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("Turn on alias analysis during testing"));
56 CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden,
57 cl::desc("Include global information in alias analysis"));
59 /// Hidden option to stress test load slicing, i.e., when this option
60 /// is enabled, load slicing bypasses most of its profitability guards.
62 StressLoadSlicing("combiner-stress-load-slicing", cl::Hidden,
63 cl::desc("Bypass the profitability model of load "
67 //------------------------------ DAGCombiner ---------------------------------//
71 const TargetLowering &TLI;
73 CodeGenOpt::Level OptLevel;
78 // Worklist of all of the nodes that need to be simplified.
80 // This has the semantics that when adding to the worklist,
81 // the item added must be next to be processed. It should
82 // also only appear once. The naive approach to this takes
85 // To reduce the insert/remove time to logarithmic, we use
86 // a set and a vector to maintain our worklist.
88 // The set contains the items on the worklist, but does not
89 // maintain the order they should be visited.
91 // The vector maintains the order nodes should be visited, but may
92 // contain duplicate or removed nodes. When choosing a node to
93 // visit, we pop off the order stack until we find an item that is
94 // also in the contents set. All operations are O(log N).
95 SmallPtrSet<SDNode*, 64> WorkListContents;
96 SmallVector<SDNode*, 64> WorkListOrder;
98 // AA - Used for DAG load/store alias analysis.
101 /// AddUsersToWorkList - When an instruction is simplified, add all users of
102 /// the instruction to the work lists because they might get more simplified
105 void AddUsersToWorkList(SDNode *N) {
106 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
111 /// visit - call the node-specific routine that knows how to fold each
112 /// particular type of node.
113 SDValue visit(SDNode *N);
116 /// AddToWorkList - Add to the work list making sure its instance is at the
117 /// back (next to be processed.)
118 void AddToWorkList(SDNode *N) {
119 WorkListContents.insert(N);
120 WorkListOrder.push_back(N);
123 /// removeFromWorkList - remove all instances of N from the worklist.
125 void removeFromWorkList(SDNode *N) {
126 WorkListContents.erase(N);
129 SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
132 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) {
133 return CombineTo(N, &Res, 1, AddTo);
136 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1,
138 SDValue To[] = { Res0, Res1 };
139 return CombineTo(N, To, 2, AddTo);
142 void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO);
146 /// SimplifyDemandedBits - Check the specified integer node value to see if
147 /// it can be simplified or if things it uses can be simplified by bit
148 /// propagation. If so, return true.
149 bool SimplifyDemandedBits(SDValue Op) {
150 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
151 APInt Demanded = APInt::getAllOnesValue(BitWidth);
152 return SimplifyDemandedBits(Op, Demanded);
155 bool SimplifyDemandedBits(SDValue Op, const APInt &Demanded);
157 bool CombineToPreIndexedLoadStore(SDNode *N);
158 bool CombineToPostIndexedLoadStore(SDNode *N);
159 bool SliceUpLoad(SDNode *N);
161 void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad);
162 SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace);
163 SDValue SExtPromoteOperand(SDValue Op, EVT PVT);
164 SDValue ZExtPromoteOperand(SDValue Op, EVT PVT);
165 SDValue PromoteIntBinOp(SDValue Op);
166 SDValue PromoteIntShiftOp(SDValue Op);
167 SDValue PromoteExtend(SDValue Op);
168 bool PromoteLoad(SDValue Op);
170 void ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
171 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
172 ISD::NodeType ExtType);
174 /// combine - call the node-specific routine that knows how to fold each
175 /// particular type of node. If that doesn't do anything, try the
176 /// target-specific DAG combines.
177 SDValue combine(SDNode *N);
179 // Visitation implementation - Implement dag node combining for different
180 // node types. The semantics are as follows:
182 // SDValue.getNode() == 0 - No change was made
183 // SDValue.getNode() == N - N was replaced, is dead and has been handled.
184 // otherwise - N should be replaced by the returned Operand.
186 SDValue visitTokenFactor(SDNode *N);
187 SDValue visitMERGE_VALUES(SDNode *N);
188 SDValue visitADD(SDNode *N);
189 SDValue visitSUB(SDNode *N);
190 SDValue visitADDC(SDNode *N);
191 SDValue visitSUBC(SDNode *N);
192 SDValue visitADDE(SDNode *N);
193 SDValue visitSUBE(SDNode *N);
194 SDValue visitMUL(SDNode *N);
195 SDValue visitSDIV(SDNode *N);
196 SDValue visitUDIV(SDNode *N);
197 SDValue visitSREM(SDNode *N);
198 SDValue visitUREM(SDNode *N);
199 SDValue visitMULHU(SDNode *N);
200 SDValue visitMULHS(SDNode *N);
201 SDValue visitSMUL_LOHI(SDNode *N);
202 SDValue visitUMUL_LOHI(SDNode *N);
203 SDValue visitSMULO(SDNode *N);
204 SDValue visitUMULO(SDNode *N);
205 SDValue visitSDIVREM(SDNode *N);
206 SDValue visitUDIVREM(SDNode *N);
207 SDValue visitAND(SDNode *N);
208 SDValue visitOR(SDNode *N);
209 SDValue visitXOR(SDNode *N);
210 SDValue SimplifyVBinOp(SDNode *N);
211 SDValue SimplifyVUnaryOp(SDNode *N);
212 SDValue visitSHL(SDNode *N);
213 SDValue visitSRA(SDNode *N);
214 SDValue visitSRL(SDNode *N);
215 SDValue visitCTLZ(SDNode *N);
216 SDValue visitCTLZ_ZERO_UNDEF(SDNode *N);
217 SDValue visitCTTZ(SDNode *N);
218 SDValue visitCTTZ_ZERO_UNDEF(SDNode *N);
219 SDValue visitCTPOP(SDNode *N);
220 SDValue visitSELECT(SDNode *N);
221 SDValue visitVSELECT(SDNode *N);
222 SDValue visitSELECT_CC(SDNode *N);
223 SDValue visitSETCC(SDNode *N);
224 SDValue visitSIGN_EXTEND(SDNode *N);
225 SDValue visitZERO_EXTEND(SDNode *N);
226 SDValue visitANY_EXTEND(SDNode *N);
227 SDValue visitSIGN_EXTEND_INREG(SDNode *N);
228 SDValue visitTRUNCATE(SDNode *N);
229 SDValue visitBITCAST(SDNode *N);
230 SDValue visitBUILD_PAIR(SDNode *N);
231 SDValue visitFADD(SDNode *N);
232 SDValue visitFSUB(SDNode *N);
233 SDValue visitFMUL(SDNode *N);
234 SDValue visitFMA(SDNode *N);
235 SDValue visitFDIV(SDNode *N);
236 SDValue visitFREM(SDNode *N);
237 SDValue visitFCOPYSIGN(SDNode *N);
238 SDValue visitSINT_TO_FP(SDNode *N);
239 SDValue visitUINT_TO_FP(SDNode *N);
240 SDValue visitFP_TO_SINT(SDNode *N);
241 SDValue visitFP_TO_UINT(SDNode *N);
242 SDValue visitFP_ROUND(SDNode *N);
243 SDValue visitFP_ROUND_INREG(SDNode *N);
244 SDValue visitFP_EXTEND(SDNode *N);
245 SDValue visitFNEG(SDNode *N);
246 SDValue visitFABS(SDNode *N);
247 SDValue visitFCEIL(SDNode *N);
248 SDValue visitFTRUNC(SDNode *N);
249 SDValue visitFFLOOR(SDNode *N);
250 SDValue visitBRCOND(SDNode *N);
251 SDValue visitBR_CC(SDNode *N);
252 SDValue visitLOAD(SDNode *N);
253 SDValue visitSTORE(SDNode *N);
254 SDValue visitINSERT_VECTOR_ELT(SDNode *N);
255 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
256 SDValue visitBUILD_VECTOR(SDNode *N);
257 SDValue visitCONCAT_VECTORS(SDNode *N);
258 SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
259 SDValue visitVECTOR_SHUFFLE(SDNode *N);
261 SDValue XformToShuffleWithZero(SDNode *N);
262 SDValue ReassociateOps(unsigned Opc, SDLoc DL, SDValue LHS, SDValue RHS);
264 SDValue visitShiftByConstant(SDNode *N, unsigned Amt);
266 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
267 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
268 SDValue SimplifySelect(SDLoc DL, SDValue N0, SDValue N1, SDValue N2);
269 SDValue SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1, SDValue N2,
270 SDValue N3, ISD::CondCode CC,
271 bool NotExtCompare = false);
272 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
273 SDLoc DL, bool foldBooleans = true);
274 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
276 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
277 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
278 SDValue BuildSDIV(SDNode *N);
279 SDValue BuildUDIV(SDNode *N);
280 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
281 bool DemandHighBits = true);
282 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
283 SDNode *MatchRotatePosNeg(SDValue Shifted, SDValue Pos, SDValue Neg,
284 SDValue InnerPos, SDValue InnerNeg,
285 unsigned PosOpcode, unsigned NegOpcode,
287 SDNode *MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL);
288 SDValue ReduceLoadWidth(SDNode *N);
289 SDValue ReduceLoadOpStoreWidth(SDNode *N);
290 SDValue TransformFPLoadStorePair(SDNode *N);
291 SDValue reduceBuildVecExtToExtBuildVec(SDNode *N);
292 SDValue reduceBuildVecConvertToConvertBuildVec(SDNode *N);
294 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
296 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
297 /// looking for aliasing nodes and adding them to the Aliases vector.
298 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
299 SmallVectorImpl<SDValue> &Aliases);
301 /// isAlias - Return true if there is any possibility that the two addresses
303 bool isAlias(SDValue Ptr1, int64_t Size1, bool IsVolatile1,
304 const Value *SrcValue1, int SrcValueOffset1,
305 unsigned SrcValueAlign1,
306 const MDNode *TBAAInfo1,
307 SDValue Ptr2, int64_t Size2, bool IsVolatile2,
308 const Value *SrcValue2, int SrcValueOffset2,
309 unsigned SrcValueAlign2,
310 const MDNode *TBAAInfo2) const;
312 /// isAlias - Return true if there is any possibility that the two addresses
314 bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1);
316 /// FindAliasInfo - Extracts the relevant alias information from the memory
317 /// node. Returns true if the operand was a load.
318 bool FindAliasInfo(SDNode *N,
319 SDValue &Ptr, int64_t &Size, bool &IsVolatile,
320 const Value *&SrcValue, int &SrcValueOffset,
321 unsigned &SrcValueAlignment,
322 const MDNode *&TBAAInfo) const;
324 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes,
325 /// looking for a better chain (aliasing node.)
326 SDValue FindBetterChain(SDNode *N, SDValue Chain);
328 /// Merge consecutive store operations into a wide store.
329 /// This optimization uses wide integers or vectors when possible.
330 /// \return True if some memory operations were changed.
331 bool MergeConsecutiveStores(StoreSDNode *N);
334 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
335 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
336 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {
337 AttributeSet FnAttrs =
338 DAG.getMachineFunction().getFunction()->getAttributes();
340 FnAttrs.hasAttribute(AttributeSet::FunctionIndex,
341 Attribute::OptimizeForSize) ||
342 FnAttrs.hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize);
345 /// Run - runs the dag combiner on all nodes in the work list
346 void Run(CombineLevel AtLevel);
348 SelectionDAG &getDAG() const { return DAG; }
350 /// getShiftAmountTy - Returns a type large enough to hold any valid
351 /// shift amount - before type legalization these can be huge.
352 EVT getShiftAmountTy(EVT LHSTy) {
353 assert(LHSTy.isInteger() && "Shift amount is not an integer type!");
354 if (LHSTy.isVector())
356 return LegalTypes ? TLI.getScalarShiftAmountTy(LHSTy)
357 : TLI.getPointerTy();
360 /// isTypeLegal - This method returns true if we are running before type
361 /// legalization or if the specified VT is legal.
362 bool isTypeLegal(const EVT &VT) {
363 if (!LegalTypes) return true;
364 return TLI.isTypeLegal(VT);
367 /// getSetCCResultType - Convenience wrapper around
368 /// TargetLowering::getSetCCResultType
369 EVT getSetCCResultType(EVT VT) const {
370 return TLI.getSetCCResultType(*DAG.getContext(), VT);
377 /// WorkListRemover - This class is a DAGUpdateListener that removes any deleted
378 /// nodes from the worklist.
379 class WorkListRemover : public SelectionDAG::DAGUpdateListener {
382 explicit WorkListRemover(DAGCombiner &dc)
383 : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {}
385 virtual void NodeDeleted(SDNode *N, SDNode *E) {
386 DC.removeFromWorkList(N);
391 //===----------------------------------------------------------------------===//
392 // TargetLowering::DAGCombinerInfo implementation
393 //===----------------------------------------------------------------------===//
395 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
396 ((DAGCombiner*)DC)->AddToWorkList(N);
399 void TargetLowering::DAGCombinerInfo::RemoveFromWorklist(SDNode *N) {
400 ((DAGCombiner*)DC)->removeFromWorkList(N);
403 SDValue TargetLowering::DAGCombinerInfo::
404 CombineTo(SDNode *N, const std::vector<SDValue> &To, bool AddTo) {
405 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
408 SDValue TargetLowering::DAGCombinerInfo::
409 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
410 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
414 SDValue TargetLowering::DAGCombinerInfo::
415 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
416 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
419 void TargetLowering::DAGCombinerInfo::
420 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
421 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
424 //===----------------------------------------------------------------------===//
426 //===----------------------------------------------------------------------===//
428 /// isNegatibleForFree - Return 1 if we can compute the negated form of the
429 /// specified expression for the same cost as the expression itself, or 2 if we
430 /// can compute the negated form more cheaply than the expression itself.
431 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
432 const TargetLowering &TLI,
433 const TargetOptions *Options,
434 unsigned Depth = 0) {
435 // fneg is removable even if it has multiple uses.
436 if (Op.getOpcode() == ISD::FNEG) return 2;
438 // Don't allow anything with multiple uses.
439 if (!Op.hasOneUse()) return 0;
441 // Don't recurse exponentially.
442 if (Depth > 6) return 0;
444 switch (Op.getOpcode()) {
445 default: return false;
446 case ISD::ConstantFP:
447 // Don't invert constant FP values after legalize. The negated constant
448 // isn't necessarily legal.
449 return LegalOperations ? 0 : 1;
451 // FIXME: determine better conditions for this xform.
452 if (!Options->UnsafeFPMath) return 0;
454 // After operation legalization, it might not be legal to create new FSUBs.
455 if (LegalOperations &&
456 !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType()))
459 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
460 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
463 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
464 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
467 // We can't turn -(A-B) into B-A when we honor signed zeros.
468 if (!Options->UnsafeFPMath) return 0;
470 // fold (fneg (fsub A, B)) -> (fsub B, A)
475 if (Options->HonorSignDependentRoundingFPMath()) return 0;
477 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
478 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
482 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
488 return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options,
493 /// GetNegatedExpression - If isNegatibleForFree returns true, this function
494 /// returns the newly negated expression.
495 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
496 bool LegalOperations, unsigned Depth = 0) {
497 // fneg is removable even if it has multiple uses.
498 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
500 // Don't allow anything with multiple uses.
501 assert(Op.hasOneUse() && "Unknown reuse!");
503 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
504 switch (Op.getOpcode()) {
505 default: llvm_unreachable("Unknown code");
506 case ISD::ConstantFP: {
507 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
509 return DAG.getConstantFP(V, Op.getValueType());
512 // FIXME: determine better conditions for this xform.
513 assert(DAG.getTarget().Options.UnsafeFPMath);
515 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
516 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
517 DAG.getTargetLoweringInfo(),
518 &DAG.getTarget().Options, Depth+1))
519 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
520 GetNegatedExpression(Op.getOperand(0), DAG,
521 LegalOperations, Depth+1),
523 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
524 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
525 GetNegatedExpression(Op.getOperand(1), DAG,
526 LegalOperations, Depth+1),
529 // We can't turn -(A-B) into B-A when we honor signed zeros.
530 assert(DAG.getTarget().Options.UnsafeFPMath);
532 // fold (fneg (fsub 0, B)) -> B
533 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
534 if (N0CFP->getValueAPF().isZero())
535 return Op.getOperand(1);
537 // fold (fneg (fsub A, B)) -> (fsub B, A)
538 return DAG.getNode(ISD::FSUB, SDLoc(Op), Op.getValueType(),
539 Op.getOperand(1), Op.getOperand(0));
543 assert(!DAG.getTarget().Options.HonorSignDependentRoundingFPMath());
545 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
546 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
547 DAG.getTargetLoweringInfo(),
548 &DAG.getTarget().Options, Depth+1))
549 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
550 GetNegatedExpression(Op.getOperand(0), DAG,
551 LegalOperations, Depth+1),
554 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
555 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
557 GetNegatedExpression(Op.getOperand(1), DAG,
558 LegalOperations, Depth+1));
562 return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(),
563 GetNegatedExpression(Op.getOperand(0), DAG,
564 LegalOperations, Depth+1));
566 return DAG.getNode(ISD::FP_ROUND, SDLoc(Op), Op.getValueType(),
567 GetNegatedExpression(Op.getOperand(0), DAG,
568 LegalOperations, Depth+1),
574 // isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
575 // that selects between the values 1 and 0, making it equivalent to a setcc.
576 // Also, set the incoming LHS, RHS, and CC references to the appropriate
577 // nodes based on the type of node we are checking. This simplifies life a
578 // bit for the callers.
579 static bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
581 if (N.getOpcode() == ISD::SETCC) {
582 LHS = N.getOperand(0);
583 RHS = N.getOperand(1);
584 CC = N.getOperand(2);
587 if (N.getOpcode() == ISD::SELECT_CC &&
588 N.getOperand(2).getOpcode() == ISD::Constant &&
589 N.getOperand(3).getOpcode() == ISD::Constant &&
590 cast<ConstantSDNode>(N.getOperand(2))->getAPIntValue() == 1 &&
591 cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) {
592 LHS = N.getOperand(0);
593 RHS = N.getOperand(1);
594 CC = N.getOperand(4);
600 // isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only
601 // one use. If this is true, it allows the users to invert the operation for
602 // free when it is profitable to do so.
603 static bool isOneUseSetCC(SDValue N) {
605 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
610 SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL,
611 SDValue N0, SDValue N1) {
612 EVT VT = N0.getValueType();
613 if (N0.getOpcode() == Opc && isa<ConstantSDNode>(N0.getOperand(1))) {
614 if (isa<ConstantSDNode>(N1)) {
615 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
617 DAG.FoldConstantArithmetic(Opc, VT,
618 cast<ConstantSDNode>(N0.getOperand(1)),
619 cast<ConstantSDNode>(N1));
620 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
622 if (N0.hasOneUse()) {
623 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one use
624 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT,
625 N0.getOperand(0), N1);
626 AddToWorkList(OpNode.getNode());
627 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
631 if (N1.getOpcode() == Opc && isa<ConstantSDNode>(N1.getOperand(1))) {
632 if (isa<ConstantSDNode>(N0)) {
633 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
635 DAG.FoldConstantArithmetic(Opc, VT,
636 cast<ConstantSDNode>(N1.getOperand(1)),
637 cast<ConstantSDNode>(N0));
638 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
640 if (N1.hasOneUse()) {
641 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one use
642 SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT,
643 N1.getOperand(0), N0);
644 AddToWorkList(OpNode.getNode());
645 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
652 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
654 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
656 DEBUG(dbgs() << "\nReplacing.1 ";
658 dbgs() << "\nWith: ";
659 To[0].getNode()->dump(&DAG);
660 dbgs() << " and " << NumTo-1 << " other values\n";
661 for (unsigned i = 0, e = NumTo; i != e; ++i)
662 assert((!To[i].getNode() ||
663 N->getValueType(i) == To[i].getValueType()) &&
664 "Cannot combine value to value of different type!"));
665 WorkListRemover DeadNodes(*this);
666 DAG.ReplaceAllUsesWith(N, To);
668 // Push the new nodes and any users onto the worklist
669 for (unsigned i = 0, e = NumTo; i != e; ++i) {
670 if (To[i].getNode()) {
671 AddToWorkList(To[i].getNode());
672 AddUsersToWorkList(To[i].getNode());
677 // Finally, if the node is now dead, remove it from the graph. The node
678 // may not be dead if the replacement process recursively simplified to
679 // something else needing this node.
680 if (N->use_empty()) {
681 // Nodes can be reintroduced into the worklist. Make sure we do not
682 // process a node that has been replaced.
683 removeFromWorkList(N);
685 // Finally, since the node is now dead, remove it from the graph.
688 return SDValue(N, 0);
692 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
693 // Replace all uses. If any nodes become isomorphic to other nodes and
694 // are deleted, make sure to remove them from our worklist.
695 WorkListRemover DeadNodes(*this);
696 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New);
698 // Push the new node and any (possibly new) users onto the worklist.
699 AddToWorkList(TLO.New.getNode());
700 AddUsersToWorkList(TLO.New.getNode());
702 // Finally, if the node is now dead, remove it from the graph. The node
703 // may not be dead if the replacement process recursively simplified to
704 // something else needing this node.
705 if (TLO.Old.getNode()->use_empty()) {
706 removeFromWorkList(TLO.Old.getNode());
708 // If the operands of this node are only used by the node, they will now
709 // be dead. Make sure to visit them first to delete dead nodes early.
710 for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands(); i != e; ++i)
711 if (TLO.Old.getNode()->getOperand(i).getNode()->hasOneUse())
712 AddToWorkList(TLO.Old.getNode()->getOperand(i).getNode());
714 DAG.DeleteNode(TLO.Old.getNode());
718 /// SimplifyDemandedBits - Check the specified integer node value to see if
719 /// it can be simplified or if things it uses can be simplified by bit
720 /// propagation. If so, return true.
721 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
722 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
723 APInt KnownZero, KnownOne;
724 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
728 AddToWorkList(Op.getNode());
730 // Replace the old value with the new one.
732 DEBUG(dbgs() << "\nReplacing.2 ";
733 TLO.Old.getNode()->dump(&DAG);
734 dbgs() << "\nWith: ";
735 TLO.New.getNode()->dump(&DAG);
738 CommitTargetLoweringOpt(TLO);
742 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
744 EVT VT = Load->getValueType(0);
745 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
747 DEBUG(dbgs() << "\nReplacing.9 ";
749 dbgs() << "\nWith: ";
750 Trunc.getNode()->dump(&DAG);
752 WorkListRemover DeadNodes(*this);
753 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc);
754 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1));
755 removeFromWorkList(Load);
756 DAG.DeleteNode(Load);
757 AddToWorkList(Trunc.getNode());
760 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
763 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
764 EVT MemVT = LD->getMemoryVT();
765 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
766 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
768 : LD->getExtensionType();
770 return DAG.getExtLoad(ExtType, dl, PVT,
771 LD->getChain(), LD->getBasePtr(),
772 MemVT, LD->getMemOperand());
775 unsigned Opc = Op.getOpcode();
778 case ISD::AssertSext:
779 return DAG.getNode(ISD::AssertSext, dl, PVT,
780 SExtPromoteOperand(Op.getOperand(0), PVT),
782 case ISD::AssertZext:
783 return DAG.getNode(ISD::AssertZext, dl, PVT,
784 ZExtPromoteOperand(Op.getOperand(0), PVT),
786 case ISD::Constant: {
788 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
789 return DAG.getNode(ExtOpc, dl, PVT, Op);
793 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
795 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
798 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
799 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
801 EVT OldVT = Op.getValueType();
803 bool Replace = false;
804 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
805 if (NewOp.getNode() == 0)
807 AddToWorkList(NewOp.getNode());
810 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
811 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
812 DAG.getValueType(OldVT));
815 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
816 EVT OldVT = Op.getValueType();
818 bool Replace = false;
819 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
820 if (NewOp.getNode() == 0)
822 AddToWorkList(NewOp.getNode());
825 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
826 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
829 /// PromoteIntBinOp - Promote the specified integer binary operation if the
830 /// target indicates it is beneficial. e.g. On x86, it's usually better to
831 /// promote i16 operations to i32 since i16 instructions are longer.
832 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
833 if (!LegalOperations)
836 EVT VT = Op.getValueType();
837 if (VT.isVector() || !VT.isInteger())
840 // If operation type is 'undesirable', e.g. i16 on x86, consider
842 unsigned Opc = Op.getOpcode();
843 if (TLI.isTypeDesirableForOp(Opc, VT))
847 // Consult target whether it is a good idea to promote this operation and
848 // what's the right type to promote it to.
849 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
850 assert(PVT != VT && "Don't know what type to promote to!");
852 bool Replace0 = false;
853 SDValue N0 = Op.getOperand(0);
854 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
855 if (NN0.getNode() == 0)
858 bool Replace1 = false;
859 SDValue N1 = Op.getOperand(1);
864 NN1 = PromoteOperand(N1, PVT, Replace1);
865 if (NN1.getNode() == 0)
869 AddToWorkList(NN0.getNode());
871 AddToWorkList(NN1.getNode());
874 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
876 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
878 DEBUG(dbgs() << "\nPromoting ";
879 Op.getNode()->dump(&DAG));
881 return DAG.getNode(ISD::TRUNCATE, dl, VT,
882 DAG.getNode(Opc, dl, PVT, NN0, NN1));
887 /// PromoteIntShiftOp - Promote the specified integer shift operation if the
888 /// target indicates it is beneficial. e.g. On x86, it's usually better to
889 /// promote i16 operations to i32 since i16 instructions are longer.
890 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
891 if (!LegalOperations)
894 EVT VT = Op.getValueType();
895 if (VT.isVector() || !VT.isInteger())
898 // If operation type is 'undesirable', e.g. i16 on x86, consider
900 unsigned Opc = Op.getOpcode();
901 if (TLI.isTypeDesirableForOp(Opc, VT))
905 // Consult target whether it is a good idea to promote this operation and
906 // what's the right type to promote it to.
907 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
908 assert(PVT != VT && "Don't know what type to promote to!");
910 bool Replace = false;
911 SDValue N0 = Op.getOperand(0);
913 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
914 else if (Opc == ISD::SRL)
915 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
917 N0 = PromoteOperand(N0, PVT, Replace);
918 if (N0.getNode() == 0)
921 AddToWorkList(N0.getNode());
923 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
925 DEBUG(dbgs() << "\nPromoting ";
926 Op.getNode()->dump(&DAG));
928 return DAG.getNode(ISD::TRUNCATE, dl, VT,
929 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
934 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
935 if (!LegalOperations)
938 EVT VT = Op.getValueType();
939 if (VT.isVector() || !VT.isInteger())
942 // If operation type is 'undesirable', e.g. i16 on x86, consider
944 unsigned Opc = Op.getOpcode();
945 if (TLI.isTypeDesirableForOp(Opc, VT))
949 // Consult target whether it is a good idea to promote this operation and
950 // what's the right type to promote it to.
951 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
952 assert(PVT != VT && "Don't know what type to promote to!");
953 // fold (aext (aext x)) -> (aext x)
954 // fold (aext (zext x)) -> (zext x)
955 // fold (aext (sext x)) -> (sext x)
956 DEBUG(dbgs() << "\nPromoting ";
957 Op.getNode()->dump(&DAG));
958 return DAG.getNode(Op.getOpcode(), SDLoc(Op), VT, Op.getOperand(0));
963 bool DAGCombiner::PromoteLoad(SDValue Op) {
964 if (!LegalOperations)
967 EVT VT = Op.getValueType();
968 if (VT.isVector() || !VT.isInteger())
971 // If operation type is 'undesirable', e.g. i16 on x86, consider
973 unsigned Opc = Op.getOpcode();
974 if (TLI.isTypeDesirableForOp(Opc, VT))
978 // Consult target whether it is a good idea to promote this operation and
979 // what's the right type to promote it to.
980 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
981 assert(PVT != VT && "Don't know what type to promote to!");
984 SDNode *N = Op.getNode();
985 LoadSDNode *LD = cast<LoadSDNode>(N);
986 EVT MemVT = LD->getMemoryVT();
987 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
988 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
990 : LD->getExtensionType();
991 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
992 LD->getChain(), LD->getBasePtr(),
993 MemVT, LD->getMemOperand());
994 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
996 DEBUG(dbgs() << "\nPromoting ";
999 Result.getNode()->dump(&DAG);
1001 WorkListRemover DeadNodes(*this);
1002 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
1003 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1));
1004 removeFromWorkList(N);
1006 AddToWorkList(Result.getNode());
1013 //===----------------------------------------------------------------------===//
1014 // Main DAG Combiner implementation
1015 //===----------------------------------------------------------------------===//
1017 void DAGCombiner::Run(CombineLevel AtLevel) {
1018 // set the instance variables, so that the various visit routines may use it.
1020 LegalOperations = Level >= AfterLegalizeVectorOps;
1021 LegalTypes = Level >= AfterLegalizeTypes;
1023 // Add all the dag nodes to the worklist.
1024 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
1025 E = DAG.allnodes_end(); I != E; ++I)
1028 // Create a dummy node (which is not added to allnodes), that adds a reference
1029 // to the root node, preventing it from being deleted, and tracking any
1030 // changes of the root.
1031 HandleSDNode Dummy(DAG.getRoot());
1033 // The root of the dag may dangle to deleted nodes until the dag combiner is
1034 // done. Set it to null to avoid confusion.
1035 DAG.setRoot(SDValue());
1037 // while the worklist isn't empty, find a node and
1038 // try and combine it.
1039 while (!WorkListContents.empty()) {
1041 // The WorkListOrder holds the SDNodes in order, but it may contain
1043 // In order to avoid a linear scan, we use a set (O(log N)) to hold what the
1044 // worklist *should* contain, and check the node we want to visit is should
1045 // actually be visited.
1047 N = WorkListOrder.pop_back_val();
1048 } while (!WorkListContents.erase(N));
1050 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1051 // N is deleted from the DAG, since they too may now be dead or may have a
1052 // reduced number of uses, allowing other xforms.
1053 if (N->use_empty() && N != &Dummy) {
1054 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1055 AddToWorkList(N->getOperand(i).getNode());
1061 SDValue RV = combine(N);
1063 if (RV.getNode() == 0)
1068 // If we get back the same node we passed in, rather than a new node or
1069 // zero, we know that the node must have defined multiple values and
1070 // CombineTo was used. Since CombineTo takes care of the worklist
1071 // mechanics for us, we have no work to do in this case.
1072 if (RV.getNode() == N)
1075 assert(N->getOpcode() != ISD::DELETED_NODE &&
1076 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1077 "Node was deleted but visit returned new node!");
1079 DEBUG(dbgs() << "\nReplacing.3 ";
1081 dbgs() << "\nWith: ";
1082 RV.getNode()->dump(&DAG);
1085 // Transfer debug value.
1086 DAG.TransferDbgValues(SDValue(N, 0), RV);
1087 WorkListRemover DeadNodes(*this);
1088 if (N->getNumValues() == RV.getNode()->getNumValues())
1089 DAG.ReplaceAllUsesWith(N, RV.getNode());
1091 assert(N->getValueType(0) == RV.getValueType() &&
1092 N->getNumValues() == 1 && "Type mismatch");
1094 DAG.ReplaceAllUsesWith(N, &OpV);
1097 // Push the new node and any users onto the worklist
1098 AddToWorkList(RV.getNode());
1099 AddUsersToWorkList(RV.getNode());
1101 // Add any uses of the old node to the worklist in case this node is the
1102 // last one that uses them. They may become dead after this node is
1104 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1105 AddToWorkList(N->getOperand(i).getNode());
1107 // Finally, if the node is now dead, remove it from the graph. The node
1108 // may not be dead if the replacement process recursively simplified to
1109 // something else needing this node.
1110 if (N->use_empty()) {
1111 // Nodes can be reintroduced into the worklist. Make sure we do not
1112 // process a node that has been replaced.
1113 removeFromWorkList(N);
1115 // Finally, since the node is now dead, remove it from the graph.
1120 // If the root changed (e.g. it was a dead load, update the root).
1121 DAG.setRoot(Dummy.getValue());
1122 DAG.RemoveDeadNodes();
1125 SDValue DAGCombiner::visit(SDNode *N) {
1126 switch (N->getOpcode()) {
1128 case ISD::TokenFactor: return visitTokenFactor(N);
1129 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1130 case ISD::ADD: return visitADD(N);
1131 case ISD::SUB: return visitSUB(N);
1132 case ISD::ADDC: return visitADDC(N);
1133 case ISD::SUBC: return visitSUBC(N);
1134 case ISD::ADDE: return visitADDE(N);
1135 case ISD::SUBE: return visitSUBE(N);
1136 case ISD::MUL: return visitMUL(N);
1137 case ISD::SDIV: return visitSDIV(N);
1138 case ISD::UDIV: return visitUDIV(N);
1139 case ISD::SREM: return visitSREM(N);
1140 case ISD::UREM: return visitUREM(N);
1141 case ISD::MULHU: return visitMULHU(N);
1142 case ISD::MULHS: return visitMULHS(N);
1143 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1144 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1145 case ISD::SMULO: return visitSMULO(N);
1146 case ISD::UMULO: return visitUMULO(N);
1147 case ISD::SDIVREM: return visitSDIVREM(N);
1148 case ISD::UDIVREM: return visitUDIVREM(N);
1149 case ISD::AND: return visitAND(N);
1150 case ISD::OR: return visitOR(N);
1151 case ISD::XOR: return visitXOR(N);
1152 case ISD::SHL: return visitSHL(N);
1153 case ISD::SRA: return visitSRA(N);
1154 case ISD::SRL: return visitSRL(N);
1155 case ISD::CTLZ: return visitCTLZ(N);
1156 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1157 case ISD::CTTZ: return visitCTTZ(N);
1158 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1159 case ISD::CTPOP: return visitCTPOP(N);
1160 case ISD::SELECT: return visitSELECT(N);
1161 case ISD::VSELECT: return visitVSELECT(N);
1162 case ISD::SELECT_CC: return visitSELECT_CC(N);
1163 case ISD::SETCC: return visitSETCC(N);
1164 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1165 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1166 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1167 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1168 case ISD::TRUNCATE: return visitTRUNCATE(N);
1169 case ISD::BITCAST: return visitBITCAST(N);
1170 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1171 case ISD::FADD: return visitFADD(N);
1172 case ISD::FSUB: return visitFSUB(N);
1173 case ISD::FMUL: return visitFMUL(N);
1174 case ISD::FMA: return visitFMA(N);
1175 case ISD::FDIV: return visitFDIV(N);
1176 case ISD::FREM: return visitFREM(N);
1177 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1178 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1179 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1180 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1181 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1182 case ISD::FP_ROUND: return visitFP_ROUND(N);
1183 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1184 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1185 case ISD::FNEG: return visitFNEG(N);
1186 case ISD::FABS: return visitFABS(N);
1187 case ISD::FFLOOR: return visitFFLOOR(N);
1188 case ISD::FCEIL: return visitFCEIL(N);
1189 case ISD::FTRUNC: return visitFTRUNC(N);
1190 case ISD::BRCOND: return visitBRCOND(N);
1191 case ISD::BR_CC: return visitBR_CC(N);
1192 case ISD::LOAD: return visitLOAD(N);
1193 case ISD::STORE: return visitSTORE(N);
1194 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1195 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1196 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1197 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1198 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1199 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1204 SDValue DAGCombiner::combine(SDNode *N) {
1205 SDValue RV = visit(N);
1207 // If nothing happened, try a target-specific DAG combine.
1208 if (RV.getNode() == 0) {
1209 assert(N->getOpcode() != ISD::DELETED_NODE &&
1210 "Node was deleted but visit returned NULL!");
1212 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1213 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1215 // Expose the DAG combiner to the target combiner impls.
1216 TargetLowering::DAGCombinerInfo
1217 DagCombineInfo(DAG, Level, false, this);
1219 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1223 // If nothing happened still, try promoting the operation.
1224 if (RV.getNode() == 0) {
1225 switch (N->getOpcode()) {
1233 RV = PromoteIntBinOp(SDValue(N, 0));
1238 RV = PromoteIntShiftOp(SDValue(N, 0));
1240 case ISD::SIGN_EXTEND:
1241 case ISD::ZERO_EXTEND:
1242 case ISD::ANY_EXTEND:
1243 RV = PromoteExtend(SDValue(N, 0));
1246 if (PromoteLoad(SDValue(N, 0)))
1252 // If N is a commutative binary node, try commuting it to enable more
1254 if (RV.getNode() == 0 &&
1255 SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1256 N->getNumValues() == 1) {
1257 SDValue N0 = N->getOperand(0);
1258 SDValue N1 = N->getOperand(1);
1260 // Constant operands are canonicalized to RHS.
1261 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1262 SDValue Ops[] = { N1, N0 };
1263 SDNode *CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(),
1266 return SDValue(CSENode, 0);
1273 /// getInputChainForNode - Given a node, return its input chain if it has one,
1274 /// otherwise return a null sd operand.
1275 static SDValue getInputChainForNode(SDNode *N) {
1276 if (unsigned NumOps = N->getNumOperands()) {
1277 if (N->getOperand(0).getValueType() == MVT::Other)
1278 return N->getOperand(0);
1279 if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1280 return N->getOperand(NumOps-1);
1281 for (unsigned i = 1; i < NumOps-1; ++i)
1282 if (N->getOperand(i).getValueType() == MVT::Other)
1283 return N->getOperand(i);
1288 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1289 // If N has two operands, where one has an input chain equal to the other,
1290 // the 'other' chain is redundant.
1291 if (N->getNumOperands() == 2) {
1292 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1293 return N->getOperand(0);
1294 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1295 return N->getOperand(1);
1298 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1299 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1300 SmallPtrSet<SDNode*, 16> SeenOps;
1301 bool Changed = false; // If we should replace this token factor.
1303 // Start out with this token factor.
1306 // Iterate through token factors. The TFs grows when new token factors are
1308 for (unsigned i = 0; i < TFs.size(); ++i) {
1309 SDNode *TF = TFs[i];
1311 // Check each of the operands.
1312 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1313 SDValue Op = TF->getOperand(i);
1315 switch (Op.getOpcode()) {
1316 case ISD::EntryToken:
1317 // Entry tokens don't need to be added to the list. They are
1322 case ISD::TokenFactor:
1323 if (Op.hasOneUse() &&
1324 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1325 // Queue up for processing.
1326 TFs.push_back(Op.getNode());
1327 // Clean up in case the token factor is removed.
1328 AddToWorkList(Op.getNode());
1335 // Only add if it isn't already in the list.
1336 if (SeenOps.insert(Op.getNode()))
1347 // If we've change things around then replace token factor.
1350 // The entry token is the only possible outcome.
1351 Result = DAG.getEntryNode();
1353 // New and improved token factor.
1354 Result = DAG.getNode(ISD::TokenFactor, SDLoc(N),
1355 MVT::Other, &Ops[0], Ops.size());
1358 // Don't add users to work list.
1359 return CombineTo(N, Result, false);
1365 /// MERGE_VALUES can always be eliminated.
1366 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1367 WorkListRemover DeadNodes(*this);
1368 // Replacing results may cause a different MERGE_VALUES to suddenly
1369 // be CSE'd with N, and carry its uses with it. Iterate until no
1370 // uses remain, to ensure that the node can be safely deleted.
1371 // First add the users of this node to the work list so that they
1372 // can be tried again once they have new operands.
1373 AddUsersToWorkList(N);
1375 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1376 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1377 } while (!N->use_empty());
1378 removeFromWorkList(N);
1380 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1384 SDValue combineShlAddConstant(SDLoc DL, SDValue N0, SDValue N1,
1385 SelectionDAG &DAG) {
1386 EVT VT = N0.getValueType();
1387 SDValue N00 = N0.getOperand(0);
1388 SDValue N01 = N0.getOperand(1);
1389 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N01);
1391 if (N01C && N00.getOpcode() == ISD::ADD && N00.getNode()->hasOneUse() &&
1392 isa<ConstantSDNode>(N00.getOperand(1))) {
1393 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1394 N0 = DAG.getNode(ISD::ADD, SDLoc(N0), VT,
1395 DAG.getNode(ISD::SHL, SDLoc(N00), VT,
1396 N00.getOperand(0), N01),
1397 DAG.getNode(ISD::SHL, SDLoc(N01), VT,
1398 N00.getOperand(1), N01));
1399 return DAG.getNode(ISD::ADD, DL, VT, N0, N1);
1405 SDValue DAGCombiner::visitADD(SDNode *N) {
1406 SDValue N0 = N->getOperand(0);
1407 SDValue N1 = N->getOperand(1);
1408 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1409 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1410 EVT VT = N0.getValueType();
1413 if (VT.isVector()) {
1414 SDValue FoldedVOp = SimplifyVBinOp(N);
1415 if (FoldedVOp.getNode()) return FoldedVOp;
1417 // fold (add x, 0) -> x, vector edition
1418 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1420 if (ISD::isBuildVectorAllZeros(N0.getNode()))
1424 // fold (add x, undef) -> undef
1425 if (N0.getOpcode() == ISD::UNDEF)
1427 if (N1.getOpcode() == ISD::UNDEF)
1429 // fold (add c1, c2) -> c1+c2
1431 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1432 // canonicalize constant to RHS
1434 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N1, N0);
1435 // fold (add x, 0) -> x
1436 if (N1C && N1C->isNullValue())
1438 // fold (add Sym, c) -> Sym+c
1439 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1440 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1441 GA->getOpcode() == ISD::GlobalAddress)
1442 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1444 (uint64_t)N1C->getSExtValue());
1445 // fold ((c1-A)+c2) -> (c1+c2)-A
1446 if (N1C && N0.getOpcode() == ISD::SUB)
1447 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1448 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1449 DAG.getConstant(N1C->getAPIntValue()+
1450 N0C->getAPIntValue(), VT),
1453 SDValue RADD = ReassociateOps(ISD::ADD, SDLoc(N), N0, N1);
1454 if (RADD.getNode() != 0)
1456 // fold ((0-A) + B) -> B-A
1457 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1458 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1459 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1, N0.getOperand(1));
1460 // fold (A + (0-B)) -> A-B
1461 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1462 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1463 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1.getOperand(1));
1464 // fold (A+(B-A)) -> B
1465 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1466 return N1.getOperand(0);
1467 // fold ((B-A)+A) -> B
1468 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1469 return N0.getOperand(0);
1470 // fold (A+(B-(A+C))) to (B-C)
1471 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1472 N0 == N1.getOperand(1).getOperand(0))
1473 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1474 N1.getOperand(1).getOperand(1));
1475 // fold (A+(B-(C+A))) to (B-C)
1476 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1477 N0 == N1.getOperand(1).getOperand(1))
1478 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1.getOperand(0),
1479 N1.getOperand(1).getOperand(0));
1480 // fold (A+((B-A)+or-C)) to (B+or-C)
1481 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1482 N1.getOperand(0).getOpcode() == ISD::SUB &&
1483 N0 == N1.getOperand(0).getOperand(1))
1484 return DAG.getNode(N1.getOpcode(), SDLoc(N), VT,
1485 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1487 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1488 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1489 SDValue N00 = N0.getOperand(0);
1490 SDValue N01 = N0.getOperand(1);
1491 SDValue N10 = N1.getOperand(0);
1492 SDValue N11 = N1.getOperand(1);
1494 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1495 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1496 DAG.getNode(ISD::ADD, SDLoc(N0), VT, N00, N10),
1497 DAG.getNode(ISD::ADD, SDLoc(N1), VT, N01, N11));
1500 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1501 return SDValue(N, 0);
1503 // fold (a+b) -> (a|b) iff a and b share no bits.
1504 if (VT.isInteger() && !VT.isVector()) {
1505 APInt LHSZero, LHSOne;
1506 APInt RHSZero, RHSOne;
1507 DAG.ComputeMaskedBits(N0, LHSZero, LHSOne);
1509 if (LHSZero.getBoolValue()) {
1510 DAG.ComputeMaskedBits(N1, RHSZero, RHSOne);
1512 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1513 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1514 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1515 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1);
1519 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1520 if (N0.getOpcode() == ISD::SHL && N0.getNode()->hasOneUse()) {
1521 SDValue Result = combineShlAddConstant(SDLoc(N), N0, N1, DAG);
1522 if (Result.getNode()) return Result;
1524 if (N1.getOpcode() == ISD::SHL && N1.getNode()->hasOneUse()) {
1525 SDValue Result = combineShlAddConstant(SDLoc(N), N1, N0, DAG);
1526 if (Result.getNode()) return Result;
1529 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1530 if (N1.getOpcode() == ISD::SHL &&
1531 N1.getOperand(0).getOpcode() == ISD::SUB)
1532 if (ConstantSDNode *C =
1533 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1534 if (C->getAPIntValue() == 0)
1535 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N0,
1536 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1537 N1.getOperand(0).getOperand(1),
1539 if (N0.getOpcode() == ISD::SHL &&
1540 N0.getOperand(0).getOpcode() == ISD::SUB)
1541 if (ConstantSDNode *C =
1542 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1543 if (C->getAPIntValue() == 0)
1544 return DAG.getNode(ISD::SUB, SDLoc(N), VT, N1,
1545 DAG.getNode(ISD::SHL, SDLoc(N), VT,
1546 N0.getOperand(0).getOperand(1),
1549 if (N1.getOpcode() == ISD::AND) {
1550 SDValue AndOp0 = N1.getOperand(0);
1551 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1552 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1553 unsigned DestBits = VT.getScalarType().getSizeInBits();
1555 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1556 // and similar xforms where the inner op is either ~0 or 0.
1557 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1559 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1563 // add (sext i1), X -> sub X, (zext i1)
1564 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1565 N0.getOperand(0).getValueType() == MVT::i1 &&
1566 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1568 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1569 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1575 SDValue DAGCombiner::visitADDC(SDNode *N) {
1576 SDValue N0 = N->getOperand(0);
1577 SDValue N1 = N->getOperand(1);
1578 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1579 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1580 EVT VT = N0.getValueType();
1582 // If the flag result is dead, turn this into an ADD.
1583 if (!N->hasAnyUseOfValue(1))
1584 return CombineTo(N, DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N1),
1585 DAG.getNode(ISD::CARRY_FALSE,
1586 SDLoc(N), MVT::Glue));
1588 // canonicalize constant to RHS.
1590 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N1, N0);
1592 // fold (addc x, 0) -> x + no carry out
1593 if (N1C && N1C->isNullValue())
1594 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1595 SDLoc(N), MVT::Glue));
1597 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1598 APInt LHSZero, LHSOne;
1599 APInt RHSZero, RHSOne;
1600 DAG.ComputeMaskedBits(N0, LHSZero, LHSOne);
1602 if (LHSZero.getBoolValue()) {
1603 DAG.ComputeMaskedBits(N1, RHSZero, RHSOne);
1605 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1606 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1607 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1608 return CombineTo(N, DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N1),
1609 DAG.getNode(ISD::CARRY_FALSE,
1610 SDLoc(N), MVT::Glue));
1616 SDValue DAGCombiner::visitADDE(SDNode *N) {
1617 SDValue N0 = N->getOperand(0);
1618 SDValue N1 = N->getOperand(1);
1619 SDValue CarryIn = N->getOperand(2);
1620 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1621 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1623 // canonicalize constant to RHS
1625 return DAG.getNode(ISD::ADDE, SDLoc(N), N->getVTList(),
1628 // fold (adde x, y, false) -> (addc x, y)
1629 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1630 return DAG.getNode(ISD::ADDC, SDLoc(N), N->getVTList(), N0, N1);
1635 // Since it may not be valid to emit a fold to zero for vector initializers
1636 // check if we can before folding.
1637 static SDValue tryFoldToZero(SDLoc DL, const TargetLowering &TLI, EVT VT,
1639 bool LegalOperations, bool LegalTypes) {
1641 return DAG.getConstant(0, VT);
1642 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
1643 return DAG.getConstant(0, VT);
1647 SDValue DAGCombiner::visitSUB(SDNode *N) {
1648 SDValue N0 = N->getOperand(0);
1649 SDValue N1 = N->getOperand(1);
1650 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1651 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1652 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? 0 :
1653 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1654 EVT VT = N0.getValueType();
1657 if (VT.isVector()) {
1658 SDValue FoldedVOp = SimplifyVBinOp(N);
1659 if (FoldedVOp.getNode()) return FoldedVOp;
1661 // fold (sub x, 0) -> x, vector edition
1662 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1666 // fold (sub x, x) -> 0
1667 // FIXME: Refactor this and xor and other similar operations together.
1669 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
1670 // fold (sub c1, c2) -> c1-c2
1672 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1673 // fold (sub x, c) -> (add x, -c)
1675 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0,
1676 DAG.getConstant(-N1C->getAPIntValue(), VT));
1677 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1678 if (N0C && N0C->isAllOnesValue())
1679 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
1680 // fold A-(A-B) -> B
1681 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1682 return N1.getOperand(1);
1683 // fold (A+B)-A -> B
1684 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1685 return N0.getOperand(1);
1686 // fold (A+B)-B -> A
1687 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1688 return N0.getOperand(0);
1689 // fold C2-(A+C1) -> (C2-C1)-A
1690 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1691 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(),
1693 return DAG.getNode(ISD::SUB, SDLoc(N), VT, NewC,
1696 // fold ((A+(B+or-C))-B) -> A+or-C
1697 if (N0.getOpcode() == ISD::ADD &&
1698 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1699 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1700 N0.getOperand(1).getOperand(0) == N1)
1701 return DAG.getNode(N0.getOperand(1).getOpcode(), SDLoc(N), VT,
1702 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1703 // fold ((A+(C+B))-B) -> A+C
1704 if (N0.getOpcode() == ISD::ADD &&
1705 N0.getOperand(1).getOpcode() == ISD::ADD &&
1706 N0.getOperand(1).getOperand(1) == N1)
1707 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1708 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1709 // fold ((A-(B-C))-C) -> A-B
1710 if (N0.getOpcode() == ISD::SUB &&
1711 N0.getOperand(1).getOpcode() == ISD::SUB &&
1712 N0.getOperand(1).getOperand(1) == N1)
1713 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1714 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1716 // If either operand of a sub is undef, the result is undef
1717 if (N0.getOpcode() == ISD::UNDEF)
1719 if (N1.getOpcode() == ISD::UNDEF)
1722 // If the relocation model supports it, consider symbol offsets.
1723 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1724 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1725 // fold (sub Sym, c) -> Sym-c
1726 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1727 return DAG.getGlobalAddress(GA->getGlobal(), SDLoc(N1C), VT,
1729 (uint64_t)N1C->getSExtValue());
1730 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1731 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1732 if (GA->getGlobal() == GB->getGlobal())
1733 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1740 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1741 SDValue N0 = N->getOperand(0);
1742 SDValue N1 = N->getOperand(1);
1743 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1744 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1745 EVT VT = N0.getValueType();
1747 // If the flag result is dead, turn this into an SUB.
1748 if (!N->hasAnyUseOfValue(1))
1749 return CombineTo(N, DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, N1),
1750 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1753 // fold (subc x, x) -> 0 + no borrow
1755 return CombineTo(N, DAG.getConstant(0, VT),
1756 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1759 // fold (subc x, 0) -> x + no borrow
1760 if (N1C && N1C->isNullValue())
1761 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1764 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1765 if (N0C && N0C->isAllOnesValue())
1766 return CombineTo(N, DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0),
1767 DAG.getNode(ISD::CARRY_FALSE, SDLoc(N),
1773 SDValue DAGCombiner::visitSUBE(SDNode *N) {
1774 SDValue N0 = N->getOperand(0);
1775 SDValue N1 = N->getOperand(1);
1776 SDValue CarryIn = N->getOperand(2);
1778 // fold (sube x, y, false) -> (subc x, y)
1779 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1780 return DAG.getNode(ISD::SUBC, SDLoc(N), N->getVTList(), N0, N1);
1785 /// isConstantSplatVector - Returns true if N is a BUILD_VECTOR node whose
1786 /// elements are all the same constant or undefined.
1787 static bool isConstantSplatVector(SDNode *N, APInt& SplatValue) {
1788 BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(N);
1793 unsigned SplatBitSize;
1795 EVT EltVT = N->getValueType(0).getVectorElementType();
1796 return (C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
1798 EltVT.getSizeInBits() >= SplatBitSize);
1801 SDValue DAGCombiner::visitMUL(SDNode *N) {
1802 SDValue N0 = N->getOperand(0);
1803 SDValue N1 = N->getOperand(1);
1804 EVT VT = N0.getValueType();
1806 // fold (mul x, undef) -> 0
1807 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1808 return DAG.getConstant(0, VT);
1810 bool N0IsConst = false;
1811 bool N1IsConst = false;
1812 APInt ConstValue0, ConstValue1;
1814 if (VT.isVector()) {
1815 SDValue FoldedVOp = SimplifyVBinOp(N);
1816 if (FoldedVOp.getNode()) return FoldedVOp;
1818 N0IsConst = isConstantSplatVector(N0.getNode(), ConstValue0);
1819 N1IsConst = isConstantSplatVector(N1.getNode(), ConstValue1);
1821 N0IsConst = dyn_cast<ConstantSDNode>(N0) != 0;
1822 ConstValue0 = N0IsConst ? (dyn_cast<ConstantSDNode>(N0))->getAPIntValue()
1824 N1IsConst = dyn_cast<ConstantSDNode>(N1) != 0;
1825 ConstValue1 = N1IsConst ? (dyn_cast<ConstantSDNode>(N1))->getAPIntValue()
1829 // fold (mul c1, c2) -> c1*c2
1830 if (N0IsConst && N1IsConst)
1831 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0.getNode(), N1.getNode());
1833 // canonicalize constant to RHS
1834 if (N0IsConst && !N1IsConst)
1835 return DAG.getNode(ISD::MUL, SDLoc(N), VT, N1, N0);
1836 // fold (mul x, 0) -> 0
1837 if (N1IsConst && ConstValue1 == 0)
1839 // We require a splat of the entire scalar bit width for non-contiguous
1842 ConstValue1.getBitWidth() == VT.getScalarType().getSizeInBits();
1843 // fold (mul x, 1) -> x
1844 if (N1IsConst && ConstValue1 == 1 && IsFullSplat)
1846 // fold (mul x, -1) -> 0-x
1847 if (N1IsConst && ConstValue1.isAllOnesValue())
1848 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1849 DAG.getConstant(0, VT), N0);
1850 // fold (mul x, (1 << c)) -> x << c
1851 if (N1IsConst && ConstValue1.isPowerOf2() && IsFullSplat)
1852 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
1853 DAG.getConstant(ConstValue1.logBase2(),
1854 getShiftAmountTy(N0.getValueType())));
1855 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
1856 if (N1IsConst && (-ConstValue1).isPowerOf2() && IsFullSplat) {
1857 unsigned Log2Val = (-ConstValue1).logBase2();
1858 // FIXME: If the input is something that is easily negated (e.g. a
1859 // single-use add), we should put the negate there.
1860 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1861 DAG.getConstant(0, VT),
1862 DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
1863 DAG.getConstant(Log2Val,
1864 getShiftAmountTy(N0.getValueType()))));
1868 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
1869 if (N1IsConst && N0.getOpcode() == ISD::SHL &&
1870 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1871 isa<ConstantSDNode>(N0.getOperand(1)))) {
1872 SDValue C3 = DAG.getNode(ISD::SHL, SDLoc(N), VT,
1873 N1, N0.getOperand(1));
1874 AddToWorkList(C3.getNode());
1875 return DAG.getNode(ISD::MUL, SDLoc(N), VT,
1876 N0.getOperand(0), C3);
1879 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
1882 SDValue Sh(0,0), Y(0,0);
1883 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
1884 if (N0.getOpcode() == ISD::SHL &&
1885 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1886 isa<ConstantSDNode>(N0.getOperand(1))) &&
1887 N0.getNode()->hasOneUse()) {
1889 } else if (N1.getOpcode() == ISD::SHL &&
1890 isa<ConstantSDNode>(N1.getOperand(1)) &&
1891 N1.getNode()->hasOneUse()) {
1896 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
1897 Sh.getOperand(0), Y);
1898 return DAG.getNode(ISD::SHL, SDLoc(N), VT,
1899 Mul, Sh.getOperand(1));
1903 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
1904 if (N1IsConst && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
1905 (isConstantSplatVector(N0.getOperand(1).getNode(), Val) ||
1906 isa<ConstantSDNode>(N0.getOperand(1))))
1907 return DAG.getNode(ISD::ADD, SDLoc(N), VT,
1908 DAG.getNode(ISD::MUL, SDLoc(N0), VT,
1909 N0.getOperand(0), N1),
1910 DAG.getNode(ISD::MUL, SDLoc(N1), VT,
1911 N0.getOperand(1), N1));
1914 SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1);
1915 if (RMUL.getNode() != 0)
1921 SDValue DAGCombiner::visitSDIV(SDNode *N) {
1922 SDValue N0 = N->getOperand(0);
1923 SDValue N1 = N->getOperand(1);
1924 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1925 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1926 EVT VT = N->getValueType(0);
1929 if (VT.isVector()) {
1930 SDValue FoldedVOp = SimplifyVBinOp(N);
1931 if (FoldedVOp.getNode()) return FoldedVOp;
1934 // fold (sdiv c1, c2) -> c1/c2
1935 if (N0C && N1C && !N1C->isNullValue())
1936 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
1937 // fold (sdiv X, 1) -> X
1938 if (N1C && N1C->getAPIntValue() == 1LL)
1940 // fold (sdiv X, -1) -> 0-X
1941 if (N1C && N1C->isAllOnesValue())
1942 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1943 DAG.getConstant(0, VT), N0);
1944 // If we know the sign bits of both operands are zero, strength reduce to a
1945 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
1946 if (!VT.isVector()) {
1947 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
1948 return DAG.getNode(ISD::UDIV, SDLoc(N), N1.getValueType(),
1951 // fold (sdiv X, pow2) -> simple ops after legalize
1952 if (N1C && !N1C->isNullValue() &&
1953 (N1C->getAPIntValue().isPowerOf2() ||
1954 (-N1C->getAPIntValue()).isPowerOf2())) {
1955 // If dividing by powers of two is cheap, then don't perform the following
1957 if (TLI.isPow2DivCheap())
1960 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
1962 // Splat the sign bit into the register
1963 SDValue SGN = DAG.getNode(ISD::SRA, SDLoc(N), VT, N0,
1964 DAG.getConstant(VT.getSizeInBits()-1,
1965 getShiftAmountTy(N0.getValueType())));
1966 AddToWorkList(SGN.getNode());
1968 // Add (N0 < 0) ? abs2 - 1 : 0;
1969 SDValue SRL = DAG.getNode(ISD::SRL, SDLoc(N), VT, SGN,
1970 DAG.getConstant(VT.getSizeInBits() - lg2,
1971 getShiftAmountTy(SGN.getValueType())));
1972 SDValue ADD = DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, SRL);
1973 AddToWorkList(SRL.getNode());
1974 AddToWorkList(ADD.getNode()); // Divide by pow2
1975 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), VT, ADD,
1976 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType())));
1978 // If we're dividing by a positive value, we're done. Otherwise, we must
1979 // negate the result.
1980 if (N1C->getAPIntValue().isNonNegative())
1983 AddToWorkList(SRA.getNode());
1984 return DAG.getNode(ISD::SUB, SDLoc(N), VT,
1985 DAG.getConstant(0, VT), SRA);
1988 // if integer divide is expensive and we satisfy the requirements, emit an
1989 // alternate sequence.
1990 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
1991 SDValue Op = BuildSDIV(N);
1992 if (Op.getNode()) return Op;
1996 if (N0.getOpcode() == ISD::UNDEF)
1997 return DAG.getConstant(0, VT);
1998 // X / undef -> undef
1999 if (N1.getOpcode() == ISD::UNDEF)
2005 SDValue DAGCombiner::visitUDIV(SDNode *N) {
2006 SDValue N0 = N->getOperand(0);
2007 SDValue N1 = N->getOperand(1);
2008 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
2009 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
2010 EVT VT = N->getValueType(0);
2013 if (VT.isVector()) {
2014 SDValue FoldedVOp = SimplifyVBinOp(N);
2015 if (FoldedVOp.getNode()) return FoldedVOp;
2018 // fold (udiv c1, c2) -> c1/c2
2019 if (N0C && N1C && !N1C->isNullValue())
2020 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
2021 // fold (udiv x, (1 << c)) -> x >>u c
2022 if (N1C && N1C->getAPIntValue().isPowerOf2())
2023 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0,
2024 DAG.getConstant(N1C->getAPIntValue().logBase2(),
2025 getShiftAmountTy(N0.getValueType())));
2026 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
2027 if (N1.getOpcode() == ISD::SHL) {
2028 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2029 if (SHC->getAPIntValue().isPowerOf2()) {
2030 EVT ADDVT = N1.getOperand(1).getValueType();
2031 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N), ADDVT,
2033 DAG.getConstant(SHC->getAPIntValue()
2036 AddToWorkList(Add.getNode());
2037 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, Add);
2041 // fold (udiv x, c) -> alternate
2042 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
2043 SDValue Op = BuildUDIV(N);
2044 if (Op.getNode()) return Op;
2048 if (N0.getOpcode() == ISD::UNDEF)
2049 return DAG.getConstant(0, VT);
2050 // X / undef -> undef
2051 if (N1.getOpcode() == ISD::UNDEF)
2057 SDValue DAGCombiner::visitSREM(SDNode *N) {
2058 SDValue N0 = N->getOperand(0);
2059 SDValue N1 = N->getOperand(1);
2060 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2061 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2062 EVT VT = N->getValueType(0);
2064 // fold (srem c1, c2) -> c1%c2
2065 if (N0C && N1C && !N1C->isNullValue())
2066 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
2067 // If we know the sign bits of both operands are zero, strength reduce to a
2068 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
2069 if (!VT.isVector()) {
2070 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2071 return DAG.getNode(ISD::UREM, SDLoc(N), VT, N0, N1);
2074 // If X/C can be simplified by the division-by-constant logic, lower
2075 // X%C to the equivalent of X-X/C*C.
2076 if (N1C && !N1C->isNullValue()) {
2077 SDValue Div = DAG.getNode(ISD::SDIV, SDLoc(N), VT, N0, N1);
2078 AddToWorkList(Div.getNode());
2079 SDValue OptimizedDiv = combine(Div.getNode());
2080 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2081 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2083 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2084 AddToWorkList(Mul.getNode());
2090 if (N0.getOpcode() == ISD::UNDEF)
2091 return DAG.getConstant(0, VT);
2092 // X % undef -> undef
2093 if (N1.getOpcode() == ISD::UNDEF)
2099 SDValue DAGCombiner::visitUREM(SDNode *N) {
2100 SDValue N0 = N->getOperand(0);
2101 SDValue N1 = N->getOperand(1);
2102 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2103 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2104 EVT VT = N->getValueType(0);
2106 // fold (urem c1, c2) -> c1%c2
2107 if (N0C && N1C && !N1C->isNullValue())
2108 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
2109 // fold (urem x, pow2) -> (and x, pow2-1)
2110 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
2111 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0,
2112 DAG.getConstant(N1C->getAPIntValue()-1,VT));
2113 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2114 if (N1.getOpcode() == ISD::SHL) {
2115 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2116 if (SHC->getAPIntValue().isPowerOf2()) {
2118 DAG.getNode(ISD::ADD, SDLoc(N), VT, N1,
2119 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
2121 AddToWorkList(Add.getNode());
2122 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, Add);
2127 // If X/C can be simplified by the division-by-constant logic, lower
2128 // X%C to the equivalent of X-X/C*C.
2129 if (N1C && !N1C->isNullValue()) {
2130 SDValue Div = DAG.getNode(ISD::UDIV, SDLoc(N), VT, N0, N1);
2131 AddToWorkList(Div.getNode());
2132 SDValue OptimizedDiv = combine(Div.getNode());
2133 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2134 SDValue Mul = DAG.getNode(ISD::MUL, SDLoc(N), VT,
2136 SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, N0, Mul);
2137 AddToWorkList(Mul.getNode());
2143 if (N0.getOpcode() == ISD::UNDEF)
2144 return DAG.getConstant(0, VT);
2145 // X % undef -> undef
2146 if (N1.getOpcode() == ISD::UNDEF)
2152 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2153 SDValue N0 = N->getOperand(0);
2154 SDValue N1 = N->getOperand(1);
2155 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2156 EVT VT = N->getValueType(0);
2159 // fold (mulhs x, 0) -> 0
2160 if (N1C && N1C->isNullValue())
2162 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2163 if (N1C && N1C->getAPIntValue() == 1)
2164 return DAG.getNode(ISD::SRA, SDLoc(N), N0.getValueType(), N0,
2165 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2166 getShiftAmountTy(N0.getValueType())));
2167 // fold (mulhs x, undef) -> 0
2168 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2169 return DAG.getConstant(0, VT);
2171 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2173 if (VT.isSimple() && !VT.isVector()) {
2174 MVT Simple = VT.getSimpleVT();
2175 unsigned SimpleSize = Simple.getSizeInBits();
2176 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2177 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2178 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2179 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2180 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2181 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2182 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2183 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2190 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2191 SDValue N0 = N->getOperand(0);
2192 SDValue N1 = N->getOperand(1);
2193 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2194 EVT VT = N->getValueType(0);
2197 // fold (mulhu x, 0) -> 0
2198 if (N1C && N1C->isNullValue())
2200 // fold (mulhu x, 1) -> 0
2201 if (N1C && N1C->getAPIntValue() == 1)
2202 return DAG.getConstant(0, N0.getValueType());
2203 // fold (mulhu x, undef) -> 0
2204 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2205 return DAG.getConstant(0, VT);
2207 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2209 if (VT.isSimple() && !VT.isVector()) {
2210 MVT Simple = VT.getSimpleVT();
2211 unsigned SimpleSize = Simple.getSizeInBits();
2212 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2213 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2214 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2215 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2216 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2217 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2218 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2219 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2226 /// SimplifyNodeWithTwoResults - Perform optimizations common to nodes that
2227 /// compute two values. LoOp and HiOp give the opcodes for the two computations
2228 /// that are being performed. Return true if a simplification was made.
2230 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2232 // If the high half is not needed, just compute the low half.
2233 bool HiExists = N->hasAnyUseOfValue(1);
2235 (!LegalOperations ||
2236 TLI.isOperationLegal(LoOp, N->getValueType(0)))) {
2237 SDValue Res = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0),
2238 N->op_begin(), N->getNumOperands());
2239 return CombineTo(N, Res, Res);
2242 // If the low half is not needed, just compute the high half.
2243 bool LoExists = N->hasAnyUseOfValue(0);
2245 (!LegalOperations ||
2246 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2247 SDValue Res = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1),
2248 N->op_begin(), N->getNumOperands());
2249 return CombineTo(N, Res, Res);
2252 // If both halves are used, return as it is.
2253 if (LoExists && HiExists)
2256 // If the two computed results can be simplified separately, separate them.
2258 SDValue Lo = DAG.getNode(LoOp, SDLoc(N), N->getValueType(0),
2259 N->op_begin(), N->getNumOperands());
2260 AddToWorkList(Lo.getNode());
2261 SDValue LoOpt = combine(Lo.getNode());
2262 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2263 (!LegalOperations ||
2264 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2265 return CombineTo(N, LoOpt, LoOpt);
2269 SDValue Hi = DAG.getNode(HiOp, SDLoc(N), N->getValueType(1),
2270 N->op_begin(), N->getNumOperands());
2271 AddToWorkList(Hi.getNode());
2272 SDValue HiOpt = combine(Hi.getNode());
2273 if (HiOpt.getNode() && HiOpt != Hi &&
2274 (!LegalOperations ||
2275 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2276 return CombineTo(N, HiOpt, HiOpt);
2282 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2283 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2284 if (Res.getNode()) return Res;
2286 EVT VT = N->getValueType(0);
2289 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2291 if (VT.isSimple() && !VT.isVector()) {
2292 MVT Simple = VT.getSimpleVT();
2293 unsigned SimpleSize = Simple.getSizeInBits();
2294 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2295 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2296 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2297 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2298 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2299 // Compute the high part as N1.
2300 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2301 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2302 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2303 // Compute the low part as N0.
2304 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2305 return CombineTo(N, Lo, Hi);
2312 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2313 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2314 if (Res.getNode()) return Res;
2316 EVT VT = N->getValueType(0);
2319 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2321 if (VT.isSimple() && !VT.isVector()) {
2322 MVT Simple = VT.getSimpleVT();
2323 unsigned SimpleSize = Simple.getSizeInBits();
2324 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2325 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2326 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2327 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2328 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2329 // Compute the high part as N1.
2330 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2331 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2332 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2333 // Compute the low part as N0.
2334 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2335 return CombineTo(N, Lo, Hi);
2342 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2343 // (smulo x, 2) -> (saddo x, x)
2344 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2345 if (C2->getAPIntValue() == 2)
2346 return DAG.getNode(ISD::SADDO, SDLoc(N), N->getVTList(),
2347 N->getOperand(0), N->getOperand(0));
2352 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2353 // (umulo x, 2) -> (uaddo x, x)
2354 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2355 if (C2->getAPIntValue() == 2)
2356 return DAG.getNode(ISD::UADDO, SDLoc(N), N->getVTList(),
2357 N->getOperand(0), N->getOperand(0));
2362 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2363 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2364 if (Res.getNode()) return Res;
2369 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2370 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2371 if (Res.getNode()) return Res;
2376 /// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with
2377 /// two operands of the same opcode, try to simplify it.
2378 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2379 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2380 EVT VT = N0.getValueType();
2381 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2383 // Bail early if none of these transforms apply.
2384 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2386 // For each of OP in AND/OR/XOR:
2387 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2388 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2389 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2390 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2392 // do not sink logical op inside of a vector extend, since it may combine
2394 EVT Op0VT = N0.getOperand(0).getValueType();
2395 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2396 N0.getOpcode() == ISD::SIGN_EXTEND ||
2397 // Avoid infinite looping with PromoteIntBinOp.
2398 (N0.getOpcode() == ISD::ANY_EXTEND &&
2399 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2400 (N0.getOpcode() == ISD::TRUNCATE &&
2401 (!TLI.isZExtFree(VT, Op0VT) ||
2402 !TLI.isTruncateFree(Op0VT, VT)) &&
2403 TLI.isTypeLegal(Op0VT))) &&
2405 Op0VT == N1.getOperand(0).getValueType() &&
2406 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2407 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2408 N0.getOperand(0).getValueType(),
2409 N0.getOperand(0), N1.getOperand(0));
2410 AddToWorkList(ORNode.getNode());
2411 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, ORNode);
2414 // For each of OP in SHL/SRL/SRA/AND...
2415 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2416 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2417 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2418 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2419 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2420 N0.getOperand(1) == N1.getOperand(1)) {
2421 SDValue ORNode = DAG.getNode(N->getOpcode(), SDLoc(N0),
2422 N0.getOperand(0).getValueType(),
2423 N0.getOperand(0), N1.getOperand(0));
2424 AddToWorkList(ORNode.getNode());
2425 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
2426 ORNode, N0.getOperand(1));
2429 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2430 // Only perform this optimization after type legalization and before
2431 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2432 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2433 // we don't want to undo this promotion.
2434 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2436 if ((N0.getOpcode() == ISD::BITCAST ||
2437 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) &&
2438 Level == AfterLegalizeTypes) {
2439 SDValue In0 = N0.getOperand(0);
2440 SDValue In1 = N1.getOperand(0);
2441 EVT In0Ty = In0.getValueType();
2442 EVT In1Ty = In1.getValueType();
2444 // If both incoming values are integers, and the original types are the
2446 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2447 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1);
2448 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op);
2449 AddToWorkList(Op.getNode());
2454 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2455 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2456 // If both shuffles use the same mask, and both shuffle within a single
2457 // vector, then it is worthwhile to move the swizzle after the operation.
2458 // The type-legalizer generates this pattern when loading illegal
2459 // vector types from memory. In many cases this allows additional shuffle
2461 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
2462 N0.getOperand(1).getOpcode() == ISD::UNDEF &&
2463 N1.getOperand(1).getOpcode() == ISD::UNDEF) {
2464 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2465 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2467 assert(N0.getOperand(0).getValueType() == N1.getOperand(1).getValueType() &&
2468 "Inputs to shuffles are not the same type");
2470 unsigned NumElts = VT.getVectorNumElements();
2472 // Check that both shuffles use the same mask. The masks are known to be of
2473 // the same length because the result vector type is the same.
2474 bool SameMask = true;
2475 for (unsigned i = 0; i != NumElts; ++i) {
2476 int Idx0 = SVN0->getMaskElt(i);
2477 int Idx1 = SVN1->getMaskElt(i);
2485 SDValue Op = DAG.getNode(N->getOpcode(), SDLoc(N), VT,
2486 N0.getOperand(0), N1.getOperand(0));
2487 AddToWorkList(Op.getNode());
2488 return DAG.getVectorShuffle(VT, SDLoc(N), Op,
2489 DAG.getUNDEF(VT), &SVN0->getMask()[0]);
2496 SDValue DAGCombiner::visitAND(SDNode *N) {
2497 SDValue N0 = N->getOperand(0);
2498 SDValue N1 = N->getOperand(1);
2499 SDValue LL, LR, RL, RR, CC0, CC1;
2500 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2501 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2502 EVT VT = N1.getValueType();
2503 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2506 if (VT.isVector()) {
2507 SDValue FoldedVOp = SimplifyVBinOp(N);
2508 if (FoldedVOp.getNode()) return FoldedVOp;
2510 // fold (and x, 0) -> 0, vector edition
2511 if (ISD::isBuildVectorAllZeros(N0.getNode()))
2513 if (ISD::isBuildVectorAllZeros(N1.getNode()))
2516 // fold (and x, -1) -> x, vector edition
2517 if (ISD::isBuildVectorAllOnes(N0.getNode()))
2519 if (ISD::isBuildVectorAllOnes(N1.getNode()))
2523 // fold (and x, undef) -> 0
2524 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2525 return DAG.getConstant(0, VT);
2526 // fold (and c1, c2) -> c1&c2
2528 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2529 // canonicalize constant to RHS
2531 return DAG.getNode(ISD::AND, SDLoc(N), VT, N1, N0);
2532 // fold (and x, -1) -> x
2533 if (N1C && N1C->isAllOnesValue())
2535 // if (and x, c) is known to be zero, return 0
2536 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2537 APInt::getAllOnesValue(BitWidth)))
2538 return DAG.getConstant(0, VT);
2540 SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1);
2541 if (RAND.getNode() != 0)
2543 // fold (and (or x, C), D) -> D if (C & D) == D
2544 if (N1C && N0.getOpcode() == ISD::OR)
2545 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2546 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2548 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2549 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2550 SDValue N0Op0 = N0.getOperand(0);
2551 APInt Mask = ~N1C->getAPIntValue();
2552 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2553 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2554 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
2555 N0.getValueType(), N0Op0);
2557 // Replace uses of the AND with uses of the Zero extend node.
2560 // We actually want to replace all uses of the any_extend with the
2561 // zero_extend, to avoid duplicating things. This will later cause this
2562 // AND to be folded.
2563 CombineTo(N0.getNode(), Zext);
2564 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2567 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2568 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2569 // already be zero by virtue of the width of the base type of the load.
2571 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2573 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2574 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2575 N0.getOpcode() == ISD::LOAD) {
2576 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2577 N0 : N0.getOperand(0) );
2579 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2580 // This can be a pure constant or a vector splat, in which case we treat the
2581 // vector as a scalar and use the splat value.
2582 APInt Constant = APInt::getNullValue(1);
2583 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2584 Constant = C->getAPIntValue();
2585 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2586 APInt SplatValue, SplatUndef;
2587 unsigned SplatBitSize;
2589 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2590 SplatBitSize, HasAnyUndefs);
2592 // Undef bits can contribute to a possible optimisation if set, so
2594 SplatValue |= SplatUndef;
2596 // The splat value may be something like "0x00FFFFFF", which means 0 for
2597 // the first vector value and FF for the rest, repeating. We need a mask
2598 // that will apply equally to all members of the vector, so AND all the
2599 // lanes of the constant together.
2600 EVT VT = Vector->getValueType(0);
2601 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2603 // If the splat value has been compressed to a bitlength lower
2604 // than the size of the vector lane, we need to re-expand it to
2606 if (BitWidth > SplatBitSize)
2607 for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
2608 SplatBitSize < BitWidth;
2609 SplatBitSize = SplatBitSize * 2)
2610 SplatValue |= SplatValue.shl(SplatBitSize);
2612 Constant = APInt::getAllOnesValue(BitWidth);
2613 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
2614 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
2618 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
2619 // actually legal and isn't going to get expanded, else this is a false
2621 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
2622 Load->getMemoryVT());
2624 // Resize the constant to the same size as the original memory access before
2625 // extension. If it is still the AllOnesValue then this AND is completely
2628 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
2631 switch (Load->getExtensionType()) {
2632 default: B = false; break;
2633 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
2635 case ISD::NON_EXTLOAD: B = true; break;
2638 if (B && Constant.isAllOnesValue()) {
2639 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
2640 // preserve semantics once we get rid of the AND.
2641 SDValue NewLoad(Load, 0);
2642 if (Load->getExtensionType() == ISD::EXTLOAD) {
2643 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
2644 Load->getValueType(0), SDLoc(Load),
2645 Load->getChain(), Load->getBasePtr(),
2646 Load->getOffset(), Load->getMemoryVT(),
2647 Load->getMemOperand());
2648 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
2649 if (Load->getNumValues() == 3) {
2650 // PRE/POST_INC loads have 3 values.
2651 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
2652 NewLoad.getValue(2) };
2653 CombineTo(Load, To, 3, true);
2655 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
2659 // Fold the AND away, taking care not to fold to the old load node if we
2661 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
2663 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2666 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2667 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2668 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2669 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2671 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2672 LL.getValueType().isInteger()) {
2673 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2674 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
2675 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2676 LR.getValueType(), LL, RL);
2677 AddToWorkList(ORNode.getNode());
2678 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
2680 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2681 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2682 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(N0),
2683 LR.getValueType(), LL, RL);
2684 AddToWorkList(ANDNode.getNode());
2685 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
2687 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2688 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2689 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(N0),
2690 LR.getValueType(), LL, RL);
2691 AddToWorkList(ORNode.getNode());
2692 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
2695 // Simplify (and (setne X, 0), (setne X, -1)) -> (setuge (add X, 1), 2)
2696 if (LL == RL && isa<ConstantSDNode>(LR) && isa<ConstantSDNode>(RR) &&
2697 Op0 == Op1 && LL.getValueType().isInteger() &&
2698 Op0 == ISD::SETNE && ((cast<ConstantSDNode>(LR)->isNullValue() &&
2699 cast<ConstantSDNode>(RR)->isAllOnesValue()) ||
2700 (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
2701 cast<ConstantSDNode>(RR)->isNullValue()))) {
2702 SDValue ADDNode = DAG.getNode(ISD::ADD, SDLoc(N0), LL.getValueType(),
2703 LL, DAG.getConstant(1, LL.getValueType()));
2704 AddToWorkList(ADDNode.getNode());
2705 return DAG.getSetCC(SDLoc(N), VT, ADDNode,
2706 DAG.getConstant(2, LL.getValueType()), ISD::SETUGE);
2708 // canonicalize equivalent to ll == rl
2709 if (LL == RR && LR == RL) {
2710 Op1 = ISD::getSetCCSwappedOperands(Op1);
2713 if (LL == RL && LR == RR) {
2714 bool isInteger = LL.getValueType().isInteger();
2715 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2716 if (Result != ISD::SETCC_INVALID &&
2717 (!LegalOperations ||
2718 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
2719 TLI.isOperationLegal(ISD::SETCC,
2720 getSetCCResultType(N0.getSimpleValueType())))))
2721 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
2726 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
2727 if (N0.getOpcode() == N1.getOpcode()) {
2728 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2729 if (Tmp.getNode()) return Tmp;
2732 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
2733 // fold (and (sra)) -> (and (srl)) when possible.
2734 if (!VT.isVector() &&
2735 SimplifyDemandedBits(SDValue(N, 0)))
2736 return SDValue(N, 0);
2738 // fold (zext_inreg (extload x)) -> (zextload x)
2739 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
2740 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2741 EVT MemVT = LN0->getMemoryVT();
2742 // If we zero all the possible extended bits, then we can turn this into
2743 // a zextload if we are running before legalize or the operation is legal.
2744 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2745 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2746 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2747 ((!LegalOperations && !LN0->isVolatile()) ||
2748 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2749 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2750 LN0->getChain(), LN0->getBasePtr(),
2751 MemVT, LN0->getMemOperand());
2753 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2754 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2757 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
2758 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
2760 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2761 EVT MemVT = LN0->getMemoryVT();
2762 // If we zero all the possible extended bits, then we can turn this into
2763 // a zextload if we are running before legalize or the operation is legal.
2764 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2765 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2766 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2767 ((!LegalOperations && !LN0->isVolatile()) ||
2768 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2769 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N0), VT,
2770 LN0->getChain(), LN0->getBasePtr(),
2771 MemVT, LN0->getMemOperand());
2773 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2774 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2778 // fold (and (load x), 255) -> (zextload x, i8)
2779 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2780 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2781 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2782 (N0.getOpcode() == ISD::ANY_EXTEND &&
2783 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2784 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2785 LoadSDNode *LN0 = HasAnyExt
2786 ? cast<LoadSDNode>(N0.getOperand(0))
2787 : cast<LoadSDNode>(N0);
2788 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2789 LN0->isUnindexed() && N0.hasOneUse() && SDValue(LN0, 0).hasOneUse()) {
2790 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2791 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2792 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2793 EVT LoadedVT = LN0->getMemoryVT();
2795 if (ExtVT == LoadedVT &&
2796 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2797 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2800 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
2801 LN0->getChain(), LN0->getBasePtr(), ExtVT,
2802 LN0->getMemOperand());
2804 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
2805 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2808 // Do not change the width of a volatile load.
2809 // Do not generate loads of non-round integer types since these can
2810 // be expensive (and would be wrong if the type is not byte sized).
2811 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
2812 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2813 EVT PtrType = LN0->getOperand(1).getValueType();
2815 unsigned Alignment = LN0->getAlignment();
2816 SDValue NewPtr = LN0->getBasePtr();
2818 // For big endian targets, we need to add an offset to the pointer
2819 // to load the correct bytes. For little endian systems, we merely
2820 // need to read fewer bytes from the same pointer.
2821 if (TLI.isBigEndian()) {
2822 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
2823 unsigned EVTStoreBytes = ExtVT.getStoreSize();
2824 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
2825 NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0), PtrType,
2826 NewPtr, DAG.getConstant(PtrOff, PtrType));
2827 Alignment = MinAlign(Alignment, PtrOff);
2830 AddToWorkList(NewPtr.getNode());
2832 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2834 DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), LoadResultTy,
2835 LN0->getChain(), NewPtr,
2836 LN0->getPointerInfo(),
2837 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2838 Alignment, LN0->getTBAAInfo());
2840 CombineTo(LN0, Load, Load.getValue(1));
2841 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2847 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
2848 VT.getSizeInBits() <= 64) {
2849 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2850 APInt ADDC = ADDI->getAPIntValue();
2851 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2852 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
2853 // immediate for an add, but it is legal if its top c2 bits are set,
2854 // transform the ADD so the immediate doesn't need to be materialized
2856 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
2857 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
2858 SRLI->getZExtValue());
2859 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
2861 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2863 DAG.getNode(ISD::ADD, SDLoc(N0), VT,
2864 N0.getOperand(0), DAG.getConstant(ADDC, VT));
2865 CombineTo(N0.getNode(), NewAdd);
2866 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2874 // fold (and (or (srl N, 8), (shl N, 8)), 0xffff) -> (srl (bswap N), const)
2875 if (N1C && N1C->getAPIntValue() == 0xffff && N0.getOpcode() == ISD::OR) {
2876 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
2877 N0.getOperand(1), false);
2878 if (BSwap.getNode())
2885 /// MatchBSwapHWord - Match (a >> 8) | (a << 8) as (bswap a) >> 16
2887 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
2888 bool DemandHighBits) {
2889 if (!LegalOperations)
2892 EVT VT = N->getValueType(0);
2893 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
2895 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
2898 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
2899 bool LookPassAnd0 = false;
2900 bool LookPassAnd1 = false;
2901 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
2903 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
2905 if (N0.getOpcode() == ISD::AND) {
2906 if (!N0.getNode()->hasOneUse())
2908 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2909 if (!N01C || N01C->getZExtValue() != 0xFF00)
2911 N0 = N0.getOperand(0);
2912 LookPassAnd0 = true;
2915 if (N1.getOpcode() == ISD::AND) {
2916 if (!N1.getNode()->hasOneUse())
2918 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
2919 if (!N11C || N11C->getZExtValue() != 0xFF)
2921 N1 = N1.getOperand(0);
2922 LookPassAnd1 = true;
2925 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
2927 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
2929 if (!N0.getNode()->hasOneUse() ||
2930 !N1.getNode()->hasOneUse())
2933 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2934 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
2937 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
2940 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
2941 SDValue N00 = N0->getOperand(0);
2942 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
2943 if (!N00.getNode()->hasOneUse())
2945 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
2946 if (!N001C || N001C->getZExtValue() != 0xFF)
2948 N00 = N00.getOperand(0);
2949 LookPassAnd0 = true;
2952 SDValue N10 = N1->getOperand(0);
2953 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
2954 if (!N10.getNode()->hasOneUse())
2956 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
2957 if (!N101C || N101C->getZExtValue() != 0xFF00)
2959 N10 = N10.getOperand(0);
2960 LookPassAnd1 = true;
2966 // Make sure everything beyond the low halfword gets set to zero since the SRL
2967 // 16 will clear the top bits.
2968 unsigned OpSizeInBits = VT.getSizeInBits();
2969 if (DemandHighBits && OpSizeInBits > 16) {
2970 // If the left-shift isn't masked out then the only way this is a bswap is
2971 // if all bits beyond the low 8 are 0. In that case the entire pattern
2972 // reduces to a left shift anyway: leave it for other parts of the combiner.
2976 // However, if the right shift isn't masked out then it might be because
2977 // it's not needed. See if we can spot that too.
2978 if (!LookPassAnd1 &&
2979 !DAG.MaskedValueIsZero(
2980 N10, APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - 16)))
2984 SDValue Res = DAG.getNode(ISD::BSWAP, SDLoc(N), VT, N00);
2985 if (OpSizeInBits > 16)
2986 Res = DAG.getNode(ISD::SRL, SDLoc(N), VT, Res,
2987 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
2991 /// isBSwapHWordElement - Return true if the specified node is an element
2992 /// that makes up a 32-bit packed halfword byteswap. i.e.
2993 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
2994 static bool isBSwapHWordElement(SDValue N, SmallVectorImpl<SDNode *> &Parts) {
2995 if (!N.getNode()->hasOneUse())
2998 unsigned Opc = N.getOpcode();
2999 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
3002 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3007 switch (N1C->getZExtValue()) {
3010 case 0xFF: Num = 0; break;
3011 case 0xFF00: Num = 1; break;
3012 case 0xFF0000: Num = 2; break;
3013 case 0xFF000000: Num = 3; break;
3016 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
3017 SDValue N0 = N.getOperand(0);
3018 if (Opc == ISD::AND) {
3019 if (Num == 0 || Num == 2) {
3021 // (x >> 8) & 0xff0000
3022 if (N0.getOpcode() != ISD::SRL)
3024 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3025 if (!C || C->getZExtValue() != 8)
3028 // (x << 8) & 0xff00
3029 // (x << 8) & 0xff000000
3030 if (N0.getOpcode() != ISD::SHL)
3032 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3033 if (!C || C->getZExtValue() != 8)
3036 } else if (Opc == ISD::SHL) {
3038 // (x & 0xff0000) << 8
3039 if (Num != 0 && Num != 2)
3041 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3042 if (!C || C->getZExtValue() != 8)
3044 } else { // Opc == ISD::SRL
3045 // (x & 0xff00) >> 8
3046 // (x & 0xff000000) >> 8
3047 if (Num != 1 && Num != 3)
3049 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
3050 if (!C || C->getZExtValue() != 8)
3057 Parts[Num] = N0.getOperand(0).getNode();
3061 /// MatchBSwapHWord - Match a 32-bit packed halfword bswap. That is
3062 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
3063 /// => (rotl (bswap x), 16)
3064 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
3065 if (!LegalOperations)
3068 EVT VT = N->getValueType(0);
3071 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
3074 SmallVector<SDNode*,4> Parts(4, (SDNode*)0);
3076 // (or (or (and), (and)), (or (and), (and)))
3077 // (or (or (or (and), (and)), (and)), (and))
3078 if (N0.getOpcode() != ISD::OR)
3080 SDValue N00 = N0.getOperand(0);
3081 SDValue N01 = N0.getOperand(1);
3083 if (N1.getOpcode() == ISD::OR &&
3084 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) {
3085 // (or (or (and), (and)), (or (and), (and)))
3086 SDValue N000 = N00.getOperand(0);
3087 if (!isBSwapHWordElement(N000, Parts))
3090 SDValue N001 = N00.getOperand(1);
3091 if (!isBSwapHWordElement(N001, Parts))
3093 SDValue N010 = N01.getOperand(0);
3094 if (!isBSwapHWordElement(N010, Parts))
3096 SDValue N011 = N01.getOperand(1);
3097 if (!isBSwapHWordElement(N011, Parts))
3100 // (or (or (or (and), (and)), (and)), (and))
3101 if (!isBSwapHWordElement(N1, Parts))
3103 if (!isBSwapHWordElement(N01, Parts))
3105 if (N00.getOpcode() != ISD::OR)
3107 SDValue N000 = N00.getOperand(0);
3108 if (!isBSwapHWordElement(N000, Parts))
3110 SDValue N001 = N00.getOperand(1);
3111 if (!isBSwapHWordElement(N001, Parts))
3115 // Make sure the parts are all coming from the same node.
3116 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
3119 SDValue BSwap = DAG.getNode(ISD::BSWAP, SDLoc(N), VT,
3120 SDValue(Parts[0],0));
3122 // Result of the bswap should be rotated by 16. If it's not legal, then
3123 // do (x << 16) | (x >> 16).
3124 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
3125 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3126 return DAG.getNode(ISD::ROTL, SDLoc(N), VT, BSwap, ShAmt);
3127 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3128 return DAG.getNode(ISD::ROTR, SDLoc(N), VT, BSwap, ShAmt);
3129 return DAG.getNode(ISD::OR, SDLoc(N), VT,
3130 DAG.getNode(ISD::SHL, SDLoc(N), VT, BSwap, ShAmt),
3131 DAG.getNode(ISD::SRL, SDLoc(N), VT, BSwap, ShAmt));
3134 SDValue DAGCombiner::visitOR(SDNode *N) {
3135 SDValue N0 = N->getOperand(0);
3136 SDValue N1 = N->getOperand(1);
3137 SDValue LL, LR, RL, RR, CC0, CC1;
3138 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3139 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3140 EVT VT = N1.getValueType();
3143 if (VT.isVector()) {
3144 SDValue FoldedVOp = SimplifyVBinOp(N);
3145 if (FoldedVOp.getNode()) return FoldedVOp;
3147 // fold (or x, 0) -> x, vector edition
3148 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3150 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3153 // fold (or x, -1) -> -1, vector edition
3154 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3156 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3160 // fold (or x, undef) -> -1
3161 if (!LegalOperations &&
3162 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3163 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3164 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
3166 // fold (or c1, c2) -> c1|c2
3168 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
3169 // canonicalize constant to RHS
3171 return DAG.getNode(ISD::OR, SDLoc(N), VT, N1, N0);
3172 // fold (or x, 0) -> x
3173 if (N1C && N1C->isNullValue())
3175 // fold (or x, -1) -> -1
3176 if (N1C && N1C->isAllOnesValue())
3178 // fold (or x, c) -> c iff (x & ~c) == 0
3179 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3182 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3183 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
3184 if (BSwap.getNode() != 0)
3186 BSwap = MatchBSwapHWordLow(N, N0, N1);
3187 if (BSwap.getNode() != 0)
3191 SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1);
3192 if (ROR.getNode() != 0)
3194 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3195 // iff (c1 & c2) == 0.
3196 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3197 isa<ConstantSDNode>(N0.getOperand(1))) {
3198 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3199 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0)
3200 return DAG.getNode(ISD::AND, SDLoc(N), VT,
3201 DAG.getNode(ISD::OR, SDLoc(N0), VT,
3202 N0.getOperand(0), N1),
3203 DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1));
3205 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3206 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3207 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3208 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3210 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3211 LL.getValueType().isInteger()) {
3212 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3213 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3214 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3215 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3216 SDValue ORNode = DAG.getNode(ISD::OR, SDLoc(LR),
3217 LR.getValueType(), LL, RL);
3218 AddToWorkList(ORNode.getNode());
3219 return DAG.getSetCC(SDLoc(N), VT, ORNode, LR, Op1);
3221 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3222 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3223 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3224 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3225 SDValue ANDNode = DAG.getNode(ISD::AND, SDLoc(LR),
3226 LR.getValueType(), LL, RL);
3227 AddToWorkList(ANDNode.getNode());
3228 return DAG.getSetCC(SDLoc(N), VT, ANDNode, LR, Op1);
3231 // canonicalize equivalent to ll == rl
3232 if (LL == RR && LR == RL) {
3233 Op1 = ISD::getSetCCSwappedOperands(Op1);
3236 if (LL == RL && LR == RR) {
3237 bool isInteger = LL.getValueType().isInteger();
3238 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3239 if (Result != ISD::SETCC_INVALID &&
3240 (!LegalOperations ||
3241 (TLI.isCondCodeLegal(Result, LL.getSimpleValueType()) &&
3242 TLI.isOperationLegal(ISD::SETCC,
3243 getSetCCResultType(N0.getValueType())))))
3244 return DAG.getSetCC(SDLoc(N), N0.getValueType(),
3249 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3250 if (N0.getOpcode() == N1.getOpcode()) {
3251 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3252 if (Tmp.getNode()) return Tmp;
3255 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3256 if (N0.getOpcode() == ISD::AND &&
3257 N1.getOpcode() == ISD::AND &&
3258 N0.getOperand(1).getOpcode() == ISD::Constant &&
3259 N1.getOperand(1).getOpcode() == ISD::Constant &&
3260 // Don't increase # computations.
3261 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3262 // We can only do this xform if we know that bits from X that are set in C2
3263 // but not in C1 are already zero. Likewise for Y.
3264 const APInt &LHSMask =
3265 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3266 const APInt &RHSMask =
3267 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3269 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3270 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3271 SDValue X = DAG.getNode(ISD::OR, SDLoc(N0), VT,
3272 N0.getOperand(0), N1.getOperand(0));
3273 return DAG.getNode(ISD::AND, SDLoc(N), VT, X,
3274 DAG.getConstant(LHSMask | RHSMask, VT));
3278 // See if this is some rotate idiom.
3279 if (SDNode *Rot = MatchRotate(N0, N1, SDLoc(N)))
3280 return SDValue(Rot, 0);
3282 // Simplify the operands using demanded-bits information.
3283 if (!VT.isVector() &&
3284 SimplifyDemandedBits(SDValue(N, 0)))
3285 return SDValue(N, 0);
3290 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
3291 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3292 if (Op.getOpcode() == ISD::AND) {
3293 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3294 Mask = Op.getOperand(1);
3295 Op = Op.getOperand(0);
3301 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3309 // Return true if we can prove that, whenever Neg and Pos are both in the
3310 // range [0, OpSize), Neg == (Pos == 0 ? 0 : OpSize - Pos). This means that
3311 // for two opposing shifts shift1 and shift2 and a value X with OpBits bits:
3313 // (or (shift1 X, Neg), (shift2 X, Pos))
3315 // reduces to a rotate in direction shift2 by Pos and a rotate in direction
3316 // shift1 by Neg. The range [0, OpSize) means that we only need to consider
3317 // shift amounts with defined behavior.
3318 static bool matchRotateSub(SDValue Pos, SDValue Neg, unsigned OpSize) {
3319 // If OpSize is a power of 2 then:
3321 // (a) (Pos == 0 ? 0 : OpSize - Pos) == (OpSize - Pos) & (OpSize - 1)
3322 // (b) Neg == Neg & (OpSize - 1) whenever Neg is in [0, OpSize).
3324 // So if OpSize is a power of 2 and Neg is (and Neg', OpSize-1), we check
3325 // for the stronger condition:
3327 // Neg & (OpSize - 1) == (OpSize - Pos) & (OpSize - 1) [A]
3329 // for all Neg and Pos. Since Neg & (OpSize - 1) == Neg' & (OpSize - 1)
3330 // we can just replace Neg with Neg' for the rest of the function.
3332 // In other cases we check for the even stronger condition:
3334 // Neg == OpSize - Pos [B]
3336 // for all Neg and Pos. Note that the (or ...) then invokes undefined
3337 // behavior if Pos == 0 (and consequently Neg == OpSize).
3339 // We could actually use [A] whenever OpSize is a power of 2, but the
3340 // only extra cases that it would match are those uninteresting ones
3341 // where Neg and Pos are never in range at the same time. E.g. for
3342 // OpSize == 32, using [A] would allow a Neg of the form (sub 64, Pos)
3343 // as well as (sub 32, Pos), but:
3345 // (or (shift1 X, (sub 64, Pos)), (shift2 X, Pos))
3347 // always invokes undefined behavior for 32-bit X.
3349 // Below, Mask == OpSize - 1 when using [A] and is all-ones otherwise.
3350 unsigned LoBits = 0;
3351 if (Neg.getOpcode() == ISD::AND &&
3352 isPowerOf2_64(OpSize) &&
3353 Neg.getOperand(1).getOpcode() == ISD::Constant &&
3354 cast<ConstantSDNode>(Neg.getOperand(1))->getAPIntValue() == OpSize - 1) {
3355 Neg = Neg.getOperand(0);
3356 LoBits = Log2_64(OpSize);
3359 // Check whether Neg has the form (sub NegC, NegOp1) for some NegC and NegOp1.
3360 if (Neg.getOpcode() != ISD::SUB)
3362 ConstantSDNode *NegC = dyn_cast<ConstantSDNode>(Neg.getOperand(0));
3365 SDValue NegOp1 = Neg.getOperand(1);
3367 // The condition we need is now:
3369 // (NegC - NegOp1) & Mask == (OpSize - Pos) & Mask
3371 // If NegOp1 == Pos then we need:
3373 // OpSize & Mask == NegC & Mask
3375 // (because "x & Mask" is a truncation and distributes through subtraction).
3378 Width = NegC->getAPIntValue();
3379 // Check for cases where Pos has the form (add NegOp1, PosC) for some PosC.
3380 // Then the condition we want to prove becomes:
3382 // (NegC - NegOp1) & Mask == (OpSize - (NegOp1 + PosC)) & Mask
3384 // which, again because "x & Mask" is a truncation, becomes:
3386 // NegC & Mask == (OpSize - PosC) & Mask
3387 // OpSize & Mask == (NegC + PosC) & Mask
3388 else if (Pos.getOpcode() == ISD::ADD &&
3389 Pos.getOperand(0) == NegOp1 &&
3390 Pos.getOperand(1).getOpcode() == ISD::Constant)
3391 Width = (cast<ConstantSDNode>(Pos.getOperand(1))->getAPIntValue() +
3392 NegC->getAPIntValue());
3396 // Now we just need to check that OpSize & Mask == Width & Mask.
3398 return Width.getLoBits(LoBits) == 0;
3399 return Width == OpSize;
3402 // A subroutine of MatchRotate used once we have found an OR of two opposite
3403 // shifts of Shifted. If Neg == <operand size> - Pos then the OR reduces
3404 // to both (PosOpcode Shifted, Pos) and (NegOpcode Shifted, Neg), with the
3405 // former being preferred if supported. InnerPos and InnerNeg are Pos and
3406 // Neg with outer conversions stripped away.
3407 SDNode *DAGCombiner::MatchRotatePosNeg(SDValue Shifted, SDValue Pos,
3408 SDValue Neg, SDValue InnerPos,
3409 SDValue InnerNeg, unsigned PosOpcode,
3410 unsigned NegOpcode, SDLoc DL) {
3411 // fold (or (shl x, (*ext y)),
3412 // (srl x, (*ext (sub 32, y)))) ->
3413 // (rotl x, y) or (rotr x, (sub 32, y))
3415 // fold (or (shl x, (*ext (sub 32, y))),
3416 // (srl x, (*ext y))) ->
3417 // (rotr x, y) or (rotl x, (sub 32, y))
3418 EVT VT = Shifted.getValueType();
3419 if (matchRotateSub(InnerPos, InnerNeg, VT.getSizeInBits())) {
3420 bool HasPos = TLI.isOperationLegalOrCustom(PosOpcode, VT);
3421 return DAG.getNode(HasPos ? PosOpcode : NegOpcode, DL, VT, Shifted,
3422 HasPos ? Pos : Neg).getNode();
3425 // fold (or (shl (*ext x), (*ext y)),
3426 // (srl (*ext x), (*ext (sub 32, y)))) ->
3427 // (*ext (rotl x, y)) or (*ext (rotr x, (sub 32, y)))
3429 // fold (or (shl (*ext x), (*ext (sub 32, y))),
3430 // (srl (*ext x), (*ext y))) ->
3431 // (*ext (rotr x, y)) or (*ext (rotl x, (sub 32, y)))
3432 if (Shifted.getOpcode() == ISD::ZERO_EXTEND ||
3433 Shifted.getOpcode() == ISD::ANY_EXTEND) {
3434 SDValue InnerShifted = Shifted.getOperand(0);
3435 EVT InnerVT = InnerShifted.getValueType();
3436 bool HasPosInner = TLI.isOperationLegalOrCustom(PosOpcode, InnerVT);
3437 if (HasPosInner || TLI.isOperationLegalOrCustom(NegOpcode, InnerVT)) {
3438 if (matchRotateSub(InnerPos, InnerNeg, InnerVT.getSizeInBits())) {
3439 SDValue V = DAG.getNode(HasPosInner ? PosOpcode : NegOpcode, DL,
3440 InnerVT, InnerShifted, HasPosInner ? Pos : Neg);
3441 return DAG.getNode(Shifted.getOpcode(), DL, VT, V).getNode();
3449 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3450 // idioms for rotate, and if the target supports rotation instructions, generate
3452 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, SDLoc DL) {
3453 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3454 EVT VT = LHS.getValueType();
3455 if (!TLI.isTypeLegal(VT)) return 0;
3457 // The target must have at least one rotate flavor.
3458 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3459 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3460 if (!HasROTL && !HasROTR) return 0;
3462 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3463 SDValue LHSShift; // The shift.
3464 SDValue LHSMask; // AND value if any.
3465 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3466 return 0; // Not part of a rotate.
3468 SDValue RHSShift; // The shift.
3469 SDValue RHSMask; // AND value if any.
3470 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3471 return 0; // Not part of a rotate.
3473 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3474 return 0; // Not shifting the same value.
3476 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3477 return 0; // Shifts must disagree.
3479 // Canonicalize shl to left side in a shl/srl pair.
3480 if (RHSShift.getOpcode() == ISD::SHL) {
3481 std::swap(LHS, RHS);
3482 std::swap(LHSShift, RHSShift);
3483 std::swap(LHSMask , RHSMask );
3486 unsigned OpSizeInBits = VT.getSizeInBits();
3487 SDValue LHSShiftArg = LHSShift.getOperand(0);
3488 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3489 SDValue RHSShiftArg = RHSShift.getOperand(0);
3490 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3492 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3493 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3494 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3495 RHSShiftAmt.getOpcode() == ISD::Constant) {
3496 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3497 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3498 if ((LShVal + RShVal) != OpSizeInBits)
3501 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3502 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
3504 // If there is an AND of either shifted operand, apply it to the result.
3505 if (LHSMask.getNode() || RHSMask.getNode()) {
3506 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3508 if (LHSMask.getNode()) {
3509 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3510 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3512 if (RHSMask.getNode()) {
3513 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3514 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3517 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3520 return Rot.getNode();
3523 // If there is a mask here, and we have a variable shift, we can't be sure
3524 // that we're masking out the right stuff.
3525 if (LHSMask.getNode() || RHSMask.getNode())
3528 // If the shift amount is sign/zext/any-extended just peel it off.
3529 SDValue LExtOp0 = LHSShiftAmt;
3530 SDValue RExtOp0 = RHSShiftAmt;
3531 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3532 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3533 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3534 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3535 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3536 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3537 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3538 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3539 LExtOp0 = LHSShiftAmt.getOperand(0);
3540 RExtOp0 = RHSShiftAmt.getOperand(0);
3543 SDNode *TryL = MatchRotatePosNeg(LHSShiftArg, LHSShiftAmt, RHSShiftAmt,
3544 LExtOp0, RExtOp0, ISD::ROTL, ISD::ROTR, DL);
3548 SDNode *TryR = MatchRotatePosNeg(RHSShiftArg, RHSShiftAmt, LHSShiftAmt,
3549 RExtOp0, LExtOp0, ISD::ROTR, ISD::ROTL, DL);
3556 SDValue DAGCombiner::visitXOR(SDNode *N) {
3557 SDValue N0 = N->getOperand(0);
3558 SDValue N1 = N->getOperand(1);
3559 SDValue LHS, RHS, CC;
3560 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3561 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3562 EVT VT = N0.getValueType();
3565 if (VT.isVector()) {
3566 SDValue FoldedVOp = SimplifyVBinOp(N);
3567 if (FoldedVOp.getNode()) return FoldedVOp;
3569 // fold (xor x, 0) -> x, vector edition
3570 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3572 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3576 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3577 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3578 return DAG.getConstant(0, VT);
3579 // fold (xor x, undef) -> undef
3580 if (N0.getOpcode() == ISD::UNDEF)
3582 if (N1.getOpcode() == ISD::UNDEF)
3584 // fold (xor c1, c2) -> c1^c2
3586 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3587 // canonicalize constant to RHS
3589 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N1, N0);
3590 // fold (xor x, 0) -> x
3591 if (N1C && N1C->isNullValue())
3594 SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1);
3595 if (RXOR.getNode() != 0)
3598 // fold !(x cc y) -> (x !cc y)
3599 if (N1C && N1C->getAPIntValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3600 bool isInt = LHS.getValueType().isInteger();
3601 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3604 if (!LegalOperations ||
3605 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) {
3606 switch (N0.getOpcode()) {
3608 llvm_unreachable("Unhandled SetCC Equivalent!");
3610 return DAG.getSetCC(SDLoc(N), VT, LHS, RHS, NotCC);
3611 case ISD::SELECT_CC:
3612 return DAG.getSelectCC(SDLoc(N), LHS, RHS, N0.getOperand(2),
3613 N0.getOperand(3), NotCC);
3618 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3619 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3620 N0.getNode()->hasOneUse() &&
3621 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3622 SDValue V = N0.getOperand(0);
3623 V = DAG.getNode(ISD::XOR, SDLoc(N0), V.getValueType(), V,
3624 DAG.getConstant(1, V.getValueType()));
3625 AddToWorkList(V.getNode());
3626 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, V);
3629 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3630 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3631 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3632 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3633 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3634 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3635 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3636 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3637 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3638 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3641 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3642 if (N1C && N1C->isAllOnesValue() &&
3643 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3644 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3645 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3646 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3647 LHS = DAG.getNode(ISD::XOR, SDLoc(LHS), VT, LHS, N1); // LHS = ~LHS
3648 RHS = DAG.getNode(ISD::XOR, SDLoc(RHS), VT, RHS, N1); // RHS = ~RHS
3649 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3650 return DAG.getNode(NewOpcode, SDLoc(N), VT, LHS, RHS);
3653 // fold (xor (and x, y), y) -> (and (not x), y)
3654 if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3655 N0->getOperand(1) == N1) {
3656 SDValue X = N0->getOperand(0);
3657 SDValue NotX = DAG.getNOT(SDLoc(X), X, VT);
3658 AddToWorkList(NotX.getNode());
3659 return DAG.getNode(ISD::AND, SDLoc(N), VT, NotX, N1);
3661 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3662 if (N1C && N0.getOpcode() == ISD::XOR) {
3663 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3664 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3666 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(1),
3667 DAG.getConstant(N1C->getAPIntValue() ^
3668 N00C->getAPIntValue(), VT));
3670 return DAG.getNode(ISD::XOR, SDLoc(N), VT, N0.getOperand(0),
3671 DAG.getConstant(N1C->getAPIntValue() ^
3672 N01C->getAPIntValue(), VT));
3674 // fold (xor x, x) -> 0
3676 return tryFoldToZero(SDLoc(N), TLI, VT, DAG, LegalOperations, LegalTypes);
3678 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
3679 if (N0.getOpcode() == N1.getOpcode()) {
3680 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3681 if (Tmp.getNode()) return Tmp;
3684 // Simplify the expression using non-local knowledge.
3685 if (!VT.isVector() &&
3686 SimplifyDemandedBits(SDValue(N, 0)))
3687 return SDValue(N, 0);
3692 /// visitShiftByConstant - Handle transforms common to the three shifts, when
3693 /// the shift amount is a constant.
3694 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, unsigned Amt) {
3695 SDNode *LHS = N->getOperand(0).getNode();
3696 if (!LHS->hasOneUse()) return SDValue();
3698 // We want to pull some binops through shifts, so that we have (and (shift))
3699 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
3700 // thing happens with address calculations, so it's important to canonicalize
3702 bool HighBitSet = false; // Can we transform this if the high bit is set?
3704 switch (LHS->getOpcode()) {
3705 default: return SDValue();
3708 HighBitSet = false; // We can only transform sra if the high bit is clear.
3711 HighBitSet = true; // We can only transform sra if the high bit is set.
3714 if (N->getOpcode() != ISD::SHL)
3715 return SDValue(); // only shl(add) not sr[al](add).
3716 HighBitSet = false; // We can only transform sra if the high bit is clear.
3720 // We require the RHS of the binop to be a constant as well.
3721 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
3722 if (!BinOpCst) return SDValue();
3724 // FIXME: disable this unless the input to the binop is a shift by a constant.
3725 // If it is not a shift, it pessimizes some common cases like:
3727 // void foo(int *X, int i) { X[i & 1235] = 1; }
3728 // int bar(int *X, int i) { return X[i & 255]; }
3729 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
3730 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
3731 BinOpLHSVal->getOpcode() != ISD::SRA &&
3732 BinOpLHSVal->getOpcode() != ISD::SRL) ||
3733 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
3736 EVT VT = N->getValueType(0);
3738 // If this is a signed shift right, and the high bit is modified by the
3739 // logical operation, do not perform the transformation. The highBitSet
3740 // boolean indicates the value of the high bit of the constant which would
3741 // cause it to be modified for this operation.
3742 if (N->getOpcode() == ISD::SRA) {
3743 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
3744 if (BinOpRHSSignSet != HighBitSet)
3748 // Fold the constants, shifting the binop RHS by the shift amount.
3749 SDValue NewRHS = DAG.getNode(N->getOpcode(), SDLoc(LHS->getOperand(1)),
3751 LHS->getOperand(1), N->getOperand(1));
3753 // Create the new shift.
3754 SDValue NewShift = DAG.getNode(N->getOpcode(),
3755 SDLoc(LHS->getOperand(0)),
3756 VT, LHS->getOperand(0), N->getOperand(1));
3758 // Create the new binop.
3759 return DAG.getNode(LHS->getOpcode(), SDLoc(N), VT, NewShift, NewRHS);
3762 SDValue DAGCombiner::visitSHL(SDNode *N) {
3763 SDValue N0 = N->getOperand(0);
3764 SDValue N1 = N->getOperand(1);
3765 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3766 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3767 EVT VT = N0.getValueType();
3768 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3771 if (VT.isVector()) {
3772 SDValue FoldedVOp = SimplifyVBinOp(N);
3773 if (FoldedVOp.getNode()) return FoldedVOp;
3776 // fold (shl c1, c2) -> c1<<c2
3778 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
3779 // fold (shl 0, x) -> 0
3780 if (N0C && N0C->isNullValue())
3782 // fold (shl x, c >= size(x)) -> undef
3783 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3784 return DAG.getUNDEF(VT);
3785 // fold (shl x, 0) -> x
3786 if (N1C && N1C->isNullValue())
3788 // fold (shl undef, x) -> 0
3789 if (N0.getOpcode() == ISD::UNDEF)
3790 return DAG.getConstant(0, VT);
3791 // if (shl x, c) is known to be zero, return 0
3792 if (DAG.MaskedValueIsZero(SDValue(N, 0),
3793 APInt::getAllOnesValue(OpSizeInBits)))
3794 return DAG.getConstant(0, VT);
3795 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
3796 if (N1.getOpcode() == ISD::TRUNCATE &&
3797 N1.getOperand(0).getOpcode() == ISD::AND &&
3798 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3799 SDValue N101 = N1.getOperand(0).getOperand(1);
3800 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3801 EVT TruncVT = N1.getValueType();
3802 SDValue N100 = N1.getOperand(0).getOperand(0);
3803 APInt TruncC = N101C->getAPIntValue();
3804 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
3805 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0,
3806 DAG.getNode(ISD::AND, SDLoc(N), TruncVT,
3807 DAG.getNode(ISD::TRUNCATE,
3810 DAG.getConstant(TruncC, TruncVT)));
3814 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3815 return SDValue(N, 0);
3817 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
3818 if (N1C && N0.getOpcode() == ISD::SHL &&
3819 N0.getOperand(1).getOpcode() == ISD::Constant) {
3820 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3821 uint64_t c2 = N1C->getZExtValue();
3822 if (c1 + c2 >= OpSizeInBits)
3823 return DAG.getConstant(0, VT);
3824 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
3825 DAG.getConstant(c1 + c2, N1.getValueType()));
3828 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
3829 // For this to be valid, the second form must not preserve any of the bits
3830 // that are shifted out by the inner shift in the first form. This means
3831 // the outer shift size must be >= the number of bits added by the ext.
3832 // As a corollary, we don't care what kind of ext it is.
3833 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
3834 N0.getOpcode() == ISD::ANY_EXTEND ||
3835 N0.getOpcode() == ISD::SIGN_EXTEND) &&
3836 N0.getOperand(0).getOpcode() == ISD::SHL &&
3837 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
3839 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
3840 uint64_t c2 = N1C->getZExtValue();
3841 EVT InnerShiftVT = N0.getOperand(0).getValueType();
3842 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
3843 if (c2 >= OpSizeInBits - InnerShiftSize) {
3844 if (c1 + c2 >= OpSizeInBits)
3845 return DAG.getConstant(0, VT);
3846 return DAG.getNode(ISD::SHL, SDLoc(N0), VT,
3847 DAG.getNode(N0.getOpcode(), SDLoc(N0), VT,
3848 N0.getOperand(0)->getOperand(0)),
3849 DAG.getConstant(c1 + c2, N1.getValueType()));
3853 // fold (shl (zext (srl x, C)), C) -> (zext (shl (srl x, C), C))
3854 // Only fold this if the inner zext has no other uses to avoid increasing
3855 // the total number of instructions.
3856 if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() &&
3857 N0.getOperand(0).getOpcode() == ISD::SRL &&
3858 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
3860 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
3861 if (c1 < VT.getSizeInBits()) {
3862 uint64_t c2 = N1C->getZExtValue();
3864 SDValue NewOp0 = N0.getOperand(0);
3865 EVT CountVT = NewOp0.getOperand(1).getValueType();
3866 SDValue NewSHL = DAG.getNode(ISD::SHL, SDLoc(N), NewOp0.getValueType(),
3867 NewOp0, DAG.getConstant(c2, CountVT));
3868 AddToWorkList(NewSHL.getNode());
3869 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
3874 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
3875 // (and (srl x, (sub c1, c2), MASK)
3876 // Only fold this if the inner shift has no other uses -- if it does, folding
3877 // this will increase the total number of instructions.
3878 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse() &&
3879 N0.getOperand(1).getOpcode() == ISD::Constant) {
3880 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3881 if (c1 < VT.getSizeInBits()) {
3882 uint64_t c2 = N1C->getZExtValue();
3883 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
3884 VT.getSizeInBits() - c1);
3887 Mask = Mask.shl(c2-c1);
3888 Shift = DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
3889 DAG.getConstant(c2-c1, N1.getValueType()));
3891 Mask = Mask.lshr(c1-c2);
3892 Shift = DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
3893 DAG.getConstant(c1-c2, N1.getValueType()));
3895 return DAG.getNode(ISD::AND, SDLoc(N0), VT, Shift,
3896 DAG.getConstant(Mask, VT));
3899 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
3900 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
3901 SDValue HiBitsMask =
3902 DAG.getConstant(APInt::getHighBitsSet(VT.getSizeInBits(),
3903 VT.getSizeInBits() -
3904 N1C->getZExtValue()),
3906 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
3911 SDValue NewSHL = visitShiftByConstant(N, N1C->getZExtValue());
3912 if (NewSHL.getNode())
3919 SDValue DAGCombiner::visitSRA(SDNode *N) {
3920 SDValue N0 = N->getOperand(0);
3921 SDValue N1 = N->getOperand(1);
3922 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3923 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3924 EVT VT = N0.getValueType();
3925 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3928 if (VT.isVector()) {
3929 SDValue FoldedVOp = SimplifyVBinOp(N);
3930 if (FoldedVOp.getNode()) return FoldedVOp;
3933 // fold (sra c1, c2) -> (sra c1, c2)
3935 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
3936 // fold (sra 0, x) -> 0
3937 if (N0C && N0C->isNullValue())
3939 // fold (sra -1, x) -> -1
3940 if (N0C && N0C->isAllOnesValue())
3942 // fold (sra x, (setge c, size(x))) -> undef
3943 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3944 return DAG.getUNDEF(VT);
3945 // fold (sra x, 0) -> x
3946 if (N1C && N1C->isNullValue())
3948 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
3950 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
3951 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
3952 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
3954 ExtVT = EVT::getVectorVT(*DAG.getContext(),
3955 ExtVT, VT.getVectorNumElements());
3956 if ((!LegalOperations ||
3957 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
3958 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
3959 N0.getOperand(0), DAG.getValueType(ExtVT));
3962 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
3963 if (N1C && N0.getOpcode() == ISD::SRA) {
3964 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3965 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
3966 if (Sum >= OpSizeInBits) Sum = OpSizeInBits-1;
3967 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0.getOperand(0),
3968 DAG.getConstant(Sum, N1C->getValueType(0)));
3972 // fold (sra (shl X, m), (sub result_size, n))
3973 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
3974 // result_size - n != m.
3975 // If truncate is free for the target sext(shl) is likely to result in better
3977 if (N0.getOpcode() == ISD::SHL) {
3978 // Get the two constanst of the shifts, CN0 = m, CN = n.
3979 const ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3981 // Determine what the truncate's result bitsize and type would be.
3983 EVT::getIntegerVT(*DAG.getContext(),
3984 OpSizeInBits - N1C->getZExtValue());
3985 // Determine the residual right-shift amount.
3986 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
3988 // If the shift is not a no-op (in which case this should be just a sign
3989 // extend already), the truncated to type is legal, sign_extend is legal
3990 // on that type, and the truncate to that type is both legal and free,
3991 // perform the transform.
3992 if ((ShiftAmt > 0) &&
3993 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
3994 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
3995 TLI.isTruncateFree(VT, TruncVT)) {
3997 SDValue Amt = DAG.getConstant(ShiftAmt,
3998 getShiftAmountTy(N0.getOperand(0).getValueType()));
3999 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0), VT,
4000 N0.getOperand(0), Amt);
4001 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), TruncVT,
4003 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N),
4004 N->getValueType(0), Trunc);
4009 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
4010 if (N1.getOpcode() == ISD::TRUNCATE &&
4011 N1.getOperand(0).getOpcode() == ISD::AND &&
4012 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
4013 SDValue N101 = N1.getOperand(0).getOperand(1);
4014 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
4015 EVT TruncVT = N1.getValueType();
4016 SDValue N100 = N1.getOperand(0).getOperand(0);
4017 APInt TruncC = N101C->getAPIntValue();
4018 TruncC = TruncC.trunc(TruncVT.getScalarType().getSizeInBits());
4019 return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0,
4020 DAG.getNode(ISD::AND, SDLoc(N),
4022 DAG.getNode(ISD::TRUNCATE,
4025 DAG.getConstant(TruncC, TruncVT)));
4029 // fold (sra (trunc (sr x, c1)), c2) -> (trunc (sra x, c1+c2))
4030 // if c1 is equal to the number of bits the trunc removes
4031 if (N0.getOpcode() == ISD::TRUNCATE &&
4032 (N0.getOperand(0).getOpcode() == ISD::SRL ||
4033 N0.getOperand(0).getOpcode() == ISD::SRA) &&
4034 N0.getOperand(0).hasOneUse() &&
4035 N0.getOperand(0).getOperand(1).hasOneUse() &&
4036 N1C && isa<ConstantSDNode>(N0.getOperand(0).getOperand(1))) {
4037 EVT LargeVT = N0.getOperand(0).getValueType();
4038 ConstantSDNode *LargeShiftAmt =
4039 cast<ConstantSDNode>(N0.getOperand(0).getOperand(1));
4041 if (LargeVT.getScalarType().getSizeInBits() - OpSizeInBits ==
4042 LargeShiftAmt->getZExtValue()) {
4044 DAG.getConstant(LargeShiftAmt->getZExtValue() + N1C->getZExtValue(),
4045 getShiftAmountTy(N0.getOperand(0).getOperand(0).getValueType()));
4046 SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), LargeVT,
4047 N0.getOperand(0).getOperand(0), Amt);
4048 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, SRA);
4052 // Simplify, based on bits shifted out of the LHS.
4053 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4054 return SDValue(N, 0);
4057 // If the sign bit is known to be zero, switch this to a SRL.
4058 if (DAG.SignBitIsZero(N0))
4059 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1);
4062 SDValue NewSRA = visitShiftByConstant(N, N1C->getZExtValue());
4063 if (NewSRA.getNode())
4070 SDValue DAGCombiner::visitSRL(SDNode *N) {
4071 SDValue N0 = N->getOperand(0);
4072 SDValue N1 = N->getOperand(1);
4073 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4074 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4075 EVT VT = N0.getValueType();
4076 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
4079 if (VT.isVector()) {
4080 SDValue FoldedVOp = SimplifyVBinOp(N);
4081 if (FoldedVOp.getNode()) return FoldedVOp;
4084 // fold (srl c1, c2) -> c1 >>u c2
4086 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
4087 // fold (srl 0, x) -> 0
4088 if (N0C && N0C->isNullValue())
4090 // fold (srl x, c >= size(x)) -> undef
4091 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
4092 return DAG.getUNDEF(VT);
4093 // fold (srl x, 0) -> x
4094 if (N1C && N1C->isNullValue())
4096 // if (srl x, c) is known to be zero, return 0
4097 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
4098 APInt::getAllOnesValue(OpSizeInBits)))
4099 return DAG.getConstant(0, VT);
4101 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
4102 if (N1C && N0.getOpcode() == ISD::SRL &&
4103 N0.getOperand(1).getOpcode() == ISD::Constant) {
4104 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
4105 uint64_t c2 = N1C->getZExtValue();
4106 if (c1 + c2 >= OpSizeInBits)
4107 return DAG.getConstant(0, VT);
4108 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
4109 DAG.getConstant(c1 + c2, N1.getValueType()));
4112 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
4113 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
4114 N0.getOperand(0).getOpcode() == ISD::SRL &&
4115 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
4117 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
4118 uint64_t c2 = N1C->getZExtValue();
4119 EVT InnerShiftVT = N0.getOperand(0).getValueType();
4120 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
4121 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
4122 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
4123 if (c1 + OpSizeInBits == InnerShiftSize) {
4124 if (c1 + c2 >= InnerShiftSize)
4125 return DAG.getConstant(0, VT);
4126 return DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT,
4127 DAG.getNode(ISD::SRL, SDLoc(N0), InnerShiftVT,
4128 N0.getOperand(0)->getOperand(0),
4129 DAG.getConstant(c1 + c2, ShiftCountVT)));
4133 // fold (srl (shl x, c), c) -> (and x, cst2)
4134 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1 &&
4135 N0.getValueSizeInBits() <= 64) {
4136 uint64_t ShAmt = N1C->getZExtValue()+64-N0.getValueSizeInBits();
4137 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
4138 DAG.getConstant(~0ULL >> ShAmt, VT));
4141 // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask)
4142 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
4143 // Shifting in all undef bits?
4144 EVT SmallVT = N0.getOperand(0).getValueType();
4145 if (N1C->getZExtValue() >= SmallVT.getSizeInBits())
4146 return DAG.getUNDEF(VT);
4148 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
4149 uint64_t ShiftAmt = N1C->getZExtValue();
4150 SDValue SmallShift = DAG.getNode(ISD::SRL, SDLoc(N0), SmallVT,
4152 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
4153 AddToWorkList(SmallShift.getNode());
4154 APInt Mask = APInt::getAllOnesValue(VT.getSizeInBits()).lshr(ShiftAmt);
4155 return DAG.getNode(ISD::AND, SDLoc(N), VT,
4156 DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, SmallShift),
4157 DAG.getConstant(Mask, VT));
4161 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
4162 // bit, which is unmodified by sra.
4163 if (N1C && N1C->getZExtValue() + 1 == VT.getSizeInBits()) {
4164 if (N0.getOpcode() == ISD::SRA)
4165 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1);
4168 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
4169 if (N1C && N0.getOpcode() == ISD::CTLZ &&
4170 N1C->getAPIntValue() == Log2_32(VT.getSizeInBits())) {
4171 APInt KnownZero, KnownOne;
4172 DAG.ComputeMaskedBits(N0.getOperand(0), KnownZero, KnownOne);
4174 // If any of the input bits are KnownOne, then the input couldn't be all
4175 // zeros, thus the result of the srl will always be zero.
4176 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
4178 // If all of the bits input the to ctlz node are known to be zero, then
4179 // the result of the ctlz is "32" and the result of the shift is one.
4180 APInt UnknownBits = ~KnownZero;
4181 if (UnknownBits == 0) return DAG.getConstant(1, VT);
4183 // Otherwise, check to see if there is exactly one bit input to the ctlz.
4184 if ((UnknownBits & (UnknownBits - 1)) == 0) {
4185 // Okay, we know that only that the single bit specified by UnknownBits
4186 // could be set on input to the CTLZ node. If this bit is set, the SRL
4187 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
4188 // to an SRL/XOR pair, which is likely to simplify more.
4189 unsigned ShAmt = UnknownBits.countTrailingZeros();
4190 SDValue Op = N0.getOperand(0);
4193 Op = DAG.getNode(ISD::SRL, SDLoc(N0), VT, Op,
4194 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
4195 AddToWorkList(Op.getNode());
4198 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
4199 Op, DAG.getConstant(1, VT));
4203 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
4204 if (N1.getOpcode() == ISD::TRUNCATE &&
4205 N1.getOperand(0).getOpcode() == ISD::AND &&
4206 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
4207 SDValue N101 = N1.getOperand(0).getOperand(1);
4208 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
4209 EVT TruncVT = N1.getValueType();
4210 SDValue N100 = N1.getOperand(0).getOperand(0);
4211 APInt TruncC = N101C->getAPIntValue();
4212 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
4213 return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0,
4214 DAG.getNode(ISD::AND, SDLoc(N),
4216 DAG.getNode(ISD::TRUNCATE,
4219 DAG.getConstant(TruncC, TruncVT)));
4223 // fold operands of srl based on knowledge that the low bits are not
4225 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
4226 return SDValue(N, 0);
4229 SDValue NewSRL = visitShiftByConstant(N, N1C->getZExtValue());
4230 if (NewSRL.getNode())
4234 // Attempt to convert a srl of a load into a narrower zero-extending load.
4235 SDValue NarrowLoad = ReduceLoadWidth(N);
4236 if (NarrowLoad.getNode())
4239 // Here is a common situation. We want to optimize:
4242 // %b = and i32 %a, 2
4243 // %c = srl i32 %b, 1
4244 // brcond i32 %c ...
4250 // %c = setcc eq %b, 0
4253 // However when after the source operand of SRL is optimized into AND, the SRL
4254 // itself may not be optimized further. Look for it and add the BRCOND into
4256 if (N->hasOneUse()) {
4257 SDNode *Use = *N->use_begin();
4258 if (Use->getOpcode() == ISD::BRCOND)
4260 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
4261 // Also look pass the truncate.
4262 Use = *Use->use_begin();
4263 if (Use->getOpcode() == ISD::BRCOND)
4271 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
4272 SDValue N0 = N->getOperand(0);
4273 EVT VT = N->getValueType(0);
4275 // fold (ctlz c1) -> c2
4276 if (isa<ConstantSDNode>(N0))
4277 return DAG.getNode(ISD::CTLZ, SDLoc(N), VT, N0);
4281 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4282 SDValue N0 = N->getOperand(0);
4283 EVT VT = N->getValueType(0);
4285 // fold (ctlz_zero_undef c1) -> c2
4286 if (isa<ConstantSDNode>(N0))
4287 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4291 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4292 SDValue N0 = N->getOperand(0);
4293 EVT VT = N->getValueType(0);
4295 // fold (cttz c1) -> c2
4296 if (isa<ConstantSDNode>(N0))
4297 return DAG.getNode(ISD::CTTZ, SDLoc(N), VT, N0);
4301 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4302 SDValue N0 = N->getOperand(0);
4303 EVT VT = N->getValueType(0);
4305 // fold (cttz_zero_undef c1) -> c2
4306 if (isa<ConstantSDNode>(N0))
4307 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, SDLoc(N), VT, N0);
4311 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4312 SDValue N0 = N->getOperand(0);
4313 EVT VT = N->getValueType(0);
4315 // fold (ctpop c1) -> c2
4316 if (isa<ConstantSDNode>(N0))
4317 return DAG.getNode(ISD::CTPOP, SDLoc(N), VT, N0);
4321 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4322 SDValue N0 = N->getOperand(0);
4323 SDValue N1 = N->getOperand(1);
4324 SDValue N2 = N->getOperand(2);
4325 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4326 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4327 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
4328 EVT VT = N->getValueType(0);
4329 EVT VT0 = N0.getValueType();
4331 // fold (select C, X, X) -> X
4334 // fold (select true, X, Y) -> X
4335 if (N0C && !N0C->isNullValue())
4337 // fold (select false, X, Y) -> Y
4338 if (N0C && N0C->isNullValue())
4340 // fold (select C, 1, X) -> (or C, X)
4341 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4342 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4343 // fold (select C, 0, 1) -> (xor C, 1)
4344 if (VT.isInteger() &&
4347 TLI.getBooleanContents(false) ==
4348 TargetLowering::ZeroOrOneBooleanContent)) &&
4349 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4352 return DAG.getNode(ISD::XOR, SDLoc(N), VT0,
4353 N0, DAG.getConstant(1, VT0));
4354 XORNode = DAG.getNode(ISD::XOR, SDLoc(N0), VT0,
4355 N0, DAG.getConstant(1, VT0));
4356 AddToWorkList(XORNode.getNode());
4358 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, XORNode);
4359 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, XORNode);
4361 // fold (select C, 0, X) -> (and (not C), X)
4362 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4363 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4364 AddToWorkList(NOTNode.getNode());
4365 return DAG.getNode(ISD::AND, SDLoc(N), VT, NOTNode, N2);
4367 // fold (select C, X, 1) -> (or (not C), X)
4368 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4369 SDValue NOTNode = DAG.getNOT(SDLoc(N0), N0, VT);
4370 AddToWorkList(NOTNode.getNode());
4371 return DAG.getNode(ISD::OR, SDLoc(N), VT, NOTNode, N1);
4373 // fold (select C, X, 0) -> (and C, X)
4374 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4375 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4376 // fold (select X, X, Y) -> (or X, Y)
4377 // fold (select X, 1, Y) -> (or X, Y)
4378 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4379 return DAG.getNode(ISD::OR, SDLoc(N), VT, N0, N2);
4380 // fold (select X, Y, X) -> (and X, Y)
4381 // fold (select X, Y, 0) -> (and X, Y)
4382 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4383 return DAG.getNode(ISD::AND, SDLoc(N), VT, N0, N1);
4385 // If we can fold this based on the true/false value, do so.
4386 if (SimplifySelectOps(N, N1, N2))
4387 return SDValue(N, 0); // Don't revisit N.
4389 // fold selects based on a setcc into other things, such as min/max/abs
4390 if (N0.getOpcode() == ISD::SETCC) {
4392 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
4393 // having to say they don't support SELECT_CC on every type the DAG knows
4394 // about, since there is no way to mark an opcode illegal at all value types
4395 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other) &&
4396 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT))
4397 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT,
4398 N0.getOperand(0), N0.getOperand(1),
4399 N1, N2, N0.getOperand(2));
4400 return SimplifySelect(SDLoc(N), N0, N1, N2);
4407 std::pair<SDValue, SDValue> SplitVSETCC(const SDNode *N, SelectionDAG &DAG) {
4410 llvm::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
4412 // Split the inputs.
4413 SDValue Lo, Hi, LL, LH, RL, RH;
4414 llvm::tie(LL, LH) = DAG.SplitVectorOperand(N, 0);
4415 llvm::tie(RL, RH) = DAG.SplitVectorOperand(N, 1);
4417 Lo = DAG.getNode(N->getOpcode(), DL, LoVT, LL, RL, N->getOperand(2));
4418 Hi = DAG.getNode(N->getOpcode(), DL, HiVT, LH, RH, N->getOperand(2));
4420 return std::make_pair(Lo, Hi);
4423 SDValue DAGCombiner::visitVSELECT(SDNode *N) {
4424 SDValue N0 = N->getOperand(0);
4425 SDValue N1 = N->getOperand(1);
4426 SDValue N2 = N->getOperand(2);
4429 // Canonicalize integer abs.
4430 // vselect (setg[te] X, 0), X, -X ->
4431 // vselect (setgt X, -1), X, -X ->
4432 // vselect (setl[te] X, 0), -X, X ->
4433 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
4434 if (N0.getOpcode() == ISD::SETCC) {
4435 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
4436 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
4438 bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
4440 if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
4441 (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) &&
4442 N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1))
4443 isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode());
4444 else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) &&
4445 N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1))
4446 isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode());
4449 EVT VT = LHS.getValueType();
4450 SDValue Shift = DAG.getNode(
4451 ISD::SRA, DL, VT, LHS,
4452 DAG.getConstant(VT.getScalarType().getSizeInBits() - 1, VT));
4453 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift);
4454 AddToWorkList(Shift.getNode());
4455 AddToWorkList(Add.getNode());
4456 return DAG.getNode(ISD::XOR, DL, VT, Add, Shift);
4460 // If the VSELECT result requires splitting and the mask is provided by a
4461 // SETCC, then split both nodes and its operands before legalization. This
4462 // prevents the type legalizer from unrolling SETCC into scalar comparisons
4463 // and enables future optimizations (e.g. min/max pattern matching on X86).
4464 if (N0.getOpcode() == ISD::SETCC) {
4465 EVT VT = N->getValueType(0);
4467 // Check if any splitting is required.
4468 if (TLI.getTypeAction(*DAG.getContext(), VT) !=
4469 TargetLowering::TypeSplitVector)
4472 SDValue Lo, Hi, CCLo, CCHi, LL, LH, RL, RH;
4473 llvm::tie(CCLo, CCHi) = SplitVSETCC(N0.getNode(), DAG);
4474 llvm::tie(LL, LH) = DAG.SplitVectorOperand(N, 1);
4475 llvm::tie(RL, RH) = DAG.SplitVectorOperand(N, 2);
4477 Lo = DAG.getNode(N->getOpcode(), DL, LL.getValueType(), CCLo, LL, RL);
4478 Hi = DAG.getNode(N->getOpcode(), DL, LH.getValueType(), CCHi, LH, RH);
4480 // Add the new VSELECT nodes to the work list in case they need to be split
4482 AddToWorkList(Lo.getNode());
4483 AddToWorkList(Hi.getNode());
4485 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
4488 // Fold (vselect (build_vector all_ones), N1, N2) -> N1
4489 if (ISD::isBuildVectorAllOnes(N0.getNode()))
4491 // Fold (vselect (build_vector all_zeros), N1, N2) -> N2
4492 if (ISD::isBuildVectorAllZeros(N0.getNode()))
4498 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
4499 SDValue N0 = N->getOperand(0);
4500 SDValue N1 = N->getOperand(1);
4501 SDValue N2 = N->getOperand(2);
4502 SDValue N3 = N->getOperand(3);
4503 SDValue N4 = N->getOperand(4);
4504 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
4506 // fold select_cc lhs, rhs, x, x, cc -> x
4510 // Determine if the condition we're dealing with is constant
4511 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
4512 N0, N1, CC, SDLoc(N), false);
4513 if (SCC.getNode()) {
4514 AddToWorkList(SCC.getNode());
4516 if (ConstantSDNode *SCCC = dyn_cast<ConstantSDNode>(SCC.getNode())) {
4517 if (!SCCC->isNullValue())
4518 return N2; // cond always true -> true val
4520 return N3; // cond always false -> false val
4523 // Fold to a simpler select_cc
4524 if (SCC.getOpcode() == ISD::SETCC)
4525 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), N2.getValueType(),
4526 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
4530 // If we can fold this based on the true/false value, do so.
4531 if (SimplifySelectOps(N, N2, N3))
4532 return SDValue(N, 0); // Don't revisit N.
4534 // fold select_cc into other things, such as min/max/abs
4535 return SimplifySelectCC(SDLoc(N), N0, N1, N2, N3, CC);
4538 SDValue DAGCombiner::visitSETCC(SDNode *N) {
4539 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
4540 cast<CondCodeSDNode>(N->getOperand(2))->get(),
4544 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
4545 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
4546 // transformation. Returns true if extension are possible and the above
4547 // mentioned transformation is profitable.
4548 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
4550 SmallVectorImpl<SDNode *> &ExtendNodes,
4551 const TargetLowering &TLI) {
4552 bool HasCopyToRegUses = false;
4553 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
4554 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
4555 UE = N0.getNode()->use_end();
4560 if (UI.getUse().getResNo() != N0.getResNo())
4562 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
4563 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
4564 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
4565 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
4566 // Sign bits will be lost after a zext.
4569 for (unsigned i = 0; i != 2; ++i) {
4570 SDValue UseOp = User->getOperand(i);
4573 if (!isa<ConstantSDNode>(UseOp))
4578 ExtendNodes.push_back(User);
4581 // If truncates aren't free and there are users we can't
4582 // extend, it isn't worthwhile.
4585 // Remember if this value is live-out.
4586 if (User->getOpcode() == ISD::CopyToReg)
4587 HasCopyToRegUses = true;
4590 if (HasCopyToRegUses) {
4591 bool BothLiveOut = false;
4592 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
4594 SDUse &Use = UI.getUse();
4595 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
4601 // Both unextended and extended values are live out. There had better be
4602 // a good reason for the transformation.
4603 return ExtendNodes.size();
4608 void DAGCombiner::ExtendSetCCUses(const SmallVectorImpl<SDNode *> &SetCCs,
4609 SDValue Trunc, SDValue ExtLoad, SDLoc DL,
4610 ISD::NodeType ExtType) {
4611 // Extend SetCC uses if necessary.
4612 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
4613 SDNode *SetCC = SetCCs[i];
4614 SmallVector<SDValue, 4> Ops;
4616 for (unsigned j = 0; j != 2; ++j) {
4617 SDValue SOp = SetCC->getOperand(j);
4619 Ops.push_back(ExtLoad);
4621 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
4624 Ops.push_back(SetCC->getOperand(2));
4625 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0),
4626 &Ops[0], Ops.size()));
4630 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
4631 SDValue N0 = N->getOperand(0);
4632 EVT VT = N->getValueType(0);
4634 // fold (sext c1) -> c1
4635 if (isa<ConstantSDNode>(N0))
4636 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N0);
4638 // fold (sext (sext x)) -> (sext x)
4639 // fold (sext (aext x)) -> (sext x)
4640 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4641 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT,
4644 if (N0.getOpcode() == ISD::TRUNCATE) {
4645 // fold (sext (truncate (load x))) -> (sext (smaller load x))
4646 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
4647 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4648 if (NarrowLoad.getNode()) {
4649 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4650 if (NarrowLoad.getNode() != N0.getNode()) {
4651 CombineTo(N0.getNode(), NarrowLoad);
4652 // CombineTo deleted the truncate, if needed, but not what's under it.
4655 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4658 // See if the value being truncated is already sign extended. If so, just
4659 // eliminate the trunc/sext pair.
4660 SDValue Op = N0.getOperand(0);
4661 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
4662 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
4663 unsigned DestBits = VT.getScalarType().getSizeInBits();
4664 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
4666 if (OpBits == DestBits) {
4667 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
4668 // bits, it is already ready.
4669 if (NumSignBits > DestBits-MidBits)
4671 } else if (OpBits < DestBits) {
4672 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
4673 // bits, just sext from i32.
4674 if (NumSignBits > OpBits-MidBits)
4675 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, Op);
4677 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
4678 // bits, just truncate to i32.
4679 if (NumSignBits > OpBits-MidBits)
4680 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
4683 // fold (sext (truncate x)) -> (sextinreg x).
4684 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
4685 N0.getValueType())) {
4686 if (OpBits < DestBits)
4687 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N0), VT, Op);
4688 else if (OpBits > DestBits)
4689 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N0), VT, Op);
4690 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, Op,
4691 DAG.getValueType(N0.getValueType()));
4695 // fold (sext (load x)) -> (sext (truncate (sextload x)))
4696 // None of the supported targets knows how to perform load and sign extend
4697 // on vectors in one instruction. We only perform this transformation on
4699 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4700 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4701 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) {
4702 bool DoXform = true;
4703 SmallVector<SDNode*, 4> SetCCs;
4704 if (!N0.hasOneUse())
4705 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
4707 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4708 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
4710 LN0->getBasePtr(), N0.getValueType(),
4711 LN0->getMemOperand());
4712 CombineTo(N, ExtLoad);
4713 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
4714 N0.getValueType(), ExtLoad);
4715 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4716 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
4718 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4722 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
4723 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
4724 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
4725 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
4726 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4727 EVT MemVT = LN0->getMemoryVT();
4728 if ((!LegalOperations && !LN0->isVolatile()) ||
4729 TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) {
4730 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
4732 LN0->getBasePtr(), MemVT,
4733 LN0->getMemOperand());
4734 CombineTo(N, ExtLoad);
4735 CombineTo(N0.getNode(),
4736 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
4737 N0.getValueType(), ExtLoad),
4738 ExtLoad.getValue(1));
4739 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4743 // fold (sext (and/or/xor (load x), cst)) ->
4744 // (and/or/xor (sextload x), (sext cst))
4745 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
4746 N0.getOpcode() == ISD::XOR) &&
4747 isa<LoadSDNode>(N0.getOperand(0)) &&
4748 N0.getOperand(1).getOpcode() == ISD::Constant &&
4749 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()) &&
4750 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
4751 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
4752 if (LN0->getExtensionType() != ISD::ZEXTLOAD) {
4753 bool DoXform = true;
4754 SmallVector<SDNode*, 4> SetCCs;
4755 if (!N0.hasOneUse())
4756 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
4759 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(LN0), VT,
4760 LN0->getChain(), LN0->getBasePtr(),
4762 LN0->getMemOperand());
4763 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4764 Mask = Mask.sext(VT.getSizeInBits());
4765 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
4766 ExtLoad, DAG.getConstant(Mask, VT));
4767 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
4768 SDLoc(N0.getOperand(0)),
4769 N0.getOperand(0).getValueType(), ExtLoad);
4771 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
4772 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
4774 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4779 if (N0.getOpcode() == ISD::SETCC) {
4780 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
4781 // Only do this before legalize for now.
4782 if (VT.isVector() && !LegalOperations &&
4783 TLI.getBooleanContents(true) ==
4784 TargetLowering::ZeroOrNegativeOneBooleanContent) {
4785 EVT N0VT = N0.getOperand(0).getValueType();
4786 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
4787 // of the same size as the compared operands. Only optimize sext(setcc())
4788 // if this is the case.
4789 EVT SVT = getSetCCResultType(N0VT);
4791 // We know that the # elements of the results is the same as the
4792 // # elements of the compare (and the # elements of the compare result
4793 // for that matter). Check to see that they are the same size. If so,
4794 // we know that the element size of the sext'd result matches the
4795 // element size of the compare operands.
4796 if (VT.getSizeInBits() == SVT.getSizeInBits())
4797 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
4799 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4801 // If the desired elements are smaller or larger than the source
4802 // elements we can use a matching integer vector type and then
4803 // truncate/sign extend
4804 EVT MatchingVectorType = N0VT.changeVectorElementTypeToInteger();
4805 if (SVT == MatchingVectorType) {
4806 SDValue VsetCC = DAG.getSetCC(SDLoc(N), MatchingVectorType,
4807 N0.getOperand(0), N0.getOperand(1),
4808 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4809 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
4813 // sext(setcc x, y, cc) -> (select_cc x, y, -1, 0, cc)
4814 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
4816 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
4818 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
4819 NegOne, DAG.getConstant(0, VT),
4820 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4821 if (SCC.getNode()) return SCC;
4822 if (!VT.isVector() &&
4823 (!LegalOperations ||
4824 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(VT)))) {
4825 return DAG.getSelect(SDLoc(N), VT,
4826 DAG.getSetCC(SDLoc(N),
4827 getSetCCResultType(VT),
4828 N0.getOperand(0), N0.getOperand(1),
4829 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
4830 NegOne, DAG.getConstant(0, VT));
4834 // fold (sext x) -> (zext x) if the sign bit is known zero.
4835 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
4836 DAG.SignBitIsZero(N0))
4837 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0);
4842 // isTruncateOf - If N is a truncate of some other value, return true, record
4843 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
4844 // This function computes KnownZero to avoid a duplicated call to
4845 // ComputeMaskedBits in the caller.
4846 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
4849 if (N->getOpcode() == ISD::TRUNCATE) {
4850 Op = N->getOperand(0);
4851 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
4855 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
4856 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
4859 SDValue Op0 = N->getOperand(0);
4860 SDValue Op1 = N->getOperand(1);
4861 assert(Op0.getValueType() == Op1.getValueType());
4863 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
4864 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
4865 if (COp0 && COp0->isNullValue())
4867 else if (COp1 && COp1->isNullValue())
4872 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
4874 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
4880 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
4881 SDValue N0 = N->getOperand(0);
4882 EVT VT = N->getValueType(0);
4884 // fold (zext c1) -> c1
4885 if (isa<ConstantSDNode>(N0))
4886 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, N0);
4887 // fold (zext (zext x)) -> (zext x)
4888 // fold (zext (aext x)) -> (zext x)
4889 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4890 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT,
4893 // fold (zext (truncate x)) -> (zext x) or
4894 // (zext (truncate x)) -> (truncate x)
4895 // This is valid when the truncated bits of x are already zero.
4896 // FIXME: We should extend this to work for vectors too.
4899 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
4900 APInt TruncatedBits =
4901 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
4902 APInt(Op.getValueSizeInBits(), 0) :
4903 APInt::getBitsSet(Op.getValueSizeInBits(),
4904 N0.getValueSizeInBits(),
4905 std::min(Op.getValueSizeInBits(),
4906 VT.getSizeInBits()));
4907 if (TruncatedBits == (KnownZero & TruncatedBits)) {
4908 if (VT.bitsGT(Op.getValueType()))
4909 return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), VT, Op);
4910 if (VT.bitsLT(Op.getValueType()))
4911 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
4917 // fold (zext (truncate (load x))) -> (zext (smaller load x))
4918 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
4919 if (N0.getOpcode() == ISD::TRUNCATE) {
4920 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4921 if (NarrowLoad.getNode()) {
4922 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4923 if (NarrowLoad.getNode() != N0.getNode()) {
4924 CombineTo(N0.getNode(), NarrowLoad);
4925 // CombineTo deleted the truncate, if needed, but not what's under it.
4928 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4932 // fold (zext (truncate x)) -> (and x, mask)
4933 if (N0.getOpcode() == ISD::TRUNCATE &&
4934 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
4936 // fold (zext (truncate (load x))) -> (zext (smaller load x))
4937 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
4938 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4939 if (NarrowLoad.getNode()) {
4940 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4941 if (NarrowLoad.getNode() != N0.getNode()) {
4942 CombineTo(N0.getNode(), NarrowLoad);
4943 // CombineTo deleted the truncate, if needed, but not what's under it.
4946 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4949 SDValue Op = N0.getOperand(0);
4950 if (Op.getValueType().bitsLT(VT)) {
4951 Op = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, Op);
4952 AddToWorkList(Op.getNode());
4953 } else if (Op.getValueType().bitsGT(VT)) {
4954 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Op);
4955 AddToWorkList(Op.getNode());
4957 return DAG.getZeroExtendInReg(Op, SDLoc(N),
4958 N0.getValueType().getScalarType());
4961 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
4962 // if either of the casts is not free.
4963 if (N0.getOpcode() == ISD::AND &&
4964 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
4965 N0.getOperand(1).getOpcode() == ISD::Constant &&
4966 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
4967 N0.getValueType()) ||
4968 !TLI.isZExtFree(N0.getValueType(), VT))) {
4969 SDValue X = N0.getOperand(0).getOperand(0);
4970 if (X.getValueType().bitsLT(VT)) {
4971 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(X), VT, X);
4972 } else if (X.getValueType().bitsGT(VT)) {
4973 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
4975 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4976 Mask = Mask.zext(VT.getSizeInBits());
4977 return DAG.getNode(ISD::AND, SDLoc(N), VT,
4978 X, DAG.getConstant(Mask, VT));
4981 // fold (zext (load x)) -> (zext (truncate (zextload x)))
4982 // None of the supported targets knows how to perform load and vector_zext
4983 // on vectors in one instruction. We only perform this transformation on
4985 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4986 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4987 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
4988 bool DoXform = true;
4989 SmallVector<SDNode*, 4> SetCCs;
4990 if (!N0.hasOneUse())
4991 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
4993 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4994 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
4996 LN0->getBasePtr(), N0.getValueType(),
4997 LN0->getMemOperand());
4998 CombineTo(N, ExtLoad);
4999 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5000 N0.getValueType(), ExtLoad);
5001 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5003 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5005 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5009 // fold (zext (and/or/xor (load x), cst)) ->
5010 // (and/or/xor (zextload x), (zext cst))
5011 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
5012 N0.getOpcode() == ISD::XOR) &&
5013 isa<LoadSDNode>(N0.getOperand(0)) &&
5014 N0.getOperand(1).getOpcode() == ISD::Constant &&
5015 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()) &&
5016 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
5017 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
5018 if (LN0->getExtensionType() != ISD::SEXTLOAD) {
5019 bool DoXform = true;
5020 SmallVector<SDNode*, 4> SetCCs;
5021 if (!N0.hasOneUse())
5022 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
5025 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(LN0), VT,
5026 LN0->getChain(), LN0->getBasePtr(),
5028 LN0->getMemOperand());
5029 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5030 Mask = Mask.zext(VT.getSizeInBits());
5031 SDValue And = DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5032 ExtLoad, DAG.getConstant(Mask, VT));
5033 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
5034 SDLoc(N0.getOperand(0)),
5035 N0.getOperand(0).getValueType(), ExtLoad);
5037 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
5038 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5040 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5045 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
5046 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
5047 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
5048 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
5049 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5050 EVT MemVT = LN0->getMemoryVT();
5051 if ((!LegalOperations && !LN0->isVolatile()) ||
5052 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) {
5053 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, SDLoc(N), VT,
5055 LN0->getBasePtr(), MemVT,
5056 LN0->getMemOperand());
5057 CombineTo(N, ExtLoad);
5058 CombineTo(N0.getNode(),
5059 DAG.getNode(ISD::TRUNCATE, SDLoc(N0), N0.getValueType(),
5061 ExtLoad.getValue(1));
5062 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5066 if (N0.getOpcode() == ISD::SETCC) {
5067 if (!LegalOperations && VT.isVector() &&
5068 N0.getValueType().getVectorElementType() == MVT::i1) {
5069 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
5070 // Only do this before legalize for now.
5071 EVT N0VT = N0.getOperand(0).getValueType();
5072 EVT EltVT = VT.getVectorElementType();
5073 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
5074 DAG.getConstant(1, EltVT));
5075 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5076 // We know that the # elements of the results is the same as the
5077 // # elements of the compare (and the # elements of the compare result
5078 // for that matter). Check to see that they are the same size. If so,
5079 // we know that the element size of the sext'd result matches the
5080 // element size of the compare operands.
5081 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5082 DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5084 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
5085 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT,
5086 &OneOps[0], OneOps.size()));
5088 // If the desired elements are smaller or larger than the source
5089 // elements we can use a matching integer vector type and then
5090 // truncate/sign extend
5091 EVT MatchingElementType =
5092 EVT::getIntegerVT(*DAG.getContext(),
5093 N0VT.getScalarType().getSizeInBits());
5094 EVT MatchingVectorType =
5095 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
5096 N0VT.getVectorNumElements());
5098 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5100 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5101 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5102 DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT),
5103 DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT,
5104 &OneOps[0], OneOps.size()));
5107 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5109 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5110 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5111 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5112 if (SCC.getNode()) return SCC;
5115 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
5116 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
5117 isa<ConstantSDNode>(N0.getOperand(1)) &&
5118 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
5120 SDValue ShAmt = N0.getOperand(1);
5121 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
5122 if (N0.getOpcode() == ISD::SHL) {
5123 SDValue InnerZExt = N0.getOperand(0);
5124 // If the original shl may be shifting out bits, do not perform this
5126 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
5127 InnerZExt.getOperand(0).getValueType().getSizeInBits();
5128 if (ShAmtVal > KnownZeroBits)
5134 // Ensure that the shift amount is wide enough for the shifted value.
5135 if (VT.getSizeInBits() >= 256)
5136 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
5138 return DAG.getNode(N0.getOpcode(), DL, VT,
5139 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
5146 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
5147 SDValue N0 = N->getOperand(0);
5148 EVT VT = N->getValueType(0);
5150 // fold (aext c1) -> c1
5151 if (isa<ConstantSDNode>(N0))
5152 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, N0);
5153 // fold (aext (aext x)) -> (aext x)
5154 // fold (aext (zext x)) -> (zext x)
5155 // fold (aext (sext x)) -> (sext x)
5156 if (N0.getOpcode() == ISD::ANY_EXTEND ||
5157 N0.getOpcode() == ISD::ZERO_EXTEND ||
5158 N0.getOpcode() == ISD::SIGN_EXTEND)
5159 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT, N0.getOperand(0));
5161 // fold (aext (truncate (load x))) -> (aext (smaller load x))
5162 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
5163 if (N0.getOpcode() == ISD::TRUNCATE) {
5164 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
5165 if (NarrowLoad.getNode()) {
5166 SDNode* oye = N0.getNode()->getOperand(0).getNode();
5167 if (NarrowLoad.getNode() != N0.getNode()) {
5168 CombineTo(N0.getNode(), NarrowLoad);
5169 // CombineTo deleted the truncate, if needed, but not what's under it.
5172 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5176 // fold (aext (truncate x))
5177 if (N0.getOpcode() == ISD::TRUNCATE) {
5178 SDValue TruncOp = N0.getOperand(0);
5179 if (TruncOp.getValueType() == VT)
5180 return TruncOp; // x iff x size == zext size.
5181 if (TruncOp.getValueType().bitsGT(VT))
5182 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, TruncOp);
5183 return DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, TruncOp);
5186 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
5187 // if the trunc is not free.
5188 if (N0.getOpcode() == ISD::AND &&
5189 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
5190 N0.getOperand(1).getOpcode() == ISD::Constant &&
5191 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
5192 N0.getValueType())) {
5193 SDValue X = N0.getOperand(0).getOperand(0);
5194 if (X.getValueType().bitsLT(VT)) {
5195 X = DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, X);
5196 } else if (X.getValueType().bitsGT(VT)) {
5197 X = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, X);
5199 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
5200 Mask = Mask.zext(VT.getSizeInBits());
5201 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5202 X, DAG.getConstant(Mask, VT));
5205 // fold (aext (load x)) -> (aext (truncate (extload x)))
5206 // None of the supported targets knows how to perform load and any_ext
5207 // on vectors in one instruction. We only perform this transformation on
5209 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
5210 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5211 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
5212 bool DoXform = true;
5213 SmallVector<SDNode*, 4> SetCCs;
5214 if (!N0.hasOneUse())
5215 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
5217 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5218 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
5220 LN0->getBasePtr(), N0.getValueType(),
5221 LN0->getMemOperand());
5222 CombineTo(N, ExtLoad);
5223 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5224 N0.getValueType(), ExtLoad);
5225 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
5226 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, SDLoc(N),
5228 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5232 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
5233 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
5234 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
5235 if (N0.getOpcode() == ISD::LOAD &&
5236 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5238 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5239 EVT MemVT = LN0->getMemoryVT();
5240 SDValue ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(N),
5241 VT, LN0->getChain(), LN0->getBasePtr(),
5242 MemVT, LN0->getMemOperand());
5243 CombineTo(N, ExtLoad);
5244 CombineTo(N0.getNode(),
5245 DAG.getNode(ISD::TRUNCATE, SDLoc(N0),
5246 N0.getValueType(), ExtLoad),
5247 ExtLoad.getValue(1));
5248 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5251 if (N0.getOpcode() == ISD::SETCC) {
5252 // aext(setcc) -> sext_in_reg(vsetcc) for vectors.
5253 // Only do this before legalize for now.
5254 if (VT.isVector() && !LegalOperations) {
5255 EVT N0VT = N0.getOperand(0).getValueType();
5256 // We know that the # elements of the results is the same as the
5257 // # elements of the compare (and the # elements of the compare result
5258 // for that matter). Check to see that they are the same size. If so,
5259 // we know that the element size of the sext'd result matches the
5260 // element size of the compare operands.
5261 if (VT.getSizeInBits() == N0VT.getSizeInBits())
5262 return DAG.getSetCC(SDLoc(N), VT, N0.getOperand(0),
5264 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5265 // If the desired elements are smaller or larger than the source
5266 // elements we can use a matching integer vector type and then
5267 // truncate/sign extend
5269 EVT MatchingElementType =
5270 EVT::getIntegerVT(*DAG.getContext(),
5271 N0VT.getScalarType().getSizeInBits());
5272 EVT MatchingVectorType =
5273 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
5274 N0VT.getVectorNumElements());
5276 DAG.getSetCC(SDLoc(N), MatchingVectorType, N0.getOperand(0),
5278 cast<CondCodeSDNode>(N0.getOperand(2))->get());
5279 return DAG.getSExtOrTrunc(VsetCC, SDLoc(N), VT);
5283 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
5285 SimplifySelectCC(SDLoc(N), N0.getOperand(0), N0.getOperand(1),
5286 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
5287 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
5295 /// GetDemandedBits - See if the specified operand can be simplified with the
5296 /// knowledge that only the bits specified by Mask are used. If so, return the
5297 /// simpler operand, otherwise return a null SDValue.
5298 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
5299 switch (V.getOpcode()) {
5301 case ISD::Constant: {
5302 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
5303 assert(CV != 0 && "Const value should be ConstSDNode.");
5304 const APInt &CVal = CV->getAPIntValue();
5305 APInt NewVal = CVal & Mask;
5307 return DAG.getConstant(NewVal, V.getValueType());
5312 // If the LHS or RHS don't contribute bits to the or, drop them.
5313 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
5314 return V.getOperand(1);
5315 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
5316 return V.getOperand(0);
5319 // Only look at single-use SRLs.
5320 if (!V.getNode()->hasOneUse())
5322 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
5323 // See if we can recursively simplify the LHS.
5324 unsigned Amt = RHSC->getZExtValue();
5326 // Watch out for shift count overflow though.
5327 if (Amt >= Mask.getBitWidth()) break;
5328 APInt NewMask = Mask << Amt;
5329 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
5330 if (SimplifyLHS.getNode())
5331 return DAG.getNode(ISD::SRL, SDLoc(V), V.getValueType(),
5332 SimplifyLHS, V.getOperand(1));
5338 /// ReduceLoadWidth - If the result of a wider load is shifted to right of N
5339 /// bits and then truncated to a narrower type and where N is a multiple
5340 /// of number of bits of the narrower type, transform it to a narrower load
5341 /// from address + N / num of bits of new type. If the result is to be
5342 /// extended, also fold the extension to form a extending load.
5343 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
5344 unsigned Opc = N->getOpcode();
5346 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
5347 SDValue N0 = N->getOperand(0);
5348 EVT VT = N->getValueType(0);
5351 // This transformation isn't valid for vector loads.
5355 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
5357 if (Opc == ISD::SIGN_EXTEND_INREG) {
5358 ExtType = ISD::SEXTLOAD;
5359 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
5360 } else if (Opc == ISD::SRL) {
5361 // Another special-case: SRL is basically zero-extending a narrower value.
5362 ExtType = ISD::ZEXTLOAD;
5364 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
5365 if (!N01) return SDValue();
5366 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
5367 VT.getSizeInBits() - N01->getZExtValue());
5369 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, ExtVT))
5372 unsigned EVTBits = ExtVT.getSizeInBits();
5374 // Do not generate loads of non-round integer types since these can
5375 // be expensive (and would be wrong if the type is not byte sized).
5376 if (!ExtVT.isRound())
5380 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
5381 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5382 ShAmt = N01->getZExtValue();
5383 // Is the shift amount a multiple of size of VT?
5384 if ((ShAmt & (EVTBits-1)) == 0) {
5385 N0 = N0.getOperand(0);
5386 // Is the load width a multiple of size of VT?
5387 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
5391 // At this point, we must have a load or else we can't do the transform.
5392 if (!isa<LoadSDNode>(N0)) return SDValue();
5394 // Because a SRL must be assumed to *need* to zero-extend the high bits
5395 // (as opposed to anyext the high bits), we can't combine the zextload
5396 // lowering of SRL and an sextload.
5397 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
5400 // If the shift amount is larger than the input type then we're not
5401 // accessing any of the loaded bytes. If the load was a zextload/extload
5402 // then the result of the shift+trunc is zero/undef (handled elsewhere).
5403 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
5408 // If the load is shifted left (and the result isn't shifted back right),
5409 // we can fold the truncate through the shift.
5410 unsigned ShLeftAmt = 0;
5411 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
5412 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
5413 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5414 ShLeftAmt = N01->getZExtValue();
5415 N0 = N0.getOperand(0);
5419 // If we haven't found a load, we can't narrow it. Don't transform one with
5420 // multiple uses, this would require adding a new load.
5421 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse())
5424 // Don't change the width of a volatile load.
5425 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5426 if (LN0->isVolatile())
5429 // Verify that we are actually reducing a load width here.
5430 if (LN0->getMemoryVT().getSizeInBits() < EVTBits)
5433 // For the transform to be legal, the load must produce only two values
5434 // (the value loaded and the chain). Don't transform a pre-increment
5435 // load, for example, which produces an extra value. Otherwise the
5436 // transformation is not equivalent, and the downstream logic to replace
5437 // uses gets things wrong.
5438 if (LN0->getNumValues() > 2)
5441 // If the load that we're shrinking is an extload and we're not just
5442 // discarding the extension we can't simply shrink the load. Bail.
5443 // TODO: It would be possible to merge the extensions in some cases.
5444 if (LN0->getExtensionType() != ISD::NON_EXTLOAD &&
5445 LN0->getMemoryVT().getSizeInBits() < ExtVT.getSizeInBits() + ShAmt)
5448 EVT PtrType = N0.getOperand(1).getValueType();
5450 if (PtrType == MVT::Untyped || PtrType.isExtended())
5451 // It's not possible to generate a constant of extended or untyped type.
5454 // For big endian targets, we need to adjust the offset to the pointer to
5455 // load the correct bytes.
5456 if (TLI.isBigEndian()) {
5457 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
5458 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
5459 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
5462 uint64_t PtrOff = ShAmt / 8;
5463 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
5464 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LN0),
5465 PtrType, LN0->getBasePtr(),
5466 DAG.getConstant(PtrOff, PtrType));
5467 AddToWorkList(NewPtr.getNode());
5470 if (ExtType == ISD::NON_EXTLOAD)
5471 Load = DAG.getLoad(VT, SDLoc(N0), LN0->getChain(), NewPtr,
5472 LN0->getPointerInfo().getWithOffset(PtrOff),
5473 LN0->isVolatile(), LN0->isNonTemporal(),
5474 LN0->isInvariant(), NewAlign, LN0->getTBAAInfo());
5476 Load = DAG.getExtLoad(ExtType, SDLoc(N0), VT, LN0->getChain(),NewPtr,
5477 LN0->getPointerInfo().getWithOffset(PtrOff),
5478 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
5479 NewAlign, LN0->getTBAAInfo());
5481 // Replace the old load's chain with the new load's chain.
5482 WorkListRemover DeadNodes(*this);
5483 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
5485 // Shift the result left, if we've swallowed a left shift.
5486 SDValue Result = Load;
5487 if (ShLeftAmt != 0) {
5488 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
5489 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
5491 // If the shift amount is as large as the result size (but, presumably,
5492 // no larger than the source) then the useful bits of the result are
5493 // zero; we can't simply return the shortened shift, because the result
5494 // of that operation is undefined.
5495 if (ShLeftAmt >= VT.getSizeInBits())
5496 Result = DAG.getConstant(0, VT);
5498 Result = DAG.getNode(ISD::SHL, SDLoc(N0), VT,
5499 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
5502 // Return the new loaded value.
5506 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
5507 SDValue N0 = N->getOperand(0);
5508 SDValue N1 = N->getOperand(1);
5509 EVT VT = N->getValueType(0);
5510 EVT EVT = cast<VTSDNode>(N1)->getVT();
5511 unsigned VTBits = VT.getScalarType().getSizeInBits();
5512 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
5514 // fold (sext_in_reg c1) -> c1
5515 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
5516 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT, N0, N1);
5518 // If the input is already sign extended, just drop the extension.
5519 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
5522 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
5523 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
5524 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT()))
5525 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
5526 N0.getOperand(0), N1);
5528 // fold (sext_in_reg (sext x)) -> (sext x)
5529 // fold (sext_in_reg (aext x)) -> (sext x)
5530 // if x is small enough.
5531 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
5532 SDValue N00 = N0.getOperand(0);
5533 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
5534 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
5535 return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
5538 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
5539 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
5540 return DAG.getZeroExtendInReg(N0, SDLoc(N), EVT);
5542 // fold operands of sext_in_reg based on knowledge that the top bits are not
5544 if (SimplifyDemandedBits(SDValue(N, 0)))
5545 return SDValue(N, 0);
5547 // fold (sext_in_reg (load x)) -> (smaller sextload x)
5548 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
5549 SDValue NarrowLoad = ReduceLoadWidth(N);
5550 if (NarrowLoad.getNode())
5553 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
5554 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
5555 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
5556 if (N0.getOpcode() == ISD::SRL) {
5557 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
5558 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
5559 // We can turn this into an SRA iff the input to the SRL is already sign
5561 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
5562 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
5563 return DAG.getNode(ISD::SRA, SDLoc(N), VT,
5564 N0.getOperand(0), N0.getOperand(1));
5568 // fold (sext_inreg (extload x)) -> (sextload x)
5569 if (ISD::isEXTLoad(N0.getNode()) &&
5570 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5571 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5572 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5573 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5574 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5575 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5577 LN0->getBasePtr(), EVT,
5578 LN0->getMemOperand());
5579 CombineTo(N, ExtLoad);
5580 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5581 AddToWorkList(ExtLoad.getNode());
5582 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5584 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
5585 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5587 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5588 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5589 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5590 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5591 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, SDLoc(N), VT,
5593 LN0->getBasePtr(), EVT,
5594 LN0->getMemOperand());
5595 CombineTo(N, ExtLoad);
5596 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5597 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5600 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
5601 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
5602 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
5603 N0.getOperand(1), false);
5604 if (BSwap.getNode() != 0)
5605 return DAG.getNode(ISD::SIGN_EXTEND_INREG, SDLoc(N), VT,
5609 // Fold a sext_inreg of a build_vector of ConstantSDNodes or undefs
5610 // into a build_vector.
5611 if (ISD::isBuildVectorOfConstantSDNodes(N0.getNode())) {
5612 SmallVector<SDValue, 8> Elts;
5613 unsigned NumElts = N0->getNumOperands();
5614 unsigned ShAmt = VTBits - EVTBits;
5616 for (unsigned i = 0; i != NumElts; ++i) {
5617 SDValue Op = N0->getOperand(i);
5618 if (Op->getOpcode() == ISD::UNDEF) {
5623 ConstantSDNode *CurrentND = cast<ConstantSDNode>(Op);
5624 const APInt &C = APInt(VTBits, CurrentND->getAPIntValue().getZExtValue());
5625 Elts.push_back(DAG.getConstant(C.shl(ShAmt).ashr(ShAmt).getZExtValue(),
5626 Op.getValueType()));
5629 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, &Elts[0], NumElts);
5635 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
5636 SDValue N0 = N->getOperand(0);
5637 EVT VT = N->getValueType(0);
5638 bool isLE = TLI.isLittleEndian();
5641 if (N0.getValueType() == N->getValueType(0))
5643 // fold (truncate c1) -> c1
5644 if (isa<ConstantSDNode>(N0))
5645 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0);
5646 // fold (truncate (truncate x)) -> (truncate x)
5647 if (N0.getOpcode() == ISD::TRUNCATE)
5648 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
5649 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
5650 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
5651 N0.getOpcode() == ISD::SIGN_EXTEND ||
5652 N0.getOpcode() == ISD::ANY_EXTEND) {
5653 if (N0.getOperand(0).getValueType().bitsLT(VT))
5654 // if the source is smaller than the dest, we still need an extend
5655 return DAG.getNode(N0.getOpcode(), SDLoc(N), VT,
5657 if (N0.getOperand(0).getValueType().bitsGT(VT))
5658 // if the source is larger than the dest, than we just need the truncate
5659 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, N0.getOperand(0));
5660 // if the source and dest are the same type, we can drop both the extend
5661 // and the truncate.
5662 return N0.getOperand(0);
5665 // Fold extract-and-trunc into a narrow extract. For example:
5666 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
5667 // i32 y = TRUNCATE(i64 x)
5669 // v16i8 b = BITCAST (v2i64 val)
5670 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
5672 // Note: We only run this optimization after type legalization (which often
5673 // creates this pattern) and before operation legalization after which
5674 // we need to be more careful about the vector instructions that we generate.
5675 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5676 LegalTypes && !LegalOperations && N0->hasOneUse()) {
5678 EVT VecTy = N0.getOperand(0).getValueType();
5679 EVT ExTy = N0.getValueType();
5680 EVT TrTy = N->getValueType(0);
5682 unsigned NumElem = VecTy.getVectorNumElements();
5683 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
5685 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
5686 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
5688 SDValue EltNo = N0->getOperand(1);
5689 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
5690 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5691 EVT IndexTy = TLI.getVectorIdxTy();
5692 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
5694 SDValue V = DAG.getNode(ISD::BITCAST, SDLoc(N),
5695 NVT, N0.getOperand(0));
5697 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
5699 DAG.getConstant(Index, IndexTy));
5703 // Fold a series of buildvector, bitcast, and truncate if possible.
5705 // (2xi32 trunc (bitcast ((4xi32)buildvector x, x, y, y) 2xi64)) to
5706 // (2xi32 (buildvector x, y)).
5707 if (Level == AfterLegalizeVectorOps && VT.isVector() &&
5708 N0.getOpcode() == ISD::BITCAST && N0.hasOneUse() &&
5709 N0.getOperand(0).getOpcode() == ISD::BUILD_VECTOR &&
5710 N0.getOperand(0).hasOneUse()) {
5712 SDValue BuildVect = N0.getOperand(0);
5713 EVT BuildVectEltTy = BuildVect.getValueType().getVectorElementType();
5714 EVT TruncVecEltTy = VT.getVectorElementType();
5716 // Check that the element types match.
5717 if (BuildVectEltTy == TruncVecEltTy) {
5718 // Now we only need to compute the offset of the truncated elements.
5719 unsigned BuildVecNumElts = BuildVect.getNumOperands();
5720 unsigned TruncVecNumElts = VT.getVectorNumElements();
5721 unsigned TruncEltOffset = BuildVecNumElts / TruncVecNumElts;
5723 assert((BuildVecNumElts % TruncVecNumElts) == 0 &&
5724 "Invalid number of elements");
5726 SmallVector<SDValue, 8> Opnds;
5727 for (unsigned i = 0, e = BuildVecNumElts; i != e; i += TruncEltOffset)
5728 Opnds.push_back(BuildVect.getOperand(i));
5730 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N), VT, &Opnds[0],
5735 // See if we can simplify the input to this truncate through knowledge that
5736 // only the low bits are being used.
5737 // For example "trunc (or (shl x, 8), y)" // -> trunc y
5738 // Currently we only perform this optimization on scalars because vectors
5739 // may have different active low bits.
5740 if (!VT.isVector()) {
5742 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
5743 VT.getSizeInBits()));
5744 if (Shorter.getNode())
5745 return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, Shorter);
5747 // fold (truncate (load x)) -> (smaller load x)
5748 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
5749 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
5750 SDValue Reduced = ReduceLoadWidth(N);
5751 if (Reduced.getNode())
5753 // Handle the case where the load remains an extending load even
5754 // after truncation.
5755 if (N0.hasOneUse() && ISD::isUNINDEXEDLoad(N0.getNode())) {
5756 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5757 if (!LN0->isVolatile() &&
5758 LN0->getMemoryVT().getStoreSizeInBits() < VT.getSizeInBits()) {
5759 SDValue NewLoad = DAG.getExtLoad(LN0->getExtensionType(), SDLoc(LN0),
5760 VT, LN0->getChain(), LN0->getBasePtr(),
5762 LN0->getMemOperand());
5763 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLoad.getValue(1));
5768 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
5769 // where ... are all 'undef'.
5770 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
5771 SmallVector<EVT, 8> VTs;
5774 unsigned NumDefs = 0;
5776 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
5777 SDValue X = N0.getOperand(i);
5778 if (X.getOpcode() != ISD::UNDEF) {
5783 // Stop if more than one members are non-undef.
5786 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
5787 VT.getVectorElementType(),
5788 X.getValueType().getVectorNumElements()));
5792 return DAG.getUNDEF(VT);
5795 assert(V.getNode() && "The single defined operand is empty!");
5796 SmallVector<SDValue, 8> Opnds;
5797 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
5799 Opnds.push_back(DAG.getUNDEF(VTs[i]));
5802 SDValue NV = DAG.getNode(ISD::TRUNCATE, SDLoc(V), VTs[i], V);
5803 AddToWorkList(NV.getNode());
5804 Opnds.push_back(NV);
5806 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT,
5807 &Opnds[0], Opnds.size());
5811 // Simplify the operands using demanded-bits information.
5812 if (!VT.isVector() &&
5813 SimplifyDemandedBits(SDValue(N, 0)))
5814 return SDValue(N, 0);
5819 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
5820 SDValue Elt = N->getOperand(i);
5821 if (Elt.getOpcode() != ISD::MERGE_VALUES)
5822 return Elt.getNode();
5823 return Elt.getOperand(Elt.getResNo()).getNode();
5826 /// CombineConsecutiveLoads - build_pair (load, load) -> load
5827 /// if load locations are consecutive.
5828 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
5829 assert(N->getOpcode() == ISD::BUILD_PAIR);
5831 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
5832 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
5833 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
5834 LD1->getPointerInfo().getAddrSpace() !=
5835 LD2->getPointerInfo().getAddrSpace())
5837 EVT LD1VT = LD1->getValueType(0);
5839 if (ISD::isNON_EXTLoad(LD2) &&
5841 // If both are volatile this would reduce the number of volatile loads.
5842 // If one is volatile it might be ok, but play conservative and bail out.
5843 !LD1->isVolatile() &&
5844 !LD2->isVolatile() &&
5845 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
5846 unsigned Align = LD1->getAlignment();
5847 unsigned NewAlign = TLI.getDataLayout()->
5848 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
5850 if (NewAlign <= Align &&
5851 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
5852 return DAG.getLoad(VT, SDLoc(N), LD1->getChain(),
5853 LD1->getBasePtr(), LD1->getPointerInfo(),
5854 false, false, false, Align);
5860 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
5861 SDValue N0 = N->getOperand(0);
5862 EVT VT = N->getValueType(0);
5864 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
5865 // Only do this before legalize, since afterward the target may be depending
5866 // on the bitconvert.
5867 // First check to see if this is all constant.
5869 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
5871 bool isSimple = true;
5872 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i)
5873 if (N0.getOperand(i).getOpcode() != ISD::UNDEF &&
5874 N0.getOperand(i).getOpcode() != ISD::Constant &&
5875 N0.getOperand(i).getOpcode() != ISD::ConstantFP) {
5880 EVT DestEltVT = N->getValueType(0).getVectorElementType();
5881 assert(!DestEltVT.isVector() &&
5882 "Element type of vector ValueType must not be vector!");
5884 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
5887 // If the input is a constant, let getNode fold it.
5888 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
5889 SDValue Res = DAG.getNode(ISD::BITCAST, SDLoc(N), VT, N0);
5890 if (Res.getNode() != N) {
5891 if (!LegalOperations ||
5892 TLI.isOperationLegal(Res.getNode()->getOpcode(), VT))
5895 // Folding it resulted in an illegal node, and it's too late to
5896 // do that. Clean up the old node and forego the transformation.
5897 // Ideally this won't happen very often, because instcombine
5898 // and the earlier dagcombine runs (where illegal nodes are
5899 // permitted) should have folded most of them already.
5900 DAG.DeleteNode(Res.getNode());
5904 // (conv (conv x, t1), t2) -> (conv x, t2)
5905 if (N0.getOpcode() == ISD::BITCAST)
5906 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT,
5909 // fold (conv (load x)) -> (load (conv*)x)
5910 // If the resultant load doesn't need a higher alignment than the original!
5911 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
5912 // Do not change the width of a volatile load.
5913 !cast<LoadSDNode>(N0)->isVolatile() &&
5914 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)) &&
5915 TLI.isLoadBitCastBeneficial(N0.getValueType(), VT)) {
5916 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5917 unsigned Align = TLI.getDataLayout()->
5918 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
5919 unsigned OrigAlign = LN0->getAlignment();
5921 if (Align <= OrigAlign) {
5922 SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(),
5923 LN0->getBasePtr(), LN0->getPointerInfo(),
5924 LN0->isVolatile(), LN0->isNonTemporal(),
5925 LN0->isInvariant(), OrigAlign,
5926 LN0->getTBAAInfo());
5928 CombineTo(N0.getNode(),
5929 DAG.getNode(ISD::BITCAST, SDLoc(N0),
5930 N0.getValueType(), Load),
5936 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
5937 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
5938 // This often reduces constant pool loads.
5939 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(N0.getValueType())) ||
5940 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(N0.getValueType()))) &&
5941 N0.getNode()->hasOneUse() && VT.isInteger() &&
5942 !VT.isVector() && !N0.getValueType().isVector()) {
5943 SDValue NewConv = DAG.getNode(ISD::BITCAST, SDLoc(N0), VT,
5945 AddToWorkList(NewConv.getNode());
5947 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
5948 if (N0.getOpcode() == ISD::FNEG)
5949 return DAG.getNode(ISD::XOR, SDLoc(N), VT,
5950 NewConv, DAG.getConstant(SignBit, VT));
5951 assert(N0.getOpcode() == ISD::FABS);
5952 return DAG.getNode(ISD::AND, SDLoc(N), VT,
5953 NewConv, DAG.getConstant(~SignBit, VT));
5956 // fold (bitconvert (fcopysign cst, x)) ->
5957 // (or (and (bitconvert x), sign), (and cst, (not sign)))
5958 // Note that we don't handle (copysign x, cst) because this can always be
5959 // folded to an fneg or fabs.
5960 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
5961 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
5962 VT.isInteger() && !VT.isVector()) {
5963 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
5964 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
5965 if (isTypeLegal(IntXVT)) {
5966 SDValue X = DAG.getNode(ISD::BITCAST, SDLoc(N0),
5967 IntXVT, N0.getOperand(1));
5968 AddToWorkList(X.getNode());
5970 // If X has a different width than the result/lhs, sext it or truncate it.
5971 unsigned VTWidth = VT.getSizeInBits();
5972 if (OrigXWidth < VTWidth) {
5973 X = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, X);
5974 AddToWorkList(X.getNode());
5975 } else if (OrigXWidth > VTWidth) {
5976 // To get the sign bit in the right place, we have to shift it right
5977 // before truncating.
5978 X = DAG.getNode(ISD::SRL, SDLoc(X),
5979 X.getValueType(), X,
5980 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
5981 AddToWorkList(X.getNode());
5982 X = DAG.getNode(ISD::TRUNCATE, SDLoc(X), VT, X);
5983 AddToWorkList(X.getNode());
5986 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
5987 X = DAG.getNode(ISD::AND, SDLoc(X), VT,
5988 X, DAG.getConstant(SignBit, VT));
5989 AddToWorkList(X.getNode());
5991 SDValue Cst = DAG.getNode(ISD::BITCAST, SDLoc(N0),
5992 VT, N0.getOperand(0));
5993 Cst = DAG.getNode(ISD::AND, SDLoc(Cst), VT,
5994 Cst, DAG.getConstant(~SignBit, VT));
5995 AddToWorkList(Cst.getNode());
5997 return DAG.getNode(ISD::OR, SDLoc(N), VT, X, Cst);
6001 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
6002 if (N0.getOpcode() == ISD::BUILD_PAIR) {
6003 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
6004 if (CombineLD.getNode())
6011 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
6012 EVT VT = N->getValueType(0);
6013 return CombineConsecutiveLoads(N, VT);
6016 /// ConstantFoldBITCASTofBUILD_VECTOR - We know that BV is a build_vector
6017 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
6018 /// destination element value type.
6019 SDValue DAGCombiner::
6020 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
6021 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
6023 // If this is already the right type, we're done.
6024 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
6026 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
6027 unsigned DstBitSize = DstEltVT.getSizeInBits();
6029 // If this is a conversion of N elements of one type to N elements of another
6030 // type, convert each element. This handles FP<->INT cases.
6031 if (SrcBitSize == DstBitSize) {
6032 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6033 BV->getValueType(0).getVectorNumElements());
6035 // Due to the FP element handling below calling this routine recursively,
6036 // we can end up with a scalar-to-vector node here.
6037 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
6038 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6039 DAG.getNode(ISD::BITCAST, SDLoc(BV),
6040 DstEltVT, BV->getOperand(0)));
6042 SmallVector<SDValue, 8> Ops;
6043 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6044 SDValue Op = BV->getOperand(i);
6045 // If the vector element type is not legal, the BUILD_VECTOR operands
6046 // are promoted and implicitly truncated. Make that explicit here.
6047 if (Op.getValueType() != SrcEltVT)
6048 Op = DAG.getNode(ISD::TRUNCATE, SDLoc(BV), SrcEltVT, Op);
6049 Ops.push_back(DAG.getNode(ISD::BITCAST, SDLoc(BV),
6051 AddToWorkList(Ops.back().getNode());
6053 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT,
6054 &Ops[0], Ops.size());
6057 // Otherwise, we're growing or shrinking the elements. To avoid having to
6058 // handle annoying details of growing/shrinking FP values, we convert them to
6060 if (SrcEltVT.isFloatingPoint()) {
6061 // Convert the input float vector to a int vector where the elements are the
6063 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
6064 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
6065 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
6069 // Now we know the input is an integer vector. If the output is a FP type,
6070 // convert to integer first, then to FP of the right size.
6071 if (DstEltVT.isFloatingPoint()) {
6072 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
6073 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
6074 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
6076 // Next, convert to FP elements of the same size.
6077 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
6080 // Okay, we know the src/dst types are both integers of differing types.
6081 // Handling growing first.
6082 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
6083 if (SrcBitSize < DstBitSize) {
6084 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
6086 SmallVector<SDValue, 8> Ops;
6087 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
6088 i += NumInputsPerOutput) {
6089 bool isLE = TLI.isLittleEndian();
6090 APInt NewBits = APInt(DstBitSize, 0);
6091 bool EltIsUndef = true;
6092 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
6093 // Shift the previously computed bits over.
6094 NewBits <<= SrcBitSize;
6095 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
6096 if (Op.getOpcode() == ISD::UNDEF) continue;
6099 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
6100 zextOrTrunc(SrcBitSize).zext(DstBitSize);
6104 Ops.push_back(DAG.getUNDEF(DstEltVT));
6106 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
6109 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
6110 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT,
6111 &Ops[0], Ops.size());
6114 // Finally, this must be the case where we are shrinking elements: each input
6115 // turns into multiple outputs.
6116 bool isS2V = ISD::isScalarToVector(BV);
6117 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
6118 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
6119 NumOutputsPerInput*BV->getNumOperands());
6120 SmallVector<SDValue, 8> Ops;
6122 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
6123 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
6124 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
6125 Ops.push_back(DAG.getUNDEF(DstEltVT));
6129 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
6130 getAPIntValue().zextOrTrunc(SrcBitSize);
6132 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
6133 APInt ThisVal = OpVal.trunc(DstBitSize);
6134 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
6135 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
6136 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
6137 return DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(BV), VT,
6139 OpVal = OpVal.lshr(DstBitSize);
6142 // For big endian targets, swap the order of the pieces of each element.
6143 if (TLI.isBigEndian())
6144 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
6147 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(BV), VT,
6148 &Ops[0], Ops.size());
6151 SDValue DAGCombiner::visitFADD(SDNode *N) {
6152 SDValue N0 = N->getOperand(0);
6153 SDValue N1 = N->getOperand(1);
6154 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6155 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6156 EVT VT = N->getValueType(0);
6159 if (VT.isVector()) {
6160 SDValue FoldedVOp = SimplifyVBinOp(N);
6161 if (FoldedVOp.getNode()) return FoldedVOp;
6164 // fold (fadd c1, c2) -> c1 + c2
6166 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N1);
6167 // canonicalize constant to RHS
6168 if (N0CFP && !N1CFP)
6169 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N0);
6170 // fold (fadd A, 0) -> A
6171 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6172 N1CFP->getValueAPF().isZero())
6174 // fold (fadd A, (fneg B)) -> (fsub A, B)
6175 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6176 isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
6177 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0,
6178 GetNegatedExpression(N1, DAG, LegalOperations));
6179 // fold (fadd (fneg A), B) -> (fsub B, A)
6180 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
6181 isNegatibleForFree(N0, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
6182 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N1,
6183 GetNegatedExpression(N0, DAG, LegalOperations));
6185 // If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
6186 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6187 N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
6188 isa<ConstantFPSDNode>(N0.getOperand(1)))
6189 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0.getOperand(0),
6190 DAG.getNode(ISD::FADD, SDLoc(N), VT,
6191 N0.getOperand(1), N1));
6193 // No FP constant should be created after legalization as Instruction
6194 // Selection pass has hard time in dealing with FP constant.
6196 // We don't need test this condition for transformation like following, as
6197 // the DAG being transformed implies it is legal to take FP constant as
6200 // (fadd (fmul c, x), x) -> (fmul c+1, x)
6202 bool AllowNewFpConst = (Level < AfterLegalizeDAG);
6204 // If allow, fold (fadd (fneg x), x) -> 0.0
6205 if (AllowNewFpConst && DAG.getTarget().Options.UnsafeFPMath &&
6206 N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1)
6207 return DAG.getConstantFP(0.0, VT);
6209 // If allow, fold (fadd x, (fneg x)) -> 0.0
6210 if (AllowNewFpConst && DAG.getTarget().Options.UnsafeFPMath &&
6211 N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0)
6212 return DAG.getConstantFP(0.0, VT);
6214 // In unsafe math mode, we can fold chains of FADD's of the same value
6215 // into multiplications. This transform is not safe in general because
6216 // we are reducing the number of rounding steps.
6217 if (DAG.getTarget().Options.UnsafeFPMath &&
6218 TLI.isOperationLegalOrCustom(ISD::FMUL, VT) &&
6220 if (N0.getOpcode() == ISD::FMUL) {
6221 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
6222 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
6224 // (fadd (fmul c, x), x) -> (fmul x, c+1)
6225 if (CFP00 && !CFP01 && N0.getOperand(1) == N1) {
6226 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6228 DAG.getConstantFP(1.0, VT));
6229 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6233 // (fadd (fmul x, c), x) -> (fmul x, c+1)
6234 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
6235 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6237 DAG.getConstantFP(1.0, VT));
6238 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6242 // (fadd (fmul c, x), (fadd x, x)) -> (fmul x, c+2)
6243 if (CFP00 && !CFP01 && N1.getOpcode() == ISD::FADD &&
6244 N1.getOperand(0) == N1.getOperand(1) &&
6245 N0.getOperand(1) == N1.getOperand(0)) {
6246 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6248 DAG.getConstantFP(2.0, VT));
6249 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6250 N0.getOperand(1), NewCFP);
6253 // (fadd (fmul x, c), (fadd x, x)) -> (fmul x, c+2)
6254 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
6255 N1.getOperand(0) == N1.getOperand(1) &&
6256 N0.getOperand(0) == N1.getOperand(0)) {
6257 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6259 DAG.getConstantFP(2.0, VT));
6260 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6261 N0.getOperand(0), NewCFP);
6265 if (N1.getOpcode() == ISD::FMUL) {
6266 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
6267 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
6269 // (fadd x, (fmul c, x)) -> (fmul x, c+1)
6270 if (CFP10 && !CFP11 && N1.getOperand(1) == N0) {
6271 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6273 DAG.getConstantFP(1.0, VT));
6274 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6278 // (fadd x, (fmul x, c)) -> (fmul x, c+1)
6279 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
6280 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6282 DAG.getConstantFP(1.0, VT));
6283 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6288 // (fadd (fadd x, x), (fmul c, x)) -> (fmul x, c+2)
6289 if (CFP10 && !CFP11 && N0.getOpcode() == ISD::FADD &&
6290 N0.getOperand(0) == N0.getOperand(1) &&
6291 N1.getOperand(1) == N0.getOperand(0)) {
6292 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6294 DAG.getConstantFP(2.0, VT));
6295 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6296 N1.getOperand(1), NewCFP);
6299 // (fadd (fadd x, x), (fmul x, c)) -> (fmul x, c+2)
6300 if (CFP11 && !CFP10 && N0.getOpcode() == ISD::FADD &&
6301 N0.getOperand(0) == N0.getOperand(1) &&
6302 N1.getOperand(0) == N0.getOperand(0)) {
6303 SDValue NewCFP = DAG.getNode(ISD::FADD, SDLoc(N), VT,
6305 DAG.getConstantFP(2.0, VT));
6306 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6307 N1.getOperand(0), NewCFP);
6311 if (N0.getOpcode() == ISD::FADD && AllowNewFpConst) {
6312 ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
6313 // (fadd (fadd x, x), x) -> (fmul x, 3.0)
6314 if (!CFP && N0.getOperand(0) == N0.getOperand(1) &&
6315 (N0.getOperand(0) == N1))
6316 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6317 N1, DAG.getConstantFP(3.0, VT));
6320 if (N1.getOpcode() == ISD::FADD && AllowNewFpConst) {
6321 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
6322 // (fadd x, (fadd x, x)) -> (fmul x, 3.0)
6323 if (!CFP10 && N1.getOperand(0) == N1.getOperand(1) &&
6324 N1.getOperand(0) == N0)
6325 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6326 N0, DAG.getConstantFP(3.0, VT));
6329 // (fadd (fadd x, x), (fadd x, x)) -> (fmul x, 4.0)
6330 if (AllowNewFpConst &&
6331 N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
6332 N0.getOperand(0) == N0.getOperand(1) &&
6333 N1.getOperand(0) == N1.getOperand(1) &&
6334 N0.getOperand(0) == N1.getOperand(0))
6335 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6337 DAG.getConstantFP(4.0, VT));
6340 // FADD -> FMA combines:
6341 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
6342 DAG.getTarget().Options.UnsafeFPMath) &&
6343 DAG.getTarget().getTargetLowering()->isFMAFasterThanFMulAndFAdd(VT) &&
6344 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
6346 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
6347 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse())
6348 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6349 N0.getOperand(0), N0.getOperand(1), N1);
6351 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
6352 // Note: Commutes FADD operands.
6353 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse())
6354 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
6355 N1.getOperand(0), N1.getOperand(1), N0);
6361 SDValue DAGCombiner::visitFSUB(SDNode *N) {
6362 SDValue N0 = N->getOperand(0);
6363 SDValue N1 = N->getOperand(1);
6364 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6365 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6366 EVT VT = N->getValueType(0);
6370 if (VT.isVector()) {
6371 SDValue FoldedVOp = SimplifyVBinOp(N);
6372 if (FoldedVOp.getNode()) return FoldedVOp;
6375 // fold (fsub c1, c2) -> c1-c2
6377 return DAG.getNode(ISD::FSUB, SDLoc(N), VT, N0, N1);
6378 // fold (fsub A, 0) -> A
6379 if (DAG.getTarget().Options.UnsafeFPMath &&
6380 N1CFP && N1CFP->getValueAPF().isZero())
6382 // fold (fsub 0, B) -> -B
6383 if (DAG.getTarget().Options.UnsafeFPMath &&
6384 N0CFP && N0CFP->getValueAPF().isZero()) {
6385 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
6386 return GetNegatedExpression(N1, DAG, LegalOperations);
6387 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6388 return DAG.getNode(ISD::FNEG, dl, VT, N1);
6390 // fold (fsub A, (fneg B)) -> (fadd A, B)
6391 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
6392 return DAG.getNode(ISD::FADD, dl, VT, N0,
6393 GetNegatedExpression(N1, DAG, LegalOperations));
6395 // If 'unsafe math' is enabled, fold
6396 // (fsub x, x) -> 0.0 &
6397 // (fsub x, (fadd x, y)) -> (fneg y) &
6398 // (fsub x, (fadd y, x)) -> (fneg y)
6399 if (DAG.getTarget().Options.UnsafeFPMath) {
6401 return DAG.getConstantFP(0.0f, VT);
6403 if (N1.getOpcode() == ISD::FADD) {
6404 SDValue N10 = N1->getOperand(0);
6405 SDValue N11 = N1->getOperand(1);
6407 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI,
6408 &DAG.getTarget().Options))
6409 return GetNegatedExpression(N11, DAG, LegalOperations);
6411 if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI,
6412 &DAG.getTarget().Options))
6413 return GetNegatedExpression(N10, DAG, LegalOperations);
6417 // FSUB -> FMA combines:
6418 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
6419 DAG.getTarget().Options.UnsafeFPMath) &&
6420 DAG.getTarget().getTargetLowering()->isFMAFasterThanFMulAndFAdd(VT) &&
6421 (!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FMA, VT))) {
6423 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
6424 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse())
6425 return DAG.getNode(ISD::FMA, dl, VT,
6426 N0.getOperand(0), N0.getOperand(1),
6427 DAG.getNode(ISD::FNEG, dl, VT, N1));
6429 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
6430 // Note: Commutes FSUB operands.
6431 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse())
6432 return DAG.getNode(ISD::FMA, dl, VT,
6433 DAG.getNode(ISD::FNEG, dl, VT,
6435 N1.getOperand(1), N0);
6437 // fold (fsub (fneg (fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
6438 if (N0.getOpcode() == ISD::FNEG &&
6439 N0.getOperand(0).getOpcode() == ISD::FMUL &&
6440 N0->hasOneUse() && N0.getOperand(0).hasOneUse()) {
6441 SDValue N00 = N0.getOperand(0).getOperand(0);
6442 SDValue N01 = N0.getOperand(0).getOperand(1);
6443 return DAG.getNode(ISD::FMA, dl, VT,
6444 DAG.getNode(ISD::FNEG, dl, VT, N00), N01,
6445 DAG.getNode(ISD::FNEG, dl, VT, N1));
6452 SDValue DAGCombiner::visitFMUL(SDNode *N) {
6453 SDValue N0 = N->getOperand(0);
6454 SDValue N1 = N->getOperand(1);
6455 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6456 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6457 EVT VT = N->getValueType(0);
6458 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6461 if (VT.isVector()) {
6462 SDValue FoldedVOp = SimplifyVBinOp(N);
6463 if (FoldedVOp.getNode()) return FoldedVOp;
6466 // fold (fmul c1, c2) -> c1*c2
6468 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0, N1);
6469 // canonicalize constant to RHS
6470 if (N0CFP && !N1CFP)
6471 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N1, N0);
6472 // fold (fmul A, 0) -> 0
6473 if (DAG.getTarget().Options.UnsafeFPMath &&
6474 N1CFP && N1CFP->getValueAPF().isZero())
6476 // fold (fmul A, 0) -> 0, vector edition.
6477 if (DAG.getTarget().Options.UnsafeFPMath &&
6478 ISD::isBuildVectorAllZeros(N1.getNode()))
6480 // fold (fmul A, 1.0) -> A
6481 if (N1CFP && N1CFP->isExactlyValue(1.0))
6483 // fold (fmul X, 2.0) -> (fadd X, X)
6484 if (N1CFP && N1CFP->isExactlyValue(+2.0))
6485 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N0);
6486 // fold (fmul X, -1.0) -> (fneg X)
6487 if (N1CFP && N1CFP->isExactlyValue(-1.0))
6488 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6489 return DAG.getNode(ISD::FNEG, SDLoc(N), VT, N0);
6491 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
6492 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
6493 &DAG.getTarget().Options)) {
6494 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
6495 &DAG.getTarget().Options)) {
6496 // Both can be negated for free, check to see if at least one is cheaper
6498 if (LHSNeg == 2 || RHSNeg == 2)
6499 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6500 GetNegatedExpression(N0, DAG, LegalOperations),
6501 GetNegatedExpression(N1, DAG, LegalOperations));
6505 // If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
6506 if (DAG.getTarget().Options.UnsafeFPMath &&
6507 N1CFP && N0.getOpcode() == ISD::FMUL &&
6508 N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
6509 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0.getOperand(0),
6510 DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6511 N0.getOperand(1), N1));
6516 SDValue DAGCombiner::visitFMA(SDNode *N) {
6517 SDValue N0 = N->getOperand(0);
6518 SDValue N1 = N->getOperand(1);
6519 SDValue N2 = N->getOperand(2);
6520 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6521 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6522 EVT VT = N->getValueType(0);
6525 if (DAG.getTarget().Options.UnsafeFPMath) {
6526 if (N0CFP && N0CFP->isZero())
6528 if (N1CFP && N1CFP->isZero())
6531 if (N0CFP && N0CFP->isExactlyValue(1.0))
6532 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N1, N2);
6533 if (N1CFP && N1CFP->isExactlyValue(1.0))
6534 return DAG.getNode(ISD::FADD, SDLoc(N), VT, N0, N2);
6536 // Canonicalize (fma c, x, y) -> (fma x, c, y)
6537 if (N0CFP && !N1CFP)
6538 return DAG.getNode(ISD::FMA, SDLoc(N), VT, N1, N0, N2);
6540 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
6541 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6542 N2.getOpcode() == ISD::FMUL &&
6543 N0 == N2.getOperand(0) &&
6544 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
6545 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6546 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
6550 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
6551 if (DAG.getTarget().Options.UnsafeFPMath &&
6552 N0.getOpcode() == ISD::FMUL && N1CFP &&
6553 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
6554 return DAG.getNode(ISD::FMA, dl, VT,
6556 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
6560 // (fma x, 1, y) -> (fadd x, y)
6561 // (fma x, -1, y) -> (fadd (fneg x), y)
6563 if (N1CFP->isExactlyValue(1.0))
6564 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
6566 if (N1CFP->isExactlyValue(-1.0) &&
6567 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
6568 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
6569 AddToWorkList(RHSNeg.getNode());
6570 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
6574 // (fma x, c, x) -> (fmul x, (c+1))
6575 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP && N0 == N2)
6576 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6577 DAG.getNode(ISD::FADD, dl, VT,
6578 N1, DAG.getConstantFP(1.0, VT)));
6580 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
6581 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6582 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0)
6583 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6584 DAG.getNode(ISD::FADD, dl, VT,
6585 N1, DAG.getConstantFP(-1.0, VT)));
6591 SDValue DAGCombiner::visitFDIV(SDNode *N) {
6592 SDValue N0 = N->getOperand(0);
6593 SDValue N1 = N->getOperand(1);
6594 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6595 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6596 EVT VT = N->getValueType(0);
6597 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6600 if (VT.isVector()) {
6601 SDValue FoldedVOp = SimplifyVBinOp(N);
6602 if (FoldedVOp.getNode()) return FoldedVOp;
6605 // fold (fdiv c1, c2) -> c1/c2
6607 return DAG.getNode(ISD::FDIV, SDLoc(N), VT, N0, N1);
6609 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
6610 if (N1CFP && DAG.getTarget().Options.UnsafeFPMath) {
6611 // Compute the reciprocal 1.0 / c2.
6612 APFloat N1APF = N1CFP->getValueAPF();
6613 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
6614 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
6615 // Only do the transform if the reciprocal is a legal fp immediate that
6616 // isn't too nasty (eg NaN, denormal, ...).
6617 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
6618 (!LegalOperations ||
6619 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
6620 // backend)... we should handle this gracefully after Legalize.
6621 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
6622 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
6623 TLI.isFPImmLegal(Recip, VT)))
6624 return DAG.getNode(ISD::FMUL, SDLoc(N), VT, N0,
6625 DAG.getConstantFP(Recip, VT));
6628 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
6629 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
6630 &DAG.getTarget().Options)) {
6631 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
6632 &DAG.getTarget().Options)) {
6633 // Both can be negated for free, check to see if at least one is cheaper
6635 if (LHSNeg == 2 || RHSNeg == 2)
6636 return DAG.getNode(ISD::FDIV, SDLoc(N), VT,
6637 GetNegatedExpression(N0, DAG, LegalOperations),
6638 GetNegatedExpression(N1, DAG, LegalOperations));
6645 SDValue DAGCombiner::visitFREM(SDNode *N) {
6646 SDValue N0 = N->getOperand(0);
6647 SDValue N1 = N->getOperand(1);
6648 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6649 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6650 EVT VT = N->getValueType(0);
6652 // fold (frem c1, c2) -> fmod(c1,c2)
6654 return DAG.getNode(ISD::FREM, SDLoc(N), VT, N0, N1);
6659 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
6660 SDValue N0 = N->getOperand(0);
6661 SDValue N1 = N->getOperand(1);
6662 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6663 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6664 EVT VT = N->getValueType(0);
6666 if (N0CFP && N1CFP) // Constant fold
6667 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT, N0, N1);
6670 const APFloat& V = N1CFP->getValueAPF();
6671 // copysign(x, c1) -> fabs(x) iff ispos(c1)
6672 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
6673 if (!V.isNegative()) {
6674 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
6675 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
6677 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6678 return DAG.getNode(ISD::FNEG, SDLoc(N), VT,
6679 DAG.getNode(ISD::FABS, SDLoc(N0), VT, N0));
6683 // copysign(fabs(x), y) -> copysign(x, y)
6684 // copysign(fneg(x), y) -> copysign(x, y)
6685 // copysign(copysign(x,z), y) -> copysign(x, y)
6686 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
6687 N0.getOpcode() == ISD::FCOPYSIGN)
6688 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
6689 N0.getOperand(0), N1);
6691 // copysign(x, abs(y)) -> abs(x)
6692 if (N1.getOpcode() == ISD::FABS)
6693 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
6695 // copysign(x, copysign(y,z)) -> copysign(x, z)
6696 if (N1.getOpcode() == ISD::FCOPYSIGN)
6697 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
6698 N0, N1.getOperand(1));
6700 // copysign(x, fp_extend(y)) -> copysign(x, y)
6701 // copysign(x, fp_round(y)) -> copysign(x, y)
6702 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
6703 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
6704 N0, N1.getOperand(0));
6709 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
6710 SDValue N0 = N->getOperand(0);
6711 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
6712 EVT VT = N->getValueType(0);
6713 EVT OpVT = N0.getValueType();
6715 // fold (sint_to_fp c1) -> c1fp
6717 // ...but only if the target supports immediate floating-point values
6718 (!LegalOperations ||
6719 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
6720 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
6722 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
6723 // but UINT_TO_FP is legal on this target, try to convert.
6724 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
6725 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
6726 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
6727 if (DAG.SignBitIsZero(N0))
6728 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
6731 // The next optimizations are desireable only if SELECT_CC can be lowered.
6732 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
6733 // having to say they don't support SELECT_CC on every type the DAG knows
6734 // about, since there is no way to mark an opcode illegal at all value types
6735 // (See also visitSELECT)
6736 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other)) {
6737 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
6738 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
6740 (!LegalOperations ||
6741 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
6743 { N0.getOperand(0), N0.getOperand(1),
6744 DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT),
6746 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops, 5);
6749 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
6750 // (select_cc x, y, 1.0, 0.0,, cc)
6751 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
6752 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
6753 (!LegalOperations ||
6754 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
6756 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
6757 DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT),
6758 N0.getOperand(0).getOperand(2) };
6759 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops, 5);
6766 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
6767 SDValue N0 = N->getOperand(0);
6768 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
6769 EVT VT = N->getValueType(0);
6770 EVT OpVT = N0.getValueType();
6772 // fold (uint_to_fp c1) -> c1fp
6774 // ...but only if the target supports immediate floating-point values
6775 (!LegalOperations ||
6776 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
6777 return DAG.getNode(ISD::UINT_TO_FP, SDLoc(N), VT, N0);
6779 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
6780 // but SINT_TO_FP is legal on this target, try to convert.
6781 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
6782 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
6783 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
6784 if (DAG.SignBitIsZero(N0))
6785 return DAG.getNode(ISD::SINT_TO_FP, SDLoc(N), VT, N0);
6788 // The next optimizations are desireable only if SELECT_CC can be lowered.
6789 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
6790 // having to say they don't support SELECT_CC on every type the DAG knows
6791 // about, since there is no way to mark an opcode illegal at all value types
6792 // (See also visitSELECT)
6793 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other)) {
6794 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
6796 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
6797 (!LegalOperations ||
6798 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
6800 { N0.getOperand(0), N0.getOperand(1),
6801 DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT),
6803 return DAG.getNode(ISD::SELECT_CC, SDLoc(N), VT, Ops, 5);
6810 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
6811 SDValue N0 = N->getOperand(0);
6812 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6813 EVT VT = N->getValueType(0);
6815 // fold (fp_to_sint c1fp) -> c1
6817 return DAG.getNode(ISD::FP_TO_SINT, SDLoc(N), VT, N0);
6822 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
6823 SDValue N0 = N->getOperand(0);
6824 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6825 EVT VT = N->getValueType(0);
6827 // fold (fp_to_uint c1fp) -> c1
6829 return DAG.getNode(ISD::FP_TO_UINT, SDLoc(N), VT, N0);
6834 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
6835 SDValue N0 = N->getOperand(0);
6836 SDValue N1 = N->getOperand(1);
6837 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6838 EVT VT = N->getValueType(0);
6840 // fold (fp_round c1fp) -> c1fp
6842 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0, N1);
6844 // fold (fp_round (fp_extend x)) -> x
6845 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
6846 return N0.getOperand(0);
6848 // fold (fp_round (fp_round x)) -> (fp_round x)
6849 if (N0.getOpcode() == ISD::FP_ROUND) {
6850 // This is a value preserving truncation if both round's are.
6851 bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
6852 N0.getNode()->getConstantOperandVal(1) == 1;
6853 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT, N0.getOperand(0),
6854 DAG.getIntPtrConstant(IsTrunc));
6857 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
6858 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
6859 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, SDLoc(N0), VT,
6860 N0.getOperand(0), N1);
6861 AddToWorkList(Tmp.getNode());
6862 return DAG.getNode(ISD::FCOPYSIGN, SDLoc(N), VT,
6863 Tmp, N0.getOperand(1));
6869 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
6870 SDValue N0 = N->getOperand(0);
6871 EVT VT = N->getValueType(0);
6872 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
6873 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6875 // fold (fp_round_inreg c1fp) -> c1fp
6876 if (N0CFP && isTypeLegal(EVT)) {
6877 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
6878 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, Round);
6884 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
6885 SDValue N0 = N->getOperand(0);
6886 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6887 EVT VT = N->getValueType(0);
6889 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
6890 if (N->hasOneUse() &&
6891 N->use_begin()->getOpcode() == ISD::FP_ROUND)
6894 // fold (fp_extend c1fp) -> c1fp
6896 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, N0);
6898 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
6900 if (N0.getOpcode() == ISD::FP_ROUND
6901 && N0.getNode()->getConstantOperandVal(1) == 1) {
6902 SDValue In = N0.getOperand(0);
6903 if (In.getValueType() == VT) return In;
6904 if (VT.bitsLT(In.getValueType()))
6905 return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT,
6906 In, N0.getOperand(1));
6907 return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In);
6910 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
6911 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6912 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6913 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
6914 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6915 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
6917 LN0->getBasePtr(), N0.getValueType(),
6918 LN0->getMemOperand());
6919 CombineTo(N, ExtLoad);
6920 CombineTo(N0.getNode(),
6921 DAG.getNode(ISD::FP_ROUND, SDLoc(N0),
6922 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
6923 ExtLoad.getValue(1));
6924 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6930 SDValue DAGCombiner::visitFNEG(SDNode *N) {
6931 SDValue N0 = N->getOperand(0);
6932 EVT VT = N->getValueType(0);
6934 if (VT.isVector()) {
6935 SDValue FoldedVOp = SimplifyVUnaryOp(N);
6936 if (FoldedVOp.getNode()) return FoldedVOp;
6939 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
6940 &DAG.getTarget().Options))
6941 return GetNegatedExpression(N0, DAG, LegalOperations);
6943 // Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading
6944 // constant pool values.
6945 if (!TLI.isFNegFree(VT) && N0.getOpcode() == ISD::BITCAST &&
6947 N0.getNode()->hasOneUse() &&
6948 N0.getOperand(0).getValueType().isInteger()) {
6949 SDValue Int = N0.getOperand(0);
6950 EVT IntVT = Int.getValueType();
6951 if (IntVT.isInteger() && !IntVT.isVector()) {
6952 Int = DAG.getNode(ISD::XOR, SDLoc(N0), IntVT, Int,
6953 DAG.getConstant(APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
6954 AddToWorkList(Int.getNode());
6955 return DAG.getNode(ISD::BITCAST, SDLoc(N),
6960 // (fneg (fmul c, x)) -> (fmul -c, x)
6961 if (N0.getOpcode() == ISD::FMUL) {
6962 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
6964 return DAG.getNode(ISD::FMUL, SDLoc(N), VT,
6966 DAG.getNode(ISD::FNEG, SDLoc(N), VT,
6973 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
6974 SDValue N0 = N->getOperand(0);
6975 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6976 EVT VT = N->getValueType(0);
6978 // fold (fceil c1) -> fceil(c1)
6980 return DAG.getNode(ISD::FCEIL, SDLoc(N), VT, N0);
6985 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
6986 SDValue N0 = N->getOperand(0);
6987 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6988 EVT VT = N->getValueType(0);
6990 // fold (ftrunc c1) -> ftrunc(c1)
6992 return DAG.getNode(ISD::FTRUNC, SDLoc(N), VT, N0);
6997 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
6998 SDValue N0 = N->getOperand(0);
6999 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7000 EVT VT = N->getValueType(0);
7002 // fold (ffloor c1) -> ffloor(c1)
7004 return DAG.getNode(ISD::FFLOOR, SDLoc(N), VT, N0);
7009 SDValue DAGCombiner::visitFABS(SDNode *N) {
7010 SDValue N0 = N->getOperand(0);
7011 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
7012 EVT VT = N->getValueType(0);
7014 if (VT.isVector()) {
7015 SDValue FoldedVOp = SimplifyVUnaryOp(N);
7016 if (FoldedVOp.getNode()) return FoldedVOp;
7019 // fold (fabs c1) -> fabs(c1)
7021 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0);
7022 // fold (fabs (fabs x)) -> (fabs x)
7023 if (N0.getOpcode() == ISD::FABS)
7024 return N->getOperand(0);
7025 // fold (fabs (fneg x)) -> (fabs x)
7026 // fold (fabs (fcopysign x, y)) -> (fabs x)
7027 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
7028 return DAG.getNode(ISD::FABS, SDLoc(N), VT, N0.getOperand(0));
7030 // Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading
7031 // constant pool values.
7032 if (!TLI.isFAbsFree(VT) &&
7033 N0.getOpcode() == ISD::BITCAST && N0.getNode()->hasOneUse() &&
7034 N0.getOperand(0).getValueType().isInteger() &&
7035 !N0.getOperand(0).getValueType().isVector()) {
7036 SDValue Int = N0.getOperand(0);
7037 EVT IntVT = Int.getValueType();
7038 if (IntVT.isInteger() && !IntVT.isVector()) {
7039 Int = DAG.getNode(ISD::AND, SDLoc(N0), IntVT, Int,
7040 DAG.getConstant(~APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
7041 AddToWorkList(Int.getNode());
7042 return DAG.getNode(ISD::BITCAST, SDLoc(N),
7043 N->getValueType(0), Int);
7050 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
7051 SDValue Chain = N->getOperand(0);
7052 SDValue N1 = N->getOperand(1);
7053 SDValue N2 = N->getOperand(2);
7055 // If N is a constant we could fold this into a fallthrough or unconditional
7056 // branch. However that doesn't happen very often in normal code, because
7057 // Instcombine/SimplifyCFG should have handled the available opportunities.
7058 // If we did this folding here, it would be necessary to update the
7059 // MachineBasicBlock CFG, which is awkward.
7061 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
7063 if (N1.getOpcode() == ISD::SETCC &&
7064 TLI.isOperationLegalOrCustom(ISD::BR_CC,
7065 N1.getOperand(0).getValueType())) {
7066 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
7067 Chain, N1.getOperand(2),
7068 N1.getOperand(0), N1.getOperand(1), N2);
7071 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
7072 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
7073 (N1.getOperand(0).hasOneUse() &&
7074 N1.getOperand(0).getOpcode() == ISD::SRL))) {
7076 if (N1.getOpcode() == ISD::TRUNCATE) {
7077 // Look pass the truncate.
7078 Trunc = N1.getNode();
7079 N1 = N1.getOperand(0);
7082 // Match this pattern so that we can generate simpler code:
7085 // %b = and i32 %a, 2
7086 // %c = srl i32 %b, 1
7087 // brcond i32 %c ...
7092 // %b = and i32 %a, 2
7093 // %c = setcc eq %b, 0
7096 // This applies only when the AND constant value has one bit set and the
7097 // SRL constant is equal to the log2 of the AND constant. The back-end is
7098 // smart enough to convert the result into a TEST/JMP sequence.
7099 SDValue Op0 = N1.getOperand(0);
7100 SDValue Op1 = N1.getOperand(1);
7102 if (Op0.getOpcode() == ISD::AND &&
7103 Op1.getOpcode() == ISD::Constant) {
7104 SDValue AndOp1 = Op0.getOperand(1);
7106 if (AndOp1.getOpcode() == ISD::Constant) {
7107 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
7109 if (AndConst.isPowerOf2() &&
7110 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
7112 DAG.getSetCC(SDLoc(N),
7113 getSetCCResultType(Op0.getValueType()),
7114 Op0, DAG.getConstant(0, Op0.getValueType()),
7117 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, SDLoc(N),
7118 MVT::Other, Chain, SetCC, N2);
7119 // Don't add the new BRCond into the worklist or else SimplifySelectCC
7120 // will convert it back to (X & C1) >> C2.
7121 CombineTo(N, NewBRCond, false);
7122 // Truncate is dead.
7124 removeFromWorkList(Trunc);
7125 DAG.DeleteNode(Trunc);
7127 // Replace the uses of SRL with SETCC
7128 WorkListRemover DeadNodes(*this);
7129 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
7130 removeFromWorkList(N1.getNode());
7131 DAG.DeleteNode(N1.getNode());
7132 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7138 // Restore N1 if the above transformation doesn't match.
7139 N1 = N->getOperand(1);
7142 // Transform br(xor(x, y)) -> br(x != y)
7143 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
7144 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
7145 SDNode *TheXor = N1.getNode();
7146 SDValue Op0 = TheXor->getOperand(0);
7147 SDValue Op1 = TheXor->getOperand(1);
7148 if (Op0.getOpcode() == Op1.getOpcode()) {
7149 // Avoid missing important xor optimizations.
7150 SDValue Tmp = visitXOR(TheXor);
7151 if (Tmp.getNode()) {
7152 if (Tmp.getNode() != TheXor) {
7153 DEBUG(dbgs() << "\nReplacing.8 ";
7155 dbgs() << "\nWith: ";
7156 Tmp.getNode()->dump(&DAG);
7158 WorkListRemover DeadNodes(*this);
7159 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
7160 removeFromWorkList(TheXor);
7161 DAG.DeleteNode(TheXor);
7162 return DAG.getNode(ISD::BRCOND, SDLoc(N),
7163 MVT::Other, Chain, Tmp, N2);
7166 // visitXOR has changed XOR's operands or replaced the XOR completely,
7168 return SDValue(N, 0);
7172 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
7174 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
7175 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
7176 Op0.getOpcode() == ISD::XOR) {
7177 TheXor = Op0.getNode();
7181 EVT SetCCVT = N1.getValueType();
7183 SetCCVT = getSetCCResultType(SetCCVT);
7184 SDValue SetCC = DAG.getSetCC(SDLoc(TheXor),
7187 Equal ? ISD::SETEQ : ISD::SETNE);
7188 // Replace the uses of XOR with SETCC
7189 WorkListRemover DeadNodes(*this);
7190 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
7191 removeFromWorkList(N1.getNode());
7192 DAG.DeleteNode(N1.getNode());
7193 return DAG.getNode(ISD::BRCOND, SDLoc(N),
7194 MVT::Other, Chain, SetCC, N2);
7201 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
7203 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
7204 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
7205 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
7207 // If N is a constant we could fold this into a fallthrough or unconditional
7208 // branch. However that doesn't happen very often in normal code, because
7209 // Instcombine/SimplifyCFG should have handled the available opportunities.
7210 // If we did this folding here, it would be necessary to update the
7211 // MachineBasicBlock CFG, which is awkward.
7213 // Use SimplifySetCC to simplify SETCC's.
7214 SDValue Simp = SimplifySetCC(getSetCCResultType(CondLHS.getValueType()),
7215 CondLHS, CondRHS, CC->get(), SDLoc(N),
7217 if (Simp.getNode()) AddToWorkList(Simp.getNode());
7219 // fold to a simpler setcc
7220 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
7221 return DAG.getNode(ISD::BR_CC, SDLoc(N), MVT::Other,
7222 N->getOperand(0), Simp.getOperand(2),
7223 Simp.getOperand(0), Simp.getOperand(1),
7229 /// canFoldInAddressingMode - Return true if 'Use' is a load or a store that
7230 /// uses N as its base pointer and that N may be folded in the load / store
7231 /// addressing mode.
7232 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
7234 const TargetLowering &TLI) {
7236 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
7237 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
7239 VT = Use->getValueType(0);
7240 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
7241 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
7243 VT = ST->getValue().getValueType();
7247 TargetLowering::AddrMode AM;
7248 if (N->getOpcode() == ISD::ADD) {
7249 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
7252 AM.BaseOffs = Offset->getSExtValue();
7256 } else if (N->getOpcode() == ISD::SUB) {
7257 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
7260 AM.BaseOffs = -Offset->getSExtValue();
7267 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
7270 /// CombineToPreIndexedLoadStore - Try turning a load / store into a
7271 /// pre-indexed load / store when the base pointer is an add or subtract
7272 /// and it has other uses besides the load / store. After the
7273 /// transformation, the new indexed load / store has effectively folded
7274 /// the add / subtract in and all of its other uses are redirected to the
7275 /// new load / store.
7276 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
7277 if (Level < AfterLegalizeDAG)
7283 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7284 if (LD->isIndexed())
7286 VT = LD->getMemoryVT();
7287 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
7288 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
7290 Ptr = LD->getBasePtr();
7291 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7292 if (ST->isIndexed())
7294 VT = ST->getMemoryVT();
7295 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
7296 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
7298 Ptr = ST->getBasePtr();
7304 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
7305 // out. There is no reason to make this a preinc/predec.
7306 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
7307 Ptr.getNode()->hasOneUse())
7310 // Ask the target to do addressing mode selection.
7313 ISD::MemIndexedMode AM = ISD::UNINDEXED;
7314 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
7317 // Backends without true r+i pre-indexed forms may need to pass a
7318 // constant base with a variable offset so that constant coercion
7319 // will work with the patterns in canonical form.
7320 bool Swapped = false;
7321 if (isa<ConstantSDNode>(BasePtr)) {
7322 std::swap(BasePtr, Offset);
7326 // Don't create a indexed load / store with zero offset.
7327 if (isa<ConstantSDNode>(Offset) &&
7328 cast<ConstantSDNode>(Offset)->isNullValue())
7331 // Try turning it into a pre-indexed load / store except when:
7332 // 1) The new base ptr is a frame index.
7333 // 2) If N is a store and the new base ptr is either the same as or is a
7334 // predecessor of the value being stored.
7335 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
7336 // that would create a cycle.
7337 // 4) All uses are load / store ops that use it as old base ptr.
7339 // Check #1. Preinc'ing a frame index would require copying the stack pointer
7340 // (plus the implicit offset) to a register to preinc anyway.
7341 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
7346 SDValue Val = cast<StoreSDNode>(N)->getValue();
7347 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
7351 // If the offset is a constant, there may be other adds of constants that
7352 // can be folded with this one. We should do this to avoid having to keep
7353 // a copy of the original base pointer.
7354 SmallVector<SDNode *, 16> OtherUses;
7355 if (isa<ConstantSDNode>(Offset))
7356 for (SDNode::use_iterator I = BasePtr.getNode()->use_begin(),
7357 E = BasePtr.getNode()->use_end(); I != E; ++I) {
7359 if (Use == Ptr.getNode())
7362 if (Use->isPredecessorOf(N))
7365 if (Use->getOpcode() != ISD::ADD && Use->getOpcode() != ISD::SUB) {
7370 SDValue Op0 = Use->getOperand(0), Op1 = Use->getOperand(1);
7371 if (Op1.getNode() == BasePtr.getNode())
7372 std::swap(Op0, Op1);
7373 assert(Op0.getNode() == BasePtr.getNode() &&
7374 "Use of ADD/SUB but not an operand");
7376 if (!isa<ConstantSDNode>(Op1)) {
7381 // FIXME: In some cases, we can be smarter about this.
7382 if (Op1.getValueType() != Offset.getValueType()) {
7387 OtherUses.push_back(Use);
7391 std::swap(BasePtr, Offset);
7393 // Now check for #3 and #4.
7394 bool RealUse = false;
7396 // Caches for hasPredecessorHelper
7397 SmallPtrSet<const SDNode *, 32> Visited;
7398 SmallVector<const SDNode *, 16> Worklist;
7400 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
7401 E = Ptr.getNode()->use_end(); I != E; ++I) {
7405 if (N->hasPredecessorHelper(Use, Visited, Worklist))
7408 // If Ptr may be folded in addressing mode of other use, then it's
7409 // not profitable to do this transformation.
7410 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
7419 Result = DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
7420 BasePtr, Offset, AM);
7422 Result = DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
7423 BasePtr, Offset, AM);
7426 DEBUG(dbgs() << "\nReplacing.4 ";
7428 dbgs() << "\nWith: ";
7429 Result.getNode()->dump(&DAG);
7431 WorkListRemover DeadNodes(*this);
7433 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
7434 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
7436 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
7439 // Finally, since the node is now dead, remove it from the graph.
7443 std::swap(BasePtr, Offset);
7445 // Replace other uses of BasePtr that can be updated to use Ptr
7446 for (unsigned i = 0, e = OtherUses.size(); i != e; ++i) {
7447 unsigned OffsetIdx = 1;
7448 if (OtherUses[i]->getOperand(OffsetIdx).getNode() == BasePtr.getNode())
7450 assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() ==
7451 BasePtr.getNode() && "Expected BasePtr operand");
7453 // We need to replace ptr0 in the following expression:
7454 // x0 * offset0 + y0 * ptr0 = t0
7456 // x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store)
7458 // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the
7459 // indexed load/store and the expresion that needs to be re-written.
7461 // Therefore, we have:
7462 // t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1
7464 ConstantSDNode *CN =
7465 cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx));
7467 APInt Offset0 = CN->getAPIntValue();
7468 APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue();
7470 X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1;
7471 Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1;
7472 X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1;
7473 Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1;
7475 unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD;
7477 APInt CNV = Offset0;
7478 if (X0 < 0) CNV = -CNV;
7479 if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1;
7480 else CNV = CNV - Offset1;
7482 // We can now generate the new expression.
7483 SDValue NewOp1 = DAG.getConstant(CNV, CN->getValueType(0));
7484 SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0);
7486 SDValue NewUse = DAG.getNode(Opcode,
7487 SDLoc(OtherUses[i]),
7488 OtherUses[i]->getValueType(0), NewOp1, NewOp2);
7489 DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse);
7490 removeFromWorkList(OtherUses[i]);
7491 DAG.DeleteNode(OtherUses[i]);
7494 // Replace the uses of Ptr with uses of the updated base value.
7495 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
7496 removeFromWorkList(Ptr.getNode());
7497 DAG.DeleteNode(Ptr.getNode());
7502 /// CombineToPostIndexedLoadStore - Try to combine a load / store with a
7503 /// add / sub of the base pointer node into a post-indexed load / store.
7504 /// The transformation folded the add / subtract into the new indexed
7505 /// load / store effectively and all of its uses are redirected to the
7506 /// new load / store.
7507 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
7508 if (Level < AfterLegalizeDAG)
7514 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
7515 if (LD->isIndexed())
7517 VT = LD->getMemoryVT();
7518 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
7519 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
7521 Ptr = LD->getBasePtr();
7522 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
7523 if (ST->isIndexed())
7525 VT = ST->getMemoryVT();
7526 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
7527 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
7529 Ptr = ST->getBasePtr();
7535 if (Ptr.getNode()->hasOneUse())
7538 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
7539 E = Ptr.getNode()->use_end(); I != E; ++I) {
7542 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
7547 ISD::MemIndexedMode AM = ISD::UNINDEXED;
7548 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
7549 // Don't create a indexed load / store with zero offset.
7550 if (isa<ConstantSDNode>(Offset) &&
7551 cast<ConstantSDNode>(Offset)->isNullValue())
7554 // Try turning it into a post-indexed load / store except when
7555 // 1) All uses are load / store ops that use it as base ptr (and
7556 // it may be folded as addressing mmode).
7557 // 2) Op must be independent of N, i.e. Op is neither a predecessor
7558 // nor a successor of N. Otherwise, if Op is folded that would
7561 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
7565 bool TryNext = false;
7566 for (SDNode::use_iterator II = BasePtr.getNode()->use_begin(),
7567 EE = BasePtr.getNode()->use_end(); II != EE; ++II) {
7569 if (Use == Ptr.getNode())
7572 // If all the uses are load / store addresses, then don't do the
7574 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
7575 bool RealUse = false;
7576 for (SDNode::use_iterator III = Use->use_begin(),
7577 EEE = Use->use_end(); III != EEE; ++III) {
7578 SDNode *UseUse = *III;
7579 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
7594 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
7595 SDValue Result = isLoad
7596 ? DAG.getIndexedLoad(SDValue(N,0), SDLoc(N),
7597 BasePtr, Offset, AM)
7598 : DAG.getIndexedStore(SDValue(N,0), SDLoc(N),
7599 BasePtr, Offset, AM);
7602 DEBUG(dbgs() << "\nReplacing.5 ";
7604 dbgs() << "\nWith: ";
7605 Result.getNode()->dump(&DAG);
7607 WorkListRemover DeadNodes(*this);
7609 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
7610 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
7612 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
7615 // Finally, since the node is now dead, remove it from the graph.
7618 // Replace the uses of Use with uses of the updated base value.
7619 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
7620 Result.getValue(isLoad ? 1 : 0));
7621 removeFromWorkList(Op);
7631 SDValue DAGCombiner::visitLOAD(SDNode *N) {
7632 LoadSDNode *LD = cast<LoadSDNode>(N);
7633 SDValue Chain = LD->getChain();
7634 SDValue Ptr = LD->getBasePtr();
7636 // If load is not volatile and there are no uses of the loaded value (and
7637 // the updated indexed value in case of indexed loads), change uses of the
7638 // chain value into uses of the chain input (i.e. delete the dead load).
7639 if (!LD->isVolatile()) {
7640 if (N->getValueType(1) == MVT::Other) {
7642 if (!N->hasAnyUseOfValue(0)) {
7643 // It's not safe to use the two value CombineTo variant here. e.g.
7644 // v1, chain2 = load chain1, loc
7645 // v2, chain3 = load chain2, loc
7647 // Now we replace use of chain2 with chain1. This makes the second load
7648 // isomorphic to the one we are deleting, and thus makes this load live.
7649 DEBUG(dbgs() << "\nReplacing.6 ";
7651 dbgs() << "\nWith chain: ";
7652 Chain.getNode()->dump(&DAG);
7654 WorkListRemover DeadNodes(*this);
7655 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
7657 if (N->use_empty()) {
7658 removeFromWorkList(N);
7662 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7666 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
7667 if (!N->hasAnyUseOfValue(0) && !N->hasAnyUseOfValue(1)) {
7668 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
7669 DEBUG(dbgs() << "\nReplacing.7 ";
7671 dbgs() << "\nWith: ";
7672 Undef.getNode()->dump(&DAG);
7673 dbgs() << " and 2 other values\n");
7674 WorkListRemover DeadNodes(*this);
7675 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
7676 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1),
7677 DAG.getUNDEF(N->getValueType(1)));
7678 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
7679 removeFromWorkList(N);
7681 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7686 // If this load is directly stored, replace the load value with the stored
7688 // TODO: Handle store large -> read small portion.
7689 // TODO: Handle TRUNCSTORE/LOADEXT
7690 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
7691 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
7692 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
7693 if (PrevST->getBasePtr() == Ptr &&
7694 PrevST->getValue().getValueType() == N->getValueType(0))
7695 return CombineTo(N, Chain.getOperand(1), Chain);
7699 // Try to infer better alignment information than the load already has.
7700 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
7701 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
7702 if (Align > LD->getMemOperand()->getBaseAlignment()) {
7704 DAG.getExtLoad(LD->getExtensionType(), SDLoc(N),
7705 LD->getValueType(0),
7706 Chain, Ptr, LD->getPointerInfo(),
7708 LD->isVolatile(), LD->isNonTemporal(), Align,
7710 return CombineTo(N, NewLoad, SDValue(NewLoad.getNode(), 1), true);
7715 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA :
7716 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
7718 // Walk up chain skipping non-aliasing memory nodes.
7719 SDValue BetterChain = FindBetterChain(N, Chain);
7721 // If there is a better chain.
7722 if (Chain != BetterChain) {
7725 // Replace the chain to void dependency.
7726 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
7727 ReplLoad = DAG.getLoad(N->getValueType(0), SDLoc(LD),
7728 BetterChain, Ptr, LD->getMemOperand());
7730 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), SDLoc(LD),
7731 LD->getValueType(0),
7732 BetterChain, Ptr, LD->getMemoryVT(),
7733 LD->getMemOperand());
7736 // Create token factor to keep old chain connected.
7737 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
7738 MVT::Other, Chain, ReplLoad.getValue(1));
7740 // Make sure the new and old chains are cleaned up.
7741 AddToWorkList(Token.getNode());
7743 // Replace uses with load result and token factor. Don't add users
7745 return CombineTo(N, ReplLoad.getValue(0), Token, false);
7749 // Try transforming N to an indexed load.
7750 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
7751 return SDValue(N, 0);
7753 // Try to slice up N to more direct loads if the slices are mapped to
7754 // different register banks or pairing can take place.
7756 return SDValue(N, 0);
7762 /// \brief Helper structure used to slice a load in smaller loads.
7763 /// Basically a slice is obtained from the following sequence:
7764 /// Origin = load Ty1, Base
7765 /// Shift = srl Ty1 Origin, CstTy Amount
7766 /// Inst = trunc Shift to Ty2
7768 /// Then, it will be rewriten into:
7769 /// Slice = load SliceTy, Base + SliceOffset
7770 /// [Inst = zext Slice to Ty2], only if SliceTy <> Ty2
7772 /// SliceTy is deduced from the number of bits that are actually used to
7774 struct LoadedSlice {
7775 /// \brief Helper structure used to compute the cost of a slice.
7777 /// Are we optimizing for code size.
7782 unsigned CrossRegisterBanksCopies;
7786 Cost(bool ForCodeSize = false)
7787 : ForCodeSize(ForCodeSize), Loads(0), Truncates(0),
7788 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {}
7790 /// \brief Get the cost of one isolated slice.
7791 Cost(const LoadedSlice &LS, bool ForCodeSize = false)
7792 : ForCodeSize(ForCodeSize), Loads(1), Truncates(0),
7793 CrossRegisterBanksCopies(0), ZExts(0), Shift(0) {
7794 EVT TruncType = LS.Inst->getValueType(0);
7795 EVT LoadedType = LS.getLoadedType();
7796 if (TruncType != LoadedType &&
7797 !LS.DAG->getTargetLoweringInfo().isZExtFree(LoadedType, TruncType))
7801 /// \brief Account for slicing gain in the current cost.
7802 /// Slicing provide a few gains like removing a shift or a
7803 /// truncate. This method allows to grow the cost of the original
7804 /// load with the gain from this slice.
7805 void addSliceGain(const LoadedSlice &LS) {
7806 // Each slice saves a truncate.
7807 const TargetLowering &TLI = LS.DAG->getTargetLoweringInfo();
7808 if (!TLI.isTruncateFree(LS.Inst->getValueType(0),
7809 LS.Inst->getOperand(0).getValueType()))
7811 // If there is a shift amount, this slice gets rid of it.
7814 // If this slice can merge a cross register bank copy, account for it.
7815 if (LS.canMergeExpensiveCrossRegisterBankCopy())
7816 ++CrossRegisterBanksCopies;
7819 Cost &operator+=(const Cost &RHS) {
7821 Truncates += RHS.Truncates;
7822 CrossRegisterBanksCopies += RHS.CrossRegisterBanksCopies;
7828 bool operator==(const Cost &RHS) const {
7829 return Loads == RHS.Loads && Truncates == RHS.Truncates &&
7830 CrossRegisterBanksCopies == RHS.CrossRegisterBanksCopies &&
7831 ZExts == RHS.ZExts && Shift == RHS.Shift;
7834 bool operator!=(const Cost &RHS) const { return !(*this == RHS); }
7836 bool operator<(const Cost &RHS) const {
7837 // Assume cross register banks copies are as expensive as loads.
7838 // FIXME: Do we want some more target hooks?
7839 unsigned ExpensiveOpsLHS = Loads + CrossRegisterBanksCopies;
7840 unsigned ExpensiveOpsRHS = RHS.Loads + RHS.CrossRegisterBanksCopies;
7841 // Unless we are optimizing for code size, consider the
7842 // expensive operation first.
7843 if (!ForCodeSize && ExpensiveOpsLHS != ExpensiveOpsRHS)
7844 return ExpensiveOpsLHS < ExpensiveOpsRHS;
7845 return (Truncates + ZExts + Shift + ExpensiveOpsLHS) <
7846 (RHS.Truncates + RHS.ZExts + RHS.Shift + ExpensiveOpsRHS);
7849 bool operator>(const Cost &RHS) const { return RHS < *this; }
7851 bool operator<=(const Cost &RHS) const { return !(RHS < *this); }
7853 bool operator>=(const Cost &RHS) const { return !(*this < RHS); }
7855 // The last instruction that represent the slice. This should be a
7856 // truncate instruction.
7858 // The original load instruction.
7860 // The right shift amount in bits from the original load.
7862 // The DAG from which Origin came from.
7863 // This is used to get some contextual information about legal types, etc.
7866 LoadedSlice(SDNode *Inst = NULL, LoadSDNode *Origin = NULL,
7867 unsigned Shift = 0, SelectionDAG *DAG = NULL)
7868 : Inst(Inst), Origin(Origin), Shift(Shift), DAG(DAG) {}
7870 LoadedSlice(const LoadedSlice &LS)
7871 : Inst(LS.Inst), Origin(LS.Origin), Shift(LS.Shift), DAG(LS.DAG) {}
7873 /// \brief Get the bits used in a chunk of bits \p BitWidth large.
7874 /// \return Result is \p BitWidth and has used bits set to 1 and
7875 /// not used bits set to 0.
7876 APInt getUsedBits() const {
7877 // Reproduce the trunc(lshr) sequence:
7878 // - Start from the truncated value.
7879 // - Zero extend to the desired bit width.
7881 assert(Origin && "No original load to compare against.");
7882 unsigned BitWidth = Origin->getValueSizeInBits(0);
7883 assert(Inst && "This slice is not bound to an instruction");
7884 assert(Inst->getValueSizeInBits(0) <= BitWidth &&
7885 "Extracted slice is bigger than the whole type!");
7886 APInt UsedBits(Inst->getValueSizeInBits(0), 0);
7887 UsedBits.setAllBits();
7888 UsedBits = UsedBits.zext(BitWidth);
7893 /// \brief Get the size of the slice to be loaded in bytes.
7894 unsigned getLoadedSize() const {
7895 unsigned SliceSize = getUsedBits().countPopulation();
7896 assert(!(SliceSize & 0x7) && "Size is not a multiple of a byte.");
7897 return SliceSize / 8;
7900 /// \brief Get the type that will be loaded for this slice.
7901 /// Note: This may not be the final type for the slice.
7902 EVT getLoadedType() const {
7903 assert(DAG && "Missing context");
7904 LLVMContext &Ctxt = *DAG->getContext();
7905 return EVT::getIntegerVT(Ctxt, getLoadedSize() * 8);
7908 /// \brief Get the alignment of the load used for this slice.
7909 unsigned getAlignment() const {
7910 unsigned Alignment = Origin->getAlignment();
7911 unsigned Offset = getOffsetFromBase();
7913 Alignment = MinAlign(Alignment, Alignment + Offset);
7917 /// \brief Check if this slice can be rewritten with legal operations.
7918 bool isLegal() const {
7919 // An invalid slice is not legal.
7920 if (!Origin || !Inst || !DAG)
7923 // Offsets are for indexed load only, we do not handle that.
7924 if (Origin->getOffset().getOpcode() != ISD::UNDEF)
7927 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
7929 // Check that the type is legal.
7930 EVT SliceType = getLoadedType();
7931 if (!TLI.isTypeLegal(SliceType))
7934 // Check that the load is legal for this type.
7935 if (!TLI.isOperationLegal(ISD::LOAD, SliceType))
7938 // Check that the offset can be computed.
7939 // 1. Check its type.
7940 EVT PtrType = Origin->getBasePtr().getValueType();
7941 if (PtrType == MVT::Untyped || PtrType.isExtended())
7944 // 2. Check that it fits in the immediate.
7945 if (!TLI.isLegalAddImmediate(getOffsetFromBase()))
7948 // 3. Check that the computation is legal.
7949 if (!TLI.isOperationLegal(ISD::ADD, PtrType))
7952 // Check that the zext is legal if it needs one.
7953 EVT TruncateType = Inst->getValueType(0);
7954 if (TruncateType != SliceType &&
7955 !TLI.isOperationLegal(ISD::ZERO_EXTEND, TruncateType))
7961 /// \brief Get the offset in bytes of this slice in the original chunk of
7963 /// \pre DAG != NULL.
7964 uint64_t getOffsetFromBase() const {
7965 assert(DAG && "Missing context.");
7967 DAG->getTargetLoweringInfo().getDataLayout()->isBigEndian();
7968 assert(!(Shift & 0x7) && "Shifts not aligned on Bytes are not supported.");
7969 uint64_t Offset = Shift / 8;
7970 unsigned TySizeInBytes = Origin->getValueSizeInBits(0) / 8;
7971 assert(!(Origin->getValueSizeInBits(0) & 0x7) &&
7972 "The size of the original loaded type is not a multiple of a"
7974 // If Offset is bigger than TySizeInBytes, it means we are loading all
7975 // zeros. This should have been optimized before in the process.
7976 assert(TySizeInBytes > Offset &&
7977 "Invalid shift amount for given loaded size");
7979 Offset = TySizeInBytes - Offset - getLoadedSize();
7983 /// \brief Generate the sequence of instructions to load the slice
7984 /// represented by this object and redirect the uses of this slice to
7985 /// this new sequence of instructions.
7986 /// \pre this->Inst && this->Origin are valid Instructions and this
7987 /// object passed the legal check: LoadedSlice::isLegal returned true.
7988 /// \return The last instruction of the sequence used to load the slice.
7989 SDValue loadSlice() const {
7990 assert(Inst && Origin && "Unable to replace a non-existing slice.");
7991 const SDValue &OldBaseAddr = Origin->getBasePtr();
7992 SDValue BaseAddr = OldBaseAddr;
7993 // Get the offset in that chunk of bytes w.r.t. the endianess.
7994 int64_t Offset = static_cast<int64_t>(getOffsetFromBase());
7995 assert(Offset >= 0 && "Offset too big to fit in int64_t!");
7997 // BaseAddr = BaseAddr + Offset.
7998 EVT ArithType = BaseAddr.getValueType();
7999 BaseAddr = DAG->getNode(ISD::ADD, SDLoc(Origin), ArithType, BaseAddr,
8000 DAG->getConstant(Offset, ArithType));
8003 // Create the type of the loaded slice according to its size.
8004 EVT SliceType = getLoadedType();
8006 // Create the load for the slice.
8007 SDValue LastInst = DAG->getLoad(
8008 SliceType, SDLoc(Origin), Origin->getChain(), BaseAddr,
8009 Origin->getPointerInfo().getWithOffset(Offset), Origin->isVolatile(),
8010 Origin->isNonTemporal(), Origin->isInvariant(), getAlignment());
8011 // If the final type is not the same as the loaded type, this means that
8012 // we have to pad with zero. Create a zero extend for that.
8013 EVT FinalType = Inst->getValueType(0);
8014 if (SliceType != FinalType)
8016 DAG->getNode(ISD::ZERO_EXTEND, SDLoc(LastInst), FinalType, LastInst);
8020 /// \brief Check if this slice can be merged with an expensive cross register
8021 /// bank copy. E.g.,
8023 /// f = bitcast i32 i to float
8024 bool canMergeExpensiveCrossRegisterBankCopy() const {
8025 if (!Inst || !Inst->hasOneUse())
8027 SDNode *Use = *Inst->use_begin();
8028 if (Use->getOpcode() != ISD::BITCAST)
8030 assert(DAG && "Missing context");
8031 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
8032 EVT ResVT = Use->getValueType(0);
8033 const TargetRegisterClass *ResRC = TLI.getRegClassFor(ResVT.getSimpleVT());
8034 const TargetRegisterClass *ArgRC =
8035 TLI.getRegClassFor(Use->getOperand(0).getValueType().getSimpleVT());
8036 if (ArgRC == ResRC || !TLI.isOperationLegal(ISD::LOAD, ResVT))
8039 // At this point, we know that we perform a cross-register-bank copy.
8040 // Check if it is expensive.
8041 const TargetRegisterInfo *TRI = TLI.getTargetMachine().getRegisterInfo();
8042 // Assume bitcasts are cheap, unless both register classes do not
8043 // explicitly share a common sub class.
8044 if (!TRI || TRI->getCommonSubClass(ArgRC, ResRC))
8047 // Check if it will be merged with the load.
8048 // 1. Check the alignment constraint.
8049 unsigned RequiredAlignment = TLI.getDataLayout()->getABITypeAlignment(
8050 ResVT.getTypeForEVT(*DAG->getContext()));
8052 if (RequiredAlignment > getAlignment())
8055 // 2. Check that the load is a legal operation for that type.
8056 if (!TLI.isOperationLegal(ISD::LOAD, ResVT))
8059 // 3. Check that we do not have a zext in the way.
8060 if (Inst->getValueType(0) != getLoadedType())
8068 /// \brief Sorts LoadedSlice according to their offset.
8069 struct LoadedSliceSorter {
8070 bool operator()(const LoadedSlice &LHS, const LoadedSlice &RHS) {
8071 assert(LHS.Origin == RHS.Origin && "Different bases not implemented.");
8072 return LHS.getOffsetFromBase() < RHS.getOffsetFromBase();
8076 /// \brief Check that all bits set in \p UsedBits form a dense region, i.e.,
8077 /// \p UsedBits looks like 0..0 1..1 0..0.
8078 static bool areUsedBitsDense(const APInt &UsedBits) {
8079 // If all the bits are one, this is dense!
8080 if (UsedBits.isAllOnesValue())
8083 // Get rid of the unused bits on the right.
8084 APInt NarrowedUsedBits = UsedBits.lshr(UsedBits.countTrailingZeros());
8085 // Get rid of the unused bits on the left.
8086 if (NarrowedUsedBits.countLeadingZeros())
8087 NarrowedUsedBits = NarrowedUsedBits.trunc(NarrowedUsedBits.getActiveBits());
8088 // Check that the chunk of bits is completely used.
8089 return NarrowedUsedBits.isAllOnesValue();
8092 /// \brief Check whether or not \p First and \p Second are next to each other
8093 /// in memory. This means that there is no hole between the bits loaded
8094 /// by \p First and the bits loaded by \p Second.
8095 static bool areSlicesNextToEachOther(const LoadedSlice &First,
8096 const LoadedSlice &Second) {
8097 assert(First.Origin == Second.Origin && First.Origin &&
8098 "Unable to match different memory origins.");
8099 APInt UsedBits = First.getUsedBits();
8100 assert((UsedBits & Second.getUsedBits()) == 0 &&
8101 "Slices are not supposed to overlap.");
8102 UsedBits |= Second.getUsedBits();
8103 return areUsedBitsDense(UsedBits);
8106 /// \brief Adjust the \p GlobalLSCost according to the target
8107 /// paring capabilities and the layout of the slices.
8108 /// \pre \p GlobalLSCost should account for at least as many loads as
8109 /// there is in the slices in \p LoadedSlices.
8110 static void adjustCostForPairing(SmallVectorImpl<LoadedSlice> &LoadedSlices,
8111 LoadedSlice::Cost &GlobalLSCost) {
8112 unsigned NumberOfSlices = LoadedSlices.size();
8113 // If there is less than 2 elements, no pairing is possible.
8114 if (NumberOfSlices < 2)
8117 // Sort the slices so that elements that are likely to be next to each
8118 // other in memory are next to each other in the list.
8119 std::sort(LoadedSlices.begin(), LoadedSlices.end(), LoadedSliceSorter());
8120 const TargetLowering &TLI = LoadedSlices[0].DAG->getTargetLoweringInfo();
8121 // First (resp. Second) is the first (resp. Second) potentially candidate
8122 // to be placed in a paired load.
8123 const LoadedSlice *First = NULL;
8124 const LoadedSlice *Second = NULL;
8125 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice,
8126 // Set the beginning of the pair.
8129 Second = &LoadedSlices[CurrSlice];
8131 // If First is NULL, it means we start a new pair.
8132 // Get to the next slice.
8136 EVT LoadedType = First->getLoadedType();
8138 // If the types of the slices are different, we cannot pair them.
8139 if (LoadedType != Second->getLoadedType())
8142 // Check if the target supplies paired loads for this type.
8143 unsigned RequiredAlignment = 0;
8144 if (!TLI.hasPairedLoad(LoadedType, RequiredAlignment)) {
8145 // move to the next pair, this type is hopeless.
8149 // Check if we meet the alignment requirement.
8150 if (RequiredAlignment > First->getAlignment())
8153 // Check that both loads are next to each other in memory.
8154 if (!areSlicesNextToEachOther(*First, *Second))
8157 assert(GlobalLSCost.Loads > 0 && "We save more loads than we created!");
8158 --GlobalLSCost.Loads;
8159 // Move to the next pair.
8164 /// \brief Check the profitability of all involved LoadedSlice.
8165 /// Currently, it is considered profitable if there is exactly two
8166 /// involved slices (1) which are (2) next to each other in memory, and
8167 /// whose cost (\see LoadedSlice::Cost) is smaller than the original load (3).
8169 /// Note: The order of the elements in \p LoadedSlices may be modified, but not
8170 /// the elements themselves.
8172 /// FIXME: When the cost model will be mature enough, we can relax
8173 /// constraints (1) and (2).
8174 static bool isSlicingProfitable(SmallVectorImpl<LoadedSlice> &LoadedSlices,
8175 const APInt &UsedBits, bool ForCodeSize) {
8176 unsigned NumberOfSlices = LoadedSlices.size();
8177 if (StressLoadSlicing)
8178 return NumberOfSlices > 1;
8181 if (NumberOfSlices != 2)
8185 if (!areUsedBitsDense(UsedBits))
8189 LoadedSlice::Cost OrigCost(ForCodeSize), GlobalSlicingCost(ForCodeSize);
8190 // The original code has one big load.
8192 for (unsigned CurrSlice = 0; CurrSlice < NumberOfSlices; ++CurrSlice) {
8193 const LoadedSlice &LS = LoadedSlices[CurrSlice];
8194 // Accumulate the cost of all the slices.
8195 LoadedSlice::Cost SliceCost(LS, ForCodeSize);
8196 GlobalSlicingCost += SliceCost;
8198 // Account as cost in the original configuration the gain obtained
8199 // with the current slices.
8200 OrigCost.addSliceGain(LS);
8203 // If the target supports paired load, adjust the cost accordingly.
8204 adjustCostForPairing(LoadedSlices, GlobalSlicingCost);
8205 return OrigCost > GlobalSlicingCost;
8208 /// \brief If the given load, \p LI, is used only by trunc or trunc(lshr)
8209 /// operations, split it in the various pieces being extracted.
8211 /// This sort of thing is introduced by SROA.
8212 /// This slicing takes care not to insert overlapping loads.
8213 /// \pre LI is a simple load (i.e., not an atomic or volatile load).
8214 bool DAGCombiner::SliceUpLoad(SDNode *N) {
8215 if (Level < AfterLegalizeDAG)
8218 LoadSDNode *LD = cast<LoadSDNode>(N);
8219 if (LD->isVolatile() || !ISD::isNormalLoad(LD) ||
8220 !LD->getValueType(0).isInteger())
8223 // Keep track of already used bits to detect overlapping values.
8224 // In that case, we will just abort the transformation.
8225 APInt UsedBits(LD->getValueSizeInBits(0), 0);
8227 SmallVector<LoadedSlice, 4> LoadedSlices;
8229 // Check if this load is used as several smaller chunks of bits.
8230 // Basically, look for uses in trunc or trunc(lshr) and record a new chain
8231 // of computation for each trunc.
8232 for (SDNode::use_iterator UI = LD->use_begin(), UIEnd = LD->use_end();
8233 UI != UIEnd; ++UI) {
8234 // Skip the uses of the chain.
8235 if (UI.getUse().getResNo() != 0)
8241 // Check if this is a trunc(lshr).
8242 if (User->getOpcode() == ISD::SRL && User->hasOneUse() &&
8243 isa<ConstantSDNode>(User->getOperand(1))) {
8244 Shift = cast<ConstantSDNode>(User->getOperand(1))->getZExtValue();
8245 User = *User->use_begin();
8248 // At this point, User is a Truncate, iff we encountered, trunc or
8250 if (User->getOpcode() != ISD::TRUNCATE)
8253 // The width of the type must be a power of 2 and greater than 8-bits.
8254 // Otherwise the load cannot be represented in LLVM IR.
8255 // Moreover, if we shifted with a non-8-bits multiple, the slice
8256 // will be accross several bytes. We do not support that.
8257 unsigned Width = User->getValueSizeInBits(0);
8258 if (Width < 8 || !isPowerOf2_32(Width) || (Shift & 0x7))
8261 // Build the slice for this chain of computations.
8262 LoadedSlice LS(User, LD, Shift, &DAG);
8263 APInt CurrentUsedBits = LS.getUsedBits();
8265 // Check if this slice overlaps with another.
8266 if ((CurrentUsedBits & UsedBits) != 0)
8268 // Update the bits used globally.
8269 UsedBits |= CurrentUsedBits;
8271 // Check if the new slice would be legal.
8275 // Record the slice.
8276 LoadedSlices.push_back(LS);
8279 // Abort slicing if it does not seem to be profitable.
8280 if (!isSlicingProfitable(LoadedSlices, UsedBits, ForCodeSize))
8285 // Rewrite each chain to use an independent load.
8286 // By construction, each chain can be represented by a unique load.
8288 // Prepare the argument for the new token factor for all the slices.
8289 SmallVector<SDValue, 8> ArgChains;
8290 for (SmallVectorImpl<LoadedSlice>::const_iterator
8291 LSIt = LoadedSlices.begin(),
8292 LSItEnd = LoadedSlices.end();
8293 LSIt != LSItEnd; ++LSIt) {
8294 SDValue SliceInst = LSIt->loadSlice();
8295 CombineTo(LSIt->Inst, SliceInst, true);
8296 if (SliceInst.getNode()->getOpcode() != ISD::LOAD)
8297 SliceInst = SliceInst.getOperand(0);
8298 assert(SliceInst->getOpcode() == ISD::LOAD &&
8299 "It takes more than a zext to get to the loaded slice!!");
8300 ArgChains.push_back(SliceInst.getValue(1));
8303 SDValue Chain = DAG.getNode(ISD::TokenFactor, SDLoc(LD), MVT::Other,
8304 &ArgChains[0], ArgChains.size());
8305 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
8309 /// CheckForMaskedLoad - Check to see if V is (and load (ptr), imm), where the
8310 /// load is having specific bytes cleared out. If so, return the byte size
8311 /// being masked out and the shift amount.
8312 static std::pair<unsigned, unsigned>
8313 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
8314 std::pair<unsigned, unsigned> Result(0, 0);
8316 // Check for the structure we're looking for.
8317 if (V->getOpcode() != ISD::AND ||
8318 !isa<ConstantSDNode>(V->getOperand(1)) ||
8319 !ISD::isNormalLoad(V->getOperand(0).getNode()))
8322 // Check the chain and pointer.
8323 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
8324 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
8326 // The store should be chained directly to the load or be an operand of a
8328 if (LD == Chain.getNode())
8330 else if (Chain->getOpcode() != ISD::TokenFactor)
8331 return Result; // Fail.
8334 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
8335 if (Chain->getOperand(i).getNode() == LD) {
8339 if (!isOk) return Result;
8342 // This only handles simple types.
8343 if (V.getValueType() != MVT::i16 &&
8344 V.getValueType() != MVT::i32 &&
8345 V.getValueType() != MVT::i64)
8348 // Check the constant mask. Invert it so that the bits being masked out are
8349 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
8350 // follow the sign bit for uniformity.
8351 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
8352 unsigned NotMaskLZ = countLeadingZeros(NotMask);
8353 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
8354 unsigned NotMaskTZ = countTrailingZeros(NotMask);
8355 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
8356 if (NotMaskLZ == 64) return Result; // All zero mask.
8358 // See if we have a continuous run of bits. If so, we have 0*1+0*
8359 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
8362 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
8363 if (V.getValueType() != MVT::i64 && NotMaskLZ)
8364 NotMaskLZ -= 64-V.getValueSizeInBits();
8366 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
8367 switch (MaskedBytes) {
8371 default: return Result; // All one mask, or 5-byte mask.
8374 // Verify that the first bit starts at a multiple of mask so that the access
8375 // is aligned the same as the access width.
8376 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
8378 Result.first = MaskedBytes;
8379 Result.second = NotMaskTZ/8;
8384 /// ShrinkLoadReplaceStoreWithStore - Check to see if IVal is something that
8385 /// provides a value as specified by MaskInfo. If so, replace the specified
8386 /// store with a narrower store of truncated IVal.
8388 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
8389 SDValue IVal, StoreSDNode *St,
8391 unsigned NumBytes = MaskInfo.first;
8392 unsigned ByteShift = MaskInfo.second;
8393 SelectionDAG &DAG = DC->getDAG();
8395 // Check to see if IVal is all zeros in the part being masked in by the 'or'
8396 // that uses this. If not, this is not a replacement.
8397 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
8398 ByteShift*8, (ByteShift+NumBytes)*8);
8399 if (!DAG.MaskedValueIsZero(IVal, Mask)) return 0;
8401 // Check that it is legal on the target to do this. It is legal if the new
8402 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
8404 MVT VT = MVT::getIntegerVT(NumBytes*8);
8405 if (!DC->isTypeLegal(VT))
8408 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
8409 // shifted by ByteShift and truncated down to NumBytes.
8411 IVal = DAG.getNode(ISD::SRL, SDLoc(IVal), IVal.getValueType(), IVal,
8412 DAG.getConstant(ByteShift*8,
8413 DC->getShiftAmountTy(IVal.getValueType())));
8415 // Figure out the offset for the store and the alignment of the access.
8417 unsigned NewAlign = St->getAlignment();
8419 if (DAG.getTargetLoweringInfo().isLittleEndian())
8420 StOffset = ByteShift;
8422 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
8424 SDValue Ptr = St->getBasePtr();
8426 Ptr = DAG.getNode(ISD::ADD, SDLoc(IVal), Ptr.getValueType(),
8427 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
8428 NewAlign = MinAlign(NewAlign, StOffset);
8431 // Truncate down to the new size.
8432 IVal = DAG.getNode(ISD::TRUNCATE, SDLoc(IVal), VT, IVal);
8435 return DAG.getStore(St->getChain(), SDLoc(St), IVal, Ptr,
8436 St->getPointerInfo().getWithOffset(StOffset),
8437 false, false, NewAlign).getNode();
8441 /// ReduceLoadOpStoreWidth - Look for sequence of load / op / store where op is
8442 /// one of 'or', 'xor', and 'and' of immediates. If 'op' is only touching some
8443 /// of the loaded bits, try narrowing the load and store if it would end up
8444 /// being a win for performance or code size.
8445 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
8446 StoreSDNode *ST = cast<StoreSDNode>(N);
8447 if (ST->isVolatile())
8450 SDValue Chain = ST->getChain();
8451 SDValue Value = ST->getValue();
8452 SDValue Ptr = ST->getBasePtr();
8453 EVT VT = Value.getValueType();
8455 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
8458 unsigned Opc = Value.getOpcode();
8460 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
8461 // is a byte mask indicating a consecutive number of bytes, check to see if
8462 // Y is known to provide just those bytes. If so, we try to replace the
8463 // load + replace + store sequence with a single (narrower) store, which makes
8465 if (Opc == ISD::OR) {
8466 std::pair<unsigned, unsigned> MaskedLoad;
8467 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
8468 if (MaskedLoad.first)
8469 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
8470 Value.getOperand(1), ST,this))
8471 return SDValue(NewST, 0);
8473 // Or is commutative, so try swapping X and Y.
8474 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
8475 if (MaskedLoad.first)
8476 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
8477 Value.getOperand(0), ST,this))
8478 return SDValue(NewST, 0);
8481 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
8482 Value.getOperand(1).getOpcode() != ISD::Constant)
8485 SDValue N0 = Value.getOperand(0);
8486 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
8487 Chain == SDValue(N0.getNode(), 1)) {
8488 LoadSDNode *LD = cast<LoadSDNode>(N0);
8489 if (LD->getBasePtr() != Ptr ||
8490 LD->getPointerInfo().getAddrSpace() !=
8491 ST->getPointerInfo().getAddrSpace())
8494 // Find the type to narrow it the load / op / store to.
8495 SDValue N1 = Value.getOperand(1);
8496 unsigned BitWidth = N1.getValueSizeInBits();
8497 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
8498 if (Opc == ISD::AND)
8499 Imm ^= APInt::getAllOnesValue(BitWidth);
8500 if (Imm == 0 || Imm.isAllOnesValue())
8502 unsigned ShAmt = Imm.countTrailingZeros();
8503 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
8504 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
8505 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
8506 while (NewBW < BitWidth &&
8507 !(TLI.isOperationLegalOrCustom(Opc, NewVT) &&
8508 TLI.isNarrowingProfitable(VT, NewVT))) {
8509 NewBW = NextPowerOf2(NewBW);
8510 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
8512 if (NewBW >= BitWidth)
8515 // If the lsb changed does not start at the type bitwidth boundary,
8516 // start at the previous one.
8518 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
8519 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt,
8520 std::min(BitWidth, ShAmt + NewBW));
8521 if ((Imm & Mask) == Imm) {
8522 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
8523 if (Opc == ISD::AND)
8524 NewImm ^= APInt::getAllOnesValue(NewBW);
8525 uint64_t PtrOff = ShAmt / 8;
8526 // For big endian targets, we need to adjust the offset to the pointer to
8527 // load the correct bytes.
8528 if (TLI.isBigEndian())
8529 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
8531 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
8532 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
8533 if (NewAlign < TLI.getDataLayout()->getABITypeAlignment(NewVTTy))
8536 SDValue NewPtr = DAG.getNode(ISD::ADD, SDLoc(LD),
8537 Ptr.getValueType(), Ptr,
8538 DAG.getConstant(PtrOff, Ptr.getValueType()));
8539 SDValue NewLD = DAG.getLoad(NewVT, SDLoc(N0),
8540 LD->getChain(), NewPtr,
8541 LD->getPointerInfo().getWithOffset(PtrOff),
8542 LD->isVolatile(), LD->isNonTemporal(),
8543 LD->isInvariant(), NewAlign,
8545 SDValue NewVal = DAG.getNode(Opc, SDLoc(Value), NewVT, NewLD,
8546 DAG.getConstant(NewImm, NewVT));
8547 SDValue NewST = DAG.getStore(Chain, SDLoc(N),
8549 ST->getPointerInfo().getWithOffset(PtrOff),
8550 false, false, NewAlign);
8552 AddToWorkList(NewPtr.getNode());
8553 AddToWorkList(NewLD.getNode());
8554 AddToWorkList(NewVal.getNode());
8555 WorkListRemover DeadNodes(*this);
8556 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
8565 /// TransformFPLoadStorePair - For a given floating point load / store pair,
8566 /// if the load value isn't used by any other operations, then consider
8567 /// transforming the pair to integer load / store operations if the target
8568 /// deems the transformation profitable.
8569 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
8570 StoreSDNode *ST = cast<StoreSDNode>(N);
8571 SDValue Chain = ST->getChain();
8572 SDValue Value = ST->getValue();
8573 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
8574 Value.hasOneUse() &&
8575 Chain == SDValue(Value.getNode(), 1)) {
8576 LoadSDNode *LD = cast<LoadSDNode>(Value);
8577 EVT VT = LD->getMemoryVT();
8578 if (!VT.isFloatingPoint() ||
8579 VT != ST->getMemoryVT() ||
8580 LD->isNonTemporal() ||
8581 ST->isNonTemporal() ||
8582 LD->getPointerInfo().getAddrSpace() != 0 ||
8583 ST->getPointerInfo().getAddrSpace() != 0)
8586 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
8587 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
8588 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
8589 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
8590 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
8593 unsigned LDAlign = LD->getAlignment();
8594 unsigned STAlign = ST->getAlignment();
8595 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
8596 unsigned ABIAlign = TLI.getDataLayout()->getABITypeAlignment(IntVTTy);
8597 if (LDAlign < ABIAlign || STAlign < ABIAlign)
8600 SDValue NewLD = DAG.getLoad(IntVT, SDLoc(Value),
8601 LD->getChain(), LD->getBasePtr(),
8602 LD->getPointerInfo(),
8603 false, false, false, LDAlign);
8605 SDValue NewST = DAG.getStore(NewLD.getValue(1), SDLoc(N),
8606 NewLD, ST->getBasePtr(),
8607 ST->getPointerInfo(),
8608 false, false, STAlign);
8610 AddToWorkList(NewLD.getNode());
8611 AddToWorkList(NewST.getNode());
8612 WorkListRemover DeadNodes(*this);
8613 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
8621 /// Helper struct to parse and store a memory address as base + index + offset.
8622 /// We ignore sign extensions when it is safe to do so.
8623 /// The following two expressions are not equivalent. To differentiate we need
8624 /// to store whether there was a sign extension involved in the index
8626 /// (load (i64 add (i64 copyfromreg %c)
8627 /// (i64 signextend (add (i8 load %index)
8631 /// (load (i64 add (i64 copyfromreg %c)
8632 /// (i64 signextend (i32 add (i32 signextend (i8 load %index))
8634 struct BaseIndexOffset {
8638 bool IsIndexSignExt;
8640 BaseIndexOffset() : Offset(0), IsIndexSignExt(false) {}
8642 BaseIndexOffset(SDValue Base, SDValue Index, int64_t Offset,
8643 bool IsIndexSignExt) :
8644 Base(Base), Index(Index), Offset(Offset), IsIndexSignExt(IsIndexSignExt) {}
8646 bool equalBaseIndex(const BaseIndexOffset &Other) {
8647 return Other.Base == Base && Other.Index == Index &&
8648 Other.IsIndexSignExt == IsIndexSignExt;
8651 /// Parses tree in Ptr for base, index, offset addresses.
8652 static BaseIndexOffset match(SDValue Ptr) {
8653 bool IsIndexSignExt = false;
8655 // We only can pattern match BASE + INDEX + OFFSET. If Ptr is not an ADD
8656 // instruction, then it could be just the BASE or everything else we don't
8657 // know how to handle. Just use Ptr as BASE and give up.
8658 if (Ptr->getOpcode() != ISD::ADD)
8659 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
8661 // We know that we have at least an ADD instruction. Try to pattern match
8662 // the simple case of BASE + OFFSET.
8663 if (isa<ConstantSDNode>(Ptr->getOperand(1))) {
8664 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
8665 return BaseIndexOffset(Ptr->getOperand(0), SDValue(), Offset,
8669 // Inside a loop the current BASE pointer is calculated using an ADD and a
8670 // MUL instruction. In this case Ptr is the actual BASE pointer.
8671 // (i64 add (i64 %array_ptr)
8672 // (i64 mul (i64 %induction_var)
8673 // (i64 %element_size)))
8674 if (Ptr->getOperand(1)->getOpcode() == ISD::MUL)
8675 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
8677 // Look at Base + Index + Offset cases.
8678 SDValue Base = Ptr->getOperand(0);
8679 SDValue IndexOffset = Ptr->getOperand(1);
8681 // Skip signextends.
8682 if (IndexOffset->getOpcode() == ISD::SIGN_EXTEND) {
8683 IndexOffset = IndexOffset->getOperand(0);
8684 IsIndexSignExt = true;
8687 // Either the case of Base + Index (no offset) or something else.
8688 if (IndexOffset->getOpcode() != ISD::ADD)
8689 return BaseIndexOffset(Base, IndexOffset, 0, IsIndexSignExt);
8691 // Now we have the case of Base + Index + offset.
8692 SDValue Index = IndexOffset->getOperand(0);
8693 SDValue Offset = IndexOffset->getOperand(1);
8695 if (!isa<ConstantSDNode>(Offset))
8696 return BaseIndexOffset(Ptr, SDValue(), 0, IsIndexSignExt);
8698 // Ignore signextends.
8699 if (Index->getOpcode() == ISD::SIGN_EXTEND) {
8700 Index = Index->getOperand(0);
8701 IsIndexSignExt = true;
8702 } else IsIndexSignExt = false;
8704 int64_t Off = cast<ConstantSDNode>(Offset)->getSExtValue();
8705 return BaseIndexOffset(Base, Index, Off, IsIndexSignExt);
8709 /// Holds a pointer to an LSBaseSDNode as well as information on where it
8710 /// is located in a sequence of memory operations connected by a chain.
8712 MemOpLink (LSBaseSDNode *N, int64_t Offset, unsigned Seq):
8713 MemNode(N), OffsetFromBase(Offset), SequenceNum(Seq) { }
8714 // Ptr to the mem node.
8715 LSBaseSDNode *MemNode;
8716 // Offset from the base ptr.
8717 int64_t OffsetFromBase;
8718 // What is the sequence number of this mem node.
8719 // Lowest mem operand in the DAG starts at zero.
8720 unsigned SequenceNum;
8723 /// Sorts store nodes in a link according to their offset from a shared
8725 struct ConsecutiveMemoryChainSorter {
8726 bool operator()(MemOpLink LHS, MemOpLink RHS) {
8727 return LHS.OffsetFromBase < RHS.OffsetFromBase;
8731 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
8732 EVT MemVT = St->getMemoryVT();
8733 int64_t ElementSizeBytes = MemVT.getSizeInBits()/8;
8734 bool NoVectors = DAG.getMachineFunction().getFunction()->getAttributes().
8735 hasAttribute(AttributeSet::FunctionIndex, Attribute::NoImplicitFloat);
8737 // Don't merge vectors into wider inputs.
8738 if (MemVT.isVector() || !MemVT.isSimple())
8741 // Perform an early exit check. Do not bother looking at stored values that
8742 // are not constants or loads.
8743 SDValue StoredVal = St->getValue();
8744 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
8745 if (!isa<ConstantSDNode>(StoredVal) && !isa<ConstantFPSDNode>(StoredVal) &&
8749 // Only look at ends of store sequences.
8750 SDValue Chain = SDValue(St, 1);
8751 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
8754 // This holds the base pointer, index, and the offset in bytes from the base
8756 BaseIndexOffset BasePtr = BaseIndexOffset::match(St->getBasePtr());
8758 // We must have a base and an offset.
8759 if (!BasePtr.Base.getNode())
8762 // Do not handle stores to undef base pointers.
8763 if (BasePtr.Base.getOpcode() == ISD::UNDEF)
8766 // Save the LoadSDNodes that we find in the chain.
8767 // We need to make sure that these nodes do not interfere with
8768 // any of the store nodes.
8769 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
8771 // Save the StoreSDNodes that we find in the chain.
8772 SmallVector<MemOpLink, 8> StoreNodes;
8774 // Walk up the chain and look for nodes with offsets from the same
8775 // base pointer. Stop when reaching an instruction with a different kind
8776 // or instruction which has a different base pointer.
8778 StoreSDNode *Index = St;
8780 // If the chain has more than one use, then we can't reorder the mem ops.
8781 if (Index != St && !SDValue(Index, 1)->hasOneUse())
8784 // Find the base pointer and offset for this memory node.
8785 BaseIndexOffset Ptr = BaseIndexOffset::match(Index->getBasePtr());
8787 // Check that the base pointer is the same as the original one.
8788 if (!Ptr.equalBaseIndex(BasePtr))
8791 // Check that the alignment is the same.
8792 if (Index->getAlignment() != St->getAlignment())
8795 // The memory operands must not be volatile.
8796 if (Index->isVolatile() || Index->isIndexed())
8800 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
8801 if (St->isTruncatingStore())
8804 // The stored memory type must be the same.
8805 if (Index->getMemoryVT() != MemVT)
8808 // We do not allow unaligned stores because we want to prevent overriding
8810 if (Index->getAlignment()*8 != MemVT.getSizeInBits())
8813 // We found a potential memory operand to merge.
8814 StoreNodes.push_back(MemOpLink(Index, Ptr.Offset, Seq++));
8816 // Find the next memory operand in the chain. If the next operand in the
8817 // chain is a store then move up and continue the scan with the next
8818 // memory operand. If the next operand is a load save it and use alias
8819 // information to check if it interferes with anything.
8820 SDNode *NextInChain = Index->getChain().getNode();
8822 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
8823 // We found a store node. Use it for the next iteration.
8826 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
8827 if (Ldn->isVolatile()) {
8832 // Save the load node for later. Continue the scan.
8833 AliasLoadNodes.push_back(Ldn);
8834 NextInChain = Ldn->getChain().getNode();
8843 // Check if there is anything to merge.
8844 if (StoreNodes.size() < 2)
8847 // Sort the memory operands according to their distance from the base pointer.
8848 std::sort(StoreNodes.begin(), StoreNodes.end(),
8849 ConsecutiveMemoryChainSorter());
8851 // Scan the memory operations on the chain and find the first non-consecutive
8852 // store memory address.
8853 unsigned LastConsecutiveStore = 0;
8854 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
8855 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) {
8857 // Check that the addresses are consecutive starting from the second
8858 // element in the list of stores.
8860 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
8861 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
8866 // Check if this store interferes with any of the loads that we found.
8867 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld)
8868 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) {
8872 // We found a load that alias with this store. Stop the sequence.
8876 // Mark this node as useful.
8877 LastConsecutiveStore = i;
8880 // The node with the lowest store address.
8881 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
8883 // Store the constants into memory as one consecutive store.
8885 unsigned LastLegalType = 0;
8886 unsigned LastLegalVectorType = 0;
8887 bool NonZero = false;
8888 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
8889 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
8890 SDValue StoredVal = St->getValue();
8892 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
8893 NonZero |= !C->isNullValue();
8894 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) {
8895 NonZero |= !C->getConstantFPValue()->isNullValue();
8901 // Find a legal type for the constant store.
8902 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
8903 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
8904 if (TLI.isTypeLegal(StoreTy))
8905 LastLegalType = i+1;
8906 // Or check whether a truncstore is legal.
8907 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
8908 TargetLowering::TypePromoteInteger) {
8909 EVT LegalizedStoredValueTy =
8910 TLI.getTypeToTransformTo(*DAG.getContext(), StoredVal.getValueType());
8911 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy))
8912 LastLegalType = i+1;
8915 // Find a legal type for the vector store.
8916 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
8917 if (TLI.isTypeLegal(Ty))
8918 LastLegalVectorType = i + 1;
8921 // We only use vectors if the constant is known to be zero and the
8922 // function is not marked with the noimplicitfloat attribute.
8923 if (NonZero || NoVectors)
8924 LastLegalVectorType = 0;
8926 // Check if we found a legal integer type to store.
8927 if (LastLegalType == 0 && LastLegalVectorType == 0)
8930 bool UseVector = (LastLegalVectorType > LastLegalType) && !NoVectors;
8931 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType;
8933 // Make sure we have something to merge.
8937 unsigned EarliestNodeUsed = 0;
8938 for (unsigned i=0; i < NumElem; ++i) {
8939 // Find a chain for the new wide-store operand. Notice that some
8940 // of the store nodes that we found may not be selected for inclusion
8941 // in the wide store. The chain we use needs to be the chain of the
8942 // earliest store node which is *used* and replaced by the wide store.
8943 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
8944 EarliestNodeUsed = i;
8947 // The earliest Node in the DAG.
8948 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
8949 SDLoc DL(StoreNodes[0].MemNode);
8953 // Find a legal type for the vector store.
8954 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
8955 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
8956 StoredVal = DAG.getConstant(0, Ty);
8958 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
8959 APInt StoreInt(StoreBW, 0);
8961 // Construct a single integer constant which is made of the smaller
8963 bool IsLE = TLI.isLittleEndian();
8964 for (unsigned i = 0; i < NumElem ; ++i) {
8965 unsigned Idx = IsLE ?(NumElem - 1 - i) : i;
8966 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
8967 SDValue Val = St->getValue();
8968 StoreInt<<=ElementSizeBytes*8;
8969 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
8970 StoreInt|=C->getAPIntValue().zext(StoreBW);
8971 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
8972 StoreInt|= C->getValueAPF().bitcastToAPInt().zext(StoreBW);
8974 assert(false && "Invalid constant element type");
8978 // Create the new Load and Store operations.
8979 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
8980 StoredVal = DAG.getConstant(StoreInt, StoreTy);
8983 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), DL, StoredVal,
8984 FirstInChain->getBasePtr(),
8985 FirstInChain->getPointerInfo(),
8987 FirstInChain->getAlignment());
8989 // Replace the first store with the new store
8990 CombineTo(EarliestOp, NewStore);
8991 // Erase all other stores.
8992 for (unsigned i = 0; i < NumElem ; ++i) {
8993 if (StoreNodes[i].MemNode == EarliestOp)
8995 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
8996 // ReplaceAllUsesWith will replace all uses that existed when it was
8997 // called, but graph optimizations may cause new ones to appear. For
8998 // example, the case in pr14333 looks like
9000 // St's chain -> St -> another store -> X
9002 // And the only difference from St to the other store is the chain.
9003 // When we change it's chain to be St's chain they become identical,
9004 // get CSEed and the net result is that X is now a use of St.
9005 // Since we know that St is redundant, just iterate.
9006 while (!St->use_empty())
9007 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain());
9008 removeFromWorkList(St);
9015 // Below we handle the case of multiple consecutive stores that
9016 // come from multiple consecutive loads. We merge them into a single
9017 // wide load and a single wide store.
9019 // Look for load nodes which are used by the stored values.
9020 SmallVector<MemOpLink, 8> LoadNodes;
9022 // Find acceptable loads. Loads need to have the same chain (token factor),
9023 // must not be zext, volatile, indexed, and they must be consecutive.
9024 BaseIndexOffset LdBasePtr;
9025 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
9026 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9027 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
9030 // Loads must only have one use.
9031 if (!Ld->hasNUsesOfValue(1, 0))
9034 // Check that the alignment is the same as the stores.
9035 if (Ld->getAlignment() != St->getAlignment())
9038 // The memory operands must not be volatile.
9039 if (Ld->isVolatile() || Ld->isIndexed())
9042 // We do not accept ext loads.
9043 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
9046 // The stored memory type must be the same.
9047 if (Ld->getMemoryVT() != MemVT)
9050 BaseIndexOffset LdPtr = BaseIndexOffset::match(Ld->getBasePtr());
9051 // If this is not the first ptr that we check.
9052 if (LdBasePtr.Base.getNode()) {
9053 // The base ptr must be the same.
9054 if (!LdPtr.equalBaseIndex(LdBasePtr))
9057 // Check that all other base pointers are the same as this one.
9061 // We found a potential memory operand to merge.
9062 LoadNodes.push_back(MemOpLink(Ld, LdPtr.Offset, 0));
9065 if (LoadNodes.size() < 2)
9068 // Scan the memory operations on the chain and find the first non-consecutive
9069 // load memory address. These variables hold the index in the store node
9071 unsigned LastConsecutiveLoad = 0;
9072 // This variable refers to the size and not index in the array.
9073 unsigned LastLegalVectorType = 0;
9074 unsigned LastLegalIntegerType = 0;
9075 StartAddress = LoadNodes[0].OffsetFromBase;
9076 SDValue FirstChain = LoadNodes[0].MemNode->getChain();
9077 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
9078 // All loads much share the same chain.
9079 if (LoadNodes[i].MemNode->getChain() != FirstChain)
9082 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
9083 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
9085 LastConsecutiveLoad = i;
9087 // Find a legal type for the vector store.
9088 EVT StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
9089 if (TLI.isTypeLegal(StoreTy))
9090 LastLegalVectorType = i + 1;
9092 // Find a legal type for the integer store.
9093 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
9094 StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9095 if (TLI.isTypeLegal(StoreTy))
9096 LastLegalIntegerType = i + 1;
9097 // Or check whether a truncstore and extload is legal.
9098 else if (TLI.getTypeAction(*DAG.getContext(), StoreTy) ==
9099 TargetLowering::TypePromoteInteger) {
9100 EVT LegalizedStoredValueTy =
9101 TLI.getTypeToTransformTo(*DAG.getContext(), StoreTy);
9102 if (TLI.isTruncStoreLegal(LegalizedStoredValueTy, StoreTy) &&
9103 TLI.isLoadExtLegal(ISD::ZEXTLOAD, StoreTy) &&
9104 TLI.isLoadExtLegal(ISD::SEXTLOAD, StoreTy) &&
9105 TLI.isLoadExtLegal(ISD::EXTLOAD, StoreTy))
9106 LastLegalIntegerType = i+1;
9110 // Only use vector types if the vector type is larger than the integer type.
9111 // If they are the same, use integers.
9112 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType && !NoVectors;
9113 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType);
9115 // We add +1 here because the LastXXX variables refer to location while
9116 // the NumElem refers to array/index size.
9117 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1;
9118 NumElem = std::min(LastLegalType, NumElem);
9123 // The earliest Node in the DAG.
9124 unsigned EarliestNodeUsed = 0;
9125 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
9126 for (unsigned i=1; i<NumElem; ++i) {
9127 // Find a chain for the new wide-store operand. Notice that some
9128 // of the store nodes that we found may not be selected for inclusion
9129 // in the wide store. The chain we use needs to be the chain of the
9130 // earliest store node which is *used* and replaced by the wide store.
9131 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
9132 EarliestNodeUsed = i;
9135 // Find if it is better to use vectors or integers to load and store
9139 JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
9141 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
9142 JointMemOpVT = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
9145 SDLoc LoadDL(LoadNodes[0].MemNode);
9146 SDLoc StoreDL(StoreNodes[0].MemNode);
9148 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
9149 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL,
9150 FirstLoad->getChain(),
9151 FirstLoad->getBasePtr(),
9152 FirstLoad->getPointerInfo(),
9153 false, false, false,
9154 FirstLoad->getAlignment());
9156 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), StoreDL, NewLoad,
9157 FirstInChain->getBasePtr(),
9158 FirstInChain->getPointerInfo(), false, false,
9159 FirstInChain->getAlignment());
9161 // Replace one of the loads with the new load.
9162 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode);
9163 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
9164 SDValue(NewLoad.getNode(), 1));
9166 // Remove the rest of the load chains.
9167 for (unsigned i = 1; i < NumElem ; ++i) {
9168 // Replace all chain users of the old load nodes with the chain of the new
9170 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
9171 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain());
9174 // Replace the first store with the new store.
9175 CombineTo(EarliestOp, NewStore);
9176 // Erase all other stores.
9177 for (unsigned i = 0; i < NumElem ; ++i) {
9178 // Remove all Store nodes.
9179 if (StoreNodes[i].MemNode == EarliestOp)
9181 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
9182 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain());
9183 removeFromWorkList(St);
9190 SDValue DAGCombiner::visitSTORE(SDNode *N) {
9191 StoreSDNode *ST = cast<StoreSDNode>(N);
9192 SDValue Chain = ST->getChain();
9193 SDValue Value = ST->getValue();
9194 SDValue Ptr = ST->getBasePtr();
9196 // If this is a store of a bit convert, store the input value if the
9197 // resultant store does not need a higher alignment than the original.
9198 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
9199 ST->isUnindexed()) {
9200 unsigned OrigAlign = ST->getAlignment();
9201 EVT SVT = Value.getOperand(0).getValueType();
9202 unsigned Align = TLI.getDataLayout()->
9203 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
9204 if (Align <= OrigAlign &&
9205 ((!LegalOperations && !ST->isVolatile()) ||
9206 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
9207 return DAG.getStore(Chain, SDLoc(N), Value.getOperand(0),
9208 Ptr, ST->getPointerInfo(), ST->isVolatile(),
9209 ST->isNonTemporal(), OrigAlign,
9213 // Turn 'store undef, Ptr' -> nothing.
9214 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
9217 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
9218 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
9219 // NOTE: If the original store is volatile, this transform must not increase
9220 // the number of stores. For example, on x86-32 an f64 can be stored in one
9221 // processor operation but an i64 (which is not legal) requires two. So the
9222 // transform should not be done in this case.
9223 if (Value.getOpcode() != ISD::TargetConstantFP) {
9225 switch (CFP->getSimpleValueType(0).SimpleTy) {
9226 default: llvm_unreachable("Unknown FP type");
9227 case MVT::f16: // We don't do this for these yet.
9233 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
9234 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
9235 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
9236 bitcastToAPInt().getZExtValue(), MVT::i32);
9237 return DAG.getStore(Chain, SDLoc(N), Tmp,
9238 Ptr, ST->getMemOperand());
9242 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
9243 !ST->isVolatile()) ||
9244 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
9245 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
9246 getZExtValue(), MVT::i64);
9247 return DAG.getStore(Chain, SDLoc(N), Tmp,
9248 Ptr, ST->getMemOperand());
9251 if (!ST->isVolatile() &&
9252 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
9253 // Many FP stores are not made apparent until after legalize, e.g. for
9254 // argument passing. Since this is so common, custom legalize the
9255 // 64-bit integer store into two 32-bit stores.
9256 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
9257 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
9258 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
9259 if (TLI.isBigEndian()) std::swap(Lo, Hi);
9261 unsigned Alignment = ST->getAlignment();
9262 bool isVolatile = ST->isVolatile();
9263 bool isNonTemporal = ST->isNonTemporal();
9264 const MDNode *TBAAInfo = ST->getTBAAInfo();
9266 SDValue St0 = DAG.getStore(Chain, SDLoc(ST), Lo,
9267 Ptr, ST->getPointerInfo(),
9268 isVolatile, isNonTemporal,
9269 ST->getAlignment(), TBAAInfo);
9270 Ptr = DAG.getNode(ISD::ADD, SDLoc(N), Ptr.getValueType(), Ptr,
9271 DAG.getConstant(4, Ptr.getValueType()));
9272 Alignment = MinAlign(Alignment, 4U);
9273 SDValue St1 = DAG.getStore(Chain, SDLoc(ST), Hi,
9274 Ptr, ST->getPointerInfo().getWithOffset(4),
9275 isVolatile, isNonTemporal,
9276 Alignment, TBAAInfo);
9277 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other,
9286 // Try to infer better alignment information than the store already has.
9287 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
9288 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
9289 if (Align > ST->getAlignment())
9290 return DAG.getTruncStore(Chain, SDLoc(N), Value,
9291 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
9292 ST->isVolatile(), ST->isNonTemporal(), Align,
9297 // Try transforming a pair floating point load / store ops to integer
9298 // load / store ops.
9299 SDValue NewST = TransformFPLoadStorePair(N);
9300 if (NewST.getNode())
9303 bool UseAA = CombinerAA.getNumOccurrences() > 0 ? CombinerAA :
9304 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
9306 // Walk up chain skipping non-aliasing memory nodes.
9307 SDValue BetterChain = FindBetterChain(N, Chain);
9309 // If there is a better chain.
9310 if (Chain != BetterChain) {
9313 // Replace the chain to avoid dependency.
9314 if (ST->isTruncatingStore()) {
9315 ReplStore = DAG.getTruncStore(BetterChain, SDLoc(N), Value, Ptr,
9316 ST->getMemoryVT(), ST->getMemOperand());
9318 ReplStore = DAG.getStore(BetterChain, SDLoc(N), Value, Ptr,
9319 ST->getMemOperand());
9322 // Create token to keep both nodes around.
9323 SDValue Token = DAG.getNode(ISD::TokenFactor, SDLoc(N),
9324 MVT::Other, Chain, ReplStore);
9326 // Make sure the new and old chains are cleaned up.
9327 AddToWorkList(Token.getNode());
9329 // Don't add users to work list.
9330 return CombineTo(N, Token, false);
9334 // Try transforming N to an indexed store.
9335 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
9336 return SDValue(N, 0);
9338 // FIXME: is there such a thing as a truncating indexed store?
9339 if (ST->isTruncatingStore() && ST->isUnindexed() &&
9340 Value.getValueType().isInteger()) {
9341 // See if we can simplify the input to this truncstore with knowledge that
9342 // only the low bits are being used. For example:
9343 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
9345 GetDemandedBits(Value,
9346 APInt::getLowBitsSet(
9347 Value.getValueType().getScalarType().getSizeInBits(),
9348 ST->getMemoryVT().getScalarType().getSizeInBits()));
9349 AddToWorkList(Value.getNode());
9350 if (Shorter.getNode())
9351 return DAG.getTruncStore(Chain, SDLoc(N), Shorter,
9352 Ptr, ST->getMemoryVT(), ST->getMemOperand());
9354 // Otherwise, see if we can simplify the operation with
9355 // SimplifyDemandedBits, which only works if the value has a single use.
9356 if (SimplifyDemandedBits(Value,
9357 APInt::getLowBitsSet(
9358 Value.getValueType().getScalarType().getSizeInBits(),
9359 ST->getMemoryVT().getScalarType().getSizeInBits())))
9360 return SDValue(N, 0);
9363 // If this is a load followed by a store to the same location, then the store
9365 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
9366 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
9367 ST->isUnindexed() && !ST->isVolatile() &&
9368 // There can't be any side effects between the load and store, such as
9370 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
9371 // The store is dead, remove it.
9376 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
9377 // truncating store. We can do this even if this is already a truncstore.
9378 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
9379 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
9380 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
9381 ST->getMemoryVT())) {
9382 return DAG.getTruncStore(Chain, SDLoc(N), Value.getOperand(0),
9383 Ptr, ST->getMemoryVT(), ST->getMemOperand());
9386 // Only perform this optimization before the types are legal, because we
9387 // don't want to perform this optimization on every DAGCombine invocation.
9389 bool EverChanged = false;
9392 // There can be multiple store sequences on the same chain.
9393 // Keep trying to merge store sequences until we are unable to do so
9394 // or until we merge the last store on the chain.
9395 bool Changed = MergeConsecutiveStores(ST);
9396 EverChanged |= Changed;
9397 if (!Changed) break;
9398 } while (ST->getOpcode() != ISD::DELETED_NODE);
9401 return SDValue(N, 0);
9404 return ReduceLoadOpStoreWidth(N);
9407 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
9408 SDValue InVec = N->getOperand(0);
9409 SDValue InVal = N->getOperand(1);
9410 SDValue EltNo = N->getOperand(2);
9413 // If the inserted element is an UNDEF, just use the input vector.
9414 if (InVal.getOpcode() == ISD::UNDEF)
9417 EVT VT = InVec.getValueType();
9419 // If we can't generate a legal BUILD_VECTOR, exit
9420 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
9423 // Check that we know which element is being inserted
9424 if (!isa<ConstantSDNode>(EltNo))
9426 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9428 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
9429 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
9431 SmallVector<SDValue, 8> Ops;
9432 // Do not combine these two vectors if the output vector will not replace
9433 // the input vector.
9434 if (InVec.getOpcode() == ISD::BUILD_VECTOR && InVec.hasOneUse()) {
9435 Ops.append(InVec.getNode()->op_begin(),
9436 InVec.getNode()->op_end());
9437 } else if (InVec.getOpcode() == ISD::UNDEF) {
9438 unsigned NElts = VT.getVectorNumElements();
9439 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
9444 // Insert the element
9445 if (Elt < Ops.size()) {
9446 // All the operands of BUILD_VECTOR must have the same type;
9447 // we enforce that here.
9448 EVT OpVT = Ops[0].getValueType();
9449 if (InVal.getValueType() != OpVT)
9450 InVal = OpVT.bitsGT(InVal.getValueType()) ?
9451 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
9452 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
9456 // Return the new vector
9457 return DAG.getNode(ISD::BUILD_VECTOR, dl,
9458 VT, &Ops[0], Ops.size());
9461 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
9462 // (vextract (scalar_to_vector val, 0) -> val
9463 SDValue InVec = N->getOperand(0);
9464 EVT VT = InVec.getValueType();
9465 EVT NVT = N->getValueType(0);
9467 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
9468 // Check if the result type doesn't match the inserted element type. A
9469 // SCALAR_TO_VECTOR may truncate the inserted element and the
9470 // EXTRACT_VECTOR_ELT may widen the extracted vector.
9471 SDValue InOp = InVec.getOperand(0);
9472 if (InOp.getValueType() != NVT) {
9473 assert(InOp.getValueType().isInteger() && NVT.isInteger());
9474 return DAG.getSExtOrTrunc(InOp, SDLoc(InVec), NVT);
9479 SDValue EltNo = N->getOperand(1);
9480 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
9482 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
9483 // We only perform this optimization before the op legalization phase because
9484 // we may introduce new vector instructions which are not backed by TD
9485 // patterns. For example on AVX, extracting elements from a wide vector
9486 // without using extract_subvector.
9487 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
9488 && ConstEltNo && !LegalOperations) {
9489 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9490 int NumElem = VT.getVectorNumElements();
9491 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
9492 // Find the new index to extract from.
9493 int OrigElt = SVOp->getMaskElt(Elt);
9495 // Extracting an undef index is undef.
9497 return DAG.getUNDEF(NVT);
9499 // Select the right vector half to extract from.
9500 if (OrigElt < NumElem) {
9501 InVec = InVec->getOperand(0);
9503 InVec = InVec->getOperand(1);
9507 EVT IndexTy = TLI.getVectorIdxTy();
9508 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), NVT,
9509 InVec, DAG.getConstant(OrigElt, IndexTy));
9512 // Perform only after legalization to ensure build_vector / vector_shuffle
9513 // optimizations have already been done.
9514 if (!LegalOperations) return SDValue();
9516 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
9517 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
9518 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
9521 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
9522 bool NewLoad = false;
9523 bool BCNumEltsChanged = false;
9524 EVT ExtVT = VT.getVectorElementType();
9527 // If the result of load has to be truncated, then it's not necessarily
9529 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
9532 if (InVec.getOpcode() == ISD::BITCAST) {
9533 // Don't duplicate a load with other uses.
9534 if (!InVec.hasOneUse())
9537 EVT BCVT = InVec.getOperand(0).getValueType();
9538 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
9540 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
9541 BCNumEltsChanged = true;
9542 InVec = InVec.getOperand(0);
9543 ExtVT = BCVT.getVectorElementType();
9547 LoadSDNode *LN0 = NULL;
9548 const ShuffleVectorSDNode *SVN = NULL;
9549 if (ISD::isNormalLoad(InVec.getNode())) {
9550 LN0 = cast<LoadSDNode>(InVec);
9551 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
9552 InVec.getOperand(0).getValueType() == ExtVT &&
9553 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
9554 // Don't duplicate a load with other uses.
9555 if (!InVec.hasOneUse())
9558 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
9559 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
9560 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
9562 // (load $addr+1*size)
9564 // Don't duplicate a load with other uses.
9565 if (!InVec.hasOneUse())
9568 // If the bit convert changed the number of elements, it is unsafe
9569 // to examine the mask.
9570 if (BCNumEltsChanged)
9573 // Select the input vector, guarding against out of range extract vector.
9574 unsigned NumElems = VT.getVectorNumElements();
9575 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
9576 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
9578 if (InVec.getOpcode() == ISD::BITCAST) {
9579 // Don't duplicate a load with other uses.
9580 if (!InVec.hasOneUse())
9583 InVec = InVec.getOperand(0);
9585 if (ISD::isNormalLoad(InVec.getNode())) {
9586 LN0 = cast<LoadSDNode>(InVec);
9587 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
9591 // Make sure we found a non-volatile load and the extractelement is
9593 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
9596 // If Idx was -1 above, Elt is going to be -1, so just return undef.
9598 return DAG.getUNDEF(LVT);
9600 unsigned Align = LN0->getAlignment();
9602 // Check the resultant load doesn't need a higher alignment than the
9606 ->getABITypeAlignment(LVT.getTypeForEVT(*DAG.getContext()));
9608 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, LVT))
9614 SDValue NewPtr = LN0->getBasePtr();
9615 unsigned PtrOff = 0;
9618 PtrOff = LVT.getSizeInBits() * Elt / 8;
9619 EVT PtrType = NewPtr.getValueType();
9620 if (TLI.isBigEndian())
9621 PtrOff = VT.getSizeInBits() / 8 - PtrOff;
9622 NewPtr = DAG.getNode(ISD::ADD, SDLoc(N), PtrType, NewPtr,
9623 DAG.getConstant(PtrOff, PtrType));
9626 // The replacement we need to do here is a little tricky: we need to
9627 // replace an extractelement of a load with a load.
9628 // Use ReplaceAllUsesOfValuesWith to do the replacement.
9629 // Note that this replacement assumes that the extractvalue is the only
9630 // use of the load; that's okay because we don't want to perform this
9631 // transformation in other cases anyway.
9634 if (NVT.bitsGT(LVT)) {
9635 // If the result type of vextract is wider than the load, then issue an
9636 // extending load instead.
9637 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, LVT)
9638 ? ISD::ZEXTLOAD : ISD::EXTLOAD;
9639 Load = DAG.getExtLoad(ExtType, SDLoc(N), NVT, LN0->getChain(),
9640 NewPtr, LN0->getPointerInfo().getWithOffset(PtrOff),
9641 LVT, LN0->isVolatile(), LN0->isNonTemporal(),
9642 Align, LN0->getTBAAInfo());
9643 Chain = Load.getValue(1);
9645 Load = DAG.getLoad(LVT, SDLoc(N), LN0->getChain(), NewPtr,
9646 LN0->getPointerInfo().getWithOffset(PtrOff),
9647 LN0->isVolatile(), LN0->isNonTemporal(),
9648 LN0->isInvariant(), Align, LN0->getTBAAInfo());
9649 Chain = Load.getValue(1);
9650 if (NVT.bitsLT(LVT))
9651 Load = DAG.getNode(ISD::TRUNCATE, SDLoc(N), NVT, Load);
9653 Load = DAG.getNode(ISD::BITCAST, SDLoc(N), NVT, Load);
9655 WorkListRemover DeadNodes(*this);
9656 SDValue From[] = { SDValue(N, 0), SDValue(LN0,1) };
9657 SDValue To[] = { Load, Chain };
9658 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
9659 // Since we're explcitly calling ReplaceAllUses, add the new node to the
9660 // worklist explicitly as well.
9661 AddToWorkList(Load.getNode());
9662 AddUsersToWorkList(Load.getNode()); // Add users too
9663 // Make sure to revisit this node to clean it up; it will usually be dead.
9665 return SDValue(N, 0);
9671 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
9672 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
9673 // We perform this optimization post type-legalization because
9674 // the type-legalizer often scalarizes integer-promoted vectors.
9675 // Performing this optimization before may create bit-casts which
9676 // will be type-legalized to complex code sequences.
9677 // We perform this optimization only before the operation legalizer because we
9678 // may introduce illegal operations.
9679 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
9682 unsigned NumInScalars = N->getNumOperands();
9684 EVT VT = N->getValueType(0);
9686 // Check to see if this is a BUILD_VECTOR of a bunch of values
9687 // which come from any_extend or zero_extend nodes. If so, we can create
9688 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
9689 // optimizations. We do not handle sign-extend because we can't fill the sign
9691 EVT SourceType = MVT::Other;
9692 bool AllAnyExt = true;
9694 for (unsigned i = 0; i != NumInScalars; ++i) {
9695 SDValue In = N->getOperand(i);
9696 // Ignore undef inputs.
9697 if (In.getOpcode() == ISD::UNDEF) continue;
9699 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
9700 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
9702 // Abort if the element is not an extension.
9703 if (!ZeroExt && !AnyExt) {
9704 SourceType = MVT::Other;
9708 // The input is a ZeroExt or AnyExt. Check the original type.
9709 EVT InTy = In.getOperand(0).getValueType();
9711 // Check that all of the widened source types are the same.
9712 if (SourceType == MVT::Other)
9715 else if (InTy != SourceType) {
9716 // Multiple income types. Abort.
9717 SourceType = MVT::Other;
9721 // Check if all of the extends are ANY_EXTENDs.
9722 AllAnyExt &= AnyExt;
9725 // In order to have valid types, all of the inputs must be extended from the
9726 // same source type and all of the inputs must be any or zero extend.
9727 // Scalar sizes must be a power of two.
9728 EVT OutScalarTy = VT.getScalarType();
9729 bool ValidTypes = SourceType != MVT::Other &&
9730 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
9731 isPowerOf2_32(SourceType.getSizeInBits());
9733 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
9734 // turn into a single shuffle instruction.
9738 bool isLE = TLI.isLittleEndian();
9739 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
9740 assert(ElemRatio > 1 && "Invalid element size ratio");
9741 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
9742 DAG.getConstant(0, SourceType);
9744 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
9745 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
9747 // Populate the new build_vector
9748 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
9749 SDValue Cast = N->getOperand(i);
9750 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
9751 Cast.getOpcode() == ISD::ZERO_EXTEND ||
9752 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
9754 if (Cast.getOpcode() == ISD::UNDEF)
9755 In = DAG.getUNDEF(SourceType);
9757 In = Cast->getOperand(0);
9758 unsigned Index = isLE ? (i * ElemRatio) :
9759 (i * ElemRatio + (ElemRatio - 1));
9761 assert(Index < Ops.size() && "Invalid index");
9765 // The type of the new BUILD_VECTOR node.
9766 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
9767 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
9768 "Invalid vector size");
9769 // Check if the new vector type is legal.
9770 if (!isTypeLegal(VecVT)) return SDValue();
9772 // Make the new BUILD_VECTOR.
9773 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], Ops.size());
9775 // The new BUILD_VECTOR node has the potential to be further optimized.
9776 AddToWorkList(BV.getNode());
9777 // Bitcast to the desired type.
9778 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
9781 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
9782 EVT VT = N->getValueType(0);
9784 unsigned NumInScalars = N->getNumOperands();
9787 EVT SrcVT = MVT::Other;
9788 unsigned Opcode = ISD::DELETED_NODE;
9789 unsigned NumDefs = 0;
9791 for (unsigned i = 0; i != NumInScalars; ++i) {
9792 SDValue In = N->getOperand(i);
9793 unsigned Opc = In.getOpcode();
9795 if (Opc == ISD::UNDEF)
9798 // If all scalar values are floats and converted from integers.
9799 if (Opcode == ISD::DELETED_NODE &&
9800 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
9807 EVT InVT = In.getOperand(0).getValueType();
9809 // If all scalar values are typed differently, bail out. It's chosen to
9810 // simplify BUILD_VECTOR of integer types.
9811 if (SrcVT == MVT::Other)
9818 // If the vector has just one element defined, it's not worth to fold it into
9819 // a vectorized one.
9823 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
9824 && "Should only handle conversion from integer to float.");
9825 assert(SrcVT != MVT::Other && "Cannot determine source type!");
9827 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
9829 if (!TLI.isOperationLegalOrCustom(Opcode, NVT))
9832 SmallVector<SDValue, 8> Opnds;
9833 for (unsigned i = 0; i != NumInScalars; ++i) {
9834 SDValue In = N->getOperand(i);
9836 if (In.getOpcode() == ISD::UNDEF)
9837 Opnds.push_back(DAG.getUNDEF(SrcVT));
9839 Opnds.push_back(In.getOperand(0));
9841 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT,
9842 &Opnds[0], Opnds.size());
9843 AddToWorkList(BV.getNode());
9845 return DAG.getNode(Opcode, dl, VT, BV);
9848 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
9849 unsigned NumInScalars = N->getNumOperands();
9851 EVT VT = N->getValueType(0);
9853 // A vector built entirely of undefs is undef.
9854 if (ISD::allOperandsUndef(N))
9855 return DAG.getUNDEF(VT);
9857 SDValue V = reduceBuildVecExtToExtBuildVec(N);
9861 V = reduceBuildVecConvertToConvertBuildVec(N);
9865 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
9866 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
9867 // at most two distinct vectors, turn this into a shuffle node.
9869 // May only combine to shuffle after legalize if shuffle is legal.
9870 if (LegalOperations &&
9871 !TLI.isOperationLegalOrCustom(ISD::VECTOR_SHUFFLE, VT))
9874 SDValue VecIn1, VecIn2;
9875 for (unsigned i = 0; i != NumInScalars; ++i) {
9876 // Ignore undef inputs.
9877 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
9879 // If this input is something other than a EXTRACT_VECTOR_ELT with a
9880 // constant index, bail out.
9881 if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
9882 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
9883 VecIn1 = VecIn2 = SDValue(0, 0);
9887 // We allow up to two distinct input vectors.
9888 SDValue ExtractedFromVec = N->getOperand(i).getOperand(0);
9889 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
9892 if (VecIn1.getNode() == 0) {
9893 VecIn1 = ExtractedFromVec;
9894 } else if (VecIn2.getNode() == 0) {
9895 VecIn2 = ExtractedFromVec;
9898 VecIn1 = VecIn2 = SDValue(0, 0);
9903 // If everything is good, we can make a shuffle operation.
9904 if (VecIn1.getNode()) {
9905 SmallVector<int, 8> Mask;
9906 for (unsigned i = 0; i != NumInScalars; ++i) {
9907 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
9912 // If extracting from the first vector, just use the index directly.
9913 SDValue Extract = N->getOperand(i);
9914 SDValue ExtVal = Extract.getOperand(1);
9915 if (Extract.getOperand(0) == VecIn1) {
9916 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
9917 if (ExtIndex > VT.getVectorNumElements())
9920 Mask.push_back(ExtIndex);
9924 // Otherwise, use InIdx + VecSize
9925 unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue();
9926 Mask.push_back(Idx+NumInScalars);
9929 // We can't generate a shuffle node with mismatched input and output types.
9930 // Attempt to transform a single input vector to the correct type.
9931 if ((VT != VecIn1.getValueType())) {
9932 // We don't support shuffeling between TWO values of different types.
9933 if (VecIn2.getNode() != 0)
9936 // We only support widening of vectors which are half the size of the
9937 // output registers. For example XMM->YMM widening on X86 with AVX.
9938 if (VecIn1.getValueType().getSizeInBits()*2 != VT.getSizeInBits())
9941 // If the input vector type has a different base type to the output
9942 // vector type, bail out.
9943 if (VecIn1.getValueType().getVectorElementType() !=
9944 VT.getVectorElementType())
9947 // Widen the input vector by adding undef values.
9948 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
9949 VecIn1, DAG.getUNDEF(VecIn1.getValueType()));
9952 // If VecIn2 is unused then change it to undef.
9953 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
9955 // Check that we were able to transform all incoming values to the same
9957 if (VecIn2.getValueType() != VecIn1.getValueType() ||
9958 VecIn1.getValueType() != VT)
9961 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
9962 if (!isTypeLegal(VT))
9965 // Return the new VECTOR_SHUFFLE node.
9969 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]);
9975 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
9976 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
9977 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
9978 // inputs come from at most two distinct vectors, turn this into a shuffle
9981 // If we only have one input vector, we don't need to do any concatenation.
9982 if (N->getNumOperands() == 1)
9983 return N->getOperand(0);
9985 // Check if all of the operands are undefs.
9986 EVT VT = N->getValueType(0);
9987 if (ISD::allOperandsUndef(N))
9988 return DAG.getUNDEF(VT);
9990 // Optimize concat_vectors where one of the vectors is undef.
9991 if (N->getNumOperands() == 2 &&
9992 N->getOperand(1)->getOpcode() == ISD::UNDEF) {
9993 SDValue In = N->getOperand(0);
9994 assert(In.getValueType().isVector() && "Must concat vectors");
9996 // Transform: concat_vectors(scalar, undef) -> scalar_to_vector(sclr).
9997 if (In->getOpcode() == ISD::BITCAST &&
9998 !In->getOperand(0)->getValueType(0).isVector()) {
9999 SDValue Scalar = In->getOperand(0);
10000 EVT SclTy = Scalar->getValueType(0);
10002 if (!SclTy.isFloatingPoint() && !SclTy.isInteger())
10005 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SclTy,
10006 VT.getSizeInBits() / SclTy.getSizeInBits());
10007 if (!TLI.isTypeLegal(NVT) || !TLI.isTypeLegal(Scalar.getValueType()))
10010 SDLoc dl = SDLoc(N);
10011 SDValue Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, NVT, Scalar);
10012 return DAG.getNode(ISD::BITCAST, dl, VT, Res);
10016 // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
10017 // nodes often generate nop CONCAT_VECTOR nodes.
10018 // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that
10019 // place the incoming vectors at the exact same location.
10020 SDValue SingleSource = SDValue();
10021 unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements();
10023 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
10024 SDValue Op = N->getOperand(i);
10026 if (Op.getOpcode() == ISD::UNDEF)
10029 // Check if this is the identity extract:
10030 if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
10033 // Find the single incoming vector for the extract_subvector.
10034 if (SingleSource.getNode()) {
10035 if (Op.getOperand(0) != SingleSource)
10038 SingleSource = Op.getOperand(0);
10040 // Check the source type is the same as the type of the result.
10041 // If not, this concat may extend the vector, so we can not
10042 // optimize it away.
10043 if (SingleSource.getValueType() != N->getValueType(0))
10047 unsigned IdentityIndex = i * PartNumElem;
10048 ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1));
10049 // The extract index must be constant.
10053 // Check that we are reading from the identity index.
10054 if (CS->getZExtValue() != IdentityIndex)
10058 if (SingleSource.getNode())
10059 return SingleSource;
10064 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
10065 EVT NVT = N->getValueType(0);
10066 SDValue V = N->getOperand(0);
10068 if (V->getOpcode() == ISD::CONCAT_VECTORS) {
10070 // (extract_subvec (concat V1, V2, ...), i)
10073 // Only operand 0 is checked as 'concat' assumes all inputs of the same
10075 if (V->getOperand(0).getValueType() != NVT)
10077 unsigned Idx = dyn_cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
10078 unsigned NumElems = NVT.getVectorNumElements();
10079 assert((Idx % NumElems) == 0 &&
10080 "IDX in concat is not a multiple of the result vector length.");
10081 return V->getOperand(Idx / NumElems);
10085 if (V->getOpcode() == ISD::BITCAST)
10086 V = V.getOperand(0);
10088 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
10090 // Handle only simple case where vector being inserted and vector
10091 // being extracted are of same type, and are half size of larger vectors.
10092 EVT BigVT = V->getOperand(0).getValueType();
10093 EVT SmallVT = V->getOperand(1).getValueType();
10094 if (!NVT.bitsEq(SmallVT) || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
10097 // Only handle cases where both indexes are constants with the same type.
10098 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
10099 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
10101 if (InsIdx && ExtIdx &&
10102 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
10103 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
10105 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
10107 // indices are equal or bit offsets are equal => V1
10108 // otherwise => (extract_subvec V1, ExtIdx)
10109 if (InsIdx->getZExtValue() * SmallVT.getScalarType().getSizeInBits() ==
10110 ExtIdx->getZExtValue() * NVT.getScalarType().getSizeInBits())
10111 return DAG.getNode(ISD::BITCAST, dl, NVT, V->getOperand(1));
10112 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, NVT,
10113 DAG.getNode(ISD::BITCAST, dl,
10114 N->getOperand(0).getValueType(),
10115 V->getOperand(0)), N->getOperand(1));
10122 // Tries to turn a shuffle of two CONCAT_VECTORS into a single concat.
10123 static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
10124 EVT VT = N->getValueType(0);
10125 unsigned NumElts = VT.getVectorNumElements();
10127 SDValue N0 = N->getOperand(0);
10128 SDValue N1 = N->getOperand(1);
10129 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
10131 SmallVector<SDValue, 4> Ops;
10132 EVT ConcatVT = N0.getOperand(0).getValueType();
10133 unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements();
10134 unsigned NumConcats = NumElts / NumElemsPerConcat;
10136 // Look at every vector that's inserted. We're looking for exact
10137 // subvector-sized copies from a concatenated vector
10138 for (unsigned I = 0; I != NumConcats; ++I) {
10139 // Make sure we're dealing with a copy.
10140 unsigned Begin = I * NumElemsPerConcat;
10141 bool AllUndef = true, NoUndef = true;
10142 for (unsigned J = Begin; J != Begin + NumElemsPerConcat; ++J) {
10143 if (SVN->getMaskElt(J) >= 0)
10150 if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0)
10153 for (unsigned J = 1; J != NumElemsPerConcat; ++J)
10154 if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J))
10157 unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat;
10158 if (FirstElt < N0.getNumOperands())
10159 Ops.push_back(N0.getOperand(FirstElt));
10161 Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands()));
10163 } else if (AllUndef) {
10164 Ops.push_back(DAG.getUNDEF(N0.getOperand(0).getValueType()));
10165 } else { // Mixed with general masks and undefs, can't do optimization.
10170 return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), VT, Ops.data(),
10174 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
10175 EVT VT = N->getValueType(0);
10176 unsigned NumElts = VT.getVectorNumElements();
10178 SDValue N0 = N->getOperand(0);
10179 SDValue N1 = N->getOperand(1);
10181 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
10183 // Canonicalize shuffle undef, undef -> undef
10184 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
10185 return DAG.getUNDEF(VT);
10187 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
10189 // Canonicalize shuffle v, v -> v, undef
10191 SmallVector<int, 8> NewMask;
10192 for (unsigned i = 0; i != NumElts; ++i) {
10193 int Idx = SVN->getMaskElt(i);
10194 if (Idx >= (int)NumElts) Idx -= NumElts;
10195 NewMask.push_back(Idx);
10197 return DAG.getVectorShuffle(VT, SDLoc(N), N0, DAG.getUNDEF(VT),
10201 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
10202 if (N0.getOpcode() == ISD::UNDEF) {
10203 SmallVector<int, 8> NewMask;
10204 for (unsigned i = 0; i != NumElts; ++i) {
10205 int Idx = SVN->getMaskElt(i);
10207 if (Idx >= (int)NumElts)
10210 Idx = -1; // remove reference to lhs
10212 NewMask.push_back(Idx);
10214 return DAG.getVectorShuffle(VT, SDLoc(N), N1, DAG.getUNDEF(VT),
10218 // Remove references to rhs if it is undef
10219 if (N1.getOpcode() == ISD::UNDEF) {
10220 bool Changed = false;
10221 SmallVector<int, 8> NewMask;
10222 for (unsigned i = 0; i != NumElts; ++i) {
10223 int Idx = SVN->getMaskElt(i);
10224 if (Idx >= (int)NumElts) {
10228 NewMask.push_back(Idx);
10231 return DAG.getVectorShuffle(VT, SDLoc(N), N0, N1, &NewMask[0]);
10234 // If it is a splat, check if the argument vector is another splat or a
10235 // build_vector with all scalar elements the same.
10236 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
10237 SDNode *V = N0.getNode();
10239 // If this is a bit convert that changes the element type of the vector but
10240 // not the number of vector elements, look through it. Be careful not to
10241 // look though conversions that change things like v4f32 to v2f64.
10242 if (V->getOpcode() == ISD::BITCAST) {
10243 SDValue ConvInput = V->getOperand(0);
10244 if (ConvInput.getValueType().isVector() &&
10245 ConvInput.getValueType().getVectorNumElements() == NumElts)
10246 V = ConvInput.getNode();
10249 if (V->getOpcode() == ISD::BUILD_VECTOR) {
10250 assert(V->getNumOperands() == NumElts &&
10251 "BUILD_VECTOR has wrong number of operands");
10253 bool AllSame = true;
10254 for (unsigned i = 0; i != NumElts; ++i) {
10255 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
10256 Base = V->getOperand(i);
10260 // Splat of <u, u, u, u>, return <u, u, u, u>
10261 if (!Base.getNode())
10263 for (unsigned i = 0; i != NumElts; ++i) {
10264 if (V->getOperand(i) != Base) {
10269 // Splat of <x, x, x, x>, return <x, x, x, x>
10275 if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
10276 Level < AfterLegalizeVectorOps &&
10277 (N1.getOpcode() == ISD::UNDEF ||
10278 (N1.getOpcode() == ISD::CONCAT_VECTORS &&
10279 N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) {
10280 SDValue V = partitionShuffleOfConcats(N, DAG);
10286 // If this shuffle node is simply a swizzle of another shuffle node,
10287 // and it reverses the swizzle of the previous shuffle then we can
10288 // optimize shuffle(shuffle(x, undef), undef) -> x.
10289 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
10290 N1.getOpcode() == ISD::UNDEF) {
10292 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
10294 // Shuffle nodes can only reverse shuffles with a single non-undef value.
10295 if (N0.getOperand(1).getOpcode() != ISD::UNDEF)
10298 // The incoming shuffle must be of the same type as the result of the
10299 // current shuffle.
10300 assert(OtherSV->getOperand(0).getValueType() == VT &&
10301 "Shuffle types don't match");
10303 for (unsigned i = 0; i != NumElts; ++i) {
10304 int Idx = SVN->getMaskElt(i);
10305 assert(Idx < (int)NumElts && "Index references undef operand");
10306 // Next, this index comes from the first value, which is the incoming
10307 // shuffle. Adopt the incoming index.
10309 Idx = OtherSV->getMaskElt(Idx);
10311 // The combined shuffle must map each index to itself.
10312 if (Idx >= 0 && (unsigned)Idx != i)
10316 return OtherSV->getOperand(0);
10322 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
10323 /// an AND to a vector_shuffle with the destination vector and a zero vector.
10324 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
10325 /// vector_shuffle V, Zero, <0, 4, 2, 4>
10326 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
10327 EVT VT = N->getValueType(0);
10329 SDValue LHS = N->getOperand(0);
10330 SDValue RHS = N->getOperand(1);
10331 if (N->getOpcode() == ISD::AND) {
10332 if (RHS.getOpcode() == ISD::BITCAST)
10333 RHS = RHS.getOperand(0);
10334 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
10335 SmallVector<int, 8> Indices;
10336 unsigned NumElts = RHS.getNumOperands();
10337 for (unsigned i = 0; i != NumElts; ++i) {
10338 SDValue Elt = RHS.getOperand(i);
10339 if (!isa<ConstantSDNode>(Elt))
10342 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
10343 Indices.push_back(i);
10344 else if (cast<ConstantSDNode>(Elt)->isNullValue())
10345 Indices.push_back(NumElts);
10350 // Let's see if the target supports this vector_shuffle.
10351 EVT RVT = RHS.getValueType();
10352 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
10355 // Return the new VECTOR_SHUFFLE node.
10356 EVT EltVT = RVT.getVectorElementType();
10357 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
10358 DAG.getConstant(0, EltVT));
10359 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
10360 RVT, &ZeroOps[0], ZeroOps.size());
10361 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
10362 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
10363 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
10370 /// SimplifyVBinOp - Visit a binary vector operation, like ADD.
10371 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
10372 assert(N->getValueType(0).isVector() &&
10373 "SimplifyVBinOp only works on vectors!");
10375 SDValue LHS = N->getOperand(0);
10376 SDValue RHS = N->getOperand(1);
10377 SDValue Shuffle = XformToShuffleWithZero(N);
10378 if (Shuffle.getNode()) return Shuffle;
10380 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
10382 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
10383 RHS.getOpcode() == ISD::BUILD_VECTOR) {
10384 SmallVector<SDValue, 8> Ops;
10385 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
10386 SDValue LHSOp = LHS.getOperand(i);
10387 SDValue RHSOp = RHS.getOperand(i);
10388 // If these two elements can't be folded, bail out.
10389 if ((LHSOp.getOpcode() != ISD::UNDEF &&
10390 LHSOp.getOpcode() != ISD::Constant &&
10391 LHSOp.getOpcode() != ISD::ConstantFP) ||
10392 (RHSOp.getOpcode() != ISD::UNDEF &&
10393 RHSOp.getOpcode() != ISD::Constant &&
10394 RHSOp.getOpcode() != ISD::ConstantFP))
10397 // Can't fold divide by zero.
10398 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
10399 N->getOpcode() == ISD::FDIV) {
10400 if ((RHSOp.getOpcode() == ISD::Constant &&
10401 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
10402 (RHSOp.getOpcode() == ISD::ConstantFP &&
10403 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
10407 EVT VT = LHSOp.getValueType();
10408 EVT RVT = RHSOp.getValueType();
10410 // Integer BUILD_VECTOR operands may have types larger than the element
10411 // size (e.g., when the element type is not legal). Prior to type
10412 // legalization, the types may not match between the two BUILD_VECTORS.
10413 // Truncate one of the operands to make them match.
10414 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
10415 RHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, RHSOp);
10417 LHSOp = DAG.getNode(ISD::TRUNCATE, SDLoc(N), RVT, LHSOp);
10421 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(LHS), VT,
10423 if (FoldOp.getOpcode() != ISD::UNDEF &&
10424 FoldOp.getOpcode() != ISD::Constant &&
10425 FoldOp.getOpcode() != ISD::ConstantFP)
10427 Ops.push_back(FoldOp);
10428 AddToWorkList(FoldOp.getNode());
10431 if (Ops.size() == LHS.getNumOperands())
10432 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
10433 LHS.getValueType(), &Ops[0], Ops.size());
10439 /// SimplifyVUnaryOp - Visit a binary vector operation, like FABS/FNEG.
10440 SDValue DAGCombiner::SimplifyVUnaryOp(SDNode *N) {
10441 assert(N->getValueType(0).isVector() &&
10442 "SimplifyVUnaryOp only works on vectors!");
10444 SDValue N0 = N->getOperand(0);
10446 if (N0.getOpcode() != ISD::BUILD_VECTOR)
10449 // Operand is a BUILD_VECTOR node, see if we can constant fold it.
10450 SmallVector<SDValue, 8> Ops;
10451 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
10452 SDValue Op = N0.getOperand(i);
10453 if (Op.getOpcode() != ISD::UNDEF &&
10454 Op.getOpcode() != ISD::ConstantFP)
10456 EVT EltVT = Op.getValueType();
10457 SDValue FoldOp = DAG.getNode(N->getOpcode(), SDLoc(N0), EltVT, Op);
10458 if (FoldOp.getOpcode() != ISD::UNDEF &&
10459 FoldOp.getOpcode() != ISD::ConstantFP)
10461 Ops.push_back(FoldOp);
10462 AddToWorkList(FoldOp.getNode());
10465 if (Ops.size() != N0.getNumOperands())
10468 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(N),
10469 N0.getValueType(), &Ops[0], Ops.size());
10472 SDValue DAGCombiner::SimplifySelect(SDLoc DL, SDValue N0,
10473 SDValue N1, SDValue N2){
10474 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
10476 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
10477 cast<CondCodeSDNode>(N0.getOperand(2))->get());
10479 // If we got a simplified select_cc node back from SimplifySelectCC, then
10480 // break it down into a new SETCC node, and a new SELECT node, and then return
10481 // the SELECT node, since we were called with a SELECT node.
10482 if (SCC.getNode()) {
10483 // Check to see if we got a select_cc back (to turn into setcc/select).
10484 // Otherwise, just return whatever node we got back, like fabs.
10485 if (SCC.getOpcode() == ISD::SELECT_CC) {
10486 SDValue SETCC = DAG.getNode(ISD::SETCC, SDLoc(N0),
10488 SCC.getOperand(0), SCC.getOperand(1),
10489 SCC.getOperand(4));
10490 AddToWorkList(SETCC.getNode());
10491 return DAG.getSelect(SDLoc(SCC), SCC.getValueType(),
10492 SCC.getOperand(2), SCC.getOperand(3), SETCC);
10500 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
10501 /// are the two values being selected between, see if we can simplify the
10502 /// select. Callers of this should assume that TheSelect is deleted if this
10503 /// returns true. As such, they should return the appropriate thing (e.g. the
10504 /// node) back to the top-level of the DAG combiner loop to avoid it being
10506 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
10509 // Cannot simplify select with vector condition
10510 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
10512 // If this is a select from two identical things, try to pull the operation
10513 // through the select.
10514 if (LHS.getOpcode() != RHS.getOpcode() ||
10515 !LHS.hasOneUse() || !RHS.hasOneUse())
10518 // If this is a load and the token chain is identical, replace the select
10519 // of two loads with a load through a select of the address to load from.
10520 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
10521 // constants have been dropped into the constant pool.
10522 if (LHS.getOpcode() == ISD::LOAD) {
10523 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
10524 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
10526 // Token chains must be identical.
10527 if (LHS.getOperand(0) != RHS.getOperand(0) ||
10528 // Do not let this transformation reduce the number of volatile loads.
10529 LLD->isVolatile() || RLD->isVolatile() ||
10530 // If this is an EXTLOAD, the VT's must match.
10531 LLD->getMemoryVT() != RLD->getMemoryVT() ||
10532 // If this is an EXTLOAD, the kind of extension must match.
10533 (LLD->getExtensionType() != RLD->getExtensionType() &&
10534 // The only exception is if one of the extensions is anyext.
10535 LLD->getExtensionType() != ISD::EXTLOAD &&
10536 RLD->getExtensionType() != ISD::EXTLOAD) ||
10537 // FIXME: this discards src value information. This is
10538 // over-conservative. It would be beneficial to be able to remember
10539 // both potential memory locations. Since we are discarding
10540 // src value info, don't do the transformation if the memory
10541 // locations are not in the default address space.
10542 LLD->getPointerInfo().getAddrSpace() != 0 ||
10543 RLD->getPointerInfo().getAddrSpace() != 0 ||
10544 !TLI.isOperationLegalOrCustom(TheSelect->getOpcode(),
10545 LLD->getBasePtr().getValueType()))
10548 // Check that the select condition doesn't reach either load. If so,
10549 // folding this will induce a cycle into the DAG. If not, this is safe to
10550 // xform, so create a select of the addresses.
10552 if (TheSelect->getOpcode() == ISD::SELECT) {
10553 SDNode *CondNode = TheSelect->getOperand(0).getNode();
10554 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
10555 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
10557 // The loads must not depend on one another.
10558 if (LLD->isPredecessorOf(RLD) ||
10559 RLD->isPredecessorOf(LLD))
10561 Addr = DAG.getSelect(SDLoc(TheSelect),
10562 LLD->getBasePtr().getValueType(),
10563 TheSelect->getOperand(0), LLD->getBasePtr(),
10564 RLD->getBasePtr());
10565 } else { // Otherwise SELECT_CC
10566 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
10567 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
10569 if ((LLD->hasAnyUseOfValue(1) &&
10570 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
10571 (RLD->hasAnyUseOfValue(1) &&
10572 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
10575 Addr = DAG.getNode(ISD::SELECT_CC, SDLoc(TheSelect),
10576 LLD->getBasePtr().getValueType(),
10577 TheSelect->getOperand(0),
10578 TheSelect->getOperand(1),
10579 LLD->getBasePtr(), RLD->getBasePtr(),
10580 TheSelect->getOperand(4));
10584 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
10585 Load = DAG.getLoad(TheSelect->getValueType(0),
10587 // FIXME: Discards pointer and TBAA info.
10588 LLD->getChain(), Addr, MachinePointerInfo(),
10589 LLD->isVolatile(), LLD->isNonTemporal(),
10590 LLD->isInvariant(), LLD->getAlignment());
10592 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
10593 RLD->getExtensionType() : LLD->getExtensionType(),
10595 TheSelect->getValueType(0),
10596 // FIXME: Discards pointer and TBAA info.
10597 LLD->getChain(), Addr, MachinePointerInfo(),
10598 LLD->getMemoryVT(), LLD->isVolatile(),
10599 LLD->isNonTemporal(), LLD->getAlignment());
10602 // Users of the select now use the result of the load.
10603 CombineTo(TheSelect, Load);
10605 // Users of the old loads now use the new load's chain. We know the
10606 // old-load value is dead now.
10607 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
10608 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
10615 /// SimplifySelectCC - Simplify an expression of the form (N0 cond N1) ? N2 : N3
10616 /// where 'cond' is the comparison specified by CC.
10617 SDValue DAGCombiner::SimplifySelectCC(SDLoc DL, SDValue N0, SDValue N1,
10618 SDValue N2, SDValue N3,
10619 ISD::CondCode CC, bool NotExtCompare) {
10620 // (x ? y : y) -> y.
10621 if (N2 == N3) return N2;
10623 EVT VT = N2.getValueType();
10624 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
10625 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
10626 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
10628 // Determine if the condition we're dealing with is constant
10629 SDValue SCC = SimplifySetCC(getSetCCResultType(N0.getValueType()),
10630 N0, N1, CC, DL, false);
10631 if (SCC.getNode()) AddToWorkList(SCC.getNode());
10632 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
10634 // fold select_cc true, x, y -> x
10635 if (SCCC && !SCCC->isNullValue())
10637 // fold select_cc false, x, y -> y
10638 if (SCCC && SCCC->isNullValue())
10641 // Check to see if we can simplify the select into an fabs node
10642 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
10643 // Allow either -0.0 or 0.0
10644 if (CFP->getValueAPF().isZero()) {
10645 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
10646 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
10647 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
10648 N2 == N3.getOperand(0))
10649 return DAG.getNode(ISD::FABS, DL, VT, N0);
10651 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
10652 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
10653 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
10654 N2.getOperand(0) == N3)
10655 return DAG.getNode(ISD::FABS, DL, VT, N3);
10659 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
10660 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
10661 // in it. This is a win when the constant is not otherwise available because
10662 // it replaces two constant pool loads with one. We only do this if the FP
10663 // type is known to be legal, because if it isn't, then we are before legalize
10664 // types an we want the other legalization to happen first (e.g. to avoid
10665 // messing with soft float) and if the ConstantFP is not legal, because if
10666 // it is legal, we may not need to store the FP constant in a constant pool.
10667 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
10668 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
10669 if (TLI.isTypeLegal(N2.getValueType()) &&
10670 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
10671 TargetLowering::Legal) &&
10672 // If both constants have multiple uses, then we won't need to do an
10673 // extra load, they are likely around in registers for other users.
10674 (TV->hasOneUse() || FV->hasOneUse())) {
10675 Constant *Elts[] = {
10676 const_cast<ConstantFP*>(FV->getConstantFPValue()),
10677 const_cast<ConstantFP*>(TV->getConstantFPValue())
10679 Type *FPTy = Elts[0]->getType();
10680 const DataLayout &TD = *TLI.getDataLayout();
10682 // Create a ConstantArray of the two constants.
10683 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
10684 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
10685 TD.getPrefTypeAlignment(FPTy));
10686 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
10688 // Get the offsets to the 0 and 1 element of the array so that we can
10689 // select between them.
10690 SDValue Zero = DAG.getIntPtrConstant(0);
10691 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
10692 SDValue One = DAG.getIntPtrConstant(EltSize);
10694 SDValue Cond = DAG.getSetCC(DL,
10695 getSetCCResultType(N0.getValueType()),
10697 AddToWorkList(Cond.getNode());
10698 SDValue CstOffset = DAG.getSelect(DL, Zero.getValueType(),
10700 AddToWorkList(CstOffset.getNode());
10701 CPIdx = DAG.getNode(ISD::ADD, DL, CPIdx.getValueType(), CPIdx,
10703 AddToWorkList(CPIdx.getNode());
10704 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
10705 MachinePointerInfo::getConstantPool(), false,
10706 false, false, Alignment);
10711 // Check to see if we can perform the "gzip trick", transforming
10712 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
10713 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
10714 (N1C->isNullValue() || // (a < 0) ? b : 0
10715 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
10716 EVT XType = N0.getValueType();
10717 EVT AType = N2.getValueType();
10718 if (XType.bitsGE(AType)) {
10719 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
10720 // single-bit constant.
10721 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
10722 unsigned ShCtV = N2C->getAPIntValue().logBase2();
10723 ShCtV = XType.getSizeInBits()-ShCtV-1;
10724 SDValue ShCt = DAG.getConstant(ShCtV,
10725 getShiftAmountTy(N0.getValueType()));
10726 SDValue Shift = DAG.getNode(ISD::SRL, SDLoc(N0),
10728 AddToWorkList(Shift.getNode());
10730 if (XType.bitsGT(AType)) {
10731 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
10732 AddToWorkList(Shift.getNode());
10735 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
10738 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0),
10740 DAG.getConstant(XType.getSizeInBits()-1,
10741 getShiftAmountTy(N0.getValueType())));
10742 AddToWorkList(Shift.getNode());
10744 if (XType.bitsGT(AType)) {
10745 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
10746 AddToWorkList(Shift.getNode());
10749 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
10753 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
10754 // where y is has a single bit set.
10755 // A plaintext description would be, we can turn the SELECT_CC into an AND
10756 // when the condition can be materialized as an all-ones register. Any
10757 // single bit-test can be materialized as an all-ones register with
10758 // shift-left and shift-right-arith.
10759 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
10760 N0->getValueType(0) == VT &&
10761 N1C && N1C->isNullValue() &&
10762 N2C && N2C->isNullValue()) {
10763 SDValue AndLHS = N0->getOperand(0);
10764 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
10765 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
10766 // Shift the tested bit over the sign bit.
10767 APInt AndMask = ConstAndRHS->getAPIntValue();
10769 DAG.getConstant(AndMask.countLeadingZeros(),
10770 getShiftAmountTy(AndLHS.getValueType()));
10771 SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(N0), VT, AndLHS, ShlAmt);
10773 // Now arithmetic right shift it all the way over, so the result is either
10774 // all-ones, or zero.
10776 DAG.getConstant(AndMask.getBitWidth()-1,
10777 getShiftAmountTy(Shl.getValueType()));
10778 SDValue Shr = DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl, ShrAmt);
10780 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
10784 // fold select C, 16, 0 -> shl C, 4
10785 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
10786 TLI.getBooleanContents(N0.getValueType().isVector()) ==
10787 TargetLowering::ZeroOrOneBooleanContent) {
10789 // If the caller doesn't want us to simplify this into a zext of a compare,
10791 if (NotExtCompare && N2C->getAPIntValue() == 1)
10794 // Get a SetCC of the condition
10795 // NOTE: Don't create a SETCC if it's not legal on this target.
10796 if (!LegalOperations ||
10797 TLI.isOperationLegal(ISD::SETCC,
10798 LegalTypes ? getSetCCResultType(N0.getValueType()) : MVT::i1)) {
10800 // cast from setcc result type to select result type
10802 SCC = DAG.getSetCC(DL, getSetCCResultType(N0.getValueType()),
10804 if (N2.getValueType().bitsLT(SCC.getValueType()))
10805 Temp = DAG.getZeroExtendInReg(SCC, SDLoc(N2),
10806 N2.getValueType());
10808 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
10809 N2.getValueType(), SCC);
10811 SCC = DAG.getSetCC(SDLoc(N0), MVT::i1, N0, N1, CC);
10812 Temp = DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N2),
10813 N2.getValueType(), SCC);
10816 AddToWorkList(SCC.getNode());
10817 AddToWorkList(Temp.getNode());
10819 if (N2C->getAPIntValue() == 1)
10822 // shl setcc result by log2 n2c
10823 return DAG.getNode(
10824 ISD::SHL, DL, N2.getValueType(), Temp,
10825 DAG.getConstant(N2C->getAPIntValue().logBase2(),
10826 getShiftAmountTy(Temp.getValueType())));
10830 // Check to see if this is the equivalent of setcc
10831 // FIXME: Turn all of these into setcc if setcc if setcc is legal
10832 // otherwise, go ahead with the folds.
10833 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
10834 EVT XType = N0.getValueType();
10835 if (!LegalOperations ||
10836 TLI.isOperationLegal(ISD::SETCC, getSetCCResultType(XType))) {
10837 SDValue Res = DAG.getSetCC(DL, getSetCCResultType(XType), N0, N1, CC);
10838 if (Res.getValueType() != VT)
10839 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
10843 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
10844 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
10845 (!LegalOperations ||
10846 TLI.isOperationLegal(ISD::CTLZ, XType))) {
10847 SDValue Ctlz = DAG.getNode(ISD::CTLZ, SDLoc(N0), XType, N0);
10848 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
10849 DAG.getConstant(Log2_32(XType.getSizeInBits()),
10850 getShiftAmountTy(Ctlz.getValueType())));
10852 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
10853 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
10854 SDValue NegN0 = DAG.getNode(ISD::SUB, SDLoc(N0),
10855 XType, DAG.getConstant(0, XType), N0);
10856 SDValue NotN0 = DAG.getNOT(SDLoc(N0), N0, XType);
10857 return DAG.getNode(ISD::SRL, DL, XType,
10858 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
10859 DAG.getConstant(XType.getSizeInBits()-1,
10860 getShiftAmountTy(XType)));
10862 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
10863 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
10864 SDValue Sign = DAG.getNode(ISD::SRL, SDLoc(N0), XType, N0,
10865 DAG.getConstant(XType.getSizeInBits()-1,
10866 getShiftAmountTy(N0.getValueType())));
10867 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
10871 // Check to see if this is an integer abs.
10872 // select_cc setg[te] X, 0, X, -X ->
10873 // select_cc setgt X, -1, X, -X ->
10874 // select_cc setl[te] X, 0, -X, X ->
10875 // select_cc setlt X, 1, -X, X ->
10876 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
10878 ConstantSDNode *SubC = NULL;
10879 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
10880 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
10881 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
10882 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
10883 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
10884 (N1C->isOne() && CC == ISD::SETLT)) &&
10885 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
10886 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
10888 EVT XType = N0.getValueType();
10889 if (SubC && SubC->isNullValue() && XType.isInteger()) {
10890 SDValue Shift = DAG.getNode(ISD::SRA, SDLoc(N0), XType,
10892 DAG.getConstant(XType.getSizeInBits()-1,
10893 getShiftAmountTy(N0.getValueType())));
10894 SDValue Add = DAG.getNode(ISD::ADD, SDLoc(N0),
10896 AddToWorkList(Shift.getNode());
10897 AddToWorkList(Add.getNode());
10898 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
10905 /// SimplifySetCC - This is a stub for TargetLowering::SimplifySetCC.
10906 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
10907 SDValue N1, ISD::CondCode Cond,
10908 SDLoc DL, bool foldBooleans) {
10909 TargetLowering::DAGCombinerInfo
10910 DagCombineInfo(DAG, Level, false, this);
10911 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
10914 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
10915 /// return a DAG expression to select that will generate the same value by
10916 /// multiplying by a magic number. See:
10917 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
10918 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
10919 std::vector<SDNode*> Built;
10920 SDValue S = TLI.BuildSDIV(N, DAG, LegalOperations, &Built);
10922 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
10924 AddToWorkList(*ii);
10928 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
10929 /// return a DAG expression to select that will generate the same value by
10930 /// multiplying by a magic number. See:
10931 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
10932 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
10933 std::vector<SDNode*> Built;
10934 SDValue S = TLI.BuildUDIV(N, DAG, LegalOperations, &Built);
10936 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
10938 AddToWorkList(*ii);
10942 /// FindBaseOffset - Return true if base is a frame index, which is known not
10943 // to alias with anything but itself. Provides base object and offset as
10945 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
10946 const GlobalValue *&GV, const void *&CV) {
10947 // Assume it is a primitive operation.
10948 Base = Ptr; Offset = 0; GV = 0; CV = 0;
10950 // If it's an adding a simple constant then integrate the offset.
10951 if (Base.getOpcode() == ISD::ADD) {
10952 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
10953 Base = Base.getOperand(0);
10954 Offset += C->getZExtValue();
10958 // Return the underlying GlobalValue, and update the Offset. Return false
10959 // for GlobalAddressSDNode since the same GlobalAddress may be represented
10960 // by multiple nodes with different offsets.
10961 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
10962 GV = G->getGlobal();
10963 Offset += G->getOffset();
10967 // Return the underlying Constant value, and update the Offset. Return false
10968 // for ConstantSDNodes since the same constant pool entry may be represented
10969 // by multiple nodes with different offsets.
10970 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
10971 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
10972 : (const void *)C->getConstVal();
10973 Offset += C->getOffset();
10976 // If it's any of the following then it can't alias with anything but itself.
10977 return isa<FrameIndexSDNode>(Base);
10980 /// isAlias - Return true if there is any possibility that the two addresses
10982 bool DAGCombiner::isAlias(SDValue Ptr1, int64_t Size1, bool IsVolatile1,
10983 const Value *SrcValue1, int SrcValueOffset1,
10984 unsigned SrcValueAlign1,
10985 const MDNode *TBAAInfo1,
10986 SDValue Ptr2, int64_t Size2, bool IsVolatile2,
10987 const Value *SrcValue2, int SrcValueOffset2,
10988 unsigned SrcValueAlign2,
10989 const MDNode *TBAAInfo2) const {
10990 // If they are the same then they must be aliases.
10991 if (Ptr1 == Ptr2) return true;
10993 // If they are both volatile then they cannot be reordered.
10994 if (IsVolatile1 && IsVolatile2) return true;
10996 // Gather base node and offset information.
10997 SDValue Base1, Base2;
10998 int64_t Offset1, Offset2;
10999 const GlobalValue *GV1, *GV2;
11000 const void *CV1, *CV2;
11001 bool isFrameIndex1 = FindBaseOffset(Ptr1, Base1, Offset1, GV1, CV1);
11002 bool isFrameIndex2 = FindBaseOffset(Ptr2, Base2, Offset2, GV2, CV2);
11004 // If they have a same base address then check to see if they overlap.
11005 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
11006 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
11008 // It is possible for different frame indices to alias each other, mostly
11009 // when tail call optimization reuses return address slots for arguments.
11010 // To catch this case, look up the actual index of frame indices to compute
11011 // the real alias relationship.
11012 if (isFrameIndex1 && isFrameIndex2) {
11013 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11014 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
11015 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
11016 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
11019 // Otherwise, if we know what the bases are, and they aren't identical, then
11020 // we know they cannot alias.
11021 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
11024 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
11025 // compared to the size and offset of the access, we may be able to prove they
11026 // do not alias. This check is conservative for now to catch cases created by
11027 // splitting vector types.
11028 if ((SrcValueAlign1 == SrcValueAlign2) &&
11029 (SrcValueOffset1 != SrcValueOffset2) &&
11030 (Size1 == Size2) && (SrcValueAlign1 > Size1)) {
11031 int64_t OffAlign1 = SrcValueOffset1 % SrcValueAlign1;
11032 int64_t OffAlign2 = SrcValueOffset2 % SrcValueAlign1;
11034 // There is no overlap between these relatively aligned accesses of similar
11035 // size, return no alias.
11036 if ((OffAlign1 + Size1) <= OffAlign2 || (OffAlign2 + Size2) <= OffAlign1)
11040 bool UseAA = CombinerGlobalAA.getNumOccurrences() > 0 ? CombinerGlobalAA :
11041 TLI.getTargetMachine().getSubtarget<TargetSubtargetInfo>().useAA();
11042 if (UseAA && SrcValue1 && SrcValue2) {
11043 // Use alias analysis information.
11044 int64_t MinOffset = std::min(SrcValueOffset1, SrcValueOffset2);
11045 int64_t Overlap1 = Size1 + SrcValueOffset1 - MinOffset;
11046 int64_t Overlap2 = Size2 + SrcValueOffset2 - MinOffset;
11047 AliasAnalysis::AliasResult AAResult =
11048 AA.alias(AliasAnalysis::Location(SrcValue1, Overlap1, TBAAInfo1),
11049 AliasAnalysis::Location(SrcValue2, Overlap2, TBAAInfo2));
11050 if (AAResult == AliasAnalysis::NoAlias)
11054 // Otherwise we have to assume they alias.
11058 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) {
11059 SDValue Ptr0, Ptr1;
11060 int64_t Size0, Size1;
11061 bool IsVolatile0, IsVolatile1;
11062 const Value *SrcValue0, *SrcValue1;
11063 int SrcValueOffset0, SrcValueOffset1;
11064 unsigned SrcValueAlign0, SrcValueAlign1;
11065 const MDNode *SrcTBAAInfo0, *SrcTBAAInfo1;
11066 FindAliasInfo(Op0, Ptr0, Size0, IsVolatile0, SrcValue0, SrcValueOffset0,
11067 SrcValueAlign0, SrcTBAAInfo0);
11068 FindAliasInfo(Op1, Ptr1, Size1, IsVolatile1, SrcValue1, SrcValueOffset1,
11069 SrcValueAlign1, SrcTBAAInfo1);
11070 return isAlias(Ptr0, Size0, IsVolatile0, SrcValue0, SrcValueOffset0,
11071 SrcValueAlign0, SrcTBAAInfo0,
11072 Ptr1, Size1, IsVolatile1, SrcValue1, SrcValueOffset1,
11073 SrcValueAlign1, SrcTBAAInfo1);
11076 /// FindAliasInfo - Extracts the relevant alias information from the memory
11077 /// node. Returns true if the operand was a nonvolatile load.
11078 bool DAGCombiner::FindAliasInfo(SDNode *N,
11079 SDValue &Ptr, int64_t &Size, bool &IsVolatile,
11080 const Value *&SrcValue,
11081 int &SrcValueOffset,
11082 unsigned &SrcValueAlign,
11083 const MDNode *&TBAAInfo) const {
11084 LSBaseSDNode *LS = cast<LSBaseSDNode>(N);
11086 Ptr = LS->getBasePtr();
11087 Size = LS->getMemoryVT().getSizeInBits() >> 3;
11088 IsVolatile = LS->isVolatile();
11089 SrcValue = LS->getSrcValue();
11090 SrcValueOffset = LS->getSrcValueOffset();
11091 SrcValueAlign = LS->getOriginalAlignment();
11092 TBAAInfo = LS->getTBAAInfo();
11093 return isa<LoadSDNode>(LS) && !IsVolatile;
11096 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
11097 /// looking for aliasing nodes and adding them to the Aliases vector.
11098 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
11099 SmallVectorImpl<SDValue> &Aliases) {
11100 SmallVector<SDValue, 8> Chains; // List of chains to visit.
11101 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
11103 // Get alias information for node.
11107 const Value *SrcValue;
11108 int SrcValueOffset;
11109 unsigned SrcValueAlign;
11110 const MDNode *SrcTBAAInfo;
11111 bool IsLoad = FindAliasInfo(N, Ptr, Size, IsVolatile, SrcValue,
11112 SrcValueOffset, SrcValueAlign, SrcTBAAInfo);
11115 Chains.push_back(OriginalChain);
11116 unsigned Depth = 0;
11118 // Look at each chain and determine if it is an alias. If so, add it to the
11119 // aliases list. If not, then continue up the chain looking for the next
11121 while (!Chains.empty()) {
11122 SDValue Chain = Chains.back();
11125 // For TokenFactor nodes, look at each operand and only continue up the
11126 // chain until we find two aliases. If we've seen two aliases, assume we'll
11127 // find more and revert to original chain since the xform is unlikely to be
11130 // FIXME: The depth check could be made to return the last non-aliasing
11131 // chain we found before we hit a tokenfactor rather than the original
11133 if (Depth > 6 || Aliases.size() == 2) {
11135 Aliases.push_back(OriginalChain);
11139 // Don't bother if we've been before.
11140 if (!Visited.insert(Chain.getNode()))
11143 switch (Chain.getOpcode()) {
11144 case ISD::EntryToken:
11145 // Entry token is ideal chain operand, but handled in FindBetterChain.
11150 // Get alias information for Chain.
11154 const Value *OpSrcValue;
11155 int OpSrcValueOffset;
11156 unsigned OpSrcValueAlign;
11157 const MDNode *OpSrcTBAAInfo;
11158 bool IsOpLoad = FindAliasInfo(Chain.getNode(), OpPtr, OpSize,
11159 OpIsVolatile, OpSrcValue, OpSrcValueOffset,
11163 // If chain is alias then stop here.
11164 if (!(IsLoad && IsOpLoad) &&
11165 isAlias(Ptr, Size, IsVolatile, SrcValue, SrcValueOffset,
11166 SrcValueAlign, SrcTBAAInfo,
11167 OpPtr, OpSize, OpIsVolatile, OpSrcValue, OpSrcValueOffset,
11168 OpSrcValueAlign, OpSrcTBAAInfo)) {
11169 Aliases.push_back(Chain);
11171 // Look further up the chain.
11172 Chains.push_back(Chain.getOperand(0));
11178 case ISD::TokenFactor:
11179 // We have to check each of the operands of the token factor for "small"
11180 // token factors, so we queue them up. Adding the operands to the queue
11181 // (stack) in reverse order maintains the original order and increases the
11182 // likelihood that getNode will find a matching token factor (CSE.)
11183 if (Chain.getNumOperands() > 16) {
11184 Aliases.push_back(Chain);
11187 for (unsigned n = Chain.getNumOperands(); n;)
11188 Chains.push_back(Chain.getOperand(--n));
11193 // For all other instructions we will just have to take what we can get.
11194 Aliases.push_back(Chain);
11200 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
11201 /// for a better chain (aliasing node.)
11202 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
11203 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
11205 // Accumulate all the aliases to this node.
11206 GatherAllAliases(N, OldChain, Aliases);
11208 // If no operands then chain to entry token.
11209 if (Aliases.size() == 0)
11210 return DAG.getEntryNode();
11212 // If a single operand then chain to it. We don't need to revisit it.
11213 if (Aliases.size() == 1)
11216 // Construct a custom tailored token factor.
11217 return DAG.getNode(ISD::TokenFactor, SDLoc(N), MVT::Other,
11218 &Aliases[0], Aliases.size());
11221 // SelectionDAG::Combine - This is the entry point for the file.
11223 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
11224 CodeGenOpt::Level OptLevel) {
11225 /// run - This is the main entry point to this class.
11227 DAGCombiner(*this, AA, OptLevel).Run(Level);