1463 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1464 N0 = DAG.getNode(ISD::ADD, SDLoc(N0), VT,
1465 DAG.getNode(ISD::SHL, SDLoc(N00), VT,
1466 N00.getOperand(0), N01),
1467 DAG.getNode(ISD::SHL, SDLoc(N01), VT,
1468 N00.getOperand(1), N01));
1469 return DAG.getNode(ISD::ADD, DL, VT, N0, N1);
1475 SDValue DAGCombiner::visitADD(SDNode *N) {
1476 SDValue N0 = N->getOperand(0);
1477 SDValue N1 = N->getOperand(1);
1478 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1479 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1480 EVT VT = N0.getValueType();
1483 if (VT.isVector()) {
1484 SDValue FoldedVOp = SimplifyVBinOp(N);
1485 if (FoldedVOp.getNode()) return FoldedVOp;
1487 // fold (add x, 0) -> x, vector edition
1488 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1490 if (ISD::isBuildVectorAllZeros(N0.getNode()))
1494 // fold (add x, undef) -> undef
1495 if (N0.getOpcode() == ISD::UNDEF)
1497 if (N1.getOpcode() == ISD::UNDEF)
1499 // fold (add c1, c2) -> c1+c2
1501 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1502 // canonicalize constant to RHS
1504 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, N0);
1505 // fold (add x, 0) -> x
1506 if (N1C && N1C->isNullValue())
1508 // fold (add Sym, c) -> Sym+c
1509 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1510 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1511 GA->getOpcode() == ISD::GlobalAddress)
1512 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1514 (uint64_t)N1C->getSExtValue());
1515 // fold ((c1-A)+c2) -> (c1+c2)-A
1516 if (N1C && N0.getOpcode() == ISD::SUB)
1517 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1518 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1519 DAG.getConstant(N1C->getAPIntValue()+
1520 N0C->getAPIntValue(), VT),
1523 SDValue RADD = ReassociateOps(ISD::ADD, SDLoc(N), N0, N1);
1524 if (RADD.getNode() != 0)
1526 // fold ((0-A) + B) -> B-A
1527 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1528 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1529 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, N0.getOperand(1));
1530 // fold (A + (0-B)) -> A-B
1531 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1532 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1533 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1.getOperand(1));
1534 // fold (A+(B-A)) -> B
1535 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1536 return N1.getOperand(0);
1537 // fold ((B-A)+A) -> B
1538 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1539 return N0.getOperand(0);
1540 // fold (A+(B-(A+C))) to (B-C)
1541 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1542 N0 == N1.getOperand(1).getOperand(0))
1543 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1544 N1.getOperand(1).getOperand(1));
1545 // fold (A+(B-(C+A))) to (B-C)
1546 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1547 N0 == N1.getOperand(1).getOperand(1))
1548 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1549 N1.getOperand(1).getOperand(0));
1550 // fold (A+((B-A)+or-C)) to (B+or-C)
1551 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1552 N1.getOperand(0).getOpcode() == ISD::SUB &&
1553 N0 == N1.getOperand(0).getOperand(1))
1554 return DAG.getNode(N1.getOpcode(), SDLoc(N), VT,
1555 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1557 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1558 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1559 SDValue N00 = N0.getOperand(0);
1560 SDValue N01 = N0.getOperand(1);
1561 SDValue N10 = N1.getOperand(0);
1562 SDValue N11 = N1.getOperand(1);
1564 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1565 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1566 DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10),
1567 DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11));
1570 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1571 return SDValue(N, 0);
1573 // fold (a+b) -> (a|b) iff a and b share no bits.
1574 if (VT.isInteger() && !VT.isVector()) {
1575 APInt LHSZero, LHSOne;
1576 APInt RHSZero, RHSOne;
1577 DAG.ComputeMaskedBits(N0, LHSZero, LHSOne);
1579 if (LHSZero.getBoolValue()) {
1580 DAG.ComputeMaskedBits(N1, RHSZero, RHSOne);
1582 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1583 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1584 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero){
1585 if (!LegalOperations || TLI.isOperationLegal(ISD::OR, VT))
1586 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1);
1591 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1592 if (N0.getOpcode() == ISD::SHL && N0.getNode()->hasOneUse()) {
1593 SDValue Result = combineShlAddConstant(SDLoc(N), N0, N1, DAG);
1594 if (Result.getNode()) return Result;
1596 if (N1.getOpcode() == ISD::SHL && N1.getNode()->hasOneUse()) {
1597 SDValue Result = combineShlAddConstant(SDLoc(N), N1, N0, DAG);
1598 if (Result.getNode()) return Result;
1601 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1602 if (N1.getOpcode() == ISD::SHL &&
1603 N1.getOperand(0).getOpcode() == ISD::SUB)
1604 if (ConstantSDNode *C =
1605 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1606 if (C->getAPIntValue() == 0)
1607 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0,
1608 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1609 N1.getOperand(0).getOperand(1),
1611 if (N0.getOpcode() == ISD::SHL &&
1612 N0.getOperand(0).getOpcode() == ISD::SUB)
1613 if (ConstantSDNode *C =
1614 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1615 if (C->getAPIntValue() == 0)
1616 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1,
1617 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1618 N0.getOperand(0).getOperand(1),
1621 if (N1.getOpcode() == ISD::AND) {
1622 SDValue AndOp0 = N1.getOperand(0);
1623 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1624 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1625 unsigned DestBits = VT.getScalarType().getSizeInBits();
1627 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1628 // and similar xforms where the inner op is either ~0 or 0.
1629 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1631 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1635 // add (sext i1), X -> sub X, (zext i1)
1636 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1637 N0.getOperand(0).getValueType() == MVT::i1 &&
1638 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1640 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1641 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1647 SDValue DAGCombiner::visitADDC(SDNode *N) {
1648 SDValue N0 = N->getOperand(0);
1649 SDValue N1 = N->getOperand(1);
1650 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1651 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1652 EVT VT = N0.getValueType();
1654 // If the flag result is dead, turn this into an ADD.
1655 if (!N->hasAnyUseOfValue(1))
1656 return CombineTo(N, DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N1),
1657 DAG.getNode(ISD::CARRY_FALSE,
1658 SDLoc(N), MVT::Glue));
1660 // canonicalize constant to RHS.
1662 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N1, N0);
1664 // fold (addc x, 0) -> x + no carry out
1665 if (N1C && N1C->isNullValue())
1666 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1667 SDLoc(N), MVT::Glue));
1669 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1670 APInt LHSZero, LHSOne;
1671 APInt RHSZero, RHSOne;
1672 DAG.ComputeMaskedBits(N0, LHSZero, LHSOne);
1674 if (LHSZero.getBoolValue()) {
1675 DAG.ComputeMaskedBits(N1, RHSZero, RHSOne);
1677 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1678 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1679 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1680 return CombineTo(N, DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1),
1681 DAG.getNode(ISD::CARRY_FALSE,
1682 SDLoc(N), MVT::Glue));
1688 SDValue DAGCombiner::visitADDE(SDNode *N) {
1689 SDValue N0 = N->getOperand(0);
1690 SDValue N1 = N->getOperand(1);
1691 SDValue CarryIn = N->getOperand(2);
1692 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1693 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1695 // canonicalize constant to RHS
1697 return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(),
1700 // fold (adde x, y, false) -> (addc x, y)
1701 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1702 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1);
1707 // Since it may not be valid to emit a fold to zero for vector initializers
1708 // check if we can before folding.
1709 static SDValue tryFoldToZero(SDLoc DL, const TargetLowering &TLI, EVT VT,
1711 bool LegalOperations, bool LegalTypes) {
1713 return DAG.getConstant(0, VT);
1714 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
1715 return DAG.getConstant(0, VT);
1719 SDValue DAGCombiner::visitSUB(SDNode *N) {
1720 SDValue N0 = N->getOperand(0);
1721 SDValue N1 = N->getOperand(1);
1722 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1723 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1724 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? 0 :
1725 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1726 EVT VT = N0.getValueType();
1729 if (VT.isVector()) {
1730 SDValue FoldedVOp = SimplifyVBinOp(N);
1731 if (FoldedVOp.getNode()) return FoldedVOp;
1733 // fold (sub x, 0) -> x, vector edition
1734 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1738 // fold (sub x, x) -> 0
1739 // FIXME: Refactor this and xor and other similar operations together.
1741 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
1742 // fold (sub c1, c2) -> c1-c2
1744 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1745 // fold (sub x, c) -> (add x, -c)
1747 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0,
1748 DAG.getConstant(-N1C->getAPIntValue(), VT));
1749 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1750 if (N0C && N0C->isAllOnesValue())
1751 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
1752 // fold A-(A-B) -> B
1753 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1754 return N1.getOperand(1);
1755 // fold (A+B)-A -> B
1756 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1757 return N0.getOperand(1);
1758 // fold (A+B)-B -> A
1759 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1760 return N0.getOperand(0);
1761 // fold C2-(A+C1) -> (C2-C1)-A
1762 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1763 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(),
1765 return DAG.getNode(ISD::SUB, SDLoc(N), VT, NewC,
1768 // fold ((A+(B+or-C))-B) -> A+or-C
1769 if (N0.getOpcode() == ISD::ADD &&
1770 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1771 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1772 N0.getOperand(1).getOperand(0) == N1)
1773 return DAG.getNode(N0.getOperand(1).getOpcode(), SDLoc(N), VT,
1774 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1775 // fold ((A+(C+B))-B) -> A+C
1776 if (N0.getOpcode() == ISD::ADD &&
1777 N0.getOperand(1).getOpcode() == ISD::ADD &&
1778 N0.getOperand(1).getOperand(1) == N1)
1779 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1780 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1781 // fold ((A-(B-C))-C) -> A-B
1782 if (N0.getOpcode() == ISD::SUB &&
1783 N0.getOperand(1).getOpcode() == ISD::SUB &&
1784 N0.getOperand(1).getOperand(1) == N1)
1785 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1786 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1788 // If either operand of a sub is undef, the result is undef
1789 if (N0.getOpcode() == ISD::UNDEF)
1791 if (N1.getOpcode() == ISD::UNDEF)
1794 // If the relocation model supports it, consider symbol offsets.
1795 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1796 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1797 // fold (sub Sym, c) -> Sym-c
1798 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1799 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1801 (uint64_t)N1C->getSExtValue());
1802 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1803 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1804 if (GA->getGlobal() == GB->getGlobal())
1805 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1812 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1813 SDValue N0 = N->getOperand(0);
1814 SDValue N1 = N->getOperand(1);
1815 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1816 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1817 EVT VT = N0.getValueType();
1819 // If the flag result is dead, turn this into an SUB.
1820 if (!N->hasAnyUseOfValue(1))
1821 return CombineTo(N, DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1),
1822 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1825 // fold (subc x, x) -> 0 + no borrow
1827 return CombineTo(N, DAG.getConstant(0, VT),
1828 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1831 // fold (subc x, 0) -> x + no borrow
1832 if (N1C && N1C->isNullValue())
1833 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1836 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1837 if (N0C && N0C->isAllOnesValue())
1838 return CombineTo(N, DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0),
1839 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1845 SDValue DAGCombiner::visitSUBE(SDNode *N) {
1846 SDValue N0 = N->getOperand(0);
1847 SDValue N1 = N->getOperand(1);
1848 SDValue CarryIn = N->getOperand(2);
1850 // fold (sube x, y, false) -> (subc x, y)
1851 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1852 return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1);
1857 SDValue DAGCombiner::visitMUL(SDNode *N) {
1858 SDValue N0 = N->getOperand(0);
1859 SDValue N1 = N->getOperand(1);
1860 EVT VT = N0.getValueType();
1862 // fold (mul x, undef) -> 0
1863 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1864 return DAG.getConstant(0, VT);
1866 bool N0IsConst = false;
1867 bool N1IsConst = false;
1868 APInt ConstValue0, ConstValue1;
1870 if (VT.isVector()) {
1871 SDValue FoldedVOp = SimplifyVBinOp(N);
1872 if (FoldedVOp.getNode()) return FoldedVOp;
1874 N0IsConst = isConstantSplatVector(N0.getNode(), ConstValue0);
1875 N1IsConst = isConstantSplatVector(N1.getNode(), ConstValue1);
1877 N0IsConst = dyn_cast<ConstantSDNode>(N0) != 0;
1878 ConstValue0 = N0IsConst ? (dyn_cast<ConstantSDNode>(N0))->getAPIntValue()
1880 N1IsConst = dyn_cast<ConstantSDNode>(N1) != 0;
1881 ConstValue1 = N1IsConst ? (dyn_cast<ConstantSDNode>(N1))->getAPIntValue()
1885 // fold (mul c1, c2) -> c1*c2
1886 if (N0IsConst && N1IsConst)
1887 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0.getNode(), N1.getNode());
1889 // canonicalize constant to RHS
1890 if (N0IsConst && !N1IsConst)
1891 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0);
1892 // fold (mul x, 0) -> 0
1893 if (N1IsConst && ConstValue1 == 0)
1895 // We require a splat of the entire scalar bit width for non-contiguous
1898 ConstValue1.getBitWidth() == VT.getScalarType().getSizeInBits();
1899 // fold (mul x, 1) -> x
1900 if (N1IsConst && ConstValue1 == 1 && IsFullSplat)
1902 // fold (mul x, -1) -> 0-x
1903 if (N1IsConst && ConstValue1.isAllOnesValue())
1904 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1905 DAG.getConstant(0, VT), N0);
1906 // fold (mul x, (1 << c)) -> x << c
1907 if (N1IsConst && ConstValue1.isPowerOf2() && IsFullSplat)
1908 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
1909 DAG.getConstant(ConstValue1.logBase2(),
1910 getShiftAmountTy(N0.getValueType())));
1911 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
1912 if (N1IsConst && (-ConstValue1).isPowerOf2() && IsFullSplat) {
1913 unsigned Log2Val = (-ConstValue1).logBase2();
1914 // FIXME: If the input is something that is easily negated (e.g. a
1915 // single-use add), we should put the negate there.
1916 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1917 DAG.getConstant(0, VT),
1918 DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
1919 DAG.getConstant(Log2Val,
1920 getShiftAmountTy(N0.getValueType()))));
1924 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
1925 if (N1IsConst && N0.getOpcode() == ISD::SHL &&
1926 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1927 isa<ConstantSDNode>(N0.getOperand(1)))) {
1928 SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT,
1929 N1, N0.getOperand(1));
1930 AddToWorkList(C3.getNode());
1931 return DAG.getNode(ISD::MUL, SDLoc(N), VT,
1932 N0.getOperand(0), C3);
1935 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
1938 SDValue Sh(0,0), Y(0,0);
1939 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
1940 if (N0.getOpcode() == ISD::SHL &&
1941 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1942 isa<ConstantSDNode>(N0.getOperand(1))) &&
1943 N0.getNode()->hasOneUse()) {
1945 } else if (N1.getOpcode() == ISD::SHL &&
1946 isa<ConstantSDNode>(N1.getOperand(1)) &&
1947 N1.getNode()->hasOneUse()) {
1952 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
1953 Sh.getOperand(0), Y);
1954 return DAG.getNode(ISD::SHL, SDLoc(N), VT,
1955 Mul, Sh.getOperand(1));
1959 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
1960 if (N1IsConst && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
1961 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1962 isa<ConstantSDNode>(N0.getOperand(1))))
1963 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1964 DAG.getNode(ISD::MUL, SDLoc(N0), VT,
1965 N0.getOperand(0), N1),
1966 DAG.getNode(ISD::MUL, SDLoc(N1), VT,
1967 N0.getOperand(1), N1));
1970 SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1);
1971 if (RMUL.getNode() != 0)
1977 SDValue DAGCombiner::visitSDIV(SDNode *N) {
1978 SDValue N0 = N->getOperand(0);
1979 SDValue N1 = N->getOperand(1);
1980 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1981 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1982 EVT VT = N->getValueType(0);
1985 if (VT.isVector()) {
1986 SDValue FoldedVOp = SimplifyVBinOp(N);
1987 if (FoldedVOp.getNode()) return FoldedVOp;
1990 // fold (sdiv c1, c2) -> c1/c2
1991 if (N0C && N1C && !N1C->isNullValue())
1992 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
1993 // fold (sdiv X, 1) -> X
1994 if (N1C && N1C->getAPIntValue() == 1LL)
1996 // fold (sdiv X, -1) -> 0-X
1997 if (N1C && N1C->isAllOnesValue())
1998 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1999 DAG.getConstant(0, VT), N0);
2000 // If we know the sign bits of both operands are zero, strength reduce to a
2001 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
2002 if (!VT.isVector()) {
2003 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2004 return DAG.getNode(ISD::UDIV, SDLoc(N), N1.getValueType(),
2007 // fold (sdiv X, pow2) -> simple ops after legalize
2008 if (N1C && !N1C->isNullValue() &&
2009 (N1C->getAPIntValue().isPowerOf2() ||
2010 (-N1C->getAPIntValue()).isPowerOf2())) {
2011 // If dividing by powers of two is cheap, then don't perform the following
2013 if (TLI.isPow2DivCheap())
2016 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
2018 // Splat the sign bit into the register
2019 SDValue SGN = DAG.getNode(ISD::SRA, SDLoc(N), VT, N0,
2020 DAG.getConstant(VT.getSizeInBits()-1,
2021 getShiftAmountTy(N0.getValueType())));
2022 AddToWorkList(SGN.getNode());
2024 // Add (N0 < 0) ? abs2 - 1 : 0;
2025 SDValue SRL = DAG.getNode(ISD::SRL, SDLoc(N), VT, SGN,
2026 DAG.getConstant(VT.getSizeInBits() - lg2,
2027 getShiftAmountTy(SGN.getValueType())));
2028 SDValue ADD = DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, SRL);
2029 AddToWorkList(SRL.getNode());
2030 AddToWorkList(ADD.getNode()); // Divide by pow2
2031 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), VT, ADD,
2032 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType())));
2034 // If we're dividing by a positive value, we're done. Otherwise, we must
2035 // negate the result.
2036 if (N1C->getAPIntValue().isNonNegative())
2039 AddToWorkList(SRA.getNode());
2040 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
2041 DAG.getConstant(0, VT), SRA);
2044 // if integer divide is expensive and we satisfy the requirements, emit an
2045 // alternate sequence.
2046 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
2047 SDValue Op = BuildSDIV(N);
2048 if (Op.getNode()) return Op;
2052 if (N0.getOpcode() == ISD::UNDEF)
2053 return DAG.getConstant(0, VT);
2054 // X / undef -> undef
2055 if (N1.getOpcode() == ISD::UNDEF)
2061 SDValue DAGCombiner::visitUDIV(SDNode *N) {
2062 SDValue N0 = N->getOperand(0);
2063 SDValue N1 = N->getOperand(1);
2064 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
2065 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
2066 EVT VT = N->getValueType(0);
2069 if (VT.isVector()) {
2070 SDValue FoldedVOp = SimplifyVBinOp(N);
2071 if (FoldedVOp.getNode()) return FoldedVOp;
2074 // fold (udiv c1, c2) -> c1/c2
2075 if (N0C && N1C && !N1C->isNullValue())
2076 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
2077 // fold (udiv x, (1 << c)) -> x >>u c
2078 if (N1C && N1C->getAPIntValue().isPowerOf2())
2079 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0,
2080 DAG.getConstant(N1C->getAPIntValue().logBase2(),
2081 getShiftAmountTy(N0.getValueType())));
2082 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
2083 if (N1.getOpcode() == ISD::SHL) {
2084 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2085 if (SHC->getAPIntValue().isPowerOf2()) {
2086 EVT ADDVT = N1.getOperand(1).getValueType();
2087 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N), ADDVT,
2089 DAG.getConstant(SHC->getAPIntValue()
2092 AddToWorkList(Add.getNode());
2093 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, Add);
2097 // fold (udiv x, c) -> alternate
2098 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
2099 SDValue Op = BuildUDIV(N);
2100 if (Op.getNode()) return Op;
2104 if (N0.getOpcode() == ISD::UNDEF)
2105 return DAG.getConstant(0, VT);
2106 // X / undef -> undef
2107 if (N1.getOpcode() == ISD::UNDEF)
2113 SDValue DAGCombiner::visitSREM(SDNode *N) {
2114 SDValue N0 = N->getOperand(0);
2115 SDValue N1 = N->getOperand(1);
2116 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2117 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2118 EVT VT = N->getValueType(0);
2120 // fold (srem c1, c2) -> c1%c2
2121 if (N0C && N1C && !N1C->isNullValue())
2122 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
2123 // If we know the sign bits of both operands are zero, strength reduce to a
2124 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
2125 if (!VT.isVector()) {
2126 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2127 return DAG.getNode(ISD::UREM, SDLoc(N), VT, N0, N1);
2130 // If X/C can be simplified by the division-by-constant logic, lower
2131 // X%C to the equivalent of X-X/C*C.
2132 if (N1C && !N1C->isNullValue()) {
2133 SDValue Div = DAG.getNode(ISD::SDIV, SDLoc(N), VT, N0, N1);
2134 AddToWorkList(Div.getNode());
2135 SDValue OptimizedDiv = combine(Div.getNode());
2136 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2137 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2139 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2140 AddToWorkList(Mul.getNode());
2146 if (N0.getOpcode() == ISD::UNDEF)
2147 return DAG.getConstant(0, VT);
2148 // X % undef -> undef
2149 if (N1.getOpcode() == ISD::UNDEF)
2155 SDValue DAGCombiner::visitUREM(SDNode *N) {
2156 SDValue N0 = N->getOperand(0);
2157 SDValue N1 = N->getOperand(1);
2158 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2159 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2160 EVT VT = N->getValueType(0);
2162 // fold (urem c1, c2) -> c1%c2
2163 if (N0C && N1C && !N1C->isNullValue())
2164 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
2165 // fold (urem x, pow2) -> (and x, pow2-1)
2166 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
2167 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0,
2168 DAG.getConstant(N1C->getAPIntValue()-1,VT));
2169 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2170 if (N1.getOpcode() == ISD::SHL) {
2171 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2172 if (SHC->getAPIntValue().isPowerOf2()) {
2174 DAG.getNode(ISD::ADD, SDLoc(N), VT, N1,
2175 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
2177 AddToWorkList(Add.getNode());
2178 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, Add);
2183 // If X/C can be simplified by the division-by-constant logic, lower
2184 // X%C to the equivalent of X-X/C*C.
2185 if (N1C && !N1C->isNullValue()) {
2186 SDValue Div = DAG.getNode(ISD::UDIV, SDLoc(N), VT, N0, N1);
2187 AddToWorkList(Div.getNode());
2188 SDValue OptimizedDiv = combine(Div.getNode());
2189 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2190 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2192 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2193 AddToWorkList(Mul.getNode());
2199 if (N0.getOpcode() == ISD::UNDEF)
2200 return DAG.getConstant(0, VT);
2201 // X % undef -> undef
2202 if (N1.getOpcode() == ISD::UNDEF)
2208 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2209 SDValue N0 = N->getOperand(0);
2210 SDValue N1 = N->getOperand(1);
2211 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2212 EVT VT = N->getValueType(0);
2215 // fold (mulhs x, 0) -> 0
2216 if (N1C && N1C->isNullValue())
2218 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2219 if (N1C && N1C->getAPIntValue() == 1)
2220 return DAG.getNode(ISD::SRA, SDLoc(N), N0.getValueType(), N0,
2221 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2222 getShiftAmountTy(N0.getValueType())));
2223 // fold (mulhs x, undef) -> 0
2224 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2225 return DAG.getConstant(0, VT);
2227 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2229 if (VT.isSimple() && !VT.isVector()) {
2230 MVT Simple = VT.getSimpleVT();
2231 unsigned SimpleSize = Simple.getSizeInBits();
2232 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2233 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2234 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2235 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2236 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2237 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2238 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2239 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2246 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2247 SDValue N0 = N->getOperand(0);
2248 SDValue N1 = N->getOperand(1);
2249 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2250 EVT VT = N->getValueType(0);
2253 // fold (mulhu x, 0) -> 0
2254 if (N1C && N1C->isNullValue())
2256 // fold (mulhu x, 1) -> 0
2257 if (N1C && N1C->getAPIntValue() == 1)
2258 return DAG.getConstant(0, N0.getValueType());
2259 // fold (mulhu x, undef) -> 0
2260 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2261 return DAG.getConstant(0, VT);
2263 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2265 if (VT.isSimple() && !VT.isVector()) {
2266 MVT Simple = VT.getSimpleVT();
2267 unsigned SimpleSize = Simple.getSizeInBits();
2268 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2269 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2270 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2271 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2272 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2273 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2274 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2275 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2282 /// SimplifyNodeWithTwoResults - Perform optimizations common to nodes that
2283 /// compute two values. LoOp and HiOp give the opcodes for the two computations
2284 /// that are being performed. Return true if a simplification was made.
2286 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2288 // If the high half is not needed, just compute the low half.
2289 bool HiExists = N->hasAnyUseOfValue(1);
2291 (!LegalOperations ||
2292 TLI.isOperationLegalOrCustom(LoOp, N->getValueType(0)))) {
2293 SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0),
2294 N->op_begin(), N->getNumOperands());
2295 return CombineTo(N, Res, Res);
2298 // If the low half is not needed, just compute the high half.
2299 bool LoExists = N->hasAnyUseOfValue(0);
2301 (!LegalOperations ||
2302 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2303 SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1),
2304 N->op_begin(), N->getNumOperands());
2305 return CombineTo(N, Res, Res);
2308 // If both halves are used, return as it is.
2309 if (LoExists && HiExists)
2312 // If the two computed results can be simplified separately, separate them.
2314 SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0),
2315 N->op_begin(), N->getNumOperands());
2316 AddToWorkList(Lo.getNode());
2317 SDValue LoOpt = combine(Lo.getNode());
2318 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2319 (!LegalOperations ||
2320 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2321 return CombineTo(N, LoOpt, LoOpt);
2325 SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1),
2326 N->op_begin(), N->getNumOperands());
2327 AddToWorkList(Hi.getNode());
2328 SDValue HiOpt = combine(Hi.getNode());
2329 if (HiOpt.getNode() && HiOpt != Hi &&
2330 (!LegalOperations ||
2331 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2332 return CombineTo(N, HiOpt, HiOpt);
2338 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2339 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2340 if (Res.getNode()) return Res;
2342 EVT VT = N->getValueType(0);
2345 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2347 if (VT.isSimple() && !VT.isVector()) {
2348 MVT Simple = VT.getSimpleVT();
2349 unsigned SimpleSize = Simple.getSizeInBits();
2350 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2351 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2352 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2353 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2354 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2355 // Compute the high part as N1.
2356 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2357 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2358 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2359 // Compute the low part as N0.
2360 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2361 return CombineTo(N, Lo, Hi);
2368 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2369 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2370 if (Res.getNode()) return Res;
2372 EVT VT = N->getValueType(0);
2375 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2377 if (VT.isSimple() && !VT.isVector()) {
2378 MVT Simple = VT.getSimpleVT();
2379 unsigned SimpleSize = Simple.getSizeInBits();
2380 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2381 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2382 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2383 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2384 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2385 // Compute the high part as N1.
2386 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2387 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2388 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2389 // Compute the low part as N0.
2390 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2391 return CombineTo(N, Lo, Hi);
2398 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2399 // (smulo x, 2) -> (saddo x, x)
2400 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2401 if (C2->getAPIntValue() == 2)
2402 return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(),
2403 N->getOperand(0), N->getOperand(0));
2408 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2409 // (umulo x, 2) -> (uaddo x, x)
2410 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2411 if (C2->getAPIntValue() == 2)
2412 return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(),
2413 N->getOperand(0), N->getOperand(0));
2418 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2419 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2420 if (Res.getNode()) return Res;
2425 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2426 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2427 if (Res.getNode()) return Res;
2432 /// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with
2433 /// two operands of the same opcode, try to simplify it.
2434 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2435 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2436 EVT VT = N0.getValueType();
2437 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2439 // Bail early if none of these transforms apply.
2440 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2442 // For each of OP in AND/OR/XOR:
2443 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2444 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2445 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2446 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2448 // do not sink logical op inside of a vector extend, since it may combine
2450 EVT Op0VT = N0.getOperand(0).getValueType();
2451 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2452 N0.getOpcode() == ISD::SIGN_EXTEND ||
2453 // Avoid infinite looping with PromoteIntBinOp.
2454 (N0.getOpcode() == ISD::ANY_EXTEND &&
2455 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2456 (N0.getOpcode() == ISD::TRUNCATE &&
2457 (!TLI.isZExtFree(VT, Op0VT) ||
2458 !TLI.isTruncateFree(Op0VT, VT)) &&
2459 TLI.isTypeLegal(Op0VT))) &&
2461 Op0VT == N1.getOperand(0).getValueType() &&
2462 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2463 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2464 N0.getOperand(0).getValueType(),
2465 N0.getOperand(0), N1.getOperand(0));
2466 AddToWorkList(ORNode.getNode());
2467 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode);
2470 // For each of OP in SHL/SRL/SRA/AND...
2471 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2472 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2473 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2474 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2475 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2476 N0.getOperand(1) == N1.getOperand(1)) {
2477 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2478 N0.getOperand(0).getValueType(),
2479 N0.getOperand(0), N1.getOperand(0));
2480 AddToWorkList(ORNode.getNode());
2481 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
2482 ORNode, N0.getOperand(1));
2485 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2486 // Only perform this optimization after type legalization and before
2487 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2488 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2489 // we don't want to undo this promotion.
2490 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2492 if ((N0.getOpcode() == ISD::BITCAST ||
2493 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) &&
2494 Level == AfterLegalizeTypes) {
2495 SDValue In0 = N0.getOperand(0);
2496 SDValue In1 = N1.getOperand(0);
2497 EVT In0Ty = In0.getValueType();
2498 EVT In1Ty = In1.getValueType();
2500 // If both incoming values are integers, and the original types are the
2502 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2503 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1);
2504 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op);
2505 AddToWorkList(Op.getNode());
2510 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2511 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2512 // If both shuffles use the same mask, and both shuffle within a single
2513 // vector, then it is worthwhile to move the swizzle after the operation.
2514 // The type-legalizer generates this pattern when loading illegal
2515 // vector types from memory. In many cases this allows additional shuffle
2517 // There are other cases where moving the shuffle after the xor/and/or
2518 // is profitable even if shuffles don't perform a swizzle.
2519 // If both shuffles use the same mask, and both shuffles have the same first
2520 // or second operand, then it might still be profitable to move the shuffle
2521 // after the xor/and/or operation.
2522 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG) {
2523 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2524 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2526 assert(N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() &&
2527 "Inputs to shuffles are not the same type");
2529 // Check that both shuffles use the same mask. The masks are known to be of
2530 // the same length because the result vector type is the same.
2531 // Check also that shuffles have only one use to avoid introducing extra
2533 if (SVN0->hasOneUse() && SVN1->hasOneUse() &&
2534 SVN0->getMask().equals(SVN1->getMask())) {
2535 SDValue ShOp = N0->getOperand(1);
2537 // Don't try to fold this node if it requires introducing a
2538 // build vector of all zeros that might be illegal at this stage.
2539 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2541 ShOp = DAG.getConstant(0, VT);
2546 // (AND (shuf (A, C), shuf (B, C)) -> shuf (AND (A, B), C)
2547 // (OR (shuf (A, C), shuf (B, C)) -> shuf (OR (A, B), C)
2548 // (XOR (shuf (A, C), shuf (B, C)) -> shuf (XOR (A, B), V_0)
2549 if (N0.getOperand(1) == N1.getOperand(1) && ShOp.getNode()) {
2550 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2551 N0->getOperand(0), N1->getOperand(0));
2552 AddToWorkList(NewNode.getNode());
2553 return DAG.getVectorShuffle(VT, SDLoc(N), NewNode, ShOp,
2554 &SVN0->getMask()[0]);
2557 // Don't try to fold this node if it requires introducing a
2558 // build vector of all zeros that might be illegal at this stage.
2559 ShOp = N0->getOperand(0);
2560 if (N->getOpcode() == ISD::XOR && ShOp.getOpcode() != ISD::UNDEF) {
2562 ShOp = DAG.getConstant(0, VT);
2567 // (AND (shuf (C, A), shuf (C, B)) -> shuf (C, AND (A, B))
2568 // (OR (shuf (C, A), shuf (C, B)) -> shuf (C, OR (A, B))
2569 // (XOR (shuf (C, A), shuf (C, B)) -> shuf (V_0, XOR (A, B))
2570 if (N0->getOperand(0) == N1->getOperand(0) && ShOp.getNode()) {
2571 SDValue NewNode = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2572 N0->getOperand(1), N1->getOperand(1));
2573 AddToWorkList(NewNode.getNode());
2574 return DAG.getVectorShuffle(VT, SDLoc(N), ShOp, NewNode,
2575 &SVN0->getMask()[0]);
2583 SDValue DAGCombiner::visitAND(SDNode *N) {
2584 SDValue N0 = N->getOperand(0);
2585 SDValue N1 = N->getOperand(1);
2586 SDValue LL, LR, RL, RR, CC0, CC1;
2587 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2588 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2589 EVT VT = N1.getValueType();
2590 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2593 if (VT.isVector()) {
2594 SDValue FoldedVOp = SimplifyVBinOp(N);
2595 if (FoldedVOp.getNode()) return FoldedVOp;
2597 // fold (and x, 0) -> 0, vector edition
2598 if (ISD::isBuildVectorAllZeros(N0.getNode()))
2600 if (ISD::isBuildVectorAllZeros(N1.getNode()))
2603 // fold (and x, -1) -> x, vector edition
2604 if (ISD::isBuildVectorAllOnes(N0.getNode()))
2606 if (ISD::isBuildVectorAllOnes(N1.getNode()))
2610 // fold (and x, undef) -> 0
2611 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2612 return DAG.getConstant(0, VT);
2613 // fold (and c1, c2) -> c1&c2
2615 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2616 // canonicalize constant to RHS
2618 return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0);
2619 // fold (and x, -1) -> x
2620 if (N1C && N1C->isAllOnesValue())
2622 // if (and x, c) is known to be zero, return 0
2623 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2624 APInt::getAllOnesValue(BitWidth)))
2625 return DAG.getConstant(0, VT);
2627 SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1);
2628 if (RAND.getNode() != 0)
2630 // fold (and (or x, C), D) -> D if (C & D) == D
2631 if (N1C && N0.getOpcode() == ISD::OR)
2632 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2633 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2635 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2636 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2637 SDValue N0Op0 = N0.getOperand(0);
2638 APInt Mask = ~N1C->getAPIntValue();
2639 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2640 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2641 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
2642 N0.getValueType(), N0Op0);
2644 // Replace uses of the AND with uses of the Zero extend node.
2647 // We actually want to replace all uses of the any_extend with the
2648 // zero_extend, to avoid duplicating things. This will later cause this
2649 // AND to be folded.
2650 CombineTo(N0.getNode(), Zext);
2651 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2654 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2655 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2656 // already be zero by virtue of the width of the base type of the load.
2658 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2660 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2661 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2662 N0.getOpcode() == ISD::LOAD) {
2663 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2664 N0 : N0.getOperand(0) );
2666 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2667 // This can be a pure constant or a vector splat, in which case we treat the
2668 // vector as a scalar and use the splat value.
2669 APInt Constant = APInt::getNullValue(1);
2670 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2671 Constant = C->getAPIntValue();
2672 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2673 APInt SplatValue, SplatUndef;
2674 unsigned SplatBitSize;
2676 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2677 SplatBitSize, HasAnyUndefs);
2679 // Undef bits can contribute to a possible optimisation if set, so
2681 SplatValue |= SplatUndef;
2683 // The splat value may be something like "0x00FFFFFF", which means 0 for
2684 // the first vector value and FF for the rest, repeating. We need a mask
2685 // that will apply equally to all members of the vector, so AND all the
2686 // lanes of the constant together.
2687 EVT VT = Vector->getValueType(0);
2688 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2690 // If the splat value has been compressed to a bitlength lower
2691 // than the size of the vector lane, we need to re-expand it to
2693 if (BitWidth > SplatBitSize)
2694 for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
2695 SplatBitSize < BitWidth;
2696 SplatBitSize = SplatBitSize * 2)
2697 SplatValue |= SplatValue.shl(SplatBitSize);
2699 Constant = APInt::getAllOnesValue(BitWidth);
2700 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
2701 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
2705 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
2706 // actually legal and isn't going to get expanded, else this is a false
2708 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
2709 Load->getMemoryVT());
2711 // Resize the constant to the same size as the original memory access before
2712 // extension. If it is still the AllOnesValue then this AND is completely
2715 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
2718 switch (Load->getExtensionType()) {
2719 default: B = false; break;
2720 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
2722 case ISD::NON_EXTLOAD: B = true; break;
2725 if (B && Constant.isAllOnesValue()) {
2726 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
2727 // preserve semantics once we get rid of the AND.
2728 SDValue NewLoad(Load, 0);
2729 if (Load->getExtensionType() == ISD::EXTLOAD) {
2730 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
2731 Load->getValueType(0), SDLoc(Load),
2732 Load->getChain(), Load->getBasePtr(),
2733 Load->getOffset(), Load->getMemoryVT(),
2734 Load->getMemOperand());
2735 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
2736 if (Load->getNumValues() == 3) {
2737 // PRE/POST_INC loads have 3 values.
2738 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
2739 NewLoad.getValue(2) };
2740 CombineTo(Load, To, 3, true);
2742 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
2746 // Fold the AND away, taking care not to fold to the old load node if we
2748 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
2750 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2753 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2754 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2755 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2756 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2758 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2759 LL.getValueType().isInteger()) {
2760 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2761 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
2762 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2763 LR.getValueType(), LL, RL);
2764 AddToWorkList(ORNode.getNode());
2765 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
2767 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2768 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2769 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(N0),
2770 LR.getValueType(), LL, RL);
2771 AddToWorkList(ANDNode.getNode());
2772 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
2774 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2775 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2776 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2777 LR.getValueType(), LL, RL);
2778 AddToWorkList(ORNode.getNode());
2779 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
2782 // Simplify (and (setne X, 0), (setne X, -1)) -> (setuge (add X, 1), 2)
2783 if (LL == RL && isa<ConstantSDNode>(LR) && isa<ConstantSDNode>(RR) &&
2784 Op0 == Op1 && LL.getValueType().isInteger() &&
2785 Op0 == ISD::SETNE && ((cast<ConstantSDNode>(LR)->isNullValue() &&
2786 cast<ConstantSDNode>(RR)->isAllOnesValue()) ||
2787 (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
2788 cast<ConstantSDNode>(RR)->isNullValue()))) {
2789 SDValue ADDNode = DAG.getNode(ISD::ADD, SDLoc(N0), LL.getValueType(),
2790 LL, DAG.getConstant(1, LL.getValueType()));
2791 AddToWorkList(ADDNode.getNode());
2792 return DAG.getSetCC(SDLoc(N), VT, ADDNode,
2793 DAG.getConstant(2, LL.getValueType()), ISD::SETUGE);
2795 // canonicalize equivalent to ll == rl
2796 if (LL == RR && LR == RL) {
2797 Op1 = ISD::getSetCCSwappedOperands(Op1);
2800 if (LL == RL && LR == RR) {
2801 bool isInteger = LL.getValueType().isInteger();
2802 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2803 if (Result != ISD::SETCC_INVALID &&
2804 (!LegalOperations ||
2805 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
2806 TLI.isOperationLegal(ISD::SETCC,
2807 getSetCCResultType(N0.getSimpleValueType())))))
2808 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
2813 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
2814 if (N0.getOpcode() == N1.getOpcode()) {
2815 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2816 if (Tmp.getNode()) return Tmp;
2819 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
2820 // fold (and (sra)) -> (and (srl)) when possible.
2821 if (!VT.isVector() &&
2822 SimplifyDemandedBits(SDValue(N, 0)))
2823 return SDValue(N, 0);
2825 // fold (zext_inreg (extload x)) -> (zextload x)
2826 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
2827 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2828 EVT MemVT = LN0->getMemoryVT();
2829 // If we zero all the possible extended bits, then we can turn this into
2830 // a zextload if we are running before legalize or the operation is legal.
2831 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2832 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2833 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2834 ((!LegalOperations && !LN0->isVolatile()) ||
2835 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2836 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2837 LN0->getChain(), LN0->getBasePtr(),
2838 MemVT, LN0->getMemOperand());
2840 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2841 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2844 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
2845 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
2847 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2848 EVT MemVT = LN0->getMemoryVT();
2849 // If we zero all the possible extended bits, then we can turn this into
2850 // a zextload if we are running before legalize or the operation is legal.
2851 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2852 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2853 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2854 ((!LegalOperations && !LN0->isVolatile()) ||
2855 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2856 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2857 LN0->getChain(), LN0->getBasePtr(),
2858 MemVT, LN0->getMemOperand());
2860 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2861 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2865 // fold (and (load x), 255) -> (zextload x, i8)
2866 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2867 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2868 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2869 (N0.getOpcode() == ISD::ANY_EXTEND &&
2870 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2871 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2872 LoadSDNode *LN0 = HasAnyExt
2873 ? cast<LoadSDNode>(N0.getOperand(0))
2874 : cast<LoadSDNode>(N0);
2875 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2876 LN0->isUnindexed() && N0.hasOneUse() && SDValue(LN0, 0).hasOneUse()) {
2877 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2878 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2879 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2880 EVT LoadedVT = LN0->getMemoryVT();
2882 if (ExtVT == LoadedVT &&
2883 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2884 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2887 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
2888 LN0->getChain(), LN0->getBasePtr(), ExtVT,
2889 LN0->getMemOperand());
2891 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
2892 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2895 // Do not change the width of a volatile load.
2896 // Do not generate loads of non-round integer types since these can
2897 // be expensive (and would be wrong if the type is not byte sized).
2898 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
2899 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2900 EVT PtrType = LN0->getOperand(1).getValueType();
2902 unsigned Alignment = LN0->getAlignment();
2903 SDValue NewPtr = LN0->getBasePtr();
2905 // For big endian targets, we need to add an offset to the pointer
2906 // to load the correct bytes. For little endian systems, we merely
2907 // need to read fewer bytes from the same pointer.
2908 if (TLI.isBigEndian()) {
2909 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
2910 unsigned EVTStoreBytes = ExtVT.getStoreSize();
2911 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
2912 NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0), PtrType,
2913 NewPtr, DAG.getConstant(PtrOff, PtrType));
2914 Alignment = MinAlign(Alignment, PtrOff);
2917 AddToWorkList(NewPtr.getNode());
2919 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2921 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
2922 LN0->getChain(), NewPtr,
2923 LN0->getPointerInfo(),
2924 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2925 Alignment, LN0->getTBAAInfo());
2927 CombineTo(LN0, Load, Load.getValue(1));
2928 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2934 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
2935 VT.getSizeInBits() <= 64) {
2936 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2937 APInt ADDC = ADDI->getAPIntValue();
2938 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2939 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
2940 // immediate for an add, but it is legal if its top c2 bits are set,
2941 // transform the ADD so the immediate doesn't need to be materialized
2943 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
2944 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
2945 SRLI->getZExtValue());
2946 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
2948 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2950 DAG.getNode(ISD::ADD, SDLoc(N0), VT,
2951 N0.getOperand(0), DAG.getConstant(ADDC, VT));
2952 CombineTo(N0.getNode(), NewAdd);
2953 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2961 // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const)
2962 if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) {
2963 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
2964 N0.getOperand(1), false);
2965 if (BSwap.getNode())
2972 /// MatchBSwapHWord - Match (a >> 8) | (a << 8) as (bswap a) >> 16
2974 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
2975 bool DemandHighBits) {
2976 if (!LegalOperations)
2979 EVT VT = N->getValueType(0);
2980 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
2982 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
2985 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
2986 bool LookPassAnd0 = false;
2987 bool LookPassAnd1 = false;
2988 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
2990 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
2992 if (N0.getOpcode() == ISD::AND) {
2993 if (!N0.getNode()->hasOneUse())
2995 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2996 if (!N01C || N01C->getZExtValue() != 0xFF00)
2998 N0 = N0.getOperand(0);
2999 LookPassAnd0 = true;
3002 if (N1.getOpcode() == ISD::AND) {
3003 if (!N1.getNode()->hasOneUse())
3005 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3006 if (!N11C || N11C->getZExtValue() != 0xFF)
3008 N1 = N1.getOperand(0);
3009 LookPassAnd1 = true;
3012 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
3014 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
3016 if (!N0.getNode()->hasOneUse() ||
3017 !N1.getNode()->hasOneUse())
3020 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3021 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
3024 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
3027 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
3028 SDValue N00 = N0->getOperand(0);
3029 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
3030 if (!N00.getNode()->hasOneUse())
3032 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
3033 if (!N001C || N001C->getZExtValue() != 0xFF)
3035 N00 = N00.getOperand(0);
3036 LookPassAnd0 = true;
3039 SDValue N10 = N1->getOperand(0);
3040 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
3041 if (!N10.getNode()->hasOneUse())
3043 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
3044 if (!N101C || N101C->getZExtValue() != 0xFF00)
3046 N10 = N10.getOperand(0);
3047 LookPassAnd1 = true;
3053 // Make sure everything beyond the low halfword gets set to zero since the SRL
3054 // 16 will clear the top bits.
3055 unsigned OpSizeInBits = VT.getSizeInBits();
3056 if (DemandHighBits && OpSizeInBits > 16) {
3057 // If the left-shift isn't masked out then the only way this is a bswap is
3058 // if all bits beyond the low 8 are 0. In that case the entire pattern
3059 // reduces to a left shift anyway: leave it for other parts of the combiner.
3063 // However, if the right shift isn't masked out then it might be because
3064 // it's not needed. See if we can spot that too.
3065 if (!LookPassAnd1 &&
3066 !DAG.MaskedValueIsZero(
3067 N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16)))
3071 SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00);
3072 if (OpSizeInBits > 16)
3073 Res = DAG.getNode(ISD::SRL, SDLoc(N), VT, Res,
3074 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
3078 /// isBSwapHWordElement - Return true if the specified node is an element
3079 /// that makes up a 32-bit packed halfword byteswap. i.e.
3080 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
3081 static bool isBSwapHWordElement(SDValue N, SmallVectorImpl<SDNode *> &Parts) {
3082 if (!N.getNode()->hasOneUse())
3085 unsigned Opc = N.getOpcode();
3086 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
3089 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3094 switch (N1C->getZExtValue()) {
3097 case 0xFF: Num = 0; break;
3098 case 0xFF00: Num = 1; break;
3099 case 0xFF0000: Num = 2; break;
3100 case 0xFF000000: Num = 3; break;
3103 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
3104 SDValue N0 = N.getOperand(0);
3105 if (Opc == ISD::AND) {
3106 if (Num == 0 || Num == 2) {
3108 // (x >> 8) & 0xff0000
3109 if (N0.getOpcode() != ISD::SRL)
3111 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3112 if (!C || C->getZExtValue() != 8)
3115 // (x << 8) & 0xff00
3116 // (x << 8) & 0xff000000
3117 if (N0.getOpcode() != ISD::SHL)
3119 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3120 if (!C || C->getZExtValue() != 8)
3123 } else if (Opc == ISD::SHL) {
3125 // (x & 0xff0000) << 8
3126 if (Num != 0 && Num != 2)
3128 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3129 if (!C || C->getZExtValue() != 8)
3131 } else { // Opc == ISD::SRL
3132 // (x & 0xff00) >> 8
3133 // (x & 0xff000000) >> 8
3134 if (Num != 1 && Num != 3)
3136 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3137 if (!C || C->getZExtValue() != 8)
3144 Parts[Num] = N0.getOperand(0).getNode();
3148 /// MatchBSwapHWord - Match a 32-bit packed halfword bswap. That is
3149 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
3150 /// => (rotl (bswap x), 16)
3151 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
3152 if (!LegalOperations)
3155 EVT VT = N->getValueType(0);
3158 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3161 SmallVector<SDNode*,4> Parts(4, (SDNode*)0);
3163 // (or (or (and), (and)), (or (and), (and)))
3164 // (or (or (or (and), (and)), (and)), (and))
3165 if (N0.getOpcode() != ISD::OR)
3167 SDValue N00 = N0.getOperand(0);
3168 SDValue N01 = N0.getOperand(1);
3170 if (N1.getOpcode() == ISD::OR &&
3171 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) {
3172 // (or (or (and), (and)), (or (and), (and)))
3173 SDValue N000 = N00.getOperand(0);
3174 if (!isBSwapHWordElement(N000, Parts))
3177 SDValue N001 = N00.getOperand(1);
3178 if (!isBSwapHWordElement(N001, Parts))
3180 SDValue N010 = N01.getOperand(0);
3181 if (!isBSwapHWordElement(N010, Parts))
3183 SDValue N011 = N01.getOperand(1);
3184 if (!isBSwapHWordElement(N011, Parts))
3187 // (or (or (or (and), (and)), (and)), (and))
3188 if (!isBSwapHWordElement(N1, Parts))
3190 if (!isBSwapHWordElement(N01, Parts))
3192 if (N00.getOpcode() != ISD::OR)
3194 SDValue N000 = N00.getOperand(0);
3195 if (!isBSwapHWordElement(N000, Parts))
3197 SDValue N001 = N00.getOperand(1);
3198 if (!isBSwapHWordElement(N001, Parts))
3202 // Make sure the parts are all coming from the same node.
3203 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
3206 SDValue BSwap = DAG.getNode(ISD::BSWAP, SDLoc(N), VT,
3207 SDValue(Parts[0],0));
3209 // Result of the bswap should be rotated by 16. If it's not legal, then
3210 // do (x << 16) | (x >> 16).
3211 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
3212 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3213 return DAG.getNode(ISD::ROTL, SDLoc(N), VT, BSwap, ShAmt);
3214 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3215 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, BSwap, ShAmt);
3216 return DAG.getNode(ISD::OR, SDLoc(N), VT,
3217 DAG.getNode(ISD::SHL, SDLoc(N), VT, BSwap, ShAmt),
3218 DAG.getNode(ISD::SRL, SDLoc(N), VT, BSwap, ShAmt));
3221 SDValue DAGCombiner::visitOR(SDNode *N) {
3222 SDValue N0 = N->getOperand(0);
3223 SDValue N1 = N->getOperand(1);
3224 SDValue LL, LR, RL, RR, CC0, CC1;
3225 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3226 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3227 EVT VT = N1.getValueType();
3230 if (VT.isVector()) {
3231 SDValue FoldedVOp = SimplifyVBinOp(N);
3232 if (FoldedVOp.getNode()) return FoldedVOp;
3234 // fold (or x, 0) -> x, vector edition
3235 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3237 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3240 // fold (or x, -1) -> -1, vector edition
3241 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3243 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3246 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf A, B, Mask1)
3247 // fold (or (shuf A, V_0, MA), (shuf B, V_0, MB)) -> (shuf B, A, Mask2)
3248 // Do this only if the resulting shuffle is legal.
3249 if (isa<ShuffleVectorSDNode>(N0) &&
3250 isa<ShuffleVectorSDNode>(N1) &&
3251 N0->getOperand(1) == N1->getOperand(1) &&
3252 ISD::isBuildVectorAllZeros(N0.getOperand(1).getNode())) {
3253 bool CanFold = true;
3254 unsigned NumElts = VT.getVectorNumElements();
3255 const ShuffleVectorSDNode *SV0 = cast<ShuffleVectorSDNode>(N0);
3256 const ShuffleVectorSDNode *SV1 = cast<ShuffleVectorSDNode>(N1);
3257 // We construct two shuffle masks:
3258 // - Mask1 is a shuffle mask for a shuffle with N0 as the first operand
3259 // and N1 as the second operand.
3260 // - Mask2 is a shuffle mask for a shuffle with N1 as the first operand
3261 // and N0 as the second operand.
3262 // We do this because OR is commutable and therefore there might be
3263 // two ways to fold this node into a shuffle.
3264 SmallVector<int,4> Mask1;
3265 SmallVector<int,4> Mask2;
3267 for (unsigned i = 0; i != NumElts && CanFold; ++i) {
3268 int M0 = SV0->getMaskElt(i);
3269 int M1 = SV1->getMaskElt(i);
3271 // Both shuffle indexes are undef. Propagate Undef.
3272 if (M0 < 0 && M1 < 0) {
3273 Mask1.push_back(M0);
3274 Mask2.push_back(M0);
3278 if (M0 < 0 || M1 < 0 ||
3279 (M0 < (int)NumElts && M1 < (int)NumElts) ||
3280 (M0 >= (int)NumElts && M1 >= (int)NumElts)) {
3285 Mask1.push_back(M0 < (int)NumElts ? M0 : M1 + NumElts);
3286 Mask2.push_back(M1 < (int)NumElts ? M1 : M0 + NumElts);
3290 // Fold this sequence only if the resulting shuffle is 'legal'.
3291 if (TLI.isShuffleMaskLegal(Mask1, VT))
3292 return DAG.getVectorShuffle(VT, SDLoc(N), N0->getOperand(0),
3293 N1->getOperand(0), &Mask1[0]);
3294 if (TLI.isShuffleMaskLegal(Mask2, VT))
3295 return DAG.getVectorShuffle(VT, SDLoc(N), N1->getOperand(0),
3296 N0->getOperand(0), &Mask2[0]);
3301 // fold (or x, undef) -> -1
3302 if (!LegalOperations &&
3303 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3304 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3305 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
3307 // fold (or c1, c2) -> c1|c2
3309 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
3310 // canonicalize constant to RHS
3312 return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0);
3313 // fold (or x, 0) -> x
3314 if (N1C && N1C->isNullValue())
3316 // fold (or x, -1) -> -1
3317 if (N1C && N1C->isAllOnesValue())
3319 // fold (or x, c) -> c iff (x & ~c) == 0
3320 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3323 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3324 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
3325 if (BSwap.getNode() != 0)
3327 BSwap = MatchBSwapHWordLow(N, N0, N1);
3328 if (BSwap.getNode() != 0)
3332 SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1);
3333 if (ROR.getNode() != 0)
3335 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3336 // iff (c1 & c2) == 0.
3337 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3338 isa<ConstantSDNode>(N0.getOperand(1))) {
3339 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3340 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) {
3341 SDValue COR = DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1);
3344 return DAG.getNode(ISD::AND, SDLoc(N), VT,
3345 DAG.getNode(ISD::OR, SDLoc(N0), VT,
3346 N0.getOperand(0), N1), COR);
3349 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3350 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3351 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3352 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3354 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3355 LL.getValueType().isInteger()) {
3356 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3357 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3358 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3359 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3360 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(LR),
3361 LR.getValueType(), LL, RL);
3362 AddToWorkList(ORNode.getNode());
3363 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
3365 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3366 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3367 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3368 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3369 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(LR),
3370 LR.getValueType(), LL, RL);
3371 AddToWorkList(ANDNode.getNode());
3372 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
3375 // canonicalize equivalent to ll == rl
3376 if (LL == RR && LR == RL) {
3377 Op1 = ISD::getSetCCSwappedOperands(Op1);
3380 if (LL == RL && LR == RR) {
3381 bool isInteger = LL.getValueType().isInteger();
3382 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3383 if (Result != ISD::SETCC_INVALID &&
3384 (!LegalOperations ||
3385 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
3386 TLI.isOperationLegal(ISD::SETCC,
3387 getSetCCResultType(N0.getValueType())))))
3388 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
3393 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3394 if (N0.getOpcode() == N1.getOpcode()) {
3395 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3396 if (Tmp.getNode()) return Tmp;
3399 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3400 if (N0.getOpcode() == ISD::AND &&
3401 N1.getOpcode() == ISD::AND &&
3402 N0.getOperand(1).getOpcode() == ISD::Constant &&
3403 N1.getOperand(1).getOpcode() == ISD::Constant &&
3404 // Don't increase # computations.
3405 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3406 // We can only do this xform if we know that bits from X that are set in C2
3407 // but not in C1 are already zero. Likewise for Y.
3408 const APInt &LHSMask =
3409 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3410 const APInt &RHSMask =
3411 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3413 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3414 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3415 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3416 N0.getOperand(0), N1.getOperand(0));
3417 return DAG.getNode(ISD::AND, SDLoc(N), VT, X,
3418 DAG.getConstant(LHSMask | RHSMask, VT));
3422 // See if this is some rotate idiom.
3423 if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N)))
3424 return SDValue(Rot, 0);
3426 // Simplify the operands using demanded-bits information.
3427 if (!VT.isVector() &&
3428 SimplifyDemandedBits(SDValue(N, 0)))
3429 return SDValue(N, 0);
3434 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
3435 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3436 if (Op.getOpcode() == ISD::AND) {
3437 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3438 Mask = Op.getOperand(1);
3439 Op = Op.getOperand(0);
3445 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3453 // Return true if we can prove that, whenever Neg and Pos are both in the
3454 // range [0, OpSize), Neg == (Pos == 0 ? 0 : OpSize - Pos). This means that
3455 // for two opposing shifts shift1 and shift2 and a value X with OpBits bits:
3457 // (or (shift1 X, Neg), (shift2 X, Pos))
3459 // reduces to a rotate in direction shift2 by Pos or (equivalently) a rotate
3460 // in direction shift1 by Neg. The range [0, OpSize) means that we only need
3461 // to consider shift amounts with defined behavior.
3462 static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned OpSize) {
3463 // If OpSize is a power of 2 then:
3465 // (a) (Pos == 0 ? 0 : OpSize - Pos) == (OpSize - Pos) & (OpSize - 1)
3466 // (b) Neg == Neg & (OpSize - 1) whenever Neg is in [0, OpSize).
3468 // So if OpSize is a power of 2 and Neg is (and Neg', OpSize-1), we check
3469 // for the stronger condition:
3471 // Neg & (OpSize - 1) == (OpSize - Pos) & (OpSize - 1) [A]
3473 // for all Neg and Pos. Since Neg & (OpSize - 1) == Neg' & (OpSize - 1)
3474 // we can just replace Neg with Neg' for the rest of the function.
3476 // In other cases we check for the even stronger condition:
3478 // Neg == OpSize - Pos [B]
3480 // for all Neg and Pos. Note that the (or ...) then invokes undefined
3481 // behavior if Pos == 0 (and consequently Neg == OpSize).
3483 // We could actually use [A] whenever OpSize is a power of 2, but the
3484 // only extra cases that it would match are those uninteresting ones
3485 // where Neg and Pos are never in range at the same time. E.g. for
3486 // OpSize == 32, using [A] would allow a Neg of the form (sub 64, Pos)
3487 // as well as (sub 32, Pos), but:
3489 // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos))
3491 // always invokes undefined behavior for 32-bit X.
3493 // Below, Mask == OpSize - 1 when using [A] and is all-ones otherwise.
3494 unsigned MaskLoBits = 0;
3495 if (Neg.getOpcode() == ISD::AND &&
3496 isPowerOf2_64(OpSize) &&
3497 Neg.getOperand(1).getOpcode() == ISD::Constant &&
3498 cast<ConstantSDNode>(Neg.getOperand(1))->getAPIntValue() == OpSize - 1) {
3499 Neg = Neg.getOperand(0);
3500 MaskLoBits = Log2_64(OpSize);
3503 // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1.
3504 if (Neg.getOpcode() != ISD::SUB)
3506 ConstantSDNode *NegC = dyn_cast<ConstantSDNode>(Neg.getOperand(0));
3509 SDValue NegOp1 = Neg.getOperand(1);
3511 // On the RHS of [A], if Pos is Pos' & (OpSize - 1), just replace Pos with
3512 // Pos'. The truncation is redundant for the purpose of the equality.
3514 Pos.getOpcode() == ISD::AND &&
3515 Pos.getOperand(1).getOpcode() == ISD::Constant &&
3516 cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() == OpSize - 1)
3517 Pos = Pos.getOperand(0);
3519 // The condition we need is now:
3521 // (NegC - NegOp1) & Mask == (OpSize - Pos) & Mask
3523 // If NegOp1 == Pos then we need:
3525 // OpSize & Mask == NegC & Mask
3527 // (because "x & Mask" is a truncation and distributes through subtraction).
3530 Width = NegC->getAPIntValue();
3531 // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC.
3532 // Then the condition we want to prove becomes:
3534 // (NegC - NegOp1) & Mask == (OpSize - (NegOp1 + PosC)) & Mask
3536 // which, again because "x & Mask" is a truncation, becomes:
3538 // NegC & Mask == (OpSize - PosC) & Mask
3539 // OpSize & Mask == (NegC + PosC) & Mask
3540 else if (Pos.getOpcode() == ISD::ADD &&
3541 Pos.getOperand(0) == NegOp1 &&
3542 Pos.getOperand(1).getOpcode() == ISD::Constant)
3543 Width = (cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() +
3544 NegC->getAPIntValue());
3548 // Now we just need to check that OpSize & Mask == Width & Mask.
3550 // Opsize & Mask is 0 since Mask is Opsize - 1.
3551 return Width.getLoBits(MaskLoBits) == 0;
3552 return Width == OpSize;
3555 // A subroutine of MatchRotate used once we have found an OR of two opposite
3556 // shifts of Shifted. If Neg == <operand size> - Pos then the OR reduces
3557 // to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the
3558 // former being preferred if supported. InnerPos and InnerNeg are Pos and
3559 // Neg with outer conversions stripped away.
3560 SDNode *DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos,
3561 SDValue Neg, SDValue InnerPos,
3562 SDValue InnerNeg, unsigned PosOpcode,
3563 unsigned NegOpcode, SDLoc DL) {
3564 // fold (or (shl x, (*ext y)),
3565 // (srl x, (*ext (sub 32, y)))) ->
3566 // (rotl x, y) or (rotr x, (sub 32, y))
3568 // fold (or (shl x, (*ext (sub 32, y))),
3569 // (srl x, (*ext y))) ->
3570 // (rotr x, y) or (rotl x, (sub 32, y))
3571 EVT VT = Shifted.getValueType();
3572 if (matchRotateSub(InnerPos, InnerNeg, VT.getSizeInBits())) {
3573 bool HasPos = TLI.isOperationLegalOrCustom(PosOpcode, VT);
3574 return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted,
3575 HasPos ? Pos : Neg).getNode();
3578 // fold (or (shl (*ext x), (*ext y)),
3579 // (srl (*ext x), (*ext (sub 32, y)))) ->
3580 // (*ext (rotl x, y)) or (*ext (rotr x, (sub 32, y)))
3582 // fold (or (shl (*ext x), (*ext (sub 32, y))),
3583 // (srl (*ext x), (*ext y))) ->
3584 // (*ext (rotr x, y)) or (*ext (rotl x, (sub 32, y)))
3585 if (Shifted.getOpcode() == ISD::ZERO_EXTEND ||
3586 Shifted.getOpcode() == ISD::ANY_EXTEND) {
3587 SDValue InnerShifted = Shifted.getOperand(0);
3588 EVT InnerVT = InnerShifted.getValueType();
3589 bool HasPosInner = TLI.isOperationLegalOrCustom(PosOpcode, InnerVT);
3590 if (HasPosInner || TLI.isOperationLegalOrCustom(NegOpcode, InnerVT)) {
3591 if (matchRotateSub(InnerPos, InnerNeg, InnerVT.getSizeInBits())) {
3592 SDValue V = DAG.getNode(HasPosInner ? PosOpcode : NegOpcode, DL,
3593 InnerVT, InnerShifted, HasPosInner ? Pos : Neg);
3594 return DAG.getNode(Shifted.getOpcode(), DL, VT, V).getNode();
3602 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3603 // idioms for rotate, and if the target supports rotation instructions, generate
3605 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL) {
3606 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3607 EVT VT = LHS.getValueType();
3608 if (!TLI.isTypeLegal(VT)) return 0;
3610 // The target must have at least one rotate flavor.
3611 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3612 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3613 if (!HasROTL && !HasROTR) return 0;
3615 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3616 SDValue LHSShift; // The shift.
3617 SDValue LHSMask; // AND value if any.
3618 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3619 return 0; // Not part of a rotate.
3621 SDValue RHSShift; // The shift.
3622 SDValue RHSMask; // AND value if any.
3623 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3624 return 0; // Not part of a rotate.
3626 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3627 return 0; // Not shifting the same value.
3629 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3630 return 0; // Shifts must disagree.
3632 // Canonicalize shl to left side in a shl/srl pair.
3633 if (RHSShift.getOpcode() == ISD::SHL) {
3634 std::swap(LHS, RHS);
3635 std::swap(LHSShift, RHSShift);
3636 std::swap(LHSMask , RHSMask );
3639 unsigned OpSizeInBits = VT.getSizeInBits();
3640 SDValue LHSShiftArg = LHSShift.getOperand(0);
3641 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3642 SDValue RHSShiftArg = RHSShift.getOperand(0);
3643 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3645 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3646 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3647 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3648 RHSShiftAmt.getOpcode() == ISD::Constant) {
3649 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3650 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3651 if ((LShVal + RShVal) != OpSizeInBits)
3654 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3655 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
3657 // If there is an AND of either shifted operand, apply it to the result.
3658 if (LHSMask.getNode() || RHSMask.getNode()) {
3659 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3661 if (LHSMask.getNode()) {
3662 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3663 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3665 if (RHSMask.getNode()) {
3666 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3667 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3670 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3673 return Rot.getNode();
3676 // If there is a mask here, and we have a variable shift, we can't be sure
3677 // that we're masking out the right stuff.
3678 if (LHSMask.getNode() || RHSMask.getNode())
3681 // If the shift amount is sign/zext/any-extended just peel it off.
3682 SDValue LExtOp0 = LHSShiftAmt;
3683 SDValue RExtOp0 = RHSShiftAmt;
3684 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3685 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3686 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3687 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3688 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3689 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3690 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3691 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3692 LExtOp0 = LHSShiftAmt.getOperand(0);
3693 RExtOp0 = RHSShiftAmt.getOperand(0);
3696 SDNode *TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt,
3697 LExtOp0, RExtOp0, ISD::ROTL, ISD::ROTR, DL);
3701 SDNode *TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt,
3702 RExtOp0, LExtOp0, ISD::ROTR, ISD::ROTL, DL);
3709 SDValue DAGCombiner::visitXOR(SDNode *N) {
3710 SDValue N0 = N->getOperand(0);
3711 SDValue N1 = N->getOperand(1);
3712 SDValue LHS, RHS, CC;
3713 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3714 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3715 EVT VT = N0.getValueType();
3718 if (VT.isVector()) {
3719 SDValue FoldedVOp = SimplifyVBinOp(N);
3720 if (FoldedVOp.getNode()) return FoldedVOp;
3722 // fold (xor x, 0) -> x, vector edition
3723 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3725 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3729 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3730 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3731 return DAG.getConstant(0, VT);
3732 // fold (xor x, undef) -> undef
3733 if (N0.getOpcode() == ISD::UNDEF)
3735 if (N1.getOpcode() == ISD::UNDEF)
3737 // fold (xor c1, c2) -> c1^c2
3739 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3740 // canonicalize constant to RHS
3742 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
3743 // fold (xor x, 0) -> x
3744 if (N1C && N1C->isNullValue())
3747 SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1);
3748 if (RXOR.getNode() != 0)
3751 // fold !(x cc y) -> (x !cc y)
3752 if (N1C && N1C->getAPIntValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3753 bool isInt = LHS.getValueType().isInteger();
3754 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3757 if (!LegalOperations ||
3758 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) {
3759 switch (N0.getOpcode()) {
3761 llvm_unreachable("Unhandled SetCC Equivalent!");
3763 return DAG.getSetCC(SDLoc(N), VT, LHS, RHS, NotCC);
3764 case ISD::SELECT_CC:
3765 return DAG.getSelectCC(SDLoc(N), LHS, RHS, N0.getOperand(2),
3766 N0.getOperand(3), NotCC);
3771 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3772 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3773 N0.getNode()->hasOneUse() &&
3774 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3775 SDValue V = N0.getOperand(0);
3776 V = DAG.getNode(ISD::XOR, SDLoc(N0), V.getValueType(), V,
3777 DAG.getConstant(1, V.getValueType()));
3778 AddToWorkList(V.getNode());
3779 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V);
3782 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3783 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3784 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3785 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3786 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3787 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3788 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3789 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3790 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3791 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3794 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3795 if (N1C && N1C->isAllOnesValue() &&
3796 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3797 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3798 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3799 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3800 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3801 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3802 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3803 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3806 // fold (xor (and x, y), y) -> (and (not x), y)
3807 if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3808 N0->getOperand(1) == N1) {
3809 SDValue X = N0->getOperand(0);
3810 SDValue NotX = DAG.getNOT(SDLoc(X), X, VT);
3811 AddToWorkList(NotX.getNode());
3812 return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1);
3814 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3815 if (N1C && N0.getOpcode() == ISD::XOR) {
3816 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3817 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3819 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(1),
3820 DAG.getConstant(N1C->getAPIntValue() ^
3821 N00C->getAPIntValue(), VT));
3823 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(0),
3824 DAG.getConstant(N1C->getAPIntValue() ^
3825 N01C->getAPIntValue(), VT));
3827 // fold (xor x, x) -> 0
3829 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
3831 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
3832 if (N0.getOpcode() == N1.getOpcode()) {
3833 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3834 if (Tmp.getNode()) return Tmp;
3837 // Simplify the expression using non-local knowledge.
3838 if (!VT.isVector() &&
3839 SimplifyDemandedBits(SDValue(N, 0)))
3840 return SDValue(N, 0);
3845 /// visitShiftByConstant - Handle transforms common to the three shifts, when
3846 /// the shift amount is a constant.
3847 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, ConstantSDNode *Amt) {
3848 // We can't and shouldn't fold opaque constants.
3849 if (Amt->isOpaque())
3852 SDNode *LHS = N->getOperand(0).getNode();
3853 if (!LHS->hasOneUse()) return SDValue();
3855 // We want to pull some binops through shifts, so that we have (and (shift))
3856 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
3857 // thing happens with address calculations, so it's important to canonicalize
3859 bool HighBitSet = false; // Can we transform this if the high bit is set?
3861 switch (LHS->getOpcode()) {
3862 default: return SDValue();
3865 HighBitSet = false; // We can only transform sra if the high bit is clear.
3868 HighBitSet = true; // We can only transform sra if the high bit is set.
3871 if (N->getOpcode() != ISD::SHL)
3872 return SDValue(); // only shl(add) not sr[al](add).
3873 HighBitSet = false; // We can only transform sra if the high bit is clear.
3877 // We require the RHS of the binop to be a constant and not opaque as well.
3878 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
3879 if (!BinOpCst || BinOpCst->isOpaque()) return SDValue();
3881 // FIXME: disable this unless the input to the binop is a shift by a constant.
3882 // If it is not a shift, it pessimizes some common cases like:
3884 // void foo(int *X, int i) { X[i & 1235] = 1; }
3885 // int bar(int *X, int i) { return X[i & 255]; }
3886 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
3887 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
3888 BinOpLHSVal->getOpcode() != ISD::SRA &&
3889 BinOpLHSVal->getOpcode() != ISD::SRL) ||
3890 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
3893 EVT VT = N->getValueType(0);
3895 // If this is a signed shift right, and the high bit is modified by the
3896 // logical operation, do not perform the transformation. The highBitSet
3897 // boolean indicates the value of the high bit of the constant which would
3898 // cause it to be modified for this operation.
3899 if (N->getOpcode() == ISD::SRA) {
3900 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
3901 if (BinOpRHSSignSet != HighBitSet)
3905 // Fold the constants, shifting the binop RHS by the shift amount.
3906 SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)),
3908 LHS->getOperand(1), N->getOperand(1));
3909 assert(isa<ConstantSDNode>(NewRHS) && "Folding was not successful!");
3911 // Create the new shift.
3912 SDValue NewShift = DAG.getNode(N->getOpcode(),
3913 SDLoc(LHS->getOperand(0)),
3914 VT, LHS->getOperand(0), N->getOperand(1));
3916 // Create the new binop.
3917 return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS);
3920 SDValue DAGCombiner::distributeTruncateThroughAnd(SDNode *N) {
3921 assert(N->getOpcode() == ISD::TRUNCATE);
3922 assert(N->getOperand(0).getOpcode() == ISD::AND);
3924 // (truncate:TruncVT (and N00, N01C)) -> (and (truncate:TruncVT N00), TruncC)
3925 if (N->hasOneUse() && N->getOperand(0).hasOneUse()) {
3926 SDValue N01 = N->getOperand(0).getOperand(1);
3928 if (ConstantSDNode *N01C = isConstOrConstSplat(N01)) {
3929 EVT TruncVT = N->getValueType(0);
3930 SDValue N00 = N->getOperand(0).getOperand(0);
3931 APInt TruncC = N01C->getAPIntValue();
3932 TruncC = TruncC.trunc(TruncVT.getScalarSizeInBits());
3934 return DAG.getNode(ISD::AND, SDLoc(N), TruncVT,
3935 DAG.getNode(ISD::TRUNCATE, SDLoc(N), TruncVT, N00),
3936 DAG.getConstant(TruncC, TruncVT));
3943 SDValue DAGCombiner::visitRotate(SDNode *N) {
3944 // fold (rot* x, (trunc (and y, c))) -> (rot* x, (and (trunc y), (trunc c))).
3945 if (N->getOperand(1).getOpcode() == ISD::TRUNCATE &&
3946 N->getOperand(1).getOperand(0).getOpcode() == ISD::AND) {
3947 SDValue NewOp1 = distributeTruncateThroughAnd(N->getOperand(1).getNode());
3948 if (NewOp1.getNode())
3949 return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0),
3950 N->getOperand(0), NewOp1);
3955 SDValue DAGCombiner::visitSHL(SDNode *N) {
3956 SDValue N0 = N->getOperand(0);
3957 SDValue N1 = N->getOperand(1);
3958 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3959 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3960 EVT VT = N0.getValueType();
3961 unsigned OpSizeInBits = VT.getScalarSizeInBits();
3964 if (VT.isVector()) {
3965 SDValue FoldedVOp = SimplifyVBinOp(N);
3966 if (FoldedVOp.getNode()) return FoldedVOp;
3968 BuildVectorSDNode *N1CV = dyn_cast<BuildVectorSDNode>(N1);
3969 // If setcc produces all-one true value then:
3970 // (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<<N1CV)
3971 if (N1CV && N1CV->isConstant()) {
3972 if (N0.getOpcode() == ISD::AND &&
3973 TLI.getBooleanContents(true) ==
3974 TargetLowering::ZeroOrNegativeOneBooleanContent) {
3975 SDValue N00 = N0->getOperand(0);
3976 SDValue N01 = N0->getOperand(1);
3977 BuildVectorSDNode *N01CV = dyn_cast<BuildVectorSDNode>(N01);
3979 if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC) {
3980 SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, VT, N01CV, N1CV);
3982 return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C);
3985 N1C = isConstOrConstSplat(N1);
3990 // fold (shl c1, c2) -> c1<<c2
3992 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
3993 // fold (shl 0, x) -> 0
3994 if (N0C && N0C->isNullValue())
3996 // fold (shl x, c >= size(x)) -> undef
3997 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3998 return DAG.getUNDEF(VT);
3999 // fold (shl x, 0) -> x
4000 if (N1C && N1C->isNullValue())
4002 // fold (shl undef, x) -> 0
4003 if (N0.getOpcode() == ISD::UNDEF)
4004 return DAG.getConstant(0, VT);
4005 // if (shl x, c) is known to be zero, return 0
4006 if (DAG.MaskedValueIsZero(SDValue(N, 0),
4007 APInt::getAllOnesValue(OpSizeInBits)))
4008 return DAG.getConstant(0, VT);
4009 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
4010 if (N1.getOpcode() == ISD::TRUNCATE &&
4011 N1.getOperand(0).getOpcode() == ISD::AND) {
4012 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4013 if (NewOp1.getNode())
4014 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0, NewOp1);
4017 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4018 return SDValue(N, 0);
4020 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
4021 if (N1C && N0.getOpcode() == ISD::SHL) {
4022 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4023 uint64_t c1 = N0C1->getZExtValue();
4024 uint64_t c2 = N1C->getZExtValue();
4025 if (c1 + c2 >= OpSizeInBits)
4026 return DAG.getConstant(0, VT);
4027 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4028 DAG.getConstant(c1 + c2, N1.getValueType()));
4032 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
4033 // For this to be valid, the second form must not preserve any of the bits
4034 // that are shifted out by the inner shift in the first form. This means
4035 // the outer shift size must be >= the number of bits added by the ext.
4036 // As a corollary, we don't care what kind of ext it is.
4037 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
4038 N0.getOpcode() == ISD::ANY_EXTEND ||
4039 N0.getOpcode() == ISD::SIGN_EXTEND) &&
4040 N0.getOperand(0).getOpcode() == ISD::SHL) {
4041 SDValue N0Op0 = N0.getOperand(0);
4042 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4043 uint64_t c1 = N0Op0C1->getZExtValue();
4044 uint64_t c2 = N1C->getZExtValue();
4045 EVT InnerShiftVT = N0Op0.getValueType();
4046 uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits();
4047 if (c2 >= OpSizeInBits - InnerShiftSize) {
4048 if (c1 + c2 >= OpSizeInBits)
4049 return DAG.getConstant(0, VT);
4050 return DAG.getNode(ISD::SHL, SDLoc(N0), VT,
4051 DAG.getNode(N0.getOpcode(), SDLoc(N0), VT,
4052 N0Op0->getOperand(0)),
4053 DAG.getConstant(c1 + c2, N1.getValueType()));
4058 // fold (shl (zext (srl x, C)), C) -> (zext (shl (srl x, C), C))
4059 // Only fold this if the inner zext has no other uses to avoid increasing
4060 // the total number of instructions.
4061 if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() &&
4062 N0.getOperand(0).getOpcode() == ISD::SRL) {
4063 SDValue N0Op0 = N0.getOperand(0);
4064 if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
4065 uint64_t c1 = N0Op0C1->getZExtValue();
4066 if (c1 < VT.getScalarSizeInBits()) {
4067 uint64_t c2 = N1C->getZExtValue();
4069 SDValue NewOp0 = N0.getOperand(0);
4070 EVT CountVT = NewOp0.getOperand(1).getValueType();
4071 SDValue NewSHL = DAG.getNode(ISD::SHL, SDLoc(N), NewOp0.getValueType(),
4072 NewOp0, DAG.getConstant(c2, CountVT));
4073 AddToWorkList(NewSHL.getNode());
4074 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
4080 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
4081 // (and (srl x, (sub c1, c2), MASK)
4082 // Only fold this if the inner shift has no other uses -- if it does, folding
4083 // this will increase the total number of instructions.
4084 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
4085 if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
4086 uint64_t c1 = N0C1->getZExtValue();
4087 if (c1 < OpSizeInBits) {
4088 uint64_t c2 = N1C->getZExtValue();
4089 APInt Mask = APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - c1);
4092 Mask = Mask.shl(c2 - c1);
4093 Shift = DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
4094 DAG.getConstant(c2 - c1, N1.getValueType()));
4096 Mask = Mask.lshr(c1 - c2);
4097 Shift = DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4098 DAG.getConstant(c1 - c2, N1.getValueType()));
4100 return DAG.getNode(ISD::AND, SDLoc(N0), VT, Shift,
4101 DAG.getConstant(Mask, VT));
4105 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
4106 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
4107 unsigned BitSize = VT.getScalarSizeInBits();
4108 SDValue HiBitsMask =
4109 DAG.getConstant(APInt::getHighBitsSet(BitSize,
4110 BitSize - N1C->getZExtValue()), VT);
4111 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4116 SDValue NewSHL = visitShiftByConstant(N, N1C);
4117 if (NewSHL.getNode())
4124 SDValue DAGCombiner::visitSRA(SDNode *N) {
4125 SDValue N0 = N->getOperand(0);
4126 SDValue N1 = N->getOperand(1);
4127 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4128 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4129 EVT VT = N0.getValueType();
4130 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4133 if (VT.isVector()) {
4134 SDValue FoldedVOp = SimplifyVBinOp(N);
4135 if (FoldedVOp.getNode()) return FoldedVOp;
4137 N1C = isConstOrConstSplat(N1);
4140 // fold (sra c1, c2) -> (sra c1, c2)
4142 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
4143 // fold (sra 0, x) -> 0
4144 if (N0C && N0C->isNullValue())
4146 // fold (sra -1, x) -> -1
4147 if (N0C && N0C->isAllOnesValue())
4149 // fold (sra x, (setge c, size(x))) -> undef
4150 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4151 return DAG.getUNDEF(VT);
4152 // fold (sra x, 0) -> x
4153 if (N1C && N1C->isNullValue())
4155 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
4157 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
4158 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
4159 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
4161 ExtVT = EVT::getVectorVT(*DAG.getContext(),
4162 ExtVT, VT.getVectorNumElements());
4163 if ((!LegalOperations ||
4164 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
4165 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
4166 N0.getOperand(0), DAG.getValueType(ExtVT));
4169 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
4170 if (N1C && N0.getOpcode() == ISD::SRA) {
4171 if (ConstantSDNode *C1 = isConstOrConstSplat(N0.getOperand(1))) {
4172 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
4173 if (Sum >= OpSizeInBits)
4174 Sum = OpSizeInBits - 1;
4175 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0.getOperand(0),
4176 DAG.getConstant(Sum, N1.getValueType()));
4180 // fold (sra (shl X, m), (sub result_size, n))
4181 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
4182 // result_size - n != m.
4183 // If truncate is free for the target sext(shl) is likely to result in better
4185 if (N0.getOpcode() == ISD::SHL && N1C) {
4186 // Get the two constanst of the shifts, CN0 = m, CN = n.
4187 const ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1));
4189 LLVMContext &Ctx = *DAG.getContext();
4190 // Determine what the truncate's result bitsize and type would be.
4191 EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - N1C->getZExtValue());
4194 TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorNumElements());
4196 // Determine the residual right-shift amount.
4197 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
4199 // If the shift is not a no-op (in which case this should be just a sign
4200 // extend already), the truncated to type is legal, sign_extend is legal
4201 // on that type, and the truncate to that type is both legal and free,
4202 // perform the transform.
4203 if ((ShiftAmt > 0) &&
4204 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
4205 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
4206 TLI.isTruncateFree(VT, TruncVT)) {
4208 SDValue Amt = DAG.getConstant(ShiftAmt,
4209 getShiftAmountTy(N0.getOperand(0).getValueType()));
4210 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0), VT,
4211 N0.getOperand(0), Amt);
4212 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), TruncVT,
4214 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N),
4215 N->getValueType(0), Trunc);
4220 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
4221 if (N1.getOpcode() == ISD::TRUNCATE &&
4222 N1.getOperand(0).getOpcode() == ISD::AND) {
4223 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4224 if (NewOp1.getNode())
4225 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, NewOp1);
4228 // fold (sra (trunc (srl x, c1)), c2) -> (trunc (sra x, c1 + c2))
4229 // if c1 is equal to the number of bits the trunc removes
4230 if (N0.getOpcode() == ISD::TRUNCATE &&
4231 (N0.getOperand(0).getOpcode() == ISD::SRL ||
4232 N0.getOperand(0).getOpcode() == ISD::SRA) &&
4233 N0.getOperand(0).hasOneUse() &&
4234 N0.getOperand(0).getOperand(1).hasOneUse() &&
4236 SDValue N0Op0 = N0.getOperand(0);
4237 if (ConstantSDNode *LargeShift = isConstOrConstSplat(N0Op0.getOperand(1))) {
4238 unsigned LargeShiftVal = LargeShift->getZExtValue();
4239 EVT LargeVT = N0Op0.getValueType();
4241 if (LargeVT.getScalarSizeInBits() - OpSizeInBits == LargeShiftVal) {
4243 DAG.getConstant(LargeShiftVal + N1C->getZExtValue(),
4244 getShiftAmountTy(N0Op0.getOperand(0).getValueType()));
4245 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), LargeVT,
4246 N0Op0.getOperand(0), Amt);
4247 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, SRA);
4252 // Simplify, based on bits shifted out of the LHS.
4253 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4254 return SDValue(N, 0);
4257 // If the sign bit is known to be zero, switch this to a SRL.
4258 if (DAG.SignBitIsZero(N0))
4259 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1);
4262 SDValue NewSRA = visitShiftByConstant(N, N1C);
4263 if (NewSRA.getNode())
4270 SDValue DAGCombiner::visitSRL(SDNode *N) {
4271 SDValue N0 = N->getOperand(0);
4272 SDValue N1 = N->getOperand(1);
4273 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4274 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4275 EVT VT = N0.getValueType();
4276 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4279 if (VT.isVector()) {
4280 SDValue FoldedVOp = SimplifyVBinOp(N);
4281 if (FoldedVOp.getNode()) return FoldedVOp;
4283 N1C = isConstOrConstSplat(N1);
4286 // fold (srl c1, c2) -> c1 >>u c2
4288 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
4289 // fold (srl 0, x) -> 0
4290 if (N0C && N0C->isNullValue())
4292 // fold (srl x, c >= size(x)) -> undef
4293 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4294 return DAG.getUNDEF(VT);
4295 // fold (srl x, 0) -> x
4296 if (N1C && N1C->isNullValue())
4298 // if (srl x, c) is known to be zero, return 0
4299 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
4300 APInt::getAllOnesValue(OpSizeInBits)))
4301 return DAG.getConstant(0, VT);
4303 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
4304 if (N1C && N0.getOpcode() == ISD::SRL) {
4305 if (ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1))) {
4306 uint64_t c1 = N01C->getZExtValue();
4307 uint64_t c2 = N1C->getZExtValue();
4308 if (c1 + c2 >= OpSizeInBits)
4309 return DAG.getConstant(0, VT);
4310 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4311 DAG.getConstant(c1 + c2, N1.getValueType()));
4315 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
4316 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
4317 N0.getOperand(0).getOpcode() == ISD::SRL &&
4318 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
4320 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
4321 uint64_t c2 = N1C->getZExtValue();
4322 EVT InnerShiftVT = N0.getOperand(0).getValueType();
4323 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
4324 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
4325 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
4326 if (c1 + OpSizeInBits == InnerShiftSize) {
4327 if (c1 + c2 >= InnerShiftSize)
4328 return DAG.getConstant(0, VT);
4329 return DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT,
4330 DAG.getNode(ISD::SRL, SDLoc(N0), InnerShiftVT,
4331 N0.getOperand(0)->getOperand(0),
4332 DAG.getConstant(c1 + c2, ShiftCountVT)));
4336 // fold (srl (shl x, c), c) -> (and x, cst2)
4337 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1) {
4338 unsigned BitSize = N0.getScalarValueSizeInBits();
4339 if (BitSize <= 64) {
4340 uint64_t ShAmt = N1C->getZExtValue() + 64 - BitSize;
4341 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4342 DAG.getConstant(~0ULL >> ShAmt, VT));
4346 // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask)
4347 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
4348 // Shifting in all undef bits?
4349 EVT SmallVT = N0.getOperand(0).getValueType();
4350 unsigned BitSize = SmallVT.getScalarSizeInBits();
4351 if (N1C->getZExtValue() >= BitSize)
4352 return DAG.getUNDEF(VT);
4354 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
4355 uint64_t ShiftAmt = N1C->getZExtValue();
4356 SDValue SmallShift = DAG.getNode(ISD::SRL, SDLoc(N0), SmallVT,
4358 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
4359 AddToWorkList(SmallShift.getNode());
4360 APInt Mask = APInt::getAllOnesValue(OpSizeInBits).lshr(ShiftAmt);
4361 return DAG.getNode(ISD::AND, SDLoc(N), VT,
4362 DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, SmallShift),
4363 DAG.getConstant(Mask, VT));
4367 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
4368 // bit, which is unmodified by sra.
4369 if (N1C && N1C->getZExtValue() + 1 == OpSizeInBits) {
4370 if (N0.getOpcode() == ISD::SRA)
4371 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1);
4374 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
4375 if (N1C && N0.getOpcode() == ISD::CTLZ &&
4376 N1C->getAPIntValue() == Log2_32(OpSizeInBits)) {
4377 APInt KnownZero, KnownOne;
4378 DAG.ComputeMaskedBits(N0.getOperand(0), KnownZero, KnownOne);
4380 // If any of the input bits are KnownOne, then the input couldn't be all
4381 // zeros, thus the result of the srl will always be zero.
4382 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
4384 // If all of the bits input the to ctlz node are known to be zero, then
4385 // the result of the ctlz is "32" and the result of the shift is one.
4386 APInt UnknownBits = ~KnownZero;
4387 if (UnknownBits == 0) return DAG.getConstant(1, VT);
4389 // Otherwise, check to see if there is exactly one bit input to the ctlz.
4390 if ((UnknownBits & (UnknownBits - 1)) == 0) {
4391 // Okay, we know that only that the single bit specified by UnknownBits
4392 // could be set on input to the CTLZ node. If this bit is set, the SRL
4393 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
4394 // to an SRL/XOR pair, which is likely to simplify more.
4395 unsigned ShAmt = UnknownBits.countTrailingZeros();
4396 SDValue Op = N0.getOperand(0);
4399 Op = DAG.getNode(ISD::SRL, SDLoc(N0), VT, Op,
4400 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
4401 AddToWorkList(Op.getNode());
4404 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
4405 Op, DAG.getConstant(1, VT));
4409 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
4410 if (N1.getOpcode() == ISD::TRUNCATE &&
4411 N1.getOperand(0).getOpcode() == ISD::AND) {
4412 SDValue NewOp1 = distributeTruncateThroughAnd(N1.getNode());
4413 if (NewOp1.getNode())
4414 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, NewOp1);
4417 // fold operands of srl based on knowledge that the low bits are not
4419 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4420 return SDValue(N, 0);
4423 SDValue NewSRL = visitShiftByConstant(N, N1C);
4424 if (NewSRL.getNode())
4428 // Attempt to convert a srl of a load into a narrower zero-extending load.
4429 SDValue NarrowLoad = ReduceLoadWidth(N);
4430 if (NarrowLoad.getNode())
4433 // Here is a common situation. We want to optimize:
4436 // %b = and i32 %a, 2
4437 // %c = srl i32 %b, 1
4438 // brcond i32 %c ...
4444 // %c = setcc eq %b, 0
4447 // However when after the source operand of SRL is optimized into AND, the SRL
4448 // itself may not be optimized further. Look for it and add the BRCOND into
4450 if (N->hasOneUse()) {
4451 SDNode *Use = *N->use_begin();
4452 if (Use->getOpcode() == ISD::BRCOND)
4454 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
4455 // Also look pass the truncate.
4456 Use = *Use->use_begin();
4457 if (Use->getOpcode() == ISD::BRCOND)
4465 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
4466 SDValue N0 = N->getOperand(0);
4467 EVT VT = N->getValueType(0);
4469 // fold (ctlz c1) -> c2
4470 if (isa<ConstantSDNode>(N0))
4471 return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0);
4475 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4476 SDValue N0 = N->getOperand(0);
4477 EVT VT = N->getValueType(0);
4479 // fold (ctlz_zero_undef c1) -> c2
4480 if (isa<ConstantSDNode>(N0))
4481 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4485 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4486 SDValue N0 = N->getOperand(0);
4487 EVT VT = N->getValueType(0);
4489 // fold (cttz c1) -> c2
4490 if (isa<ConstantSDNode>(N0))
4491 return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0);
4495 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4496 SDValue N0 = N->getOperand(0);
4497 EVT VT = N->getValueType(0);
4499 // fold (cttz_zero_undef c1) -> c2
4500 if (isa<ConstantSDNode>(N0))
4501 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4505 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4506 SDValue N0 = N->getOperand(0);
4507 EVT VT = N->getValueType(0);
4509 // fold (ctpop c1) -> c2
4510 if (isa<ConstantSDNode>(N0))
4511 return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0);
4515 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4516 SDValue N0 = N->getOperand(0);
4517 SDValue N1 = N->getOperand(1);
4518 SDValue N2 = N->getOperand(2);
4519 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4520 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4521 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
4522 EVT VT = N->getValueType(0);
4523 EVT VT0 = N0.getValueType();
4525 // fold (select C, X, X) -> X
4528 // fold (select true, X, Y) -> X
4529 if (N0C && !N0C->isNullValue())
4531 // fold (select false, X, Y) -> Y
4532 if (N0C && N0C->isNullValue())
4534 // fold (select C, 1, X) -> (or C, X)
4535 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4536 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4537 // fold (select C, 0, 1) -> (xor C, 1)
4538 if (VT.isInteger() &&
4541 TLI.getBooleanContents(false) ==
4542 TargetLowering::ZeroOrOneBooleanContent)) &&
4543 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4546 return DAG.getNode(ISD::XOR, SDLoc(N), VT0,
4547 N0, DAG.getConstant(1, VT0));
4548 XORNode = DAG.getNode(ISD::XOR, SDLoc(N0), VT0,
4549 N0, DAG.getConstant(1, VT0));
4550 AddToWorkList(XORNode.getNode());
4552 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, XORNode);
4553 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, XORNode);
4555 // fold (select C, 0, X) -> (and (not C), X)
4556 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4557 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4558 AddToWorkList(NOTNode.getNode());
4559 return DAG.getNode(ISD::AND, SDLoc(N), VT, NOTNode, N2);
4561 // fold (select C, X, 1) -> (or (not C), X)
4562 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4563 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4564 AddToWorkList(NOTNode.getNode());
4565 return DAG.getNode(ISD::OR, SDLoc(N), VT, NOTNode, N1);
4567 // fold (select C, X, 0) -> (and C, X)
4568 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4569 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4570 // fold (select X, X, Y) -> (or X, Y)
4571 // fold (select X, 1, Y) -> (or X, Y)
4572 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4573 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4574 // fold (select X, Y, X) -> (and X, Y)
4575 // fold (select X, Y, 0) -> (and X, Y)
4576 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4577 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4579 // If we can fold this based on the true/false value, do so.
4580 if (SimplifySelectOps(N, N1, N2))
4581 return SDValue(N, 0); // Don't revisit N.
4583 // fold selects based on a setcc into other things, such as min/max/abs
4584 if (N0.getOpcode() == ISD::SETCC) {
4586 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
4587 // having to say they don't support SELECT_CC on every type the DAG knows
4588 // about, since there is no way to mark an opcode illegal at all value types
4589 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other) &&
4590 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT))
4591 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT,
4592 N0.getOperand(0), N0.getOperand(1),
4593 N1, N2, N0.getOperand(2));
4594 return SimplifySelect(SDLoc(N), N0, N1, N2);
4601 std::pair<SDValue, SDValue> SplitVSETCC(const SDNode *N, SelectionDAG &DAG) {
4604 std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
4606 // Split the inputs.
4607 SDValue Lo, Hi, LL, LH, RL, RH;
4608 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
4609 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
4611 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));
4612 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));
4614 return std::make_pair(Lo, Hi);
4617 SDValue DAGCombiner::visitVSELECT(SDNode *N) {
4618 SDValue N0 = N->getOperand(0);
4619 SDValue N1 = N->getOperand(1);
4620 SDValue N2 = N->getOperand(2);
4623 // Canonicalize integer abs.
4624 // vselect (setg[te] X, 0), X, -X ->
4625 // vselect (setgt X, -1), X, -X ->
4626 // vselect (setl[te] X, 0), -X, X ->
4627 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
4628 if (N0.getOpcode() == ISD::SETCC) {
4629 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
4630 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
4632 bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
4634 if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
4635 (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) &&
4636 N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1))
4637 isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode());
4638 else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) &&
4639 N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1))
4640 isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode());
4643 EVT VT = LHS.getValueType();
4644 SDValue Shift = DAG.getNode(
4645 ISD::SRA, DL, VT, LHS,
4646 DAG.getConstant(VT.getScalarType().getSizeInBits() - 1, VT));
4647 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift);
4648 AddToWorkList(Shift.getNode());
4649 AddToWorkList(Add.getNode());
4650 return DAG.getNode(ISD::XOR, DL, VT, Add, Shift);
4654 // If the VSELECT result requires splitting and the mask is provided by a
4655 // SETCC, then split both nodes and its operands before legalization. This
4656 // prevents the type legalizer from unrolling SETCC into scalar comparisons
4657 // and enables future optimizations (e.g. min/max pattern matching on X86).
4658 if (N0.getOpcode() == ISD::SETCC) {
4659 EVT VT = N->getValueType(0);
4661 // Check if any splitting is required.
4662 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
4663 TargetLowering::TypeSplitVector)
4666 SDValue Lo, Hi, CCLo, CCHi, LL, LH, RL, RH;
4667 std::tie(CCLo, CCHi) = SplitVSETCC(N0.getNode(), DAG);
4668 std::tie(LL, LH) = DAG.SplitVectorOperand(N, 1);
4669 std::tie(RL, RH) = DAG.SplitVectorOperand(N, 2);
4671 Lo = DAG.getNode(N->getOpcode(), DL, LL.getValueType(), CCLo, LL, RL);
4672 Hi = DAG.getNode(N->getOpcode(), DL, LH.getValueType(), CCHi, LH, RH);
4674 // Add the new VSELECT nodes to the work list in case they need to be split
4676 AddToWorkList(Lo.getNode());
4677 AddToWorkList(Hi.getNode());
4679 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
4682 // Fold (vselect (build_vector all_ones), N1, N2) -> N1
4683 if (ISD::isBuildVectorAllOnes(N0.getNode()))
4685 // Fold (vselect (build_vector all_zeros), N1, N2) -> N2
4686 if (ISD::isBuildVectorAllZeros(N0.getNode()))
4692 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
4693 SDValue N0 = N->getOperand(0);
4694 SDValue N1 = N->getOperand(1);
4695 SDValue N2 = N->getOperand(2);
4696 SDValue N3 = N->getOperand(3);
4697 SDValue N4 = N->getOperand(4);
4698 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
4700 // fold select_cc lhs, rhs, x, x, cc -> x
4704 // Determine if the condition we're dealing with is constant
4705 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
4706 N0, N1, CC, SDLoc(N), false);
4707 if (SCC.getNode()) {
4708 AddToWorkList(SCC.getNode());
4710 if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) {
4711 if (!SCCC->isNullValue())
4712 return N2; // cond always true -> true val
4714 return N3; // cond always false -> false val
4717 // Fold to a simpler select_cc
4718 if (SCC.getOpcode() == ISD::SETCC)
4719 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(),
4720 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
4724 // If we can fold this based on the true/false value, do so.
4725 if (SimplifySelectOps(N, N2, N3))
4726 return SDValue(N, 0); // Don't revisit N.
4728 // fold select_cc into other things, such as min/max/abs
4729 return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC);
4732 SDValue DAGCombiner::visitSETCC(SDNode *N) {
4733 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
4734 cast<CondCodeSDNode>(N->getOperand(2))->get(),
4738 // tryToFoldExtendOfConstant - Try to fold a sext/zext/aext
4739 // dag node into a ConstantSDNode or a build_vector of constants.
4740 // This function is called by the DAGCombiner when visiting sext/zext/aext
4741 // dag nodes (see for example method DAGCombiner::visitSIGN_EXTEND).
4742 // Vector extends are not folded if operations are legal; this is to
4743 // avoid introducing illegal build_vector dag nodes.
4744 static SDNode *tryToFoldExtendOfConstant(SDNode *N, const TargetLowering &TLI,
4745 SelectionDAG &DAG, bool LegalTypes,
4746 bool LegalOperations) {
4747 unsigned Opcode = N->getOpcode();
4748 SDValue N0 = N->getOperand(0);
4749 EVT VT = N->getValueType(0);
4751 assert((Opcode == ISD::SIGN_EXTEND || Opcode == ISD::ZERO_EXTEND ||
4752 Opcode == ISD::ANY_EXTEND) && "Expected EXTEND dag node in input!");
4754 // fold (sext c1) -> c1
4755 // fold (zext c1) -> c1
4756 // fold (aext c1) -> c1
4757 if (isa<ConstantSDNode>(N0))
4758 return DAG.getNode(Opcode, SDLoc(N), VT, N0).getNode();
4760 // fold (sext (build_vector AllConstants) -> (build_vector AllConstants)
4761 // fold (zext (build_vector AllConstants) -> (build_vector AllConstants)
4762 // fold (aext (build_vector AllConstants) -> (build_vector AllConstants)
4763 EVT SVT = VT.getScalarType();
4764 if (!(VT.isVector() &&
4765 (!LegalTypes || (!LegalOperations && TLI.isTypeLegal(SVT))) &&
4766 ISD::isBuildVectorOfConstantSDNodes(N0.getNode())))
4769 // We can fold this node into a build_vector.
4770 unsigned VTBits = SVT.getSizeInBits();
4771 unsigned EVTBits = N0->getValueType(0).getScalarType().getSizeInBits();
4772 unsigned ShAmt = VTBits - EVTBits;
4773 SmallVector<SDValue, 8> Elts;
4774 unsigned NumElts = N0->getNumOperands();
4777 for (unsigned i=0; i != NumElts; ++i) {
4778 SDValue Op = N0->getOperand(i);
4779 if (Op->getOpcode() == ISD::UNDEF) {
4780 Elts.push_back(DAG.getUNDEF(SVT));
4784 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
4785 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
4786 if (Opcode == ISD::SIGN_EXTEND)
4787 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
4790 Elts.push_back(DAG.getConstant(C.shl(ShAmt).lshr(ShAmt).getZExtValue(),
4794 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, &Elts[0], NumElts).getNode();
4797 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
4798 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
4799 // transformation. Returns true if extension are possible and the above
4800 // mentioned transformation is profitable.
4801 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
4803 SmallVectorImpl<SDNode *> &ExtendNodes,
4804 const TargetLowering &TLI) {
4805 bool HasCopyToRegUses = false;
4806 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
4807 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
4808 UE = N0.getNode()->use_end();
4813 if (UI.getUse().getResNo() != N0.getResNo())
4815 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
4816 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
4817 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
4818 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
4819 // Sign bits will be lost after a zext.
4822 for (unsigned i = 0; i != 2; ++i) {
4823 SDValue UseOp = User->getOperand(i);
4826 if (!isa<ConstantSDNode>(UseOp))
4831 ExtendNodes.push_back(User);
4834 // If truncates aren't free and there are users we can't
4835 // extend, it isn't worthwhile.
4838 // Remember if this value is live-out.
4839 if (User->getOpcode() == ISD::CopyToReg)
4840 HasCopyToRegUses = true;
4843 if (HasCopyToRegUses) {
4844 bool BothLiveOut = false;
4845 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
4847 SDUse &Use = UI.getUse();
4848 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
4854 // Both unextended and extended values are live out. There had better be
4855 // a good reason for the transformation.
4856 return ExtendNodes.size();
4861 void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
4862 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
4863 ISD::NodeType ExtType) {
4864 // Extend SetCC uses if necessary.
4865 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
4866 SDNode *SetCC = SetCCs[i];
4867 SmallVector<SDValue, 4> Ops;
4869 for (unsigned j = 0; j != 2; ++j) {
4870 SDValue SOp = SetCC->getOperand(j);
4872 Ops.push_back(ExtLoad);
4874 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
4877 Ops.push_back(SetCC->getOperand(2));
4878 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0),
4879 &Ops[0], Ops.size()));
4883 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
4884 SDValue N0 = N->getOperand(0);
4885 EVT VT = N->getValueType(0);
4887 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
4889 return SDValue(Res, 0);
4891 // fold (sext (sext x)) -> (sext x)
4892 // fold (sext (aext x)) -> (sext x)
4893 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4894 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT,
4897 if (N0.getOpcode() == ISD::TRUNCATE) {
4898 // fold (sext (truncate (load x))) -> (sext (smaller load x))
4899 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
4900 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4901 if (NarrowLoad.getNode()) {
4902 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4903 if (NarrowLoad.getNode() != N0.getNode()) {
4904 CombineTo(N0.getNode(), NarrowLoad);
4905 // CombineTo deleted the truncate, if needed, but not what's under it.
4908 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4911 // See if the value being truncated is already sign extended. If so, just
4912 // eliminate the trunc/sext pair.
4913 SDValue Op = N0.getOperand(0);
4914 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
4915 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
4916 unsigned DestBits = VT.getScalarType().getSizeInBits();
4917 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
4919 if (OpBits == DestBits) {
4920 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
4921 // bits, it is already ready.
4922 if (NumSignBits > DestBits-MidBits)
4924 } else if (OpBits < DestBits) {
4925 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
4926 // bits, just sext from i32.
4927 if (NumSignBits > OpBits-MidBits)
4928 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, Op);
4930 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
4931 // bits, just truncate to i32.
4932 if (NumSignBits > OpBits-MidBits)
4933 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
4936 // fold (sext (truncate x)) -> (sextinreg x).
4937 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
4938 N0.getValueType())) {
4939 if (OpBits < DestBits)
4940 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op);
4941 else if (OpBits > DestBits)
4942 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op);
4943 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, Op,
4944 DAG.getValueType(N0.getValueType()));
4948 // fold (sext (load x)) -> (sext (truncate (sextload x)))
4949 // None of the supported targets knows how to perform load and sign extend
4950 // on vectors in one instruction. We only perform this transformation on
4952 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4953 ISD::isUNINDEXEDLoad(N0.getNode()) &&
4954 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4955 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) {
4956 bool DoXform = true;
4957 SmallVector<SDNode*, 4> SetCCs;
4958 if (!N0.hasOneUse())
4959 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
4961 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4962 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
4964 LN0->getBasePtr(), N0.getValueType(),
4965 LN0->getMemOperand());
4966 CombineTo(N, ExtLoad);
4967 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
4968 N0.getValueType(), ExtLoad);
4969 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4970 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
4972 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4976 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
4977 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
4978 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
4979 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
4980 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4981 EVT MemVT = LN0->getMemoryVT();
4982 if ((!LegalOperations && !LN0->isVolatile()) ||
4983 TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) {
4984 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
4986 LN0->getBasePtr(), MemVT,
4987 LN0->getMemOperand());
4988 CombineTo(N, ExtLoad);
4989 CombineTo(N0.getNode(),
4990 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
4991 N0.getValueType(), ExtLoad),
4992 ExtLoad.getValue(1));
4993 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4997 // fold (sext (and/or/xor (load x), cst)) ->
4998 // (and/or/xor (sextload x), (sext cst))
4999 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5000 N0.getOpcode() == ISD::XOR) &&
5001 isa<LoadSDNode>(N0.getOperand(0)) &&
5002 N0.getOperand(1).getOpcode() == ISD::Constant &&
5003 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()) &&
5004 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5005 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5006 if (LN0->getExtensionType() != ISD::ZEXTLOAD && LN0->isUnindexed()) {
5007 bool DoXform = true;
5008 SmallVector<SDNode*, 4> SetCCs;
5009 if (!N0.hasOneUse())
5010 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
5013 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN0), VT,
5014 LN0->getChain(), LN0->getBasePtr(),
5016 LN0->getMemOperand());
5017 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5018 Mask = Mask.sext(VT.getSizeInBits());
5019 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5020 ExtLoad, DAG.getConstant(Mask, VT));
5021 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5022 SDLoc(N0.getOperand(0)),
5023 N0.getOperand(0).getValueType(), ExtLoad);
5025 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5026 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5028 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5033 if (N0.getOpcode() == ISD::SETCC) {
5034 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
5035 // Only do this before legalize for now.
5036 if (VT.isVector() && !LegalOperations &&
5037 TLI.getBooleanContents(true) ==
5038 TargetLowering::ZeroOrNegativeOneBooleanContent) {
5039 EVT N0VT = N0.getOperand(0).getValueType();
5040 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
5041 // of the same size as the compared operands. Only optimize sext(setcc())
5042 // if this is the case.
5043 EVT SVT = getSetCCResultType(N0VT);
5045 // We know that the # elements of the results is the same as the
5046 // # elements of the compare (and the # elements of the compare result
5047 // for that matter). Check to see that they are the same size. If so,
5048 // we know that the element size of the sext'd result matches the
5049 // element size of the compare operands.
5050 if (VT.getSizeInBits() == SVT.getSizeInBits())
5051 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5053 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5055 // If the desired elements are smaller or larger than the source
5056 // elements we can use a matching integer vector type and then
5057 // truncate/sign extend
5058 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
5059 if (SVT == MatchingVectorType) {
5060 SDValue VsetCC = DAG.getSetCC(SDLoc(N), MatchingVectorType,
5061 N0.getOperand(0), N0.getOperand(1),
5062 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5063 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
5067 // sext(setcc x, y, cc) -> (select (setcc x, y, cc), -1, 0)
5068 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
5070 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
5072 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5073 NegOne, DAG.getConstant(0, VT),
5074 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5075 if (SCC.getNode()) return SCC;
5077 if (!VT.isVector()) {
5078 EVT SetCCVT = getSetCCResultType(N0.getOperand(0).getValueType());
5079 if (!LegalOperations || TLI.isOperationLegal(ISD::SETCC, SetCCVT)) {
5081 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
5082 SDValue SetCC = DAG.getSetCC(DL,
5084 N0.getOperand(0), N0.getOperand(1), CC);
5085 EVT SelectVT = getSetCCResultType(VT);
5086 return DAG.getSelect(DL, VT,
5087 DAG.getSExtOrTrunc(SetCC, DL, SelectVT),
5088 NegOne, DAG.getConstant(0, VT));
5094 // fold (sext x) -> (zext x) if the sign bit is known zero.
5095 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
5096 DAG.SignBitIsZero(N0))
5097 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0);
5102 // isTruncateOf - If N is a truncate of some other value, return true, record
5103 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
5104 // This function computes KnownZero to avoid a duplicated call to
5105 // ComputeMaskedBits in the caller.
5106 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
5109 if (N->getOpcode() == ISD::TRUNCATE) {
5110 Op = N->getOperand(0);
5111 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
5115 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
5116 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
5119 SDValue Op0 = N->getOperand(0);
5120 SDValue Op1 = N->getOperand(1);
5121 assert(Op0.getValueType() == Op1.getValueType());
5123 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
5124 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
5125 if (COp0 && COp0->isNullValue())
5127 else if (COp1 && COp1->isNullValue())
5132 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
5134 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
5140 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
5141 SDValue N0 = N->getOperand(0);
5142 EVT VT = N->getValueType(0);
5144 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5146 return SDValue(Res, 0);
5148 // fold (zext (zext x)) -> (zext x)
5149 // fold (zext (aext x)) -> (zext x)
5150 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
5151 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT,
5154 // fold (zext (truncate x)) -> (zext x) or
5155 // (zext (truncate x)) -> (truncate x)
5156 // This is valid when the truncated bits of x are already zero.
5157 // FIXME: We should extend this to work for vectors too.
5160 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
5161 APInt TruncatedBits =
5162 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
5163 APInt(Op.getValueSizeInBits(), 0) :
5164 APInt::getBitsSet(Op.getValueSizeInBits(),
5165 N0.getValueSizeInBits(),
5166 std::min(Op.getValueSizeInBits(),
5167 VT.getSizeInBits()));
5168 if (TruncatedBits == (KnownZero & TruncatedBits)) {
5169 if (VT.bitsGT(Op.getValueType()))
5170 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, Op);
5171 if (VT.bitsLT(Op.getValueType()))
5172 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5178 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5179 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
5180 if (N0.getOpcode() == ISD::TRUNCATE) {
5181 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5182 if (NarrowLoad.getNode()) {
5183 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5184 if (NarrowLoad.getNode() != N0.getNode()) {
5185 CombineTo(N0.getNode(), NarrowLoad);
5186 // CombineTo deleted the truncate, if needed, but not what's under it.
5189 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5193 // fold (zext (truncate x)) -> (and x, mask)
5194 if (N0.getOpcode() == ISD::TRUNCATE &&
5195 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
5197 // fold (zext (truncate (load x))) -> (zext (smaller load x))
5198 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
5199 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5200 if (NarrowLoad.getNode()) {
5201 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5202 if (NarrowLoad.getNode() != N0.getNode()) {
5203 CombineTo(N0.getNode(), NarrowLoad);
5204 // CombineTo deleted the truncate, if needed, but not what's under it.
5207 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5210 SDValue Op = N0.getOperand(0);
5211 if (Op.getValueType().bitsLT(VT)) {
5212 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, Op);
5213 AddToWorkList(Op.getNode());
5214 } else if (Op.getValueType().bitsGT(VT)) {
5215 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
5216 AddToWorkList(Op.getNode());
5218 return DAG.getZeroExtendInReg(Op, SDLoc(N),
5219 N0.getValueType().getScalarType());
5222 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
5223 // if either of the casts is not free.
5224 if (N0.getOpcode() == ISD::AND &&
5225 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5226 N0.getOperand(1).getOpcode() == ISD::Constant &&
5227 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5228 N0.getValueType()) ||
5229 !TLI.isZExtFree(N0.getValueType(), VT))) {
5230 SDValue X = N0.getOperand(0).getOperand(0);
5231 if (X.getValueType().bitsLT(VT)) {
5232 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(X), VT, X);
5233 } else if (X.getValueType().bitsGT(VT)) {
5234 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
5236 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5237 Mask = Mask.zext(VT.getSizeInBits());
5238 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5239 X, DAG.getConstant(Mask, VT));
5242 // fold (zext (load x)) -> (zext (truncate (zextload x)))
5243 // None of the supported targets knows how to perform load and vector_zext
5244 // on vectors in one instruction. We only perform this transformation on
5246 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5247 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5248 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5249 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
5250 bool DoXform = true;
5251 SmallVector<SDNode*, 4> SetCCs;
5252 if (!N0.hasOneUse())
5253 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
5255 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5256 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5258 LN0->getBasePtr(), N0.getValueType(),
5259 LN0->getMemOperand());
5260 CombineTo(N, ExtLoad);
5261 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5262 N0.getValueType(), ExtLoad);
5263 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5265 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5267 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5271 // fold (zext (and/or/xor (load x), cst)) ->
5272 // (and/or/xor (zextload x), (zext cst))
5273 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5274 N0.getOpcode() == ISD::XOR) &&
5275 isa<LoadSDNode>(N0.getOperand(0)) &&
5276 N0.getOperand(1).getOpcode() == ISD::Constant &&
5277 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()) &&
5278 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5279 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5280 if (LN0->getExtensionType() != ISD::SEXTLOAD && LN0->isUnindexed()) {
5281 bool DoXform = true;
5282 SmallVector<SDNode*, 4> SetCCs;
5283 if (!N0.hasOneUse())
5284 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
5287 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), VT,
5288 LN0->getChain(), LN0->getBasePtr(),
5290 LN0->getMemOperand());
5291 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5292 Mask = Mask.zext(VT.getSizeInBits());
5293 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5294 ExtLoad, DAG.getConstant(Mask, VT));
5295 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5296 SDLoc(N0.getOperand(0)),
5297 N0.getOperand(0).getValueType(), ExtLoad);
5299 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5300 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5302 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5307 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
5308 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
5309 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5310 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5311 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5312 EVT MemVT = LN0->getMemoryVT();
5313 if ((!LegalOperations && !LN0->isVolatile()) ||
5314 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) {
5315 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5317 LN0->getBasePtr(), MemVT,
5318 LN0->getMemOperand());
5319 CombineTo(N, ExtLoad);
5320 CombineTo(N0.getNode(),
5321 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(),
5323 ExtLoad.getValue(1));
5324 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5328 if (N0.getOpcode() == ISD::SETCC) {
5329 if (!LegalOperations && VT.isVector() &&
5330 N0.getValueType().getVectorElementType() == MVT::i1) {
5331 EVT N0VT = N0.getOperand(0).getValueType();
5332 if (getSetCCResultType(N0VT) == N0.getValueType())
5335 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
5336 // Only do this before legalize for now.
5337 EVT EltVT = VT.getVectorElementType();
5338 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
5339 DAG.getConstant(1, EltVT));
5340 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5341 // We know that the # elements of the results is the same as the
5342 // # elements of the compare (and the # elements of the compare result
5343 // for that matter). Check to see that they are the same size. If so,
5344 // we know that the element size of the sext'd result matches the
5345 // element size of the compare operands.
5346 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5347 DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5349 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
5350 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT,
5351 &OneOps[0], OneOps.size()));
5353 // If the desired elements are smaller or larger than the source
5354 // elements we can use a matching integer vector type and then
5355 // truncate/sign extend
5356 EVT MatchingElementType =
5357 EVT::getIntegerVT(*DAG.getContext(),
5358 N0VT.getScalarType().getSizeInBits());
5359 EVT MatchingVectorType =
5360 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
5361 N0VT.getVectorNumElements());
5363 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5365 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5366 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5367 DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT),
5368 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT,
5369 &OneOps[0], OneOps.size()));
5372 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5374 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5375 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5376 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5377 if (SCC.getNode()) return SCC;
5380 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
5381 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
5382 isa<ConstantSDNode>(N0.getOperand(1)) &&
5383 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
5385 SDValue ShAmt = N0.getOperand(1);
5386 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
5387 if (N0.getOpcode() == ISD::SHL) {
5388 SDValue InnerZExt = N0.getOperand(0);
5389 // If the original shl may be shifting out bits, do not perform this
5391 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
5392 InnerZExt.getOperand(0).getValueType().getSizeInBits();
5393 if (ShAmtVal > KnownZeroBits)
5399 // Ensure that the shift amount is wide enough for the shifted value.
5400 if (VT.getSizeInBits() >= 256)
5401 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
5403 return DAG.getNode(N0.getOpcode(), DL, VT,
5404 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
5411 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
5412 SDValue N0 = N->getOperand(0);
5413 EVT VT = N->getValueType(0);
5415 if (SDNode *Res = tryToFoldExtendOfConstant(N, TLI, DAG, LegalTypes,
5417 return SDValue(Res, 0);
5419 // fold (aext (aext x)) -> (aext x)
5420 // fold (aext (zext x)) -> (zext x)
5421 // fold (aext (sext x)) -> (sext x)
5422 if (N0.getOpcode() == ISD::ANY_EXTEND ||
5423 N0.getOpcode() == ISD::ZERO_EXTEND ||
5424 N0.getOpcode() == ISD::SIGN_EXTEND)
5425 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0));
5427 // fold (aext (truncate (load x))) -> (aext (smaller load x))
5428 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
5429 if (N0.getOpcode() == ISD::TRUNCATE) {
5430 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5431 if (NarrowLoad.getNode()) {
5432 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5433 if (NarrowLoad.getNode() != N0.getNode()) {
5434 CombineTo(N0.getNode(), NarrowLoad);
5435 // CombineTo deleted the truncate, if needed, but not what's under it.
5438 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5442 // fold (aext (truncate x))
5443 if (N0.getOpcode() == ISD::TRUNCATE) {
5444 SDValue TruncOp = N0.getOperand(0);
5445 if (TruncOp.getValueType() == VT)
5446 return TruncOp; // x iff x size == zext size.
5447 if (TruncOp.getValueType().bitsGT(VT))
5448 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, TruncOp);
5449 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, TruncOp);
5452 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
5453 // if the trunc is not free.
5454 if (N0.getOpcode() == ISD::AND &&
5455 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5456 N0.getOperand(1).getOpcode() == ISD::Constant &&
5457 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5458 N0.getValueType())) {
5459 SDValue X = N0.getOperand(0).getOperand(0);
5460 if (X.getValueType().bitsLT(VT)) {
5461 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, X);
5462 } else if (X.getValueType().bitsGT(VT)) {
5463 X = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, X);
5465 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5466 Mask = Mask.zext(VT.getSizeInBits());
5467 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5468 X, DAG.getConstant(Mask, VT));
5471 // fold (aext (load x)) -> (aext (truncate (extload x)))
5472 // None of the supported targets knows how to perform load and any_ext
5473 // on vectors in one instruction. We only perform this transformation on
5475 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5476 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5477 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5478 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
5479 bool DoXform = true;
5480 SmallVector<SDNode*, 4> SetCCs;
5481 if (!N0.hasOneUse())
5482 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
5484 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5485 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
5487 LN0->getBasePtr(), N0.getValueType(),
5488 LN0->getMemOperand());
5489 CombineTo(N, ExtLoad);
5490 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5491 N0.getValueType(), ExtLoad);
5492 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5493 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5495 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5499 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
5500 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
5501 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
5502 if (N0.getOpcode() == ISD::LOAD &&
5503 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5505 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5506 ISD::LoadExtType ExtType = LN0->getExtensionType();
5507 EVT MemVT = LN0->getMemoryVT();
5508 if (!LegalOperations || TLI.isLoadExtLegal(ExtType, MemVT)) {
5509 SDValue ExtLoad = DAG.getExtLoad(ExtType, SDLoc(N),
5510 VT, LN0->getChain(), LN0->getBasePtr(),
5511 MemVT, LN0->getMemOperand());
5512 CombineTo(N, ExtLoad);
5513 CombineTo(N0.getNode(),
5514 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5515 N0.getValueType(), ExtLoad),
5516 ExtLoad.getValue(1));
5517 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5521 if (N0.getOpcode() == ISD::SETCC) {
5522 // aext(setcc) -> sext_in_reg(vsetcc) for vectors.
5523 // Only do this before legalize for now.
5524 if (VT.isVector() && !LegalOperations) {
5525 EVT N0VT = N0.getOperand(0).getValueType();
5526 // We know that the # elements of the results is the same as the
5527 // # elements of the compare (and the # elements of the compare result
5528 // for that matter). Check to see that they are the same size. If so,
5529 // we know that the element size of the sext'd result matches the
5530 // element size of the compare operands.
5531 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5532 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5534 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5535 // If the desired elements are smaller or larger than the source
5536 // elements we can use a matching integer vector type and then
5537 // truncate/sign extend
5539 EVT MatchingElementType =
5540 EVT::getIntegerVT(*DAG.getContext(),
5541 N0VT.getScalarType().getSizeInBits());
5542 EVT MatchingVectorType =
5543 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
5544 N0VT.getVectorNumElements());
5546 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5548 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5549 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
5553 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5555 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5556 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5557 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5565 /// GetDemandedBits - See if the specified operand can be simplified with the
5566 /// knowledge that only the bits specified by Mask are used. If so, return the
5567 /// simpler operand, otherwise return a null SDValue.
5568 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
5569 switch (V.getOpcode()) {
5571 case ISD::Constant: {
5572 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
5573 assert(CV != 0 && "Const value should be ConstSDNode.");
5574 const APInt &CVal = CV->getAPIntValue();
5575 APInt NewVal = CVal & Mask;
5577 return DAG.getConstant(NewVal, V.getValueType());
5582 // If the LHS or RHS don't contribute bits to the or, drop them.
5583 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
5584 return V.getOperand(1);
5585 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
5586 return V.getOperand(0);
5589 // Only look at single-use SRLs.
5590 if (!V.getNode()->hasOneUse())
5592 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
5593 // See if we can recursively simplify the LHS.
5594 unsigned Amt = RHSC->getZExtValue();
5596 // Watch out for shift count overflow though.
5597 if (Amt >= Mask.getBitWidth()) break;
5598 APInt NewMask = Mask << Amt;
5599 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
5600 if (SimplifyLHS.getNode())
5601 return DAG.getNode(ISD::SRL, SDLoc(V), V.getValueType(),
5602 SimplifyLHS, V.getOperand(1));
5608 /// ReduceLoadWidth - If the result of a wider load is shifted to right of N
5609 /// bits and then truncated to a narrower type and where N is a multiple
5610 /// of number of bits of the narrower type, transform it to a narrower load
5611 /// from address + N / num of bits of new type. If the result is to be
5612 /// extended, also fold the extension to form a extending load.
5613 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
5614 unsigned Opc = N->getOpcode();
5616 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
5617 SDValue N0 = N->getOperand(0);
5618 EVT VT = N->getValueType(0);
5621 // This transformation isn't valid for vector loads.
5625 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
5627 if (Opc == ISD::SIGN_EXTEND_INREG) {
5628 ExtType = ISD::SEXTLOAD;
5629 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
5630 } else if (Opc == ISD::SRL) {
5631 // Another special-case: SRL is basically zero-extending a narrower value.
5632 ExtType = ISD::ZEXTLOAD;
5634 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
5635 if (!N01) return SDValue();
5636 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
5637 VT.getSizeInBits() - N01->getZExtValue());
5639 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, ExtVT))
5642 unsigned EVTBits = ExtVT.getSizeInBits();
5644 // Do not generate loads of non-round integer types since these can
5645 // be expensive (and would be wrong if the type is not byte sized).
5646 if (!ExtVT.isRound())
5650 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
5651 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5652 ShAmt = N01->getZExtValue();
5653 // Is the shift amount a multiple of size of VT?
5654 if ((ShAmt & (EVTBits-1)) == 0) {
5655 N0 = N0.getOperand(0);
5656 // Is the load width a multiple of size of VT?
5657 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
5661 // At this point, we must have a load or else we can't do the transform.
5662 if (!isa<LoadSDNode>(N0)) return SDValue();
5664 // Because a SRL must be assumed to *need* to zero-extend the high bits
5665 // (as opposed to anyext the high bits), we can't combine the zextload
5666 // lowering of SRL and an sextload.
5667 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
5670 // If the shift amount is larger than the input type then we're not
5671 // accessing any of the loaded bytes. If the load was a zextload/extload
5672 // then the result of the shift+trunc is zero/undef (handled elsewhere).
5673 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
5678 // If the load is shifted left (and the result isn't shifted back right),
5679 // we can fold the truncate through the shift.
5680 unsigned ShLeftAmt = 0;
5681 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
5682 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
5683 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5684 ShLeftAmt = N01->getZExtValue();
5685 N0 = N0.getOperand(0);
5689 // If we haven't found a load, we can't narrow it. Don't transform one with
5690 // multiple uses, this would require adding a new load.
5691 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse())
5694 // Don't change the width of a volatile load.
5695 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5696 if (LN0->isVolatile())
5699 // Verify that we are actually reducing a load width here.
5700 if (LN0->getMemoryVT().getSizeInBits() < EVTBits)
5703 // For the transform to be legal, the load must produce only two values
5704 // (the value loaded and the chain). Don't transform a pre-increment
5705 // load, for example, which produces an extra value. Otherwise the
5706 // transformation is not equivalent, and the downstream logic to replace
5707 // uses gets things wrong.
5708 if (LN0->getNumValues() > 2)
5711 // If the load that we're shrinking is an extload and we're not just
5712 // discarding the extension we can't simply shrink the load. Bail.
5713 // TODO: It would be possible to merge the extensions in some cases.
5714 if (LN0->getExtensionType() != ISD::NON_EXTLOAD &&
5715 LN0->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits() + ShAmt)
5718 EVT PtrType = N0.getOperand(1).getValueType();
5720 if (PtrType == MVT::Untyped || PtrType.isExtended())
5721 // It's not possible to generate a constant of extended or untyped type.
5724 // For big endian targets, we need to adjust the offset to the pointer to
5725 // load the correct bytes.
5726 if (TLI.isBigEndian()) {
5727 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
5728 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
5729 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
5732 uint64_t PtrOff = ShAmt / 8;
5733 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
5734 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0),
5735 PtrType, LN0->getBasePtr(),
5736 DAG.getConstant(PtrOff, PtrType));
5737 AddToWorkList(NewPtr.getNode());
5740 if (ExtType == ISD::NON_EXTLOAD)
5741 Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr,
5742 LN0->getPointerInfo().getWithOffset(PtrOff),
5743 LN0->isVolatile(), LN0->isNonTemporal(),
5744 LN0->isInvariant(), NewAlign, LN0->getTBAAInfo());
5746 Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(),NewPtr,
5747 LN0->getPointerInfo().getWithOffset(PtrOff),
5748 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
5749 NewAlign, LN0->getTBAAInfo());
5751 // Replace the old load's chain with the new load's chain.
5752 WorkListRemover DeadNodes(*this);
5753 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
5755 // Shift the result left, if we've swallowed a left shift.
5756 SDValue Result = Load;
5757 if (ShLeftAmt != 0) {
5758 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
5759 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
5761 // If the shift amount is as large as the result size (but, presumably,
5762 // no larger than the source) then the useful bits of the result are
5763 // zero; we can't simply return the shortened shift, because the result
5764 // of that operation is undefined.
5765 if (ShLeftAmt >= VT.getSizeInBits())
5766 Result = DAG.getConstant(0, VT);
5768 Result = DAG.getNode(ISD::SHL, SDLoc(N0), VT,
5769 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
5772 // Return the new loaded value.
5776 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
5777 SDValue N0 = N->getOperand(0);
5778 SDValue N1 = N->getOperand(1);
5779 EVT VT = N->getValueType(0);
5780 EVT EVT = cast<VTSDNode>(N1)->getVT();
5781 unsigned VTBits = VT.getScalarType().getSizeInBits();
5782 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
5784 // fold (sext_in_reg c1) -> c1
5785 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
5786 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1);
5788 // If the input is already sign extended, just drop the extension.
5789 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
5792 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
5793 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
5794 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT()))
5795 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
5796 N0.getOperand(0), N1);
5798 // fold (sext_in_reg (sext x)) -> (sext x)
5799 // fold (sext_in_reg (aext x)) -> (sext x)
5800 // if x is small enough.
5801 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
5802 SDValue N00 = N0.getOperand(0);
5803 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
5804 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
5805 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
5808 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
5809 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
5810 return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT);
5812 // fold operands of sext_in_reg based on knowledge that the top bits are not
5814 if (SimplifyDemandedBits(SDValue(N, 0)))
5815 return SDValue(N, 0);
5817 // fold (sext_in_reg (load x)) -> (smaller sextload x)
5818 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
5819 SDValue NarrowLoad = ReduceLoadWidth(N);
5820 if (NarrowLoad.getNode())
5823 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
5824 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
5825 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
5826 if (N0.getOpcode() == ISD::SRL) {
5827 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
5828 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
5829 // We can turn this into an SRA iff the input to the SRL is already sign
5831 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
5832 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
5833 return DAG.getNode(ISD::SRA, SDLoc(N), VT,
5834 N0.getOperand(0), N0.getOperand(1));
5838 // fold (sext_inreg (extload x)) -> (sextload x)
5839 if (ISD::isEXTLoad(N0.getNode()) &&
5840 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5841 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5842 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5843 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5844 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5845 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5847 LN0->getBasePtr(), EVT,
5848 LN0->getMemOperand());
5849 CombineTo(N, ExtLoad);
5850 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5851 AddToWorkList(ExtLoad.getNode());
5852 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5854 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
5855 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5857 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5858 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5859 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5860 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5861 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5863 LN0->getBasePtr(), EVT,
5864 LN0->getMemOperand());
5865 CombineTo(N, ExtLoad);
5866 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5867 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5870 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
5871 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
5872 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
5873 N0.getOperand(1), false);
5874 if (BSwap.getNode() != 0)
5875 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
5879 // Fold a sext_inreg of a build_vector of ConstantSDNodes or undefs
5880 // into a build_vector.
5881 if (ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
5882 SmallVector<SDValue, 8> Elts;
5883 unsigned NumElts = N0->getNumOperands();
5884 unsigned ShAmt = VTBits - EVTBits;
5886 for (unsigned i = 0; i != NumElts; ++i) {
5887 SDValue Op = N0->getOperand(i);
5888 if (Op->getOpcode() == ISD::UNDEF) {
5893 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
5894 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
5895 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
5896 Op.getValueType()));
5899 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, &Elts[0], NumElts);
5905 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
5906 SDValue N0 = N->getOperand(0);
5907 EVT VT = N->getValueType(0);
5908 bool isLE = TLI.isLittleEndian();
5911 if (N0.getValueType() == N->getValueType(0))
5913 // fold (truncate c1) -> c1
5914 if (isa<ConstantSDNode>(N0))
5915 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0);
5916 // fold (truncate (truncate x)) -> (truncate x)
5917 if (N0.getOpcode() == ISD::TRUNCATE)
5918 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
5919 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
5920 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
5921 N0.getOpcode() == ISD::SIGN_EXTEND ||
5922 N0.getOpcode() == ISD::ANY_EXTEND) {
5923 if (N0.getOperand(0).getValueType().bitsLT(VT))
5924 // if the source is smaller than the dest, we still need an extend
5925 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5927 if (N0.getOperand(0).getValueType().bitsGT(VT))
5928 // if the source is larger than the dest, than we just need the truncate
5929 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
5930 // if the source and dest are the same type, we can drop both the extend
5931 // and the truncate.
5932 return N0.getOperand(0);
5935 // Fold extract-and-trunc into a narrow extract. For example:
5936 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
5937 // i32 y = TRUNCATE(i64 x)
5939 // v16i8 b = BITCAST (v2i64 val)
5940 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
5942 // Note: We only run this optimization after type legalization (which often
5943 // creates this pattern) and before operation legalization after which
5944 // we need to be more careful about the vector instructions that we generate.
5945 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5946 LegalTypes && !LegalOperations && N0->hasOneUse() && VT != MVT::i1) {
5948 EVT VecTy = N0.getOperand(0).getValueType();
5949 EVT ExTy = N0.getValueType();
5950 EVT TrTy = N->getValueType(0);
5952 unsigned NumElem = VecTy.getVectorNumElements();
5953 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
5955 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
5956 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
5958 SDValue EltNo = N0->getOperand(1);
5959 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
5960 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5961 EVT IndexTy = TLI.getVectorIdxTy();
5962 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
5964 SDValue V = DAG.getNode(ISD::BITCAST, SDLoc(N),
5965 NVT, N0.getOperand(0));
5967 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
5969 DAG.getConstant(Index, IndexTy));
5973 // Fold a series of buildvector, bitcast, and truncate if possible.
5975 // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to
5976 // (2xi32 (buildvector x, y)).
5977 if (Level == AfterLegalizeVectorOps && VT.isVector() &&
5978 N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
5979 N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
5980 N0.getOperand(0).hasOneUse()) {
5982 SDValue BuildVect = N0.getOperand(0);
5983 EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType();
5984 EVT TruncVecEltTy = VT.getVectorElementType();
5986 // Check that the element types match.
5987 if (BuildVectEltTy == TruncVecEltTy) {
5988 // Now we only need to compute the offset of the truncated elements.
5989 unsigned BuildVecNumElts = BuildVect.getNumOperands();
5990 unsigned TruncVecNumElts = VT.getVectorNumElements();
5991 unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts;
5993 assert((BuildVecNumElts % TruncVecNumElts) == 0 &&
5994 "Invalid number of elements");
5996 SmallVector<SDValue, 8> Opnds;
5997 for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset)
5998 Opnds.push_back(BuildVect.getOperand(i));
6000 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, &Opnds[0],
6005 // See if we can simplify the input to this truncate through knowledge that
6006 // only the low bits are being used.
6007 // For example "trunc (or (shl x, 8), y)" // -> trunc y
6008 // Currently we only perform this optimization on scalars because vectors
6009 // may have different active low bits.
6010 if (!VT.isVector()) {
6012 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
6013 VT.getSizeInBits()));
6014 if (Shorter.getNode())
6015 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter);
6017 // fold (truncate (load x)) -> (smaller load x)
6018 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
6019 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
6020 SDValue Reduced = ReduceLoadWidth(N);
6021 if (Reduced.getNode())
6023 // Handle the case where the load remains an extending load even
6024 // after truncation.
6025 if (N0.hasOneUse() && ISD::isUNINDEXEDLoad(N0.getNode())) {
6026 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6027 if (!LN0->isVolatile() &&
6028 LN0->getMemoryVT().getStoreSizeInBits() < VT.getSizeInBits()) {
6029 SDValue NewLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(LN0),
6030 VT, LN0->getChain(), LN0->getBasePtr(),
6032 LN0->getMemOperand());
6033 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLoad.getValue(1));
6038 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
6039 // where ... are all 'undef'.
6040 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
6041 SmallVector<EVT, 8> VTs;
6044 unsigned NumDefs = 0;
6046 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
6047 SDValue X = N0.getOperand(i);
6048 if (X.getOpcode() != ISD::UNDEF) {
6053 // Stop if more than one members are non-undef.
6056 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
6057 VT.getVectorElementType(),
6058 X.getValueType().getVectorNumElements()));
6062 return DAG.getUNDEF(VT);
6065 assert(V.getNode() && "The single defined operand is empty!");
6066 SmallVector<SDValue, 8> Opnds;
6067 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
6069 Opnds.push_back(DAG.getUNDEF(VTs[i]));
6072 SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V);
6073 AddToWorkList(NV.getNode());
6074 Opnds.push_back(NV);
6076 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
6077 &Opnds[0], Opnds.size());
6081 // Simplify the operands using demanded-bits information.
6082 if (!VT.isVector() &&
6083 SimplifyDemandedBits(SDValue(N, 0)))
6084 return SDValue(N, 0);
6089 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
6090 SDValue Elt = N->getOperand(i);
6091 if (Elt.getOpcode() != ISD::MERGE_VALUES)
6092 return Elt.getNode();
6093 return Elt.getOperand(Elt.getResNo()).getNode();
6096 /// CombineConsecutiveLoads - build_pair (load, load) -> load
6097 /// if load locations are consecutive.
6098 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
6099 assert(N->getOpcode() == ISD::BUILD_PAIR);
6101 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
6102 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
6103 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
6104 LD1->getAddressSpace() != LD2->getAddressSpace())
6106 EVT LD1VT = LD1->getValueType(0);
6108 if (ISD::isNON_EXTLoad(LD2) &&
6110 // If both are volatile this would reduce the number of volatile loads.
6111 // If one is volatile it might be ok, but play conservative and bail out.
6112 !LD1->isVolatile() &&
6113 !LD2->isVolatile() &&
6114 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
6115 unsigned Align = LD1->getAlignment();
6116 unsigned NewAlign = TLI.getDataLayout()->
6117 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6119 if (NewAlign <= Align &&
6120 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
6121 return DAG.getLoad(VT, SDLoc(N), LD1->getChain(),
6122 LD1->getBasePtr(), LD1->getPointerInfo(),
6123 false, false, false, Align);
6129 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
6130 SDValue N0 = N->getOperand(0);
6131 EVT VT = N->getValueType(0);
6133 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
6134 // Only do this before legalize, since afterward the target may be depending
6135 // on the bitconvert.
6136 // First check to see if this is all constant.
6138 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
6140 bool isSimple = cast<BuildVectorSDNode>(N0)->isConstant();
6142 EVT DestEltVT = N->getValueType(0).getVectorElementType();
6143 assert(!DestEltVT.isVector() &&
6144 "Element type of vector ValueType must not be vector!");
6146 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
6149 // If the input is a constant, let getNode fold it.
6150 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
6151 SDValue Res = DAG.getNode(ISD::BITCAST, SDLoc(N), VT, N0);
6152 if (Res.getNode() != N) {
6153 if (!LegalOperations ||
6154 TLI.isOperationLegal(Res.getNode()->getOpcode(), VT))
6157 // Folding it resulted in an illegal node, and it's too late to
6158 // do that. Clean up the old node and forego the transformation.
6159 // Ideally this won't happen very often, because instcombine
6160 // and the earlier dagcombine runs (where illegal nodes are
6161 // permitted) should have folded most of them already.
6162 DAG.DeleteNode(Res.getNode());
6166 // (conv (conv x, t1), t2) -> (conv x, t2)
6167 if (N0.getOpcode() == ISD::BITCAST)
6168 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT,
6171 // fold (conv (load x)) -> (load (conv*)x)
6172 // If the resultant load doesn't need a higher alignment than the original!
6173 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6174 // Do not change the width of a volatile load.
6175 !cast<LoadSDNode>(N0)->isVolatile() &&
6176 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)) &&
6177 TLI.isLoadBitCastBeneficial(N0.getValueType(), VT)) {
6178 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6179 unsigned Align = TLI.getDataLayout()->
6180 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
6181 unsigned OrigAlign = LN0->getAlignment();
6183 if (Align <= OrigAlign) {
6184 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(),
6185 LN0->getBasePtr(), LN0->getPointerInfo(),
6186 LN0->isVolatile(), LN0->isNonTemporal(),
6187 LN0->isInvariant(), OrigAlign,
6188 LN0->getTBAAInfo());
6190 CombineTo(N0.getNode(),
6191 DAG.getNode(ISD::BITCAST, SDLoc(N0),
6192 N0.getValueType(), Load),
6198 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
6199 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
6200 // This often reduces constant pool loads.
6201 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) ||
6202 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) &&
6203 N0.getNode()->hasOneUse() && VT.isInteger() &&
6204 !VT.isVector() && !N0.getValueType().isVector()) {
6205 SDValue NewConv = DAG.getNode(ISD::BITCAST, SDLoc(N0), VT,
6207 AddToWorkList(NewConv.getNode());
6209 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6210 if (N0.getOpcode() == ISD::FNEG)
6211 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
6212 NewConv, DAG.getConstant(SignBit, VT));
6213 assert(N0.getOpcode() == ISD::FABS);
6214 return DAG.getNode(ISD::AND, SDLoc(N), VT,
6215 NewConv, DAG.getConstant(~SignBit, VT));
6218 // fold (bitconvert (fcopysign cst, x)) ->
6219 // (or (and (bitconvert x), sign), (and cst, (not sign)))
6220 // Note that we don't handle (copysign x, cst) because this can always be
6221 // folded to an fneg or fabs.
6222 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
6223 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
6224 VT.isInteger() && !VT.isVector()) {
6225 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
6226 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
6227 if (isTypeLegal(IntXVT)) {
6228 SDValue X = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6229 IntXVT, N0.getOperand(1));
6230 AddToWorkList(X.getNode());
6232 // If X has a different width than the result/lhs, sext it or truncate it.
6233 unsigned VTWidth = VT.getSizeInBits();
6234 if (OrigXWidth < VTWidth) {
6235 X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X);
6236 AddToWorkList(X.getNode());
6237 } else if (OrigXWidth > VTWidth) {
6238 // To get the sign bit in the right place, we have to shift it right
6239 // before truncating.
6240 X = DAG.getNode(ISD::SRL, SDLoc(X),
6241 X.getValueType(), X,
6242 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
6243 AddToWorkList(X.getNode());
6244 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
6245 AddToWorkList(X.getNode());
6248 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
6249 X = DAG.getNode(ISD::AND, SDLoc(X), VT,
6250 X, DAG.getConstant(SignBit, VT));
6251 AddToWorkList(X.getNode());
6253 SDValue Cst = DAG.getNode(ISD::BITCAST, SDLoc(N0),
6254 VT, N0.getOperand(0));
6255 Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT,
6256 Cst, DAG.getConstant(~SignBit, VT));
6257 AddToWorkList(Cst.getNode());
6259 return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst);
6263 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
6264 if (N0.getOpcode() == ISD::BUILD_PAIR) {
6265 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
6266 if (CombineLD.getNode())
6273 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
6274 EVT VT = N->getValueType(0);
6275 return CombineConsecutiveLoads(N, VT);
6278 /// ConstantFoldBITCASTofBUILD_VECTOR - We know that BV is a build_vector
6279 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
6280 /// destination element value type.
6281 SDValue DAGCombiner::
6282 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
6283 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
6285 // If this is already the right type, we're done.
6286 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
6288 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
6289 unsigned DstBitSize = DstEltVT.getSizeInBits();
6291 // If this is a conversion of N elements of one type to N elements of another
6292 // type, convert each element. This handles FP<->INT cases.
6293 if (SrcBitSize == DstBitSize) {
6294 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6295 BV->getValueType(0).getVectorNumElements());
6297 // Due to the FP element handling below calling this routine recursively,
6298 // we can end up with a scalar-to-vector node here.
6299 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
6300 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6301 DAG.getNode(ISD::BITCAST, SDLoc(BV),
6302 DstEltVT, BV->getOperand(0)));
6304 SmallVector<SDValue, 8> Ops;
6305 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6306 SDValue Op = BV->getOperand(i);
6307 // If the vector element type is not legal, the BUILD_VECTOR operands
6308 // are promoted and implicitly truncated. Make that explicit here.
6309 if (Op.getValueType() != SrcEltVT)
6310 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op);
6311 Ops.push_back(DAG.getNode(ISD::BITCAST, SDLoc(BV),
6313 AddToWorkList(Ops.back().getNode());
6315 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT,
6316 &Ops[0], Ops.size());
6319 // Otherwise, we're growing or shrinking the elements. To avoid having to
6320 // handle annoying details of growing/shrinking FP values, we convert them to
6322 if (SrcEltVT.isFloatingPoint()) {
6323 // Convert the input float vector to a int vector where the elements are the
6325 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
6326 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
6327 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
6331 // Now we know the input is an integer vector. If the output is a FP type,
6332 // convert to integer first, then to FP of the right size.
6333 if (DstEltVT.isFloatingPoint()) {
6334 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
6335 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
6336 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
6338 // Next, convert to FP elements of the same size.
6339 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
6342 // Okay, we know the src/dst types are both integers of differing types.
6343 // Handling growing first.
6344 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
6345 if (SrcBitSize < DstBitSize) {
6346 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
6348 SmallVector<SDValue, 8> Ops;
6349 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
6350 i += NumInputsPerOutput) {
6351 bool isLE = TLI.isLittleEndian();
6352 APInt NewBits = APInt(DstBitSize, 0);
6353 bool EltIsUndef = true;
6354 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
6355 // Shift the previously computed bits over.
6356 NewBits <<= SrcBitSize;
6357 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
6358 if (Op.getOpcode() == ISD::UNDEF) continue;
6361 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
6362 zextOrTrunc(SrcBitSize).zext(DstBitSize);
6366 Ops.push_back(DAG.getUNDEF(DstEltVT));
6368 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
6371 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
6372 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT,
6373 &Ops[0], Ops.size());
6376 // Finally, this must be the case where we are shrinking elements: each input
6377 // turns into multiple outputs.
6378 bool isS2V = ISD::isScalarToVector(BV);
6379 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
6380 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6381 NumOutputsPerInput*BV->getNumOperands());
6382 SmallVector<SDValue, 8> Ops;
6384 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6385 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
6386 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
6387 Ops.push_back(DAG.getUNDEF(DstEltVT));
6391 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
6392 getAPIntValue().zextOrTrunc(SrcBitSize);
6394 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
6395 APInt ThisVal = OpVal.trunc(DstBitSize);
6396 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
6397 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
6398 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
6399 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6401 OpVal = OpVal.lshr(DstBitSize);
6404 // For big endian targets, swap the order of the pieces of each element.
6405 if (TLI.isBigEndian())
6406 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
6409 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT,
6410 &Ops[0], Ops.size());
6413 SDValue DAGCombiner::visitFADD(SDNode *N) {
6414 SDValue N0 = N->getOperand(0);
6415 SDValue N1 = N->getOperand(1);
6416 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6417 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6418 EVT VT = N->getValueType(0);
6421 if (VT.isVector()) {
6422 SDValue FoldedVOp = SimplifyVBinOp(N);
6423 if (FoldedVOp.getNode()) return FoldedVOp;
6426 // fold (fadd c1, c2) -> c1 + c2
6428 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N1);
6429 // canonicalize constant to RHS
6430 if (N0CFP && !N1CFP)
6431 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N0);
6432 // fold (fadd A, 0) -> A
6433 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6434 N1CFP->getValueAPF().isZero())
6436 // fold (fadd A, (fneg B)) -> (fsub A, B)
6437 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6438 isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
6439 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0,
6440 GetNegatedExpression(N1, DAG, LegalOperations));
6441 // fold (fadd (fneg A), B) -> (fsub B, A)
6442 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6443 isNegatibleForFree(N0, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
6444 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N1,
6445 GetNegatedExpression(N0, DAG, LegalOperations));
6447 // If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
6448 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6449 N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
6450 isa<ConstantFPSDNode>(N0.getOperand(1)))
6451 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0.getOperand(0),
6452 DAG.getNode(ISD::FADD, SDLoc(N), VT,
6453 N0.getOperand(1), N1));
6455 // No FP constant should be created after legalization as Instruction
6456 // Selection pass has hard time in dealing with FP constant.
6458 // We don't need test this condition for transformation like following, as
6459 // the DAG being transformed implies it is legal to take FP constant as
6462 // (fadd (fmul c, x), x) -> (fmul c+1, x)
6464 bool AllowNewFpConst = (Level < AfterLegalizeDAG);
6466 // If allow, fold (fadd (fneg x), x) -> 0.0
6467 if (AllowNewFpConst && DAG.getTarget().Options.UnsafeFPMath &&
6468 N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1)
6469 return DAG.getConstantFP(0.0, VT);
6471 // If allow, fold (fadd x, (fneg x)) -> 0.0
6472 if (AllowNewFpConst && DAG.getTarget().Options.UnsafeFPMath &&
6473 N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0)
6474 return DAG.getConstantFP(0.0, VT);
6476 // In unsafe math mode, we can fold chains of FADD's of the same value
6477 // into multiplications. This transform is not safe in general because
6478 // we are reducing the number of rounding steps.
6479 if (DAG.getTarget().Options.UnsafeFPMath &&
6480 TLI.isOperationLegalOrCustom(ISD::FMUL, VT) &&
6482 if (N0.getOpcode() == ISD::FMUL) {
6483 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
6484 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
6486 // (fadd (fmul c, x), x) -> (fmul x, c+1)
6487 if (CFP00 && !CFP01 && N0.getOperand(1) == N1) {
6488 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6490 DAG.getConstantFP(1.0, VT));
6491 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6495 // (fadd (fmul x, c), x) -> (fmul x, c+1)
6496 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
6497 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6499 DAG.getConstantFP(1.0, VT));
6500 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6504 // (fadd (fmul c, x), (fadd x, x)) -> (fmul x, c+2)
6505 if (CFP00 && !CFP01 && N1.getOpcode() == ISD::FADD &&
6506 N1.getOperand(0) == N1.getOperand(1) &&
6507 N0.getOperand(1) == N1.getOperand(0)) {
6508 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6510 DAG.getConstantFP(2.0, VT));
6511 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6512 N0.getOperand(1), NewCFP);
6515 // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2)
6516 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
6517 N1.getOperand(0) == N1.getOperand(1) &&
6518 N0.getOperand(0) == N1.getOperand(0)) {
6519 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6521 DAG.getConstantFP(2.0, VT));
6522 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6523 N0.getOperand(0), NewCFP);
6527 if (N1.getOpcode() == ISD::FMUL) {
6528 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
6529 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
6531 // (fadd x, (fmul c, x)) -> (fmul x, c+1)
6532 if (CFP10 && !CFP11 && N1.getOperand(1) == N0) {
6533 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6535 DAG.getConstantFP(1.0, VT));
6536 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6540 // (fadd x, (fmul x, c)) -> (fmul x, c+1)
6541 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
6542 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6544 DAG.getConstantFP(1.0, VT));
6545 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6550 // (fadd (fadd x, x), (fmul c, x)) -> (fmul x, c+2)
6551 if (CFP10 && !CFP11 && N0.getOpcode() == ISD::FADD &&
6552 N0.getOperand(0) == N0.getOperand(1) &&
6553 N1.getOperand(1) == N0.getOperand(0)) {
6554 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6556 DAG.getConstantFP(2.0, VT));
6557 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6558 N1.getOperand(1), NewCFP);
6561 // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2)
6562 if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD &&
6563 N0.getOperand(0) == N0.getOperand(1) &&
6564 N1.getOperand(0) == N0.getOperand(0)) {
6565 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6567 DAG.getConstantFP(2.0, VT));
6568 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6569 N1.getOperand(0), NewCFP);
6573 if (N0.getOpcode() == ISD::FADD && AllowNewFpConst) {
6574 ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
6575 // (fadd (fadd x, x), x) -> (fmul x, 3.0)
6576 if (!CFP && N0.getOperand(0) == N0.getOperand(1) &&
6577 (N0.getOperand(0) == N1))
6578 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6579 N1, DAG.getConstantFP(3.0, VT));
6582 if (N1.getOpcode() == ISD::FADD && AllowNewFpConst) {
6583 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
6584 // (fadd x, (fadd x, x)) -> (fmul x, 3.0)
6585 if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) &&
6586 N1.getOperand(0) == N0)
6587 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6588 N0, DAG.getConstantFP(3.0, VT));
6591 // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0)
6592 if (AllowNewFpConst &&
6593 N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
6594 N0.getOperand(0) == N0.getOperand(1) &&
6595 N1.getOperand(0) == N1.getOperand(1) &&
6596 N0.getOperand(0) == N1.getOperand(0))
6597 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6599 DAG.getConstantFP(4.0, VT));
6602 // FADD -> FMA combines:
6603 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
6604 DAG.getTarget().Options.UnsafeFPMath) &&
6605 DAG.getTarget().getTargetLowering()->isFMAFasterThanFMulAndFAdd(VT) &&
6606 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
6608 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
6609 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse())
6610 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6611 N0.getOperand(0), N0.getOperand(1), N1);
6613 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
6614 // Note: Commutes FADD operands.
6615 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse())
6616 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6617 N1.getOperand(0), N1.getOperand(1), N0);
6623 SDValue DAGCombiner::visitFSUB(SDNode *N) {
6624 SDValue N0 = N->getOperand(0);
6625 SDValue N1 = N->getOperand(1);
6626 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6627 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6628 EVT VT = N->getValueType(0);
6632 if (VT.isVector()) {
6633 SDValue FoldedVOp = SimplifyVBinOp(N);
6634 if (FoldedVOp.getNode()) return FoldedVOp;
6637 // fold (fsub c1, c2) -> c1-c2
6639 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0, N1);
6640 // fold (fsub A, 0) -> A
6641 if (DAG.getTarget().Options.UnsafeFPMath &&
6642 N1CFP && N1CFP->getValueAPF().isZero())
6644 // fold (fsub 0, B) -> -B
6645 if (DAG.getTarget().Options.UnsafeFPMath &&
6646 N0CFP && N0CFP->getValueAPF().isZero()) {
6647 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
6648 return GetNegatedExpression(N1, DAG, LegalOperations);
6649 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6650 return DAG.getNode(ISD::FNEG, dl, VT, N1);
6652 // fold (fsub A, (fneg B)) -> (fadd A, B)
6653 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
6654 return DAG.getNode(ISD::FADD, dl, VT, N0,
6655 GetNegatedExpression(N1, DAG, LegalOperations));
6657 // If 'unsafe math' is enabled, fold
6658 // (fsub x, x) -> 0.0 &
6659 // (fsub x, (fadd x, y)) -> (fneg y) &
6660 // (fsub x, (fadd y, x)) -> (fneg y)
6661 if (DAG.getTarget().Options.UnsafeFPMath) {
6663 return DAG.getConstantFP(0.0f, VT);
6665 if (N1.getOpcode() == ISD::FADD) {
6666 SDValue N10 = N1->getOperand(0);
6667 SDValue N11 = N1->getOperand(1);
6669 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI,
6670 &DAG.getTarget().Options))
6671 return GetNegatedExpression(N11, DAG, LegalOperations);
6673 if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI,
6674 &DAG.getTarget().Options))
6675 return GetNegatedExpression(N10, DAG, LegalOperations);
6679 // FSUB -> FMA combines:
6680 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
6681 DAG.getTarget().Options.UnsafeFPMath) &&
6682 DAG.getTarget().getTargetLowering()->isFMAFasterThanFMulAndFAdd(VT) &&
6683 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
6685 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
6686 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse())
6687 return DAG.getNode(ISD::FMA, dl, VT,
6688 N0.getOperand(0), N0.getOperand(1),
6689 DAG.getNode(ISD::FNEG, dl, VT, N1));
6691 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
6692 // Note: Commutes FSUB operands.
6693 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse())
6694 return DAG.getNode(ISD::FMA, dl, VT,
6695 DAG.getNode(ISD::FNEG, dl, VT,
6697 N1.getOperand(1), N0);
6699 // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
6700 if (N0.getOpcode() == ISD::FNEG &&
6701 N0.getOperand(0).getOpcode() == ISD::FMUL &&
6702 N0->hasOneUse() && N0.getOperand(0).hasOneUse()) {
6703 SDValue N00 = N0.getOperand(0).getOperand(0);
6704 SDValue N01 = N0.getOperand(0).getOperand(1);
6705 return DAG.getNode(ISD::FMA, dl, VT,
6706 DAG.getNode(ISD::FNEG, dl, VT, N00), N01,
6707 DAG.getNode(ISD::FNEG, dl, VT, N1));
6714 SDValue DAGCombiner::visitFMUL(SDNode *N) {
6715 SDValue N0 = N->getOperand(0);
6716 SDValue N1 = N->getOperand(1);
6717 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6718 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6719 EVT VT = N->getValueType(0);
6720 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6723 if (VT.isVector()) {
6724 SDValue FoldedVOp = SimplifyVBinOp(N);
6725 if (FoldedVOp.getNode()) return FoldedVOp;
6728 // fold (fmul c1, c2) -> c1*c2
6730 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, N1);
6731 // canonicalize constant to RHS
6732 if (N0CFP && !N1CFP)
6733 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, N0);
6734 // fold (fmul A, 0) -> 0
6735 if (DAG.getTarget().Options.UnsafeFPMath &&
6736 N1CFP && N1CFP->getValueAPF().isZero())
6738 // fold (fmul A, 0) -> 0, vector edition.
6739 if (DAG.getTarget().Options.UnsafeFPMath &&
6740 ISD::isBuildVectorAllZeros(N1.getNode()))
6742 // fold (fmul A, 1.0) -> A
6743 if (N1CFP && N1CFP->isExactlyValue(1.0))
6745 // fold (fmul X, 2.0) -> (fadd X, X)
6746 if (N1CFP && N1CFP->isExactlyValue(+2.0))
6747 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N0);
6748 // fold (fmul X, -1.0) -> (fneg X)
6749 if (N1CFP && N1CFP->isExactlyValue(-1.0))
6750 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6751 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
6753 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
6754 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
6755 &DAG.getTarget().Options)) {
6756 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
6757 &DAG.getTarget().Options)) {
6758 // Both can be negated for free, check to see if at least one is cheaper
6760 if (LHSNeg == 2 || RHSNeg == 2)
6761 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6762 GetNegatedExpression(N0, DAG, LegalOperations),
6763 GetNegatedExpression(N1, DAG, LegalOperations));
6767 // If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
6768 if (DAG.getTarget().Options.UnsafeFPMath &&
6769 N1CFP && N0.getOpcode() == ISD::FMUL &&
6770 N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
6771 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
6772 DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6773 N0.getOperand(1), N1));
6778 SDValue DAGCombiner::visitFMA(SDNode *N) {
6779 SDValue N0 = N->getOperand(0);
6780 SDValue N1 = N->getOperand(1);
6781 SDValue N2 = N->getOperand(2);
6782 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6783 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6784 EVT VT = N->getValueType(0);
6787 if (DAG.getTarget().Options.UnsafeFPMath) {
6788 if (N0CFP && N0CFP->isZero())
6790 if (N1CFP && N1CFP->isZero())
6793 if (N0CFP && N0CFP->isExactlyValue(1.0))
6794 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2);
6795 if (N1CFP && N1CFP->isExactlyValue(1.0))
6796 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2);
6798 // Canonicalize (fma c, x, y) -> (fma x, c, y)
6799 if (N0CFP && !N1CFP)
6800 return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2);
6802 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
6803 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6804 N2.getOpcode() == ISD::FMUL &&
6805 N0 == N2.getOperand(0) &&
6806 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
6807 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6808 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
6812 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
6813 if (DAG.getTarget().Options.UnsafeFPMath &&
6814 N0.getOpcode() == ISD::FMUL && N1CFP &&
6815 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
6816 return DAG.getNode(ISD::FMA, dl, VT,
6818 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
6822 // (fma x, 1, y) -> (fadd x, y)
6823 // (fma x, -1, y) -> (fadd (fneg x), y)
6825 if (N1CFP->isExactlyValue(1.0))
6826 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
6828 if (N1CFP->isExactlyValue(-1.0) &&
6829 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
6830 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
6831 AddToWorkList(RHSNeg.getNode());
6832 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
6836 // (fma x, c, x) -> (fmul x, (c+1))
6837 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP && N0 == N2)
6838 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6839 DAG.getNode(ISD::FADD, dl, VT,
6840 N1, DAG.getConstantFP(1.0, VT)));
6842 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
6843 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6844 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0)
6845 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6846 DAG.getNode(ISD::FADD, dl, VT,
6847 N1, DAG.getConstantFP(-1.0, VT)));
6853 SDValue DAGCombiner::visitFDIV(SDNode *N) {
6854 SDValue N0 = N->getOperand(0);
6855 SDValue N1 = N->getOperand(1);
6856 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6857 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6858 EVT VT = N->getValueType(0);
6859 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6862 if (VT.isVector()) {
6863 SDValue FoldedVOp = SimplifyVBinOp(N);
6864 if (FoldedVOp.getNode()) return FoldedVOp;
6867 // fold (fdiv c1, c2) -> c1/c2
6869 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1);
6871 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
6872 if (N1CFP && DAG.getTarget().Options.UnsafeFPMath) {
6873 // Compute the reciprocal 1.0 / c2.
6874 APFloat N1APF = N1CFP->getValueAPF();
6875 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
6876 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
6877 // Only do the transform if the reciprocal is a legal fp immediate that
6878 // isn't too nasty (eg NaN, denormal, ...).
6879 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
6880 (!LegalOperations ||
6881 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
6882 // backend)... we should handle this gracefully after Legalize.
6883 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
6884 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
6885 TLI.isFPImmLegal(Recip, VT)))
6886 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0,
6887 DAG.getConstantFP(Recip, VT));
6890 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
6891 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
6892 &DAG.getTarget().Options)) {
6893 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
6894 &DAG.getTarget().Options)) {
6895 // Both can be negated for free, check to see if at least one is cheaper
6897 if (LHSNeg == 2 || RHSNeg == 2)
6898 return DAG.getNode(ISD::FDIV, SDLoc(N), VT,
6899 GetNegatedExpression(N0, DAG, LegalOperations),
6900 GetNegatedExpression(N1, DAG, LegalOperations));
6907 SDValue DAGCombiner::visitFREM(SDNode *N) {
6908 SDValue N0 = N->getOperand(0);
6909 SDValue N1 = N->getOperand(1);
6910 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6911 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6912 EVT VT = N->getValueType(0);
6914 // fold (frem c1, c2) -> fmod(c1,c2)
6916 return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1);
6921 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
6922 SDValue N0 = N->getOperand(0);
6923 SDValue N1 = N->getOperand(1);
6924 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6925 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6926 EVT VT = N->getValueType(0);
6928 if (N0CFP && N1CFP) // Constant fold
6929 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1);
6932 const APFloat& V = N1CFP->getValueAPF();
6933 // copysign(x, c1) -> fabs(x) iff ispos(c1)
6934 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
6935 if (!V.isNegative()) {
6936 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
6937 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
6939 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6940 return DAG.getNode(ISD::FNEG, SDLoc(N), VT,
6941 DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0));
6945 // copysign(fabs(x), y) -> copysign(x, y)
6946 // copysign(fneg(x), y) -> copysign(x, y)
6947 // copysign(copysign(x,z), y) -> copysign(x, y)
6948 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
6949 N0.getOpcode() == ISD::FCOPYSIGN)
6950 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
6951 N0.getOperand(0), N1);
6953 // copysign(x, abs(y)) -> abs(x)
6954 if (N1.getOpcode() == ISD::FABS)
6955 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
6957 // copysign(x, copysign(y,z)) -> copysign(x, z)
6958 if (N1.getOpcode() == ISD::FCOPYSIGN)
6959 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
6960 N0, N1.getOperand(1));
6962 // copysign(x, fp_extend(y)) -> copysign(x, y)
6963 // copysign(x, fp_round(y)) -> copysign(x, y)
6964 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
6965 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
6966 N0, N1.getOperand(0));
6971 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
6972 SDValue N0 = N->getOperand(0);
6973 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
6974 EVT VT = N->getValueType(0);
6975 EVT OpVT = N0.getValueType();
6977 // fold (sint_to_fp c1) -> c1fp
6979 // ...but only if the target supports immediate floating-point values
6980 (!LegalOperations ||
6981 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
6982 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
6984 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
6985 // but UINT_TO_FP is legal on this target, try to convert.
6986 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
6987 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
6988 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
6989 if (DAG.SignBitIsZero(N0))
6990 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
6993 // The next optimizations are desirable only if SELECT_CC can be lowered.
6994 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
6995 // having to say they don't support SELECT_CC on every type the DAG knows
6996 // about, since there is no way to mark an opcode illegal at all value types
6997 // (See also visitSELECT)
6998 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other)) {
6999 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7000 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
7002 (!LegalOperations ||
7003 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7005 { N0.getOperand(0), N0.getOperand(1),
7006 DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT),
7008 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops, 5);
7011 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
7012 // (select_cc x, y, 1.0, 0.0,, cc)
7013 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
7014 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
7015 (!LegalOperations ||
7016 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7018 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
7019 DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT),
7020 N0.getOperand(0).getOperand(2) };
7021 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops, 5);
7028 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
7029 SDValue N0 = N->getOperand(0);
7030 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
7031 EVT VT = N->getValueType(0);
7032 EVT OpVT = N0.getValueType();
7034 // fold (uint_to_fp c1) -> c1fp
7036 // ...but only if the target supports immediate floating-point values
7037 (!LegalOperations ||
7038 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
7039 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
7041 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
7042 // but SINT_TO_FP is legal on this target, try to convert.
7043 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
7044 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
7045 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
7046 if (DAG.SignBitIsZero(N0))
7047 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
7050 // The next optimizations are desirable only if SELECT_CC can be lowered.
7051 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
7052 // having to say they don't support SELECT_CC on every type the DAG knows
7053 // about, since there is no way to mark an opcode illegal at all value types
7054 // (See also visitSELECT)
7055 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other)) {
7056 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
7058 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
7059 (!LegalOperations ||
7060 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
7062 { N0.getOperand(0), N0.getOperand(1),
7063 DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT),
7065 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops, 5);
7072 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
7073 SDValue N0 = N->getOperand(0);
7074 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7075 EVT VT = N->getValueType(0);
7077 // fold (fp_to_sint c1fp) -> c1
7079 return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0);
7084 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
7085 SDValue N0 = N->getOperand(0);
7086 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7087 EVT VT = N->getValueType(0);
7089 // fold (fp_to_uint c1fp) -> c1
7091 return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0);
7096 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
7097 SDValue N0 = N->getOperand(0);
7098 SDValue N1 = N->getOperand(1);
7099 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7100 EVT VT = N->getValueType(0);
7102 // fold (fp_round c1fp) -> c1fp
7104 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1);
7106 // fold (fp_round (fp_extend x)) -> x
7107 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
7108 return N0.getOperand(0);
7110 // fold (fp_round (fp_round x)) -> (fp_round x)
7111 if (N0.getOpcode() == ISD::FP_ROUND) {
7112 // This is a value preserving truncation if both round's are.
7113 bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
7114 N0.getNode()->getConstantOperandVal(1) == 1;
7115 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0.getOperand(0),
7116 DAG.getIntPtrConstant(IsTrunc));
7119 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
7120 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
7121 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT,
7122 N0.getOperand(0), N1);
7123 AddToWorkList(Tmp.getNode());
7124 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
7125 Tmp, N0.getOperand(1));
7131 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
7132 SDValue N0 = N->getOperand(0);
7133 EVT VT = N->getValueType(0);
7134 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
7135 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7137 // fold (fp_round_inreg c1fp) -> c1fp
7138 if (N0CFP && isTypeLegal(EVT)) {
7139 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
7140 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, Round);
7146 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
7147 SDValue N0 = N->getOperand(0);
7148 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7149 EVT VT = N->getValueType(0);
7151 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
7152 if (N->hasOneUse() &&
7153 N->use_begin()->getOpcode() == ISD::FP_ROUND)
7156 // fold (fp_extend c1fp) -> c1fp
7158 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0);
7160 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
7162 if (N0.getOpcode() == ISD::FP_ROUND
7163 && N0.getNode()->getConstantOperandVal(1) == 1) {
7164 SDValue In = N0.getOperand(0);
7165 if (In.getValueType() == VT) return In;
7166 if (VT.bitsLT(In.getValueType()))
7167 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT,
7168 In, N0.getOperand(1));
7169 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In);
7172 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
7173 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7174 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
7175 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
7176 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
7177 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
7179 LN0->getBasePtr(), N0.getValueType(),
7180 LN0->getMemOperand());
7181 CombineTo(N, ExtLoad);
7182 CombineTo(N0.getNode(),
7183 DAG.getNode(ISD::FP_ROUND, SDLoc(N0),
7184 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
7185 ExtLoad.getValue(1));
7186 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7192 SDValue DAGCombiner::visitFNEG(SDNode *N) {
7193 SDValue N0 = N->getOperand(0);
7194 EVT VT = N->getValueType(0);
7196 if (VT.isVector()) {
7197 SDValue FoldedVOp = SimplifyVUnaryOp(N);
7198 if (FoldedVOp.getNode()) return FoldedVOp;
7201 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
7202 &DAG.getTarget().Options))
7203 return GetNegatedExpression(N0, DAG, LegalOperations);
7205 // Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading
7206 // constant pool values.
7207 if (!TLI.isFNegFree(VT) && N0.getOpcode() == ISD::BITCAST &&
7209 N0.getNode()->hasOneUse() &&
7210 N0.getOperand(0).getValueType().isInteger()) {
7211 SDValue Int = N0.getOperand(0);
7212 EVT IntVT = Int.getValueType();
7213 if (IntVT.isInteger() && !IntVT.isVector()) {
7214 Int = DAG.getNode(ISD::XOR, SDLoc(N0), IntVT, Int,
7215 DAG.getConstant(APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
7216 AddToWorkList(Int.getNode());
7217 return DAG.getNode(ISD::BITCAST, SDLoc(N),
7222 // (fneg (fmul c, x)) -> (fmul -c, x)
7223 if (N0.getOpcode() == ISD::FMUL) {
7224 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
7226 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
7228 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
7235 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
7236 SDValue N0 = N->getOperand(0);
7237 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7238 EVT VT = N->getValueType(0);
7240 // fold (fceil c1) -> fceil(c1)
7242 return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0);
7247 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
7248 SDValue N0 = N->getOperand(0);
7249 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7250 EVT VT = N->getValueType(0);
7252 // fold (ftrunc c1) -> ftrunc(c1)
7254 return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0);
7259 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
7260 SDValue N0 = N->getOperand(0);
7261 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7262 EVT VT = N->getValueType(0);
7264 // fold (ffloor c1) -> ffloor(c1)
7266 return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0);
7271 SDValue DAGCombiner::visitFABS(SDNode *N) {
7272 SDValue N0 = N->getOperand(0);
7273 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7274 EVT VT = N->getValueType(0);
7276 if (VT.isVector()) {
7277 SDValue FoldedVOp = SimplifyVUnaryOp(N);
7278 if (FoldedVOp.getNode()) return FoldedVOp;
7281 // fold (fabs c1) -> fabs(c1)
7283 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7284 // fold (fabs (fabs x)) -> (fabs x)
7285 if (N0.getOpcode() == ISD::FABS)
7286 return N->getOperand(0);
7287 // fold (fabs (fneg x)) -> (fabs x)
7288 // fold (fabs (fcopysign x, y)) -> (fabs x)
7289 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
7290 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0));
7292 // Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading
7293 // constant pool values.
7294 if (!TLI.isFAbsFree(VT) &&
7295 N0.getOpcode() == ISD::BITCAST && N0.getNode()->hasOneUse() &&
7296 N0.getOperand(0).getValueType().isInteger() &&
7297 !N0.getOperand(0).getValueType().isVector()) {
7298 SDValue Int = N0.getOperand(0);
7299 EVT IntVT = Int.getValueType();
7300 if (IntVT.isInteger() && !IntVT.isVector()) {
7301 Int = DAG.getNode(ISD::AND, SDLoc(N0), IntVT, Int,
7302 DAG.getConstant(~APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
7303 AddToWorkList(Int.getNode());
7304 return DAG.getNode(ISD::BITCAST, SDLoc(N),
7305 N->getValueType(0), Int);
7312 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
7313 SDValue Chain = N->getOperand(0);
7314 SDValue N1 = N->getOperand(1);
7315 SDValue N2 = N->getOperand(2);
7317 // If N is a constant we could fold this into a fallthrough or unconditional
7318 // branch. However that doesn't happen very often in normal code, because
7319 // Instcombine/SimplifyCFG should have handled the available opportunities.
7320 // If we did this folding here, it would be necessary to update the
7321 // MachineBasicBlock CFG, which is awkward.
7323 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
7325 if (N1.getOpcode() == ISD::SETCC &&
7326 TLI.isOperationLegalOrCustom(ISD::BR_CC,
7327 N1.getOperand(0).getValueType())) {
7328 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
7329 Chain, N1.getOperand(2),
7330 N1.getOperand(0), N1.getOperand(1), N2);
7333 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
7334 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
7335 (N1.getOperand(0).hasOneUse() &&
7336 N1.getOperand(0).getOpcode() == ISD::SRL))) {
7338 if (N1.getOpcode() == ISD::TRUNCATE) {
7339 // Look pass the truncate.
7340 Trunc = N1.getNode();
7341 N1 = N1.getOperand(0);
7344 // Match this pattern so that we can generate simpler code:
7347 // %b = and i32 %a, 2
7348 // %c = srl i32 %b, 1
7349 // brcond i32 %c ...
7354 // %b = and i32 %a, 2
7355 // %c = setcc eq %b, 0
7358 // This applies only when the AND constant value has one bit set and the
7359 // SRL constant is equal to the log2 of the AND constant. The back-end is
7360 // smart enough to convert the result into a TEST/JMP sequence.
7361 SDValue Op0 = N1.getOperand(0);
7362 SDValue Op1 = N1.getOperand(1);
7364 if (Op0.getOpcode() == ISD::AND &&
7365 Op1.getOpcode() == ISD::Constant) {
7366 SDValue AndOp1 = Op0.getOperand(1);
7368 if (AndOp1.getOpcode() == ISD::Constant) {
7369 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
7371 if (AndConst.isPowerOf2() &&
7372 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
7374 DAG.getSetCC(SDLoc(N),
7375 getSetCCResultType(Op0.getValueType()),
7376 Op0, DAG.getConstant(0, Op0.getValueType()),
7379 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, SDLoc(N),
7380 MVT::Other, Chain, SetCC, N2);
7381 // Don't add the new BRCond into the worklist or else SimplifySelectCC
7382 // will convert it back to (X & C1) >> C2.
7383 CombineTo(N, NewBRCond, false);
7384 // Truncate is dead.
7386 removeFromWorkList(Trunc);
7387 DAG.DeleteNode(Trunc);
7389 // Replace the uses of SRL with SETCC
7390 WorkListRemover DeadNodes(*this);
7391 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
7392 removeFromWorkList(N1.getNode());
7393 DAG.DeleteNode(N1.getNode());
7394 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7400 // Restore N1 if the above transformation doesn't match.
7401 N1 = N->getOperand(1);
7404 // Transform br(xor(x, y)) -> br(x != y)
7405 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
7406 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
7407 SDNode *TheXor = N1.getNode();
7408 SDValue Op0 = TheXor->getOperand(0);
7409 SDValue Op1 = TheXor->getOperand(1);
7410 if (Op0.getOpcode() == Op1.getOpcode()) {
7411 // Avoid missing important xor optimizations.
7412 SDValue Tmp = visitXOR(TheXor);
7413 if (Tmp.getNode()) {
7414 if (Tmp.getNode() != TheXor) {
7415 DEBUG(dbgs() << "\nReplacing.8 ";
7417 dbgs() << "\nWith: ";
7418 Tmp.getNode()->dump(&DAG);
7420 WorkListRemover DeadNodes(*this);
7421 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
7422 removeFromWorkList(TheXor);
7423 DAG.DeleteNode(TheXor);
7424 return DAG.getNode(ISD::BRCOND, SDLoc(N),
7425 MVT::Other, Chain, Tmp, N2);
7428 // visitXOR has changed XOR's operands or replaced the XOR completely,
7430 return SDValue(N, 0);
7434 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
7436 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
7437 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
7438 Op0.getOpcode() == ISD::XOR) {
7439 TheXor = Op0.getNode();
7443 EVT SetCCVT = N1.getValueType();
7445 SetCCVT = getSetCCResultType(SetCCVT);
7446 SDValue SetCC = DAG.getSetCC(SDLoc(TheXor),
7449 Equal ? ISD::SETEQ : ISD::SETNE);
7450 // Replace the uses of XOR with SETCC
7451 WorkListRemover DeadNodes(*this);
7452 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
7453 removeFromWorkList(N1.getNode());
7454 DAG.DeleteNode(N1.getNode());
7455 return DAG.getNode(ISD::BRCOND, SDLoc(N),
7456 MVT::Other, Chain, SetCC, N2);
7463 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
7465 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
7466 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
7467 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
7469 // If N is a constant we could fold this into a fallthrough or unconditional
7470 // branch. However that doesn't happen very often in normal code, because
7471 // Instcombine/SimplifyCFG should have handled the available opportunities.
7472 // If we did this folding here, it would be necessary to update the
7473 // MachineBasicBlock CFG, which is awkward.
7475 // Use SimplifySetCC to simplify SETCC's.
7476 SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()),
7477 CondLHS, CondRHS, CC->get(), SDLoc(N),
7479 if (Simp.getNode()) AddToWorkList(Simp.getNode());
7481 // fold to a simpler setcc
7482 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
7483 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
7484 N->getOperand(0), Simp.getOperand(2),
7485 Simp.getOperand(0), Simp.getOperand(1),
7491 /// canFoldInAddressingMode - Return true if 'Use' is a load or a store that
7492 /// uses N as its base pointer and that N may be folded in the load / store
7493 /// addressing mode.
7494 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
7496 const TargetLowering &TLI) {
7498 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
7499 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
7501 VT = Use->getValueType(0);
7502 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
7503 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
7505 VT = ST->getValue().getValueType();
7509 TargetLowering::AddrMode AM;
7510 if (N->getOpcode() == ISD::ADD) {
7511 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
7514 AM.BaseOffs = Offset->getSExtValue();
7518 } else if (N->getOpcode() == ISD::SUB) {
7519 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
7522 AM.BaseOffs = -Offset->getSExtValue();
7529 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
7532 /// CombineToPreIndexedLoadStore - Try turning a load / store into a
7533 /// pre-indexed load / store when the base pointer is an add or subtract
7534 /// and it has other uses besides the load / store. After the
7535 /// transformation, the new indexed load / store has effectively folded
7536 /// the add / subtract in and all of its other uses are redirected to the
7537 /// new load / store.
7538 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
7539 if (Level < AfterLegalizeDAG)
7545 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7546 if (LD->isIndexed())
7548 VT = LD->getMemoryVT();
7549 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
7550 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
7552 Ptr = LD->getBasePtr();
7553 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7554 if (ST->isIndexed())
7556 VT = ST->getMemoryVT();
7557 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
7558 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
7560 Ptr = ST->getBasePtr();
7566 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
7567 // out. There is no reason to make this a preinc/predec.
7568 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
7569 Ptr.getNode()->hasOneUse())
7572 // Ask the target to do addressing mode selection.
7575 ISD::MemIndexedMode AM = ISD::UNINDEXED;
7576 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
7579 // Backends without true r+i pre-indexed forms may need to pass a
7580 // constant base with a variable offset so that constant coercion
7581 // will work with the patterns in canonical form.
7582 bool Swapped = false;
7583 if (isa<ConstantSDNode>(BasePtr)) {
7584 std::swap(BasePtr, Offset);
7588 // Don't create a indexed load / store with zero offset.
7589 if (isa<ConstantSDNode>(Offset) &&
7590 cast<ConstantSDNode>(Offset)->isNullValue())
7593 // Try turning it into a pre-indexed load / store except when:
7594 // 1) The new base ptr is a frame index.
7595 // 2) If N is a store and the new base ptr is either the same as or is a
7596 // predecessor of the value being stored.
7597 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
7598 // that would create a cycle.
7599 // 4) All uses are load / store ops that use it as old base ptr.
7601 // Check #1. Preinc'ing a frame index would require copying the stack pointer
7602 // (plus the implicit offset) to a register to preinc anyway.
7603 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
7608 SDValue Val = cast<StoreSDNode>(N)->getValue();
7609 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
7613 // If the offset is a constant, there may be other adds of constants that
7614 // can be folded with this one. We should do this to avoid having to keep
7615 // a copy of the original base pointer.
7616 SmallVector<SDNode *, 16> OtherUses;
7617 if (isa<ConstantSDNode>(Offset))
7618 for (SDNode *Use : BasePtr.getNode()->uses()) {
7619 if (Use == Ptr.getNode())
7622 if (Use->isPredecessorOf(N))
7625 if (Use->getOpcode() != ISD::ADD && Use->getOpcode() != ISD::SUB) {
7630 SDValue Op0 = Use->getOperand(0), Op1 = Use->getOperand(1);
7631 if (Op1.getNode() == BasePtr.getNode())
7632 std::swap(Op0, Op1);
7633 assert(Op0.getNode() == BasePtr.getNode() &&
7634 "Use of ADD/SUB but not an operand");
7636 if (!isa<ConstantSDNode>(Op1)) {
7641 // FIXME: In some cases, we can be smarter about this.
7642 if (Op1.getValueType() != Offset.getValueType()) {
7647 OtherUses.push_back(Use);
7651 std::swap(BasePtr, Offset);
7653 // Now check for #3 and #4.
7654 bool RealUse = false;
7656 // Caches for hasPredecessorHelper
7657 SmallPtrSet<const SDNode *, 32> Visited;
7658 SmallVector<const SDNode *, 16> Worklist;
7660 for (SDNode *Use : Ptr.getNode()->uses()) {
7663 if (N->hasPredecessorHelper(Use, Visited, Worklist))
7666 // If Ptr may be folded in addressing mode of other use, then it's
7667 // not profitable to do this transformation.
7668 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
7677 Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
7678 BasePtr, Offset, AM);
7680 Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
7681 BasePtr, Offset, AM);
7684 DEBUG(dbgs() << "\nReplacing.4 ";
7686 dbgs() << "\nWith: ";
7687 Result.getNode()->dump(&DAG);
7689 WorkListRemover DeadNodes(*this);
7691 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
7692 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
7694 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
7697 // Finally, since the node is now dead, remove it from the graph.
7701 std::swap(BasePtr, Offset);
7703 // Replace other uses of BasePtr that can be updated to use Ptr
7704 for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) {
7705 unsigned OffsetIdx = 1;
7706 if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode())
7708 assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() ==
7709 BasePtr.getNode() && "Expected BasePtr operand");
7711 // We need to replace ptr0 in the following expression:
7712 // x0 * offset0 + y0 * ptr0 = t0
7714 // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store)
7716 // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the
7717 // indexed load/store and the expresion that needs to be re-written.
7719 // Therefore, we have:
7720 // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1
7722 ConstantSDNode *CN =
7723 cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx));
7725 APInt Offset0 = CN->getAPIntValue();
7726 APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue();
7728 X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1;
7729 Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1;
7730 X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1;
7731 Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1;
7733 unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD;
7735 APInt CNV = Offset0;
7736 if (X0 < 0) CNV = -CNV;
7737 if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1;
7738 else CNV = CNV - Offset1;
7740 // We can now generate the new expression.
7741 SDValue NewOp1 = DAG.getConstant(CNV, CN->getValueType(0));
7742 SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0);
7744 SDValue NewUse = DAG.getNode(Opcode,
7745 SDLoc(OtherUses[i]),
7746 OtherUses[i]->getValueType(0), NewOp1, NewOp2);
7747 DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse);
7748 removeFromWorkList(OtherUses[i]);
7749 DAG.DeleteNode(OtherUses[i]);
7752 // Replace the uses of Ptr with uses of the updated base value.
7753 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
7754 removeFromWorkList(Ptr.getNode());
7755 DAG.DeleteNode(Ptr.getNode());
7760 /// CombineToPostIndexedLoadStore - Try to combine a load / store with a
7761 /// add / sub of the base pointer node into a post-indexed load / store.
7762 /// The transformation folded the add / subtract into the new indexed
7763 /// load / store effectively and all of its uses are redirected to the
7764 /// new load / store.
7765 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
7766 if (Level < AfterLegalizeDAG)
7772 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7773 if (LD->isIndexed())
7775 VT = LD->getMemoryVT();
7776 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
7777 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
7779 Ptr = LD->getBasePtr();
7780 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7781 if (ST->isIndexed())
7783 VT = ST->getMemoryVT();
7784 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
7785 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
7787 Ptr = ST->getBasePtr();
7793 if (Ptr.getNode()->hasOneUse())
7796 for (SDNode *Op : Ptr.getNode()->uses()) {
7798 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
7803 ISD::MemIndexedMode AM = ISD::UNINDEXED;
7804 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
7805 // Don't create a indexed load / store with zero offset.
7806 if (isa<ConstantSDNode>(Offset) &&
7807 cast<ConstantSDNode>(Offset)->isNullValue())
7810 // Try turning it into a post-indexed load / store except when
7811 // 1) All uses are load / store ops that use it as base ptr (and
7812 // it may be folded as addressing mmode).
7813 // 2) Op must be independent of N, i.e. Op is neither a predecessor
7814 // nor a successor of N. Otherwise, if Op is folded that would
7817 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
7821 bool TryNext = false;
7822 for (SDNode *Use : BasePtr.getNode()->uses()) {
7823 if (Use == Ptr.getNode())
7826 // If all the uses are load / store addresses, then don't do the
7828 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
7829 bool RealUse = false;
7830 for (SDNode *UseUse : Use->uses()) {
7831 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
7846 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
7847 SDValue Result = isLoad
7848 ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
7849 BasePtr, Offset, AM)
7850 : DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
7851 BasePtr, Offset, AM);
7854 DEBUG(dbgs() << "\nReplacing.5 ";
7856 dbgs() << "\nWith: ";
7857 Result.getNode()->dump(&DAG);
7859 WorkListRemover DeadNodes(*this);
7861 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
7862 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
7864 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
7867 // Finally, since the node is now dead, remove it from the graph.
7870 // Replace the uses of Use with uses of the updated base value.
7871 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
7872 Result.getValue(isLoad ? 1 : 0));
7873 removeFromWorkList(Op);
7883 SDValue DAGCombiner::visitLOAD(SDNode *N) {
7884 LoadSDNode *LD = cast<LoadSDNode>(N);
7885 SDValue Chain = LD->getChain();
7886 SDValue Ptr = LD->getBasePtr();
7888 // If load is not volatile and there are no uses of the loaded value (and
7889 // the updated indexed value in case of indexed loads), change uses of the
7890 // chain value into uses of the chain input (i.e. delete the dead load).
7891 if (!LD->isVolatile()) {
7892 if (N->getValueType(1) == MVT::Other) {
7894 if (!N->hasAnyUseOfValue(0)) {
7895 // It's not safe to use the two value CombineTo variant here. e.g.
7896 // v1, chain2 = load chain1, loc
7897 // v2, chain3 = load chain2, loc
7899 // Now we replace use of chain2 with chain1. This makes the second load
7900 // isomorphic to the one we are deleting, and thus makes this load live.
7901 DEBUG(dbgs() << "\nReplacing.6 ";
7903 dbgs() << "\nWith chain: ";
7904 Chain.getNode()->dump(&DAG);
7906 WorkListRemover DeadNodes(*this);
7907 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
7909 if (N->use_empty()) {
7910 removeFromWorkList(N);
7914 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7918 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
7919 if (!N->hasAnyUseOfValue(0) && !N->hasAnyUseOfValue(1)) {
7920 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
7921 DEBUG(dbgs() << "\nReplacing.7 ";
7923 dbgs() << "\nWith: ";
7924 Undef.getNode()->dump(&DAG);
7925 dbgs() << " and 2 other values\n");
7926 WorkListRemover DeadNodes(*this);
7927 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
7928 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1),
7929 DAG.getUNDEF(N->getValueType(1)));
7930 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
7931 removeFromWorkList(N);
7933 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7938 // If this load is directly stored, replace the load value with the stored
7940 // TODO: Handle store large -> read small portion.
7941 // TODO: Handle TRUNCSTORE/LOADEXT
7942 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
7943 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
7944 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
7945 if (PrevST->getBasePtr() == Ptr &&
7946 PrevST->getValue().getValueType() == N->getValueType(0))
7947 return CombineTo(N, Chain.getOperand(1), Chain);
7951 // Try to infer better alignment information than the load already has.
7952 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
7953 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
7954 if (Align > LD->getMemOperand()->getBaseAlignment()) {
7956 DAG.getExtLoad(LD->getExtensionType(), SDLoc(N),
7957 LD->getValueType(0),
7958 Chain, Ptr, LD->getPointerInfo(),
7960 LD->isVolatile(), LD->isNonTemporal(), Align,
7962 return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true);
7967 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA :
7968 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
7970 if (CombinerAAOnlyFunc.getNumOccurrences() &&
7971 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
7974 if (UseAA && LD->isUnindexed()) {
7975 // Walk up chain skipping non-aliasing memory nodes.
7976 SDValue BetterChain = FindBetterChain(N, Chain);
7978 // If there is a better chain.
7979 if (Chain != BetterChain) {
7982 // Replace the chain to void dependency.
7983 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
7984 ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD),
7985 BetterChain, Ptr, LD->getMemOperand());
7987 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD),
7988 LD->getValueType(0),
7989 BetterChain, Ptr, LD->getMemoryVT(),
7990 LD->getMemOperand());
7993 // Create token factor to keep old chain connected.
7994 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
7995 MVT::Other, Chain, ReplLoad.getValue(1));
7997 // Make sure the new and old chains are cleaned up.
7998 AddToWorkList(Token.getNode());
8000 // Replace uses with load result and token factor. Don't add users
8002 return CombineTo(N, ReplLoad.getValue(0), Token, false);
8006 // Try transforming N to an indexed load.
8007 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
8008 return SDValue(N, 0);
8010 // Try to slice up N to more direct loads if the slices are mapped to
8011 // different register banks or pairing can take place.
8013 return SDValue(N, 0);
8019 /// \brief Helper structure used to slice a load in smaller loads.
8020 /// Basically a slice is obtained from the following sequence:
8021 /// Origin = load Ty1, Base
8022 /// Shift = srl Ty1 Origin, CstTy Amount
8023 /// Inst = trunc Shift to Ty2
8025 /// Then, it will be rewriten into:
8026 /// Slice = load SliceTy, Base + SliceOffset
8027 /// [Inst = zext Slice to Ty2], only if SliceTy <> Ty2
8029 /// SliceTy is deduced from the number of bits that are actually used to
8031 struct LoadedSlice {
8032 /// \brief Helper structure used to compute the cost of a slice.
8034 /// Are we optimizing for code size.
8039 unsigned CrossRegisterBanksCopies;
8043 Cost(bool ForCodeSize = false)
8044 : ForCodeSize(ForCodeSize), Loads(0), Truncates(0),
8045 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {}
8047 /// \brief Get the cost of one isolated slice.
8048 Cost(const LoadedSlice &LS, bool ForCodeSize = false)
8049 : ForCodeSize(ForCodeSize), Loads(1), Truncates(0),
8050 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {
8051 EVT TruncType = LS.Inst->getValueType(0);
8052 EVT LoadedType = LS.getLoadedType();
8053 if (TruncType != LoadedType &&
8054 !LS.DAG->getTargetLoweringInfo().isZExtFree(LoadedType, TruncType))
8058 /// \brief Account for slicing gain in the current cost.
8059 /// Slicing provide a few gains like removing a shift or a
8060 /// truncate. This method allows to grow the cost of the original
8061 /// load with the gain from this slice.
8062 void addSliceGain(const LoadedSlice &LS) {
8063 // Each slice saves a truncate.
8064 const TargetLowering &TLI = LS.DAG->getTargetLoweringInfo();
8065 if (!TLI.isTruncateFree(LS.Inst->getValueType(0),
8066 LS.Inst->getOperand(0).getValueType()))
8068 // If there is a shift amount, this slice gets rid of it.
8071 // If this slice can merge a cross register bank copy, account for it.
8072 if (LS.canMergeExpensiveCrossRegisterBankCopy())
8073 ++CrossRegisterBanksCopies;
8076 Cost &operator+=(const Cost &RHS) {
8078 Truncates += RHS.Truncates;
8079 CrossRegisterBanksCopies += RHS.CrossRegisterBanksCopies;
8085 bool operator==(const Cost &RHS) const {
8086 return Loads == RHS.Loads && Truncates == RHS.Truncates &&
8087 CrossRegisterBanksCopies == RHS.CrossRegisterBanksCopies &&
8088 ZExts == RHS.ZExts && Shift == RHS.Shift;
8091 bool operator!=(const Cost &RHS) const { return !(*this == RHS); }
8093 bool operator<(const Cost &RHS) const {
8094 // Assume cross register banks copies are as expensive as loads.
8095 // FIXME: Do we want some more target hooks?
8096 unsigned ExpensiveOpsLHS = Loads + CrossRegisterBanksCopies;
8097 unsigned ExpensiveOpsRHS = RHS.Loads + RHS.CrossRegisterBanksCopies;
8098 // Unless we are optimizing for code size, consider the
8099 // expensive operation first.
8100 if (!ForCodeSize && ExpensiveOpsLHS != ExpensiveOpsRHS)
8101 return ExpensiveOpsLHS < ExpensiveOpsRHS;
8102 return (Truncates + ZExts + Shift + ExpensiveOpsLHS) <
8103 (RHS.Truncates + RHS.ZExts + RHS.Shift + ExpensiveOpsRHS);
8106 bool operator>(const Cost &RHS) const { return RHS < *this; }
8108 bool operator<=(const Cost &RHS) const { return !(RHS < *this); }
8110 bool operator>=(const Cost &RHS) const { return !(*this < RHS); }
8112 // The last instruction that represent the slice. This should be a
8113 // truncate instruction.
8115 // The original load instruction.
8117 // The right shift amount in bits from the original load.
8119 // The DAG from which Origin came from.
8120 // This is used to get some contextual information about legal types, etc.
8123 LoadedSlice(SDNode *Inst = NULL, LoadSDNode *Origin = NULL,
8124 unsigned Shift = 0, SelectionDAG *DAG = NULL)
8125 : Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {}
8127 LoadedSlice(const LoadedSlice &LS)
8128 : Inst(LS.Inst), Origin(LS.Origin), Shift(LS.Shift), DAG(LS.DAG) {}
8130 /// \brief Get the bits used in a chunk of bits \p BitWidth large.
8131 /// \return Result is \p BitWidth and has used bits set to 1 and
8132 /// not used bits set to 0.
8133 APInt getUsedBits() const {
8134 // Reproduce the trunc(lshr) sequence:
8135 // - Start from the truncated value.
8136 // - Zero extend to the desired bit width.
8138 assert(Origin && "No original load to compare against.");
8139 unsigned BitWidth = Origin->getValueSizeInBits(0);
8140 assert(Inst && "This slice is not bound to an instruction");
8141 assert(Inst->getValueSizeInBits(0) <= BitWidth &&
8142 "Extracted slice is bigger than the whole type!");
8143 APInt UsedBits(Inst->getValueSizeInBits(0), 0);
8144 UsedBits.setAllBits();
8145 UsedBits = UsedBits.zext(BitWidth);
8150 /// \brief Get the size of the slice to be loaded in bytes.
8151 unsigned getLoadedSize() const {
8152 unsigned SliceSize = getUsedBits().countPopulation();
8153 assert(!(SliceSize & 0x7) && "Size is not a multiple of a byte.");
8154 return SliceSize / 8;
8157 /// \brief Get the type that will be loaded for this slice.
8158 /// Note: This may not be the final type for the slice.
8159 EVT getLoadedType() const {
8160 assert(DAG && "Missing context");
8161 LLVMContext &Ctxt = *DAG->getContext();
8162 return EVT::getIntegerVT(Ctxt, getLoadedSize() * 8);
8165 /// \brief Get the alignment of the load used for this slice.
8166 unsigned getAlignment() const {
8167 unsigned Alignment = Origin->getAlignment();
8168 unsigned Offset = getOffsetFromBase();
8170 Alignment = MinAlign(Alignment, Alignment + Offset);
8174 /// \brief Check if this slice can be rewritten with legal operations.
8175 bool isLegal() const {
8176 // An invalid slice is not legal.
8177 if (!Origin || !Inst || !DAG)
8180 // Offsets are for indexed load only, we do not handle that.
8181 if (Origin->getOffset().getOpcode() != ISD::UNDEF)
8184 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
8186 // Check that the type is legal.
8187 EVT SliceType = getLoadedType();
8188 if (!TLI.isTypeLegal(SliceType))
8191 // Check that the load is legal for this type.
8192 if (!TLI.isOperationLegal(ISD::LOAD, SliceType))
8195 // Check that the offset can be computed.
8196 // 1. Check its type.
8197 EVT PtrType = Origin->getBasePtr().getValueType();
8198 if (PtrType == MVT::Untyped || PtrType.isExtended())
8201 // 2. Check that it fits in the immediate.
8202 if (!TLI.isLegalAddImmediate(getOffsetFromBase()))
8205 // 3. Check that the computation is legal.
8206 if (!TLI.isOperationLegal(ISD::ADD, PtrType))
8209 // Check that the zext is legal if it needs one.
8210 EVT TruncateType = Inst->getValueType(0);
8211 if (TruncateType != SliceType &&
8212 !TLI.isOperationLegal(ISD::ZERO_EXTEND, TruncateType))
8218 /// \brief Get the offset in bytes of this slice in the original chunk of
8220 /// \pre DAG != NULL.
8221 uint64_t getOffsetFromBase() const {
8222 assert(DAG && "Missing context.");
8224 DAG->getTargetLoweringInfo().getDataLayout()->isBigEndian();
8225 assert(!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported.");
8226 uint64_t Offset = Shift / 8;
8227 unsigned TySizeInBytes = Origin->getValueSizeInBits(0) / 8;
8228 assert(!(Origin->getValueSizeInBits(0) & 0x7) &&
8229 "The size of the original loaded type is not a multiple of a"
8231 // If Offset is bigger than TySizeInBytes, it means we are loading all
8232 // zeros. This should have been optimized before in the process.
8233 assert(TySizeInBytes > Offset &&
8234 "Invalid shift amount for given loaded size");
8236 Offset = TySizeInBytes - Offset - getLoadedSize();
8240 /// \brief Generate the sequence of instructions to load the slice
8241 /// represented by this object and redirect the uses of this slice to
8242 /// this new sequence of instructions.
8243 /// \pre this->Inst && this->Origin are valid Instructions and this
8244 /// object passed the legal check: LoadedSlice::isLegal returned true.
8245 /// \return The last instruction of the sequence used to load the slice.
8246 SDValue loadSlice() const {
8247 assert(Inst && Origin && "Unable to replace a non-existing slice.");
8248 const SDValue &OldBaseAddr = Origin->getBasePtr();
8249 SDValue BaseAddr = OldBaseAddr;
8250 // Get the offset in that chunk of bytes w.r.t. the endianess.
8251 int64_t Offset = static_cast<int64_t>(getOffsetFromBase());
8252 assert(Offset >= 0 && "Offset too big to fit in int64_t!");
8254 // BaseAddr = BaseAddr + Offset.
8255 EVT ArithType = BaseAddr.getValueType();
8256 BaseAddr = DAG->getNode(ISD::ADD, SDLoc(Origin), ArithType, BaseAddr,
8257 DAG->getConstant(Offset, ArithType));
8260 // Create the type of the loaded slice according to its size.
8261 EVT SliceType = getLoadedType();
8263 // Create the load for the slice.
8264 SDValue LastInst = DAG->getLoad(
8265 SliceType, SDLoc(Origin), Origin->getChain(), BaseAddr,
8266 Origin->getPointerInfo().getWithOffset(Offset), Origin->isVolatile(),
8267 Origin->isNonTemporal(), Origin->isInvariant(), getAlignment());
8268 // If the final type is not the same as the loaded type, this means that
8269 // we have to pad with zero. Create a zero extend for that.
8270 EVT FinalType = Inst->getValueType(0);
8271 if (SliceType != FinalType)
8273 DAG->getNode(ISD::ZERO_EXTEND, SDLoc(LastInst), FinalType, LastInst);
8277 /// \brief Check if this slice can be merged with an expensive cross register
8278 /// bank copy. E.g.,
8280 /// f = bitcast i32 i to float
8281 bool canMergeExpensiveCrossRegisterBankCopy() const {
8282 if (!Inst || !Inst->hasOneUse())
8284 SDNode *Use = *Inst->use_begin();
8285 if (Use->getOpcode() != ISD::BITCAST)
8287 assert(DAG && "Missing context");
8288 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
8289 EVT ResVT = Use->getValueType(0);
8290 const TargetRegisterClass *ResRC = TLI.getRegClassFor(ResVT.getSimpleVT());
8291 const TargetRegisterClass *ArgRC =
8292 TLI.getRegClassFor(Use->getOperand(0).getValueType().getSimpleVT());
8293 if (ArgRC == ResRC || !TLI.isOperationLegal(ISD::LOAD, ResVT))
8296 // At this point, we know that we perform a cross-register-bank copy.
8297 // Check if it is expensive.
8298 const TargetRegisterInfo *TRI = TLI.getTargetMachine().getRegisterInfo();
8299 // Assume bitcasts are cheap, unless both register classes do not
8300 // explicitly share a common sub class.
8301 if (!TRI || TRI->getCommonSubClass(ArgRC, ResRC))
8304 // Check if it will be merged with the load.
8305 // 1. Check the alignment constraint.
8306 unsigned RequiredAlignment = TLI.getDataLayout()->getABITypeAlignment(
8307 ResVT.getTypeForEVT(*DAG->getContext()));
8309 if (RequiredAlignment > getAlignment())
8312 // 2. Check that the load is a legal operation for that type.
8313 if (!TLI.isOperationLegal(ISD::LOAD, ResVT))
8316 // 3. Check that we do not have a zext in the way.
8317 if (Inst->getValueType(0) != getLoadedType())
8325 /// \brief Check that all bits set in \p UsedBits form a dense region, i.e.,
8326 /// \p UsedBits looks like 0..0 1..1 0..0.
8327 static bool areUsedBitsDense(const APInt &UsedBits) {
8328 // If all the bits are one, this is dense!
8329 if (UsedBits.isAllOnesValue())
8332 // Get rid of the unused bits on the right.
8333 APInt NarrowedUsedBits = UsedBits.lshr(UsedBits.countTrailingZeros());
8334 // Get rid of the unused bits on the left.
8335 if (NarrowedUsedBits.countLeadingZeros())
8336 NarrowedUsedBits = NarrowedUsedBits.trunc(NarrowedUsedBits.getActiveBits());
8337 // Check that the chunk of bits is completely used.
8338 return NarrowedUsedBits.isAllOnesValue();
8341 /// \brief Check whether or not \p First and \p Second are next to each other
8342 /// in memory. This means that there is no hole between the bits loaded
8343 /// by \p First and the bits loaded by \p Second.
8344 static bool areSlicesNextToEachOther(const LoadedSlice &First,
8345 const LoadedSlice &Second) {
8346 assert(First.Origin == Second.Origin && First.Origin &&
8347 "Unable to match different memory origins.");
8348 APInt UsedBits = First.getUsedBits();
8349 assert((UsedBits & Second.getUsedBits()) == 0 &&
8350 "Slices are not supposed to overlap.");
8351 UsedBits |= Second.getUsedBits();
8352 return areUsedBitsDense(UsedBits);
8355 /// \brief Adjust the \p GlobalLSCost according to the target
8356 /// paring capabilities and the layout of the slices.
8357 /// \pre \p GlobalLSCost should account for at least as many loads as
8358 /// there is in the slices in \p LoadedSlices.
8359 static void adjustCostForPairing(SmallVectorImpl<LoadedSlice> &LoadedSlices,
8360 LoadedSlice::Cost &GlobalLSCost) {
8361 unsigned NumberOfSlices = LoadedSlices.size();
8362 // If there is less than 2 elements, no pairing is possible.
8363 if (NumberOfSlices < 2)
8366 // Sort the slices so that elements that are likely to be next to each
8367 // other in memory are next to each other in the list.
8368 std::sort(LoadedSlices.begin(), LoadedSlices.end(),
8369 [](const LoadedSlice &LHS, const LoadedSlice &RHS) {
8370 assert(LHS.Origin == RHS.Origin && "Different bases not implemented.");
8371 return LHS.getOffsetFromBase() < RHS.getOffsetFromBase();
8373 const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo();
8374 // First (resp. Second) is the first (resp. Second) potentially candidate
8375 // to be placed in a paired load.
8376 const LoadedSlice *First = NULL;
8377 const LoadedSlice *Second = NULL;
8378 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice,
8379 // Set the beginning of the pair.
8382 Second = &LoadedSlices[CurrSlice];
8384 // If First is NULL, it means we start a new pair.
8385 // Get to the next slice.
8389 EVT LoadedType = First->getLoadedType();
8391 // If the types of the slices are different, we cannot pair them.
8392 if (LoadedType != Second->getLoadedType())
8395 // Check if the target supplies paired loads for this type.
8396 unsigned RequiredAlignment = 0;
8397 if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) {
8398 // move to the next pair, this type is hopeless.
8402 // Check if we meet the alignment requirement.
8403 if (RequiredAlignment > First->getAlignment())
8406 // Check that both loads are next to each other in memory.
8407 if (!areSlicesNextToEachOther(*First, *Second))
8410 assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!");
8411 --GlobalLSCost.Loads;
8412 // Move to the next pair.
8417 /// \brief Check the profitability of all involved LoadedSlice.
8418 /// Currently, it is considered profitable if there is exactly two
8419 /// involved slices (1) which are (2) next to each other in memory, and
8420 /// whose cost (\see LoadedSlice::Cost) is smaller than the original load (3).
8422 /// Note: The order of the elements in \p LoadedSlices may be modified, but not
8423 /// the elements themselves.
8425 /// FIXME: When the cost model will be mature enough, we can relax
8426 /// constraints (1) and (2).
8427 static bool isSlicingProfitable(SmallVectorImpl<LoadedSlice> &LoadedSlices,
8428 const APInt &UsedBits, bool ForCodeSize) {
8429 unsigned NumberOfSlices = LoadedSlices.size();
8430 if (StressLoadSlicing)
8431 return NumberOfSlices > 1;
8434 if (NumberOfSlices != 2)
8438 if (!areUsedBitsDense(UsedBits))
8442 LoadedSlice::Cost OrigCost(ForCodeSize), GlobalSlicingCost(ForCodeSize);
8443 // The original code has one big load.
8445 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice) {
8446 const LoadedSlice &LS = LoadedSlices[CurrSlice];
8447 // Accumulate the cost of all the slices.
8448 LoadedSlice::Cost SliceCost(LS, ForCodeSize);
8449 GlobalSlicingCost += SliceCost;
8451 // Account as cost in the original configuration the gain obtained
8452 // with the current slices.
8453 OrigCost.addSliceGain(LS);
8456 // If the target supports paired load, adjust the cost accordingly.
8457 adjustCostForPairing(LoadedSlices, GlobalSlicingCost);
8458 return OrigCost > GlobalSlicingCost;
8461 /// \brief If the given load, \p LI, is used only by trunc or trunc(lshr)
8462 /// operations, split it in the various pieces being extracted.
8464 /// This sort of thing is introduced by SROA.
8465 /// This slicing takes care not to insert overlapping loads.
8466 /// \pre LI is a simple load (i.e., not an atomic or volatile load).
8467 bool DAGCombiner::SliceUpLoad(SDNode *N) {
8468 if (Level < AfterLegalizeDAG)
8471 LoadSDNode *LD = cast<LoadSDNode>(N);
8472 if (LD->isVolatile() || !ISD::isNormalLoad(LD) ||
8473 !LD->getValueType(0).isInteger())
8476 // Keep track of already used bits to detect overlapping values.
8477 // In that case, we will just abort the transformation.
8478 APInt UsedBits(LD->getValueSizeInBits(0), 0);
8480 SmallVector<LoadedSlice, 4> LoadedSlices;
8482 // Check if this load is used as several smaller chunks of bits.
8483 // Basically, look for uses in trunc or trunc(lshr) and record a new chain
8484 // of computation for each trunc.
8485 for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end();
8486 UI != UIEnd; ++UI) {
8487 // Skip the uses of the chain.
8488 if (UI.getUse().getResNo() != 0)
8494 // Check if this is a trunc(lshr).
8495 if (User->getOpcode() == ISD::SRL && User->hasOneUse() &&
8496 isa<ConstantSDNode>(User->getOperand(1))) {
8497 Shift = cast<ConstantSDNode>(User->getOperand(1))->getZExtValue();
8498 User = *User->use_begin();
8501 // At this point, User is a Truncate, iff we encountered, trunc or
8503 if (User->getOpcode() != ISD::TRUNCATE)
8506 // The width of the type must be a power of 2 and greater than 8-bits.
8507 // Otherwise the load cannot be represented in LLVM IR.
8508 // Moreover, if we shifted with a non-8-bits multiple, the slice
8509 // will be across several bytes. We do not support that.
8510 unsigned Width = User->getValueSizeInBits(0);
8511 if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7))
8514 // Build the slice for this chain of computations.
8515 LoadedSlice LS(User, LD, Shift, &DAG);
8516 APInt CurrentUsedBits = LS.getUsedBits();
8518 // Check if this slice overlaps with another.
8519 if ((CurrentUsedBits & UsedBits) != 0)
8521 // Update the bits used globally.
8522 UsedBits |= CurrentUsedBits;
8524 // Check if the new slice would be legal.
8528 // Record the slice.
8529 LoadedSlices.push_back(LS);
8532 // Abort slicing if it does not seem to be profitable.
8533 if (!isSlicingProfitable(LoadedSlices, UsedBits, ForCodeSize))
8538 // Rewrite each chain to use an independent load.
8539 // By construction, each chain can be represented by a unique load.
8541 // Prepare the argument for the new token factor for all the slices.
8542 SmallVector<SDValue, 8> ArgChains;
8543 for (SmallVectorImpl<LoadedSlice>::const_iterator
8544 LSIt = LoadedSlices.begin(),
8545 LSItEnd = LoadedSlices.end();
8546 LSIt != LSItEnd; ++LSIt) {
8547 SDValue SliceInst = LSIt->loadSlice();
8548 CombineTo(LSIt->Inst, SliceInst, true);
8549 if (SliceInst.getNode()->getOpcode() != ISD::LOAD)
8550 SliceInst = SliceInst.getOperand(0);
8551 assert(SliceInst->getOpcode() == ISD::LOAD &&
8552 "It takes more than a zext to get to the loaded slice!!");
8553 ArgChains.push_back(SliceInst.getValue(1));
8556 SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other,
8557 &ArgChains[0], ArgChains.size());
8558 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
8562 /// CheckForMaskedLoad - Check to see if V is (and load (ptr), imm), where the
8563 /// load is having specific bytes cleared out. If so, return the byte size
8564 /// being masked out and the shift amount.
8565 static std::pair<unsigned, unsigned>
8566 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
8567 std::pair<unsigned, unsigned> Result(0, 0);
8569 // Check for the structure we're looking for.
8570 if (V->getOpcode() != ISD::AND ||
8571 !isa<ConstantSDNode>(V->getOperand(1)) ||
8572 !ISD::isNormalLoad(V->getOperand(0).getNode()))
8575 // Check the chain and pointer.
8576 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
8577 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
8579 // The store should be chained directly to the load or be an operand of a
8581 if (LD == Chain.getNode())
8583 else if (Chain->getOpcode() != ISD::TokenFactor)
8584 return Result; // Fail.
8587 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
8588 if (Chain->getOperand(i).getNode() == LD) {
8592 if (!isOk) return Result;
8595 // This only handles simple types.
8596 if (V.getValueType() != MVT::i16 &&
8597 V.getValueType() != MVT::i32 &&
8598 V.getValueType() != MVT::i64)
8601 // Check the constant mask. Invert it so that the bits being masked out are
8602 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
8603 // follow the sign bit for uniformity.
8604 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
8605 unsigned NotMaskLZ = countLeadingZeros(NotMask);
8606 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
8607 unsigned NotMaskTZ = countTrailingZeros(NotMask);
8608 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
8609 if (NotMaskLZ == 64) return Result; // All zero mask.
8611 // See if we have a continuous run of bits. If so, we have 0*1+0*
8612 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
8615 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
8616 if (V.getValueType() != MVT::i64 && NotMaskLZ)
8617 NotMaskLZ -= 64-V.getValueSizeInBits();
8619 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
8620 switch (MaskedBytes) {
8624 default: return Result; // All one mask, or 5-byte mask.
8627 // Verify that the first bit starts at a multiple of mask so that the access
8628 // is aligned the same as the access width.
8629 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
8631 Result.first = MaskedBytes;
8632 Result.second = NotMaskTZ/8;
8637 /// ShrinkLoadReplaceStoreWithStore - Check to see if IVal is something that
8638 /// provides a value as specified by MaskInfo. If so, replace the specified
8639 /// store with a narrower store of truncated IVal.
8641 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
8642 SDValue IVal, StoreSDNode *St,
8644 unsigned NumBytes = MaskInfo.first;
8645 unsigned ByteShift = MaskInfo.second;
8646 SelectionDAG &DAG = DC->getDAG();
8648 // Check to see if IVal is all zeros in the part being masked in by the 'or'
8649 // that uses this. If not, this is not a replacement.
8650 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
8651 ByteShift*8, (ByteShift+NumBytes)*8);
8652 if (!DAG.MaskedValueIsZero(IVal, Mask)) return 0;
8654 // Check that it is legal on the target to do this. It is legal if the new
8655 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
8657 MVT VT = MVT::getIntegerVT(NumBytes*8);
8658 if (!DC->isTypeLegal(VT))
8661 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
8662 // shifted by ByteShift and truncated down to NumBytes.
8664 IVal = DAG.getNode(ISD::SRL, SDLoc(IVal), IVal.getValueType(), IVal,
8665 DAG.getConstant(ByteShift*8,
8666 DC->getShiftAmountTy(IVal.getValueType())));
8668 // Figure out the offset for the store and the alignment of the access.
8670 unsigned NewAlign = St->getAlignment();
8672 if (DAG.getTargetLoweringInfo().isLittleEndian())
8673 StOffset = ByteShift;
8675 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
8677 SDValue Ptr = St->getBasePtr();
8679 Ptr = DAG.getNode(ISD::ADD, SDLoc(IVal), Ptr.getValueType(),
8680 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
8681 NewAlign = MinAlign(NewAlign, StOffset);
8684 // Truncate down to the new size.
8685 IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal);
8688 return DAG.getStore(St->getChain(), SDLoc(St), IVal, Ptr,
8689 St->getPointerInfo().getWithOffset(StOffset),
8690 false, false, NewAlign).getNode();
8694 /// ReduceLoadOpStoreWidth - Look for sequence of load / op / store where op is
8695 /// one of 'or', 'xor', and 'and' of immediates. If 'op' is only touching some
8696 /// of the loaded bits, try narrowing the load and store if it would end up
8697 /// being a win for performance or code size.
8698 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
8699 StoreSDNode *ST = cast<StoreSDNode>(N);
8700 if (ST->isVolatile())
8703 SDValue Chain = ST->getChain();
8704 SDValue Value = ST->getValue();
8705 SDValue Ptr = ST->getBasePtr();
8706 EVT VT = Value.getValueType();
8708 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
8711 unsigned Opc = Value.getOpcode();
8713 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
8714 // is a byte mask indicating a consecutive number of bytes, check to see if
8715 // Y is known to provide just those bytes. If so, we try to replace the
8716 // load + replace + store sequence with a single (narrower) store, which makes
8718 if (Opc == ISD::OR) {
8719 std::pair<unsigned, unsigned> MaskedLoad;
8720 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
8721 if (MaskedLoad.first)
8722 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
8723 Value.getOperand(1), ST,this))
8724 return SDValue(NewST, 0);
8726 // Or is commutative, so try swapping X and Y.
8727 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
8728 if (MaskedLoad.first)
8729 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
8730 Value.getOperand(0), ST,this))
8731 return SDValue(NewST, 0);
8734 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
8735 Value.getOperand(1).getOpcode() != ISD::Constant)
8738 SDValue N0 = Value.getOperand(0);
8739 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
8740 Chain == SDValue(N0.getNode(), 1)) {
8741 LoadSDNode *LD = cast<LoadSDNode>(N0);
8742 if (LD->getBasePtr() != Ptr ||
8743 LD->getPointerInfo().getAddrSpace() !=
8744 ST->getPointerInfo().getAddrSpace())
8747 // Find the type to narrow it the load / op / store to.
8748 SDValue N1 = Value.getOperand(1);
8749 unsigned BitWidth = N1.getValueSizeInBits();
8750 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
8751 if (Opc == ISD::AND)
8752 Imm ^= APInt::getAllOnesValue(BitWidth);
8753 if (Imm == 0 || Imm.isAllOnesValue())
8755 unsigned ShAmt = Imm.countTrailingZeros();
8756 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
8757 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
8758 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
8759 while (NewBW < BitWidth &&
8760 !(TLI.isOperationLegalOrCustom(Opc, NewVT) &&
8761 TLI.isNarrowingProfitable(VT, NewVT))) {
8762 NewBW = NextPowerOf2(NewBW);
8763 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
8765 if (NewBW >= BitWidth)
8768 // If the lsb changed does not start at the type bitwidth boundary,
8769 // start at the previous one.
8771 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
8772 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt,
8773 std::min(BitWidth, ShAmt + NewBW));
8774 if ((Imm & Mask) == Imm) {
8775 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
8776 if (Opc == ISD::AND)
8777 NewImm ^= APInt::getAllOnesValue(NewBW);
8778 uint64_t PtrOff = ShAmt / 8;
8779 // For big endian targets, we need to adjust the offset to the pointer to
8780 // load the correct bytes.
8781 if (TLI.isBigEndian())
8782 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
8784 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
8785 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
8786 if (NewAlign < TLI.getDataLayout()->getABITypeAlignment(NewVTTy))
8789 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD),
8790 Ptr.getValueType(), Ptr,
8791 DAG.getConstant(PtrOff, Ptr.getValueType()));
8792 SDValue NewLD = DAG.getLoad(NewVT, SDLoc(N0),
8793 LD->getChain(), NewPtr,
8794 LD->getPointerInfo().getWithOffset(PtrOff),
8795 LD->isVolatile(), LD->isNonTemporal(),
8796 LD->isInvariant(), NewAlign,
8798 SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD,
8799 DAG.getConstant(NewImm, NewVT));
8800 SDValue NewST = DAG.getStore(Chain, SDLoc(N),
8802 ST->getPointerInfo().getWithOffset(PtrOff),
8803 false, false, NewAlign);
8805 AddToWorkList(NewPtr.getNode());
8806 AddToWorkList(NewLD.getNode());
8807 AddToWorkList(NewVal.getNode());
8808 WorkListRemover DeadNodes(*this);
8809 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
8818 /// TransformFPLoadStorePair - For a given floating point load / store pair,
8819 /// if the load value isn't used by any other operations, then consider
8820 /// transforming the pair to integer load / store operations if the target
8821 /// deems the transformation profitable.
8822 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
8823 StoreSDNode *ST = cast<StoreSDNode>(N);
8824 SDValue Chain = ST->getChain();
8825 SDValue Value = ST->getValue();
8826 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
8827 Value.hasOneUse() &&
8828 Chain == SDValue(Value.getNode(), 1)) {
8829 LoadSDNode *LD = cast<LoadSDNode>(Value);
8830 EVT VT = LD->getMemoryVT();
8831 if (!VT.isFloatingPoint() ||
8832 VT != ST->getMemoryVT() ||
8833 LD->isNonTemporal() ||
8834 ST->isNonTemporal() ||
8835 LD->getPointerInfo().getAddrSpace() != 0 ||
8836 ST->getPointerInfo().getAddrSpace() != 0)
8839 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
8840 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
8841 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
8842 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
8843 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
8846 unsigned LDAlign = LD->getAlignment();
8847 unsigned STAlign = ST->getAlignment();
8848 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
8849 unsigned ABIAlign = TLI.getDataLayout()->getABITypeAlignment(IntVTTy);
8850 if (LDAlign < ABIAlign || STAlign < ABIAlign)
8853 SDValue NewLD = DAG.getLoad(IntVT, SDLoc(Value),
8854 LD->getChain(), LD->getBasePtr(),
8855 LD->getPointerInfo(),
8856 false, false, false, LDAlign);
8858 SDValue NewST = DAG.getStore(NewLD.getValue(1), SDLoc(N),
8859 NewLD, ST->getBasePtr(),
8860 ST->getPointerInfo(),
8861 false, false, STAlign);
8863 AddToWorkList(NewLD.getNode());
8864 AddToWorkList(NewST.getNode());
8865 WorkListRemover DeadNodes(*this);
8866 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
8874 /// Helper struct to parse and store a memory address as base + index + offset.
8875 /// We ignore sign extensions when it is safe to do so.
8876 /// The following two expressions are not equivalent. To differentiate we need
8877 /// to store whether there was a sign extension involved in the index
8879 /// (load (i64 add (i64 copyfromreg %c)
8880 /// (i64 signextend (add (i8 load %index)
8884 /// (load (i64 add (i64 copyfromreg %c)
8885 /// (i64 signextend (i32 add (i32 signextend (i8 load %index))
8887 struct BaseIndexOffset {
8891 bool IsIndexSignExt;
8893 BaseIndexOffset() : Offset(0), IsIndexSignExt(false) {}
8895 BaseIndexOffset(SDValue Base, SDValue Index, int64_t Offset,
8896 bool IsIndexSignExt) :
8897 Base(Base), Index(Index), Offset(Offset), IsIndexSignExt(IsIndexSignExt) {}
8899 bool equalBaseIndex(const BaseIndexOffset &Other) {
8900 return Other.Base == Base && Other.Index == Index &&
8901 Other.IsIndexSignExt == IsIndexSignExt;
8904 /// Parses tree in Ptr for base, index, offset addresses.
8905 static BaseIndexOffset match(SDValue Ptr) {
8906 bool IsIndexSignExt = false;
8908 // We only can pattern match BASE + INDEX + OFFSET. If Ptr is not an ADD
8909 // instruction, then it could be just the BASE or everything else we don't
8910 // know how to handle. Just use Ptr as BASE and give up.
8911 if (Ptr->getOpcode() != ISD::ADD)
8912 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
8914 // We know that we have at least an ADD instruction. Try to pattern match
8915 // the simple case of BASE + OFFSET.
8916 if (isa<ConstantSDNode>(Ptr->getOperand(1))) {
8917 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
8918 return BaseIndexOffset(Ptr->getOperand(0), SDValue(), Offset,
8922 // Inside a loop the current BASE pointer is calculated using an ADD and a
8923 // MUL instruction. In this case Ptr is the actual BASE pointer.
8924 // (i64 add (i64 %array_ptr)
8925 // (i64 mul (i64 %induction_var)
8926 // (i64 %element_size)))
8927 if (Ptr->getOperand(1)->getOpcode() == ISD::MUL)
8928 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
8930 // Look at Base + Index + Offset cases.
8931 SDValue Base = Ptr->getOperand(0);
8932 SDValue IndexOffset = Ptr->getOperand(1);
8934 // Skip signextends.
8935 if (IndexOffset->getOpcode() == ISD::SIGN_EXTEND) {
8936 IndexOffset = IndexOffset->getOperand(0);
8937 IsIndexSignExt = true;
8940 // Either the case of Base + Index (no offset) or something else.
8941 if (IndexOffset->getOpcode() != ISD::ADD)
8942 return BaseIndexOffset(Base, IndexOffset, 0, IsIndexSignExt);
8944 // Now we have the case of Base + Index + offset.
8945 SDValue Index = IndexOffset->getOperand(0);
8946 SDValue Offset = IndexOffset->getOperand(1);
8948 if (!isa<ConstantSDNode>(Offset))
8949 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
8951 // Ignore signextends.
8952 if (Index->getOpcode() == ISD::SIGN_EXTEND) {
8953 Index = Index->getOperand(0);
8954 IsIndexSignExt = true;
8955 } else IsIndexSignExt = false;
8957 int64_t Off = cast<ConstantSDNode>(Offset)->getSExtValue();
8958 return BaseIndexOffset(Base, Index, Off, IsIndexSignExt);
8962 /// Holds a pointer to an LSBaseSDNode as well as information on where it
8963 /// is located in a sequence of memory operations connected by a chain.
8965 MemOpLink (LSBaseSDNode *N, int64_t Offset, unsigned Seq):
8966 MemNode(N), OffsetFromBase(Offset), SequenceNum(Seq) { }
8967 // Ptr to the mem node.
8968 LSBaseSDNode *MemNode;
8969 // Offset from the base ptr.
8970 int64_t OffsetFromBase;
8971 // What is the sequence number of this mem node.
8972 // Lowest mem operand in the DAG starts at zero.
8973 unsigned SequenceNum;
8976 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
8977 EVT MemVT = St->getMemoryVT();
8978 int64_t ElementSizeBytes = MemVT.getSizeInBits()/8;
8979 bool NoVectors = DAG.getMachineFunction().getFunction()->getAttributes().
8980 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
8982 // Don't merge vectors into wider inputs.
8983 if (MemVT.isVector() || !MemVT.isSimple())
8986 // Perform an early exit check. Do not bother looking at stored values that
8987 // are not constants or loads.
8988 SDValue StoredVal = St->getValue();
8989 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
8990 if (!isa<ConstantSDNode>(StoredVal) && !isa<ConstantFPSDNode>(StoredVal) &&
8994 // Only look at ends of store sequences.
8995 SDValue Chain = SDValue(St, 1);
8996 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
8999 // This holds the base pointer, index, and the offset in bytes from the base
9001 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr());
9003 // We must have a base and an offset.
9004 if (!BasePtr.Base.getNode())
9007 // Do not handle stores to undef base pointers.
9008 if (BasePtr.Base.getOpcode() == ISD::UNDEF)
9011 // Save the LoadSDNodes that we find in the chain.
9012 // We need to make sure that these nodes do not interfere with
9013 // any of the store nodes.
9014 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
9016 // Save the StoreSDNodes that we find in the chain.
9017 SmallVector<MemOpLink, 8> StoreNodes;
9019 // Walk up the chain and look for nodes with offsets from the same
9020 // base pointer. Stop when reaching an instruction with a different kind
9021 // or instruction which has a different base pointer.
9023 StoreSDNode *Index = St;
9025 // If the chain has more than one use, then we can't reorder the mem ops.
9026 if (Index != St && !SDValue(Index, 1)->hasOneUse())
9029 // Find the base pointer and offset for this memory node.
9030 BaseIndexOffset Ptr = BaseIndexOffset::match(Index->getBasePtr());
9032 // Check that the base pointer is the same as the original one.
9033 if (!Ptr.equalBaseIndex(BasePtr))
9036 // Check that the alignment is the same.
9037 if (Index->getAlignment() != St->getAlignment())
9040 // The memory operands must not be volatile.
9041 if (Index->isVolatile() || Index->isIndexed())
9045 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
9046 if (St->isTruncatingStore())
9049 // The stored memory type must be the same.
9050 if (Index->getMemoryVT() != MemVT)
9053 // We do not allow unaligned stores because we want to prevent overriding
9055 if (Index->getAlignment()*8 != MemVT.getSizeInBits())
9058 // We found a potential memory operand to merge.
9059 StoreNodes.push_back(MemOpLink(Index, Ptr.Offset, Seq++));
9061 // Find the next memory operand in the chain. If the next operand in the
9062 // chain is a store then move up and continue the scan with the next
9063 // memory operand. If the next operand is a load save it and use alias
9064 // information to check if it interferes with anything.
9065 SDNode *NextInChain = Index->getChain().getNode();
9067 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
9068 // We found a store node. Use it for the next iteration.
9071 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
9072 if (Ldn->isVolatile()) {
9077 // Save the load node for later. Continue the scan.
9078 AliasLoadNodes.push_back(Ldn);
9079 NextInChain = Ldn->getChain().getNode();
9088 // Check if there is anything to merge.
9089 if (StoreNodes.size() < 2)
9092 // Sort the memory operands according to their distance from the base pointer.
9093 std::sort(StoreNodes.begin(), StoreNodes.end(),
9094 [](MemOpLink LHS, MemOpLink RHS) {
9095 return LHS.OffsetFromBase < RHS.OffsetFromBase ||
9096 (LHS.OffsetFromBase == RHS.OffsetFromBase &&
9097 LHS.SequenceNum > RHS.SequenceNum);
9100 // Scan the memory operations on the chain and find the first non-consecutive
9101 // store memory address.
9102 unsigned LastConsecutiveStore = 0;
9103 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
9104 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) {
9106 // Check that the addresses are consecutive starting from the second
9107 // element in the list of stores.
9109 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
9110 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
9115 // Check if this store interferes with any of the loads that we found.
9116 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld)
9117 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) {
9121 // We found a load that alias with this store. Stop the sequence.
9125 // Mark this node as useful.
9126 LastConsecutiveStore = i;
9129 // The node with the lowest store address.
9130 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
9132 // Store the constants into memory as one consecutive store.
9134 unsigned LastLegalType = 0;
9135 unsigned LastLegalVectorType = 0;
9136 bool NonZero = false;
9137 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
9138 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9139 SDValue StoredVal = St->getValue();
9141 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
9142 NonZero |= !C->isNullValue();
9143 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) {
9144 NonZero |= !C->getConstantFPValue()->isNullValue();
9150 // Find a legal type for the constant store.
9151 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
9152 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9153 if (TLI.isTypeLegal(StoreTy))
9154 LastLegalType = i+1;
9155 // Or check whether a truncstore is legal.
9156 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
9157 TargetLowering::TypePromoteInteger) {
9158 EVT LegalizedStoredValueTy =
9159 TLI.getTypeToTransformTo(*DAG.getContext(), StoredVal.getValueType());
9160 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy))
9161 LastLegalType = i+1;
9164 // Find a legal type for the vector store.
9165 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
9166 if (TLI.isTypeLegal(Ty))
9167 LastLegalVectorType = i + 1;
9170 // We only use vectors if the constant is known to be zero and the
9171 // function is not marked with the noimplicitfloat attribute.
9172 if (NonZero || NoVectors)
9173 LastLegalVectorType = 0;
9175 // Check if we found a legal integer type to store.
9176 if (LastLegalType == 0 && LastLegalVectorType == 0)
9179 bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors;
9180 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType;
9182 // Make sure we have something to merge.
9186 unsigned EarliestNodeUsed = 0;
9187 for (unsigned i=0; i < NumElem; ++i) {
9188 // Find a chain for the new wide-store operand. Notice that some
9189 // of the store nodes that we found may not be selected for inclusion
9190 // in the wide store. The chain we use needs to be the chain of the
9191 // earliest store node which is *used* and replaced by the wide store.
9192 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
9193 EarliestNodeUsed = i;
9196 // The earliest Node in the DAG.
9197 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
9198 SDLoc DL(StoreNodes[0].MemNode);
9202 // Find a legal type for the vector store.
9203 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
9204 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
9205 StoredVal = DAG.getConstant(0, Ty);
9207 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
9208 APInt StoreInt(StoreBW, 0);
9210 // Construct a single integer constant which is made of the smaller
9212 bool IsLE = TLI.isLittleEndian();
9213 for (unsigned i = 0; i < NumElem ; ++i) {
9214 unsigned Idx = IsLE ?(NumElem - 1 - i) : i;
9215 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
9216 SDValue Val = St->getValue();
9217 StoreInt<<=ElementSizeBytes*8;
9218 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
9219 StoreInt|=C->getAPIntValue().zext(StoreBW);
9220 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
9221 StoreInt|= C->getValueAPF().bitcastToAPInt().zext(StoreBW);
9223 assert(false && "Invalid constant element type");
9227 // Create the new Load and Store operations.
9228 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9229 StoredVal = DAG.getConstant(StoreInt, StoreTy);
9232 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), DL, StoredVal,
9233 FirstInChain->getBasePtr(),
9234 FirstInChain->getPointerInfo(),
9236 FirstInChain->getAlignment());
9238 // Replace the first store with the new store
9239 CombineTo(EarliestOp, NewStore);
9240 // Erase all other stores.
9241 for (unsigned i = 0; i < NumElem ; ++i) {
9242 if (StoreNodes[i].MemNode == EarliestOp)
9244 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9245 // ReplaceAllUsesWith will replace all uses that existed when it was
9246 // called, but graph optimizations may cause new ones to appear. For
9247 // example, the case in pr14333 looks like
9249 // St's chain -> St -> another store -> X
9251 // And the only difference from St to the other store is the chain.
9252 // When we change it's chain to be St's chain they become identical,
9253 // get CSEed and the net result is that X is now a use of St.
9254 // Since we know that St is redundant, just iterate.
9255 while (!St->use_empty())
9256 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain());
9257 removeFromWorkList(St);
9264 // Below we handle the case of multiple consecutive stores that
9265 // come from multiple consecutive loads. We merge them into a single
9266 // wide load and a single wide store.
9268 // Look for load nodes which are used by the stored values.
9269 SmallVector<MemOpLink, 8> LoadNodes;
9271 // Find acceptable loads. Loads need to have the same chain (token factor),
9272 // must not be zext, volatile, indexed, and they must be consecutive.
9273 BaseIndexOffset LdBasePtr;
9274 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
9275 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9276 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
9279 // Loads must only have one use.
9280 if (!Ld->hasNUsesOfValue(1, 0))
9283 // Check that the alignment is the same as the stores.
9284 if (Ld->getAlignment() != St->getAlignment())
9287 // The memory operands must not be volatile.
9288 if (Ld->isVolatile() || Ld->isIndexed())
9291 // We do not accept ext loads.
9292 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
9295 // The stored memory type must be the same.
9296 if (Ld->getMemoryVT() != MemVT)
9299 BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld->getBasePtr());
9300 // If this is not the first ptr that we check.
9301 if (LdBasePtr.Base.getNode()) {
9302 // The base ptr must be the same.
9303 if (!LdPtr.equalBaseIndex(LdBasePtr))
9306 // Check that all other base pointers are the same as this one.
9310 // We found a potential memory operand to merge.
9311 LoadNodes.push_back(MemOpLink(Ld, LdPtr.Offset, 0));
9314 if (LoadNodes.size() < 2)
9317 // Scan the memory operations on the chain and find the first non-consecutive
9318 // load memory address. These variables hold the index in the store node
9320 unsigned LastConsecutiveLoad = 0;
9321 // This variable refers to the size and not index in the array.
9322 unsigned LastLegalVectorType = 0;
9323 unsigned LastLegalIntegerType = 0;
9324 StartAddress = LoadNodes[0].OffsetFromBase;
9325 SDValue FirstChain = LoadNodes[0].MemNode->getChain();
9326 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
9327 // All loads much share the same chain.
9328 if (LoadNodes[i].MemNode->getChain() != FirstChain)
9331 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
9332 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
9334 LastConsecutiveLoad = i;
9336 // Find a legal type for the vector store.
9337 EVT StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
9338 if (TLI.isTypeLegal(StoreTy))
9339 LastLegalVectorType = i + 1;
9341 // Find a legal type for the integer store.
9342 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
9343 StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9344 if (TLI.isTypeLegal(StoreTy))
9345 LastLegalIntegerType = i + 1;
9346 // Or check whether a truncstore and extload is legal.
9347 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
9348 TargetLowering::TypePromoteInteger) {
9349 EVT LegalizedStoredValueTy =
9350 TLI.getTypeToTransformTo(*DAG.getContext(), StoreTy);
9351 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
9352 TLI.isLoadExtLegal(ISD::ZEXTLOAD, StoreTy) &&
9353 TLI.isLoadExtLegal(ISD::SEXTLOAD, StoreTy) &&
9354 TLI.isLoadExtLegal(ISD::EXTLOAD, StoreTy))
9355 LastLegalIntegerType = i+1;
9359 // Only use vector types if the vector type is larger than the integer type.
9360 // If they are the same, use integers.
9361 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors;
9362 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType);
9364 // We add +1 here because the LastXXX variables refer to location while
9365 // the NumElem refers to array/index size.
9366 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1;
9367 NumElem = std::min(LastLegalType, NumElem);
9372 // The earliest Node in the DAG.
9373 unsigned EarliestNodeUsed = 0;
9374 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
9375 for (unsigned i=1; i<NumElem; ++i) {
9376 // Find a chain for the new wide-store operand. Notice that some
9377 // of the store nodes that we found may not be selected for inclusion
9378 // in the wide store. The chain we use needs to be the chain of the
9379 // earliest store node which is *used* and replaced by the wide store.
9380 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
9381 EarliestNodeUsed = i;
9384 // Find if it is better to use vectors or integers to load and store
9388 JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
9390 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
9391 JointMemOpVT = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9394 SDLoc LoadDL(LoadNodes[0].MemNode);
9395 SDLoc StoreDL(StoreNodes[0].MemNode);
9397 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
9398 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL,
9399 FirstLoad->getChain(),
9400 FirstLoad->getBasePtr(),
9401 FirstLoad->getPointerInfo(),
9402 false, false, false,
9403 FirstLoad->getAlignment());
9405 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), StoreDL, NewLoad,
9406 FirstInChain->getBasePtr(),
9407 FirstInChain->getPointerInfo(), false, false,
9408 FirstInChain->getAlignment());
9410 // Replace one of the loads with the new load.
9411 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode);
9412 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
9413 SDValue(NewLoad.getNode(), 1));
9415 // Remove the rest of the load chains.
9416 for (unsigned i = 1; i < NumElem ; ++i) {
9417 // Replace all chain users of the old load nodes with the chain of the new
9419 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
9420 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain());
9423 // Replace the first store with the new store.
9424 CombineTo(EarliestOp, NewStore);
9425 // Erase all other stores.
9426 for (unsigned i = 0; i < NumElem ; ++i) {
9427 // Remove all Store nodes.
9428 if (StoreNodes[i].MemNode == EarliestOp)
9430 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9431 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain());
9432 removeFromWorkList(St);
9439 SDValue DAGCombiner::visitSTORE(SDNode *N) {
9440 StoreSDNode *ST = cast<StoreSDNode>(N);
9441 SDValue Chain = ST->getChain();
9442 SDValue Value = ST->getValue();
9443 SDValue Ptr = ST->getBasePtr();
9445 // If this is a store of a bit convert, store the input value if the
9446 // resultant store does not need a higher alignment than the original.
9447 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
9448 ST->isUnindexed()) {
9449 unsigned OrigAlign = ST->getAlignment();
9450 EVT SVT = Value.getOperand(0).getValueType();
9451 unsigned Align = TLI.getDataLayout()->
9452 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
9453 if (Align <= OrigAlign &&
9454 ((!LegalOperations && !ST->isVolatile()) ||
9455 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
9456 return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0),
9457 Ptr, ST->getPointerInfo(), ST->isVolatile(),
9458 ST->isNonTemporal(), OrigAlign,
9462 // Turn 'store undef, Ptr' -> nothing.
9463 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
9466 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
9467 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
9468 // NOTE: If the original store is volatile, this transform must not increase
9469 // the number of stores. For example, on x86-32 an f64 can be stored in one
9470 // processor operation but an i64 (which is not legal) requires two. So the
9471 // transform should not be done in this case.
9472 if (Value.getOpcode() != ISD::TargetConstantFP) {
9474 switch (CFP->getSimpleValueType(0).SimpleTy) {
9475 default: llvm_unreachable("Unknown FP type");
9476 case MVT::f16: // We don't do this for these yet.
9482 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
9483 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
9484 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
9485 bitcastToAPInt().getZExtValue(), MVT::i32);
9486 return DAG.getStore(Chain, SDLoc(N), Tmp,
9487 Ptr, ST->getMemOperand());
9491 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
9492 !ST->isVolatile()) ||
9493 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
9494 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
9495 getZExtValue(), MVT::i64);
9496 return DAG.getStore(Chain, SDLoc(N), Tmp,
9497 Ptr, ST->getMemOperand());
9500 if (!ST->isVolatile() &&
9501 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
9502 // Many FP stores are not made apparent until after legalize, e.g. for
9503 // argument passing. Since this is so common, custom legalize the
9504 // 64-bit integer store into two 32-bit stores.
9505 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
9506 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
9507 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
9508 if (TLI.isBigEndian()) std::swap(Lo, Hi);
9510 unsigned Alignment = ST->getAlignment();
9511 bool isVolatile = ST->isVolatile();
9512 bool isNonTemporal = ST->isNonTemporal();
9513 const MDNode *TBAAInfo = ST->getTBAAInfo();
9515 SDValue St0 = DAG.getStore(Chain, SDLoc(ST), Lo,
9516 Ptr, ST->getPointerInfo(),
9517 isVolatile, isNonTemporal,
9518 ST->getAlignment(), TBAAInfo);
9519 Ptr = DAG.getNode(ISD::ADD, SDLoc(N), Ptr.getValueType(), Ptr,
9520 DAG.getConstant(4, Ptr.getValueType()));
9521 Alignment = MinAlign(Alignment, 4U);
9522 SDValue St1 = DAG.getStore(Chain, SDLoc(ST), Hi,
9523 Ptr, ST->getPointerInfo().getWithOffset(4),
9524 isVolatile, isNonTemporal,
9525 Alignment, TBAAInfo);
9526 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other,
9535 // Try to infer better alignment information than the store already has.
9536 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
9537 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
9538 if (Align > ST->getAlignment())
9539 return DAG.getTruncStore(Chain, SDLoc(N), Value,
9540 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
9541 ST->isVolatile(), ST->isNonTemporal(), Align,
9546 // Try transforming a pair floating point load / store ops to integer
9547 // load / store ops.
9548 SDValue NewST = TransformFPLoadStorePair(N);
9549 if (NewST.getNode())
9552 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA :
9553 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
9555 if (CombinerAAOnlyFunc.getNumOccurrences() &&
9556 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
9559 if (UseAA && ST->isUnindexed()) {
9560 // Walk up chain skipping non-aliasing memory nodes.
9561 SDValue BetterChain = FindBetterChain(N, Chain);
9563 // If there is a better chain.
9564 if (Chain != BetterChain) {
9567 // Replace the chain to avoid dependency.
9568 if (ST->isTruncatingStore()) {
9569 ReplStore = DAG.getTruncStore(BetterChain, SDLoc(N), Value, Ptr,
9570 ST->getMemoryVT(), ST->getMemOperand());
9572 ReplStore = DAG.getStore(BetterChain, SDLoc(N), Value, Ptr,
9573 ST->getMemOperand());
9576 // Create token to keep both nodes around.
9577 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
9578 MVT::Other, Chain, ReplStore);
9580 // Make sure the new and old chains are cleaned up.
9581 AddToWorkList(Token.getNode());
9583 // Don't add users to work list.
9584 return CombineTo(N, Token, false);
9588 // Try transforming N to an indexed store.
9589 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
9590 return SDValue(N, 0);
9592 // FIXME: is there such a thing as a truncating indexed store?
9593 if (ST->isTruncatingStore() && ST->isUnindexed() &&
9594 Value.getValueType().isInteger()) {
9595 // See if we can simplify the input to this truncstore with knowledge that
9596 // only the low bits are being used. For example:
9597 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
9599 GetDemandedBits(Value,
9600 APInt::getLowBitsSet(
9601 Value.getValueType().getScalarType().getSizeInBits(),
9602 ST->getMemoryVT().getScalarType().getSizeInBits()));
9603 AddToWorkList(Value.getNode());
9604 if (Shorter.getNode())
9605 return DAG.getTruncStore(Chain, SDLoc(N), Shorter,
9606 Ptr, ST->getMemoryVT(), ST->getMemOperand());
9608 // Otherwise, see if we can simplify the operation with
9609 // SimplifyDemandedBits, which only works if the value has a single use.
9610 if (SimplifyDemandedBits(Value,
9611 APInt::getLowBitsSet(
9612 Value.getValueType().getScalarType().getSizeInBits(),
9613 ST->getMemoryVT().getScalarType().getSizeInBits())))
9614 return SDValue(N, 0);
9617 // If this is a load followed by a store to the same location, then the store
9619 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
9620 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
9621 ST->isUnindexed() && !ST->isVolatile() &&
9622 // There can't be any side effects between the load and store, such as
9624 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
9625 // The store is dead, remove it.
9630 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
9631 // truncating store. We can do this even if this is already a truncstore.
9632 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
9633 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
9634 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
9635 ST->getMemoryVT())) {
9636 return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0),
9637 Ptr, ST->getMemoryVT(), ST->getMemOperand());
9640 // Only perform this optimization before the types are legal, because we
9641 // don't want to perform this optimization on every DAGCombine invocation.
9643 bool EverChanged = false;
9646 // There can be multiple store sequences on the same chain.
9647 // Keep trying to merge store sequences until we are unable to do so
9648 // or until we merge the last store on the chain.
9649 bool Changed = MergeConsecutiveStores(ST);
9650 EverChanged |= Changed;
9651 if (!Changed) break;
9652 } while (ST->getOpcode() != ISD::DELETED_NODE);
9655 return SDValue(N, 0);
9658 return ReduceLoadOpStoreWidth(N);
9661 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
9662 SDValue InVec = N->getOperand(0);
9663 SDValue InVal = N->getOperand(1);
9664 SDValue EltNo = N->getOperand(2);
9667 // If the inserted element is an UNDEF, just use the input vector.
9668 if (InVal.getOpcode() == ISD::UNDEF)
9671 EVT VT = InVec.getValueType();
9673 // If we can't generate a legal BUILD_VECTOR, exit
9674 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
9677 // Check that we know which element is being inserted
9678 if (!isa<ConstantSDNode>(EltNo))
9680 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9682 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
9683 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
9685 SmallVector<SDValue, 8> Ops;
9686 // Do not combine these two vectors if the output vector will not replace
9687 // the input vector.
9688 if (InVec.getOpcode() == ISD::BUILD_VECTOR && InVec.hasOneUse()) {
9689 Ops.append(InVec.getNode()->op_begin(),
9690 InVec.getNode()->op_end());
9691 } else if (InVec.getOpcode() == ISD::UNDEF) {
9692 unsigned NElts = VT.getVectorNumElements();
9693 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
9698 // Insert the element
9699 if (Elt < Ops.size()) {
9700 // All the operands of BUILD_VECTOR must have the same type;
9701 // we enforce that here.
9702 EVT OpVT = Ops[0].getValueType();
9703 if (InVal.getValueType() != OpVT)
9704 InVal = OpVT.bitsGT(InVal.getValueType()) ?
9705 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
9706 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
9710 // Return the new vector
9711 return DAG.getNode(ISD::BUILD_VECTOR, dl,
9712 VT, &Ops[0], Ops.size());
9715 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
9716 // (vextract (scalar_to_vector val, 0) -> val
9717 SDValue InVec = N->getOperand(0);
9718 EVT VT = InVec.getValueType();
9719 EVT NVT = N->getValueType(0);
9721 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
9722 // Check if the result type doesn't match the inserted element type. A
9723 // SCALAR_TO_VECTOR may truncate the inserted element and the
9724 // EXTRACT_VECTOR_ELT may widen the extracted vector.
9725 SDValue InOp = InVec.getOperand(0);
9726 if (InOp.getValueType() != NVT) {
9727 assert(InOp.getValueType().isInteger() && NVT.isInteger());
9728 return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT);
9733 SDValue EltNo = N->getOperand(1);
9734 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
9736 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
9737 // We only perform this optimization before the op legalization phase because
9738 // we may introduce new vector instructions which are not backed by TD
9739 // patterns. For example on AVX, extracting elements from a wide vector
9740 // without using extract_subvector. However, if we can find an underlying
9741 // scalar value, then we can always use that.
9742 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
9744 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9745 int NumElem = VT.getVectorNumElements();
9746 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
9747 // Find the new index to extract from.
9748 int OrigElt = SVOp->getMaskElt(Elt);
9750 // Extracting an undef index is undef.
9752 return DAG.getUNDEF(NVT);
9754 // Select the right vector half to extract from.
9756 if (OrigElt < NumElem) {
9757 SVInVec = InVec->getOperand(0);
9759 SVInVec = InVec->getOperand(1);
9763 if (SVInVec.getOpcode() == ISD::BUILD_VECTOR) {
9764 SDValue InOp = SVInVec.getOperand(OrigElt);
9765 if (InOp.getValueType() != NVT) {
9766 assert(InOp.getValueType().isInteger() && NVT.isInteger());
9767 InOp = DAG.getSExtOrTrunc(InOp, SDLoc(SVInVec), NVT);
9773 // FIXME: We should handle recursing on other vector shuffles and
9774 // scalar_to_vector here as well.
9776 if (!LegalOperations) {
9777 EVT IndexTy = TLI.getVectorIdxTy();
9778 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT,
9779 SVInVec, DAG.getConstant(OrigElt, IndexTy));
9783 // Perform only after legalization to ensure build_vector / vector_shuffle
9784 // optimizations have already been done.
9785 if (!LegalOperations) return SDValue();
9787 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
9788 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
9789 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
9792 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9793 bool NewLoad = false;
9794 bool BCNumEltsChanged = false;
9795 EVT ExtVT = VT.getVectorElementType();
9798 // If the result of load has to be truncated, then it's not necessarily
9800 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
9803 if (InVec.getOpcode() == ISD::BITCAST) {
9804 // Don't duplicate a load with other uses.
9805 if (!InVec.hasOneUse())
9808 EVT BCVT = InVec.getOperand(0).getValueType();
9809 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
9811 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
9812 BCNumEltsChanged = true;
9813 InVec = InVec.getOperand(0);
9814 ExtVT = BCVT.getVectorElementType();
9818 LoadSDNode *LN0 = NULL;
9819 const ShuffleVectorSDNode *SVN = NULL;
9820 if (ISD::isNormalLoad(InVec.getNode())) {
9821 LN0 = cast<LoadSDNode>(InVec);
9822 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
9823 InVec.getOperand(0).getValueType() == ExtVT &&
9824 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
9825 // Don't duplicate a load with other uses.
9826 if (!InVec.hasOneUse())
9829 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
9830 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
9831 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
9833 // (load $addr+1*size)
9835 // Don't duplicate a load with other uses.
9836 if (!InVec.hasOneUse())
9839 // If the bit convert changed the number of elements, it is unsafe
9840 // to examine the mask.
9841 if (BCNumEltsChanged)
9844 // Select the input vector, guarding against out of range extract vector.
9845 unsigned NumElems = VT.getVectorNumElements();
9846 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
9847 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
9849 if (InVec.getOpcode() == ISD::BITCAST) {
9850 // Don't duplicate a load with other uses.
9851 if (!InVec.hasOneUse())
9854 InVec = InVec.getOperand(0);
9856 if (ISD::isNormalLoad(InVec.getNode())) {
9857 LN0 = cast<LoadSDNode>(InVec);
9858 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
9862 // Make sure we found a non-volatile load and the extractelement is
9864 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
9867 // If Idx was -1 above, Elt is going to be -1, so just return undef.
9869 return DAG.getUNDEF(LVT);
9871 unsigned Align = LN0->getAlignment();
9873 // Check the resultant load doesn't need a higher alignment than the
9877 ->getABITypeAlignment(LVT.getTypeForEVT(*DAG.getContext()));
9879 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, LVT))
9885 SDValue NewPtr = LN0->getBasePtr();
9886 unsigned PtrOff = 0;
9889 PtrOff = LVT.getSizeInBits() * Elt / 8;
9890 EVT PtrType = NewPtr.getValueType();
9891 if (TLI.isBigEndian())
9892 PtrOff = VT.getSizeInBits() / 8 - PtrOff;
9893 NewPtr = DAG.getNode(ISD::ADD, SDLoc(N), PtrType, NewPtr,
9894 DAG.getConstant(PtrOff, PtrType));
9897 // The replacement we need to do here is a little tricky: we need to
9898 // replace an extractelement of a load with a load.
9899 // Use ReplaceAllUsesOfValuesWith to do the replacement.
9900 // Note that this replacement assumes that the extractvalue is the only
9901 // use of the load; that's okay because we don't want to perform this
9902 // transformation in other cases anyway.
9905 if (NVT.bitsGT(LVT)) {
9906 // If the result type of vextract is wider than the load, then issue an
9907 // extending load instead.
9908 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, LVT)
9909 ? ISD::ZEXTLOAD : ISD::EXTLOAD;
9910 Load = DAG.getExtLoad(ExtType, SDLoc(N), NVT, LN0->getChain(),
9911 NewPtr, LN0->getPointerInfo().getWithOffset(PtrOff),
9912 LVT, LN0->isVolatile(), LN0->isNonTemporal(),
9913 Align, LN0->getTBAAInfo());
9914 Chain = Load.getValue(1);
9916 Load = DAG.getLoad(LVT, SDLoc(N), LN0->getChain(), NewPtr,
9917 LN0->getPointerInfo().getWithOffset(PtrOff),
9918 LN0->isVolatile(), LN0->isNonTemporal(),
9919 LN0->isInvariant(), Align, LN0->getTBAAInfo());
9920 Chain = Load.getValue(1);
9921 if (NVT.bitsLT(LVT))
9922 Load = DAG.getNode(ISD::TRUNCATE, SDLoc(N), NVT, Load);
9924 Load = DAG.getNode(ISD::BITCAST, SDLoc(N), NVT, Load);
9926 WorkListRemover DeadNodes(*this);
9927 SDValue From[] = { SDValue(N, 0), SDValue(LN0,1) };
9928 SDValue To[] = { Load, Chain };
9929 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
9930 // Since we're explcitly calling ReplaceAllUses, add the new node to the
9931 // worklist explicitly as well.
9932 AddToWorkList(Load.getNode());
9933 AddUsersToWorkList(Load.getNode()); // Add users too
9934 // Make sure to revisit this node to clean it up; it will usually be dead.
9936 return SDValue(N, 0);
9942 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
9943 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
9944 // We perform this optimization post type-legalization because
9945 // the type-legalizer often scalarizes integer-promoted vectors.
9946 // Performing this optimization before may create bit-casts which
9947 // will be type-legalized to complex code sequences.
9948 // We perform this optimization only before the operation legalizer because we
9949 // may introduce illegal operations.
9950 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
9953 unsigned NumInScalars = N->getNumOperands();
9955 EVT VT = N->getValueType(0);
9957 // Check to see if this is a BUILD_VECTOR of a bunch of values
9958 // which come from any_extend or zero_extend nodes. If so, we can create
9959 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
9960 // optimizations. We do not handle sign-extend because we can't fill the sign
9962 EVT SourceType = MVT::Other;
9963 bool AllAnyExt = true;
9965 for (unsigned i = 0; i != NumInScalars; ++i) {
9966 SDValue In = N->getOperand(i);
9967 // Ignore undef inputs.
9968 if (In.getOpcode() == ISD::UNDEF) continue;
9970 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
9971 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
9973 // Abort if the element is not an extension.
9974 if (!ZeroExt && !AnyExt) {
9975 SourceType = MVT::Other;
9979 // The input is a ZeroExt or AnyExt. Check the original type.
9980 EVT InTy = In.getOperand(0).getValueType();
9982 // Check that all of the widened source types are the same.
9983 if (SourceType == MVT::Other)
9986 else if (InTy != SourceType) {
9987 // Multiple income types. Abort.
9988 SourceType = MVT::Other;
9992 // Check if all of the extends are ANY_EXTENDs.
9993 AllAnyExt &= AnyExt;
9996 // In order to have valid types, all of the inputs must be extended from the
9997 // same source type and all of the inputs must be any or zero extend.
9998 // Scalar sizes must be a power of two.
9999 EVT OutScalarTy = VT.getScalarType();
10000 bool ValidTypes = SourceType != MVT::Other &&
10001 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
10002 isPowerOf2_32(SourceType.getSizeInBits());
10004 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
10005 // turn into a single shuffle instruction.
10009 bool isLE = TLI.isLittleEndian();
10010 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
10011 assert(ElemRatio > 1 && "Invalid element size ratio");
10012 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
10013 DAG.getConstant(0, SourceType);
10015 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
10016 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
10018 // Populate the new build_vector
10019 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
10020 SDValue Cast = N->getOperand(i);
10021 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
10022 Cast.getOpcode() == ISD::ZERO_EXTEND ||
10023 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
10025 if (Cast.getOpcode() == ISD::UNDEF)
10026 In = DAG.getUNDEF(SourceType);
10028 In = Cast->getOperand(0);
10029 unsigned Index = isLE ? (i * ElemRatio) :
10030 (i * ElemRatio + (ElemRatio - 1));
10032 assert(Index < Ops.size() && "Invalid index");
10036 // The type of the new BUILD_VECTOR node.
10037 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
10038 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
10039 "Invalid vector size");
10040 // Check if the new vector type is legal.
10041 if (!isTypeLegal(VecVT)) return SDValue();
10043 // Make the new BUILD_VECTOR.
10044 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], Ops.size());
10046 // The new BUILD_VECTOR node has the potential to be further optimized.
10047 AddToWorkList(BV.getNode());
10048 // Bitcast to the desired type.
10049 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
10052 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
10053 EVT VT = N->getValueType(0);
10055 unsigned NumInScalars = N->getNumOperands();
10058 EVT SrcVT = MVT::Other;
10059 unsigned Opcode = ISD::DELETED_NODE;
10060 unsigned NumDefs = 0;
10062 for (unsigned i = 0; i != NumInScalars; ++i) {
10063 SDValue In = N->getOperand(i);
10064 unsigned Opc = In.getOpcode();
10066 if (Opc == ISD::UNDEF)
10069 // If all scalar values are floats and converted from integers.
10070 if (Opcode == ISD::DELETED_NODE &&
10071 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
10078 EVT InVT = In.getOperand(0).getValueType();
10080 // If all scalar values are typed differently, bail out. It's chosen to
10081 // simplify BUILD_VECTOR of integer types.
10082 if (SrcVT == MVT::Other)
10089 // If the vector has just one element defined, it's not worth to fold it into
10090 // a vectorized one.
10094 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
10095 && "Should only handle conversion from integer to float.");
10096 assert(SrcVT != MVT::Other && "Cannot determine source type!");
10098 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
10100 if (!TLI.isOperationLegalOrCustom(Opcode, NVT))
10103 SmallVector<SDValue, 8> Opnds;
10104 for (unsigned i = 0; i != NumInScalars; ++i) {
10105 SDValue In = N->getOperand(i);
10107 if (In.getOpcode() == ISD::UNDEF)
10108 Opnds.push_back(DAG.getUNDEF(SrcVT));
10110 Opnds.push_back(In.getOperand(0));
10112 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT,
10113 &Opnds[0], Opnds.size());
10114 AddToWorkList(BV.getNode());
10116 return DAG.getNode(Opcode, dl, VT, BV);
10119 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
10120 unsigned NumInScalars = N->getNumOperands();
10122 EVT VT = N->getValueType(0);
10124 // A vector built entirely of undefs is undef.
10125 if (ISD::allOperandsUndef(N))
10126 return DAG.getUNDEF(VT);
10128 SDValue V = reduceBuildVecExtToExtBuildVec(N);
10132 V = reduceBuildVecConvertToConvertBuildVec(N);
10136 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
10137 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
10138 // at most two distinct vectors, turn this into a shuffle node.
10140 // May only combine to shuffle after legalize if shuffle is legal.
10141 if (LegalOperations &&
10142 !TLI.isOperationLegalOrCustom(ISD::VECTOR_SHUFFLE, VT))
10145 SDValue VecIn1, VecIn2;
10146 for (unsigned i = 0; i != NumInScalars; ++i) {
10147 // Ignore undef inputs.
10148 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
10150 // If this input is something other than a EXTRACT_VECTOR_ELT with a
10151 // constant index, bail out.
10152 if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
10153 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
10154 VecIn1 = VecIn2 = SDValue(0, 0);
10158 // We allow up to two distinct input vectors.
10159 SDValue ExtractedFromVec = N->getOperand(i).getOperand(0);
10160 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
10163 if (VecIn1.getNode() == 0) {
10164 VecIn1 = ExtractedFromVec;
10165 } else if (VecIn2.getNode() == 0) {
10166 VecIn2 = ExtractedFromVec;
10168 // Too many inputs.
10169 VecIn1 = VecIn2 = SDValue(0, 0);
10174 // If everything is good, we can make a shuffle operation.
10175 if (VecIn1.getNode()) {
10176 SmallVector<int, 8> Mask;
10177 for (unsigned i = 0; i != NumInScalars; ++i) {
10178 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
10179 Mask.push_back(-1);
10183 // If extracting from the first vector, just use the index directly.
10184 SDValue Extract = N->getOperand(i);
10185 SDValue ExtVal = Extract.getOperand(1);
10186 if (Extract.getOperand(0) == VecIn1) {
10187 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
10188 if (ExtIndex > VT.getVectorNumElements())
10191 Mask.push_back(ExtIndex);
10195 // Otherwise, use InIdx + VecSize
10196 unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue();
10197 Mask.push_back(Idx+NumInScalars);
10200 // We can't generate a shuffle node with mismatched input and output types.
10201 // Attempt to transform a single input vector to the correct type.
10202 if ((VT != VecIn1.getValueType())) {
10203 // We don't support shuffeling between TWO values of different types.
10204 if (VecIn2.getNode() != 0)
10207 // We only support widening of vectors which are half the size of the
10208 // output registers. For example XMM->YMM widening on X86 with AVX.
10209 if (VecIn1.getValueType().getSizeInBits()*2 != VT.getSizeInBits())
10212 // If the input vector type has a different base type to the output
10213 // vector type, bail out.
10214 if (VecIn1.getValueType().getVectorElementType() !=
10215 VT.getVectorElementType())
10218 // Widen the input vector by adding undef values.
10219 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
10220 VecIn1, DAG.getUNDEF(VecIn1.getValueType()));
10223 // If VecIn2 is unused then change it to undef.
10224 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
10226 // Check that we were able to transform all incoming values to the same
10228 if (VecIn2.getValueType() != VecIn1.getValueType() ||
10229 VecIn1.getValueType() != VT)
10232 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
10233 if (!isTypeLegal(VT))
10236 // Return the new VECTOR_SHUFFLE node.
10240 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]);
10246 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
10247 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
10248 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
10249 // inputs come from at most two distinct vectors, turn this into a shuffle
10252 // If we only have one input vector, we don't need to do any concatenation.
10253 if (N->getNumOperands() == 1)
10254 return N->getOperand(0);
10256 // Check if all of the operands are undefs.
10257 EVT VT = N->getValueType(0);
10258 if (ISD::allOperandsUndef(N))
10259 return DAG.getUNDEF(VT);
10261 // Optimize concat_vectors where one of the vectors is undef.
10262 if (N->getNumOperands() == 2 &&
10263 N->getOperand(1)->getOpcode() == ISD::UNDEF) {
10264 SDValue In = N->getOperand(0);
10265 assert(In.getValueType().isVector() && "Must concat vectors");
10267 // Transform: concat_vectors(scalar, undef) -> scalar_to_vector(sclr).
10268 if (In->getOpcode() == ISD::BITCAST &&
10269 !In->getOperand(0)->getValueType(0).isVector()) {
10270 SDValue Scalar = In->getOperand(0);
10271 EVT SclTy = Scalar->getValueType(0);
10273 if (!SclTy.isFloatingPoint() && !SclTy.isInteger())
10276 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SclTy,
10277 VT.getSizeInBits() / SclTy.getSizeInBits());
10278 if (!TLI.isTypeLegal(NVT) || !TLI.isTypeLegal(Scalar.getValueType()))
10281 SDLoc dl = SDLoc(N);
10282 SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NVT, Scalar);
10283 return DAG.getNode(ISD::BITCAST, dl, VT, Res);
10287 // fold (concat_vectors (BUILD_VECTOR A, B, ...), (BUILD_VECTOR C, D, ...))
10288 // -> (BUILD_VECTOR A, B, ..., C, D, ...)
10289 if (N->getNumOperands() == 2 &&
10290 N->getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
10291 N->getOperand(1).getOpcode() == ISD::BUILD_VECTOR) {
10292 EVT VT = N->getValueType(0);
10293 SDValue N0 = N->getOperand(0);
10294 SDValue N1 = N->getOperand(1);
10295 SmallVector<SDValue, 8> Opnds;
10296 unsigned BuildVecNumElts = N0.getNumOperands();
10298 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10299 Opnds.push_back(N0.getOperand(i));
10300 for (unsigned i = 0; i != BuildVecNumElts; ++i)
10301 Opnds.push_back(N1.getOperand(i));
10303 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, &Opnds[0],
10307 // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
10308 // nodes often generate nop CONCAT_VECTOR nodes.
10309 // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that
10310 // place the incoming vectors at the exact same location.
10311 SDValue SingleSource = SDValue();
10312 unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements();
10314 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
10315 SDValue Op = N->getOperand(i);
10317 if (Op.getOpcode() == ISD::UNDEF)
10320 // Check if this is the identity extract:
10321 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
10324 // Find the single incoming vector for the extract_subvector.
10325 if (SingleSource.getNode()) {
10326 if (Op.getOperand(0) != SingleSource)
10329 SingleSource = Op.getOperand(0);
10331 // Check the source type is the same as the type of the result.
10332 // If not, this concat may extend the vector, so we can not
10333 // optimize it away.
10334 if (SingleSource.getValueType() != N->getValueType(0))
10338 unsigned IdentityIndex = i * PartNumElem;
10339 ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1));
10340 // The extract index must be constant.
10344 // Check that we are reading from the identity index.
10345 if (CS->getZExtValue() != IdentityIndex)
10349 if (SingleSource.getNode())
10350 return SingleSource;
10355 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
10356 EVT NVT = N->getValueType(0);
10357 SDValue V = N->getOperand(0);
10359 if (V->getOpcode() == ISD::CONCAT_VECTORS) {
10361 // (extract_subvec (concat V1, V2, ...), i)
10364 // Only operand 0 is checked as 'concat' assumes all inputs of the same
10366 if (V->getOperand(0).getValueType() != NVT)
10368 unsigned Idx = dyn_cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
10369 unsigned NumElems = NVT.getVectorNumElements();
10370 assert((Idx % NumElems) == 0 &&
10371 "IDX in concat is not a multiple of the result vector length.");
10372 return V->getOperand(Idx / NumElems);
10376 if (V->getOpcode() == ISD::BITCAST)
10377 V = V.getOperand(0);
10379 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
10381 // Handle only simple case where vector being inserted and vector
10382 // being extracted are of same type, and are half size of larger vectors.
10383 EVT BigVT = V->getOperand(0).getValueType();
10384 EVT SmallVT = V->getOperand(1).getValueType();
10385 if (!NVT.bitsEq(SmallVT) || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
10388 // Only handle cases where both indexes are constants with the same type.
10389 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
10390 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
10392 if (InsIdx && ExtIdx &&
10393 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
10394 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
10396 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
10398 // indices are equal or bit offsets are equal => V1
10399 // otherwise => (extract_subvec V1, ExtIdx)
10400 if (InsIdx->getZExtValue() * SmallVT.getScalarType().getSizeInBits() ==
10401 ExtIdx->getZExtValue() * NVT.getScalarType().getSizeInBits())
10402 return DAG.getNode(ISD::BITCAST, dl, NVT, V->getOperand(1));
10403 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT,
10404 DAG.getNode(ISD::BITCAST, dl,
10405 N->getOperand(0).getValueType(),
10406 V->getOperand(0)), N->getOperand(1));
10413 // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat.
10414 static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
10415 EVT VT = N->getValueType(0);
10416 unsigned NumElts = VT.getVectorNumElements();
10418 SDValue N0 = N->getOperand(0);
10419 SDValue N1 = N->getOperand(1);
10420 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
10422 SmallVector<SDValue, 4> Ops;
10423 EVT ConcatVT = N0.getOperand(0).getValueType();
10424 unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements();
10425 unsigned NumConcats = NumElts / NumElemsPerConcat;
10427 // Look at every vector that's inserted. We're looking for exact
10428 // subvector-sized copies from a concatenated vector
10429 for (unsigned I = 0; I != NumConcats; ++I) {
10430 // Make sure we're dealing with a copy.
10431 unsigned Begin = I * NumElemsPerConcat;
10432 bool AllUndef = true, NoUndef = true;
10433 for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) {
10434 if (SVN->getMaskElt(J) >= 0)
10441 if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0)
10444 for (unsigned J = 1; J != NumElemsPerConcat; ++J)
10445 if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J))
10448 unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat;
10449 if (FirstElt < N0.getNumOperands())
10450 Ops.push_back(N0.getOperand(FirstElt));
10452 Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands()));
10454 } else if (AllUndef) {
10455 Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType()));
10456 } else { // Mixed with general masks and undefs, can't do optimization.
10461 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops.data(),
10465 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
10466 EVT VT = N->getValueType(0);
10467 unsigned NumElts = VT.getVectorNumElements();
10469 SDValue N0 = N->getOperand(0);
10470 SDValue N1 = N->getOperand(1);
10472 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
10474 // Canonicalize shuffle undef, undef -> undef
10475 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
10476 return DAG.getUNDEF(VT);
10478 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
10480 // Canonicalize shuffle v, v -> v, undef
10482 SmallVector<int, 8> NewMask;
10483 for (unsigned i = 0; i != NumElts; ++i) {
10484 int Idx = SVN->getMaskElt(i);
10485 if (Idx >= (int)NumElts) Idx -= NumElts;
10486 NewMask.push_back(Idx);
10488 return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT),
10492 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
10493 if (N0.getOpcode() == ISD::UNDEF) {
10494 SmallVector<int, 8> NewMask;
10495 for (unsigned i = 0; i != NumElts; ++i) {
10496 int Idx = SVN->getMaskElt(i);
10498 if (Idx >= (int)NumElts)
10501 Idx = -1; // remove reference to lhs
10503 NewMask.push_back(Idx);
10505 return DAG.getVectorShuffle(VT, SDLoc(N), N1, DAG.getUNDEF(VT),
10509 // Remove references to rhs if it is undef
10510 if (N1.getOpcode() == ISD::UNDEF) {
10511 bool Changed = false;
10512 SmallVector<int, 8> NewMask;
10513 for (unsigned i = 0; i != NumElts; ++i) {
10514 int Idx = SVN->getMaskElt(i);
10515 if (Idx >= (int)NumElts) {
10519 NewMask.push_back(Idx);
10522 return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, &NewMask[0]);
10525 // If it is a splat, check if the argument vector is another splat or a
10526 // build_vector with all scalar elements the same.
10527 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
10528 SDNode *V = N0.getNode();
10530 // If this is a bit convert that changes the element type of the vector but
10531 // not the number of vector elements, look through it. Be careful not to
10532 // look though conversions that change things like v4f32 to v2f64.
10533 if (V->getOpcode() == ISD::BITCAST) {
10534 SDValue ConvInput = V->getOperand(0);
10535 if (ConvInput.getValueType().isVector() &&
10536 ConvInput.getValueType().getVectorNumElements() == NumElts)
10537 V = ConvInput.getNode();
10540 if (V->getOpcode() == ISD::BUILD_VECTOR) {
10541 assert(V->getNumOperands() == NumElts &&
10542 "BUILD_VECTOR has wrong number of operands");
10544 bool AllSame = true;
10545 for (unsigned i = 0; i != NumElts; ++i) {
10546 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
10547 Base = V->getOperand(i);
10551 // Splat of <u, u, u, u>, return <u, u, u, u>
10552 if (!Base.getNode())
10554 for (unsigned i = 0; i != NumElts; ++i) {
10555 if (V->getOperand(i) != Base) {
10560 // Splat of <x, x, x, x>, return <x, x, x, x>
10566 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
10567 Level < AfterLegalizeVectorOps &&
10568 (N1.getOpcode() == ISD::UNDEF ||
10569 (N1.getOpcode() == ISD::CONCAT_VECTORS &&
10570 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) {
10571 SDValue V = partitionShuffleOfConcats(N, DAG);
10577 // If this shuffle node is simply a swizzle of another shuffle node,
10578 // and it reverses the swizzle of the previous shuffle then we can
10579 // optimize shuffle(shuffle(x, undef), undef) -> x.
10580 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
10581 N1.getOpcode() == ISD::UNDEF) {
10583 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
10585 // Shuffle nodes can only reverse shuffles with a single non-undef value.
10586 if (N0.getOperand(1).getOpcode() != ISD::UNDEF)
10589 // The incoming shuffle must be of the same type as the result of the
10590 // current shuffle.
10591 assert(OtherSV->getOperand(0).getValueType() == VT &&
10592 "Shuffle types don't match");
10594 for (unsigned i = 0; i != NumElts; ++i) {
10595 int Idx = SVN->getMaskElt(i);
10596 assert(Idx < (int)NumElts && "Index references undef operand");
10597 // Next, this index comes from the first value, which is the incoming
10598 // shuffle. Adopt the incoming index.
10600 Idx = OtherSV->getMaskElt(Idx);
10602 // The combined shuffle must map each index to itself.
10603 if (Idx >= 0 && (unsigned)Idx != i)
10607 return OtherSV->getOperand(0);
10613 SDValue DAGCombiner::visitINSERT_SUBVECTOR(SDNode *N) {
10614 SDValue N0 = N->getOperand(0);
10615 SDValue N2 = N->getOperand(2);
10617 // If the input vector is a concatenation, and the insert replaces
10618 // one of the halves, we can optimize into a single concat_vectors.
10619 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
10620 N0->getNumOperands() == 2 && N2.getOpcode() == ISD::Constant) {
10621 APInt InsIdx = cast<ConstantSDNode>(N2)->getAPIntValue();
10622 EVT VT = N->getValueType(0);
10624 // Lower half: fold (insert_subvector (concat_vectors X, Y), Z) ->
10625 // (concat_vectors Z, Y)
10627 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
10628 N->getOperand(1), N0.getOperand(1));
10630 // Upper half: fold (insert_subvector (concat_vectors X, Y), Z) ->
10631 // (concat_vectors X, Z)
10632 if (InsIdx == VT.getVectorNumElements()/2)
10633 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
10634 N0.getOperand(0), N->getOperand(1));
10640 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
10641 /// an AND to a vector_shuffle with the destination vector and a zero vector.
10642 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
10643 /// vector_shuffle V, Zero, <0, 4, 2, 4>
10644 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
10645 EVT VT = N->getValueType(0);
10647 SDValue LHS = N->getOperand(0);
10648 SDValue RHS = N->getOperand(1);
10649 if (N->getOpcode() == ISD::AND) {
10650 if (RHS.getOpcode() == ISD::BITCAST)
10651 RHS = RHS.getOperand(0);
10652 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
10653 SmallVector<int, 8> Indices;
10654 unsigned NumElts = RHS.getNumOperands();
10655 for (unsigned i = 0; i != NumElts; ++i) {
10656 SDValue Elt = RHS.getOperand(i);
10657 if (!isa<ConstantSDNode>(Elt))
10660 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
10661 Indices.push_back(i);
10662 else if (cast<ConstantSDNode>(Elt)->isNullValue())
10663 Indices.push_back(NumElts);
10668 // Let's see if the target supports this vector_shuffle.
10669 EVT RVT = RHS.getValueType();
10670 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
10673 // Return the new VECTOR_SHUFFLE node.
10674 EVT EltVT = RVT.getVectorElementType();
10675 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
10676 DAG.getConstant(0, EltVT));
10677 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
10678 RVT, &ZeroOps[0], ZeroOps.size());
10679 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
10680 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
10681 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
10688 /// SimplifyVBinOp - Visit a binary vector operation, like ADD.
10689 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
10690 assert(N->getValueType(0).isVector() &&
10691 "SimplifyVBinOp only works on vectors!");
10693 SDValue LHS = N->getOperand(0);
10694 SDValue RHS = N->getOperand(1);
10695 SDValue Shuffle = XformToShuffleWithZero(N);
10696 if (Shuffle.getNode()) return Shuffle;
10698 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
10700 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
10701 RHS.getOpcode() == ISD::BUILD_VECTOR) {
10702 // Check if both vectors are constants. If not bail out.
10703 if (!(cast<BuildVectorSDNode>(LHS)->isConstant() &&
10704 cast<BuildVectorSDNode>(RHS)->isConstant()))
10707 SmallVector<SDValue, 8> Ops;
10708 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
10709 SDValue LHSOp = LHS.getOperand(i);
10710 SDValue RHSOp = RHS.getOperand(i);
10712 // Can't fold divide by zero.
10713 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
10714 N->getOpcode() == ISD::FDIV) {
10715 if ((RHSOp.getOpcode() == ISD::Constant &&
10716 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
10717 (RHSOp.getOpcode() == ISD::ConstantFP &&
10718 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
10722 EVT VT = LHSOp.getValueType();
10723 EVT RVT = RHSOp.getValueType();
10725 // Integer BUILD_VECTOR operands may have types larger than the element
10726 // size (e.g., when the element type is not legal). Prior to type
10727 // legalization, the types may not match between the two BUILD_VECTORS.
10728 // Truncate one of the operands to make them match.
10729 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
10730 RHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, RHSOp);
10732 LHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), RVT, LHSOp);
10736 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(LHS), VT,
10738 if (FoldOp.getOpcode() != ISD::UNDEF &&
10739 FoldOp.getOpcode() != ISD::Constant &&
10740 FoldOp.getOpcode() != ISD::ConstantFP)
10742 Ops.push_back(FoldOp);
10743 AddToWorkList(FoldOp.getNode());
10746 if (Ops.size() == LHS.getNumOperands())
10747 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
10748 LHS.getValueType(), &Ops[0], Ops.size());
10754 /// SimplifyVUnaryOp - Visit a binary vector operation, like FABS/FNEG.
10755 SDValue DAGCombiner::SimplifyVUnaryOp(SDNode *N) {
10756 assert(N->getValueType(0).isVector() &&
10757 "SimplifyVUnaryOp only works on vectors!");
10759 SDValue N0 = N->getOperand(0);
10761 if (N0.getOpcode() != ISD::BUILD_VECTOR)
10764 // Operand is a BUILD_VECTOR node, see if we can constant fold it.
10765 SmallVector<SDValue, 8> Ops;
10766 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
10767 SDValue Op = N0.getOperand(i);
10768 if (Op.getOpcode() != ISD::UNDEF &&
10769 Op.getOpcode() != ISD::ConstantFP)
10771 EVT EltVT = Op.getValueType();
10772 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(N0), EltVT, Op);
10773 if (FoldOp.getOpcode() != ISD::UNDEF &&
10774 FoldOp.getOpcode() != ISD::ConstantFP)
10776 Ops.push_back(FoldOp);
10777 AddToWorkList(FoldOp.getNode());
10780 if (Ops.size() != N0.getNumOperands())
10783 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
10784 N0.getValueType(), &Ops[0], Ops.size());
10787 SDValue DAGCombiner::SimplifySelect(SDLoc DL, SDValue N0,
10788 SDValue N1, SDValue N2){
10789 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
10791 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
10792 cast<CondCodeSDNode>(N0.getOperand(2))->get());
10794 // If we got a simplified select_cc node back from SimplifySelectCC, then
10795 // break it down into a new SETCC node, and a new SELECT node, and then return
10796 // the SELECT node, since we were called with a SELECT node.
10797 if (SCC.getNode()) {
10798 // Check to see if we got a select_cc back (to turn into setcc/select).
10799 // Otherwise, just return whatever node we got back, like fabs.
10800 if (SCC.getOpcode() == ISD::SELECT_CC) {
10801 SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0),
10803 SCC.getOperand(0), SCC.getOperand(1),
10804 SCC.getOperand(4));
10805 AddToWorkList(SETCC.getNode());
10806 return DAG.getSelect(SDLoc(SCC), SCC.getValueType(),
10807 SCC.getOperand(2), SCC.getOperand(3), SETCC);
10815 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
10816 /// are the two values being selected between, see if we can simplify the
10817 /// select. Callers of this should assume that TheSelect is deleted if this
10818 /// returns true. As such, they should return the appropriate thing (e.g. the
10819 /// node) back to the top-level of the DAG combiner loop to avoid it being
10821 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
10824 // Cannot simplify select with vector condition
10825 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
10827 // If this is a select from two identical things, try to pull the operation
10828 // through the select.
10829 if (LHS.getOpcode() != RHS.getOpcode() ||
10830 !LHS.hasOneUse() || !RHS.hasOneUse())
10833 // If this is a load and the token chain is identical, replace the select
10834 // of two loads with a load through a select of the address to load from.
10835 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
10836 // constants have been dropped into the constant pool.
10837 if (LHS.getOpcode() == ISD::LOAD) {
10838 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
10839 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
10841 // Token chains must be identical.
10842 if (LHS.getOperand(0) != RHS.getOperand(0) ||
10843 // Do not let this transformation reduce the number of volatile loads.
10844 LLD->isVolatile() || RLD->isVolatile() ||
10845 // If this is an EXTLOAD, the VT's must match.
10846 LLD->getMemoryVT() != RLD->getMemoryVT() ||
10847 // If this is an EXTLOAD, the kind of extension must match.
10848 (LLD->getExtensionType() != RLD->getExtensionType() &&
10849 // The only exception is if one of the extensions is anyext.
10850 LLD->getExtensionType() != ISD::EXTLOAD &&
10851 RLD->getExtensionType() != ISD::EXTLOAD) ||
10852 // FIXME: this discards src value information. This is
10853 // over-conservative. It would be beneficial to be able to remember
10854 // both potential memory locations. Since we are discarding
10855 // src value info, don't do the transformation if the memory
10856 // locations are not in the default address space.
10857 LLD->getPointerInfo().getAddrSpace() != 0 ||
10858 RLD->getPointerInfo().getAddrSpace() != 0 ||
10859 !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(),
10860 LLD->getBasePtr().getValueType()))
10863 // Check that the select condition doesn't reach either load. If so,
10864 // folding this will induce a cycle into the DAG. If not, this is safe to
10865 // xform, so create a select of the addresses.
10867 if (TheSelect->getOpcode() == ISD::SELECT) {
10868 SDNode *CondNode = TheSelect->getOperand(0).getNode();
10869 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
10870 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
10872 // The loads must not depend on one another.
10873 if (LLD->isPredecessorOf(RLD) ||
10874 RLD->isPredecessorOf(LLD))
10876 Addr = DAG.getSelect(SDLoc(TheSelect),
10877 LLD->getBasePtr().getValueType(),
10878 TheSelect->getOperand(0), LLD->getBasePtr(),
10879 RLD->getBasePtr());
10880 } else { // Otherwise SELECT_CC
10881 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
10882 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
10884 if ((LLD->hasAnyUseOfValue(1) &&
10885 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
10886 (RLD->hasAnyUseOfValue(1) &&
10887 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
10890 Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect),
10891 LLD->getBasePtr().getValueType(),
10892 TheSelect->getOperand(0),
10893 TheSelect->getOperand(1),
10894 LLD->getBasePtr(), RLD->getBasePtr(),
10895 TheSelect->getOperand(4));
10899 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
10900 Load = DAG.getLoad(TheSelect->getValueType(0),
10902 // FIXME: Discards pointer and TBAA info.
10903 LLD->getChain(), Addr, MachinePointerInfo(),
10904 LLD->isVolatile(), LLD->isNonTemporal(),
10905 LLD->isInvariant(), LLD->getAlignment());
10907 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
10908 RLD->getExtensionType() : LLD->getExtensionType(),
10910 TheSelect->getValueType(0),
10911 // FIXME: Discards pointer and TBAA info.
10912 LLD->getChain(), Addr, MachinePointerInfo(),
10913 LLD->getMemoryVT(), LLD->isVolatile(),
10914 LLD->isNonTemporal(), LLD->getAlignment());
10917 // Users of the select now use the result of the load.
10918 CombineTo(TheSelect, Load);
10920 // Users of the old loads now use the new load's chain. We know the
10921 // old-load value is dead now.
10922 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
10923 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
10930 /// SimplifySelectCC - Simplify an expression of the form (N0 cond N1) ? N2 : N3
10931 /// where 'cond' is the comparison specified by CC.
10932 SDValue DAGCombiner::SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1,
10933 SDValue N2, SDValue N3,
10934 ISD::CondCode CC, bool NotExtCompare) {
10935 // (x ? y : y) -> y.
10936 if (N2 == N3) return N2;
10938 EVT VT = N2.getValueType();
10939 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
10940 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
10941 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
10943 // Determine if the condition we're dealing with is constant
10944 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
10945 N0, N1, CC, DL, false);
10946 if (SCC.getNode()) AddToWorkList(SCC.getNode());
10947 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
10949 // fold select_cc true, x, y -> x
10950 if (SCCC && !SCCC->isNullValue())
10952 // fold select_cc false, x, y -> y
10953 if (SCCC && SCCC->isNullValue())
10956 // Check to see if we can simplify the select into an fabs node
10957 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
10958 // Allow either -0.0 or 0.0
10959 if (CFP->getValueAPF().isZero()) {
10960 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
10961 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
10962 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
10963 N2 == N3.getOperand(0))
10964 return DAG.getNode(ISD::FABS, DL, VT, N0);
10966 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
10967 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
10968 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
10969 N2.getOperand(0) == N3)
10970 return DAG.getNode(ISD::FABS, DL, VT, N3);
10974 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
10975 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
10976 // in it. This is a win when the constant is not otherwise available because
10977 // it replaces two constant pool loads with one. We only do this if the FP
10978 // type is known to be legal, because if it isn't, then we are before legalize
10979 // types an we want the other legalization to happen first (e.g. to avoid
10980 // messing with soft float) and if the ConstantFP is not legal, because if
10981 // it is legal, we may not need to store the FP constant in a constant pool.
10982 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
10983 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
10984 if (TLI.isTypeLegal(N2.getValueType()) &&
10985 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
10986 TargetLowering::Legal) &&
10987 // If both constants have multiple uses, then we won't need to do an
10988 // extra load, they are likely around in registers for other users.
10989 (TV->hasOneUse() || FV->hasOneUse())) {
10990 Constant *Elts[] = {
10991 const_cast<ConstantFP*>(FV->getConstantFPValue()),
10992 const_cast<ConstantFP*>(TV->getConstantFPValue())
10994 Type *FPTy = Elts[0]->getType();
10995 const DataLayout &TD = *TLI.getDataLayout();
10997 // Create a ConstantArray of the two constants.
10998 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
10999 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
11000 TD.getPrefTypeAlignment(FPTy));
11001 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
11003 // Get the offsets to the 0 and 1 element of the array so that we can
11004 // select between them.
11005 SDValue Zero = DAG.getIntPtrConstant(0);
11006 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
11007 SDValue One = DAG.getIntPtrConstant(EltSize);
11009 SDValue Cond = DAG.getSetCC(DL,
11010 getSetCCResultType(N0.getValueType()),
11012 AddToWorkList(Cond.getNode());
11013 SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(),
11015 AddToWorkList(CstOffset.getNode());
11016 CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx,
11018 AddToWorkList(CPIdx.getNode());
11019 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
11020 MachinePointerInfo::getConstantPool(), false,
11021 false, false, Alignment);
11026 // Check to see if we can perform the "gzip trick", transforming
11027 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
11028 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
11029 (N1C->isNullValue() || // (a < 0) ? b : 0
11030 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
11031 EVT XType = N0.getValueType();
11032 EVT AType = N2.getValueType();
11033 if (XType.bitsGE(AType)) {
11034 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
11035 // single-bit constant.
11036 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
11037 unsigned ShCtV = N2C->getAPIntValue().logBase2();
11038 ShCtV = XType.getSizeInBits()-ShCtV-1;
11039 SDValue ShCt = DAG.getConstant(ShCtV,
11040 getShiftAmountTy(N0.getValueType()));
11041 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0),
11043 AddToWorkList(Shift.getNode());
11045 if (XType.bitsGT(AType)) {
11046 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
11047 AddToWorkList(Shift.getNode());
11050 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
11053 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0),
11055 DAG.getConstant(XType.getSizeInBits()-1,
11056 getShiftAmountTy(N0.getValueType())));
11057 AddToWorkList(Shift.getNode());
11059 if (XType.bitsGT(AType)) {
11060 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
11061 AddToWorkList(Shift.getNode());
11064 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
11068 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
11069 // where y is has a single bit set.
11070 // A plaintext description would be, we can turn the SELECT_CC into an AND
11071 // when the condition can be materialized as an all-ones register. Any
11072 // single bit-test can be materialized as an all-ones register with
11073 // shift-left and shift-right-arith.
11074 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
11075 N0->getValueType(0) == VT &&
11076 N1C && N1C->isNullValue() &&
11077 N2C && N2C->isNullValue()) {
11078 SDValue AndLHS = N0->getOperand(0);
11079 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
11080 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
11081 // Shift the tested bit over the sign bit.
11082 APInt AndMask = ConstAndRHS->getAPIntValue();
11084 DAG.getConstant(AndMask.countLeadingZeros(),
11085 getShiftAmountTy(AndLHS.getValueType()));
11086 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt);
11088 // Now arithmetic right shift it all the way over, so the result is either
11089 // all-ones, or zero.
11091 DAG.getConstant(AndMask.getBitWidth()-1,
11092 getShiftAmountTy(Shl.getValueType()));
11093 SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt);
11095 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
11099 // fold select C, 16, 0 -> shl C, 4
11100 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
11101 TLI.getBooleanContents(N0.getValueType().isVector()) ==
11102 TargetLowering::ZeroOrOneBooleanContent) {
11104 // If the caller doesn't want us to simplify this into a zext of a compare,
11106 if (NotExtCompare && N2C->getAPIntValue() == 1)
11109 // Get a SetCC of the condition
11110 // NOTE: Don't create a SETCC if it's not legal on this target.
11111 if (!LegalOperations ||
11112 TLI.isOperationLegal(ISD::SETCC,
11113 LegalTypes ? getSetCCResultType(N0.getValueType()) : MVT::i1)) {
11115 // cast from setcc result type to select result type
11117 SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()),
11119 if (N2.getValueType().bitsLT(SCC.getValueType()))
11120 Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2),
11121 N2.getValueType());
11123 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
11124 N2.getValueType(), SCC);
11126 SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC);
11127 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
11128 N2.getValueType(), SCC);
11131 AddToWorkList(SCC.getNode());
11132 AddToWorkList(Temp.getNode());
11134 if (N2C->getAPIntValue() == 1)
11137 // shl setcc result by log2 n2c
11138 return DAG.getNode(
11139 ISD::SHL, DL, N2.getValueType(), Temp,
11140 DAG.getConstant(N2C->getAPIntValue().logBase2(),
11141 getShiftAmountTy(Temp.getValueType())));
11145 // Check to see if this is the equivalent of setcc
11146 // FIXME: Turn all of these into setcc if setcc if setcc is legal
11147 // otherwise, go ahead with the folds.
11148 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
11149 EVT XType = N0.getValueType();
11150 if (!LegalOperations ||
11151 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(XType))) {
11152 SDValue Res = DAG.getSetCC(DL, getSetCCResultType(XType), N0, N1, CC);
11153 if (Res.getValueType() != VT)
11154 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
11158 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
11159 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
11160 (!LegalOperations ||
11161 TLI.isOperationLegal(ISD::CTLZ, XType))) {
11162 SDValue Ctlz = DAG.getNode(ISD::CTLZ, SDLoc(N0), XType, N0);
11163 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
11164 DAG.getConstant(Log2_32(XType.getSizeInBits()),
11165 getShiftAmountTy(Ctlz.getValueType())));
11167 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
11168 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
11169 SDValue NegN0 = DAG.getNode(ISD::SUB, SDLoc(N0),
11170 XType, DAG.getConstant(0, XType), N0);
11171 SDValue NotN0 = DAG.getNOT(SDLoc(N0), N0, XType);
11172 return DAG.getNode(ISD::SRL, DL, XType,
11173 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
11174 DAG.getConstant(XType.getSizeInBits()-1,
11175 getShiftAmountTy(XType)));
11177 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
11178 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
11179 SDValue Sign = DAG.getNode(ISD::SRL, SDLoc(N0), XType, N0,
11180 DAG.getConstant(XType.getSizeInBits()-1,
11181 getShiftAmountTy(N0.getValueType())));
11182 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
11186 // Check to see if this is an integer abs.
11187 // select_cc setg[te] X, 0, X, -X ->
11188 // select_cc setgt X, -1, X, -X ->
11189 // select_cc setl[te] X, 0, -X, X ->
11190 // select_cc setlt X, 1, -X, X ->
11191 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
11193 ConstantSDNode *SubC = NULL;
11194 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
11195 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
11196 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
11197 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
11198 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
11199 (N1C->isOne() && CC == ISD::SETLT)) &&
11200 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
11201 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
11203 EVT XType = N0.getValueType();
11204 if (SubC && SubC->isNullValue() && XType.isInteger()) {
11205 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0), XType,
11207 DAG.getConstant(XType.getSizeInBits()-1,
11208 getShiftAmountTy(N0.getValueType())));
11209 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0),
11211 AddToWorkList(Shift.getNode());
11212 AddToWorkList(Add.getNode());
11213 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
11220 /// SimplifySetCC - This is a stub for TargetLowering::SimplifySetCC.
11221 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
11222 SDValue N1, ISD::CondCode Cond,
11223 SDLoc DL, bool foldBooleans) {
11224 TargetLowering::DAGCombinerInfo
11225 DagCombineInfo(DAG, Level, false, this);
11226 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
11229 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
11230 /// return a DAG expression to select that will generate the same value by
11231 /// multiplying by a magic number. See:
11232 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
11233 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
11234 std::vector<SDNode*> Built;
11235 SDValue S = TLI.BuildSDIV(N, DAG, LegalOperations, &Built);
11237 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
11239 AddToWorkList(*ii);
11243 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
11244 /// return a DAG expression to select that will generate the same value by
11245 /// multiplying by a magic number. See:
11246 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
11247 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
11248 std::vector<SDNode*> Built;
11249 SDValue S = TLI.BuildUDIV(N, DAG, LegalOperations, &Built);
11251 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
11253 AddToWorkList(*ii);
11257 /// FindBaseOffset - Return true if base is a frame index, which is known not
11258 // to alias with anything but itself. Provides base object and offset as
11260 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
11261 const GlobalValue *&GV, const void *&CV) {
11262 // Assume it is a primitive operation.
11263 Base = Ptr; Offset = 0; GV = 0; CV = 0;
11265 // If it's an adding a simple constant then integrate the offset.
11266 if (Base.getOpcode() == ISD::ADD) {
11267 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
11268 Base = Base.getOperand(0);
11269 Offset += C->getZExtValue();
11273 // Return the underlying GlobalValue, and update the Offset. Return false
11274 // for GlobalAddressSDNode since the same GlobalAddress may be represented
11275 // by multiple nodes with different offsets.
11276 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
11277 GV = G->getGlobal();
11278 Offset += G->getOffset();
11282 // Return the underlying Constant value, and update the Offset. Return false
11283 // for ConstantSDNodes since the same constant pool entry may be represented
11284 // by multiple nodes with different offsets.
11285 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
11286 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
11287 : (const void *)C->getConstVal();
11288 Offset += C->getOffset();
11291 // If it's any of the following then it can't alias with anything but itself.
11292 return isa<FrameIndexSDNode>(Base);
11295 /// isAlias - Return true if there is any possibility that the two addresses
11297 bool DAGCombiner::isAlias(SDValue Ptr1, int64_t Size1, bool IsVolatile1,
11298 const Value *SrcValue1, int SrcValueOffset1,
11299 unsigned SrcValueAlign1,
11300 const MDNode *TBAAInfo1,
11301 SDValue Ptr2, int64_t Size2, bool IsVolatile2,
11302 const Value *SrcValue2, int SrcValueOffset2,
11303 unsigned SrcValueAlign2,
11304 const MDNode *TBAAInfo2) const {
11305 // If they are the same then they must be aliases.
11306 if (Ptr1 == Ptr2) return true;
11308 // If they are both volatile then they cannot be reordered.
11309 if (IsVolatile1 && IsVolatile2) return true;
11311 // Gather base node and offset information.
11312 SDValue Base1, Base2;
11313 int64_t Offset1, Offset2;
11314 const GlobalValue *GV1, *GV2;
11315 const void *CV1, *CV2;
11316 bool isFrameIndex1 = FindBaseOffset(Ptr1, Base1, Offset1, GV1, CV1);
11317 bool isFrameIndex2 = FindBaseOffset(Ptr2, Base2, Offset2, GV2, CV2);
11319 // If they have a same base address then check to see if they overlap.
11320 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
11321 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
11323 // It is possible for different frame indices to alias each other, mostly
11324 // when tail call optimization reuses return address slots for arguments.
11325 // To catch this case, look up the actual index of frame indices to compute
11326 // the real alias relationship.
11327 if (isFrameIndex1 && isFrameIndex2) {
11328 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11329 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
11330 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
11331 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
11334 // Otherwise, if we know what the bases are, and they aren't identical, then
11335 // we know they cannot alias.
11336 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
11339 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
11340 // compared to the size and offset of the access, we may be able to prove they
11341 // do not alias. This check is conservative for now to catch cases created by
11342 // splitting vector types.
11343 if ((SrcValueAlign1 == SrcValueAlign2) &&
11344 (SrcValueOffset1 != SrcValueOffset2) &&
11345 (Size1 == Size2) && (SrcValueAlign1 > Size1)) {
11346 int64_t OffAlign1 = SrcValueOffset1 % SrcValueAlign1;
11347 int64_t OffAlign2 = SrcValueOffset2 % SrcValueAlign1;
11349 // There is no overlap between these relatively aligned accesses of similar
11350 // size, return no alias.
11351 if ((OffAlign1 + Size1) <= OffAlign2 || (OffAlign2 + Size2) <= OffAlign1)
11355 bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0 ? CombinerGlobalAA :
11356 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
11358 if (CombinerAAOnlyFunc.getNumOccurrences() &&
11359 CombinerAAOnlyFunc != DAG.getMachineFunction().getName())
11362 if (UseAA && SrcValue1 && SrcValue2) {
11363 // Use alias analysis information.
11364 int64_t MinOffset = std::min(SrcValueOffset1, SrcValueOffset2);
11365 int64_t Overlap1 = Size1 + SrcValueOffset1 - MinOffset;
11366 int64_t Overlap2 = Size2 + SrcValueOffset2 - MinOffset;
11367 AliasAnalysis::AliasResult AAResult =
11368 AA.alias(AliasAnalysis::Location(SrcValue1, Overlap1,
11369 UseTBAA ? TBAAInfo1 : 0),
11370 AliasAnalysis::Location(SrcValue2, Overlap2,
11371 UseTBAA ? TBAAInfo2 : 0));
11372 if (AAResult == AliasAnalysis::NoAlias)
11376 // Otherwise we have to assume they alias.
11380 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) {
11381 SDValue Ptr0, Ptr1;
11382 int64_t Size0, Size1;
11383 bool IsVolatile0, IsVolatile1;
11384 const Value *SrcValue0, *SrcValue1;
11385 int SrcValueOffset0, SrcValueOffset1;
11386 unsigned SrcValueAlign0, SrcValueAlign1;
11387 const MDNode *SrcTBAAInfo0, *SrcTBAAInfo1;
11388 FindAliasInfo(Op0, Ptr0, Size0, IsVolatile0, SrcValue0, SrcValueOffset0,
11389 SrcValueAlign0, SrcTBAAInfo0);
11390 FindAliasInfo(Op1, Ptr1, Size1, IsVolatile1, SrcValue1, SrcValueOffset1,
11391 SrcValueAlign1, SrcTBAAInfo1);
11392 return isAlias(Ptr0, Size0, IsVolatile0, SrcValue0, SrcValueOffset0,
11393 SrcValueAlign0, SrcTBAAInfo0,
11394 Ptr1, Size1, IsVolatile1, SrcValue1, SrcValueOffset1,
11395 SrcValueAlign1, SrcTBAAInfo1);
11398 /// FindAliasInfo - Extracts the relevant alias information from the memory
11399 /// node. Returns true if the operand was a nonvolatile load.
11400 bool DAGCombiner::FindAliasInfo(SDNode *N,
11401 SDValue &Ptr, int64_t &Size, bool &IsVolatile,
11402 const Value *&SrcValue,
11403 int &SrcValueOffset,
11404 unsigned &SrcValueAlign,
11405 const MDNode *&TBAAInfo) const {
11406 LSBaseSDNode *LS = cast<LSBaseSDNode>(N);
11408 Ptr = LS->getBasePtr();
11409 Size = LS->getMemoryVT().getSizeInBits() >> 3;
11410 IsVolatile = LS->isVolatile();
11411 SrcValue = LS->getSrcValue();
11412 SrcValueOffset = LS->getSrcValueOffset();
11413 SrcValueAlign = LS->getOriginalAlignment();
11414 TBAAInfo = LS->getTBAAInfo();
11415 return isa<LoadSDNode>(LS) && !IsVolatile;
11418 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
11419 /// looking for aliasing nodes and adding them to the Aliases vector.
11420 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
11421 SmallVectorImpl<SDValue> &Aliases) {
11422 SmallVector<SDValue, 8> Chains; // List of chains to visit.
11423 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
11425 // Get alias information for node.
11429 const Value *SrcValue;
11430 int SrcValueOffset;
11431 unsigned SrcValueAlign;
11432 const MDNode *SrcTBAAInfo;
11433 bool IsLoad = FindAliasInfo(N, Ptr, Size, IsVolatile, SrcValue,
11434 SrcValueOffset, SrcValueAlign, SrcTBAAInfo);
11437 Chains.push_back(OriginalChain);
11438 unsigned Depth = 0;
11440 // Look at each chain and determine if it is an alias. If so, add it to the
11441 // aliases list. If not, then continue up the chain looking for the next
11443 while (!Chains.empty()) {
11444 SDValue Chain = Chains.back();
11447 // For TokenFactor nodes, look at each operand and only continue up the
11448 // chain until we find two aliases. If we've seen two aliases, assume we'll
11449 // find more and revert to original chain since the xform is unlikely to be
11452 // FIXME: The depth check could be made to return the last non-aliasing
11453 // chain we found before we hit a tokenfactor rather than the original
11455 if (Depth > 6 || Aliases.size() == 2) {
11457 Aliases.push_back(OriginalChain);
11461 // Don't bother if we've been before.
11462 if (!Visited.insert(Chain.getNode()))
11465 switch (Chain.getOpcode()) {
11466 case ISD::EntryToken:
11467 // Entry token is ideal chain operand, but handled in FindBetterChain.
11472 // Get alias information for Chain.
11476 const Value *OpSrcValue;
11477 int OpSrcValueOffset;
11478 unsigned OpSrcValueAlign;
11479 const MDNode *OpSrcTBAAInfo;
11480 bool IsOpLoad = FindAliasInfo(Chain.getNode(), OpPtr, OpSize,
11481 OpIsVolatile, OpSrcValue, OpSrcValueOffset,
11485 // If chain is alias then stop here.
11486 if (!(IsLoad && IsOpLoad) &&
11487 isAlias(Ptr, Size, IsVolatile, SrcValue, SrcValueOffset,
11488 SrcValueAlign, SrcTBAAInfo,
11489 OpPtr, OpSize, OpIsVolatile, OpSrcValue, OpSrcValueOffset,
11490 OpSrcValueAlign, OpSrcTBAAInfo)) {
11491 Aliases.push_back(Chain);
11493 // Look further up the chain.
11494 Chains.push_back(Chain.getOperand(0));
11500 case ISD::TokenFactor:
11501 // We have to check each of the operands of the token factor for "small"
11502 // token factors, so we queue them up. Adding the operands to the queue
11503 // (stack) in reverse order maintains the original order and increases the
11504 // likelihood that getNode will find a matching token factor (CSE.)
11505 if (Chain.getNumOperands() > 16) {
11506 Aliases.push_back(Chain);
11509 for (unsigned n = Chain.getNumOperands(); n;)
11510 Chains.push_back(Chain.getOperand(--n));
11515 // For all other instructions we will just have to take what we can get.
11516 Aliases.push_back(Chain);
11521 // We need to be careful here to also search for aliases through the
11522 // value operand of a store, etc. Consider the following situation:
11524 // L1 = load Token1, %52
11525 // S1 = store Token1, L1, %51
11526 // L2 = load Token1, %52+8
11527 // S2 = store Token1, L2, %51+8
11528 // Token2 = Token(S1, S2)
11529 // L3 = load Token2, %53
11530 // S3 = store Token2, L3, %52
11531 // L4 = load Token2, %53+8
11532 // S4 = store Token2, L4, %52+8
11533 // If we search for aliases of S3 (which loads address %52), and we look
11534 // only through the chain, then we'll miss the trivial dependence on L1
11535 // (which also loads from %52). We then might change all loads and
11536 // stores to use Token1 as their chain operand, which could result in
11537 // copying %53 into %52 before copying %52 into %51 (which should
11540 // The problem is, however, that searching for such data dependencies
11541 // can become expensive, and the cost is not directly related to the
11542 // chain depth. Instead, we'll rule out such configurations here by
11543 // insisting that we've visited all chain users (except for users
11544 // of the original chain, which is not necessary). When doing this,
11545 // we need to look through nodes we don't care about (otherwise, things
11546 // like register copies will interfere with trivial cases).
11548 SmallVector<const SDNode *, 16> Worklist;
11549 for (SmallPtrSet<SDNode *, 16>::iterator I = Visited.begin(),
11550 IE = Visited.end(); I != IE; ++I)
11551 if (*I != OriginalChain.getNode())
11552 Worklist.push_back(*I);
11554 while (!Worklist.empty()) {
11555 const SDNode *M = Worklist.pop_back_val();
11557 // We have already visited M, and want to make sure we've visited any uses
11558 // of M that we care about. For uses that we've not visisted, and don't
11559 // care about, queue them to the worklist.
11561 for (SDNode::use_iterator UI = M->use_begin(),
11562 UIE = M->use_end(); UI != UIE; ++UI)
11563 if (UI.getUse().getValueType() == MVT::Other && Visited.insert(*UI)) {
11564 if (isa<MemIntrinsicSDNode>(*UI) || isa<MemSDNode>(*UI)) {
11565 // We've not visited this use, and we care about it (it could have an
11566 // ordering dependency with the original node).
11568 Aliases.push_back(OriginalChain);
11572 // We've not visited this use, but we don't care about it. Mark it as
11573 // visited and enqueue it to the worklist.
11574 Worklist.push_back(*UI);
11579 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
11580 /// for a better chain (aliasing node.)
11581 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
11582 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
11584 // Accumulate all the aliases to this node.
11585 GatherAllAliases(N, OldChain, Aliases);
11587 // If no operands then chain to entry token.
11588 if (Aliases.size() == 0)
11589 return DAG.getEntryNode();
11591 // If a single operand then chain to it. We don't need to revisit it.
11592 if (Aliases.size() == 1)
11595 // Construct a custom tailored token factor.
11596 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other,
11597 &Aliases[0], Aliases.size());
11600 // SelectionDAG::Combine - This is the entry point for the file.
11602 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
11603 CodeGenOpt::Level OptLevel) {
11604 /// run - This is the main entry point to this class.
11606 DAGCombiner(*this, AA, OptLevel).Run(Level);