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/DataLayout.h"
27 #include "llvm/DerivedTypes.h"
28 #include "llvm/LLVMContext.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/ErrorHandling.h"
32 #include "llvm/Support/MathExtras.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetLowering.h"
35 #include "llvm/Target/TargetMachine.h"
36 #include "llvm/Target/TargetOptions.h"
40 STATISTIC(NodesCombined , "Number of dag nodes combined");
41 STATISTIC(PreIndexedNodes , "Number of pre-indexed nodes created");
42 STATISTIC(PostIndexedNodes, "Number of post-indexed nodes created");
43 STATISTIC(OpsNarrowed , "Number of load/op/store narrowed");
44 STATISTIC(LdStFP2Int , "Number of fp load/store pairs transformed to int");
48 CombinerAA("combiner-alias-analysis", cl::Hidden,
49 cl::desc("Turn on alias analysis during testing"));
52 CombinerGlobalAA("combiner-global-alias-analysis", cl::Hidden,
53 cl::desc("Include global information in alias analysis"));
55 //------------------------------ DAGCombiner ---------------------------------//
59 const TargetLowering &TLI;
61 CodeGenOpt::Level OptLevel;
65 // Worklist of all of the nodes that need to be simplified.
67 // This has the semantics that when adding to the worklist,
68 // the item added must be next to be processed. It should
69 // also only appear once. The naive approach to this takes
72 // To reduce the insert/remove time to logarithmic, we use
73 // a set and a vector to maintain our worklist.
75 // The set contains the items on the worklist, but does not
76 // maintain the order they should be visited.
78 // The vector maintains the order nodes should be visited, but may
79 // contain duplicate or removed nodes. When choosing a node to
80 // visit, we pop off the order stack until we find an item that is
81 // also in the contents set. All operations are O(log N).
82 SmallPtrSet<SDNode*, 64> WorkListContents;
83 SmallVector<SDNode*, 64> WorkListOrder;
85 // AA - Used for DAG load/store alias analysis.
88 /// AddUsersToWorkList - When an instruction is simplified, add all users of
89 /// the instruction to the work lists because they might get more simplified
92 void AddUsersToWorkList(SDNode *N) {
93 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
98 /// visit - call the node-specific routine that knows how to fold each
99 /// particular type of node.
100 SDValue visit(SDNode *N);
103 /// AddToWorkList - Add to the work list making sure its instance is at the
104 /// back (next to be processed.)
105 void AddToWorkList(SDNode *N) {
106 WorkListContents.insert(N);
107 WorkListOrder.push_back(N);
110 /// removeFromWorkList - remove all instances of N from the worklist.
112 void removeFromWorkList(SDNode *N) {
113 WorkListContents.erase(N);
116 SDValue CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
119 SDValue CombineTo(SDNode *N, SDValue Res, bool AddTo = true) {
120 return CombineTo(N, &Res, 1, AddTo);
123 SDValue CombineTo(SDNode *N, SDValue Res0, SDValue Res1,
125 SDValue To[] = { Res0, Res1 };
126 return CombineTo(N, To, 2, AddTo);
129 void CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO);
133 /// SimplifyDemandedBits - Check the specified integer node value to see if
134 /// it can be simplified or if things it uses can be simplified by bit
135 /// propagation. If so, return true.
136 bool SimplifyDemandedBits(SDValue Op) {
137 unsigned BitWidth = Op.getValueType().getScalarType().getSizeInBits();
138 APInt Demanded = APInt::getAllOnesValue(BitWidth);
139 return SimplifyDemandedBits(Op, Demanded);
142 bool SimplifyDemandedBits(SDValue Op, const APInt &Demanded);
144 bool CombineToPreIndexedLoadStore(SDNode *N);
145 bool CombineToPostIndexedLoadStore(SDNode *N);
147 void ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad);
148 SDValue PromoteOperand(SDValue Op, EVT PVT, bool &Replace);
149 SDValue SExtPromoteOperand(SDValue Op, EVT PVT);
150 SDValue ZExtPromoteOperand(SDValue Op, EVT PVT);
151 SDValue PromoteIntBinOp(SDValue Op);
152 SDValue PromoteIntShiftOp(SDValue Op);
153 SDValue PromoteExtend(SDValue Op);
154 bool PromoteLoad(SDValue Op);
156 void ExtendSetCCUses(SmallVector<SDNode*, 4> SetCCs,
157 SDValue Trunc, SDValue ExtLoad, DebugLoc DL,
158 ISD::NodeType ExtType);
160 /// combine - call the node-specific routine that knows how to fold each
161 /// particular type of node. If that doesn't do anything, try the
162 /// target-specific DAG combines.
163 SDValue combine(SDNode *N);
165 // Visitation implementation - Implement dag node combining for different
166 // node types. The semantics are as follows:
168 // SDValue.getNode() == 0 - No change was made
169 // SDValue.getNode() == N - N was replaced, is dead and has been handled.
170 // otherwise - N should be replaced by the returned Operand.
172 SDValue visitTokenFactor(SDNode *N);
173 SDValue visitMERGE_VALUES(SDNode *N);
174 SDValue visitADD(SDNode *N);
175 SDValue visitSUB(SDNode *N);
176 SDValue visitADDC(SDNode *N);
177 SDValue visitSUBC(SDNode *N);
178 SDValue visitADDE(SDNode *N);
179 SDValue visitSUBE(SDNode *N);
180 SDValue visitMUL(SDNode *N);
181 SDValue visitSDIV(SDNode *N);
182 SDValue visitUDIV(SDNode *N);
183 SDValue visitSREM(SDNode *N);
184 SDValue visitUREM(SDNode *N);
185 SDValue visitMULHU(SDNode *N);
186 SDValue visitMULHS(SDNode *N);
187 SDValue visitSMUL_LOHI(SDNode *N);
188 SDValue visitUMUL_LOHI(SDNode *N);
189 SDValue visitSMULO(SDNode *N);
190 SDValue visitUMULO(SDNode *N);
191 SDValue visitSDIVREM(SDNode *N);
192 SDValue visitUDIVREM(SDNode *N);
193 SDValue visitAND(SDNode *N);
194 SDValue visitOR(SDNode *N);
195 SDValue visitXOR(SDNode *N);
196 SDValue SimplifyVBinOp(SDNode *N);
197 SDValue SimplifyVUnaryOp(SDNode *N);
198 SDValue visitSHL(SDNode *N);
199 SDValue visitSRA(SDNode *N);
200 SDValue visitSRL(SDNode *N);
201 SDValue visitCTLZ(SDNode *N);
202 SDValue visitCTLZ_ZERO_UNDEF(SDNode *N);
203 SDValue visitCTTZ(SDNode *N);
204 SDValue visitCTTZ_ZERO_UNDEF(SDNode *N);
205 SDValue visitCTPOP(SDNode *N);
206 SDValue visitSELECT(SDNode *N);
207 SDValue visitSELECT_CC(SDNode *N);
208 SDValue visitSETCC(SDNode *N);
209 SDValue visitSIGN_EXTEND(SDNode *N);
210 SDValue visitZERO_EXTEND(SDNode *N);
211 SDValue visitANY_EXTEND(SDNode *N);
212 SDValue visitSIGN_EXTEND_INREG(SDNode *N);
213 SDValue visitTRUNCATE(SDNode *N);
214 SDValue visitBITCAST(SDNode *N);
215 SDValue visitBUILD_PAIR(SDNode *N);
216 SDValue visitFADD(SDNode *N);
217 SDValue visitFSUB(SDNode *N);
218 SDValue visitFMUL(SDNode *N);
219 SDValue visitFMA(SDNode *N);
220 SDValue visitFDIV(SDNode *N);
221 SDValue visitFREM(SDNode *N);
222 SDValue visitFCOPYSIGN(SDNode *N);
223 SDValue visitSINT_TO_FP(SDNode *N);
224 SDValue visitUINT_TO_FP(SDNode *N);
225 SDValue visitFP_TO_SINT(SDNode *N);
226 SDValue visitFP_TO_UINT(SDNode *N);
227 SDValue visitFP_ROUND(SDNode *N);
228 SDValue visitFP_ROUND_INREG(SDNode *N);
229 SDValue visitFP_EXTEND(SDNode *N);
230 SDValue visitFNEG(SDNode *N);
231 SDValue visitFABS(SDNode *N);
232 SDValue visitFCEIL(SDNode *N);
233 SDValue visitFTRUNC(SDNode *N);
234 SDValue visitFFLOOR(SDNode *N);
235 SDValue visitBRCOND(SDNode *N);
236 SDValue visitBR_CC(SDNode *N);
237 SDValue visitLOAD(SDNode *N);
238 SDValue visitSTORE(SDNode *N);
239 SDValue visitINSERT_VECTOR_ELT(SDNode *N);
240 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
241 SDValue visitBUILD_VECTOR(SDNode *N);
242 SDValue visitCONCAT_VECTORS(SDNode *N);
243 SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
244 SDValue visitVECTOR_SHUFFLE(SDNode *N);
245 SDValue visitMEMBARRIER(SDNode *N);
247 SDValue XformToShuffleWithZero(SDNode *N);
248 SDValue ReassociateOps(unsigned Opc, DebugLoc DL, SDValue LHS, SDValue RHS);
250 SDValue visitShiftByConstant(SDNode *N, unsigned Amt);
252 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
253 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
254 SDValue SimplifySelect(DebugLoc DL, SDValue N0, SDValue N1, SDValue N2);
255 SDValue SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1, SDValue N2,
256 SDValue N3, ISD::CondCode CC,
257 bool NotExtCompare = false);
258 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
259 DebugLoc DL, bool foldBooleans = true);
260 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
262 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
263 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
264 SDValue BuildSDIV(SDNode *N);
265 SDValue BuildUDIV(SDNode *N);
266 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
267 bool DemandHighBits = true);
268 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
269 SDNode *MatchRotate(SDValue LHS, SDValue RHS, DebugLoc DL);
270 SDValue ReduceLoadWidth(SDNode *N);
271 SDValue ReduceLoadOpStoreWidth(SDNode *N);
272 SDValue TransformFPLoadStorePair(SDNode *N);
273 SDValue reduceBuildVecExtToExtBuildVec(SDNode *N);
274 SDValue reduceBuildVecConvertToConvertBuildVec(SDNode *N);
276 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
278 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
279 /// looking for aliasing nodes and adding them to the Aliases vector.
280 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
281 SmallVector<SDValue, 8> &Aliases);
283 /// isAlias - Return true if there is any possibility that the two addresses
285 bool isAlias(SDValue Ptr1, int64_t Size1,
286 const Value *SrcValue1, int SrcValueOffset1,
287 unsigned SrcValueAlign1,
288 const MDNode *TBAAInfo1,
289 SDValue Ptr2, int64_t Size2,
290 const Value *SrcValue2, int SrcValueOffset2,
291 unsigned SrcValueAlign2,
292 const MDNode *TBAAInfo2) const;
294 /// isAlias - Return true if there is any possibility that the two addresses
296 bool isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1);
298 /// FindAliasInfo - Extracts the relevant alias information from the memory
299 /// node. Returns true if the operand was a load.
300 bool FindAliasInfo(SDNode *N,
301 SDValue &Ptr, int64_t &Size,
302 const Value *&SrcValue, int &SrcValueOffset,
303 unsigned &SrcValueAlignment,
304 const MDNode *&TBAAInfo) const;
306 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes,
307 /// looking for a better chain (aliasing node.)
308 SDValue FindBetterChain(SDNode *N, SDValue Chain);
310 /// Merge consecutive store operations into a wide store.
311 /// This optimization uses wide integers or vectors when possible.
312 /// \return True if some memory operations were changed.
313 bool MergeConsecutiveStores(StoreSDNode *N);
316 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
317 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
318 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {}
320 /// Run - runs the dag combiner on all nodes in the work list
321 void Run(CombineLevel AtLevel);
323 SelectionDAG &getDAG() const { return DAG; }
325 /// getShiftAmountTy - Returns a type large enough to hold any valid
326 /// shift amount - before type legalization these can be huge.
327 EVT getShiftAmountTy(EVT LHSTy) {
328 return LegalTypes ? TLI.getShiftAmountTy(LHSTy) : TLI.getPointerTy();
331 /// isTypeLegal - This method returns true if we are running before type
332 /// legalization or if the specified VT is legal.
333 bool isTypeLegal(const EVT &VT) {
334 if (!LegalTypes) return true;
335 return TLI.isTypeLegal(VT);
342 /// WorkListRemover - This class is a DAGUpdateListener that removes any deleted
343 /// nodes from the worklist.
344 class WorkListRemover : public SelectionDAG::DAGUpdateListener {
347 explicit WorkListRemover(DAGCombiner &dc)
348 : SelectionDAG::DAGUpdateListener(dc.getDAG()), DC(dc) {}
350 virtual void NodeDeleted(SDNode *N, SDNode *E) {
351 DC.removeFromWorkList(N);
356 //===----------------------------------------------------------------------===//
357 // TargetLowering::DAGCombinerInfo implementation
358 //===----------------------------------------------------------------------===//
360 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
361 ((DAGCombiner*)DC)->AddToWorkList(N);
364 void TargetLowering::DAGCombinerInfo::RemoveFromWorklist(SDNode *N) {
365 ((DAGCombiner*)DC)->removeFromWorkList(N);
368 SDValue TargetLowering::DAGCombinerInfo::
369 CombineTo(SDNode *N, const std::vector<SDValue> &To, bool AddTo) {
370 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
373 SDValue TargetLowering::DAGCombinerInfo::
374 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
375 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
379 SDValue TargetLowering::DAGCombinerInfo::
380 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
381 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
384 void TargetLowering::DAGCombinerInfo::
385 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
386 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
389 //===----------------------------------------------------------------------===//
391 //===----------------------------------------------------------------------===//
393 /// isNegatibleForFree - Return 1 if we can compute the negated form of the
394 /// specified expression for the same cost as the expression itself, or 2 if we
395 /// can compute the negated form more cheaply than the expression itself.
396 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
397 const TargetLowering &TLI,
398 const TargetOptions *Options,
399 unsigned Depth = 0) {
400 // fneg is removable even if it has multiple uses.
401 if (Op.getOpcode() == ISD::FNEG) return 2;
403 // Don't allow anything with multiple uses.
404 if (!Op.hasOneUse()) return 0;
406 // Don't recurse exponentially.
407 if (Depth > 6) return 0;
409 switch (Op.getOpcode()) {
410 default: return false;
411 case ISD::ConstantFP:
412 // Don't invert constant FP values after legalize. The negated constant
413 // isn't necessarily legal.
414 return LegalOperations ? 0 : 1;
416 // FIXME: determine better conditions for this xform.
417 if (!Options->UnsafeFPMath) return 0;
419 // After operation legalization, it might not be legal to create new FSUBs.
420 if (LegalOperations &&
421 !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType()))
424 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
425 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
428 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
429 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
432 // We can't turn -(A-B) into B-A when we honor signed zeros.
433 if (!Options->UnsafeFPMath) return 0;
435 // fold (fneg (fsub A, B)) -> (fsub B, A)
440 if (Options->HonorSignDependentRoundingFPMath()) return 0;
442 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
443 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
447 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
453 return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options,
458 /// GetNegatedExpression - If isNegatibleForFree returns true, this function
459 /// returns the newly negated expression.
460 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
461 bool LegalOperations, unsigned Depth = 0) {
462 // fneg is removable even if it has multiple uses.
463 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
465 // Don't allow anything with multiple uses.
466 assert(Op.hasOneUse() && "Unknown reuse!");
468 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
469 switch (Op.getOpcode()) {
470 default: llvm_unreachable("Unknown code");
471 case ISD::ConstantFP: {
472 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
474 return DAG.getConstantFP(V, Op.getValueType());
477 // FIXME: determine better conditions for this xform.
478 assert(DAG.getTarget().Options.UnsafeFPMath);
480 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
481 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
482 DAG.getTargetLoweringInfo(),
483 &DAG.getTarget().Options, Depth+1))
484 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
485 GetNegatedExpression(Op.getOperand(0), DAG,
486 LegalOperations, Depth+1),
488 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
489 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
490 GetNegatedExpression(Op.getOperand(1), DAG,
491 LegalOperations, Depth+1),
494 // We can't turn -(A-B) into B-A when we honor signed zeros.
495 assert(DAG.getTarget().Options.UnsafeFPMath);
497 // fold (fneg (fsub 0, B)) -> B
498 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
499 if (N0CFP->getValueAPF().isZero())
500 return Op.getOperand(1);
502 // fold (fneg (fsub A, B)) -> (fsub B, A)
503 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
504 Op.getOperand(1), Op.getOperand(0));
508 assert(!DAG.getTarget().Options.HonorSignDependentRoundingFPMath());
510 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
511 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
512 DAG.getTargetLoweringInfo(),
513 &DAG.getTarget().Options, Depth+1))
514 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
515 GetNegatedExpression(Op.getOperand(0), DAG,
516 LegalOperations, Depth+1),
519 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
520 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
522 GetNegatedExpression(Op.getOperand(1), DAG,
523 LegalOperations, Depth+1));
527 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
528 GetNegatedExpression(Op.getOperand(0), DAG,
529 LegalOperations, Depth+1));
531 return DAG.getNode(ISD::FP_ROUND, Op.getDebugLoc(), Op.getValueType(),
532 GetNegatedExpression(Op.getOperand(0), DAG,
533 LegalOperations, Depth+1),
539 // isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
540 // that selects between the values 1 and 0, making it equivalent to a setcc.
541 // Also, set the incoming LHS, RHS, and CC references to the appropriate
542 // nodes based on the type of node we are checking. This simplifies life a
543 // bit for the callers.
544 static bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
546 if (N.getOpcode() == ISD::SETCC) {
547 LHS = N.getOperand(0);
548 RHS = N.getOperand(1);
549 CC = N.getOperand(2);
552 if (N.getOpcode() == ISD::SELECT_CC &&
553 N.getOperand(2).getOpcode() == ISD::Constant &&
554 N.getOperand(3).getOpcode() == ISD::Constant &&
555 cast<ConstantSDNode>(N.getOperand(2))->getAPIntValue() == 1 &&
556 cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) {
557 LHS = N.getOperand(0);
558 RHS = N.getOperand(1);
559 CC = N.getOperand(4);
565 // isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only
566 // one use. If this is true, it allows the users to invert the operation for
567 // free when it is profitable to do so.
568 static bool isOneUseSetCC(SDValue N) {
570 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
575 SDValue DAGCombiner::ReassociateOps(unsigned Opc, DebugLoc DL,
576 SDValue N0, SDValue N1) {
577 EVT VT = N0.getValueType();
578 if (N0.getOpcode() == Opc && isa<ConstantSDNode>(N0.getOperand(1))) {
579 if (isa<ConstantSDNode>(N1)) {
580 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
582 DAG.FoldConstantArithmetic(Opc, VT,
583 cast<ConstantSDNode>(N0.getOperand(1)),
584 cast<ConstantSDNode>(N1));
585 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
587 if (N0.hasOneUse()) {
588 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one use
589 SDValue OpNode = DAG.getNode(Opc, N0.getDebugLoc(), VT,
590 N0.getOperand(0), N1);
591 AddToWorkList(OpNode.getNode());
592 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
596 if (N1.getOpcode() == Opc && isa<ConstantSDNode>(N1.getOperand(1))) {
597 if (isa<ConstantSDNode>(N0)) {
598 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
600 DAG.FoldConstantArithmetic(Opc, VT,
601 cast<ConstantSDNode>(N1.getOperand(1)),
602 cast<ConstantSDNode>(N0));
603 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
605 if (N1.hasOneUse()) {
606 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one use
607 SDValue OpNode = DAG.getNode(Opc, N0.getDebugLoc(), VT,
608 N1.getOperand(0), N0);
609 AddToWorkList(OpNode.getNode());
610 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
617 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
619 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
621 DEBUG(dbgs() << "\nReplacing.1 ";
623 dbgs() << "\nWith: ";
624 To[0].getNode()->dump(&DAG);
625 dbgs() << " and " << NumTo-1 << " other values\n";
626 for (unsigned i = 0, e = NumTo; i != e; ++i)
627 assert((!To[i].getNode() ||
628 N->getValueType(i) == To[i].getValueType()) &&
629 "Cannot combine value to value of different type!"));
630 WorkListRemover DeadNodes(*this);
631 DAG.ReplaceAllUsesWith(N, To);
633 // Push the new nodes and any users onto the worklist
634 for (unsigned i = 0, e = NumTo; i != e; ++i) {
635 if (To[i].getNode()) {
636 AddToWorkList(To[i].getNode());
637 AddUsersToWorkList(To[i].getNode());
642 // Finally, if the node is now dead, remove it from the graph. The node
643 // may not be dead if the replacement process recursively simplified to
644 // something else needing this node.
645 if (N->use_empty()) {
646 // Nodes can be reintroduced into the worklist. Make sure we do not
647 // process a node that has been replaced.
648 removeFromWorkList(N);
650 // Finally, since the node is now dead, remove it from the graph.
653 return SDValue(N, 0);
657 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
658 // Replace all uses. If any nodes become isomorphic to other nodes and
659 // are deleted, make sure to remove them from our worklist.
660 WorkListRemover DeadNodes(*this);
661 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New);
663 // Push the new node and any (possibly new) users onto the worklist.
664 AddToWorkList(TLO.New.getNode());
665 AddUsersToWorkList(TLO.New.getNode());
667 // Finally, if the node is now dead, remove it from the graph. The node
668 // may not be dead if the replacement process recursively simplified to
669 // something else needing this node.
670 if (TLO.Old.getNode()->use_empty()) {
671 removeFromWorkList(TLO.Old.getNode());
673 // If the operands of this node are only used by the node, they will now
674 // be dead. Make sure to visit them first to delete dead nodes early.
675 for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands(); i != e; ++i)
676 if (TLO.Old.getNode()->getOperand(i).getNode()->hasOneUse())
677 AddToWorkList(TLO.Old.getNode()->getOperand(i).getNode());
679 DAG.DeleteNode(TLO.Old.getNode());
683 /// SimplifyDemandedBits - Check the specified integer node value to see if
684 /// it can be simplified or if things it uses can be simplified by bit
685 /// propagation. If so, return true.
686 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
687 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
688 APInt KnownZero, KnownOne;
689 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
693 AddToWorkList(Op.getNode());
695 // Replace the old value with the new one.
697 DEBUG(dbgs() << "\nReplacing.2 ";
698 TLO.Old.getNode()->dump(&DAG);
699 dbgs() << "\nWith: ";
700 TLO.New.getNode()->dump(&DAG);
703 CommitTargetLoweringOpt(TLO);
707 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
708 DebugLoc dl = Load->getDebugLoc();
709 EVT VT = Load->getValueType(0);
710 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
712 DEBUG(dbgs() << "\nReplacing.9 ";
714 dbgs() << "\nWith: ";
715 Trunc.getNode()->dump(&DAG);
717 WorkListRemover DeadNodes(*this);
718 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc);
719 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1));
720 removeFromWorkList(Load);
721 DAG.DeleteNode(Load);
722 AddToWorkList(Trunc.getNode());
725 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
727 DebugLoc dl = Op.getDebugLoc();
728 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
729 EVT MemVT = LD->getMemoryVT();
730 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
731 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
733 : LD->getExtensionType();
735 return DAG.getExtLoad(ExtType, dl, PVT,
736 LD->getChain(), LD->getBasePtr(),
737 LD->getPointerInfo(),
738 MemVT, LD->isVolatile(),
739 LD->isNonTemporal(), LD->getAlignment());
742 unsigned Opc = Op.getOpcode();
745 case ISD::AssertSext:
746 return DAG.getNode(ISD::AssertSext, dl, PVT,
747 SExtPromoteOperand(Op.getOperand(0), PVT),
749 case ISD::AssertZext:
750 return DAG.getNode(ISD::AssertZext, dl, PVT,
751 ZExtPromoteOperand(Op.getOperand(0), PVT),
753 case ISD::Constant: {
755 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
756 return DAG.getNode(ExtOpc, dl, PVT, Op);
760 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
762 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
765 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
766 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
768 EVT OldVT = Op.getValueType();
769 DebugLoc dl = Op.getDebugLoc();
770 bool Replace = false;
771 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
772 if (NewOp.getNode() == 0)
774 AddToWorkList(NewOp.getNode());
777 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
778 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
779 DAG.getValueType(OldVT));
782 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
783 EVT OldVT = Op.getValueType();
784 DebugLoc dl = Op.getDebugLoc();
785 bool Replace = false;
786 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
787 if (NewOp.getNode() == 0)
789 AddToWorkList(NewOp.getNode());
792 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
793 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
796 /// PromoteIntBinOp - Promote the specified integer binary operation if the
797 /// target indicates it is beneficial. e.g. On x86, it's usually better to
798 /// promote i16 operations to i32 since i16 instructions are longer.
799 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
800 if (!LegalOperations)
803 EVT VT = Op.getValueType();
804 if (VT.isVector() || !VT.isInteger())
807 // If operation type is 'undesirable', e.g. i16 on x86, consider
809 unsigned Opc = Op.getOpcode();
810 if (TLI.isTypeDesirableForOp(Opc, VT))
814 // Consult target whether it is a good idea to promote this operation and
815 // what's the right type to promote it to.
816 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
817 assert(PVT != VT && "Don't know what type to promote to!");
819 bool Replace0 = false;
820 SDValue N0 = Op.getOperand(0);
821 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
822 if (NN0.getNode() == 0)
825 bool Replace1 = false;
826 SDValue N1 = Op.getOperand(1);
831 NN1 = PromoteOperand(N1, PVT, Replace1);
832 if (NN1.getNode() == 0)
836 AddToWorkList(NN0.getNode());
838 AddToWorkList(NN1.getNode());
841 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
843 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
845 DEBUG(dbgs() << "\nPromoting ";
846 Op.getNode()->dump(&DAG));
847 DebugLoc dl = Op.getDebugLoc();
848 return DAG.getNode(ISD::TRUNCATE, dl, VT,
849 DAG.getNode(Opc, dl, PVT, NN0, NN1));
854 /// PromoteIntShiftOp - Promote the specified integer shift operation if the
855 /// target indicates it is beneficial. e.g. On x86, it's usually better to
856 /// promote i16 operations to i32 since i16 instructions are longer.
857 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
858 if (!LegalOperations)
861 EVT VT = Op.getValueType();
862 if (VT.isVector() || !VT.isInteger())
865 // If operation type is 'undesirable', e.g. i16 on x86, consider
867 unsigned Opc = Op.getOpcode();
868 if (TLI.isTypeDesirableForOp(Opc, VT))
872 // Consult target whether it is a good idea to promote this operation and
873 // what's the right type to promote it to.
874 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
875 assert(PVT != VT && "Don't know what type to promote to!");
877 bool Replace = false;
878 SDValue N0 = Op.getOperand(0);
880 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
881 else if (Opc == ISD::SRL)
882 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
884 N0 = PromoteOperand(N0, PVT, Replace);
885 if (N0.getNode() == 0)
888 AddToWorkList(N0.getNode());
890 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
892 DEBUG(dbgs() << "\nPromoting ";
893 Op.getNode()->dump(&DAG));
894 DebugLoc dl = Op.getDebugLoc();
895 return DAG.getNode(ISD::TRUNCATE, dl, VT,
896 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
901 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
902 if (!LegalOperations)
905 EVT VT = Op.getValueType();
906 if (VT.isVector() || !VT.isInteger())
909 // If operation type is 'undesirable', e.g. i16 on x86, consider
911 unsigned Opc = Op.getOpcode();
912 if (TLI.isTypeDesirableForOp(Opc, VT))
916 // Consult target whether it is a good idea to promote this operation and
917 // what's the right type to promote it to.
918 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
919 assert(PVT != VT && "Don't know what type to promote to!");
920 // fold (aext (aext x)) -> (aext x)
921 // fold (aext (zext x)) -> (zext x)
922 // fold (aext (sext x)) -> (sext x)
923 DEBUG(dbgs() << "\nPromoting ";
924 Op.getNode()->dump(&DAG));
925 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), VT, Op.getOperand(0));
930 bool DAGCombiner::PromoteLoad(SDValue Op) {
931 if (!LegalOperations)
934 EVT VT = Op.getValueType();
935 if (VT.isVector() || !VT.isInteger())
938 // If operation type is 'undesirable', e.g. i16 on x86, consider
940 unsigned Opc = Op.getOpcode();
941 if (TLI.isTypeDesirableForOp(Opc, VT))
945 // Consult target whether it is a good idea to promote this operation and
946 // what's the right type to promote it to.
947 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
948 assert(PVT != VT && "Don't know what type to promote to!");
950 DebugLoc dl = Op.getDebugLoc();
951 SDNode *N = Op.getNode();
952 LoadSDNode *LD = cast<LoadSDNode>(N);
953 EVT MemVT = LD->getMemoryVT();
954 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
955 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
957 : LD->getExtensionType();
958 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
959 LD->getChain(), LD->getBasePtr(),
960 LD->getPointerInfo(),
961 MemVT, LD->isVolatile(),
962 LD->isNonTemporal(), LD->getAlignment());
963 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
965 DEBUG(dbgs() << "\nPromoting ";
968 Result.getNode()->dump(&DAG);
970 WorkListRemover DeadNodes(*this);
971 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result);
972 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1));
973 removeFromWorkList(N);
975 AddToWorkList(Result.getNode());
982 //===----------------------------------------------------------------------===//
983 // Main DAG Combiner implementation
984 //===----------------------------------------------------------------------===//
986 void DAGCombiner::Run(CombineLevel AtLevel) {
987 // set the instance variables, so that the various visit routines may use it.
989 LegalOperations = Level >= AfterLegalizeVectorOps;
990 LegalTypes = Level >= AfterLegalizeTypes;
992 // Add all the dag nodes to the worklist.
993 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
994 E = DAG.allnodes_end(); I != E; ++I)
997 // Create a dummy node (which is not added to allnodes), that adds a reference
998 // to the root node, preventing it from being deleted, and tracking any
999 // changes of the root.
1000 HandleSDNode Dummy(DAG.getRoot());
1002 // The root of the dag may dangle to deleted nodes until the dag combiner is
1003 // done. Set it to null to avoid confusion.
1004 DAG.setRoot(SDValue());
1006 // while the worklist isn't empty, find a node and
1007 // try and combine it.
1008 while (!WorkListContents.empty()) {
1010 // The WorkListOrder holds the SDNodes in order, but it may contain duplicates.
1011 // In order to avoid a linear scan, we use a set (O(log N)) to hold what the
1012 // worklist *should* contain, and check the node we want to visit is should
1013 // actually be visited.
1015 N = WorkListOrder.pop_back_val();
1016 } while (!WorkListContents.erase(N));
1018 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1019 // N is deleted from the DAG, since they too may now be dead or may have a
1020 // reduced number of uses, allowing other xforms.
1021 if (N->use_empty() && N != &Dummy) {
1022 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1023 AddToWorkList(N->getOperand(i).getNode());
1029 SDValue RV = combine(N);
1031 if (RV.getNode() == 0)
1036 // If we get back the same node we passed in, rather than a new node or
1037 // zero, we know that the node must have defined multiple values and
1038 // CombineTo was used. Since CombineTo takes care of the worklist
1039 // mechanics for us, we have no work to do in this case.
1040 if (RV.getNode() == N)
1043 assert(N->getOpcode() != ISD::DELETED_NODE &&
1044 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1045 "Node was deleted but visit returned new node!");
1047 DEBUG(dbgs() << "\nReplacing.3 ";
1049 dbgs() << "\nWith: ";
1050 RV.getNode()->dump(&DAG);
1053 // Transfer debug value.
1054 DAG.TransferDbgValues(SDValue(N, 0), RV);
1055 WorkListRemover DeadNodes(*this);
1056 if (N->getNumValues() == RV.getNode()->getNumValues())
1057 DAG.ReplaceAllUsesWith(N, RV.getNode());
1059 assert(N->getValueType(0) == RV.getValueType() &&
1060 N->getNumValues() == 1 && "Type mismatch");
1062 DAG.ReplaceAllUsesWith(N, &OpV);
1065 // Push the new node and any users onto the worklist
1066 AddToWorkList(RV.getNode());
1067 AddUsersToWorkList(RV.getNode());
1069 // Add any uses of the old node to the worklist in case this node is the
1070 // last one that uses them. They may become dead after this node is
1072 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1073 AddToWorkList(N->getOperand(i).getNode());
1075 // Finally, if the node is now dead, remove it from the graph. The node
1076 // may not be dead if the replacement process recursively simplified to
1077 // something else needing this node.
1078 if (N->use_empty()) {
1079 // Nodes can be reintroduced into the worklist. Make sure we do not
1080 // process a node that has been replaced.
1081 removeFromWorkList(N);
1083 // Finally, since the node is now dead, remove it from the graph.
1088 // If the root changed (e.g. it was a dead load, update the root).
1089 DAG.setRoot(Dummy.getValue());
1090 DAG.RemoveDeadNodes();
1093 SDValue DAGCombiner::visit(SDNode *N) {
1094 switch (N->getOpcode()) {
1096 case ISD::TokenFactor: return visitTokenFactor(N);
1097 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1098 case ISD::ADD: return visitADD(N);
1099 case ISD::SUB: return visitSUB(N);
1100 case ISD::ADDC: return visitADDC(N);
1101 case ISD::SUBC: return visitSUBC(N);
1102 case ISD::ADDE: return visitADDE(N);
1103 case ISD::SUBE: return visitSUBE(N);
1104 case ISD::MUL: return visitMUL(N);
1105 case ISD::SDIV: return visitSDIV(N);
1106 case ISD::UDIV: return visitUDIV(N);
1107 case ISD::SREM: return visitSREM(N);
1108 case ISD::UREM: return visitUREM(N);
1109 case ISD::MULHU: return visitMULHU(N);
1110 case ISD::MULHS: return visitMULHS(N);
1111 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1112 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1113 case ISD::SMULO: return visitSMULO(N);
1114 case ISD::UMULO: return visitUMULO(N);
1115 case ISD::SDIVREM: return visitSDIVREM(N);
1116 case ISD::UDIVREM: return visitUDIVREM(N);
1117 case ISD::AND: return visitAND(N);
1118 case ISD::OR: return visitOR(N);
1119 case ISD::XOR: return visitXOR(N);
1120 case ISD::SHL: return visitSHL(N);
1121 case ISD::SRA: return visitSRA(N);
1122 case ISD::SRL: return visitSRL(N);
1123 case ISD::CTLZ: return visitCTLZ(N);
1124 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1125 case ISD::CTTZ: return visitCTTZ(N);
1126 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1127 case ISD::CTPOP: return visitCTPOP(N);
1128 case ISD::SELECT: return visitSELECT(N);
1129 case ISD::SELECT_CC: return visitSELECT_CC(N);
1130 case ISD::SETCC: return visitSETCC(N);
1131 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1132 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1133 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1134 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1135 case ISD::TRUNCATE: return visitTRUNCATE(N);
1136 case ISD::BITCAST: return visitBITCAST(N);
1137 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1138 case ISD::FADD: return visitFADD(N);
1139 case ISD::FSUB: return visitFSUB(N);
1140 case ISD::FMUL: return visitFMUL(N);
1141 case ISD::FMA: return visitFMA(N);
1142 case ISD::FDIV: return visitFDIV(N);
1143 case ISD::FREM: return visitFREM(N);
1144 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1145 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1146 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1147 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1148 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1149 case ISD::FP_ROUND: return visitFP_ROUND(N);
1150 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1151 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1152 case ISD::FNEG: return visitFNEG(N);
1153 case ISD::FABS: return visitFABS(N);
1154 case ISD::FFLOOR: return visitFFLOOR(N);
1155 case ISD::FCEIL: return visitFCEIL(N);
1156 case ISD::FTRUNC: return visitFTRUNC(N);
1157 case ISD::BRCOND: return visitBRCOND(N);
1158 case ISD::BR_CC: return visitBR_CC(N);
1159 case ISD::LOAD: return visitLOAD(N);
1160 case ISD::STORE: return visitSTORE(N);
1161 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1162 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1163 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1164 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1165 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1166 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1167 case ISD::MEMBARRIER: return visitMEMBARRIER(N);
1172 SDValue DAGCombiner::combine(SDNode *N) {
1173 SDValue RV = visit(N);
1175 // If nothing happened, try a target-specific DAG combine.
1176 if (RV.getNode() == 0) {
1177 assert(N->getOpcode() != ISD::DELETED_NODE &&
1178 "Node was deleted but visit returned NULL!");
1180 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1181 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1183 // Expose the DAG combiner to the target combiner impls.
1184 TargetLowering::DAGCombinerInfo
1185 DagCombineInfo(DAG, !LegalTypes, !LegalOperations, false, this);
1187 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1191 // If nothing happened still, try promoting the operation.
1192 if (RV.getNode() == 0) {
1193 switch (N->getOpcode()) {
1201 RV = PromoteIntBinOp(SDValue(N, 0));
1206 RV = PromoteIntShiftOp(SDValue(N, 0));
1208 case ISD::SIGN_EXTEND:
1209 case ISD::ZERO_EXTEND:
1210 case ISD::ANY_EXTEND:
1211 RV = PromoteExtend(SDValue(N, 0));
1214 if (PromoteLoad(SDValue(N, 0)))
1220 // If N is a commutative binary node, try commuting it to enable more
1222 if (RV.getNode() == 0 &&
1223 SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1224 N->getNumValues() == 1) {
1225 SDValue N0 = N->getOperand(0);
1226 SDValue N1 = N->getOperand(1);
1228 // Constant operands are canonicalized to RHS.
1229 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1230 SDValue Ops[] = { N1, N0 };
1231 SDNode *CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(),
1234 return SDValue(CSENode, 0);
1241 /// getInputChainForNode - Given a node, return its input chain if it has one,
1242 /// otherwise return a null sd operand.
1243 static SDValue getInputChainForNode(SDNode *N) {
1244 if (unsigned NumOps = N->getNumOperands()) {
1245 if (N->getOperand(0).getValueType() == MVT::Other)
1246 return N->getOperand(0);
1247 else if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1248 return N->getOperand(NumOps-1);
1249 for (unsigned i = 1; i < NumOps-1; ++i)
1250 if (N->getOperand(i).getValueType() == MVT::Other)
1251 return N->getOperand(i);
1256 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1257 // If N has two operands, where one has an input chain equal to the other,
1258 // the 'other' chain is redundant.
1259 if (N->getNumOperands() == 2) {
1260 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1261 return N->getOperand(0);
1262 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1263 return N->getOperand(1);
1266 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1267 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1268 SmallPtrSet<SDNode*, 16> SeenOps;
1269 bool Changed = false; // If we should replace this token factor.
1271 // Start out with this token factor.
1274 // Iterate through token factors. The TFs grows when new token factors are
1276 for (unsigned i = 0; i < TFs.size(); ++i) {
1277 SDNode *TF = TFs[i];
1279 // Check each of the operands.
1280 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1281 SDValue Op = TF->getOperand(i);
1283 switch (Op.getOpcode()) {
1284 case ISD::EntryToken:
1285 // Entry tokens don't need to be added to the list. They are
1290 case ISD::TokenFactor:
1291 if (Op.hasOneUse() &&
1292 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1293 // Queue up for processing.
1294 TFs.push_back(Op.getNode());
1295 // Clean up in case the token factor is removed.
1296 AddToWorkList(Op.getNode());
1303 // Only add if it isn't already in the list.
1304 if (SeenOps.insert(Op.getNode()))
1315 // If we've change things around then replace token factor.
1318 // The entry token is the only possible outcome.
1319 Result = DAG.getEntryNode();
1321 // New and improved token factor.
1322 Result = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
1323 MVT::Other, &Ops[0], Ops.size());
1326 // Don't add users to work list.
1327 return CombineTo(N, Result, false);
1333 /// MERGE_VALUES can always be eliminated.
1334 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1335 WorkListRemover DeadNodes(*this);
1336 // Replacing results may cause a different MERGE_VALUES to suddenly
1337 // be CSE'd with N, and carry its uses with it. Iterate until no
1338 // uses remain, to ensure that the node can be safely deleted.
1339 // First add the users of this node to the work list so that they
1340 // can be tried again once they have new operands.
1341 AddUsersToWorkList(N);
1343 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1344 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i));
1345 } while (!N->use_empty());
1346 removeFromWorkList(N);
1348 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1352 SDValue combineShlAddConstant(DebugLoc DL, SDValue N0, SDValue N1,
1353 SelectionDAG &DAG) {
1354 EVT VT = N0.getValueType();
1355 SDValue N00 = N0.getOperand(0);
1356 SDValue N01 = N0.getOperand(1);
1357 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N01);
1359 if (N01C && N00.getOpcode() == ISD::ADD && N00.getNode()->hasOneUse() &&
1360 isa<ConstantSDNode>(N00.getOperand(1))) {
1361 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1362 N0 = DAG.getNode(ISD::ADD, N0.getDebugLoc(), VT,
1363 DAG.getNode(ISD::SHL, N00.getDebugLoc(), VT,
1364 N00.getOperand(0), N01),
1365 DAG.getNode(ISD::SHL, N01.getDebugLoc(), VT,
1366 N00.getOperand(1), N01));
1367 return DAG.getNode(ISD::ADD, DL, VT, N0, N1);
1373 SDValue DAGCombiner::visitADD(SDNode *N) {
1374 SDValue N0 = N->getOperand(0);
1375 SDValue N1 = N->getOperand(1);
1376 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1377 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1378 EVT VT = N0.getValueType();
1381 if (VT.isVector()) {
1382 SDValue FoldedVOp = SimplifyVBinOp(N);
1383 if (FoldedVOp.getNode()) return FoldedVOp;
1385 // fold (add x, 0) -> x, vector edition
1386 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1388 if (ISD::isBuildVectorAllZeros(N0.getNode()))
1392 // fold (add x, undef) -> undef
1393 if (N0.getOpcode() == ISD::UNDEF)
1395 if (N1.getOpcode() == ISD::UNDEF)
1397 // fold (add c1, c2) -> c1+c2
1399 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1400 // canonicalize constant to RHS
1402 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N1, N0);
1403 // fold (add x, 0) -> x
1404 if (N1C && N1C->isNullValue())
1406 // fold (add Sym, c) -> Sym+c
1407 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1408 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1409 GA->getOpcode() == ISD::GlobalAddress)
1410 return DAG.getGlobalAddress(GA->getGlobal(), N1C->getDebugLoc(), VT,
1412 (uint64_t)N1C->getSExtValue());
1413 // fold ((c1-A)+c2) -> (c1+c2)-A
1414 if (N1C && N0.getOpcode() == ISD::SUB)
1415 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1416 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1417 DAG.getConstant(N1C->getAPIntValue()+
1418 N0C->getAPIntValue(), VT),
1421 SDValue RADD = ReassociateOps(ISD::ADD, N->getDebugLoc(), N0, N1);
1422 if (RADD.getNode() != 0)
1424 // fold ((0-A) + B) -> B-A
1425 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1426 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1427 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1, N0.getOperand(1));
1428 // fold (A + (0-B)) -> A-B
1429 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1430 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1431 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, N1.getOperand(1));
1432 // fold (A+(B-A)) -> B
1433 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1434 return N1.getOperand(0);
1435 // fold ((B-A)+A) -> B
1436 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1437 return N0.getOperand(0);
1438 // fold (A+(B-(A+C))) to (B-C)
1439 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1440 N0 == N1.getOperand(1).getOperand(0))
1441 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1.getOperand(0),
1442 N1.getOperand(1).getOperand(1));
1443 // fold (A+(B-(C+A))) to (B-C)
1444 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1445 N0 == N1.getOperand(1).getOperand(1))
1446 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1.getOperand(0),
1447 N1.getOperand(1).getOperand(0));
1448 // fold (A+((B-A)+or-C)) to (B+or-C)
1449 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1450 N1.getOperand(0).getOpcode() == ISD::SUB &&
1451 N0 == N1.getOperand(0).getOperand(1))
1452 return DAG.getNode(N1.getOpcode(), N->getDebugLoc(), VT,
1453 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1455 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1456 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1457 SDValue N00 = N0.getOperand(0);
1458 SDValue N01 = N0.getOperand(1);
1459 SDValue N10 = N1.getOperand(0);
1460 SDValue N11 = N1.getOperand(1);
1462 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1463 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1464 DAG.getNode(ISD::ADD, N0.getDebugLoc(), VT, N00, N10),
1465 DAG.getNode(ISD::ADD, N1.getDebugLoc(), VT, N01, N11));
1468 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1469 return SDValue(N, 0);
1471 // fold (a+b) -> (a|b) iff a and b share no bits.
1472 if (VT.isInteger() && !VT.isVector()) {
1473 APInt LHSZero, LHSOne;
1474 APInt RHSZero, RHSOne;
1475 DAG.ComputeMaskedBits(N0, LHSZero, LHSOne);
1477 if (LHSZero.getBoolValue()) {
1478 DAG.ComputeMaskedBits(N1, RHSZero, RHSOne);
1480 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1481 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1482 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1483 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N1);
1487 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1488 if (N0.getOpcode() == ISD::SHL && N0.getNode()->hasOneUse()) {
1489 SDValue Result = combineShlAddConstant(N->getDebugLoc(), N0, N1, DAG);
1490 if (Result.getNode()) return Result;
1492 if (N1.getOpcode() == ISD::SHL && N1.getNode()->hasOneUse()) {
1493 SDValue Result = combineShlAddConstant(N->getDebugLoc(), N1, N0, DAG);
1494 if (Result.getNode()) return Result;
1497 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1498 if (N1.getOpcode() == ISD::SHL &&
1499 N1.getOperand(0).getOpcode() == ISD::SUB)
1500 if (ConstantSDNode *C =
1501 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1502 if (C->getAPIntValue() == 0)
1503 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0,
1504 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1505 N1.getOperand(0).getOperand(1),
1507 if (N0.getOpcode() == ISD::SHL &&
1508 N0.getOperand(0).getOpcode() == ISD::SUB)
1509 if (ConstantSDNode *C =
1510 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1511 if (C->getAPIntValue() == 0)
1512 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1,
1513 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1514 N0.getOperand(0).getOperand(1),
1517 if (N1.getOpcode() == ISD::AND) {
1518 SDValue AndOp0 = N1.getOperand(0);
1519 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1520 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1521 unsigned DestBits = VT.getScalarType().getSizeInBits();
1523 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1524 // and similar xforms where the inner op is either ~0 or 0.
1525 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1526 DebugLoc DL = N->getDebugLoc();
1527 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1531 // add (sext i1), X -> sub X, (zext i1)
1532 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1533 N0.getOperand(0).getValueType() == MVT::i1 &&
1534 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1535 DebugLoc DL = N->getDebugLoc();
1536 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1537 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1543 SDValue DAGCombiner::visitADDC(SDNode *N) {
1544 SDValue N0 = N->getOperand(0);
1545 SDValue N1 = N->getOperand(1);
1546 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1547 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1548 EVT VT = N0.getValueType();
1550 // If the flag result is dead, turn this into an ADD.
1551 if (!N->hasAnyUseOfValue(1))
1552 return CombineTo(N, DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0, N1),
1553 DAG.getNode(ISD::CARRY_FALSE,
1554 N->getDebugLoc(), MVT::Glue));
1556 // canonicalize constant to RHS.
1558 return DAG.getNode(ISD::ADDC, N->getDebugLoc(), N->getVTList(), N1, N0);
1560 // fold (addc x, 0) -> x + no carry out
1561 if (N1C && N1C->isNullValue())
1562 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1563 N->getDebugLoc(), MVT::Glue));
1565 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1566 APInt LHSZero, LHSOne;
1567 APInt RHSZero, RHSOne;
1568 DAG.ComputeMaskedBits(N0, LHSZero, LHSOne);
1570 if (LHSZero.getBoolValue()) {
1571 DAG.ComputeMaskedBits(N1, RHSZero, RHSOne);
1573 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1574 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1575 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1576 return CombineTo(N, DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N1),
1577 DAG.getNode(ISD::CARRY_FALSE,
1578 N->getDebugLoc(), MVT::Glue));
1584 SDValue DAGCombiner::visitADDE(SDNode *N) {
1585 SDValue N0 = N->getOperand(0);
1586 SDValue N1 = N->getOperand(1);
1587 SDValue CarryIn = N->getOperand(2);
1588 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1589 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1591 // canonicalize constant to RHS
1593 return DAG.getNode(ISD::ADDE, N->getDebugLoc(), N->getVTList(),
1596 // fold (adde x, y, false) -> (addc x, y)
1597 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1598 return DAG.getNode(ISD::ADDC, N->getDebugLoc(), N->getVTList(), N0, N1);
1603 // Since it may not be valid to emit a fold to zero for vector initializers
1604 // check if we can before folding.
1605 static SDValue tryFoldToZero(DebugLoc DL, const TargetLowering &TLI, EVT VT,
1606 SelectionDAG &DAG, bool LegalOperations) {
1607 if (!VT.isVector()) {
1608 return DAG.getConstant(0, VT);
1610 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) {
1611 // Produce a vector of zeros.
1612 SDValue El = DAG.getConstant(0, VT.getVectorElementType());
1613 std::vector<SDValue> Ops(VT.getVectorNumElements(), El);
1614 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT,
1615 &Ops[0], Ops.size());
1620 SDValue DAGCombiner::visitSUB(SDNode *N) {
1621 SDValue N0 = N->getOperand(0);
1622 SDValue N1 = N->getOperand(1);
1623 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1624 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1625 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? 0 :
1626 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1627 EVT VT = N0.getValueType();
1630 if (VT.isVector()) {
1631 SDValue FoldedVOp = SimplifyVBinOp(N);
1632 if (FoldedVOp.getNode()) return FoldedVOp;
1634 // fold (sub x, 0) -> x, vector edition
1635 if (ISD::isBuildVectorAllZeros(N1.getNode()))
1639 // fold (sub x, x) -> 0
1640 // FIXME: Refactor this and xor and other similar operations together.
1642 return tryFoldToZero(N->getDebugLoc(), TLI, VT, DAG, LegalOperations);
1643 // fold (sub c1, c2) -> c1-c2
1645 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1646 // fold (sub x, c) -> (add x, -c)
1648 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0,
1649 DAG.getConstant(-N1C->getAPIntValue(), VT));
1650 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1651 if (N0C && N0C->isAllOnesValue())
1652 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0);
1653 // fold A-(A-B) -> B
1654 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1655 return N1.getOperand(1);
1656 // fold (A+B)-A -> B
1657 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1658 return N0.getOperand(1);
1659 // fold (A+B)-B -> A
1660 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1661 return N0.getOperand(0);
1662 // fold C2-(A+C1) -> (C2-C1)-A
1663 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1664 SDValue NewC = DAG.getConstant(N0C->getAPIntValue() - N1C1->getAPIntValue(),
1666 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, NewC,
1669 // fold ((A+(B+or-C))-B) -> A+or-C
1670 if (N0.getOpcode() == ISD::ADD &&
1671 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1672 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1673 N0.getOperand(1).getOperand(0) == N1)
1674 return DAG.getNode(N0.getOperand(1).getOpcode(), N->getDebugLoc(), VT,
1675 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1676 // fold ((A+(C+B))-B) -> A+C
1677 if (N0.getOpcode() == ISD::ADD &&
1678 N0.getOperand(1).getOpcode() == ISD::ADD &&
1679 N0.getOperand(1).getOperand(1) == N1)
1680 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT,
1681 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1682 // fold ((A-(B-C))-C) -> A-B
1683 if (N0.getOpcode() == ISD::SUB &&
1684 N0.getOperand(1).getOpcode() == ISD::SUB &&
1685 N0.getOperand(1).getOperand(1) == N1)
1686 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1687 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1689 // If either operand of a sub is undef, the result is undef
1690 if (N0.getOpcode() == ISD::UNDEF)
1692 if (N1.getOpcode() == ISD::UNDEF)
1695 // If the relocation model supports it, consider symbol offsets.
1696 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1697 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1698 // fold (sub Sym, c) -> Sym-c
1699 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1700 return DAG.getGlobalAddress(GA->getGlobal(), N1C->getDebugLoc(), VT,
1702 (uint64_t)N1C->getSExtValue());
1703 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1704 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1705 if (GA->getGlobal() == GB->getGlobal())
1706 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1713 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1714 SDValue N0 = N->getOperand(0);
1715 SDValue N1 = N->getOperand(1);
1716 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1717 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1718 EVT VT = N0.getValueType();
1720 // If the flag result is dead, turn this into an SUB.
1721 if (!N->hasAnyUseOfValue(1))
1722 return CombineTo(N, DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, N1),
1723 DAG.getNode(ISD::CARRY_FALSE, N->getDebugLoc(),
1726 // fold (subc x, x) -> 0 + no borrow
1728 return CombineTo(N, DAG.getConstant(0, VT),
1729 DAG.getNode(ISD::CARRY_FALSE, N->getDebugLoc(),
1732 // fold (subc x, 0) -> x + no borrow
1733 if (N1C && N1C->isNullValue())
1734 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, N->getDebugLoc(),
1737 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1738 if (N0C && N0C->isAllOnesValue())
1739 return CombineTo(N, DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0),
1740 DAG.getNode(ISD::CARRY_FALSE, N->getDebugLoc(),
1746 SDValue DAGCombiner::visitSUBE(SDNode *N) {
1747 SDValue N0 = N->getOperand(0);
1748 SDValue N1 = N->getOperand(1);
1749 SDValue CarryIn = N->getOperand(2);
1751 // fold (sube x, y, false) -> (subc x, y)
1752 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1753 return DAG.getNode(ISD::SUBC, N->getDebugLoc(), N->getVTList(), N0, N1);
1758 SDValue DAGCombiner::visitMUL(SDNode *N) {
1759 SDValue N0 = N->getOperand(0);
1760 SDValue N1 = N->getOperand(1);
1761 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1762 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1763 EVT VT = N0.getValueType();
1766 if (VT.isVector()) {
1767 SDValue FoldedVOp = SimplifyVBinOp(N);
1768 if (FoldedVOp.getNode()) return FoldedVOp;
1771 // fold (mul x, undef) -> 0
1772 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1773 return DAG.getConstant(0, VT);
1774 // fold (mul c1, c2) -> c1*c2
1776 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0C, N1C);
1777 // canonicalize constant to RHS
1779 return DAG.getNode(ISD::MUL, N->getDebugLoc(), VT, N1, N0);
1780 // fold (mul x, 0) -> 0
1781 if (N1C && N1C->isNullValue())
1783 // fold (mul x, -1) -> 0-x
1784 if (N1C && N1C->isAllOnesValue())
1785 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1786 DAG.getConstant(0, VT), N0);
1787 // fold (mul x, (1 << c)) -> x << c
1788 if (N1C && N1C->getAPIntValue().isPowerOf2())
1789 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
1790 DAG.getConstant(N1C->getAPIntValue().logBase2(),
1791 getShiftAmountTy(N0.getValueType())));
1792 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
1793 if (N1C && (-N1C->getAPIntValue()).isPowerOf2()) {
1794 unsigned Log2Val = (-N1C->getAPIntValue()).logBase2();
1795 // FIXME: If the input is something that is easily negated (e.g. a
1796 // single-use add), we should put the negate there.
1797 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1798 DAG.getConstant(0, VT),
1799 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
1800 DAG.getConstant(Log2Val,
1801 getShiftAmountTy(N0.getValueType()))));
1803 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
1804 if (N1C && N0.getOpcode() == ISD::SHL &&
1805 isa<ConstantSDNode>(N0.getOperand(1))) {
1806 SDValue C3 = DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1807 N1, N0.getOperand(1));
1808 AddToWorkList(C3.getNode());
1809 return DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1810 N0.getOperand(0), C3);
1813 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
1816 SDValue Sh(0,0), Y(0,0);
1817 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
1818 if (N0.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N0.getOperand(1)) &&
1819 N0.getNode()->hasOneUse()) {
1821 } else if (N1.getOpcode() == ISD::SHL &&
1822 isa<ConstantSDNode>(N1.getOperand(1)) &&
1823 N1.getNode()->hasOneUse()) {
1828 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1829 Sh.getOperand(0), Y);
1830 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1831 Mul, Sh.getOperand(1));
1835 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
1836 if (N1C && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
1837 isa<ConstantSDNode>(N0.getOperand(1)))
1838 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT,
1839 DAG.getNode(ISD::MUL, N0.getDebugLoc(), VT,
1840 N0.getOperand(0), N1),
1841 DAG.getNode(ISD::MUL, N1.getDebugLoc(), VT,
1842 N0.getOperand(1), N1));
1845 SDValue RMUL = ReassociateOps(ISD::MUL, N->getDebugLoc(), N0, N1);
1846 if (RMUL.getNode() != 0)
1852 SDValue DAGCombiner::visitSDIV(SDNode *N) {
1853 SDValue N0 = N->getOperand(0);
1854 SDValue N1 = N->getOperand(1);
1855 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1856 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1857 EVT VT = N->getValueType(0);
1860 if (VT.isVector()) {
1861 SDValue FoldedVOp = SimplifyVBinOp(N);
1862 if (FoldedVOp.getNode()) return FoldedVOp;
1865 // fold (sdiv c1, c2) -> c1/c2
1866 if (N0C && N1C && !N1C->isNullValue())
1867 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
1868 // fold (sdiv X, 1) -> X
1869 if (N1C && N1C->getAPIntValue() == 1LL)
1871 // fold (sdiv X, -1) -> 0-X
1872 if (N1C && N1C->isAllOnesValue())
1873 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1874 DAG.getConstant(0, VT), N0);
1875 // If we know the sign bits of both operands are zero, strength reduce to a
1876 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
1877 if (!VT.isVector()) {
1878 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
1879 return DAG.getNode(ISD::UDIV, N->getDebugLoc(), N1.getValueType(),
1882 // fold (sdiv X, pow2) -> simple ops after legalize
1883 if (N1C && !N1C->isNullValue() &&
1884 (N1C->getAPIntValue().isPowerOf2() ||
1885 (-N1C->getAPIntValue()).isPowerOf2())) {
1886 // If dividing by powers of two is cheap, then don't perform the following
1888 if (TLI.isPow2DivCheap())
1891 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
1893 // Splat the sign bit into the register
1894 SDValue SGN = DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0,
1895 DAG.getConstant(VT.getSizeInBits()-1,
1896 getShiftAmountTy(N0.getValueType())));
1897 AddToWorkList(SGN.getNode());
1899 // Add (N0 < 0) ? abs2 - 1 : 0;
1900 SDValue SRL = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, SGN,
1901 DAG.getConstant(VT.getSizeInBits() - lg2,
1902 getShiftAmountTy(SGN.getValueType())));
1903 SDValue ADD = DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0, SRL);
1904 AddToWorkList(SRL.getNode());
1905 AddToWorkList(ADD.getNode()); // Divide by pow2
1906 SDValue SRA = DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, ADD,
1907 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType())));
1909 // If we're dividing by a positive value, we're done. Otherwise, we must
1910 // negate the result.
1911 if (N1C->getAPIntValue().isNonNegative())
1914 AddToWorkList(SRA.getNode());
1915 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1916 DAG.getConstant(0, VT), SRA);
1919 // if integer divide is expensive and we satisfy the requirements, emit an
1920 // alternate sequence.
1921 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
1922 SDValue Op = BuildSDIV(N);
1923 if (Op.getNode()) return Op;
1927 if (N0.getOpcode() == ISD::UNDEF)
1928 return DAG.getConstant(0, VT);
1929 // X / undef -> undef
1930 if (N1.getOpcode() == ISD::UNDEF)
1936 SDValue DAGCombiner::visitUDIV(SDNode *N) {
1937 SDValue N0 = N->getOperand(0);
1938 SDValue N1 = N->getOperand(1);
1939 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1940 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1941 EVT VT = N->getValueType(0);
1944 if (VT.isVector()) {
1945 SDValue FoldedVOp = SimplifyVBinOp(N);
1946 if (FoldedVOp.getNode()) return FoldedVOp;
1949 // fold (udiv c1, c2) -> c1/c2
1950 if (N0C && N1C && !N1C->isNullValue())
1951 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
1952 // fold (udiv x, (1 << c)) -> x >>u c
1953 if (N1C && N1C->getAPIntValue().isPowerOf2())
1954 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0,
1955 DAG.getConstant(N1C->getAPIntValue().logBase2(),
1956 getShiftAmountTy(N0.getValueType())));
1957 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
1958 if (N1.getOpcode() == ISD::SHL) {
1959 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
1960 if (SHC->getAPIntValue().isPowerOf2()) {
1961 EVT ADDVT = N1.getOperand(1).getValueType();
1962 SDValue Add = DAG.getNode(ISD::ADD, N->getDebugLoc(), ADDVT,
1964 DAG.getConstant(SHC->getAPIntValue()
1967 AddToWorkList(Add.getNode());
1968 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0, Add);
1972 // fold (udiv x, c) -> alternate
1973 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
1974 SDValue Op = BuildUDIV(N);
1975 if (Op.getNode()) return Op;
1979 if (N0.getOpcode() == ISD::UNDEF)
1980 return DAG.getConstant(0, VT);
1981 // X / undef -> undef
1982 if (N1.getOpcode() == ISD::UNDEF)
1988 SDValue DAGCombiner::visitSREM(SDNode *N) {
1989 SDValue N0 = N->getOperand(0);
1990 SDValue N1 = N->getOperand(1);
1991 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1992 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1993 EVT VT = N->getValueType(0);
1995 // fold (srem c1, c2) -> c1%c2
1996 if (N0C && N1C && !N1C->isNullValue())
1997 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
1998 // If we know the sign bits of both operands are zero, strength reduce to a
1999 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
2000 if (!VT.isVector()) {
2001 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
2002 return DAG.getNode(ISD::UREM, N->getDebugLoc(), VT, N0, N1);
2005 // If X/C can be simplified by the division-by-constant logic, lower
2006 // X%C to the equivalent of X-X/C*C.
2007 if (N1C && !N1C->isNullValue()) {
2008 SDValue Div = DAG.getNode(ISD::SDIV, N->getDebugLoc(), VT, N0, N1);
2009 AddToWorkList(Div.getNode());
2010 SDValue OptimizedDiv = combine(Div.getNode());
2011 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2012 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
2014 SDValue Sub = DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, Mul);
2015 AddToWorkList(Mul.getNode());
2021 if (N0.getOpcode() == ISD::UNDEF)
2022 return DAG.getConstant(0, VT);
2023 // X % undef -> undef
2024 if (N1.getOpcode() == ISD::UNDEF)
2030 SDValue DAGCombiner::visitUREM(SDNode *N) {
2031 SDValue N0 = N->getOperand(0);
2032 SDValue N1 = N->getOperand(1);
2033 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2034 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2035 EVT VT = N->getValueType(0);
2037 // fold (urem c1, c2) -> c1%c2
2038 if (N0C && N1C && !N1C->isNullValue())
2039 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
2040 // fold (urem x, pow2) -> (and x, pow2-1)
2041 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
2042 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0,
2043 DAG.getConstant(N1C->getAPIntValue()-1,VT));
2044 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2045 if (N1.getOpcode() == ISD::SHL) {
2046 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2047 if (SHC->getAPIntValue().isPowerOf2()) {
2049 DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N1,
2050 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
2052 AddToWorkList(Add.getNode());
2053 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, Add);
2058 // If X/C can be simplified by the division-by-constant logic, lower
2059 // X%C to the equivalent of X-X/C*C.
2060 if (N1C && !N1C->isNullValue()) {
2061 SDValue Div = DAG.getNode(ISD::UDIV, N->getDebugLoc(), VT, N0, N1);
2062 AddToWorkList(Div.getNode());
2063 SDValue OptimizedDiv = combine(Div.getNode());
2064 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2065 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
2067 SDValue Sub = DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, Mul);
2068 AddToWorkList(Mul.getNode());
2074 if (N0.getOpcode() == ISD::UNDEF)
2075 return DAG.getConstant(0, VT);
2076 // X % undef -> undef
2077 if (N1.getOpcode() == ISD::UNDEF)
2083 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2084 SDValue N0 = N->getOperand(0);
2085 SDValue N1 = N->getOperand(1);
2086 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2087 EVT VT = N->getValueType(0);
2088 DebugLoc DL = N->getDebugLoc();
2090 // fold (mulhs x, 0) -> 0
2091 if (N1C && N1C->isNullValue())
2093 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2094 if (N1C && N1C->getAPIntValue() == 1)
2095 return DAG.getNode(ISD::SRA, N->getDebugLoc(), N0.getValueType(), N0,
2096 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2097 getShiftAmountTy(N0.getValueType())));
2098 // fold (mulhs x, undef) -> 0
2099 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2100 return DAG.getConstant(0, VT);
2102 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2104 if (VT.isSimple() && !VT.isVector()) {
2105 MVT Simple = VT.getSimpleVT();
2106 unsigned SimpleSize = Simple.getSizeInBits();
2107 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2108 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2109 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2110 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2111 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2112 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2113 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2114 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2121 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2122 SDValue N0 = N->getOperand(0);
2123 SDValue N1 = N->getOperand(1);
2124 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2125 EVT VT = N->getValueType(0);
2126 DebugLoc DL = N->getDebugLoc();
2128 // fold (mulhu x, 0) -> 0
2129 if (N1C && N1C->isNullValue())
2131 // fold (mulhu x, 1) -> 0
2132 if (N1C && N1C->getAPIntValue() == 1)
2133 return DAG.getConstant(0, N0.getValueType());
2134 // fold (mulhu x, undef) -> 0
2135 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2136 return DAG.getConstant(0, VT);
2138 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2140 if (VT.isSimple() && !VT.isVector()) {
2141 MVT Simple = VT.getSimpleVT();
2142 unsigned SimpleSize = Simple.getSizeInBits();
2143 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2144 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2145 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2146 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2147 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2148 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2149 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2150 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2157 /// SimplifyNodeWithTwoResults - Perform optimizations common to nodes that
2158 /// compute two values. LoOp and HiOp give the opcodes for the two computations
2159 /// that are being performed. Return true if a simplification was made.
2161 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2163 // If the high half is not needed, just compute the low half.
2164 bool HiExists = N->hasAnyUseOfValue(1);
2166 (!LegalOperations ||
2167 TLI.isOperationLegal(LoOp, N->getValueType(0)))) {
2168 SDValue Res = DAG.getNode(LoOp, N->getDebugLoc(), N->getValueType(0),
2169 N->op_begin(), N->getNumOperands());
2170 return CombineTo(N, Res, Res);
2173 // If the low half is not needed, just compute the high half.
2174 bool LoExists = N->hasAnyUseOfValue(0);
2176 (!LegalOperations ||
2177 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2178 SDValue Res = DAG.getNode(HiOp, N->getDebugLoc(), N->getValueType(1),
2179 N->op_begin(), N->getNumOperands());
2180 return CombineTo(N, Res, Res);
2183 // If both halves are used, return as it is.
2184 if (LoExists && HiExists)
2187 // If the two computed results can be simplified separately, separate them.
2189 SDValue Lo = DAG.getNode(LoOp, N->getDebugLoc(), N->getValueType(0),
2190 N->op_begin(), N->getNumOperands());
2191 AddToWorkList(Lo.getNode());
2192 SDValue LoOpt = combine(Lo.getNode());
2193 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2194 (!LegalOperations ||
2195 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2196 return CombineTo(N, LoOpt, LoOpt);
2200 SDValue Hi = DAG.getNode(HiOp, N->getDebugLoc(), N->getValueType(1),
2201 N->op_begin(), N->getNumOperands());
2202 AddToWorkList(Hi.getNode());
2203 SDValue HiOpt = combine(Hi.getNode());
2204 if (HiOpt.getNode() && HiOpt != Hi &&
2205 (!LegalOperations ||
2206 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2207 return CombineTo(N, HiOpt, HiOpt);
2213 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2214 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2215 if (Res.getNode()) return Res;
2217 EVT VT = N->getValueType(0);
2218 DebugLoc DL = N->getDebugLoc();
2220 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2222 if (VT.isSimple() && !VT.isVector()) {
2223 MVT Simple = VT.getSimpleVT();
2224 unsigned SimpleSize = Simple.getSizeInBits();
2225 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2226 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2227 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2228 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2229 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2230 // Compute the high part as N1.
2231 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2232 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2233 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2234 // Compute the low part as N0.
2235 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2236 return CombineTo(N, Lo, Hi);
2243 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2244 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2245 if (Res.getNode()) return Res;
2247 EVT VT = N->getValueType(0);
2248 DebugLoc DL = N->getDebugLoc();
2250 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2252 if (VT.isSimple() && !VT.isVector()) {
2253 MVT Simple = VT.getSimpleVT();
2254 unsigned SimpleSize = Simple.getSizeInBits();
2255 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2256 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2257 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2258 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2259 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2260 // Compute the high part as N1.
2261 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2262 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2263 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2264 // Compute the low part as N0.
2265 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2266 return CombineTo(N, Lo, Hi);
2273 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2274 // (smulo x, 2) -> (saddo x, x)
2275 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2276 if (C2->getAPIntValue() == 2)
2277 return DAG.getNode(ISD::SADDO, N->getDebugLoc(), N->getVTList(),
2278 N->getOperand(0), N->getOperand(0));
2283 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2284 // (umulo x, 2) -> (uaddo x, x)
2285 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2286 if (C2->getAPIntValue() == 2)
2287 return DAG.getNode(ISD::UADDO, N->getDebugLoc(), N->getVTList(),
2288 N->getOperand(0), N->getOperand(0));
2293 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2294 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2295 if (Res.getNode()) return Res;
2300 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2301 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2302 if (Res.getNode()) return Res;
2307 /// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with
2308 /// two operands of the same opcode, try to simplify it.
2309 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2310 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2311 EVT VT = N0.getValueType();
2312 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2314 // Bail early if none of these transforms apply.
2315 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2317 // For each of OP in AND/OR/XOR:
2318 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2319 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2320 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2321 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2323 // do not sink logical op inside of a vector extend, since it may combine
2325 EVT Op0VT = N0.getOperand(0).getValueType();
2326 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2327 N0.getOpcode() == ISD::SIGN_EXTEND ||
2328 // Avoid infinite looping with PromoteIntBinOp.
2329 (N0.getOpcode() == ISD::ANY_EXTEND &&
2330 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2331 (N0.getOpcode() == ISD::TRUNCATE &&
2332 (!TLI.isZExtFree(VT, Op0VT) ||
2333 !TLI.isTruncateFree(Op0VT, VT)) &&
2334 TLI.isTypeLegal(Op0VT))) &&
2336 Op0VT == N1.getOperand(0).getValueType() &&
2337 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2338 SDValue ORNode = DAG.getNode(N->getOpcode(), N0.getDebugLoc(),
2339 N0.getOperand(0).getValueType(),
2340 N0.getOperand(0), N1.getOperand(0));
2341 AddToWorkList(ORNode.getNode());
2342 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, ORNode);
2345 // For each of OP in SHL/SRL/SRA/AND...
2346 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2347 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2348 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2349 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2350 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2351 N0.getOperand(1) == N1.getOperand(1)) {
2352 SDValue ORNode = DAG.getNode(N->getOpcode(), N0.getDebugLoc(),
2353 N0.getOperand(0).getValueType(),
2354 N0.getOperand(0), N1.getOperand(0));
2355 AddToWorkList(ORNode.getNode());
2356 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
2357 ORNode, N0.getOperand(1));
2360 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2361 // Only perform this optimization after type legalization and before
2362 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2363 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2364 // we don't want to undo this promotion.
2365 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2367 if ((N0.getOpcode() == ISD::BITCAST ||
2368 N0.getOpcode() == ISD::SCALAR_TO_VECTOR) &&
2369 Level == AfterLegalizeTypes) {
2370 SDValue In0 = N0.getOperand(0);
2371 SDValue In1 = N1.getOperand(0);
2372 EVT In0Ty = In0.getValueType();
2373 EVT In1Ty = In1.getValueType();
2374 DebugLoc DL = N->getDebugLoc();
2375 // If both incoming values are integers, and the original types are the
2377 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2378 SDValue Op = DAG.getNode(N->getOpcode(), DL, In0Ty, In0, In1);
2379 SDValue BC = DAG.getNode(N0.getOpcode(), DL, VT, Op);
2380 AddToWorkList(Op.getNode());
2385 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2386 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2387 // If both shuffles use the same mask, and both shuffle within a single
2388 // vector, then it is worthwhile to move the swizzle after the operation.
2389 // The type-legalizer generates this pattern when loading illegal
2390 // vector types from memory. In many cases this allows additional shuffle
2392 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
2393 N0.getOperand(1).getOpcode() == ISD::UNDEF &&
2394 N1.getOperand(1).getOpcode() == ISD::UNDEF) {
2395 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2396 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2398 assert(N0.getOperand(0).getValueType() == N1.getOperand(1).getValueType() &&
2399 "Inputs to shuffles are not the same type");
2401 unsigned NumElts = VT.getVectorNumElements();
2403 // Check that both shuffles use the same mask. The masks are known to be of
2404 // the same length because the result vector type is the same.
2405 bool SameMask = true;
2406 for (unsigned i = 0; i != NumElts; ++i) {
2407 int Idx0 = SVN0->getMaskElt(i);
2408 int Idx1 = SVN1->getMaskElt(i);
2416 SDValue Op = DAG.getNode(N->getOpcode(), N->getDebugLoc(), VT,
2417 N0.getOperand(0), N1.getOperand(0));
2418 AddToWorkList(Op.getNode());
2419 return DAG.getVectorShuffle(VT, N->getDebugLoc(), Op,
2420 DAG.getUNDEF(VT), &SVN0->getMask()[0]);
2427 SDValue DAGCombiner::visitAND(SDNode *N) {
2428 SDValue N0 = N->getOperand(0);
2429 SDValue N1 = N->getOperand(1);
2430 SDValue LL, LR, RL, RR, CC0, CC1;
2431 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2432 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2433 EVT VT = N1.getValueType();
2434 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2437 if (VT.isVector()) {
2438 SDValue FoldedVOp = SimplifyVBinOp(N);
2439 if (FoldedVOp.getNode()) return FoldedVOp;
2441 // fold (and x, 0) -> 0, vector edition
2442 if (ISD::isBuildVectorAllZeros(N0.getNode()))
2444 if (ISD::isBuildVectorAllZeros(N1.getNode()))
2447 // fold (and x, -1) -> x, vector edition
2448 if (ISD::isBuildVectorAllOnes(N0.getNode()))
2450 if (ISD::isBuildVectorAllOnes(N1.getNode()))
2454 // fold (and x, undef) -> 0
2455 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2456 return DAG.getConstant(0, VT);
2457 // fold (and c1, c2) -> c1&c2
2459 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2460 // canonicalize constant to RHS
2462 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N1, N0);
2463 // fold (and x, -1) -> x
2464 if (N1C && N1C->isAllOnesValue())
2466 // if (and x, c) is known to be zero, return 0
2467 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2468 APInt::getAllOnesValue(BitWidth)))
2469 return DAG.getConstant(0, VT);
2471 SDValue RAND = ReassociateOps(ISD::AND, N->getDebugLoc(), N0, N1);
2472 if (RAND.getNode() != 0)
2474 // fold (and (or x, C), D) -> D if (C & D) == D
2475 if (N1C && N0.getOpcode() == ISD::OR)
2476 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2477 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2479 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2480 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2481 SDValue N0Op0 = N0.getOperand(0);
2482 APInt Mask = ~N1C->getAPIntValue();
2483 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2484 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2485 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(),
2486 N0.getValueType(), N0Op0);
2488 // Replace uses of the AND with uses of the Zero extend node.
2491 // We actually want to replace all uses of the any_extend with the
2492 // zero_extend, to avoid duplicating things. This will later cause this
2493 // AND to be folded.
2494 CombineTo(N0.getNode(), Zext);
2495 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2498 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2499 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2500 // already be zero by virtue of the width of the base type of the load.
2502 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2504 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2505 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2506 N0.getOpcode() == ISD::LOAD) {
2507 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2508 N0 : N0.getOperand(0) );
2510 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2511 // This can be a pure constant or a vector splat, in which case we treat the
2512 // vector as a scalar and use the splat value.
2513 APInt Constant = APInt::getNullValue(1);
2514 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2515 Constant = C->getAPIntValue();
2516 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2517 APInt SplatValue, SplatUndef;
2518 unsigned SplatBitSize;
2520 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2521 SplatBitSize, HasAnyUndefs);
2523 // Undef bits can contribute to a possible optimisation if set, so
2525 SplatValue |= SplatUndef;
2527 // The splat value may be something like "0x00FFFFFF", which means 0 for
2528 // the first vector value and FF for the rest, repeating. We need a mask
2529 // that will apply equally to all members of the vector, so AND all the
2530 // lanes of the constant together.
2531 EVT VT = Vector->getValueType(0);
2532 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2534 // If the splat value has been compressed to a bitlength lower
2535 // than the size of the vector lane, we need to re-expand it to
2537 if (BitWidth > SplatBitSize)
2538 for (SplatValue = SplatValue.zextOrTrunc(BitWidth);
2539 SplatBitSize < BitWidth;
2540 SplatBitSize = SplatBitSize * 2)
2541 SplatValue |= SplatValue.shl(SplatBitSize);
2543 Constant = APInt::getAllOnesValue(BitWidth);
2544 for (unsigned i = 0, n = SplatBitSize/BitWidth; i < n; ++i)
2545 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
2549 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
2550 // actually legal and isn't going to get expanded, else this is a false
2552 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
2553 Load->getMemoryVT());
2555 // Resize the constant to the same size as the original memory access before
2556 // extension. If it is still the AllOnesValue then this AND is completely
2559 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
2562 switch (Load->getExtensionType()) {
2563 default: B = false; break;
2564 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
2566 case ISD::NON_EXTLOAD: B = true; break;
2569 if (B && Constant.isAllOnesValue()) {
2570 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
2571 // preserve semantics once we get rid of the AND.
2572 SDValue NewLoad(Load, 0);
2573 if (Load->getExtensionType() == ISD::EXTLOAD) {
2574 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
2575 Load->getValueType(0), Load->getDebugLoc(),
2576 Load->getChain(), Load->getBasePtr(),
2577 Load->getOffset(), Load->getMemoryVT(),
2578 Load->getMemOperand());
2579 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
2580 if (Load->getNumValues() == 3) {
2581 // PRE/POST_INC loads have 3 values.
2582 SDValue To[] = { NewLoad.getValue(0), NewLoad.getValue(1),
2583 NewLoad.getValue(2) };
2584 CombineTo(Load, To, 3, true);
2586 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
2590 // Fold the AND away, taking care not to fold to the old load node if we
2592 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
2594 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2597 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2598 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2599 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2600 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2602 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2603 LL.getValueType().isInteger()) {
2604 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2605 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
2606 SDValue ORNode = DAG.getNode(ISD::OR, N0.getDebugLoc(),
2607 LR.getValueType(), LL, RL);
2608 AddToWorkList(ORNode.getNode());
2609 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
2611 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2612 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2613 SDValue ANDNode = DAG.getNode(ISD::AND, N0.getDebugLoc(),
2614 LR.getValueType(), LL, RL);
2615 AddToWorkList(ANDNode.getNode());
2616 return DAG.getSetCC(N->getDebugLoc(), VT, ANDNode, LR, Op1);
2618 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2619 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2620 SDValue ORNode = DAG.getNode(ISD::OR, N0.getDebugLoc(),
2621 LR.getValueType(), LL, RL);
2622 AddToWorkList(ORNode.getNode());
2623 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
2626 // canonicalize equivalent to ll == rl
2627 if (LL == RR && LR == RL) {
2628 Op1 = ISD::getSetCCSwappedOperands(Op1);
2631 if (LL == RL && LR == RR) {
2632 bool isInteger = LL.getValueType().isInteger();
2633 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2634 if (Result != ISD::SETCC_INVALID &&
2635 (!LegalOperations ||
2636 TLI.isCondCodeLegal(Result, LL.getSimpleValueType())))
2637 return DAG.getSetCC(N->getDebugLoc(), N0.getValueType(),
2642 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
2643 if (N0.getOpcode() == N1.getOpcode()) {
2644 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2645 if (Tmp.getNode()) return Tmp;
2648 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
2649 // fold (and (sra)) -> (and (srl)) when possible.
2650 if (!VT.isVector() &&
2651 SimplifyDemandedBits(SDValue(N, 0)))
2652 return SDValue(N, 0);
2654 // fold (zext_inreg (extload x)) -> (zextload x)
2655 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
2656 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2657 EVT MemVT = LN0->getMemoryVT();
2658 // If we zero all the possible extended bits, then we can turn this into
2659 // a zextload if we are running before legalize or the operation is legal.
2660 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2661 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2662 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2663 ((!LegalOperations && !LN0->isVolatile()) ||
2664 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2665 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
2666 LN0->getChain(), LN0->getBasePtr(),
2667 LN0->getPointerInfo(), MemVT,
2668 LN0->isVolatile(), LN0->isNonTemporal(),
2669 LN0->getAlignment());
2671 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2672 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2675 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
2676 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
2678 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2679 EVT MemVT = LN0->getMemoryVT();
2680 // If we zero all the possible extended bits, then we can turn this into
2681 // a zextload if we are running before legalize or the operation is legal.
2682 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2683 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2684 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2685 ((!LegalOperations && !LN0->isVolatile()) ||
2686 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2687 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
2689 LN0->getBasePtr(), LN0->getPointerInfo(),
2691 LN0->isVolatile(), LN0->isNonTemporal(),
2692 LN0->getAlignment());
2694 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2695 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2699 // fold (and (load x), 255) -> (zextload x, i8)
2700 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2701 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2702 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2703 (N0.getOpcode() == ISD::ANY_EXTEND &&
2704 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2705 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2706 LoadSDNode *LN0 = HasAnyExt
2707 ? cast<LoadSDNode>(N0.getOperand(0))
2708 : cast<LoadSDNode>(N0);
2709 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2710 LN0->isUnindexed() && N0.hasOneUse() && LN0->hasOneUse()) {
2711 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2712 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2713 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2714 EVT LoadedVT = LN0->getMemoryVT();
2716 if (ExtVT == LoadedVT &&
2717 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2718 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2721 DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy,
2722 LN0->getChain(), LN0->getBasePtr(),
2723 LN0->getPointerInfo(),
2724 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2725 LN0->getAlignment());
2727 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
2728 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2731 // Do not change the width of a volatile load.
2732 // Do not generate loads of non-round integer types since these can
2733 // be expensive (and would be wrong if the type is not byte sized).
2734 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
2735 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2736 EVT PtrType = LN0->getOperand(1).getValueType();
2738 unsigned Alignment = LN0->getAlignment();
2739 SDValue NewPtr = LN0->getBasePtr();
2741 // For big endian targets, we need to add an offset to the pointer
2742 // to load the correct bytes. For little endian systems, we merely
2743 // need to read fewer bytes from the same pointer.
2744 if (TLI.isBigEndian()) {
2745 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
2746 unsigned EVTStoreBytes = ExtVT.getStoreSize();
2747 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
2748 NewPtr = DAG.getNode(ISD::ADD, LN0->getDebugLoc(), PtrType,
2749 NewPtr, DAG.getConstant(PtrOff, PtrType));
2750 Alignment = MinAlign(Alignment, PtrOff);
2753 AddToWorkList(NewPtr.getNode());
2755 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2757 DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy,
2758 LN0->getChain(), NewPtr,
2759 LN0->getPointerInfo(),
2760 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2763 CombineTo(LN0, Load, Load.getValue(1));
2764 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2770 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
2771 VT.getSizeInBits() <= 64) {
2772 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2773 APInt ADDC = ADDI->getAPIntValue();
2774 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2775 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
2776 // immediate for an add, but it is legal if its top c2 bits are set,
2777 // transform the ADD so the immediate doesn't need to be materialized
2779 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
2780 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
2781 SRLI->getZExtValue());
2782 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
2784 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2786 DAG.getNode(ISD::ADD, N0.getDebugLoc(), VT,
2787 N0.getOperand(0), DAG.getConstant(ADDC, VT));
2788 CombineTo(N0.getNode(), NewAdd);
2789 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2800 /// MatchBSwapHWord - Match (a >> 8) | (a << 8) as (bswap a) >> 16
2802 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
2803 bool DemandHighBits) {
2804 if (!LegalOperations)
2807 EVT VT = N->getValueType(0);
2808 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
2810 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
2813 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
2814 bool LookPassAnd0 = false;
2815 bool LookPassAnd1 = false;
2816 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
2818 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
2820 if (N0.getOpcode() == ISD::AND) {
2821 if (!N0.getNode()->hasOneUse())
2823 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2824 if (!N01C || N01C->getZExtValue() != 0xFF00)
2826 N0 = N0.getOperand(0);
2827 LookPassAnd0 = true;
2830 if (N1.getOpcode() == ISD::AND) {
2831 if (!N1.getNode()->hasOneUse())
2833 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
2834 if (!N11C || N11C->getZExtValue() != 0xFF)
2836 N1 = N1.getOperand(0);
2837 LookPassAnd1 = true;
2840 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
2842 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
2844 if (!N0.getNode()->hasOneUse() ||
2845 !N1.getNode()->hasOneUse())
2848 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2849 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
2852 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
2855 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
2856 SDValue N00 = N0->getOperand(0);
2857 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
2858 if (!N00.getNode()->hasOneUse())
2860 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
2861 if (!N001C || N001C->getZExtValue() != 0xFF)
2863 N00 = N00.getOperand(0);
2864 LookPassAnd0 = true;
2867 SDValue N10 = N1->getOperand(0);
2868 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
2869 if (!N10.getNode()->hasOneUse())
2871 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
2872 if (!N101C || N101C->getZExtValue() != 0xFF00)
2874 N10 = N10.getOperand(0);
2875 LookPassAnd1 = true;
2881 // Make sure everything beyond the low halfword is zero since the SRL 16
2882 // will clear the top bits.
2883 unsigned OpSizeInBits = VT.getSizeInBits();
2884 if (DemandHighBits && OpSizeInBits > 16 &&
2885 (!LookPassAnd0 || !LookPassAnd1) &&
2886 !DAG.MaskedValueIsZero(N10, APInt::getHighBitsSet(OpSizeInBits, 16)))
2889 SDValue Res = DAG.getNode(ISD::BSWAP, N->getDebugLoc(), VT, N00);
2890 if (OpSizeInBits > 16)
2891 Res = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, Res,
2892 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
2896 /// isBSwapHWordElement - Return true if the specified node is an element
2897 /// that makes up a 32-bit packed halfword byteswap. i.e.
2898 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
2899 static bool isBSwapHWordElement(SDValue N, SmallVector<SDNode*,4> &Parts) {
2900 if (!N.getNode()->hasOneUse())
2903 unsigned Opc = N.getOpcode();
2904 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
2907 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2912 switch (N1C->getZExtValue()) {
2915 case 0xFF: Num = 0; break;
2916 case 0xFF00: Num = 1; break;
2917 case 0xFF0000: Num = 2; break;
2918 case 0xFF000000: Num = 3; break;
2921 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
2922 SDValue N0 = N.getOperand(0);
2923 if (Opc == ISD::AND) {
2924 if (Num == 0 || Num == 2) {
2926 // (x >> 8) & 0xff0000
2927 if (N0.getOpcode() != ISD::SRL)
2929 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2930 if (!C || C->getZExtValue() != 8)
2933 // (x << 8) & 0xff00
2934 // (x << 8) & 0xff000000
2935 if (N0.getOpcode() != ISD::SHL)
2937 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2938 if (!C || C->getZExtValue() != 8)
2941 } else if (Opc == ISD::SHL) {
2943 // (x & 0xff0000) << 8
2944 if (Num != 0 && Num != 2)
2946 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2947 if (!C || C->getZExtValue() != 8)
2949 } else { // Opc == ISD::SRL
2950 // (x & 0xff00) >> 8
2951 // (x & 0xff000000) >> 8
2952 if (Num != 1 && Num != 3)
2954 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2955 if (!C || C->getZExtValue() != 8)
2962 Parts[Num] = N0.getOperand(0).getNode();
2966 /// MatchBSwapHWord - Match a 32-bit packed halfword bswap. That is
2967 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
2968 /// => (rotl (bswap x), 16)
2969 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
2970 if (!LegalOperations)
2973 EVT VT = N->getValueType(0);
2976 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
2979 SmallVector<SDNode*,4> Parts(4, (SDNode*)0);
2981 // (or (or (and), (and)), (or (and), (and)))
2982 // (or (or (or (and), (and)), (and)), (and))
2983 if (N0.getOpcode() != ISD::OR)
2985 SDValue N00 = N0.getOperand(0);
2986 SDValue N01 = N0.getOperand(1);
2988 if (N1.getOpcode() == ISD::OR) {
2989 // (or (or (and), (and)), (or (and), (and)))
2990 SDValue N000 = N00.getOperand(0);
2991 if (!isBSwapHWordElement(N000, Parts))
2994 SDValue N001 = N00.getOperand(1);
2995 if (!isBSwapHWordElement(N001, Parts))
2997 SDValue N010 = N01.getOperand(0);
2998 if (!isBSwapHWordElement(N010, Parts))
3000 SDValue N011 = N01.getOperand(1);
3001 if (!isBSwapHWordElement(N011, Parts))
3004 // (or (or (or (and), (and)), (and)), (and))
3005 if (!isBSwapHWordElement(N1, Parts))
3007 if (!isBSwapHWordElement(N01, Parts))
3009 if (N00.getOpcode() != ISD::OR)
3011 SDValue N000 = N00.getOperand(0);
3012 if (!isBSwapHWordElement(N000, Parts))
3014 SDValue N001 = N00.getOperand(1);
3015 if (!isBSwapHWordElement(N001, Parts))
3019 // Make sure the parts are all coming from the same node.
3020 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
3023 SDValue BSwap = DAG.getNode(ISD::BSWAP, N->getDebugLoc(), VT,
3024 SDValue(Parts[0],0));
3026 // Result of the bswap should be rotated by 16. If it's not legal, than
3027 // do (x << 16) | (x >> 16).
3028 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
3029 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3030 return DAG.getNode(ISD::ROTL, N->getDebugLoc(), VT, BSwap, ShAmt);
3031 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3032 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, BSwap, ShAmt);
3033 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT,
3034 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, BSwap, ShAmt),
3035 DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, BSwap, ShAmt));
3038 SDValue DAGCombiner::visitOR(SDNode *N) {
3039 SDValue N0 = N->getOperand(0);
3040 SDValue N1 = N->getOperand(1);
3041 SDValue LL, LR, RL, RR, CC0, CC1;
3042 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3043 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3044 EVT VT = N1.getValueType();
3047 if (VT.isVector()) {
3048 SDValue FoldedVOp = SimplifyVBinOp(N);
3049 if (FoldedVOp.getNode()) return FoldedVOp;
3051 // fold (or x, 0) -> x, vector edition
3052 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3054 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3057 // fold (or x, -1) -> -1, vector edition
3058 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3060 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3064 // fold (or x, undef) -> -1
3065 if (!LegalOperations &&
3066 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3067 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3068 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
3070 // fold (or c1, c2) -> c1|c2
3072 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
3073 // canonicalize constant to RHS
3075 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N1, N0);
3076 // fold (or x, 0) -> x
3077 if (N1C && N1C->isNullValue())
3079 // fold (or x, -1) -> -1
3080 if (N1C && N1C->isAllOnesValue())
3082 // fold (or x, c) -> c iff (x & ~c) == 0
3083 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3086 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3087 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
3088 if (BSwap.getNode() != 0)
3090 BSwap = MatchBSwapHWordLow(N, N0, N1);
3091 if (BSwap.getNode() != 0)
3095 SDValue ROR = ReassociateOps(ISD::OR, N->getDebugLoc(), N0, N1);
3096 if (ROR.getNode() != 0)
3098 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3099 // iff (c1 & c2) == 0.
3100 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3101 isa<ConstantSDNode>(N0.getOperand(1))) {
3102 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3103 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0)
3104 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
3105 DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
3106 N0.getOperand(0), N1),
3107 DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1));
3109 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3110 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3111 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3112 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3114 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3115 LL.getValueType().isInteger()) {
3116 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3117 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3118 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3119 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3120 SDValue ORNode = DAG.getNode(ISD::OR, LR.getDebugLoc(),
3121 LR.getValueType(), LL, RL);
3122 AddToWorkList(ORNode.getNode());
3123 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
3125 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3126 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3127 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3128 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3129 SDValue ANDNode = DAG.getNode(ISD::AND, LR.getDebugLoc(),
3130 LR.getValueType(), LL, RL);
3131 AddToWorkList(ANDNode.getNode());
3132 return DAG.getSetCC(N->getDebugLoc(), VT, ANDNode, LR, Op1);
3135 // canonicalize equivalent to ll == rl
3136 if (LL == RR && LR == RL) {
3137 Op1 = ISD::getSetCCSwappedOperands(Op1);
3140 if (LL == RL && LR == RR) {
3141 bool isInteger = LL.getValueType().isInteger();
3142 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3143 if (Result != ISD::SETCC_INVALID &&
3144 (!LegalOperations ||
3145 TLI.isCondCodeLegal(Result, LL.getSimpleValueType())))
3146 return DAG.getSetCC(N->getDebugLoc(), N0.getValueType(),
3151 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3152 if (N0.getOpcode() == N1.getOpcode()) {
3153 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3154 if (Tmp.getNode()) return Tmp;
3157 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3158 if (N0.getOpcode() == ISD::AND &&
3159 N1.getOpcode() == ISD::AND &&
3160 N0.getOperand(1).getOpcode() == ISD::Constant &&
3161 N1.getOperand(1).getOpcode() == ISD::Constant &&
3162 // Don't increase # computations.
3163 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3164 // We can only do this xform if we know that bits from X that are set in C2
3165 // but not in C1 are already zero. Likewise for Y.
3166 const APInt &LHSMask =
3167 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3168 const APInt &RHSMask =
3169 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3171 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3172 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3173 SDValue X = DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
3174 N0.getOperand(0), N1.getOperand(0));
3175 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, X,
3176 DAG.getConstant(LHSMask | RHSMask, VT));
3180 // See if this is some rotate idiom.
3181 if (SDNode *Rot = MatchRotate(N0, N1, N->getDebugLoc()))
3182 return SDValue(Rot, 0);
3184 // Simplify the operands using demanded-bits information.
3185 if (!VT.isVector() &&
3186 SimplifyDemandedBits(SDValue(N, 0)))
3187 return SDValue(N, 0);
3192 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
3193 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3194 if (Op.getOpcode() == ISD::AND) {
3195 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3196 Mask = Op.getOperand(1);
3197 Op = Op.getOperand(0);
3203 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3211 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3212 // idioms for rotate, and if the target supports rotation instructions, generate
3214 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, DebugLoc DL) {
3215 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3216 EVT VT = LHS.getValueType();
3217 if (!TLI.isTypeLegal(VT)) return 0;
3219 // The target must have at least one rotate flavor.
3220 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3221 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3222 if (!HasROTL && !HasROTR) return 0;
3224 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3225 SDValue LHSShift; // The shift.
3226 SDValue LHSMask; // AND value if any.
3227 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3228 return 0; // Not part of a rotate.
3230 SDValue RHSShift; // The shift.
3231 SDValue RHSMask; // AND value if any.
3232 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3233 return 0; // Not part of a rotate.
3235 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3236 return 0; // Not shifting the same value.
3238 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3239 return 0; // Shifts must disagree.
3241 // Canonicalize shl to left side in a shl/srl pair.
3242 if (RHSShift.getOpcode() == ISD::SHL) {
3243 std::swap(LHS, RHS);
3244 std::swap(LHSShift, RHSShift);
3245 std::swap(LHSMask , RHSMask );
3248 unsigned OpSizeInBits = VT.getSizeInBits();
3249 SDValue LHSShiftArg = LHSShift.getOperand(0);
3250 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3251 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3253 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3254 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3255 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3256 RHSShiftAmt.getOpcode() == ISD::Constant) {
3257 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3258 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3259 if ((LShVal + RShVal) != OpSizeInBits)
3262 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3263 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
3265 // If there is an AND of either shifted operand, apply it to the result.
3266 if (LHSMask.getNode() || RHSMask.getNode()) {
3267 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3269 if (LHSMask.getNode()) {
3270 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3271 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3273 if (RHSMask.getNode()) {
3274 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3275 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3278 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3281 return Rot.getNode();
3284 // If there is a mask here, and we have a variable shift, we can't be sure
3285 // that we're masking out the right stuff.
3286 if (LHSMask.getNode() || RHSMask.getNode())
3289 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y)
3290 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotr x, (sub 32, y))
3291 if (RHSShiftAmt.getOpcode() == ISD::SUB &&
3292 LHSShiftAmt == RHSShiftAmt.getOperand(1)) {
3293 if (ConstantSDNode *SUBC =
3294 dyn_cast<ConstantSDNode>(RHSShiftAmt.getOperand(0))) {
3295 if (SUBC->getAPIntValue() == OpSizeInBits) {
3296 return DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT, LHSShiftArg,
3297 HasROTL ? LHSShiftAmt : RHSShiftAmt).getNode();
3302 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y)
3303 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotl x, (sub 32, y))
3304 if (LHSShiftAmt.getOpcode() == ISD::SUB &&
3305 RHSShiftAmt == LHSShiftAmt.getOperand(1)) {
3306 if (ConstantSDNode *SUBC =
3307 dyn_cast<ConstantSDNode>(LHSShiftAmt.getOperand(0))) {
3308 if (SUBC->getAPIntValue() == OpSizeInBits) {
3309 return DAG.getNode(HasROTR ? ISD::ROTR : ISD::ROTL, DL, VT, LHSShiftArg,
3310 HasROTR ? RHSShiftAmt : LHSShiftAmt).getNode();
3315 // Look for sign/zext/any-extended or truncate cases:
3316 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3317 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3318 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3319 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3320 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3321 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3322 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3323 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3324 SDValue LExtOp0 = LHSShiftAmt.getOperand(0);
3325 SDValue RExtOp0 = RHSShiftAmt.getOperand(0);
3326 if (RExtOp0.getOpcode() == ISD::SUB &&
3327 RExtOp0.getOperand(1) == LExtOp0) {
3328 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
3330 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
3331 // (rotr x, (sub 32, y))
3332 if (ConstantSDNode *SUBC =
3333 dyn_cast<ConstantSDNode>(RExtOp0.getOperand(0))) {
3334 if (SUBC->getAPIntValue() == OpSizeInBits) {
3335 return DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3337 HasROTL ? LHSShiftAmt : RHSShiftAmt).getNode();
3340 } else if (LExtOp0.getOpcode() == ISD::SUB &&
3341 RExtOp0 == LExtOp0.getOperand(1)) {
3342 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
3344 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
3345 // (rotl x, (sub 32, y))
3346 if (ConstantSDNode *SUBC =
3347 dyn_cast<ConstantSDNode>(LExtOp0.getOperand(0))) {
3348 if (SUBC->getAPIntValue() == OpSizeInBits) {
3349 return DAG.getNode(HasROTR ? ISD::ROTR : ISD::ROTL, DL, VT,
3351 HasROTR ? RHSShiftAmt : LHSShiftAmt).getNode();
3360 SDValue DAGCombiner::visitXOR(SDNode *N) {
3361 SDValue N0 = N->getOperand(0);
3362 SDValue N1 = N->getOperand(1);
3363 SDValue LHS, RHS, CC;
3364 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3365 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3366 EVT VT = N0.getValueType();
3369 if (VT.isVector()) {
3370 SDValue FoldedVOp = SimplifyVBinOp(N);
3371 if (FoldedVOp.getNode()) return FoldedVOp;
3373 // fold (xor x, 0) -> x, vector edition
3374 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3376 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3380 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3381 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3382 return DAG.getConstant(0, VT);
3383 // fold (xor x, undef) -> undef
3384 if (N0.getOpcode() == ISD::UNDEF)
3386 if (N1.getOpcode() == ISD::UNDEF)
3388 // fold (xor c1, c2) -> c1^c2
3390 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3391 // canonicalize constant to RHS
3393 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0);
3394 // fold (xor x, 0) -> x
3395 if (N1C && N1C->isNullValue())
3398 SDValue RXOR = ReassociateOps(ISD::XOR, N->getDebugLoc(), N0, N1);
3399 if (RXOR.getNode() != 0)
3402 // fold !(x cc y) -> (x !cc y)
3403 if (N1C && N1C->getAPIntValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3404 bool isInt = LHS.getValueType().isInteger();
3405 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3408 if (!LegalOperations ||
3409 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) {
3410 switch (N0.getOpcode()) {
3412 llvm_unreachable("Unhandled SetCC Equivalent!");
3414 return DAG.getSetCC(N->getDebugLoc(), VT, LHS, RHS, NotCC);
3415 case ISD::SELECT_CC:
3416 return DAG.getSelectCC(N->getDebugLoc(), LHS, RHS, N0.getOperand(2),
3417 N0.getOperand(3), NotCC);
3422 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3423 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3424 N0.getNode()->hasOneUse() &&
3425 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3426 SDValue V = N0.getOperand(0);
3427 V = DAG.getNode(ISD::XOR, N0.getDebugLoc(), V.getValueType(), V,
3428 DAG.getConstant(1, V.getValueType()));
3429 AddToWorkList(V.getNode());
3430 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, V);
3433 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3434 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3435 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3436 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3437 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3438 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3439 LHS = DAG.getNode(ISD::XOR, LHS.getDebugLoc(), VT, LHS, N1); // LHS = ~LHS
3440 RHS = DAG.getNode(ISD::XOR, RHS.getDebugLoc(), VT, RHS, N1); // RHS = ~RHS
3441 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3442 return DAG.getNode(NewOpcode, N->getDebugLoc(), VT, LHS, RHS);
3445 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3446 if (N1C && N1C->isAllOnesValue() &&
3447 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3448 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3449 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3450 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3451 LHS = DAG.getNode(ISD::XOR, LHS.getDebugLoc(), VT, LHS, N1); // LHS = ~LHS
3452 RHS = DAG.getNode(ISD::XOR, RHS.getDebugLoc(), VT, RHS, N1); // RHS = ~RHS
3453 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3454 return DAG.getNode(NewOpcode, N->getDebugLoc(), VT, LHS, RHS);
3457 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3458 if (N1C && N0.getOpcode() == ISD::XOR) {
3459 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3460 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3462 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N0.getOperand(1),
3463 DAG.getConstant(N1C->getAPIntValue() ^
3464 N00C->getAPIntValue(), VT));
3466 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N0.getOperand(0),
3467 DAG.getConstant(N1C->getAPIntValue() ^
3468 N01C->getAPIntValue(), VT));
3470 // fold (xor x, x) -> 0
3472 return tryFoldToZero(N->getDebugLoc(), TLI, VT, DAG, LegalOperations);
3474 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
3475 if (N0.getOpcode() == N1.getOpcode()) {
3476 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3477 if (Tmp.getNode()) return Tmp;
3480 // Simplify the expression using non-local knowledge.
3481 if (!VT.isVector() &&
3482 SimplifyDemandedBits(SDValue(N, 0)))
3483 return SDValue(N, 0);
3488 /// visitShiftByConstant - Handle transforms common to the three shifts, when
3489 /// the shift amount is a constant.
3490 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, unsigned Amt) {
3491 SDNode *LHS = N->getOperand(0).getNode();
3492 if (!LHS->hasOneUse()) return SDValue();
3494 // We want to pull some binops through shifts, so that we have (and (shift))
3495 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
3496 // thing happens with address calculations, so it's important to canonicalize
3498 bool HighBitSet = false; // Can we transform this if the high bit is set?
3500 switch (LHS->getOpcode()) {
3501 default: return SDValue();
3504 HighBitSet = false; // We can only transform sra if the high bit is clear.
3507 HighBitSet = true; // We can only transform sra if the high bit is set.
3510 if (N->getOpcode() != ISD::SHL)
3511 return SDValue(); // only shl(add) not sr[al](add).
3512 HighBitSet = false; // We can only transform sra if the high bit is clear.
3516 // We require the RHS of the binop to be a constant as well.
3517 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
3518 if (!BinOpCst) return SDValue();
3520 // FIXME: disable this unless the input to the binop is a shift by a constant.
3521 // If it is not a shift, it pessimizes some common cases like:
3523 // void foo(int *X, int i) { X[i & 1235] = 1; }
3524 // int bar(int *X, int i) { return X[i & 255]; }
3525 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
3526 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
3527 BinOpLHSVal->getOpcode() != ISD::SRA &&
3528 BinOpLHSVal->getOpcode() != ISD::SRL) ||
3529 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
3532 EVT VT = N->getValueType(0);
3534 // If this is a signed shift right, and the high bit is modified by the
3535 // logical operation, do not perform the transformation. The highBitSet
3536 // boolean indicates the value of the high bit of the constant which would
3537 // cause it to be modified for this operation.
3538 if (N->getOpcode() == ISD::SRA) {
3539 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
3540 if (BinOpRHSSignSet != HighBitSet)
3544 // Fold the constants, shifting the binop RHS by the shift amount.
3545 SDValue NewRHS = DAG.getNode(N->getOpcode(), LHS->getOperand(1).getDebugLoc(),
3547 LHS->getOperand(1), N->getOperand(1));
3549 // Create the new shift.
3550 SDValue NewShift = DAG.getNode(N->getOpcode(),
3551 LHS->getOperand(0).getDebugLoc(),
3552 VT, LHS->getOperand(0), N->getOperand(1));
3554 // Create the new binop.
3555 return DAG.getNode(LHS->getOpcode(), N->getDebugLoc(), VT, NewShift, NewRHS);
3558 SDValue DAGCombiner::visitSHL(SDNode *N) {
3559 SDValue N0 = N->getOperand(0);
3560 SDValue N1 = N->getOperand(1);
3561 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3562 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3563 EVT VT = N0.getValueType();
3564 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3566 // fold (shl c1, c2) -> c1<<c2
3568 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
3569 // fold (shl 0, x) -> 0
3570 if (N0C && N0C->isNullValue())
3572 // fold (shl x, c >= size(x)) -> undef
3573 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3574 return DAG.getUNDEF(VT);
3575 // fold (shl x, 0) -> x
3576 if (N1C && N1C->isNullValue())
3578 // fold (shl undef, x) -> 0
3579 if (N0.getOpcode() == ISD::UNDEF)
3580 return DAG.getConstant(0, VT);
3581 // if (shl x, c) is known to be zero, return 0
3582 if (DAG.MaskedValueIsZero(SDValue(N, 0),
3583 APInt::getAllOnesValue(OpSizeInBits)))
3584 return DAG.getConstant(0, VT);
3585 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
3586 if (N1.getOpcode() == ISD::TRUNCATE &&
3587 N1.getOperand(0).getOpcode() == ISD::AND &&
3588 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3589 SDValue N101 = N1.getOperand(0).getOperand(1);
3590 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3591 EVT TruncVT = N1.getValueType();
3592 SDValue N100 = N1.getOperand(0).getOperand(0);
3593 APInt TruncC = N101C->getAPIntValue();
3594 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
3595 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
3596 DAG.getNode(ISD::AND, N->getDebugLoc(), TruncVT,
3597 DAG.getNode(ISD::TRUNCATE,
3600 DAG.getConstant(TruncC, TruncVT)));
3604 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3605 return SDValue(N, 0);
3607 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
3608 if (N1C && N0.getOpcode() == ISD::SHL &&
3609 N0.getOperand(1).getOpcode() == ISD::Constant) {
3610 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3611 uint64_t c2 = N1C->getZExtValue();
3612 if (c1 + c2 >= OpSizeInBits)
3613 return DAG.getConstant(0, VT);
3614 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0.getOperand(0),
3615 DAG.getConstant(c1 + c2, N1.getValueType()));
3618 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
3619 // For this to be valid, the second form must not preserve any of the bits
3620 // that are shifted out by the inner shift in the first form. This means
3621 // the outer shift size must be >= the number of bits added by the ext.
3622 // As a corollary, we don't care what kind of ext it is.
3623 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
3624 N0.getOpcode() == ISD::ANY_EXTEND ||
3625 N0.getOpcode() == ISD::SIGN_EXTEND) &&
3626 N0.getOperand(0).getOpcode() == ISD::SHL &&
3627 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
3629 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
3630 uint64_t c2 = N1C->getZExtValue();
3631 EVT InnerShiftVT = N0.getOperand(0).getValueType();
3632 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
3633 if (c2 >= OpSizeInBits - InnerShiftSize) {
3634 if (c1 + c2 >= OpSizeInBits)
3635 return DAG.getConstant(0, VT);
3636 return DAG.getNode(ISD::SHL, N0->getDebugLoc(), VT,
3637 DAG.getNode(N0.getOpcode(), N0->getDebugLoc(), VT,
3638 N0.getOperand(0)->getOperand(0)),
3639 DAG.getConstant(c1 + c2, N1.getValueType()));
3643 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
3644 // (and (srl x, (sub c1, c2), MASK)
3645 // Only fold this if the inner shift has no other uses -- if it does, folding
3646 // this will increase the total number of instructions.
3647 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse() &&
3648 N0.getOperand(1).getOpcode() == ISD::Constant) {
3649 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3650 if (c1 < VT.getSizeInBits()) {
3651 uint64_t c2 = N1C->getZExtValue();
3652 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
3653 VT.getSizeInBits() - c1);
3656 Mask = Mask.shl(c2-c1);
3657 Shift = DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0.getOperand(0),
3658 DAG.getConstant(c2-c1, N1.getValueType()));
3660 Mask = Mask.lshr(c1-c2);
3661 Shift = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0),
3662 DAG.getConstant(c1-c2, N1.getValueType()));
3664 return DAG.getNode(ISD::AND, N0.getDebugLoc(), VT, Shift,
3665 DAG.getConstant(Mask, VT));
3668 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
3669 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
3670 SDValue HiBitsMask =
3671 DAG.getConstant(APInt::getHighBitsSet(VT.getSizeInBits(),
3672 VT.getSizeInBits() -
3673 N1C->getZExtValue()),
3675 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
3680 SDValue NewSHL = visitShiftByConstant(N, N1C->getZExtValue());
3681 if (NewSHL.getNode())
3688 SDValue DAGCombiner::visitSRA(SDNode *N) {
3689 SDValue N0 = N->getOperand(0);
3690 SDValue N1 = N->getOperand(1);
3691 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3692 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3693 EVT VT = N0.getValueType();
3694 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3696 // fold (sra c1, c2) -> (sra c1, c2)
3698 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
3699 // fold (sra 0, x) -> 0
3700 if (N0C && N0C->isNullValue())
3702 // fold (sra -1, x) -> -1
3703 if (N0C && N0C->isAllOnesValue())
3705 // fold (sra x, (setge c, size(x))) -> undef
3706 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3707 return DAG.getUNDEF(VT);
3708 // fold (sra x, 0) -> x
3709 if (N1C && N1C->isNullValue())
3711 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
3713 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
3714 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
3715 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
3717 ExtVT = EVT::getVectorVT(*DAG.getContext(),
3718 ExtVT, VT.getVectorNumElements());
3719 if ((!LegalOperations ||
3720 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
3721 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
3722 N0.getOperand(0), DAG.getValueType(ExtVT));
3725 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
3726 if (N1C && N0.getOpcode() == ISD::SRA) {
3727 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3728 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
3729 if (Sum >= OpSizeInBits) Sum = OpSizeInBits-1;
3730 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0.getOperand(0),
3731 DAG.getConstant(Sum, N1C->getValueType(0)));
3735 // fold (sra (shl X, m), (sub result_size, n))
3736 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
3737 // result_size - n != m.
3738 // If truncate is free for the target sext(shl) is likely to result in better
3740 if (N0.getOpcode() == ISD::SHL) {
3741 // Get the two constanst of the shifts, CN0 = m, CN = n.
3742 const ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3744 // Determine what the truncate's result bitsize and type would be.
3746 EVT::getIntegerVT(*DAG.getContext(),
3747 OpSizeInBits - N1C->getZExtValue());
3748 // Determine the residual right-shift amount.
3749 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
3751 // If the shift is not a no-op (in which case this should be just a sign
3752 // extend already), the truncated to type is legal, sign_extend is legal
3753 // on that type, and the truncate to that type is both legal and free,
3754 // perform the transform.
3755 if ((ShiftAmt > 0) &&
3756 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
3757 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
3758 TLI.isTruncateFree(VT, TruncVT)) {
3760 SDValue Amt = DAG.getConstant(ShiftAmt,
3761 getShiftAmountTy(N0.getOperand(0).getValueType()));
3762 SDValue Shift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), VT,
3763 N0.getOperand(0), Amt);
3764 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), TruncVT,
3766 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(),
3767 N->getValueType(0), Trunc);
3772 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
3773 if (N1.getOpcode() == ISD::TRUNCATE &&
3774 N1.getOperand(0).getOpcode() == ISD::AND &&
3775 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3776 SDValue N101 = N1.getOperand(0).getOperand(1);
3777 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3778 EVT TruncVT = N1.getValueType();
3779 SDValue N100 = N1.getOperand(0).getOperand(0);
3780 APInt TruncC = N101C->getAPIntValue();
3781 TruncC = TruncC.trunc(TruncVT.getScalarType().getSizeInBits());
3782 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0,
3783 DAG.getNode(ISD::AND, N->getDebugLoc(),
3785 DAG.getNode(ISD::TRUNCATE,
3788 DAG.getConstant(TruncC, TruncVT)));
3792 // fold (sra (trunc (sr x, c1)), c2) -> (trunc (sra x, c1+c2))
3793 // if c1 is equal to the number of bits the trunc removes
3794 if (N0.getOpcode() == ISD::TRUNCATE &&
3795 (N0.getOperand(0).getOpcode() == ISD::SRL ||
3796 N0.getOperand(0).getOpcode() == ISD::SRA) &&
3797 N0.getOperand(0).hasOneUse() &&
3798 N0.getOperand(0).getOperand(1).hasOneUse() &&
3799 N1C && isa<ConstantSDNode>(N0.getOperand(0).getOperand(1))) {
3800 EVT LargeVT = N0.getOperand(0).getValueType();
3801 ConstantSDNode *LargeShiftAmt =
3802 cast<ConstantSDNode>(N0.getOperand(0).getOperand(1));
3804 if (LargeVT.getScalarType().getSizeInBits() - OpSizeInBits ==
3805 LargeShiftAmt->getZExtValue()) {
3807 DAG.getConstant(LargeShiftAmt->getZExtValue() + N1C->getZExtValue(),
3808 getShiftAmountTy(N0.getOperand(0).getOperand(0).getValueType()));
3809 SDValue SRA = DAG.getNode(ISD::SRA, N->getDebugLoc(), LargeVT,
3810 N0.getOperand(0).getOperand(0), Amt);
3811 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, SRA);
3815 // Simplify, based on bits shifted out of the LHS.
3816 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3817 return SDValue(N, 0);
3820 // If the sign bit is known to be zero, switch this to a SRL.
3821 if (DAG.SignBitIsZero(N0))
3822 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0, N1);
3825 SDValue NewSRA = visitShiftByConstant(N, N1C->getZExtValue());
3826 if (NewSRA.getNode())
3833 SDValue DAGCombiner::visitSRL(SDNode *N) {
3834 SDValue N0 = N->getOperand(0);
3835 SDValue N1 = N->getOperand(1);
3836 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3837 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3838 EVT VT = N0.getValueType();
3839 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3841 // fold (srl c1, c2) -> c1 >>u c2
3843 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
3844 // fold (srl 0, x) -> 0
3845 if (N0C && N0C->isNullValue())
3847 // fold (srl x, c >= size(x)) -> undef
3848 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3849 return DAG.getUNDEF(VT);
3850 // fold (srl x, 0) -> x
3851 if (N1C && N1C->isNullValue())
3853 // if (srl x, c) is known to be zero, return 0
3854 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
3855 APInt::getAllOnesValue(OpSizeInBits)))
3856 return DAG.getConstant(0, VT);
3858 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
3859 if (N1C && N0.getOpcode() == ISD::SRL &&
3860 N0.getOperand(1).getOpcode() == ISD::Constant) {
3861 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3862 uint64_t c2 = N1C->getZExtValue();
3863 if (c1 + c2 >= OpSizeInBits)
3864 return DAG.getConstant(0, VT);
3865 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0),
3866 DAG.getConstant(c1 + c2, N1.getValueType()));
3869 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
3870 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
3871 N0.getOperand(0).getOpcode() == ISD::SRL &&
3872 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
3874 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
3875 uint64_t c2 = N1C->getZExtValue();
3876 EVT InnerShiftVT = N0.getOperand(0).getValueType();
3877 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
3878 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
3879 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
3880 if (c1 + OpSizeInBits == InnerShiftSize) {
3881 if (c1 + c2 >= InnerShiftSize)
3882 return DAG.getConstant(0, VT);
3883 return DAG.getNode(ISD::TRUNCATE, N0->getDebugLoc(), VT,
3884 DAG.getNode(ISD::SRL, N0->getDebugLoc(), InnerShiftVT,
3885 N0.getOperand(0)->getOperand(0),
3886 DAG.getConstant(c1 + c2, ShiftCountVT)));
3890 // fold (srl (shl x, c), c) -> (and x, cst2)
3891 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1 &&
3892 N0.getValueSizeInBits() <= 64) {
3893 uint64_t ShAmt = N1C->getZExtValue()+64-N0.getValueSizeInBits();
3894 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
3895 DAG.getConstant(~0ULL >> ShAmt, VT));
3899 // fold (srl (anyextend x), c) -> (anyextend (srl x, c))
3900 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
3901 // Shifting in all undef bits?
3902 EVT SmallVT = N0.getOperand(0).getValueType();
3903 if (N1C->getZExtValue() >= SmallVT.getSizeInBits())
3904 return DAG.getUNDEF(VT);
3906 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
3907 uint64_t ShiftAmt = N1C->getZExtValue();
3908 SDValue SmallShift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), SmallVT,
3910 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
3911 AddToWorkList(SmallShift.getNode());
3912 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, SmallShift);
3916 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
3917 // bit, which is unmodified by sra.
3918 if (N1C && N1C->getZExtValue() + 1 == VT.getSizeInBits()) {
3919 if (N0.getOpcode() == ISD::SRA)
3920 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0), N1);
3923 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
3924 if (N1C && N0.getOpcode() == ISD::CTLZ &&
3925 N1C->getAPIntValue() == Log2_32(VT.getSizeInBits())) {
3926 APInt KnownZero, KnownOne;
3927 DAG.ComputeMaskedBits(N0.getOperand(0), KnownZero, KnownOne);
3929 // If any of the input bits are KnownOne, then the input couldn't be all
3930 // zeros, thus the result of the srl will always be zero.
3931 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
3933 // If all of the bits input the to ctlz node are known to be zero, then
3934 // the result of the ctlz is "32" and the result of the shift is one.
3935 APInt UnknownBits = ~KnownZero;
3936 if (UnknownBits == 0) return DAG.getConstant(1, VT);
3938 // Otherwise, check to see if there is exactly one bit input to the ctlz.
3939 if ((UnknownBits & (UnknownBits - 1)) == 0) {
3940 // Okay, we know that only that the single bit specified by UnknownBits
3941 // could be set on input to the CTLZ node. If this bit is set, the SRL
3942 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
3943 // to an SRL/XOR pair, which is likely to simplify more.
3944 unsigned ShAmt = UnknownBits.countTrailingZeros();
3945 SDValue Op = N0.getOperand(0);
3948 Op = DAG.getNode(ISD::SRL, N0.getDebugLoc(), VT, Op,
3949 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
3950 AddToWorkList(Op.getNode());
3953 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
3954 Op, DAG.getConstant(1, VT));
3958 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
3959 if (N1.getOpcode() == ISD::TRUNCATE &&
3960 N1.getOperand(0).getOpcode() == ISD::AND &&
3961 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3962 SDValue N101 = N1.getOperand(0).getOperand(1);
3963 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3964 EVT TruncVT = N1.getValueType();
3965 SDValue N100 = N1.getOperand(0).getOperand(0);
3966 APInt TruncC = N101C->getAPIntValue();
3967 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
3968 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0,
3969 DAG.getNode(ISD::AND, N->getDebugLoc(),
3971 DAG.getNode(ISD::TRUNCATE,
3974 DAG.getConstant(TruncC, TruncVT)));
3978 // fold operands of srl based on knowledge that the low bits are not
3980 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3981 return SDValue(N, 0);
3984 SDValue NewSRL = visitShiftByConstant(N, N1C->getZExtValue());
3985 if (NewSRL.getNode())
3989 // Attempt to convert a srl of a load into a narrower zero-extending load.
3990 SDValue NarrowLoad = ReduceLoadWidth(N);
3991 if (NarrowLoad.getNode())
3994 // Here is a common situation. We want to optimize:
3997 // %b = and i32 %a, 2
3998 // %c = srl i32 %b, 1
3999 // brcond i32 %c ...
4005 // %c = setcc eq %b, 0
4008 // However when after the source operand of SRL is optimized into AND, the SRL
4009 // itself may not be optimized further. Look for it and add the BRCOND into
4011 if (N->hasOneUse()) {
4012 SDNode *Use = *N->use_begin();
4013 if (Use->getOpcode() == ISD::BRCOND)
4015 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
4016 // Also look pass the truncate.
4017 Use = *Use->use_begin();
4018 if (Use->getOpcode() == ISD::BRCOND)
4026 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
4027 SDValue N0 = N->getOperand(0);
4028 EVT VT = N->getValueType(0);
4030 // fold (ctlz c1) -> c2
4031 if (isa<ConstantSDNode>(N0))
4032 return DAG.getNode(ISD::CTLZ, N->getDebugLoc(), VT, N0);
4036 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4037 SDValue N0 = N->getOperand(0);
4038 EVT VT = N->getValueType(0);
4040 // fold (ctlz_zero_undef c1) -> c2
4041 if (isa<ConstantSDNode>(N0))
4042 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, N->getDebugLoc(), VT, N0);
4046 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4047 SDValue N0 = N->getOperand(0);
4048 EVT VT = N->getValueType(0);
4050 // fold (cttz c1) -> c2
4051 if (isa<ConstantSDNode>(N0))
4052 return DAG.getNode(ISD::CTTZ, N->getDebugLoc(), VT, N0);
4056 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4057 SDValue N0 = N->getOperand(0);
4058 EVT VT = N->getValueType(0);
4060 // fold (cttz_zero_undef c1) -> c2
4061 if (isa<ConstantSDNode>(N0))
4062 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, N->getDebugLoc(), VT, N0);
4066 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4067 SDValue N0 = N->getOperand(0);
4068 EVT VT = N->getValueType(0);
4070 // fold (ctpop c1) -> c2
4071 if (isa<ConstantSDNode>(N0))
4072 return DAG.getNode(ISD::CTPOP, N->getDebugLoc(), VT, N0);
4076 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4077 SDValue N0 = N->getOperand(0);
4078 SDValue N1 = N->getOperand(1);
4079 SDValue N2 = N->getOperand(2);
4080 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4081 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4082 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
4083 EVT VT = N->getValueType(0);
4084 EVT VT0 = N0.getValueType();
4086 // fold (select C, X, X) -> X
4089 // fold (select true, X, Y) -> X
4090 if (N0C && !N0C->isNullValue())
4092 // fold (select false, X, Y) -> Y
4093 if (N0C && N0C->isNullValue())
4095 // fold (select C, 1, X) -> (or C, X)
4096 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4097 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N2);
4098 // fold (select C, 0, 1) -> (xor C, 1)
4099 if (VT.isInteger() &&
4102 TLI.getBooleanContents(false) ==
4103 TargetLowering::ZeroOrOneBooleanContent)) &&
4104 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4107 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT0,
4108 N0, DAG.getConstant(1, VT0));
4109 XORNode = DAG.getNode(ISD::XOR, N0.getDebugLoc(), VT0,
4110 N0, DAG.getConstant(1, VT0));
4111 AddToWorkList(XORNode.getNode());
4113 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, XORNode);
4114 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, XORNode);
4116 // fold (select C, 0, X) -> (and (not C), X)
4117 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4118 SDValue NOTNode = DAG.getNOT(N0.getDebugLoc(), N0, VT);
4119 AddToWorkList(NOTNode.getNode());
4120 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, NOTNode, N2);
4122 // fold (select C, X, 1) -> (or (not C), X)
4123 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4124 SDValue NOTNode = DAG.getNOT(N0.getDebugLoc(), N0, VT);
4125 AddToWorkList(NOTNode.getNode());
4126 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, NOTNode, N1);
4128 // fold (select C, X, 0) -> (and C, X)
4129 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4130 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, N1);
4131 // fold (select X, X, Y) -> (or X, Y)
4132 // fold (select X, 1, Y) -> (or X, Y)
4133 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4134 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N2);
4135 // fold (select X, Y, X) -> (and X, Y)
4136 // fold (select X, Y, 0) -> (and X, Y)
4137 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4138 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, N1);
4140 // If we can fold this based on the true/false value, do so.
4141 if (SimplifySelectOps(N, N1, N2))
4142 return SDValue(N, 0); // Don't revisit N.
4144 // fold selects based on a setcc into other things, such as min/max/abs
4145 if (N0.getOpcode() == ISD::SETCC) {
4147 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
4148 // having to say they don't support SELECT_CC on every type the DAG knows
4149 // about, since there is no way to mark an opcode illegal at all value types
4150 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other) &&
4151 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT))
4152 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT,
4153 N0.getOperand(0), N0.getOperand(1),
4154 N1, N2, N0.getOperand(2));
4155 return SimplifySelect(N->getDebugLoc(), N0, N1, N2);
4161 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
4162 SDValue N0 = N->getOperand(0);
4163 SDValue N1 = N->getOperand(1);
4164 SDValue N2 = N->getOperand(2);
4165 SDValue N3 = N->getOperand(3);
4166 SDValue N4 = N->getOperand(4);
4167 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
4169 // fold select_cc lhs, rhs, x, x, cc -> x
4173 // Determine if the condition we're dealing with is constant
4174 SDValue SCC = SimplifySetCC(TLI.getSetCCResultType(N0.getValueType()),
4175 N0, N1, CC, N->getDebugLoc(), false);
4176 if (SCC.getNode()) AddToWorkList(SCC.getNode());
4178 if (ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode())) {
4179 if (!SCCC->isNullValue())
4180 return N2; // cond always true -> true val
4182 return N3; // cond always false -> false val
4185 // Fold to a simpler select_cc
4186 if (SCC.getNode() && SCC.getOpcode() == ISD::SETCC)
4187 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), N2.getValueType(),
4188 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
4191 // If we can fold this based on the true/false value, do so.
4192 if (SimplifySelectOps(N, N2, N3))
4193 return SDValue(N, 0); // Don't revisit N.
4195 // fold select_cc into other things, such as min/max/abs
4196 return SimplifySelectCC(N->getDebugLoc(), N0, N1, N2, N3, CC);
4199 SDValue DAGCombiner::visitSETCC(SDNode *N) {
4200 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
4201 cast<CondCodeSDNode>(N->getOperand(2))->get(),
4205 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
4206 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
4207 // transformation. Returns true if extension are possible and the above
4208 // mentioned transformation is profitable.
4209 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
4211 SmallVector<SDNode*, 4> &ExtendNodes,
4212 const TargetLowering &TLI) {
4213 bool HasCopyToRegUses = false;
4214 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
4215 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
4216 UE = N0.getNode()->use_end();
4221 if (UI.getUse().getResNo() != N0.getResNo())
4223 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
4224 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
4225 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
4226 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
4227 // Sign bits will be lost after a zext.
4230 for (unsigned i = 0; i != 2; ++i) {
4231 SDValue UseOp = User->getOperand(i);
4234 if (!isa<ConstantSDNode>(UseOp))
4239 ExtendNodes.push_back(User);
4242 // If truncates aren't free and there are users we can't
4243 // extend, it isn't worthwhile.
4246 // Remember if this value is live-out.
4247 if (User->getOpcode() == ISD::CopyToReg)
4248 HasCopyToRegUses = true;
4251 if (HasCopyToRegUses) {
4252 bool BothLiveOut = false;
4253 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
4255 SDUse &Use = UI.getUse();
4256 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
4262 // Both unextended and extended values are live out. There had better be
4263 // a good reason for the transformation.
4264 return ExtendNodes.size();
4269 void DAGCombiner::ExtendSetCCUses(SmallVector<SDNode*, 4> SetCCs,
4270 SDValue Trunc, SDValue ExtLoad, DebugLoc DL,
4271 ISD::NodeType ExtType) {
4272 // Extend SetCC uses if necessary.
4273 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
4274 SDNode *SetCC = SetCCs[i];
4275 SmallVector<SDValue, 4> Ops;
4277 for (unsigned j = 0; j != 2; ++j) {
4278 SDValue SOp = SetCC->getOperand(j);
4280 Ops.push_back(ExtLoad);
4282 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
4285 Ops.push_back(SetCC->getOperand(2));
4286 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0),
4287 &Ops[0], Ops.size()));
4291 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
4292 SDValue N0 = N->getOperand(0);
4293 EVT VT = N->getValueType(0);
4295 // fold (sext c1) -> c1
4296 if (isa<ConstantSDNode>(N0))
4297 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N0);
4299 // fold (sext (sext x)) -> (sext x)
4300 // fold (sext (aext x)) -> (sext x)
4301 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4302 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT,
4305 if (N0.getOpcode() == ISD::TRUNCATE) {
4306 // fold (sext (truncate (load x))) -> (sext (smaller load x))
4307 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
4308 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4309 if (NarrowLoad.getNode()) {
4310 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4311 if (NarrowLoad.getNode() != N0.getNode()) {
4312 CombineTo(N0.getNode(), NarrowLoad);
4313 // CombineTo deleted the truncate, if needed, but not what's under it.
4316 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4319 // See if the value being truncated is already sign extended. If so, just
4320 // eliminate the trunc/sext pair.
4321 SDValue Op = N0.getOperand(0);
4322 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
4323 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
4324 unsigned DestBits = VT.getScalarType().getSizeInBits();
4325 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
4327 if (OpBits == DestBits) {
4328 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
4329 // bits, it is already ready.
4330 if (NumSignBits > DestBits-MidBits)
4332 } else if (OpBits < DestBits) {
4333 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
4334 // bits, just sext from i32.
4335 if (NumSignBits > OpBits-MidBits)
4336 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, Op);
4338 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
4339 // bits, just truncate to i32.
4340 if (NumSignBits > OpBits-MidBits)
4341 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4344 // fold (sext (truncate x)) -> (sextinreg x).
4345 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
4346 N0.getValueType())) {
4347 if (OpBits < DestBits)
4348 Op = DAG.getNode(ISD::ANY_EXTEND, N0.getDebugLoc(), VT, Op);
4349 else if (OpBits > DestBits)
4350 Op = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), VT, Op);
4351 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT, Op,
4352 DAG.getValueType(N0.getValueType()));
4356 // fold (sext (load x)) -> (sext (truncate (sextload x)))
4357 // None of the supported targets knows how to perform load and sign extend
4358 // on vectors in one instruction. We only perform this transformation on
4360 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4361 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4362 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) {
4363 bool DoXform = true;
4364 SmallVector<SDNode*, 4> SetCCs;
4365 if (!N0.hasOneUse())
4366 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
4368 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4369 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
4371 LN0->getBasePtr(), LN0->getPointerInfo(),
4373 LN0->isVolatile(), LN0->isNonTemporal(),
4374 LN0->getAlignment());
4375 CombineTo(N, ExtLoad);
4376 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4377 N0.getValueType(), ExtLoad);
4378 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4379 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4381 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4385 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
4386 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
4387 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
4388 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
4389 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4390 EVT MemVT = LN0->getMemoryVT();
4391 if ((!LegalOperations && !LN0->isVolatile()) ||
4392 TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) {
4393 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
4395 LN0->getBasePtr(), LN0->getPointerInfo(),
4397 LN0->isVolatile(), LN0->isNonTemporal(),
4398 LN0->getAlignment());
4399 CombineTo(N, ExtLoad);
4400 CombineTo(N0.getNode(),
4401 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4402 N0.getValueType(), ExtLoad),
4403 ExtLoad.getValue(1));
4404 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4408 // fold (sext (and/or/xor (load x), cst)) ->
4409 // (and/or/xor (sextload x), (sext cst))
4410 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
4411 N0.getOpcode() == ISD::XOR) &&
4412 isa<LoadSDNode>(N0.getOperand(0)) &&
4413 N0.getOperand(1).getOpcode() == ISD::Constant &&
4414 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()) &&
4415 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
4416 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
4417 if (LN0->getExtensionType() != ISD::ZEXTLOAD) {
4418 bool DoXform = true;
4419 SmallVector<SDNode*, 4> SetCCs;
4420 if (!N0.hasOneUse())
4421 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
4424 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, LN0->getDebugLoc(), VT,
4425 LN0->getChain(), LN0->getBasePtr(),
4426 LN0->getPointerInfo(),
4429 LN0->isNonTemporal(),
4430 LN0->getAlignment());
4431 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4432 Mask = Mask.sext(VT.getSizeInBits());
4433 SDValue And = DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
4434 ExtLoad, DAG.getConstant(Mask, VT));
4435 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
4436 N0.getOperand(0).getDebugLoc(),
4437 N0.getOperand(0).getValueType(), ExtLoad);
4439 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
4440 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4442 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4447 if (N0.getOpcode() == ISD::SETCC) {
4448 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
4449 // Only do this before legalize for now.
4450 if (VT.isVector() && !LegalOperations) {
4451 EVT N0VT = N0.getOperand(0).getValueType();
4452 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
4453 // of the same size as the compared operands. Only optimize sext(setcc())
4454 // if this is the case.
4455 EVT SVT = TLI.getSetCCResultType(N0VT);
4457 // We know that the # elements of the results is the same as the
4458 // # elements of the compare (and the # elements of the compare result
4459 // for that matter). Check to see that they are the same size. If so,
4460 // we know that the element size of the sext'd result matches the
4461 // element size of the compare operands.
4462 if (VT.getSizeInBits() == SVT.getSizeInBits())
4463 return DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4465 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4466 // If the desired elements are smaller or larger than the source
4467 // elements we can use a matching integer vector type and then
4468 // truncate/sign extend
4469 EVT MatchingElementType =
4470 EVT::getIntegerVT(*DAG.getContext(),
4471 N0VT.getScalarType().getSizeInBits());
4472 EVT MatchingVectorType =
4473 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4474 N0VT.getVectorNumElements());
4476 if (SVT == MatchingVectorType) {
4477 SDValue VsetCC = DAG.getSetCC(N->getDebugLoc(), MatchingVectorType,
4478 N0.getOperand(0), N0.getOperand(1),
4479 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4480 return DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT);
4484 // sext(setcc x, y, cc) -> (select_cc x, y, -1, 0, cc)
4485 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
4487 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
4489 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4490 NegOne, DAG.getConstant(0, VT),
4491 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4492 if (SCC.getNode()) return SCC;
4493 if (!LegalOperations ||
4494 TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(VT)))
4495 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4496 DAG.getSetCC(N->getDebugLoc(),
4497 TLI.getSetCCResultType(VT),
4498 N0.getOperand(0), N0.getOperand(1),
4499 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
4500 NegOne, DAG.getConstant(0, VT));
4503 // fold (sext x) -> (zext x) if the sign bit is known zero.
4504 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
4505 DAG.SignBitIsZero(N0))
4506 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
4511 // isTruncateOf - If N is a truncate of some other value, return true, record
4512 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
4513 // This function computes KnownZero to avoid a duplicated call to
4514 // ComputeMaskedBits in the caller.
4515 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
4518 if (N->getOpcode() == ISD::TRUNCATE) {
4519 Op = N->getOperand(0);
4520 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
4524 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
4525 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
4528 SDValue Op0 = N->getOperand(0);
4529 SDValue Op1 = N->getOperand(1);
4530 assert(Op0.getValueType() == Op1.getValueType());
4532 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
4533 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
4534 if (COp0 && COp0->isNullValue())
4536 else if (COp1 && COp1->isNullValue())
4541 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
4543 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
4549 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
4550 SDValue N0 = N->getOperand(0);
4551 EVT VT = N->getValueType(0);
4553 // fold (zext c1) -> c1
4554 if (isa<ConstantSDNode>(N0))
4555 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
4556 // fold (zext (zext x)) -> (zext x)
4557 // fold (zext (aext x)) -> (zext x)
4558 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4559 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT,
4562 // fold (zext (truncate x)) -> (zext x) or
4563 // (zext (truncate x)) -> (truncate x)
4564 // This is valid when the truncated bits of x are already zero.
4565 // FIXME: We should extend this to work for vectors too.
4568 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
4569 APInt TruncatedBits =
4570 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
4571 APInt(Op.getValueSizeInBits(), 0) :
4572 APInt::getBitsSet(Op.getValueSizeInBits(),
4573 N0.getValueSizeInBits(),
4574 std::min(Op.getValueSizeInBits(),
4575 VT.getSizeInBits()));
4576 if (TruncatedBits == (KnownZero & TruncatedBits)) {
4577 if (VT.bitsGT(Op.getValueType()))
4578 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, Op);
4579 if (VT.bitsLT(Op.getValueType()))
4580 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4586 // fold (zext (truncate (load x))) -> (zext (smaller load x))
4587 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
4588 if (N0.getOpcode() == ISD::TRUNCATE) {
4589 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4590 if (NarrowLoad.getNode()) {
4591 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4592 if (NarrowLoad.getNode() != N0.getNode()) {
4593 CombineTo(N0.getNode(), NarrowLoad);
4594 // CombineTo deleted the truncate, if needed, but not what's under it.
4597 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4601 // fold (zext (truncate x)) -> (and x, mask)
4602 if (N0.getOpcode() == ISD::TRUNCATE &&
4603 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
4605 // fold (zext (truncate (load x))) -> (zext (smaller load x))
4606 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
4607 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4608 if (NarrowLoad.getNode()) {
4609 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4610 if (NarrowLoad.getNode() != N0.getNode()) {
4611 CombineTo(N0.getNode(), NarrowLoad);
4612 // CombineTo deleted the truncate, if needed, but not what's under it.
4615 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4618 SDValue Op = N0.getOperand(0);
4619 if (Op.getValueType().bitsLT(VT)) {
4620 Op = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, Op);
4621 AddToWorkList(Op.getNode());
4622 } else if (Op.getValueType().bitsGT(VT)) {
4623 Op = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4624 AddToWorkList(Op.getNode());
4626 return DAG.getZeroExtendInReg(Op, N->getDebugLoc(),
4627 N0.getValueType().getScalarType());
4630 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
4631 // if either of the casts is not free.
4632 if (N0.getOpcode() == ISD::AND &&
4633 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
4634 N0.getOperand(1).getOpcode() == ISD::Constant &&
4635 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
4636 N0.getValueType()) ||
4637 !TLI.isZExtFree(N0.getValueType(), VT))) {
4638 SDValue X = N0.getOperand(0).getOperand(0);
4639 if (X.getValueType().bitsLT(VT)) {
4640 X = DAG.getNode(ISD::ANY_EXTEND, X.getDebugLoc(), VT, X);
4641 } else if (X.getValueType().bitsGT(VT)) {
4642 X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
4644 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4645 Mask = Mask.zext(VT.getSizeInBits());
4646 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4647 X, DAG.getConstant(Mask, VT));
4650 // fold (zext (load x)) -> (zext (truncate (zextload x)))
4651 // None of the supported targets knows how to perform load and vector_zext
4652 // on vectors in one instruction. We only perform this transformation on
4654 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4655 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4656 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
4657 bool DoXform = true;
4658 SmallVector<SDNode*, 4> SetCCs;
4659 if (!N0.hasOneUse())
4660 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
4662 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4663 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N->getDebugLoc(), VT,
4665 LN0->getBasePtr(), LN0->getPointerInfo(),
4667 LN0->isVolatile(), LN0->isNonTemporal(),
4668 LN0->getAlignment());
4669 CombineTo(N, ExtLoad);
4670 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4671 N0.getValueType(), ExtLoad);
4672 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4674 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4676 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4680 // fold (zext (and/or/xor (load x), cst)) ->
4681 // (and/or/xor (zextload x), (zext cst))
4682 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
4683 N0.getOpcode() == ISD::XOR) &&
4684 isa<LoadSDNode>(N0.getOperand(0)) &&
4685 N0.getOperand(1).getOpcode() == ISD::Constant &&
4686 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()) &&
4687 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
4688 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
4689 if (LN0->getExtensionType() != ISD::SEXTLOAD) {
4690 bool DoXform = true;
4691 SmallVector<SDNode*, 4> SetCCs;
4692 if (!N0.hasOneUse())
4693 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
4696 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), VT,
4697 LN0->getChain(), LN0->getBasePtr(),
4698 LN0->getPointerInfo(),
4701 LN0->isNonTemporal(),
4702 LN0->getAlignment());
4703 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4704 Mask = Mask.zext(VT.getSizeInBits());
4705 SDValue And = DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
4706 ExtLoad, DAG.getConstant(Mask, VT));
4707 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
4708 N0.getOperand(0).getDebugLoc(),
4709 N0.getOperand(0).getValueType(), ExtLoad);
4711 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
4712 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4714 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4719 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
4720 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
4721 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
4722 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
4723 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4724 EVT MemVT = LN0->getMemoryVT();
4725 if ((!LegalOperations && !LN0->isVolatile()) ||
4726 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) {
4727 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N->getDebugLoc(), VT,
4729 LN0->getBasePtr(), LN0->getPointerInfo(),
4731 LN0->isVolatile(), LN0->isNonTemporal(),
4732 LN0->getAlignment());
4733 CombineTo(N, ExtLoad);
4734 CombineTo(N0.getNode(),
4735 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), N0.getValueType(),
4737 ExtLoad.getValue(1));
4738 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4742 if (N0.getOpcode() == ISD::SETCC) {
4743 if (!LegalOperations && VT.isVector()) {
4744 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
4745 // Only do this before legalize for now.
4746 EVT N0VT = N0.getOperand(0).getValueType();
4747 EVT EltVT = VT.getVectorElementType();
4748 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
4749 DAG.getConstant(1, EltVT));
4750 if (VT.getSizeInBits() == N0VT.getSizeInBits())
4751 // We know that the # elements of the results is the same as the
4752 // # elements of the compare (and the # elements of the compare result
4753 // for that matter). Check to see that they are the same size. If so,
4754 // we know that the element size of the sext'd result matches the
4755 // element size of the compare operands.
4756 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4757 DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4759 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
4760 DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
4761 &OneOps[0], OneOps.size()));
4763 // If the desired elements are smaller or larger than the source
4764 // elements we can use a matching integer vector type and then
4765 // truncate/sign extend
4766 EVT MatchingElementType =
4767 EVT::getIntegerVT(*DAG.getContext(),
4768 N0VT.getScalarType().getSizeInBits());
4769 EVT MatchingVectorType =
4770 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4771 N0VT.getVectorNumElements());
4773 DAG.getSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
4775 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4776 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4777 DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT),
4778 DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
4779 &OneOps[0], OneOps.size()));
4782 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
4784 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4785 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
4786 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4787 if (SCC.getNode()) return SCC;
4790 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
4791 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
4792 isa<ConstantSDNode>(N0.getOperand(1)) &&
4793 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
4795 SDValue ShAmt = N0.getOperand(1);
4796 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
4797 if (N0.getOpcode() == ISD::SHL) {
4798 SDValue InnerZExt = N0.getOperand(0);
4799 // If the original shl may be shifting out bits, do not perform this
4801 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
4802 InnerZExt.getOperand(0).getValueType().getSizeInBits();
4803 if (ShAmtVal > KnownZeroBits)
4807 DebugLoc DL = N->getDebugLoc();
4809 // Ensure that the shift amount is wide enough for the shifted value.
4810 if (VT.getSizeInBits() >= 256)
4811 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
4813 return DAG.getNode(N0.getOpcode(), DL, VT,
4814 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
4821 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
4822 SDValue N0 = N->getOperand(0);
4823 EVT VT = N->getValueType(0);
4825 // fold (aext c1) -> c1
4826 if (isa<ConstantSDNode>(N0))
4827 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, N0);
4828 // fold (aext (aext x)) -> (aext x)
4829 // fold (aext (zext x)) -> (zext x)
4830 // fold (aext (sext x)) -> (sext x)
4831 if (N0.getOpcode() == ISD::ANY_EXTEND ||
4832 N0.getOpcode() == ISD::ZERO_EXTEND ||
4833 N0.getOpcode() == ISD::SIGN_EXTEND)
4834 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, N0.getOperand(0));
4836 // fold (aext (truncate (load x))) -> (aext (smaller load x))
4837 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
4838 if (N0.getOpcode() == ISD::TRUNCATE) {
4839 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4840 if (NarrowLoad.getNode()) {
4841 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4842 if (NarrowLoad.getNode() != N0.getNode()) {
4843 CombineTo(N0.getNode(), NarrowLoad);
4844 // CombineTo deleted the truncate, if needed, but not what's under it.
4847 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4851 // fold (aext (truncate x))
4852 if (N0.getOpcode() == ISD::TRUNCATE) {
4853 SDValue TruncOp = N0.getOperand(0);
4854 if (TruncOp.getValueType() == VT)
4855 return TruncOp; // x iff x size == zext size.
4856 if (TruncOp.getValueType().bitsGT(VT))
4857 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, TruncOp);
4858 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, TruncOp);
4861 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
4862 // if the trunc is not free.
4863 if (N0.getOpcode() == ISD::AND &&
4864 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
4865 N0.getOperand(1).getOpcode() == ISD::Constant &&
4866 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
4867 N0.getValueType())) {
4868 SDValue X = N0.getOperand(0).getOperand(0);
4869 if (X.getValueType().bitsLT(VT)) {
4870 X = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, X);
4871 } else if (X.getValueType().bitsGT(VT)) {
4872 X = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, X);
4874 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4875 Mask = Mask.zext(VT.getSizeInBits());
4876 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4877 X, DAG.getConstant(Mask, VT));
4880 // fold (aext (load x)) -> (aext (truncate (extload x)))
4881 // None of the supported targets knows how to perform load and any_ext
4882 // on vectors in one instruction. We only perform this transformation on
4884 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4885 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4886 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
4887 bool DoXform = true;
4888 SmallVector<SDNode*, 4> SetCCs;
4889 if (!N0.hasOneUse())
4890 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
4892 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4893 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
4895 LN0->getBasePtr(), LN0->getPointerInfo(),
4897 LN0->isVolatile(), LN0->isNonTemporal(),
4898 LN0->getAlignment());
4899 CombineTo(N, ExtLoad);
4900 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4901 N0.getValueType(), ExtLoad);
4902 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4903 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4905 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4909 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
4910 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
4911 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
4912 if (N0.getOpcode() == ISD::LOAD &&
4913 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
4915 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4916 EVT MemVT = LN0->getMemoryVT();
4917 SDValue ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), N->getDebugLoc(),
4918 VT, LN0->getChain(), LN0->getBasePtr(),
4919 LN0->getPointerInfo(), MemVT,
4920 LN0->isVolatile(), LN0->isNonTemporal(),
4921 LN0->getAlignment());
4922 CombineTo(N, ExtLoad);
4923 CombineTo(N0.getNode(),
4924 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4925 N0.getValueType(), ExtLoad),
4926 ExtLoad.getValue(1));
4927 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4930 if (N0.getOpcode() == ISD::SETCC) {
4931 // aext(setcc) -> sext_in_reg(vsetcc) for vectors.
4932 // Only do this before legalize for now.
4933 if (VT.isVector() && !LegalOperations) {
4934 EVT N0VT = N0.getOperand(0).getValueType();
4935 // We know that the # elements of the results is the same as the
4936 // # elements of the compare (and the # elements of the compare result
4937 // for that matter). Check to see that they are the same size. If so,
4938 // we know that the element size of the sext'd result matches the
4939 // element size of the compare operands.
4940 if (VT.getSizeInBits() == N0VT.getSizeInBits())
4941 return DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4943 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4944 // If the desired elements are smaller or larger than the source
4945 // elements we can use a matching integer vector type and then
4946 // truncate/sign extend
4948 EVT MatchingElementType =
4949 EVT::getIntegerVT(*DAG.getContext(),
4950 N0VT.getScalarType().getSizeInBits());
4951 EVT MatchingVectorType =
4952 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4953 N0VT.getVectorNumElements());
4955 DAG.getSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
4957 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4958 return DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT);
4962 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
4964 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4965 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
4966 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4974 /// GetDemandedBits - See if the specified operand can be simplified with the
4975 /// knowledge that only the bits specified by Mask are used. If so, return the
4976 /// simpler operand, otherwise return a null SDValue.
4977 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
4978 switch (V.getOpcode()) {
4980 case ISD::Constant: {
4981 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
4982 assert(CV != 0 && "Const value should be ConstSDNode.");
4983 const APInt &CVal = CV->getAPIntValue();
4984 APInt NewVal = CVal & Mask;
4985 if (NewVal != CVal) {
4986 return DAG.getConstant(NewVal, V.getValueType());
4992 // If the LHS or RHS don't contribute bits to the or, drop them.
4993 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
4994 return V.getOperand(1);
4995 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
4996 return V.getOperand(0);
4999 // Only look at single-use SRLs.
5000 if (!V.getNode()->hasOneUse())
5002 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
5003 // See if we can recursively simplify the LHS.
5004 unsigned Amt = RHSC->getZExtValue();
5006 // Watch out for shift count overflow though.
5007 if (Amt >= Mask.getBitWidth()) break;
5008 APInt NewMask = Mask << Amt;
5009 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
5010 if (SimplifyLHS.getNode())
5011 return DAG.getNode(ISD::SRL, V.getDebugLoc(), V.getValueType(),
5012 SimplifyLHS, V.getOperand(1));
5018 /// ReduceLoadWidth - If the result of a wider load is shifted to right of N
5019 /// bits and then truncated to a narrower type and where N is a multiple
5020 /// of number of bits of the narrower type, transform it to a narrower load
5021 /// from address + N / num of bits of new type. If the result is to be
5022 /// extended, also fold the extension to form a extending load.
5023 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
5024 unsigned Opc = N->getOpcode();
5026 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
5027 SDValue N0 = N->getOperand(0);
5028 EVT VT = N->getValueType(0);
5031 // This transformation isn't valid for vector loads.
5035 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
5037 if (Opc == ISD::SIGN_EXTEND_INREG) {
5038 ExtType = ISD::SEXTLOAD;
5039 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
5040 } else if (Opc == ISD::SRL) {
5041 // Another special-case: SRL is basically zero-extending a narrower value.
5042 ExtType = ISD::ZEXTLOAD;
5044 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
5045 if (!N01) return SDValue();
5046 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
5047 VT.getSizeInBits() - N01->getZExtValue());
5049 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, ExtVT))
5052 unsigned EVTBits = ExtVT.getSizeInBits();
5054 // Do not generate loads of non-round integer types since these can
5055 // be expensive (and would be wrong if the type is not byte sized).
5056 if (!ExtVT.isRound())
5060 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
5061 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5062 ShAmt = N01->getZExtValue();
5063 // Is the shift amount a multiple of size of VT?
5064 if ((ShAmt & (EVTBits-1)) == 0) {
5065 N0 = N0.getOperand(0);
5066 // Is the load width a multiple of size of VT?
5067 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
5071 // At this point, we must have a load or else we can't do the transform.
5072 if (!isa<LoadSDNode>(N0)) return SDValue();
5074 // Because a SRL must be assumed to *need* to zero-extend the high bits
5075 // (as opposed to anyext the high bits), we can't combine the zextload
5076 // lowering of SRL and an sextload.
5077 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
5080 // If the shift amount is larger than the input type then we're not
5081 // accessing any of the loaded bytes. If the load was a zextload/extload
5082 // then the result of the shift+trunc is zero/undef (handled elsewhere).
5083 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
5088 // If the load is shifted left (and the result isn't shifted back right),
5089 // we can fold the truncate through the shift.
5090 unsigned ShLeftAmt = 0;
5091 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
5092 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
5093 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5094 ShLeftAmt = N01->getZExtValue();
5095 N0 = N0.getOperand(0);
5099 // If we haven't found a load, we can't narrow it. Don't transform one with
5100 // multiple uses, this would require adding a new load.
5101 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse() ||
5102 // Don't change the width of a volatile load.
5103 cast<LoadSDNode>(N0)->isVolatile())
5106 // Verify that we are actually reducing a load width here.
5107 if (cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits() < EVTBits)
5110 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5111 EVT PtrType = N0.getOperand(1).getValueType();
5113 if (PtrType == MVT::Untyped || PtrType.isExtended())
5114 // It's not possible to generate a constant of extended or untyped type.
5117 // For big endian targets, we need to adjust the offset to the pointer to
5118 // load the correct bytes.
5119 if (TLI.isBigEndian()) {
5120 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
5121 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
5122 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
5125 uint64_t PtrOff = ShAmt / 8;
5126 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
5127 SDValue NewPtr = DAG.getNode(ISD::ADD, LN0->getDebugLoc(),
5128 PtrType, LN0->getBasePtr(),
5129 DAG.getConstant(PtrOff, PtrType));
5130 AddToWorkList(NewPtr.getNode());
5133 if (ExtType == ISD::NON_EXTLOAD)
5134 Load = DAG.getLoad(VT, N0.getDebugLoc(), LN0->getChain(), NewPtr,
5135 LN0->getPointerInfo().getWithOffset(PtrOff),
5136 LN0->isVolatile(), LN0->isNonTemporal(),
5137 LN0->isInvariant(), NewAlign);
5139 Load = DAG.getExtLoad(ExtType, N0.getDebugLoc(), VT, LN0->getChain(),NewPtr,
5140 LN0->getPointerInfo().getWithOffset(PtrOff),
5141 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
5144 // Replace the old load's chain with the new load's chain.
5145 WorkListRemover DeadNodes(*this);
5146 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
5148 // Shift the result left, if we've swallowed a left shift.
5149 SDValue Result = Load;
5150 if (ShLeftAmt != 0) {
5151 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
5152 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
5154 Result = DAG.getNode(ISD::SHL, N0.getDebugLoc(), VT,
5155 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
5158 // Return the new loaded value.
5162 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
5163 SDValue N0 = N->getOperand(0);
5164 SDValue N1 = N->getOperand(1);
5165 EVT VT = N->getValueType(0);
5166 EVT EVT = cast<VTSDNode>(N1)->getVT();
5167 unsigned VTBits = VT.getScalarType().getSizeInBits();
5168 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
5170 // fold (sext_in_reg c1) -> c1
5171 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
5172 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT, N0, N1);
5174 // If the input is already sign extended, just drop the extension.
5175 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
5178 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
5179 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
5180 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT())) {
5181 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
5182 N0.getOperand(0), N1);
5185 // fold (sext_in_reg (sext x)) -> (sext x)
5186 // fold (sext_in_reg (aext x)) -> (sext x)
5187 // if x is small enough.
5188 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
5189 SDValue N00 = N0.getOperand(0);
5190 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
5191 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
5192 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N00, N1);
5195 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
5196 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
5197 return DAG.getZeroExtendInReg(N0, N->getDebugLoc(), EVT);
5199 // fold operands of sext_in_reg based on knowledge that the top bits are not
5201 if (SimplifyDemandedBits(SDValue(N, 0)))
5202 return SDValue(N, 0);
5204 // fold (sext_in_reg (load x)) -> (smaller sextload x)
5205 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
5206 SDValue NarrowLoad = ReduceLoadWidth(N);
5207 if (NarrowLoad.getNode())
5210 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
5211 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
5212 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
5213 if (N0.getOpcode() == ISD::SRL) {
5214 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
5215 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
5216 // We can turn this into an SRA iff the input to the SRL is already sign
5218 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
5219 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
5220 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT,
5221 N0.getOperand(0), N0.getOperand(1));
5225 // fold (sext_inreg (extload x)) -> (sextload x)
5226 if (ISD::isEXTLoad(N0.getNode()) &&
5227 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5228 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5229 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5230 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5231 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5232 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
5234 LN0->getBasePtr(), LN0->getPointerInfo(),
5236 LN0->isVolatile(), LN0->isNonTemporal(),
5237 LN0->getAlignment());
5238 CombineTo(N, ExtLoad);
5239 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5240 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5242 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
5243 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5245 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5246 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5247 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5248 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5249 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
5251 LN0->getBasePtr(), LN0->getPointerInfo(),
5253 LN0->isVolatile(), LN0->isNonTemporal(),
5254 LN0->getAlignment());
5255 CombineTo(N, ExtLoad);
5256 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5257 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5260 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
5261 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
5262 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
5263 N0.getOperand(1), false);
5264 if (BSwap.getNode() != 0)
5265 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
5272 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
5273 SDValue N0 = N->getOperand(0);
5274 EVT VT = N->getValueType(0);
5275 bool isLE = TLI.isLittleEndian();
5278 if (N0.getValueType() == N->getValueType(0))
5280 // fold (truncate c1) -> c1
5281 if (isa<ConstantSDNode>(N0))
5282 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0);
5283 // fold (truncate (truncate x)) -> (truncate x)
5284 if (N0.getOpcode() == ISD::TRUNCATE)
5285 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0.getOperand(0));
5286 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
5287 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
5288 N0.getOpcode() == ISD::SIGN_EXTEND ||
5289 N0.getOpcode() == ISD::ANY_EXTEND) {
5290 if (N0.getOperand(0).getValueType().bitsLT(VT))
5291 // if the source is smaller than the dest, we still need an extend
5292 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
5294 if (N0.getOperand(0).getValueType().bitsGT(VT))
5295 // if the source is larger than the dest, than we just need the truncate
5296 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0.getOperand(0));
5297 // if the source and dest are the same type, we can drop both the extend
5298 // and the truncate.
5299 return N0.getOperand(0);
5302 // Fold extract-and-trunc into a narrow extract. For example:
5303 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
5304 // i32 y = TRUNCATE(i64 x)
5306 // v16i8 b = BITCAST (v2i64 val)
5307 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
5309 // Note: We only run this optimization after type legalization (which often
5310 // creates this pattern) and before operation legalization after which
5311 // we need to be more careful about the vector instructions that we generate.
5312 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5313 LegalTypes && !LegalOperations && N0->hasOneUse()) {
5315 EVT VecTy = N0.getOperand(0).getValueType();
5316 EVT ExTy = N0.getValueType();
5317 EVT TrTy = N->getValueType(0);
5319 unsigned NumElem = VecTy.getVectorNumElements();
5320 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
5322 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
5323 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
5325 SDValue EltNo = N0->getOperand(1);
5326 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
5327 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5328 EVT IndexTy = N0->getOperand(1).getValueType();
5329 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
5331 SDValue V = DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
5332 NVT, N0.getOperand(0));
5334 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
5335 N->getDebugLoc(), TrTy, V,
5336 DAG.getConstant(Index, IndexTy));
5340 // See if we can simplify the input to this truncate through knowledge that
5341 // only the low bits are being used.
5342 // For example "trunc (or (shl x, 8), y)" // -> trunc y
5343 // Currently we only perform this optimization on scalars because vectors
5344 // may have different active low bits.
5345 if (!VT.isVector()) {
5347 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
5348 VT.getSizeInBits()));
5349 if (Shorter.getNode())
5350 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Shorter);
5352 // fold (truncate (load x)) -> (smaller load x)
5353 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
5354 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
5355 SDValue Reduced = ReduceLoadWidth(N);
5356 if (Reduced.getNode())
5359 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
5360 // where ... are all 'undef'.
5361 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
5362 SmallVector<EVT, 8> VTs;
5365 unsigned NumDefs = 0;
5367 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
5368 SDValue X = N0.getOperand(i);
5369 if (X.getOpcode() != ISD::UNDEF) {
5374 // Stop if more than one members are non-undef.
5377 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
5378 VT.getVectorElementType(),
5379 X.getValueType().getVectorNumElements()));
5383 return DAG.getUNDEF(VT);
5386 assert(V.getNode() && "The single defined operand is empty!");
5387 SmallVector<SDValue, 8> Opnds;
5388 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
5390 Opnds.push_back(DAG.getUNDEF(VTs[i]));
5393 SDValue NV = DAG.getNode(ISD::TRUNCATE, V.getDebugLoc(), VTs[i], V);
5394 AddToWorkList(NV.getNode());
5395 Opnds.push_back(NV);
5397 return DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
5398 &Opnds[0], Opnds.size());
5402 // Simplify the operands using demanded-bits information.
5403 if (!VT.isVector() &&
5404 SimplifyDemandedBits(SDValue(N, 0)))
5405 return SDValue(N, 0);
5410 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
5411 SDValue Elt = N->getOperand(i);
5412 if (Elt.getOpcode() != ISD::MERGE_VALUES)
5413 return Elt.getNode();
5414 return Elt.getOperand(Elt.getResNo()).getNode();
5417 /// CombineConsecutiveLoads - build_pair (load, load) -> load
5418 /// if load locations are consecutive.
5419 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
5420 assert(N->getOpcode() == ISD::BUILD_PAIR);
5422 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
5423 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
5424 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
5425 LD1->getPointerInfo().getAddrSpace() !=
5426 LD2->getPointerInfo().getAddrSpace())
5428 EVT LD1VT = LD1->getValueType(0);
5430 if (ISD::isNON_EXTLoad(LD2) &&
5432 // If both are volatile this would reduce the number of volatile loads.
5433 // If one is volatile it might be ok, but play conservative and bail out.
5434 !LD1->isVolatile() &&
5435 !LD2->isVolatile() &&
5436 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
5437 unsigned Align = LD1->getAlignment();
5438 unsigned NewAlign = TLI.getDataLayout()->
5439 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
5441 if (NewAlign <= Align &&
5442 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
5443 return DAG.getLoad(VT, N->getDebugLoc(), LD1->getChain(),
5444 LD1->getBasePtr(), LD1->getPointerInfo(),
5445 false, false, false, Align);
5451 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
5452 SDValue N0 = N->getOperand(0);
5453 EVT VT = N->getValueType(0);
5455 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
5456 // Only do this before legalize, since afterward the target may be depending
5457 // on the bitconvert.
5458 // First check to see if this is all constant.
5460 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
5462 bool isSimple = true;
5463 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i)
5464 if (N0.getOperand(i).getOpcode() != ISD::UNDEF &&
5465 N0.getOperand(i).getOpcode() != ISD::Constant &&
5466 N0.getOperand(i).getOpcode() != ISD::ConstantFP) {
5471 EVT DestEltVT = N->getValueType(0).getVectorElementType();
5472 assert(!DestEltVT.isVector() &&
5473 "Element type of vector ValueType must not be vector!");
5475 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
5478 // If the input is a constant, let getNode fold it.
5479 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
5480 SDValue Res = DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, N0);
5481 if (Res.getNode() != N) {
5482 if (!LegalOperations ||
5483 TLI.isOperationLegal(Res.getNode()->getOpcode(), VT))
5486 // Folding it resulted in an illegal node, and it's too late to
5487 // do that. Clean up the old node and forego the transformation.
5488 // Ideally this won't happen very often, because instcombine
5489 // and the earlier dagcombine runs (where illegal nodes are
5490 // permitted) should have folded most of them already.
5491 DAG.DeleteNode(Res.getNode());
5495 // (conv (conv x, t1), t2) -> (conv x, t2)
5496 if (N0.getOpcode() == ISD::BITCAST)
5497 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT,
5500 // fold (conv (load x)) -> (load (conv*)x)
5501 // If the resultant load doesn't need a higher alignment than the original!
5502 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
5503 // Do not change the width of a volatile load.
5504 !cast<LoadSDNode>(N0)->isVolatile() &&
5505 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT))) {
5506 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5507 unsigned Align = TLI.getDataLayout()->
5508 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
5509 unsigned OrigAlign = LN0->getAlignment();
5511 if (Align <= OrigAlign) {
5512 SDValue Load = DAG.getLoad(VT, N->getDebugLoc(), LN0->getChain(),
5513 LN0->getBasePtr(), LN0->getPointerInfo(),
5514 LN0->isVolatile(), LN0->isNonTemporal(),
5515 LN0->isInvariant(), OrigAlign);
5517 CombineTo(N0.getNode(),
5518 DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5519 N0.getValueType(), Load),
5525 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
5526 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
5527 // This often reduces constant pool loads.
5528 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(VT)) ||
5529 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(VT))) &&
5530 N0.getNode()->hasOneUse() && VT.isInteger() &&
5531 !VT.isVector() && !N0.getValueType().isVector()) {
5532 SDValue NewConv = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(), VT,
5534 AddToWorkList(NewConv.getNode());
5536 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
5537 if (N0.getOpcode() == ISD::FNEG)
5538 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
5539 NewConv, DAG.getConstant(SignBit, VT));
5540 assert(N0.getOpcode() == ISD::FABS);
5541 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
5542 NewConv, DAG.getConstant(~SignBit, VT));
5545 // fold (bitconvert (fcopysign cst, x)) ->
5546 // (or (and (bitconvert x), sign), (and cst, (not sign)))
5547 // Note that we don't handle (copysign x, cst) because this can always be
5548 // folded to an fneg or fabs.
5549 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
5550 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
5551 VT.isInteger() && !VT.isVector()) {
5552 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
5553 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
5554 if (isTypeLegal(IntXVT)) {
5555 SDValue X = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5556 IntXVT, N0.getOperand(1));
5557 AddToWorkList(X.getNode());
5559 // If X has a different width than the result/lhs, sext it or truncate it.
5560 unsigned VTWidth = VT.getSizeInBits();
5561 if (OrigXWidth < VTWidth) {
5562 X = DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, X);
5563 AddToWorkList(X.getNode());
5564 } else if (OrigXWidth > VTWidth) {
5565 // To get the sign bit in the right place, we have to shift it right
5566 // before truncating.
5567 X = DAG.getNode(ISD::SRL, X.getDebugLoc(),
5568 X.getValueType(), X,
5569 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
5570 AddToWorkList(X.getNode());
5571 X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
5572 AddToWorkList(X.getNode());
5575 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
5576 X = DAG.getNode(ISD::AND, X.getDebugLoc(), VT,
5577 X, DAG.getConstant(SignBit, VT));
5578 AddToWorkList(X.getNode());
5580 SDValue Cst = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5581 VT, N0.getOperand(0));
5582 Cst = DAG.getNode(ISD::AND, Cst.getDebugLoc(), VT,
5583 Cst, DAG.getConstant(~SignBit, VT));
5584 AddToWorkList(Cst.getNode());
5586 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, X, Cst);
5590 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
5591 if (N0.getOpcode() == ISD::BUILD_PAIR) {
5592 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
5593 if (CombineLD.getNode())
5600 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
5601 EVT VT = N->getValueType(0);
5602 return CombineConsecutiveLoads(N, VT);
5605 /// ConstantFoldBITCASTofBUILD_VECTOR - We know that BV is a build_vector
5606 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
5607 /// destination element value type.
5608 SDValue DAGCombiner::
5609 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
5610 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
5612 // If this is already the right type, we're done.
5613 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
5615 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
5616 unsigned DstBitSize = DstEltVT.getSizeInBits();
5618 // If this is a conversion of N elements of one type to N elements of another
5619 // type, convert each element. This handles FP<->INT cases.
5620 if (SrcBitSize == DstBitSize) {
5621 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
5622 BV->getValueType(0).getVectorNumElements());
5624 // Due to the FP element handling below calling this routine recursively,
5625 // we can end up with a scalar-to-vector node here.
5626 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
5627 return DAG.getNode(ISD::SCALAR_TO_VECTOR, BV->getDebugLoc(), VT,
5628 DAG.getNode(ISD::BITCAST, BV->getDebugLoc(),
5629 DstEltVT, BV->getOperand(0)));
5631 SmallVector<SDValue, 8> Ops;
5632 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
5633 SDValue Op = BV->getOperand(i);
5634 // If the vector element type is not legal, the BUILD_VECTOR operands
5635 // are promoted and implicitly truncated. Make that explicit here.
5636 if (Op.getValueType() != SrcEltVT)
5637 Op = DAG.getNode(ISD::TRUNCATE, BV->getDebugLoc(), SrcEltVT, Op);
5638 Ops.push_back(DAG.getNode(ISD::BITCAST, BV->getDebugLoc(),
5640 AddToWorkList(Ops.back().getNode());
5642 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5643 &Ops[0], Ops.size());
5646 // Otherwise, we're growing or shrinking the elements. To avoid having to
5647 // handle annoying details of growing/shrinking FP values, we convert them to
5649 if (SrcEltVT.isFloatingPoint()) {
5650 // Convert the input float vector to a int vector where the elements are the
5652 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
5653 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
5654 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
5658 // Now we know the input is an integer vector. If the output is a FP type,
5659 // convert to integer first, then to FP of the right size.
5660 if (DstEltVT.isFloatingPoint()) {
5661 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
5662 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
5663 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
5665 // Next, convert to FP elements of the same size.
5666 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
5669 // Okay, we know the src/dst types are both integers of differing types.
5670 // Handling growing first.
5671 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
5672 if (SrcBitSize < DstBitSize) {
5673 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
5675 SmallVector<SDValue, 8> Ops;
5676 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
5677 i += NumInputsPerOutput) {
5678 bool isLE = TLI.isLittleEndian();
5679 APInt NewBits = APInt(DstBitSize, 0);
5680 bool EltIsUndef = true;
5681 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
5682 // Shift the previously computed bits over.
5683 NewBits <<= SrcBitSize;
5684 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
5685 if (Op.getOpcode() == ISD::UNDEF) continue;
5688 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
5689 zextOrTrunc(SrcBitSize).zext(DstBitSize);
5693 Ops.push_back(DAG.getUNDEF(DstEltVT));
5695 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
5698 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
5699 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5700 &Ops[0], Ops.size());
5703 // Finally, this must be the case where we are shrinking elements: each input
5704 // turns into multiple outputs.
5705 bool isS2V = ISD::isScalarToVector(BV);
5706 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
5707 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
5708 NumOutputsPerInput*BV->getNumOperands());
5709 SmallVector<SDValue, 8> Ops;
5711 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
5712 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
5713 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
5714 Ops.push_back(DAG.getUNDEF(DstEltVT));
5718 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
5719 getAPIntValue().zextOrTrunc(SrcBitSize);
5721 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
5722 APInt ThisVal = OpVal.trunc(DstBitSize);
5723 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
5724 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
5725 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
5726 return DAG.getNode(ISD::SCALAR_TO_VECTOR, BV->getDebugLoc(), VT,
5728 OpVal = OpVal.lshr(DstBitSize);
5731 // For big endian targets, swap the order of the pieces of each element.
5732 if (TLI.isBigEndian())
5733 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
5736 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5737 &Ops[0], Ops.size());
5740 SDValue DAGCombiner::visitFADD(SDNode *N) {
5741 SDValue N0 = N->getOperand(0);
5742 SDValue N1 = N->getOperand(1);
5743 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5744 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5745 EVT VT = N->getValueType(0);
5748 if (VT.isVector()) {
5749 SDValue FoldedVOp = SimplifyVBinOp(N);
5750 if (FoldedVOp.getNode()) return FoldedVOp;
5753 // fold (fadd c1, c2) -> c1 + c2
5755 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N1);
5756 // canonicalize constant to RHS
5757 if (N0CFP && !N1CFP)
5758 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N1, N0);
5759 // fold (fadd A, 0) -> A
5760 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
5761 N1CFP->getValueAPF().isZero())
5763 // fold (fadd A, (fneg B)) -> (fsub A, B)
5764 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
5765 isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
5766 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N0,
5767 GetNegatedExpression(N1, DAG, LegalOperations));
5768 // fold (fadd (fneg A), B) -> (fsub B, A)
5769 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
5770 isNegatibleForFree(N0, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
5771 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N1,
5772 GetNegatedExpression(N0, DAG, LegalOperations));
5774 // If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
5775 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
5776 N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
5777 isa<ConstantFPSDNode>(N0.getOperand(1)))
5778 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0.getOperand(0),
5779 DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5780 N0.getOperand(1), N1));
5782 // If allow, fold (fadd (fneg x), x) -> 0.0
5783 if (DAG.getTarget().Options.UnsafeFPMath &&
5784 N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1) {
5785 return DAG.getConstantFP(0.0, VT);
5788 // If allow, fold (fadd x, (fneg x)) -> 0.0
5789 if (DAG.getTarget().Options.UnsafeFPMath &&
5790 N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0) {
5791 return DAG.getConstantFP(0.0, VT);
5794 // In unsafe math mode, we can fold chains of FADD's of the same value
5795 // into multiplications. This transform is not safe in general because
5796 // we are reducing the number of rounding steps.
5797 if (DAG.getTarget().Options.UnsafeFPMath &&
5798 TLI.isOperationLegalOrCustom(ISD::FMUL, VT) &&
5800 if (N0.getOpcode() == ISD::FMUL) {
5801 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
5802 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
5804 // (fadd (fmul c, x), x) -> (fmul c+1, x)
5805 if (CFP00 && !CFP01 && N0.getOperand(1) == N1) {
5806 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5808 DAG.getConstantFP(1.0, VT));
5809 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5813 // (fadd (fmul x, c), x) -> (fmul c+1, x)
5814 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
5815 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5817 DAG.getConstantFP(1.0, VT));
5818 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5822 // (fadd (fadd x, x), x) -> (fmul 3.0, x)
5823 if (!CFP00 && !CFP01 && N0.getOperand(0) == N0.getOperand(1) &&
5824 N0.getOperand(0) == N1) {
5825 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5826 N1, DAG.getConstantFP(3.0, VT));
5829 // (fadd (fmul c, x), (fadd x, x)) -> (fmul c+2, x)
5830 if (CFP00 && !CFP01 && N1.getOpcode() == ISD::FADD &&
5831 N1.getOperand(0) == N1.getOperand(1) &&
5832 N0.getOperand(1) == N1.getOperand(0)) {
5833 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5835 DAG.getConstantFP(2.0, VT));
5836 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5837 N0.getOperand(1), NewCFP);
5840 // (fadd (fmul x, c), (fadd x, x)) -> (fmul c+2, x)
5841 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
5842 N1.getOperand(0) == N1.getOperand(1) &&
5843 N0.getOperand(0) == N1.getOperand(0)) {
5844 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5846 DAG.getConstantFP(2.0, VT));
5847 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5848 N0.getOperand(0), NewCFP);
5852 if (N1.getOpcode() == ISD::FMUL) {
5853 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
5854 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
5856 // (fadd x, (fmul c, x)) -> (fmul c+1, x)
5857 if (CFP10 && !CFP11 && N1.getOperand(1) == N0) {
5858 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5860 DAG.getConstantFP(1.0, VT));
5861 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5865 // (fadd x, (fmul x, c)) -> (fmul c+1, x)
5866 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
5867 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5869 DAG.getConstantFP(1.0, VT));
5870 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5874 // (fadd x, (fadd x, x)) -> (fmul 3.0, x)
5875 if (!CFP10 && !CFP11 && N1.getOperand(0) == N1.getOperand(1) &&
5876 N1.getOperand(0) == N0) {
5877 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5878 N0, DAG.getConstantFP(3.0, VT));
5881 // (fadd (fadd x, x), (fmul c, x)) -> (fmul c+2, x)
5882 if (CFP10 && !CFP11 && N1.getOpcode() == ISD::FADD &&
5883 N1.getOperand(0) == N1.getOperand(1) &&
5884 N0.getOperand(1) == N1.getOperand(0)) {
5885 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5887 DAG.getConstantFP(2.0, VT));
5888 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5889 N0.getOperand(1), NewCFP);
5892 // (fadd (fadd x, x), (fmul x, c)) -> (fmul c+2, x)
5893 if (CFP11 && !CFP10 && N1.getOpcode() == ISD::FADD &&
5894 N1.getOperand(0) == N1.getOperand(1) &&
5895 N0.getOperand(0) == N1.getOperand(0)) {
5896 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5898 DAG.getConstantFP(2.0, VT));
5899 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5900 N0.getOperand(0), NewCFP);
5904 // (fadd (fadd x, x), (fadd x, x)) -> (fmul 4.0, x)
5905 if (N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
5906 N0.getOperand(0) == N0.getOperand(1) &&
5907 N1.getOperand(0) == N1.getOperand(1) &&
5908 N0.getOperand(0) == N1.getOperand(0)) {
5909 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5911 DAG.getConstantFP(4.0, VT));
5915 // FADD -> FMA combines:
5916 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
5917 DAG.getTarget().Options.UnsafeFPMath) &&
5918 DAG.getTarget().getTargetLowering()->isFMAFasterThanMulAndAdd(VT) &&
5919 TLI.isOperationLegalOrCustom(ISD::FMA, VT)) {
5921 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
5922 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse()) {
5923 return DAG.getNode(ISD::FMA, N->getDebugLoc(), VT,
5924 N0.getOperand(0), N0.getOperand(1), N1);
5927 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
5928 // Note: Commutes FADD operands.
5929 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse()) {
5930 return DAG.getNode(ISD::FMA, N->getDebugLoc(), VT,
5931 N1.getOperand(0), N1.getOperand(1), N0);
5938 SDValue DAGCombiner::visitFSUB(SDNode *N) {
5939 SDValue N0 = N->getOperand(0);
5940 SDValue N1 = N->getOperand(1);
5941 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5942 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5943 EVT VT = N->getValueType(0);
5944 DebugLoc dl = N->getDebugLoc();
5947 if (VT.isVector()) {
5948 SDValue FoldedVOp = SimplifyVBinOp(N);
5949 if (FoldedVOp.getNode()) return FoldedVOp;
5952 // fold (fsub c1, c2) -> c1-c2
5954 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N0, N1);
5955 // fold (fsub A, 0) -> A
5956 if (DAG.getTarget().Options.UnsafeFPMath &&
5957 N1CFP && N1CFP->getValueAPF().isZero())
5959 // fold (fsub 0, B) -> -B
5960 if (DAG.getTarget().Options.UnsafeFPMath &&
5961 N0CFP && N0CFP->getValueAPF().isZero()) {
5962 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
5963 return GetNegatedExpression(N1, DAG, LegalOperations);
5964 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
5965 return DAG.getNode(ISD::FNEG, dl, VT, N1);
5967 // fold (fsub A, (fneg B)) -> (fadd A, B)
5968 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
5969 return DAG.getNode(ISD::FADD, dl, VT, N0,
5970 GetNegatedExpression(N1, DAG, LegalOperations));
5972 // If 'unsafe math' is enabled, fold
5973 // (fsub x, x) -> 0.0 &
5974 // (fsub x, (fadd x, y)) -> (fneg y) &
5975 // (fsub x, (fadd y, x)) -> (fneg y)
5976 if (DAG.getTarget().Options.UnsafeFPMath) {
5978 return DAG.getConstantFP(0.0f, VT);
5980 if (N1.getOpcode() == ISD::FADD) {
5981 SDValue N10 = N1->getOperand(0);
5982 SDValue N11 = N1->getOperand(1);
5984 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI,
5985 &DAG.getTarget().Options))
5986 return GetNegatedExpression(N11, DAG, LegalOperations);
5987 else if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI,
5988 &DAG.getTarget().Options))
5989 return GetNegatedExpression(N10, DAG, LegalOperations);
5993 // FSUB -> FMA combines:
5994 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
5995 DAG.getTarget().Options.UnsafeFPMath) &&
5996 DAG.getTarget().getTargetLowering()->isFMAFasterThanMulAndAdd(VT) &&
5997 TLI.isOperationLegalOrCustom(ISD::FMA, VT)) {
5999 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
6000 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse()) {
6001 return DAG.getNode(ISD::FMA, dl, VT,
6002 N0.getOperand(0), N0.getOperand(1),
6003 DAG.getNode(ISD::FNEG, dl, VT, N1));
6006 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
6007 // Note: Commutes FSUB operands.
6008 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse()) {
6009 return DAG.getNode(ISD::FMA, dl, VT,
6010 DAG.getNode(ISD::FNEG, dl, VT,
6012 N1.getOperand(1), N0);
6015 // fold (fsub (-(fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
6016 if (N0.getOpcode() == ISD::FNEG &&
6017 N0.getOperand(0).getOpcode() == ISD::FMUL &&
6018 N0->hasOneUse() && N0.getOperand(0).hasOneUse()) {
6019 SDValue N00 = N0.getOperand(0).getOperand(0);
6020 SDValue N01 = N0.getOperand(0).getOperand(1);
6021 return DAG.getNode(ISD::FMA, dl, VT,
6022 DAG.getNode(ISD::FNEG, dl, VT, N00), N01,
6023 DAG.getNode(ISD::FNEG, dl, VT, N1));
6030 SDValue DAGCombiner::visitFMUL(SDNode *N) {
6031 SDValue N0 = N->getOperand(0);
6032 SDValue N1 = N->getOperand(1);
6033 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6034 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6035 EVT VT = N->getValueType(0);
6036 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6039 if (VT.isVector()) {
6040 SDValue FoldedVOp = SimplifyVBinOp(N);
6041 if (FoldedVOp.getNode()) return FoldedVOp;
6044 // fold (fmul c1, c2) -> c1*c2
6046 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0, N1);
6047 // canonicalize constant to RHS
6048 if (N0CFP && !N1CFP)
6049 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N1, N0);
6050 // fold (fmul A, 0) -> 0
6051 if (DAG.getTarget().Options.UnsafeFPMath &&
6052 N1CFP && N1CFP->getValueAPF().isZero())
6054 // fold (fmul A, 0) -> 0, vector edition.
6055 if (DAG.getTarget().Options.UnsafeFPMath &&
6056 ISD::isBuildVectorAllZeros(N1.getNode()))
6058 // fold (fmul A, 1.0) -> A
6059 if (N1CFP && N1CFP->isExactlyValue(1.0))
6061 // fold (fmul X, 2.0) -> (fadd X, X)
6062 if (N1CFP && N1CFP->isExactlyValue(+2.0))
6063 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N0);
6064 // fold (fmul X, -1.0) -> (fneg X)
6065 if (N1CFP && N1CFP->isExactlyValue(-1.0))
6066 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6067 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT, N0);
6069 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
6070 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
6071 &DAG.getTarget().Options)) {
6072 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
6073 &DAG.getTarget().Options)) {
6074 // Both can be negated for free, check to see if at least one is cheaper
6076 if (LHSNeg == 2 || RHSNeg == 2)
6077 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
6078 GetNegatedExpression(N0, DAG, LegalOperations),
6079 GetNegatedExpression(N1, DAG, LegalOperations));
6083 // If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
6084 if (DAG.getTarget().Options.UnsafeFPMath &&
6085 N1CFP && N0.getOpcode() == ISD::FMUL &&
6086 N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
6087 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0.getOperand(0),
6088 DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
6089 N0.getOperand(1), N1));
6094 SDValue DAGCombiner::visitFMA(SDNode *N) {
6095 SDValue N0 = N->getOperand(0);
6096 SDValue N1 = N->getOperand(1);
6097 SDValue N2 = N->getOperand(2);
6098 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6099 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6100 EVT VT = N->getValueType(0);
6101 DebugLoc dl = N->getDebugLoc();
6103 if (DAG.getTarget().Options.UnsafeFPMath) {
6104 if (N0CFP && N0CFP->isZero())
6106 if (N1CFP && N1CFP->isZero())
6109 if (N0CFP && N0CFP->isExactlyValue(1.0))
6110 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N1, N2);
6111 if (N1CFP && N1CFP->isExactlyValue(1.0))
6112 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N2);
6114 // Canonicalize (fma c, x, y) -> (fma x, c, y)
6115 if (N0CFP && !N1CFP)
6116 return DAG.getNode(ISD::FMA, N->getDebugLoc(), VT, N1, N0, N2);
6118 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
6119 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6120 N2.getOpcode() == ISD::FMUL &&
6121 N0 == N2.getOperand(0) &&
6122 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
6123 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6124 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
6128 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
6129 if (DAG.getTarget().Options.UnsafeFPMath &&
6130 N0.getOpcode() == ISD::FMUL && N1CFP &&
6131 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
6132 return DAG.getNode(ISD::FMA, dl, VT,
6134 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
6138 // (fma x, 1, y) -> (fadd x, y)
6139 // (fma x, -1, y) -> (fadd (fneg x), y)
6141 if (N1CFP->isExactlyValue(1.0))
6142 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
6144 if (N1CFP->isExactlyValue(-1.0) &&
6145 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
6146 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
6147 AddToWorkList(RHSNeg.getNode());
6148 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
6152 // (fma x, c, x) -> (fmul x, (c+1))
6153 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP && N0 == N2) {
6154 return DAG.getNode(ISD::FMUL, dl, VT,
6156 DAG.getNode(ISD::FADD, dl, VT,
6157 N1, DAG.getConstantFP(1.0, VT)));
6160 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
6161 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6162 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0) {
6163 return DAG.getNode(ISD::FMUL, dl, VT,
6165 DAG.getNode(ISD::FADD, dl, VT,
6166 N1, DAG.getConstantFP(-1.0, VT)));
6173 SDValue DAGCombiner::visitFDIV(SDNode *N) {
6174 SDValue N0 = N->getOperand(0);
6175 SDValue N1 = N->getOperand(1);
6176 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6177 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6178 EVT VT = N->getValueType(0);
6179 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6182 if (VT.isVector()) {
6183 SDValue FoldedVOp = SimplifyVBinOp(N);
6184 if (FoldedVOp.getNode()) return FoldedVOp;
6187 // fold (fdiv c1, c2) -> c1/c2
6189 return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT, N0, N1);
6191 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
6192 if (N1CFP && DAG.getTarget().Options.UnsafeFPMath) {
6193 // Compute the reciprocal 1.0 / c2.
6194 APFloat N1APF = N1CFP->getValueAPF();
6195 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
6196 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
6197 // Only do the transform if the reciprocal is a legal fp immediate that
6198 // isn't too nasty (eg NaN, denormal, ...).
6199 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
6200 (!LegalOperations ||
6201 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
6202 // backend)... we should handle this gracefully after Legalize.
6203 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
6204 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
6205 TLI.isFPImmLegal(Recip, VT)))
6206 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0,
6207 DAG.getConstantFP(Recip, VT));
6210 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
6211 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
6212 &DAG.getTarget().Options)) {
6213 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
6214 &DAG.getTarget().Options)) {
6215 // Both can be negated for free, check to see if at least one is cheaper
6217 if (LHSNeg == 2 || RHSNeg == 2)
6218 return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT,
6219 GetNegatedExpression(N0, DAG, LegalOperations),
6220 GetNegatedExpression(N1, DAG, LegalOperations));
6227 SDValue DAGCombiner::visitFREM(SDNode *N) {
6228 SDValue N0 = N->getOperand(0);
6229 SDValue N1 = N->getOperand(1);
6230 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6231 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6232 EVT VT = N->getValueType(0);
6234 // fold (frem c1, c2) -> fmod(c1,c2)
6236 return DAG.getNode(ISD::FREM, N->getDebugLoc(), VT, N0, N1);
6241 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
6242 SDValue N0 = N->getOperand(0);
6243 SDValue N1 = N->getOperand(1);
6244 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6245 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6246 EVT VT = N->getValueType(0);
6248 if (N0CFP && N1CFP) // Constant fold
6249 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT, N0, N1);
6252 const APFloat& V = N1CFP->getValueAPF();
6253 // copysign(x, c1) -> fabs(x) iff ispos(c1)
6254 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
6255 if (!V.isNegative()) {
6256 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
6257 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
6259 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6260 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT,
6261 DAG.getNode(ISD::FABS, N0.getDebugLoc(), VT, N0));
6265 // copysign(fabs(x), y) -> copysign(x, y)
6266 // copysign(fneg(x), y) -> copysign(x, y)
6267 // copysign(copysign(x,z), y) -> copysign(x, y)
6268 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
6269 N0.getOpcode() == ISD::FCOPYSIGN)
6270 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
6271 N0.getOperand(0), N1);
6273 // copysign(x, abs(y)) -> abs(x)
6274 if (N1.getOpcode() == ISD::FABS)
6275 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
6277 // copysign(x, copysign(y,z)) -> copysign(x, z)
6278 if (N1.getOpcode() == ISD::FCOPYSIGN)
6279 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
6280 N0, N1.getOperand(1));
6282 // copysign(x, fp_extend(y)) -> copysign(x, y)
6283 // copysign(x, fp_round(y)) -> copysign(x, y)
6284 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
6285 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
6286 N0, N1.getOperand(0));
6291 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
6292 SDValue N0 = N->getOperand(0);
6293 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
6294 EVT VT = N->getValueType(0);
6295 EVT OpVT = N0.getValueType();
6297 // fold (sint_to_fp c1) -> c1fp
6299 // ...but only if the target supports immediate floating-point values
6300 (!LegalOperations ||
6301 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
6302 return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
6304 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
6305 // but UINT_TO_FP is legal on this target, try to convert.
6306 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
6307 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
6308 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
6309 if (DAG.SignBitIsZero(N0))
6310 return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
6313 // The next optimizations are desireable only if SELECT_CC can be lowered.
6314 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
6315 // having to say they don't support SELECT_CC on every type the DAG knows
6316 // about, since there is no way to mark an opcode illegal at all value types
6317 // (See also visitSELECT)
6318 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other)) {
6319 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
6320 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
6322 (!LegalOperations ||
6323 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
6325 { N0.getOperand(0), N0.getOperand(1),
6326 DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT),
6328 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT, Ops, 5);
6331 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
6332 // (select_cc x, y, 1.0, 0.0,, cc)
6333 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
6334 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
6335 (!LegalOperations ||
6336 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
6338 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
6339 DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT),
6340 N0.getOperand(0).getOperand(2) };
6341 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT, Ops, 5);
6348 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
6349 SDValue N0 = N->getOperand(0);
6350 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
6351 EVT VT = N->getValueType(0);
6352 EVT OpVT = N0.getValueType();
6354 // fold (uint_to_fp c1) -> c1fp
6356 // ...but only if the target supports immediate floating-point values
6357 (!LegalOperations ||
6358 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
6359 return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
6361 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
6362 // but SINT_TO_FP is legal on this target, try to convert.
6363 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
6364 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
6365 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
6366 if (DAG.SignBitIsZero(N0))
6367 return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
6370 // The next optimizations are desireable only if SELECT_CC can be lowered.
6371 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
6372 // having to say they don't support SELECT_CC on every type the DAG knows
6373 // about, since there is no way to mark an opcode illegal at all value types
6374 // (See also visitSELECT)
6375 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other)) {
6376 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
6378 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
6379 (!LegalOperations ||
6380 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
6382 { N0.getOperand(0), N0.getOperand(1),
6383 DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT),
6385 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT, Ops, 5);
6392 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
6393 SDValue N0 = N->getOperand(0);
6394 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6395 EVT VT = N->getValueType(0);
6397 // fold (fp_to_sint c1fp) -> c1
6399 return DAG.getNode(ISD::FP_TO_SINT, N->getDebugLoc(), VT, N0);
6404 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
6405 SDValue N0 = N->getOperand(0);
6406 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6407 EVT VT = N->getValueType(0);
6409 // fold (fp_to_uint c1fp) -> c1
6411 return DAG.getNode(ISD::FP_TO_UINT, N->getDebugLoc(), VT, N0);
6416 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
6417 SDValue N0 = N->getOperand(0);
6418 SDValue N1 = N->getOperand(1);
6419 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6420 EVT VT = N->getValueType(0);
6422 // fold (fp_round c1fp) -> c1fp
6424 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT, N0, N1);
6426 // fold (fp_round (fp_extend x)) -> x
6427 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
6428 return N0.getOperand(0);
6430 // fold (fp_round (fp_round x)) -> (fp_round x)
6431 if (N0.getOpcode() == ISD::FP_ROUND) {
6432 // This is a value preserving truncation if both round's are.
6433 bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
6434 N0.getNode()->getConstantOperandVal(1) == 1;
6435 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT, N0.getOperand(0),
6436 DAG.getIntPtrConstant(IsTrunc));
6439 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
6440 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
6441 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(), VT,
6442 N0.getOperand(0), N1);
6443 AddToWorkList(Tmp.getNode());
6444 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
6445 Tmp, N0.getOperand(1));
6451 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
6452 SDValue N0 = N->getOperand(0);
6453 EVT VT = N->getValueType(0);
6454 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
6455 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6457 // fold (fp_round_inreg c1fp) -> c1fp
6458 if (N0CFP && isTypeLegal(EVT)) {
6459 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
6460 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, Round);
6466 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
6467 SDValue N0 = N->getOperand(0);
6468 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6469 EVT VT = N->getValueType(0);
6471 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
6472 if (N->hasOneUse() &&
6473 N->use_begin()->getOpcode() == ISD::FP_ROUND)
6476 // fold (fp_extend c1fp) -> c1fp
6478 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, N0);
6480 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
6482 if (N0.getOpcode() == ISD::FP_ROUND
6483 && N0.getNode()->getConstantOperandVal(1) == 1) {
6484 SDValue In = N0.getOperand(0);
6485 if (In.getValueType() == VT) return In;
6486 if (VT.bitsLT(In.getValueType()))
6487 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT,
6488 In, N0.getOperand(1));
6489 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, In);
6492 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
6493 if (ISD::isNON_EXTLoad(N0.getNode()) && N0.hasOneUse() &&
6494 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6495 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
6496 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6497 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
6499 LN0->getBasePtr(), LN0->getPointerInfo(),
6501 LN0->isVolatile(), LN0->isNonTemporal(),
6502 LN0->getAlignment());
6503 CombineTo(N, ExtLoad);
6504 CombineTo(N0.getNode(),
6505 DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(),
6506 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
6507 ExtLoad.getValue(1));
6508 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6514 SDValue DAGCombiner::visitFNEG(SDNode *N) {
6515 SDValue N0 = N->getOperand(0);
6516 EVT VT = N->getValueType(0);
6518 if (VT.isVector()) {
6519 SDValue FoldedVOp = SimplifyVUnaryOp(N);
6520 if (FoldedVOp.getNode()) return FoldedVOp;
6523 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
6524 &DAG.getTarget().Options))
6525 return GetNegatedExpression(N0, DAG, LegalOperations);
6527 // Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading
6528 // constant pool values.
6529 if (!TLI.isFNegFree(VT) && N0.getOpcode() == ISD::BITCAST &&
6531 N0.getNode()->hasOneUse() &&
6532 N0.getOperand(0).getValueType().isInteger()) {
6533 SDValue Int = N0.getOperand(0);
6534 EVT IntVT = Int.getValueType();
6535 if (IntVT.isInteger() && !IntVT.isVector()) {
6536 Int = DAG.getNode(ISD::XOR, N0.getDebugLoc(), IntVT, Int,
6537 DAG.getConstant(APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
6538 AddToWorkList(Int.getNode());
6539 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
6544 // (fneg (fmul c, x)) -> (fmul -c, x)
6545 if (N0.getOpcode() == ISD::FMUL) {
6546 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
6548 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
6550 DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT,
6558 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
6559 SDValue N0 = N->getOperand(0);
6560 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6561 EVT VT = N->getValueType(0);
6563 // fold (fceil c1) -> fceil(c1)
6565 return DAG.getNode(ISD::FCEIL, N->getDebugLoc(), VT, N0);
6570 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
6571 SDValue N0 = N->getOperand(0);
6572 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6573 EVT VT = N->getValueType(0);
6575 // fold (ftrunc c1) -> ftrunc(c1)
6577 return DAG.getNode(ISD::FTRUNC, N->getDebugLoc(), VT, N0);
6582 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
6583 SDValue N0 = N->getOperand(0);
6584 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6585 EVT VT = N->getValueType(0);
6587 // fold (ffloor c1) -> ffloor(c1)
6589 return DAG.getNode(ISD::FFLOOR, N->getDebugLoc(), VT, N0);
6594 SDValue DAGCombiner::visitFABS(SDNode *N) {
6595 SDValue N0 = N->getOperand(0);
6596 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6597 EVT VT = N->getValueType(0);
6599 if (VT.isVector()) {
6600 SDValue FoldedVOp = SimplifyVUnaryOp(N);
6601 if (FoldedVOp.getNode()) return FoldedVOp;
6604 // fold (fabs c1) -> fabs(c1)
6606 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
6607 // fold (fabs (fabs x)) -> (fabs x)
6608 if (N0.getOpcode() == ISD::FABS)
6609 return N->getOperand(0);
6610 // fold (fabs (fneg x)) -> (fabs x)
6611 // fold (fabs (fcopysign x, y)) -> (fabs x)
6612 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
6613 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0.getOperand(0));
6615 // Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading
6616 // constant pool values.
6617 if (!TLI.isFAbsFree(VT) &&
6618 N0.getOpcode() == ISD::BITCAST && N0.getNode()->hasOneUse() &&
6619 N0.getOperand(0).getValueType().isInteger() &&
6620 !N0.getOperand(0).getValueType().isVector()) {
6621 SDValue Int = N0.getOperand(0);
6622 EVT IntVT = Int.getValueType();
6623 if (IntVT.isInteger() && !IntVT.isVector()) {
6624 Int = DAG.getNode(ISD::AND, N0.getDebugLoc(), IntVT, Int,
6625 DAG.getConstant(~APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
6626 AddToWorkList(Int.getNode());
6627 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
6628 N->getValueType(0), Int);
6635 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
6636 SDValue Chain = N->getOperand(0);
6637 SDValue N1 = N->getOperand(1);
6638 SDValue N2 = N->getOperand(2);
6640 // If N is a constant we could fold this into a fallthrough or unconditional
6641 // branch. However that doesn't happen very often in normal code, because
6642 // Instcombine/SimplifyCFG should have handled the available opportunities.
6643 // If we did this folding here, it would be necessary to update the
6644 // MachineBasicBlock CFG, which is awkward.
6646 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
6648 if (N1.getOpcode() == ISD::SETCC &&
6649 TLI.isOperationLegalOrCustom(ISD::BR_CC, MVT::Other)) {
6650 return DAG.getNode(ISD::BR_CC, N->getDebugLoc(), MVT::Other,
6651 Chain, N1.getOperand(2),
6652 N1.getOperand(0), N1.getOperand(1), N2);
6655 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
6656 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
6657 (N1.getOperand(0).hasOneUse() &&
6658 N1.getOperand(0).getOpcode() == ISD::SRL))) {
6660 if (N1.getOpcode() == ISD::TRUNCATE) {
6661 // Look pass the truncate.
6662 Trunc = N1.getNode();
6663 N1 = N1.getOperand(0);
6666 // Match this pattern so that we can generate simpler code:
6669 // %b = and i32 %a, 2
6670 // %c = srl i32 %b, 1
6671 // brcond i32 %c ...
6676 // %b = and i32 %a, 2
6677 // %c = setcc eq %b, 0
6680 // This applies only when the AND constant value has one bit set and the
6681 // SRL constant is equal to the log2 of the AND constant. The back-end is
6682 // smart enough to convert the result into a TEST/JMP sequence.
6683 SDValue Op0 = N1.getOperand(0);
6684 SDValue Op1 = N1.getOperand(1);
6686 if (Op0.getOpcode() == ISD::AND &&
6687 Op1.getOpcode() == ISD::Constant) {
6688 SDValue AndOp1 = Op0.getOperand(1);
6690 if (AndOp1.getOpcode() == ISD::Constant) {
6691 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
6693 if (AndConst.isPowerOf2() &&
6694 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
6696 DAG.getSetCC(N->getDebugLoc(),
6697 TLI.getSetCCResultType(Op0.getValueType()),
6698 Op0, DAG.getConstant(0, Op0.getValueType()),
6701 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
6702 MVT::Other, Chain, SetCC, N2);
6703 // Don't add the new BRCond into the worklist or else SimplifySelectCC
6704 // will convert it back to (X & C1) >> C2.
6705 CombineTo(N, NewBRCond, false);
6706 // Truncate is dead.
6708 removeFromWorkList(Trunc);
6709 DAG.DeleteNode(Trunc);
6711 // Replace the uses of SRL with SETCC
6712 WorkListRemover DeadNodes(*this);
6713 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
6714 removeFromWorkList(N1.getNode());
6715 DAG.DeleteNode(N1.getNode());
6716 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6722 // Restore N1 if the above transformation doesn't match.
6723 N1 = N->getOperand(1);
6726 // Transform br(xor(x, y)) -> br(x != y)
6727 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
6728 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
6729 SDNode *TheXor = N1.getNode();
6730 SDValue Op0 = TheXor->getOperand(0);
6731 SDValue Op1 = TheXor->getOperand(1);
6732 if (Op0.getOpcode() == Op1.getOpcode()) {
6733 // Avoid missing important xor optimizations.
6734 SDValue Tmp = visitXOR(TheXor);
6735 if (Tmp.getNode() && Tmp.getNode() != TheXor) {
6736 DEBUG(dbgs() << "\nReplacing.8 ";
6738 dbgs() << "\nWith: ";
6739 Tmp.getNode()->dump(&DAG);
6741 WorkListRemover DeadNodes(*this);
6742 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
6743 removeFromWorkList(TheXor);
6744 DAG.DeleteNode(TheXor);
6745 return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
6746 MVT::Other, Chain, Tmp, N2);
6750 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
6752 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
6753 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
6754 Op0.getOpcode() == ISD::XOR) {
6755 TheXor = Op0.getNode();
6759 EVT SetCCVT = N1.getValueType();
6761 SetCCVT = TLI.getSetCCResultType(SetCCVT);
6762 SDValue SetCC = DAG.getSetCC(TheXor->getDebugLoc(),
6765 Equal ? ISD::SETEQ : ISD::SETNE);
6766 // Replace the uses of XOR with SETCC
6767 WorkListRemover DeadNodes(*this);
6768 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
6769 removeFromWorkList(N1.getNode());
6770 DAG.DeleteNode(N1.getNode());
6771 return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
6772 MVT::Other, Chain, SetCC, N2);
6779 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
6781 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
6782 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
6783 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
6785 // If N is a constant we could fold this into a fallthrough or unconditional
6786 // branch. However that doesn't happen very often in normal code, because
6787 // Instcombine/SimplifyCFG should have handled the available opportunities.
6788 // If we did this folding here, it would be necessary to update the
6789 // MachineBasicBlock CFG, which is awkward.
6791 // Use SimplifySetCC to simplify SETCC's.
6792 SDValue Simp = SimplifySetCC(TLI.getSetCCResultType(CondLHS.getValueType()),
6793 CondLHS, CondRHS, CC->get(), N->getDebugLoc(),
6795 if (Simp.getNode()) AddToWorkList(Simp.getNode());
6797 // fold to a simpler setcc
6798 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
6799 return DAG.getNode(ISD::BR_CC, N->getDebugLoc(), MVT::Other,
6800 N->getOperand(0), Simp.getOperand(2),
6801 Simp.getOperand(0), Simp.getOperand(1),
6807 /// canFoldInAddressingMode - Return true if 'Use' is a load or a store that
6808 /// uses N as its base pointer and that N may be folded in the load / store
6809 /// addressing mode.
6810 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
6812 const TargetLowering &TLI) {
6814 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
6815 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
6817 VT = Use->getValueType(0);
6818 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
6819 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
6821 VT = ST->getValue().getValueType();
6826 if (N->getOpcode() == ISD::ADD) {
6827 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
6830 AM.BaseOffs = Offset->getSExtValue();
6834 } else if (N->getOpcode() == ISD::SUB) {
6835 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
6838 AM.BaseOffs = -Offset->getSExtValue();
6845 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
6848 /// CombineToPreIndexedLoadStore - Try turning a load / store into a
6849 /// pre-indexed load / store when the base pointer is an add or subtract
6850 /// and it has other uses besides the load / store. After the
6851 /// transformation, the new indexed load / store has effectively folded
6852 /// the add / subtract in and all of its other uses are redirected to the
6853 /// new load / store.
6854 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
6855 if (Level < AfterLegalizeDAG)
6861 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6862 if (LD->isIndexed())
6864 VT = LD->getMemoryVT();
6865 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
6866 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
6868 Ptr = LD->getBasePtr();
6869 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6870 if (ST->isIndexed())
6872 VT = ST->getMemoryVT();
6873 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
6874 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
6876 Ptr = ST->getBasePtr();
6882 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
6883 // out. There is no reason to make this a preinc/predec.
6884 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
6885 Ptr.getNode()->hasOneUse())
6888 // Ask the target to do addressing mode selection.
6891 ISD::MemIndexedMode AM = ISD::UNINDEXED;
6892 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
6894 // Don't create a indexed load / store with zero offset.
6895 if (isa<ConstantSDNode>(Offset) &&
6896 cast<ConstantSDNode>(Offset)->isNullValue())
6899 // Try turning it into a pre-indexed load / store except when:
6900 // 1) The new base ptr is a frame index.
6901 // 2) If N is a store and the new base ptr is either the same as or is a
6902 // predecessor of the value being stored.
6903 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
6904 // that would create a cycle.
6905 // 4) All uses are load / store ops that use it as old base ptr.
6907 // Check #1. Preinc'ing a frame index would require copying the stack pointer
6908 // (plus the implicit offset) to a register to preinc anyway.
6909 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
6914 SDValue Val = cast<StoreSDNode>(N)->getValue();
6915 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
6919 // Now check for #3 and #4.
6920 bool RealUse = false;
6922 // Caches for hasPredecessorHelper
6923 SmallPtrSet<const SDNode *, 32> Visited;
6924 SmallVector<const SDNode *, 16> Worklist;
6926 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
6927 E = Ptr.getNode()->use_end(); I != E; ++I) {
6931 if (N->hasPredecessorHelper(Use, Visited, Worklist))
6934 // If Ptr may be folded in addressing mode of other use, then it's
6935 // not profitable to do this transformation.
6936 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
6945 Result = DAG.getIndexedLoad(SDValue(N,0), N->getDebugLoc(),
6946 BasePtr, Offset, AM);
6948 Result = DAG.getIndexedStore(SDValue(N,0), N->getDebugLoc(),
6949 BasePtr, Offset, AM);
6952 DEBUG(dbgs() << "\nReplacing.4 ";
6954 dbgs() << "\nWith: ";
6955 Result.getNode()->dump(&DAG);
6957 WorkListRemover DeadNodes(*this);
6959 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
6960 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
6962 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
6965 // Finally, since the node is now dead, remove it from the graph.
6968 // Replace the uses of Ptr with uses of the updated base value.
6969 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
6970 removeFromWorkList(Ptr.getNode());
6971 DAG.DeleteNode(Ptr.getNode());
6976 /// CombineToPostIndexedLoadStore - Try to combine a load / store with a
6977 /// add / sub of the base pointer node into a post-indexed load / store.
6978 /// The transformation folded the add / subtract into the new indexed
6979 /// load / store effectively and all of its uses are redirected to the
6980 /// new load / store.
6981 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
6982 if (Level < AfterLegalizeDAG)
6988 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6989 if (LD->isIndexed())
6991 VT = LD->getMemoryVT();
6992 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
6993 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
6995 Ptr = LD->getBasePtr();
6996 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6997 if (ST->isIndexed())
6999 VT = ST->getMemoryVT();
7000 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
7001 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
7003 Ptr = ST->getBasePtr();
7009 if (Ptr.getNode()->hasOneUse())
7012 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
7013 E = Ptr.getNode()->use_end(); I != E; ++I) {
7016 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
7021 ISD::MemIndexedMode AM = ISD::UNINDEXED;
7022 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
7023 // Don't create a indexed load / store with zero offset.
7024 if (isa<ConstantSDNode>(Offset) &&
7025 cast<ConstantSDNode>(Offset)->isNullValue())
7028 // Try turning it into a post-indexed load / store except when
7029 // 1) All uses are load / store ops that use it as base ptr (and
7030 // it may be folded as addressing mmode).
7031 // 2) Op must be independent of N, i.e. Op is neither a predecessor
7032 // nor a successor of N. Otherwise, if Op is folded that would
7035 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
7039 bool TryNext = false;
7040 for (SDNode::use_iterator II = BasePtr.getNode()->use_begin(),
7041 EE = BasePtr.getNode()->use_end(); II != EE; ++II) {
7043 if (Use == Ptr.getNode())
7046 // If all the uses are load / store addresses, then don't do the
7048 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
7049 bool RealUse = false;
7050 for (SDNode::use_iterator III = Use->use_begin(),
7051 EEE = Use->use_end(); III != EEE; ++III) {
7052 SDNode *UseUse = *III;
7053 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
7068 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
7069 SDValue Result = isLoad
7070 ? DAG.getIndexedLoad(SDValue(N,0), N->getDebugLoc(),
7071 BasePtr, Offset, AM)
7072 : DAG.getIndexedStore(SDValue(N,0), N->getDebugLoc(),
7073 BasePtr, Offset, AM);
7076 DEBUG(dbgs() << "\nReplacing.5 ";
7078 dbgs() << "\nWith: ";
7079 Result.getNode()->dump(&DAG);
7081 WorkListRemover DeadNodes(*this);
7083 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
7084 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
7086 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
7089 // Finally, since the node is now dead, remove it from the graph.
7092 // Replace the uses of Use with uses of the updated base value.
7093 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
7094 Result.getValue(isLoad ? 1 : 0));
7095 removeFromWorkList(Op);
7105 SDValue DAGCombiner::visitLOAD(SDNode *N) {
7106 LoadSDNode *LD = cast<LoadSDNode>(N);
7107 SDValue Chain = LD->getChain();
7108 SDValue Ptr = LD->getBasePtr();
7110 // If load is not volatile and there are no uses of the loaded value (and
7111 // the updated indexed value in case of indexed loads), change uses of the
7112 // chain value into uses of the chain input (i.e. delete the dead load).
7113 if (!LD->isVolatile()) {
7114 if (N->getValueType(1) == MVT::Other) {
7116 if (!N->hasAnyUseOfValue(0)) {
7117 // It's not safe to use the two value CombineTo variant here. e.g.
7118 // v1, chain2 = load chain1, loc
7119 // v2, chain3 = load chain2, loc
7121 // Now we replace use of chain2 with chain1. This makes the second load
7122 // isomorphic to the one we are deleting, and thus makes this load live.
7123 DEBUG(dbgs() << "\nReplacing.6 ";
7125 dbgs() << "\nWith chain: ";
7126 Chain.getNode()->dump(&DAG);
7128 WorkListRemover DeadNodes(*this);
7129 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
7131 if (N->use_empty()) {
7132 removeFromWorkList(N);
7136 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7140 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
7141 if (!N->hasAnyUseOfValue(0) && !N->hasAnyUseOfValue(1)) {
7142 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
7143 DEBUG(dbgs() << "\nReplacing.7 ";
7145 dbgs() << "\nWith: ";
7146 Undef.getNode()->dump(&DAG);
7147 dbgs() << " and 2 other values\n");
7148 WorkListRemover DeadNodes(*this);
7149 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
7150 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1),
7151 DAG.getUNDEF(N->getValueType(1)));
7152 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
7153 removeFromWorkList(N);
7155 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7160 // If this load is directly stored, replace the load value with the stored
7162 // TODO: Handle store large -> read small portion.
7163 // TODO: Handle TRUNCSTORE/LOADEXT
7164 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
7165 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
7166 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
7167 if (PrevST->getBasePtr() == Ptr &&
7168 PrevST->getValue().getValueType() == N->getValueType(0))
7169 return CombineTo(N, Chain.getOperand(1), Chain);
7173 // Try to infer better alignment information than the load already has.
7174 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
7175 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
7176 if (Align > LD->getAlignment())
7177 return DAG.getExtLoad(LD->getExtensionType(), N->getDebugLoc(),
7178 LD->getValueType(0),
7179 Chain, Ptr, LD->getPointerInfo(),
7181 LD->isVolatile(), LD->isNonTemporal(), Align);
7186 // Walk up chain skipping non-aliasing memory nodes.
7187 SDValue BetterChain = FindBetterChain(N, Chain);
7189 // If there is a better chain.
7190 if (Chain != BetterChain) {
7193 // Replace the chain to void dependency.
7194 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
7195 ReplLoad = DAG.getLoad(N->getValueType(0), LD->getDebugLoc(),
7196 BetterChain, Ptr, LD->getPointerInfo(),
7197 LD->isVolatile(), LD->isNonTemporal(),
7198 LD->isInvariant(), LD->getAlignment());
7200 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), LD->getDebugLoc(),
7201 LD->getValueType(0),
7202 BetterChain, Ptr, LD->getPointerInfo(),
7205 LD->isNonTemporal(),
7206 LD->getAlignment());
7209 // Create token factor to keep old chain connected.
7210 SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
7211 MVT::Other, Chain, ReplLoad.getValue(1));
7213 // Make sure the new and old chains are cleaned up.
7214 AddToWorkList(Token.getNode());
7216 // Replace uses with load result and token factor. Don't add users
7218 return CombineTo(N, ReplLoad.getValue(0), Token, false);
7222 // Try transforming N to an indexed load.
7223 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
7224 return SDValue(N, 0);
7229 /// CheckForMaskedLoad - Check to see if V is (and load (ptr), imm), where the
7230 /// load is having specific bytes cleared out. If so, return the byte size
7231 /// being masked out and the shift amount.
7232 static std::pair<unsigned, unsigned>
7233 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
7234 std::pair<unsigned, unsigned> Result(0, 0);
7236 // Check for the structure we're looking for.
7237 if (V->getOpcode() != ISD::AND ||
7238 !isa<ConstantSDNode>(V->getOperand(1)) ||
7239 !ISD::isNormalLoad(V->getOperand(0).getNode()))
7242 // Check the chain and pointer.
7243 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
7244 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
7246 // The store should be chained directly to the load or be an operand of a
7248 if (LD == Chain.getNode())
7250 else if (Chain->getOpcode() != ISD::TokenFactor)
7251 return Result; // Fail.
7254 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
7255 if (Chain->getOperand(i).getNode() == LD) {
7259 if (!isOk) return Result;
7262 // This only handles simple types.
7263 if (V.getValueType() != MVT::i16 &&
7264 V.getValueType() != MVT::i32 &&
7265 V.getValueType() != MVT::i64)
7268 // Check the constant mask. Invert it so that the bits being masked out are
7269 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
7270 // follow the sign bit for uniformity.
7271 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
7272 unsigned NotMaskLZ = CountLeadingZeros_64(NotMask);
7273 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
7274 unsigned NotMaskTZ = CountTrailingZeros_64(NotMask);
7275 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
7276 if (NotMaskLZ == 64) return Result; // All zero mask.
7278 // See if we have a continuous run of bits. If so, we have 0*1+0*
7279 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
7282 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
7283 if (V.getValueType() != MVT::i64 && NotMaskLZ)
7284 NotMaskLZ -= 64-V.getValueSizeInBits();
7286 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
7287 switch (MaskedBytes) {
7291 default: return Result; // All one mask, or 5-byte mask.
7294 // Verify that the first bit starts at a multiple of mask so that the access
7295 // is aligned the same as the access width.
7296 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
7298 Result.first = MaskedBytes;
7299 Result.second = NotMaskTZ/8;
7304 /// ShrinkLoadReplaceStoreWithStore - Check to see if IVal is something that
7305 /// provides a value as specified by MaskInfo. If so, replace the specified
7306 /// store with a narrower store of truncated IVal.
7308 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
7309 SDValue IVal, StoreSDNode *St,
7311 unsigned NumBytes = MaskInfo.first;
7312 unsigned ByteShift = MaskInfo.second;
7313 SelectionDAG &DAG = DC->getDAG();
7315 // Check to see if IVal is all zeros in the part being masked in by the 'or'
7316 // that uses this. If not, this is not a replacement.
7317 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
7318 ByteShift*8, (ByteShift+NumBytes)*8);
7319 if (!DAG.MaskedValueIsZero(IVal, Mask)) return 0;
7321 // Check that it is legal on the target to do this. It is legal if the new
7322 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
7324 MVT VT = MVT::getIntegerVT(NumBytes*8);
7325 if (!DC->isTypeLegal(VT))
7328 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
7329 // shifted by ByteShift and truncated down to NumBytes.
7331 IVal = DAG.getNode(ISD::SRL, IVal->getDebugLoc(), IVal.getValueType(), IVal,
7332 DAG.getConstant(ByteShift*8,
7333 DC->getShiftAmountTy(IVal.getValueType())));
7335 // Figure out the offset for the store and the alignment of the access.
7337 unsigned NewAlign = St->getAlignment();
7339 if (DAG.getTargetLoweringInfo().isLittleEndian())
7340 StOffset = ByteShift;
7342 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
7344 SDValue Ptr = St->getBasePtr();
7346 Ptr = DAG.getNode(ISD::ADD, IVal->getDebugLoc(), Ptr.getValueType(),
7347 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
7348 NewAlign = MinAlign(NewAlign, StOffset);
7351 // Truncate down to the new size.
7352 IVal = DAG.getNode(ISD::TRUNCATE, IVal->getDebugLoc(), VT, IVal);
7355 return DAG.getStore(St->getChain(), St->getDebugLoc(), IVal, Ptr,
7356 St->getPointerInfo().getWithOffset(StOffset),
7357 false, false, NewAlign).getNode();
7361 /// ReduceLoadOpStoreWidth - Look for sequence of load / op / store where op is
7362 /// one of 'or', 'xor', and 'and' of immediates. If 'op' is only touching some
7363 /// of the loaded bits, try narrowing the load and store if it would end up
7364 /// being a win for performance or code size.
7365 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
7366 StoreSDNode *ST = cast<StoreSDNode>(N);
7367 if (ST->isVolatile())
7370 SDValue Chain = ST->getChain();
7371 SDValue Value = ST->getValue();
7372 SDValue Ptr = ST->getBasePtr();
7373 EVT VT = Value.getValueType();
7375 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
7378 unsigned Opc = Value.getOpcode();
7380 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
7381 // is a byte mask indicating a consecutive number of bytes, check to see if
7382 // Y is known to provide just those bytes. If so, we try to replace the
7383 // load + replace + store sequence with a single (narrower) store, which makes
7385 if (Opc == ISD::OR) {
7386 std::pair<unsigned, unsigned> MaskedLoad;
7387 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
7388 if (MaskedLoad.first)
7389 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
7390 Value.getOperand(1), ST,this))
7391 return SDValue(NewST, 0);
7393 // Or is commutative, so try swapping X and Y.
7394 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
7395 if (MaskedLoad.first)
7396 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
7397 Value.getOperand(0), ST,this))
7398 return SDValue(NewST, 0);
7401 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
7402 Value.getOperand(1).getOpcode() != ISD::Constant)
7405 SDValue N0 = Value.getOperand(0);
7406 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7407 Chain == SDValue(N0.getNode(), 1)) {
7408 LoadSDNode *LD = cast<LoadSDNode>(N0);
7409 if (LD->getBasePtr() != Ptr ||
7410 LD->getPointerInfo().getAddrSpace() !=
7411 ST->getPointerInfo().getAddrSpace())
7414 // Find the type to narrow it the load / op / store to.
7415 SDValue N1 = Value.getOperand(1);
7416 unsigned BitWidth = N1.getValueSizeInBits();
7417 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
7418 if (Opc == ISD::AND)
7419 Imm ^= APInt::getAllOnesValue(BitWidth);
7420 if (Imm == 0 || Imm.isAllOnesValue())
7422 unsigned ShAmt = Imm.countTrailingZeros();
7423 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
7424 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
7425 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
7426 while (NewBW < BitWidth &&
7427 !(TLI.isOperationLegalOrCustom(Opc, NewVT) &&
7428 TLI.isNarrowingProfitable(VT, NewVT))) {
7429 NewBW = NextPowerOf2(NewBW);
7430 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
7432 if (NewBW >= BitWidth)
7435 // If the lsb changed does not start at the type bitwidth boundary,
7436 // start at the previous one.
7438 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
7439 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt, ShAmt + NewBW);
7440 if ((Imm & Mask) == Imm) {
7441 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
7442 if (Opc == ISD::AND)
7443 NewImm ^= APInt::getAllOnesValue(NewBW);
7444 uint64_t PtrOff = ShAmt / 8;
7445 // For big endian targets, we need to adjust the offset to the pointer to
7446 // load the correct bytes.
7447 if (TLI.isBigEndian())
7448 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
7450 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
7451 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
7452 if (NewAlign < TLI.getDataLayout()->getABITypeAlignment(NewVTTy))
7455 SDValue NewPtr = DAG.getNode(ISD::ADD, LD->getDebugLoc(),
7456 Ptr.getValueType(), Ptr,
7457 DAG.getConstant(PtrOff, Ptr.getValueType()));
7458 SDValue NewLD = DAG.getLoad(NewVT, N0.getDebugLoc(),
7459 LD->getChain(), NewPtr,
7460 LD->getPointerInfo().getWithOffset(PtrOff),
7461 LD->isVolatile(), LD->isNonTemporal(),
7462 LD->isInvariant(), NewAlign);
7463 SDValue NewVal = DAG.getNode(Opc, Value.getDebugLoc(), NewVT, NewLD,
7464 DAG.getConstant(NewImm, NewVT));
7465 SDValue NewST = DAG.getStore(Chain, N->getDebugLoc(),
7467 ST->getPointerInfo().getWithOffset(PtrOff),
7468 false, false, NewAlign);
7470 AddToWorkList(NewPtr.getNode());
7471 AddToWorkList(NewLD.getNode());
7472 AddToWorkList(NewVal.getNode());
7473 WorkListRemover DeadNodes(*this);
7474 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
7483 /// TransformFPLoadStorePair - For a given floating point load / store pair,
7484 /// if the load value isn't used by any other operations, then consider
7485 /// transforming the pair to integer load / store operations if the target
7486 /// deems the transformation profitable.
7487 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
7488 StoreSDNode *ST = cast<StoreSDNode>(N);
7489 SDValue Chain = ST->getChain();
7490 SDValue Value = ST->getValue();
7491 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
7492 Value.hasOneUse() &&
7493 Chain == SDValue(Value.getNode(), 1)) {
7494 LoadSDNode *LD = cast<LoadSDNode>(Value);
7495 EVT VT = LD->getMemoryVT();
7496 if (!VT.isFloatingPoint() ||
7497 VT != ST->getMemoryVT() ||
7498 LD->isNonTemporal() ||
7499 ST->isNonTemporal() ||
7500 LD->getPointerInfo().getAddrSpace() != 0 ||
7501 ST->getPointerInfo().getAddrSpace() != 0)
7504 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
7505 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
7506 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
7507 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
7508 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
7511 unsigned LDAlign = LD->getAlignment();
7512 unsigned STAlign = ST->getAlignment();
7513 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
7514 unsigned ABIAlign = TLI.getDataLayout()->getABITypeAlignment(IntVTTy);
7515 if (LDAlign < ABIAlign || STAlign < ABIAlign)
7518 SDValue NewLD = DAG.getLoad(IntVT, Value.getDebugLoc(),
7519 LD->getChain(), LD->getBasePtr(),
7520 LD->getPointerInfo(),
7521 false, false, false, LDAlign);
7523 SDValue NewST = DAG.getStore(NewLD.getValue(1), N->getDebugLoc(),
7524 NewLD, ST->getBasePtr(),
7525 ST->getPointerInfo(),
7526 false, false, STAlign);
7528 AddToWorkList(NewLD.getNode());
7529 AddToWorkList(NewST.getNode());
7530 WorkListRemover DeadNodes(*this);
7531 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
7539 /// Returns the base pointer and an integer offset from that object.
7540 static std::pair<SDValue, int64_t> GetPointerBaseAndOffset(SDValue Ptr) {
7541 if (Ptr->getOpcode() == ISD::ADD && isa<ConstantSDNode>(Ptr->getOperand(1))) {
7542 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
7543 SDValue Base = Ptr->getOperand(0);
7544 return std::make_pair(Base, Offset);
7547 return std::make_pair(Ptr, 0);
7550 /// Holds a pointer to an LSBaseSDNode as well as information on where it
7551 /// is located in a sequence of memory operations connected by a chain.
7553 MemOpLink (LSBaseSDNode *N, int64_t Offset, unsigned Seq):
7554 MemNode(N), OffsetFromBase(Offset), SequenceNum(Seq) { }
7555 // Ptr to the mem node.
7556 LSBaseSDNode *MemNode;
7557 // Offset from the base ptr.
7558 int64_t OffsetFromBase;
7559 // What is the sequence number of this mem node.
7560 // Lowest mem operand in the DAG starts at zero.
7561 unsigned SequenceNum;
7564 /// Sorts store nodes in a link according to their offset from a shared
7566 struct ConsecutiveMemoryChainSorter {
7567 bool operator()(MemOpLink LHS, MemOpLink RHS) {
7568 return LHS.OffsetFromBase < RHS.OffsetFromBase;
7572 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
7573 EVT MemVT = St->getMemoryVT();
7574 int64_t ElementSizeBytes = MemVT.getSizeInBits()/8;
7576 // Don't merge vectors into wider inputs.
7577 if (MemVT.isVector() || !MemVT.isSimple())
7580 // Perform an early exit check. Do not bother looking at stored values that
7581 // are not constants or loads.
7582 SDValue StoredVal = St->getValue();
7583 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
7584 if (!isa<ConstantSDNode>(StoredVal) && !isa<ConstantFPSDNode>(StoredVal) &&
7588 // Only look at ends of store sequences.
7589 SDValue Chain = SDValue(St, 1);
7590 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
7593 // This holds the base pointer and the offset in bytes from the base pointer.
7594 std::pair<SDValue, int64_t> BasePtr =
7595 GetPointerBaseAndOffset(St->getBasePtr());
7597 // We must have a base and an offset.
7598 if (!BasePtr.first.getNode())
7601 // Do not handle stores to undef base pointers.
7602 if (BasePtr.first.getOpcode() == ISD::UNDEF)
7605 // Save the LoadSDNodes that we find in the chain.
7606 // We need to make sure that these nodes do not interfere with
7607 // any of the store nodes.
7608 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
7610 // Save the StoreSDNodes that we find in the chain.
7611 SmallVector<MemOpLink, 8> StoreNodes;
7613 // Walk up the chain and look for nodes with offsets from the same
7614 // base pointer. Stop when reaching an instruction with a different kind
7615 // or instruction which has a different base pointer.
7617 StoreSDNode *Index = St;
7619 // If the chain has more than one use, then we can't reorder the mem ops.
7620 if (Index != St && !SDValue(Index, 1)->hasOneUse())
7623 // Find the base pointer and offset for this memory node.
7624 std::pair<SDValue, int64_t> Ptr =
7625 GetPointerBaseAndOffset(Index->getBasePtr());
7627 // Check that the base pointer is the same as the original one.
7628 if (Ptr.first.getNode() != BasePtr.first.getNode())
7631 // Check that the alignment is the same.
7632 if (Index->getAlignment() != St->getAlignment())
7635 // The memory operands must not be volatile.
7636 if (Index->isVolatile() || Index->isIndexed())
7640 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
7641 if (St->isTruncatingStore())
7644 // The stored memory type must be the same.
7645 if (Index->getMemoryVT() != MemVT)
7648 // We do not allow unaligned stores because we want to prevent overriding
7650 if (Index->getAlignment()*8 != MemVT.getSizeInBits())
7653 // We found a potential memory operand to merge.
7654 StoreNodes.push_back(MemOpLink(Index, Ptr.second, Seq++));
7656 // Find the next memory operand in the chain. If the next operand in the
7657 // chain is a store then move up and continue the scan with the next
7658 // memory operand. If the next operand is a load save it and use alias
7659 // information to check if it interferes with anything.
7660 SDNode *NextInChain = Index->getChain().getNode();
7662 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
7663 // We found a store node. Use it for the next iteration.
7666 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
7667 // Save the load node for later. Continue the scan.
7668 AliasLoadNodes.push_back(Ldn);
7669 NextInChain = Ldn->getChain().getNode();
7678 // Check if there is anything to merge.
7679 if (StoreNodes.size() < 2)
7682 // Sort the memory operands according to their distance from the base pointer.
7683 std::sort(StoreNodes.begin(), StoreNodes.end(),
7684 ConsecutiveMemoryChainSorter());
7686 // Scan the memory operations on the chain and find the first non-consecutive
7687 // store memory address.
7688 unsigned LastConsecutiveStore = 0;
7689 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
7690 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) {
7692 // Check that the addresses are consecutive starting from the second
7693 // element in the list of stores.
7695 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
7696 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
7701 // Check if this store interferes with any of the loads that we found.
7702 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld)
7703 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) {
7707 // We found a load that alias with this store. Stop the sequence.
7711 // Mark this node as useful.
7712 LastConsecutiveStore = i;
7715 // The node with the lowest store address.
7716 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
7718 // Store the constants into memory as one consecutive store.
7720 unsigned LastLegalType = 0;
7721 unsigned LastLegalVectorType = 0;
7722 bool NonZero = false;
7723 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
7724 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
7725 SDValue StoredVal = St->getValue();
7727 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
7728 NonZero |= !C->isNullValue();
7729 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) {
7730 NonZero |= !C->getConstantFPValue()->isNullValue();
7736 // Find a legal type for the constant store.
7737 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
7738 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
7739 if (TLI.isTypeLegal(StoreTy))
7740 LastLegalType = i+1;
7742 // Find a legal type for the vector store.
7743 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
7744 if (TLI.isTypeLegal(Ty))
7745 LastLegalVectorType = i + 1;
7748 // We only use vectors if the constant is known to be zero.
7750 LastLegalVectorType = 0;
7752 // Check if we found a legal integer type to store.
7753 if (LastLegalType == 0 && LastLegalVectorType == 0)
7756 bool UseVector = LastLegalVectorType > LastLegalType;
7757 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType;
7759 // Make sure we have something to merge.
7763 unsigned EarliestNodeUsed = 0;
7764 for (unsigned i=0; i < NumElem; ++i) {
7765 // Find a chain for the new wide-store operand. Notice that some
7766 // of the store nodes that we found may not be selected for inclusion
7767 // in the wide store. The chain we use needs to be the chain of the
7768 // earliest store node which is *used* and replaced by the wide store.
7769 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
7770 EarliestNodeUsed = i;
7773 // The earliest Node in the DAG.
7774 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
7775 DebugLoc DL = StoreNodes[0].MemNode->getDebugLoc();
7779 // Find a legal type for the vector store.
7780 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
7781 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
7782 StoredVal = DAG.getConstant(0, Ty);
7784 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
7785 APInt StoreInt(StoreBW, 0);
7787 // Construct a single integer constant which is made of the smaller
7789 bool IsLE = TLI.isLittleEndian();
7790 for (unsigned i = 0; i < NumElem ; ++i) {
7791 unsigned Idx = IsLE ?(NumElem - 1 - i) : i;
7792 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
7793 SDValue Val = St->getValue();
7794 StoreInt<<=ElementSizeBytes*8;
7795 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
7796 StoreInt|=C->getAPIntValue().zext(StoreBW);
7797 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
7798 StoreInt|= C->getValueAPF().bitcastToAPInt().zext(StoreBW);
7800 assert(false && "Invalid constant element type");
7804 // Create the new Load and Store operations.
7805 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
7806 StoredVal = DAG.getConstant(StoreInt, StoreTy);
7809 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), DL, StoredVal,
7810 FirstInChain->getBasePtr(),
7811 FirstInChain->getPointerInfo(),
7813 FirstInChain->getAlignment());
7815 // Replace the first store with the new store
7816 CombineTo(EarliestOp, NewStore);
7817 // Erase all other stores.
7818 for (unsigned i = 0; i < NumElem ; ++i) {
7819 if (StoreNodes[i].MemNode == EarliestOp)
7821 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
7822 // ReplaceAllUsesWith will replace all uses that existed when it was
7823 // called, but graph optimizations may cause new ones to appear. For
7824 // example, the case in pr14333 looks like
7826 // St's chain -> St -> another store -> X
7828 // And the only difference from St to the other store is the chain.
7829 // When we change it's chain to be St's chain they become identical,
7830 // get CSEed and the net result is that X is now a use of St.
7831 // Since we know that St is redundant, just iterate.
7832 while (!St->use_empty())
7833 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain());
7834 removeFromWorkList(St);
7841 // Below we handle the case of multiple consecutive stores that
7842 // come from multiple consecutive loads. We merge them into a single
7843 // wide load and a single wide store.
7845 // Look for load nodes which are used by the stored values.
7846 SmallVector<MemOpLink, 8> LoadNodes;
7848 // Find acceptable loads. Loads need to have the same chain (token factor),
7849 // must not be zext, volatile, indexed, and they must be consecutive.
7851 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
7852 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
7853 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
7856 // Loads must only have one use.
7857 if (!Ld->hasNUsesOfValue(1, 0))
7860 // Check that the alignment is the same as the stores.
7861 if (Ld->getAlignment() != St->getAlignment())
7864 // The memory operands must not be volatile.
7865 if (Ld->isVolatile() || Ld->isIndexed())
7868 // We do not accept ext loads.
7869 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
7872 // The stored memory type must be the same.
7873 if (Ld->getMemoryVT() != MemVT)
7876 std::pair<SDValue, int64_t> LdPtr =
7877 GetPointerBaseAndOffset(Ld->getBasePtr());
7879 // If this is not the first ptr that we check.
7880 if (LdBasePtr.getNode()) {
7881 // The base ptr must be the same.
7882 if (LdPtr.first != LdBasePtr)
7885 // Check that all other base pointers are the same as this one.
7886 LdBasePtr = LdPtr.first;
7889 // We found a potential memory operand to merge.
7890 LoadNodes.push_back(MemOpLink(Ld, LdPtr.second, 0));
7893 if (LoadNodes.size() < 2)
7896 // Scan the memory operations on the chain and find the first non-consecutive
7897 // load memory address. These variables hold the index in the store node
7899 unsigned LastConsecutiveLoad = 0;
7900 // This variable refers to the size and not index in the array.
7901 unsigned LastLegalVectorType = 0;
7902 unsigned LastLegalIntegerType = 0;
7903 StartAddress = LoadNodes[0].OffsetFromBase;
7904 SDValue FirstChain = LoadNodes[0].MemNode->getChain();
7905 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
7906 // All loads much share the same chain.
7907 if (LoadNodes[i].MemNode->getChain() != FirstChain)
7910 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
7911 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
7913 LastConsecutiveLoad = i;
7915 // Find a legal type for the vector store.
7916 EVT StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
7917 if (TLI.isTypeLegal(StoreTy))
7918 LastLegalVectorType = i + 1;
7920 // Find a legal type for the integer store.
7921 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
7922 StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
7923 if (TLI.isTypeLegal(StoreTy))
7924 LastLegalIntegerType = i + 1;
7927 // Only use vector types if the vector type is larger than the integer type.
7928 // If they are the same, use integers.
7929 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType;
7930 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType);
7932 // We add +1 here because the LastXXX variables refer to location while
7933 // the NumElem refers to array/index size.
7934 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1;
7935 NumElem = std::min(LastLegalType, NumElem);
7940 // The earliest Node in the DAG.
7941 unsigned EarliestNodeUsed = 0;
7942 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
7943 for (unsigned i=1; i<NumElem; ++i) {
7944 // Find a chain for the new wide-store operand. Notice that some
7945 // of the store nodes that we found may not be selected for inclusion
7946 // in the wide store. The chain we use needs to be the chain of the
7947 // earliest store node which is *used* and replaced by the wide store.
7948 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
7949 EarliestNodeUsed = i;
7952 // Find if it is better to use vectors or integers to load and store
7956 JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
7958 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
7959 JointMemOpVT = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
7962 DebugLoc LoadDL = LoadNodes[0].MemNode->getDebugLoc();
7963 DebugLoc StoreDL = StoreNodes[0].MemNode->getDebugLoc();
7965 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
7966 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL,
7967 FirstLoad->getChain(),
7968 FirstLoad->getBasePtr(),
7969 FirstLoad->getPointerInfo(),
7970 false, false, false,
7971 FirstLoad->getAlignment());
7973 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), StoreDL, NewLoad,
7974 FirstInChain->getBasePtr(),
7975 FirstInChain->getPointerInfo(), false, false,
7976 FirstInChain->getAlignment());
7978 // Replace one of the loads with the new load.
7979 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode);
7980 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
7981 SDValue(NewLoad.getNode(), 1));
7983 // Remove the rest of the load chains.
7984 for (unsigned i = 1; i < NumElem ; ++i) {
7985 // Replace all chain users of the old load nodes with the chain of the new
7987 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
7988 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain());
7991 // Replace the first store with the new store.
7992 CombineTo(EarliestOp, NewStore);
7993 // Erase all other stores.
7994 for (unsigned i = 0; i < NumElem ; ++i) {
7995 // Remove all Store nodes.
7996 if (StoreNodes[i].MemNode == EarliestOp)
7998 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
7999 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain());
8000 removeFromWorkList(St);
8007 SDValue DAGCombiner::visitSTORE(SDNode *N) {
8008 StoreSDNode *ST = cast<StoreSDNode>(N);
8009 SDValue Chain = ST->getChain();
8010 SDValue Value = ST->getValue();
8011 SDValue Ptr = ST->getBasePtr();
8013 // If this is a store of a bit convert, store the input value if the
8014 // resultant store does not need a higher alignment than the original.
8015 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
8016 ST->isUnindexed()) {
8017 unsigned OrigAlign = ST->getAlignment();
8018 EVT SVT = Value.getOperand(0).getValueType();
8019 unsigned Align = TLI.getDataLayout()->
8020 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
8021 if (Align <= OrigAlign &&
8022 ((!LegalOperations && !ST->isVolatile()) ||
8023 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
8024 return DAG.getStore(Chain, N->getDebugLoc(), Value.getOperand(0),
8025 Ptr, ST->getPointerInfo(), ST->isVolatile(),
8026 ST->isNonTemporal(), OrigAlign);
8029 // Turn 'store undef, Ptr' -> nothing.
8030 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
8033 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
8034 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
8035 // NOTE: If the original store is volatile, this transform must not increase
8036 // the number of stores. For example, on x86-32 an f64 can be stored in one
8037 // processor operation but an i64 (which is not legal) requires two. So the
8038 // transform should not be done in this case.
8039 if (Value.getOpcode() != ISD::TargetConstantFP) {
8041 switch (CFP->getValueType(0).getSimpleVT().SimpleTy) {
8042 default: llvm_unreachable("Unknown FP type");
8043 case MVT::f16: // We don't do this for these yet.
8049 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
8050 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
8051 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
8052 bitcastToAPInt().getZExtValue(), MVT::i32);
8053 return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
8054 Ptr, ST->getPointerInfo(), ST->isVolatile(),
8055 ST->isNonTemporal(), ST->getAlignment());
8059 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
8060 !ST->isVolatile()) ||
8061 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
8062 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
8063 getZExtValue(), MVT::i64);
8064 return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
8065 Ptr, ST->getPointerInfo(), ST->isVolatile(),
8066 ST->isNonTemporal(), ST->getAlignment());
8069 if (!ST->isVolatile() &&
8070 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
8071 // Many FP stores are not made apparent until after legalize, e.g. for
8072 // argument passing. Since this is so common, custom legalize the
8073 // 64-bit integer store into two 32-bit stores.
8074 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
8075 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
8076 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
8077 if (TLI.isBigEndian()) std::swap(Lo, Hi);
8079 unsigned Alignment = ST->getAlignment();
8080 bool isVolatile = ST->isVolatile();
8081 bool isNonTemporal = ST->isNonTemporal();
8083 SDValue St0 = DAG.getStore(Chain, ST->getDebugLoc(), Lo,
8084 Ptr, ST->getPointerInfo(),
8085 isVolatile, isNonTemporal,
8086 ST->getAlignment());
8087 Ptr = DAG.getNode(ISD::ADD, N->getDebugLoc(), Ptr.getValueType(), Ptr,
8088 DAG.getConstant(4, Ptr.getValueType()));
8089 Alignment = MinAlign(Alignment, 4U);
8090 SDValue St1 = DAG.getStore(Chain, ST->getDebugLoc(), Hi,
8091 Ptr, ST->getPointerInfo().getWithOffset(4),
8092 isVolatile, isNonTemporal,
8094 return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
8103 // Try to infer better alignment information than the store already has.
8104 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
8105 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
8106 if (Align > ST->getAlignment())
8107 return DAG.getTruncStore(Chain, N->getDebugLoc(), Value,
8108 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
8109 ST->isVolatile(), ST->isNonTemporal(), Align);
8113 // Try transforming a pair floating point load / store ops to integer
8114 // load / store ops.
8115 SDValue NewST = TransformFPLoadStorePair(N);
8116 if (NewST.getNode())
8120 // Walk up chain skipping non-aliasing memory nodes.
8121 SDValue BetterChain = FindBetterChain(N, Chain);
8123 // If there is a better chain.
8124 if (Chain != BetterChain) {
8127 // Replace the chain to avoid dependency.
8128 if (ST->isTruncatingStore()) {
8129 ReplStore = DAG.getTruncStore(BetterChain, N->getDebugLoc(), Value, Ptr,
8130 ST->getPointerInfo(),
8131 ST->getMemoryVT(), ST->isVolatile(),
8132 ST->isNonTemporal(), ST->getAlignment());
8134 ReplStore = DAG.getStore(BetterChain, N->getDebugLoc(), Value, Ptr,
8135 ST->getPointerInfo(),
8136 ST->isVolatile(), ST->isNonTemporal(),
8137 ST->getAlignment());
8140 // Create token to keep both nodes around.
8141 SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
8142 MVT::Other, Chain, ReplStore);
8144 // Make sure the new and old chains are cleaned up.
8145 AddToWorkList(Token.getNode());
8147 // Don't add users to work list.
8148 return CombineTo(N, Token, false);
8152 // Try transforming N to an indexed store.
8153 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
8154 return SDValue(N, 0);
8156 // FIXME: is there such a thing as a truncating indexed store?
8157 if (ST->isTruncatingStore() && ST->isUnindexed() &&
8158 Value.getValueType().isInteger()) {
8159 // See if we can simplify the input to this truncstore with knowledge that
8160 // only the low bits are being used. For example:
8161 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
8163 GetDemandedBits(Value,
8164 APInt::getLowBitsSet(
8165 Value.getValueType().getScalarType().getSizeInBits(),
8166 ST->getMemoryVT().getScalarType().getSizeInBits()));
8167 AddToWorkList(Value.getNode());
8168 if (Shorter.getNode())
8169 return DAG.getTruncStore(Chain, N->getDebugLoc(), Shorter,
8170 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
8171 ST->isVolatile(), ST->isNonTemporal(),
8172 ST->getAlignment());
8174 // Otherwise, see if we can simplify the operation with
8175 // SimplifyDemandedBits, which only works if the value has a single use.
8176 if (SimplifyDemandedBits(Value,
8177 APInt::getLowBitsSet(
8178 Value.getValueType().getScalarType().getSizeInBits(),
8179 ST->getMemoryVT().getScalarType().getSizeInBits())))
8180 return SDValue(N, 0);
8183 // If this is a load followed by a store to the same location, then the store
8185 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
8186 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
8187 ST->isUnindexed() && !ST->isVolatile() &&
8188 // There can't be any side effects between the load and store, such as
8190 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
8191 // The store is dead, remove it.
8196 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
8197 // truncating store. We can do this even if this is already a truncstore.
8198 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
8199 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
8200 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
8201 ST->getMemoryVT())) {
8202 return DAG.getTruncStore(Chain, N->getDebugLoc(), Value.getOperand(0),
8203 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
8204 ST->isVolatile(), ST->isNonTemporal(),
8205 ST->getAlignment());
8208 // Only perform this optimization before the types are legal, because we
8209 // don't want to perform this optimization on every DAGCombine invocation.
8211 bool EverChanged = false;
8214 // There can be multiple store sequences on the same chain.
8215 // Keep trying to merge store sequences until we are unable to do so
8216 // or until we merge the last store on the chain.
8217 bool Changed = MergeConsecutiveStores(ST);
8218 EverChanged |= Changed;
8219 if (!Changed) break;
8220 } while (ST->getOpcode() != ISD::DELETED_NODE);
8223 return SDValue(N, 0);
8226 return ReduceLoadOpStoreWidth(N);
8229 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
8230 SDValue InVec = N->getOperand(0);
8231 SDValue InVal = N->getOperand(1);
8232 SDValue EltNo = N->getOperand(2);
8233 DebugLoc dl = N->getDebugLoc();
8235 // If the inserted element is an UNDEF, just use the input vector.
8236 if (InVal.getOpcode() == ISD::UNDEF)
8239 EVT VT = InVec.getValueType();
8241 // If we can't generate a legal BUILD_VECTOR, exit
8242 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
8245 // Check that we know which element is being inserted
8246 if (!isa<ConstantSDNode>(EltNo))
8248 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
8250 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
8251 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
8253 SmallVector<SDValue, 8> Ops;
8254 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
8255 Ops.append(InVec.getNode()->op_begin(),
8256 InVec.getNode()->op_end());
8257 } else if (InVec.getOpcode() == ISD::UNDEF) {
8258 unsigned NElts = VT.getVectorNumElements();
8259 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
8264 // Insert the element
8265 if (Elt < Ops.size()) {
8266 // All the operands of BUILD_VECTOR must have the same type;
8267 // we enforce that here.
8268 EVT OpVT = Ops[0].getValueType();
8269 if (InVal.getValueType() != OpVT)
8270 InVal = OpVT.bitsGT(InVal.getValueType()) ?
8271 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
8272 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
8276 // Return the new vector
8277 return DAG.getNode(ISD::BUILD_VECTOR, dl,
8278 VT, &Ops[0], Ops.size());
8281 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
8282 // (vextract (scalar_to_vector val, 0) -> val
8283 SDValue InVec = N->getOperand(0);
8284 EVT VT = InVec.getValueType();
8285 EVT NVT = N->getValueType(0);
8287 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
8288 // Check if the result type doesn't match the inserted element type. A
8289 // SCALAR_TO_VECTOR may truncate the inserted element and the
8290 // EXTRACT_VECTOR_ELT may widen the extracted vector.
8291 SDValue InOp = InVec.getOperand(0);
8292 if (InOp.getValueType() != NVT) {
8293 assert(InOp.getValueType().isInteger() && NVT.isInteger());
8294 return DAG.getSExtOrTrunc(InOp, InVec.getDebugLoc(), NVT);
8299 SDValue EltNo = N->getOperand(1);
8300 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
8302 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
8303 // We only perform this optimization before the op legalization phase because
8304 // we may introduce new vector instructions which are not backed by TD
8305 // patterns. For example on AVX, extracting elements from a wide vector
8306 // without using extract_subvector.
8307 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
8308 && ConstEltNo && !LegalOperations) {
8309 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
8310 int NumElem = VT.getVectorNumElements();
8311 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
8312 // Find the new index to extract from.
8313 int OrigElt = SVOp->getMaskElt(Elt);
8315 // Extracting an undef index is undef.
8317 return DAG.getUNDEF(NVT);
8319 // Select the right vector half to extract from.
8320 if (OrigElt < NumElem) {
8321 InVec = InVec->getOperand(0);
8323 InVec = InVec->getOperand(1);
8327 EVT IndexTy = N->getOperand(1).getValueType();
8328 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, N->getDebugLoc(), NVT,
8329 InVec, DAG.getConstant(OrigElt, IndexTy));
8332 // Perform only after legalization to ensure build_vector / vector_shuffle
8333 // optimizations have already been done.
8334 if (!LegalOperations) return SDValue();
8336 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
8337 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
8338 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
8341 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
8342 bool NewLoad = false;
8343 bool BCNumEltsChanged = false;
8344 EVT ExtVT = VT.getVectorElementType();
8347 // If the result of load has to be truncated, then it's not necessarily
8349 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
8352 if (InVec.getOpcode() == ISD::BITCAST) {
8353 // Don't duplicate a load with other uses.
8354 if (!InVec.hasOneUse())
8357 EVT BCVT = InVec.getOperand(0).getValueType();
8358 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
8360 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
8361 BCNumEltsChanged = true;
8362 InVec = InVec.getOperand(0);
8363 ExtVT = BCVT.getVectorElementType();
8367 LoadSDNode *LN0 = NULL;
8368 const ShuffleVectorSDNode *SVN = NULL;
8369 if (ISD::isNormalLoad(InVec.getNode())) {
8370 LN0 = cast<LoadSDNode>(InVec);
8371 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
8372 InVec.getOperand(0).getValueType() == ExtVT &&
8373 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
8374 // Don't duplicate a load with other uses.
8375 if (!InVec.hasOneUse())
8378 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
8379 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
8380 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
8382 // (load $addr+1*size)
8384 // Don't duplicate a load with other uses.
8385 if (!InVec.hasOneUse())
8388 // If the bit convert changed the number of elements, it is unsafe
8389 // to examine the mask.
8390 if (BCNumEltsChanged)
8393 // Select the input vector, guarding against out of range extract vector.
8394 unsigned NumElems = VT.getVectorNumElements();
8395 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
8396 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
8398 if (InVec.getOpcode() == ISD::BITCAST) {
8399 // Don't duplicate a load with other uses.
8400 if (!InVec.hasOneUse())
8403 InVec = InVec.getOperand(0);
8405 if (ISD::isNormalLoad(InVec.getNode())) {
8406 LN0 = cast<LoadSDNode>(InVec);
8407 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
8411 // Make sure we found a non-volatile load and the extractelement is
8413 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
8416 // If Idx was -1 above, Elt is going to be -1, so just return undef.
8418 return DAG.getUNDEF(LVT);
8420 unsigned Align = LN0->getAlignment();
8422 // Check the resultant load doesn't need a higher alignment than the
8426 ->getABITypeAlignment(LVT.getTypeForEVT(*DAG.getContext()));
8428 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, LVT))
8434 SDValue NewPtr = LN0->getBasePtr();
8435 unsigned PtrOff = 0;
8438 PtrOff = LVT.getSizeInBits() * Elt / 8;
8439 EVT PtrType = NewPtr.getValueType();
8440 if (TLI.isBigEndian())
8441 PtrOff = VT.getSizeInBits() / 8 - PtrOff;
8442 NewPtr = DAG.getNode(ISD::ADD, N->getDebugLoc(), PtrType, NewPtr,
8443 DAG.getConstant(PtrOff, PtrType));
8446 // The replacement we need to do here is a little tricky: we need to
8447 // replace an extractelement of a load with a load.
8448 // Use ReplaceAllUsesOfValuesWith to do the replacement.
8449 // Note that this replacement assumes that the extractvalue is the only
8450 // use of the load; that's okay because we don't want to perform this
8451 // transformation in other cases anyway.
8454 if (NVT.bitsGT(LVT)) {
8455 // If the result type of vextract is wider than the load, then issue an
8456 // extending load instead.
8457 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, LVT)
8458 ? ISD::ZEXTLOAD : ISD::EXTLOAD;
8459 Load = DAG.getExtLoad(ExtType, N->getDebugLoc(), NVT, LN0->getChain(),
8460 NewPtr, LN0->getPointerInfo().getWithOffset(PtrOff),
8461 LVT, LN0->isVolatile(), LN0->isNonTemporal(),Align);
8462 Chain = Load.getValue(1);
8464 Load = DAG.getLoad(LVT, N->getDebugLoc(), LN0->getChain(), NewPtr,
8465 LN0->getPointerInfo().getWithOffset(PtrOff),
8466 LN0->isVolatile(), LN0->isNonTemporal(),
8467 LN0->isInvariant(), Align);
8468 Chain = Load.getValue(1);
8469 if (NVT.bitsLT(LVT))
8470 Load = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), NVT, Load);
8472 Load = DAG.getNode(ISD::BITCAST, N->getDebugLoc(), NVT, Load);
8474 WorkListRemover DeadNodes(*this);
8475 SDValue From[] = { SDValue(N, 0), SDValue(LN0,1) };
8476 SDValue To[] = { Load, Chain };
8477 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
8478 // Since we're explcitly calling ReplaceAllUses, add the new node to the
8479 // worklist explicitly as well.
8480 AddToWorkList(Load.getNode());
8481 AddUsersToWorkList(Load.getNode()); // Add users too
8482 // Make sure to revisit this node to clean it up; it will usually be dead.
8484 return SDValue(N, 0);
8490 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
8491 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
8492 // We perform this optimization post type-legalization because
8493 // the type-legalizer often scalarizes integer-promoted vectors.
8494 // Performing this optimization before may create bit-casts which
8495 // will be type-legalized to complex code sequences.
8496 // We perform this optimization only before the operation legalizer because we
8497 // may introduce illegal operations.
8498 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
8501 unsigned NumInScalars = N->getNumOperands();
8502 DebugLoc dl = N->getDebugLoc();
8503 EVT VT = N->getValueType(0);
8505 // Check to see if this is a BUILD_VECTOR of a bunch of values
8506 // which come from any_extend or zero_extend nodes. If so, we can create
8507 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
8508 // optimizations. We do not handle sign-extend because we can't fill the sign
8510 EVT SourceType = MVT::Other;
8511 bool AllAnyExt = true;
8513 for (unsigned i = 0; i != NumInScalars; ++i) {
8514 SDValue In = N->getOperand(i);
8515 // Ignore undef inputs.
8516 if (In.getOpcode() == ISD::UNDEF) continue;
8518 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
8519 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
8521 // Abort if the element is not an extension.
8522 if (!ZeroExt && !AnyExt) {
8523 SourceType = MVT::Other;
8527 // The input is a ZeroExt or AnyExt. Check the original type.
8528 EVT InTy = In.getOperand(0).getValueType();
8530 // Check that all of the widened source types are the same.
8531 if (SourceType == MVT::Other)
8534 else if (InTy != SourceType) {
8535 // Multiple income types. Abort.
8536 SourceType = MVT::Other;
8540 // Check if all of the extends are ANY_EXTENDs.
8541 AllAnyExt &= AnyExt;
8544 // In order to have valid types, all of the inputs must be extended from the
8545 // same source type and all of the inputs must be any or zero extend.
8546 // Scalar sizes must be a power of two.
8547 EVT OutScalarTy = VT.getScalarType();
8548 bool ValidTypes = SourceType != MVT::Other &&
8549 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
8550 isPowerOf2_32(SourceType.getSizeInBits());
8552 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
8553 // turn into a single shuffle instruction.
8557 bool isLE = TLI.isLittleEndian();
8558 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
8559 assert(ElemRatio > 1 && "Invalid element size ratio");
8560 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
8561 DAG.getConstant(0, SourceType);
8563 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
8564 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
8566 // Populate the new build_vector
8567 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
8568 SDValue Cast = N->getOperand(i);
8569 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
8570 Cast.getOpcode() == ISD::ZERO_EXTEND ||
8571 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
8573 if (Cast.getOpcode() == ISD::UNDEF)
8574 In = DAG.getUNDEF(SourceType);
8576 In = Cast->getOperand(0);
8577 unsigned Index = isLE ? (i * ElemRatio) :
8578 (i * ElemRatio + (ElemRatio - 1));
8580 assert(Index < Ops.size() && "Invalid index");
8584 // The type of the new BUILD_VECTOR node.
8585 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
8586 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
8587 "Invalid vector size");
8588 // Check if the new vector type is legal.
8589 if (!isTypeLegal(VecVT)) return SDValue();
8591 // Make the new BUILD_VECTOR.
8592 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], Ops.size());
8594 // The new BUILD_VECTOR node has the potential to be further optimized.
8595 AddToWorkList(BV.getNode());
8596 // Bitcast to the desired type.
8597 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8600 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
8601 EVT VT = N->getValueType(0);
8603 unsigned NumInScalars = N->getNumOperands();
8604 DebugLoc dl = N->getDebugLoc();
8606 EVT SrcVT = MVT::Other;
8607 unsigned Opcode = ISD::DELETED_NODE;
8608 unsigned NumDefs = 0;
8610 for (unsigned i = 0; i != NumInScalars; ++i) {
8611 SDValue In = N->getOperand(i);
8612 unsigned Opc = In.getOpcode();
8614 if (Opc == ISD::UNDEF)
8617 // If all scalar values are floats and converted from integers.
8618 if (Opcode == ISD::DELETED_NODE &&
8619 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
8621 // If not supported by target, bail out.
8622 if (TLI.getOperationAction(Opcode, VT) != TargetLowering::Legal &&
8623 TLI.getOperationAction(Opcode, VT) != TargetLowering::Custom)
8629 EVT InVT = In.getOperand(0).getValueType();
8631 // If all scalar values are typed differently, bail out. It's chosen to
8632 // simplify BUILD_VECTOR of integer types.
8633 if (SrcVT == MVT::Other)
8640 // If the vector has just one element defined, it's not worth to fold it into
8641 // a vectorized one.
8645 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
8646 && "Should only handle conversion from integer to float.");
8647 assert(SrcVT != MVT::Other && "Cannot determine source type!");
8649 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
8650 SmallVector<SDValue, 8> Opnds;
8651 for (unsigned i = 0; i != NumInScalars; ++i) {
8652 SDValue In = N->getOperand(i);
8654 if (In.getOpcode() == ISD::UNDEF)
8655 Opnds.push_back(DAG.getUNDEF(SrcVT));
8657 Opnds.push_back(In.getOperand(0));
8659 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT,
8660 &Opnds[0], Opnds.size());
8661 AddToWorkList(BV.getNode());
8663 return DAG.getNode(Opcode, dl, VT, BV);
8666 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
8667 unsigned NumInScalars = N->getNumOperands();
8668 DebugLoc dl = N->getDebugLoc();
8669 EVT VT = N->getValueType(0);
8671 // A vector built entirely of undefs is undef.
8672 if (ISD::allOperandsUndef(N))
8673 return DAG.getUNDEF(VT);
8675 SDValue V = reduceBuildVecExtToExtBuildVec(N);
8679 V = reduceBuildVecConvertToConvertBuildVec(N);
8683 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
8684 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
8685 // at most two distinct vectors, turn this into a shuffle node.
8687 // May only combine to shuffle after legalize if shuffle is legal.
8688 if (LegalOperations &&
8689 !TLI.isOperationLegalOrCustom(ISD::VECTOR_SHUFFLE, VT))
8692 SDValue VecIn1, VecIn2;
8693 for (unsigned i = 0; i != NumInScalars; ++i) {
8694 // Ignore undef inputs.
8695 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
8697 // If this input is something other than a EXTRACT_VECTOR_ELT with a
8698 // constant index, bail out.
8699 if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
8700 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
8701 VecIn1 = VecIn2 = SDValue(0, 0);
8705 // We allow up to two distinct input vectors.
8706 SDValue ExtractedFromVec = N->getOperand(i).getOperand(0);
8707 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
8710 if (VecIn1.getNode() == 0) {
8711 VecIn1 = ExtractedFromVec;
8712 } else if (VecIn2.getNode() == 0) {
8713 VecIn2 = ExtractedFromVec;
8716 VecIn1 = VecIn2 = SDValue(0, 0);
8721 // If everything is good, we can make a shuffle operation.
8722 if (VecIn1.getNode()) {
8723 SmallVector<int, 8> Mask;
8724 for (unsigned i = 0; i != NumInScalars; ++i) {
8725 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
8730 // If extracting from the first vector, just use the index directly.
8731 SDValue Extract = N->getOperand(i);
8732 SDValue ExtVal = Extract.getOperand(1);
8733 if (Extract.getOperand(0) == VecIn1) {
8734 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
8735 if (ExtIndex > VT.getVectorNumElements())
8738 Mask.push_back(ExtIndex);
8742 // Otherwise, use InIdx + VecSize
8743 unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue();
8744 Mask.push_back(Idx+NumInScalars);
8747 // We can't generate a shuffle node with mismatched input and output types.
8748 // Attempt to transform a single input vector to the correct type.
8749 if ((VT != VecIn1.getValueType())) {
8750 // We don't support shuffeling between TWO values of different types.
8751 if (VecIn2.getNode() != 0)
8754 // We only support widening of vectors which are half the size of the
8755 // output registers. For example XMM->YMM widening on X86 with AVX.
8756 if (VecIn1.getValueType().getSizeInBits()*2 != VT.getSizeInBits())
8759 // If the input vector type has a different base type to the output
8760 // vector type, bail out.
8761 if (VecIn1.getValueType().getVectorElementType() !=
8762 VT.getVectorElementType())
8765 // Widen the input vector by adding undef values.
8766 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
8767 VecIn1, DAG.getUNDEF(VecIn1.getValueType()));
8770 // If VecIn2 is unused then change it to undef.
8771 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
8773 // Check that we were able to transform all incoming values to the same
8775 if (VecIn2.getValueType() != VecIn1.getValueType() ||
8776 VecIn1.getValueType() != VT)
8779 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
8780 if (!isTypeLegal(VT))
8783 // Return the new VECTOR_SHUFFLE node.
8787 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]);
8793 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
8794 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
8795 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
8796 // inputs come from at most two distinct vectors, turn this into a shuffle
8799 // If we only have one input vector, we don't need to do any concatenation.
8800 if (N->getNumOperands() == 1)
8801 return N->getOperand(0);
8803 // Check if all of the operands are undefs.
8804 if (ISD::allOperandsUndef(N))
8805 return DAG.getUNDEF(N->getValueType(0));
8810 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
8811 EVT NVT = N->getValueType(0);
8812 SDValue V = N->getOperand(0);
8814 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
8815 // Handle only simple case where vector being inserted and vector
8816 // being extracted are of same type, and are half size of larger vectors.
8817 EVT BigVT = V->getOperand(0).getValueType();
8818 EVT SmallVT = V->getOperand(1).getValueType();
8819 if (NVT != SmallVT || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
8822 // Only handle cases where both indexes are constants with the same type.
8823 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
8824 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
8826 if (InsIdx && ExtIdx &&
8827 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
8828 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
8830 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
8832 // indices are equal => V1
8833 // otherwise => (extract_subvec V1, ExtIdx)
8834 if (InsIdx->getZExtValue() == ExtIdx->getZExtValue())
8835 return V->getOperand(1);
8836 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, N->getDebugLoc(), NVT,
8837 V->getOperand(0), N->getOperand(1));
8841 if (V->getOpcode() == ISD::CONCAT_VECTORS) {
8843 // (extract_subvec (concat V1, V2, ...), i)
8846 // Only operand 0 is checked as 'concat' assumes all inputs of the same type.
8847 if (V->getOperand(0).getValueType() != NVT)
8849 unsigned Idx = dyn_cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8850 unsigned NumElems = NVT.getVectorNumElements();
8851 assert((Idx % NumElems) == 0 &&
8852 "IDX in concat is not a multiple of the result vector length.");
8853 return V->getOperand(Idx / NumElems);
8859 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
8860 EVT VT = N->getValueType(0);
8861 unsigned NumElts = VT.getVectorNumElements();
8863 SDValue N0 = N->getOperand(0);
8864 SDValue N1 = N->getOperand(1);
8866 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
8868 // Canonicalize shuffle undef, undef -> undef
8869 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
8870 return DAG.getUNDEF(VT);
8872 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8874 // Canonicalize shuffle v, v -> v, undef
8876 SmallVector<int, 8> NewMask;
8877 for (unsigned i = 0; i != NumElts; ++i) {
8878 int Idx = SVN->getMaskElt(i);
8879 if (Idx >= (int)NumElts) Idx -= NumElts;
8880 NewMask.push_back(Idx);
8882 return DAG.getVectorShuffle(VT, N->getDebugLoc(), N0, DAG.getUNDEF(VT),
8886 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
8887 if (N0.getOpcode() == ISD::UNDEF) {
8888 SmallVector<int, 8> NewMask;
8889 for (unsigned i = 0; i != NumElts; ++i) {
8890 int Idx = SVN->getMaskElt(i);
8892 if (Idx < (int)NumElts)
8897 NewMask.push_back(Idx);
8899 return DAG.getVectorShuffle(VT, N->getDebugLoc(), N1, DAG.getUNDEF(VT),
8903 // Remove references to rhs if it is undef
8904 if (N1.getOpcode() == ISD::UNDEF) {
8905 bool Changed = false;
8906 SmallVector<int, 8> NewMask;
8907 for (unsigned i = 0; i != NumElts; ++i) {
8908 int Idx = SVN->getMaskElt(i);
8909 if (Idx >= (int)NumElts) {
8913 NewMask.push_back(Idx);
8916 return DAG.getVectorShuffle(VT, N->getDebugLoc(), N0, N1, &NewMask[0]);
8919 // If it is a splat, check if the argument vector is another splat or a
8920 // build_vector with all scalar elements the same.
8921 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
8922 SDNode *V = N0.getNode();
8924 // If this is a bit convert that changes the element type of the vector but
8925 // not the number of vector elements, look through it. Be careful not to
8926 // look though conversions that change things like v4f32 to v2f64.
8927 if (V->getOpcode() == ISD::BITCAST) {
8928 SDValue ConvInput = V->getOperand(0);
8929 if (ConvInput.getValueType().isVector() &&
8930 ConvInput.getValueType().getVectorNumElements() == NumElts)
8931 V = ConvInput.getNode();
8934 if (V->getOpcode() == ISD::BUILD_VECTOR) {
8935 assert(V->getNumOperands() == NumElts &&
8936 "BUILD_VECTOR has wrong number of operands");
8938 bool AllSame = true;
8939 for (unsigned i = 0; i != NumElts; ++i) {
8940 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
8941 Base = V->getOperand(i);
8945 // Splat of <u, u, u, u>, return <u, u, u, u>
8946 if (!Base.getNode())
8948 for (unsigned i = 0; i != NumElts; ++i) {
8949 if (V->getOperand(i) != Base) {
8954 // Splat of <x, x, x, x>, return <x, x, x, x>
8960 // If this shuffle node is simply a swizzle of another shuffle node,
8961 // and it reverses the swizzle of the previous shuffle then we can
8962 // optimize shuffle(shuffle(x, undef), undef) -> x.
8963 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
8964 N1.getOpcode() == ISD::UNDEF) {
8966 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
8968 // Shuffle nodes can only reverse shuffles with a single non-undef value.
8969 if (N0.getOperand(1).getOpcode() != ISD::UNDEF)
8972 // The incoming shuffle must be of the same type as the result of the
8974 assert(OtherSV->getOperand(0).getValueType() == VT &&
8975 "Shuffle types don't match");
8977 for (unsigned i = 0; i != NumElts; ++i) {
8978 int Idx = SVN->getMaskElt(i);
8979 assert(Idx < (int)NumElts && "Index references undef operand");
8980 // Next, this index comes from the first value, which is the incoming
8981 // shuffle. Adopt the incoming index.
8983 Idx = OtherSV->getMaskElt(Idx);
8985 // The combined shuffle must map each index to itself.
8986 if (Idx >= 0 && (unsigned)Idx != i)
8990 return OtherSV->getOperand(0);
8996 SDValue DAGCombiner::visitMEMBARRIER(SDNode* N) {
8997 if (!TLI.getShouldFoldAtomicFences())
9000 SDValue atomic = N->getOperand(0);
9001 switch (atomic.getOpcode()) {
9002 case ISD::ATOMIC_CMP_SWAP:
9003 case ISD::ATOMIC_SWAP:
9004 case ISD::ATOMIC_LOAD_ADD:
9005 case ISD::ATOMIC_LOAD_SUB:
9006 case ISD::ATOMIC_LOAD_AND:
9007 case ISD::ATOMIC_LOAD_OR:
9008 case ISD::ATOMIC_LOAD_XOR:
9009 case ISD::ATOMIC_LOAD_NAND:
9010 case ISD::ATOMIC_LOAD_MIN:
9011 case ISD::ATOMIC_LOAD_MAX:
9012 case ISD::ATOMIC_LOAD_UMIN:
9013 case ISD::ATOMIC_LOAD_UMAX:
9019 SDValue fence = atomic.getOperand(0);
9020 if (fence.getOpcode() != ISD::MEMBARRIER)
9023 switch (atomic.getOpcode()) {
9024 case ISD::ATOMIC_CMP_SWAP:
9025 return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
9026 fence.getOperand(0),
9027 atomic.getOperand(1), atomic.getOperand(2),
9028 atomic.getOperand(3)), atomic.getResNo());
9029 case ISD::ATOMIC_SWAP:
9030 case ISD::ATOMIC_LOAD_ADD:
9031 case ISD::ATOMIC_LOAD_SUB:
9032 case ISD::ATOMIC_LOAD_AND:
9033 case ISD::ATOMIC_LOAD_OR:
9034 case ISD::ATOMIC_LOAD_XOR:
9035 case ISD::ATOMIC_LOAD_NAND:
9036 case ISD::ATOMIC_LOAD_MIN:
9037 case ISD::ATOMIC_LOAD_MAX:
9038 case ISD::ATOMIC_LOAD_UMIN:
9039 case ISD::ATOMIC_LOAD_UMAX:
9040 return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
9041 fence.getOperand(0),
9042 atomic.getOperand(1), atomic.getOperand(2)),
9049 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
9050 /// an AND to a vector_shuffle with the destination vector and a zero vector.
9051 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
9052 /// vector_shuffle V, Zero, <0, 4, 2, 4>
9053 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
9054 EVT VT = N->getValueType(0);
9055 DebugLoc dl = N->getDebugLoc();
9056 SDValue LHS = N->getOperand(0);
9057 SDValue RHS = N->getOperand(1);
9058 if (N->getOpcode() == ISD::AND) {
9059 if (RHS.getOpcode() == ISD::BITCAST)
9060 RHS = RHS.getOperand(0);
9061 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
9062 SmallVector<int, 8> Indices;
9063 unsigned NumElts = RHS.getNumOperands();
9064 for (unsigned i = 0; i != NumElts; ++i) {
9065 SDValue Elt = RHS.getOperand(i);
9066 if (!isa<ConstantSDNode>(Elt))
9069 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
9070 Indices.push_back(i);
9071 else if (cast<ConstantSDNode>(Elt)->isNullValue())
9072 Indices.push_back(NumElts);
9077 // Let's see if the target supports this vector_shuffle.
9078 EVT RVT = RHS.getValueType();
9079 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
9082 // Return the new VECTOR_SHUFFLE node.
9083 EVT EltVT = RVT.getVectorElementType();
9084 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
9085 DAG.getConstant(0, EltVT));
9086 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
9087 RVT, &ZeroOps[0], ZeroOps.size());
9088 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
9089 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
9090 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
9097 /// SimplifyVBinOp - Visit a binary vector operation, like ADD.
9098 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
9099 // After legalize, the target may be depending on adds and other
9100 // binary ops to provide legal ways to construct constants or other
9101 // things. Simplifying them may result in a loss of legality.
9102 if (LegalOperations) return SDValue();
9104 assert(N->getValueType(0).isVector() &&
9105 "SimplifyVBinOp only works on vectors!");
9107 SDValue LHS = N->getOperand(0);
9108 SDValue RHS = N->getOperand(1);
9109 SDValue Shuffle = XformToShuffleWithZero(N);
9110 if (Shuffle.getNode()) return Shuffle;
9112 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
9114 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
9115 RHS.getOpcode() == ISD::BUILD_VECTOR) {
9116 SmallVector<SDValue, 8> Ops;
9117 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
9118 SDValue LHSOp = LHS.getOperand(i);
9119 SDValue RHSOp = RHS.getOperand(i);
9120 // If these two elements can't be folded, bail out.
9121 if ((LHSOp.getOpcode() != ISD::UNDEF &&
9122 LHSOp.getOpcode() != ISD::Constant &&
9123 LHSOp.getOpcode() != ISD::ConstantFP) ||
9124 (RHSOp.getOpcode() != ISD::UNDEF &&
9125 RHSOp.getOpcode() != ISD::Constant &&
9126 RHSOp.getOpcode() != ISD::ConstantFP))
9129 // Can't fold divide by zero.
9130 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
9131 N->getOpcode() == ISD::FDIV) {
9132 if ((RHSOp.getOpcode() == ISD::Constant &&
9133 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
9134 (RHSOp.getOpcode() == ISD::ConstantFP &&
9135 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
9139 EVT VT = LHSOp.getValueType();
9140 EVT RVT = RHSOp.getValueType();
9142 // Integer BUILD_VECTOR operands may have types larger than the element
9143 // size (e.g., when the element type is not legal). Prior to type
9144 // legalization, the types may not match between the two BUILD_VECTORS.
9145 // Truncate one of the operands to make them match.
9146 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
9147 RHSOp = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, RHSOp);
9149 LHSOp = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), RVT, LHSOp);
9153 SDValue FoldOp = DAG.getNode(N->getOpcode(), LHS.getDebugLoc(), VT,
9155 if (FoldOp.getOpcode() != ISD::UNDEF &&
9156 FoldOp.getOpcode() != ISD::Constant &&
9157 FoldOp.getOpcode() != ISD::ConstantFP)
9159 Ops.push_back(FoldOp);
9160 AddToWorkList(FoldOp.getNode());
9163 if (Ops.size() == LHS.getNumOperands())
9164 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
9165 LHS.getValueType(), &Ops[0], Ops.size());
9171 /// SimplifyVUnaryOp - Visit a binary vector operation, like FABS/FNEG.
9172 SDValue DAGCombiner::SimplifyVUnaryOp(SDNode *N) {
9173 // After legalize, the target may be depending on adds and other
9174 // binary ops to provide legal ways to construct constants or other
9175 // things. Simplifying them may result in a loss of legality.
9176 if (LegalOperations) return SDValue();
9178 assert(N->getValueType(0).isVector() &&
9179 "SimplifyVUnaryOp only works on vectors!");
9181 SDValue N0 = N->getOperand(0);
9183 if (N0.getOpcode() != ISD::BUILD_VECTOR)
9186 // Operand is a BUILD_VECTOR node, see if we can constant fold it.
9187 SmallVector<SDValue, 8> Ops;
9188 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
9189 SDValue Op = N0.getOperand(i);
9190 if (Op.getOpcode() != ISD::UNDEF &&
9191 Op.getOpcode() != ISD::ConstantFP)
9193 EVT EltVT = Op.getValueType();
9194 SDValue FoldOp = DAG.getNode(N->getOpcode(), N0.getDebugLoc(), EltVT, Op);
9195 if (FoldOp.getOpcode() != ISD::UNDEF &&
9196 FoldOp.getOpcode() != ISD::ConstantFP)
9198 Ops.push_back(FoldOp);
9199 AddToWorkList(FoldOp.getNode());
9202 if (Ops.size() != N0.getNumOperands())
9205 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
9206 N0.getValueType(), &Ops[0], Ops.size());
9209 SDValue DAGCombiner::SimplifySelect(DebugLoc DL, SDValue N0,
9210 SDValue N1, SDValue N2){
9211 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
9213 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
9214 cast<CondCodeSDNode>(N0.getOperand(2))->get());
9216 // If we got a simplified select_cc node back from SimplifySelectCC, then
9217 // break it down into a new SETCC node, and a new SELECT node, and then return
9218 // the SELECT node, since we were called with a SELECT node.
9219 if (SCC.getNode()) {
9220 // Check to see if we got a select_cc back (to turn into setcc/select).
9221 // Otherwise, just return whatever node we got back, like fabs.
9222 if (SCC.getOpcode() == ISD::SELECT_CC) {
9223 SDValue SETCC = DAG.getNode(ISD::SETCC, N0.getDebugLoc(),
9225 SCC.getOperand(0), SCC.getOperand(1),
9227 AddToWorkList(SETCC.getNode());
9228 return DAG.getNode(ISD::SELECT, SCC.getDebugLoc(), SCC.getValueType(),
9229 SCC.getOperand(2), SCC.getOperand(3), SETCC);
9237 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
9238 /// are the two values being selected between, see if we can simplify the
9239 /// select. Callers of this should assume that TheSelect is deleted if this
9240 /// returns true. As such, they should return the appropriate thing (e.g. the
9241 /// node) back to the top-level of the DAG combiner loop to avoid it being
9243 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
9246 // Cannot simplify select with vector condition
9247 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
9249 // If this is a select from two identical things, try to pull the operation
9250 // through the select.
9251 if (LHS.getOpcode() != RHS.getOpcode() ||
9252 !LHS.hasOneUse() || !RHS.hasOneUse())
9255 // If this is a load and the token chain is identical, replace the select
9256 // of two loads with a load through a select of the address to load from.
9257 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
9258 // constants have been dropped into the constant pool.
9259 if (LHS.getOpcode() == ISD::LOAD) {
9260 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
9261 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
9263 // Token chains must be identical.
9264 if (LHS.getOperand(0) != RHS.getOperand(0) ||
9265 // Do not let this transformation reduce the number of volatile loads.
9266 LLD->isVolatile() || RLD->isVolatile() ||
9267 // If this is an EXTLOAD, the VT's must match.
9268 LLD->getMemoryVT() != RLD->getMemoryVT() ||
9269 // If this is an EXTLOAD, the kind of extension must match.
9270 (LLD->getExtensionType() != RLD->getExtensionType() &&
9271 // The only exception is if one of the extensions is anyext.
9272 LLD->getExtensionType() != ISD::EXTLOAD &&
9273 RLD->getExtensionType() != ISD::EXTLOAD) ||
9274 // FIXME: this discards src value information. This is
9275 // over-conservative. It would be beneficial to be able to remember
9276 // both potential memory locations. Since we are discarding
9277 // src value info, don't do the transformation if the memory
9278 // locations are not in the default address space.
9279 LLD->getPointerInfo().getAddrSpace() != 0 ||
9280 RLD->getPointerInfo().getAddrSpace() != 0)
9283 // Check that the select condition doesn't reach either load. If so,
9284 // folding this will induce a cycle into the DAG. If not, this is safe to
9285 // xform, so create a select of the addresses.
9287 if (TheSelect->getOpcode() == ISD::SELECT) {
9288 SDNode *CondNode = TheSelect->getOperand(0).getNode();
9289 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
9290 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
9292 // The loads must not depend on one another.
9293 if (LLD->isPredecessorOf(RLD) ||
9294 RLD->isPredecessorOf(LLD))
9296 Addr = DAG.getNode(ISD::SELECT, TheSelect->getDebugLoc(),
9297 LLD->getBasePtr().getValueType(),
9298 TheSelect->getOperand(0), LLD->getBasePtr(),
9300 } else { // Otherwise SELECT_CC
9301 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
9302 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
9304 if ((LLD->hasAnyUseOfValue(1) &&
9305 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
9306 (RLD->hasAnyUseOfValue(1) &&
9307 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
9310 Addr = DAG.getNode(ISD::SELECT_CC, TheSelect->getDebugLoc(),
9311 LLD->getBasePtr().getValueType(),
9312 TheSelect->getOperand(0),
9313 TheSelect->getOperand(1),
9314 LLD->getBasePtr(), RLD->getBasePtr(),
9315 TheSelect->getOperand(4));
9319 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
9320 Load = DAG.getLoad(TheSelect->getValueType(0),
9321 TheSelect->getDebugLoc(),
9322 // FIXME: Discards pointer info.
9323 LLD->getChain(), Addr, MachinePointerInfo(),
9324 LLD->isVolatile(), LLD->isNonTemporal(),
9325 LLD->isInvariant(), LLD->getAlignment());
9327 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
9328 RLD->getExtensionType() : LLD->getExtensionType(),
9329 TheSelect->getDebugLoc(),
9330 TheSelect->getValueType(0),
9331 // FIXME: Discards pointer info.
9332 LLD->getChain(), Addr, MachinePointerInfo(),
9333 LLD->getMemoryVT(), LLD->isVolatile(),
9334 LLD->isNonTemporal(), LLD->getAlignment());
9337 // Users of the select now use the result of the load.
9338 CombineTo(TheSelect, Load);
9340 // Users of the old loads now use the new load's chain. We know the
9341 // old-load value is dead now.
9342 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
9343 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
9350 /// SimplifySelectCC - Simplify an expression of the form (N0 cond N1) ? N2 : N3
9351 /// where 'cond' is the comparison specified by CC.
9352 SDValue DAGCombiner::SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1,
9353 SDValue N2, SDValue N3,
9354 ISD::CondCode CC, bool NotExtCompare) {
9355 // (x ? y : y) -> y.
9356 if (N2 == N3) return N2;
9358 EVT VT = N2.getValueType();
9359 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
9360 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
9361 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
9363 // Determine if the condition we're dealing with is constant
9364 SDValue SCC = SimplifySetCC(TLI.getSetCCResultType(N0.getValueType()),
9365 N0, N1, CC, DL, false);
9366 if (SCC.getNode()) AddToWorkList(SCC.getNode());
9367 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
9369 // fold select_cc true, x, y -> x
9370 if (SCCC && !SCCC->isNullValue())
9372 // fold select_cc false, x, y -> y
9373 if (SCCC && SCCC->isNullValue())
9376 // Check to see if we can simplify the select into an fabs node
9377 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
9378 // Allow either -0.0 or 0.0
9379 if (CFP->getValueAPF().isZero()) {
9380 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
9381 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
9382 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
9383 N2 == N3.getOperand(0))
9384 return DAG.getNode(ISD::FABS, DL, VT, N0);
9386 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
9387 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
9388 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
9389 N2.getOperand(0) == N3)
9390 return DAG.getNode(ISD::FABS, DL, VT, N3);
9394 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
9395 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
9396 // in it. This is a win when the constant is not otherwise available because
9397 // it replaces two constant pool loads with one. We only do this if the FP
9398 // type is known to be legal, because if it isn't, then we are before legalize
9399 // types an we want the other legalization to happen first (e.g. to avoid
9400 // messing with soft float) and if the ConstantFP is not legal, because if
9401 // it is legal, we may not need to store the FP constant in a constant pool.
9402 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
9403 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
9404 if (TLI.isTypeLegal(N2.getValueType()) &&
9405 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
9406 TargetLowering::Legal) &&
9407 // If both constants have multiple uses, then we won't need to do an
9408 // extra load, they are likely around in registers for other users.
9409 (TV->hasOneUse() || FV->hasOneUse())) {
9410 Constant *Elts[] = {
9411 const_cast<ConstantFP*>(FV->getConstantFPValue()),
9412 const_cast<ConstantFP*>(TV->getConstantFPValue())
9414 Type *FPTy = Elts[0]->getType();
9415 const DataLayout &TD = *TLI.getDataLayout();
9417 // Create a ConstantArray of the two constants.
9418 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
9419 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
9420 TD.getPrefTypeAlignment(FPTy));
9421 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
9423 // Get the offsets to the 0 and 1 element of the array so that we can
9424 // select between them.
9425 SDValue Zero = DAG.getIntPtrConstant(0);
9426 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
9427 SDValue One = DAG.getIntPtrConstant(EltSize);
9429 SDValue Cond = DAG.getSetCC(DL,
9430 TLI.getSetCCResultType(N0.getValueType()),
9432 AddToWorkList(Cond.getNode());
9433 SDValue CstOffset = DAG.getNode(ISD::SELECT, DL, Zero.getValueType(),
9435 AddToWorkList(CstOffset.getNode());
9436 CPIdx = DAG.getNode(ISD::ADD, DL, TLI.getPointerTy(), CPIdx,
9438 AddToWorkList(CPIdx.getNode());
9439 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
9440 MachinePointerInfo::getConstantPool(), false,
9441 false, false, Alignment);
9446 // Check to see if we can perform the "gzip trick", transforming
9447 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
9448 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
9449 (N1C->isNullValue() || // (a < 0) ? b : 0
9450 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
9451 EVT XType = N0.getValueType();
9452 EVT AType = N2.getValueType();
9453 if (XType.bitsGE(AType)) {
9454 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
9455 // single-bit constant.
9456 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
9457 unsigned ShCtV = N2C->getAPIntValue().logBase2();
9458 ShCtV = XType.getSizeInBits()-ShCtV-1;
9459 SDValue ShCt = DAG.getConstant(ShCtV,
9460 getShiftAmountTy(N0.getValueType()));
9461 SDValue Shift = DAG.getNode(ISD::SRL, N0.getDebugLoc(),
9463 AddToWorkList(Shift.getNode());
9465 if (XType.bitsGT(AType)) {
9466 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
9467 AddToWorkList(Shift.getNode());
9470 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
9473 SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(),
9475 DAG.getConstant(XType.getSizeInBits()-1,
9476 getShiftAmountTy(N0.getValueType())));
9477 AddToWorkList(Shift.getNode());
9479 if (XType.bitsGT(AType)) {
9480 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
9481 AddToWorkList(Shift.getNode());
9484 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
9488 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
9489 // where y is has a single bit set.
9490 // A plaintext description would be, we can turn the SELECT_CC into an AND
9491 // when the condition can be materialized as an all-ones register. Any
9492 // single bit-test can be materialized as an all-ones register with
9493 // shift-left and shift-right-arith.
9494 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
9495 N0->getValueType(0) == VT &&
9496 N1C && N1C->isNullValue() &&
9497 N2C && N2C->isNullValue()) {
9498 SDValue AndLHS = N0->getOperand(0);
9499 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
9500 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
9501 // Shift the tested bit over the sign bit.
9502 APInt AndMask = ConstAndRHS->getAPIntValue();
9504 DAG.getConstant(AndMask.countLeadingZeros(),
9505 getShiftAmountTy(AndLHS.getValueType()));
9506 SDValue Shl = DAG.getNode(ISD::SHL, N0.getDebugLoc(), VT, AndLHS, ShlAmt);
9508 // Now arithmetic right shift it all the way over, so the result is either
9509 // all-ones, or zero.
9511 DAG.getConstant(AndMask.getBitWidth()-1,
9512 getShiftAmountTy(Shl.getValueType()));
9513 SDValue Shr = DAG.getNode(ISD::SRA, N0.getDebugLoc(), VT, Shl, ShrAmt);
9515 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
9519 // fold select C, 16, 0 -> shl C, 4
9520 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
9521 TLI.getBooleanContents(N0.getValueType().isVector()) ==
9522 TargetLowering::ZeroOrOneBooleanContent) {
9524 // If the caller doesn't want us to simplify this into a zext of a compare,
9526 if (NotExtCompare && N2C->getAPIntValue() == 1)
9529 // Get a SetCC of the condition
9530 // NOTE: Don't create a SETCC if it's not legal on this target.
9531 if (!LegalOperations ||
9532 TLI.isOperationLegal(ISD::SETCC,
9533 LegalTypes ? TLI.getSetCCResultType(N0.getValueType()) : MVT::i1)) {
9535 // cast from setcc result type to select result type
9537 SCC = DAG.getSetCC(DL, TLI.getSetCCResultType(N0.getValueType()),
9539 if (N2.getValueType().bitsLT(SCC.getValueType()))
9540 Temp = DAG.getZeroExtendInReg(SCC, N2.getDebugLoc(),
9543 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getDebugLoc(),
9544 N2.getValueType(), SCC);
9546 SCC = DAG.getSetCC(N0.getDebugLoc(), MVT::i1, N0, N1, CC);
9547 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getDebugLoc(),
9548 N2.getValueType(), SCC);
9551 AddToWorkList(SCC.getNode());
9552 AddToWorkList(Temp.getNode());
9554 if (N2C->getAPIntValue() == 1)
9557 // shl setcc result by log2 n2c
9558 return DAG.getNode(ISD::SHL, DL, N2.getValueType(), Temp,
9559 DAG.getConstant(N2C->getAPIntValue().logBase2(),
9560 getShiftAmountTy(Temp.getValueType())));
9564 // Check to see if this is the equivalent of setcc
9565 // FIXME: Turn all of these into setcc if setcc if setcc is legal
9566 // otherwise, go ahead with the folds.
9567 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
9568 EVT XType = N0.getValueType();
9569 if (!LegalOperations ||
9570 TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(XType))) {
9571 SDValue Res = DAG.getSetCC(DL, TLI.getSetCCResultType(XType), N0, N1, CC);
9572 if (Res.getValueType() != VT)
9573 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
9577 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
9578 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
9579 (!LegalOperations ||
9580 TLI.isOperationLegal(ISD::CTLZ, XType))) {
9581 SDValue Ctlz = DAG.getNode(ISD::CTLZ, N0.getDebugLoc(), XType, N0);
9582 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
9583 DAG.getConstant(Log2_32(XType.getSizeInBits()),
9584 getShiftAmountTy(Ctlz.getValueType())));
9586 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
9587 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
9588 SDValue NegN0 = DAG.getNode(ISD::SUB, N0.getDebugLoc(),
9589 XType, DAG.getConstant(0, XType), N0);
9590 SDValue NotN0 = DAG.getNOT(N0.getDebugLoc(), N0, XType);
9591 return DAG.getNode(ISD::SRL, DL, XType,
9592 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
9593 DAG.getConstant(XType.getSizeInBits()-1,
9594 getShiftAmountTy(XType)));
9596 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
9597 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
9598 SDValue Sign = DAG.getNode(ISD::SRL, N0.getDebugLoc(), XType, N0,
9599 DAG.getConstant(XType.getSizeInBits()-1,
9600 getShiftAmountTy(N0.getValueType())));
9601 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
9605 // Check to see if this is an integer abs.
9606 // select_cc setg[te] X, 0, X, -X ->
9607 // select_cc setgt X, -1, X, -X ->
9608 // select_cc setl[te] X, 0, -X, X ->
9609 // select_cc setlt X, 1, -X, X ->
9610 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
9612 ConstantSDNode *SubC = NULL;
9613 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
9614 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
9615 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
9616 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
9617 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
9618 (N1C->isOne() && CC == ISD::SETLT)) &&
9619 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
9620 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
9622 EVT XType = N0.getValueType();
9623 if (SubC && SubC->isNullValue() && XType.isInteger()) {
9624 SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(), XType,
9626 DAG.getConstant(XType.getSizeInBits()-1,
9627 getShiftAmountTy(N0.getValueType())));
9628 SDValue Add = DAG.getNode(ISD::ADD, N0.getDebugLoc(),
9630 AddToWorkList(Shift.getNode());
9631 AddToWorkList(Add.getNode());
9632 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
9639 /// SimplifySetCC - This is a stub for TargetLowering::SimplifySetCC.
9640 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
9641 SDValue N1, ISD::CondCode Cond,
9642 DebugLoc DL, bool foldBooleans) {
9643 TargetLowering::DAGCombinerInfo
9644 DagCombineInfo(DAG, !LegalTypes, !LegalOperations, false, this);
9645 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
9648 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
9649 /// return a DAG expression to select that will generate the same value by
9650 /// multiplying by a magic number. See:
9651 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
9652 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
9653 std::vector<SDNode*> Built;
9654 SDValue S = TLI.BuildSDIV(N, DAG, LegalOperations, &Built);
9656 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
9662 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
9663 /// return a DAG expression to select that will generate the same value by
9664 /// multiplying by a magic number. See:
9665 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
9666 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
9667 std::vector<SDNode*> Built;
9668 SDValue S = TLI.BuildUDIV(N, DAG, LegalOperations, &Built);
9670 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
9676 /// FindBaseOffset - Return true if base is a frame index, which is known not
9677 // to alias with anything but itself. Provides base object and offset as
9679 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
9680 const GlobalValue *&GV, const void *&CV) {
9681 // Assume it is a primitive operation.
9682 Base = Ptr; Offset = 0; GV = 0; CV = 0;
9684 // If it's an adding a simple constant then integrate the offset.
9685 if (Base.getOpcode() == ISD::ADD) {
9686 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
9687 Base = Base.getOperand(0);
9688 Offset += C->getZExtValue();
9692 // Return the underlying GlobalValue, and update the Offset. Return false
9693 // for GlobalAddressSDNode since the same GlobalAddress may be represented
9694 // by multiple nodes with different offsets.
9695 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
9696 GV = G->getGlobal();
9697 Offset += G->getOffset();
9701 // Return the underlying Constant value, and update the Offset. Return false
9702 // for ConstantSDNodes since the same constant pool entry may be represented
9703 // by multiple nodes with different offsets.
9704 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
9705 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
9706 : (const void *)C->getConstVal();
9707 Offset += C->getOffset();
9710 // If it's any of the following then it can't alias with anything but itself.
9711 return isa<FrameIndexSDNode>(Base);
9714 /// isAlias - Return true if there is any possibility that the two addresses
9716 bool DAGCombiner::isAlias(SDValue Ptr1, int64_t Size1,
9717 const Value *SrcValue1, int SrcValueOffset1,
9718 unsigned SrcValueAlign1,
9719 const MDNode *TBAAInfo1,
9720 SDValue Ptr2, int64_t Size2,
9721 const Value *SrcValue2, int SrcValueOffset2,
9722 unsigned SrcValueAlign2,
9723 const MDNode *TBAAInfo2) const {
9724 // If they are the same then they must be aliases.
9725 if (Ptr1 == Ptr2) return true;
9727 // Gather base node and offset information.
9728 SDValue Base1, Base2;
9729 int64_t Offset1, Offset2;
9730 const GlobalValue *GV1, *GV2;
9731 const void *CV1, *CV2;
9732 bool isFrameIndex1 = FindBaseOffset(Ptr1, Base1, Offset1, GV1, CV1);
9733 bool isFrameIndex2 = FindBaseOffset(Ptr2, Base2, Offset2, GV2, CV2);
9735 // If they have a same base address then check to see if they overlap.
9736 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
9737 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
9739 // It is possible for different frame indices to alias each other, mostly
9740 // when tail call optimization reuses return address slots for arguments.
9741 // To catch this case, look up the actual index of frame indices to compute
9742 // the real alias relationship.
9743 if (isFrameIndex1 && isFrameIndex2) {
9744 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
9745 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
9746 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
9747 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
9750 // Otherwise, if we know what the bases are, and they aren't identical, then
9751 // we know they cannot alias.
9752 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
9755 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
9756 // compared to the size and offset of the access, we may be able to prove they
9757 // do not alias. This check is conservative for now to catch cases created by
9758 // splitting vector types.
9759 if ((SrcValueAlign1 == SrcValueAlign2) &&
9760 (SrcValueOffset1 != SrcValueOffset2) &&
9761 (Size1 == Size2) && (SrcValueAlign1 > Size1)) {
9762 int64_t OffAlign1 = SrcValueOffset1 % SrcValueAlign1;
9763 int64_t OffAlign2 = SrcValueOffset2 % SrcValueAlign1;
9765 // There is no overlap between these relatively aligned accesses of similar
9766 // size, return no alias.
9767 if ((OffAlign1 + Size1) <= OffAlign2 || (OffAlign2 + Size2) <= OffAlign1)
9771 if (CombinerGlobalAA) {
9772 // Use alias analysis information.
9773 int64_t MinOffset = std::min(SrcValueOffset1, SrcValueOffset2);
9774 int64_t Overlap1 = Size1 + SrcValueOffset1 - MinOffset;
9775 int64_t Overlap2 = Size2 + SrcValueOffset2 - MinOffset;
9776 AliasAnalysis::AliasResult AAResult =
9777 AA.alias(AliasAnalysis::Location(SrcValue1, Overlap1, TBAAInfo1),
9778 AliasAnalysis::Location(SrcValue2, Overlap2, TBAAInfo2));
9779 if (AAResult == AliasAnalysis::NoAlias)
9783 // Otherwise we have to assume they alias.
9787 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) {
9789 int64_t Size0, Size1;
9790 const Value *SrcValue0, *SrcValue1;
9791 int SrcValueOffset0, SrcValueOffset1;
9792 unsigned SrcValueAlign0, SrcValueAlign1;
9793 const MDNode *SrcTBAAInfo0, *SrcTBAAInfo1;
9794 FindAliasInfo(Op0, Ptr0, Size0, SrcValue0, SrcValueOffset0,
9795 SrcValueAlign0, SrcTBAAInfo0);
9796 FindAliasInfo(Op1, Ptr1, Size1, SrcValue1, SrcValueOffset1,
9797 SrcValueAlign1, SrcTBAAInfo1);
9798 return isAlias(Ptr0, Size0, SrcValue0, SrcValueOffset0,
9799 SrcValueAlign0, SrcTBAAInfo0,
9800 Ptr1, Size1, SrcValue1, SrcValueOffset1,
9801 SrcValueAlign1, SrcTBAAInfo1);
9804 /// FindAliasInfo - Extracts the relevant alias information from the memory
9805 /// node. Returns true if the operand was a load.
9806 bool DAGCombiner::FindAliasInfo(SDNode *N,
9807 SDValue &Ptr, int64_t &Size,
9808 const Value *&SrcValue,
9809 int &SrcValueOffset,
9810 unsigned &SrcValueAlign,
9811 const MDNode *&TBAAInfo) const {
9812 LSBaseSDNode *LS = cast<LSBaseSDNode>(N);
9814 Ptr = LS->getBasePtr();
9815 Size = LS->getMemoryVT().getSizeInBits() >> 3;
9816 SrcValue = LS->getSrcValue();
9817 SrcValueOffset = LS->getSrcValueOffset();
9818 SrcValueAlign = LS->getOriginalAlignment();
9819 TBAAInfo = LS->getTBAAInfo();
9820 return isa<LoadSDNode>(LS);
9823 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
9824 /// looking for aliasing nodes and adding them to the Aliases vector.
9825 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
9826 SmallVector<SDValue, 8> &Aliases) {
9827 SmallVector<SDValue, 8> Chains; // List of chains to visit.
9828 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
9830 // Get alias information for node.
9833 const Value *SrcValue;
9835 unsigned SrcValueAlign;
9836 const MDNode *SrcTBAAInfo;
9837 bool IsLoad = FindAliasInfo(N, Ptr, Size, SrcValue, SrcValueOffset,
9838 SrcValueAlign, SrcTBAAInfo);
9841 Chains.push_back(OriginalChain);
9844 // Look at each chain and determine if it is an alias. If so, add it to the
9845 // aliases list. If not, then continue up the chain looking for the next
9847 while (!Chains.empty()) {
9848 SDValue Chain = Chains.back();
9851 // For TokenFactor nodes, look at each operand and only continue up the
9852 // chain until we find two aliases. If we've seen two aliases, assume we'll
9853 // find more and revert to original chain since the xform is unlikely to be
9856 // FIXME: The depth check could be made to return the last non-aliasing
9857 // chain we found before we hit a tokenfactor rather than the original
9859 if (Depth > 6 || Aliases.size() == 2) {
9861 Aliases.push_back(OriginalChain);
9865 // Don't bother if we've been before.
9866 if (!Visited.insert(Chain.getNode()))
9869 switch (Chain.getOpcode()) {
9870 case ISD::EntryToken:
9871 // Entry token is ideal chain operand, but handled in FindBetterChain.
9876 // Get alias information for Chain.
9879 const Value *OpSrcValue;
9880 int OpSrcValueOffset;
9881 unsigned OpSrcValueAlign;
9882 const MDNode *OpSrcTBAAInfo;
9883 bool IsOpLoad = FindAliasInfo(Chain.getNode(), OpPtr, OpSize,
9884 OpSrcValue, OpSrcValueOffset,
9888 // If chain is alias then stop here.
9889 if (!(IsLoad && IsOpLoad) &&
9890 isAlias(Ptr, Size, SrcValue, SrcValueOffset, SrcValueAlign,
9892 OpPtr, OpSize, OpSrcValue, OpSrcValueOffset,
9893 OpSrcValueAlign, OpSrcTBAAInfo)) {
9894 Aliases.push_back(Chain);
9896 // Look further up the chain.
9897 Chains.push_back(Chain.getOperand(0));
9903 case ISD::TokenFactor:
9904 // We have to check each of the operands of the token factor for "small"
9905 // token factors, so we queue them up. Adding the operands to the queue
9906 // (stack) in reverse order maintains the original order and increases the
9907 // likelihood that getNode will find a matching token factor (CSE.)
9908 if (Chain.getNumOperands() > 16) {
9909 Aliases.push_back(Chain);
9912 for (unsigned n = Chain.getNumOperands(); n;)
9913 Chains.push_back(Chain.getOperand(--n));
9918 // For all other instructions we will just have to take what we can get.
9919 Aliases.push_back(Chain);
9925 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
9926 /// for a better chain (aliasing node.)
9927 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
9928 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
9930 // Accumulate all the aliases to this node.
9931 GatherAllAliases(N, OldChain, Aliases);
9933 // If no operands then chain to entry token.
9934 if (Aliases.size() == 0)
9935 return DAG.getEntryNode();
9937 // If a single operand then chain to it. We don't need to revisit it.
9938 if (Aliases.size() == 1)
9941 // Construct a custom tailored token factor.
9942 return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
9943 &Aliases[0], Aliases.size());
9946 // SelectionDAG::Combine - This is the entry point for the file.
9948 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
9949 CodeGenOpt::Level OptLevel) {
9950 /// run - This is the main entry point to this class.
9952 DAGCombiner(*this, AA, OptLevel).Run(Level);