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 N00.getNumOperands() == 2 && N01.getNumOperands() == 2) {
2990 // (or (or (and), (and)), (or (and), (and)))
2991 SDValue N000 = N00.getOperand(0);
2992 if (!isBSwapHWordElement(N000, Parts))
2995 SDValue N001 = N00.getOperand(1);
2996 if (!isBSwapHWordElement(N001, Parts))
2998 SDValue N010 = N01.getOperand(0);
2999 if (!isBSwapHWordElement(N010, Parts))
3001 SDValue N011 = N01.getOperand(1);
3002 if (!isBSwapHWordElement(N011, Parts))
3005 // (or (or (or (and), (and)), (and)), (and))
3006 if (!isBSwapHWordElement(N1, Parts))
3008 if (!isBSwapHWordElement(N01, Parts))
3010 if (N00.getOpcode() != ISD::OR)
3012 SDValue N000 = N00.getOperand(0);
3013 if (!isBSwapHWordElement(N000, Parts))
3015 SDValue N001 = N00.getOperand(1);
3016 if (!isBSwapHWordElement(N001, Parts))
3020 // Make sure the parts are all coming from the same node.
3021 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
3024 SDValue BSwap = DAG.getNode(ISD::BSWAP, N->getDebugLoc(), VT,
3025 SDValue(Parts[0],0));
3027 // Result of the bswap should be rotated by 16. If it's not legal, than
3028 // do (x << 16) | (x >> 16).
3029 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
3030 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3031 return DAG.getNode(ISD::ROTL, N->getDebugLoc(), VT, BSwap, ShAmt);
3032 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3033 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, BSwap, ShAmt);
3034 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT,
3035 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, BSwap, ShAmt),
3036 DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, BSwap, ShAmt));
3039 SDValue DAGCombiner::visitOR(SDNode *N) {
3040 SDValue N0 = N->getOperand(0);
3041 SDValue N1 = N->getOperand(1);
3042 SDValue LL, LR, RL, RR, CC0, CC1;
3043 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3044 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3045 EVT VT = N1.getValueType();
3048 if (VT.isVector()) {
3049 SDValue FoldedVOp = SimplifyVBinOp(N);
3050 if (FoldedVOp.getNode()) return FoldedVOp;
3052 // fold (or x, 0) -> x, vector edition
3053 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3055 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3058 // fold (or x, -1) -> -1, vector edition
3059 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3061 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3065 // fold (or x, undef) -> -1
3066 if (!LegalOperations &&
3067 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3068 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3069 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
3071 // fold (or c1, c2) -> c1|c2
3073 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
3074 // canonicalize constant to RHS
3076 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N1, N0);
3077 // fold (or x, 0) -> x
3078 if (N1C && N1C->isNullValue())
3080 // fold (or x, -1) -> -1
3081 if (N1C && N1C->isAllOnesValue())
3083 // fold (or x, c) -> c iff (x & ~c) == 0
3084 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3087 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3088 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
3089 if (BSwap.getNode() != 0)
3091 BSwap = MatchBSwapHWordLow(N, N0, N1);
3092 if (BSwap.getNode() != 0)
3096 SDValue ROR = ReassociateOps(ISD::OR, N->getDebugLoc(), N0, N1);
3097 if (ROR.getNode() != 0)
3099 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3100 // iff (c1 & c2) == 0.
3101 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3102 isa<ConstantSDNode>(N0.getOperand(1))) {
3103 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3104 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0)
3105 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
3106 DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
3107 N0.getOperand(0), N1),
3108 DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1));
3110 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3111 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3112 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3113 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3115 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3116 LL.getValueType().isInteger()) {
3117 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3118 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3119 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3120 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3121 SDValue ORNode = DAG.getNode(ISD::OR, LR.getDebugLoc(),
3122 LR.getValueType(), LL, RL);
3123 AddToWorkList(ORNode.getNode());
3124 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
3126 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3127 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3128 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3129 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3130 SDValue ANDNode = DAG.getNode(ISD::AND, LR.getDebugLoc(),
3131 LR.getValueType(), LL, RL);
3132 AddToWorkList(ANDNode.getNode());
3133 return DAG.getSetCC(N->getDebugLoc(), VT, ANDNode, LR, Op1);
3136 // canonicalize equivalent to ll == rl
3137 if (LL == RR && LR == RL) {
3138 Op1 = ISD::getSetCCSwappedOperands(Op1);
3141 if (LL == RL && LR == RR) {
3142 bool isInteger = LL.getValueType().isInteger();
3143 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3144 if (Result != ISD::SETCC_INVALID &&
3145 (!LegalOperations ||
3146 TLI.isCondCodeLegal(Result, LL.getSimpleValueType())))
3147 return DAG.getSetCC(N->getDebugLoc(), N0.getValueType(),
3152 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3153 if (N0.getOpcode() == N1.getOpcode()) {
3154 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3155 if (Tmp.getNode()) return Tmp;
3158 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3159 if (N0.getOpcode() == ISD::AND &&
3160 N1.getOpcode() == ISD::AND &&
3161 N0.getOperand(1).getOpcode() == ISD::Constant &&
3162 N1.getOperand(1).getOpcode() == ISD::Constant &&
3163 // Don't increase # computations.
3164 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3165 // We can only do this xform if we know that bits from X that are set in C2
3166 // but not in C1 are already zero. Likewise for Y.
3167 const APInt &LHSMask =
3168 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3169 const APInt &RHSMask =
3170 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3172 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3173 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3174 SDValue X = DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
3175 N0.getOperand(0), N1.getOperand(0));
3176 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, X,
3177 DAG.getConstant(LHSMask | RHSMask, VT));
3181 // See if this is some rotate idiom.
3182 if (SDNode *Rot = MatchRotate(N0, N1, N->getDebugLoc()))
3183 return SDValue(Rot, 0);
3185 // Simplify the operands using demanded-bits information.
3186 if (!VT.isVector() &&
3187 SimplifyDemandedBits(SDValue(N, 0)))
3188 return SDValue(N, 0);
3193 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
3194 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3195 if (Op.getOpcode() == ISD::AND) {
3196 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3197 Mask = Op.getOperand(1);
3198 Op = Op.getOperand(0);
3204 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3212 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3213 // idioms for rotate, and if the target supports rotation instructions, generate
3215 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, DebugLoc DL) {
3216 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3217 EVT VT = LHS.getValueType();
3218 if (!TLI.isTypeLegal(VT)) return 0;
3220 // The target must have at least one rotate flavor.
3221 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3222 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3223 if (!HasROTL && !HasROTR) return 0;
3225 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3226 SDValue LHSShift; // The shift.
3227 SDValue LHSMask; // AND value if any.
3228 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3229 return 0; // Not part of a rotate.
3231 SDValue RHSShift; // The shift.
3232 SDValue RHSMask; // AND value if any.
3233 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3234 return 0; // Not part of a rotate.
3236 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3237 return 0; // Not shifting the same value.
3239 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3240 return 0; // Shifts must disagree.
3242 // Canonicalize shl to left side in a shl/srl pair.
3243 if (RHSShift.getOpcode() == ISD::SHL) {
3244 std::swap(LHS, RHS);
3245 std::swap(LHSShift, RHSShift);
3246 std::swap(LHSMask , RHSMask );
3249 unsigned OpSizeInBits = VT.getSizeInBits();
3250 SDValue LHSShiftArg = LHSShift.getOperand(0);
3251 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3252 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3254 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3255 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3256 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3257 RHSShiftAmt.getOpcode() == ISD::Constant) {
3258 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3259 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3260 if ((LShVal + RShVal) != OpSizeInBits)
3263 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3264 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
3266 // If there is an AND of either shifted operand, apply it to the result.
3267 if (LHSMask.getNode() || RHSMask.getNode()) {
3268 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3270 if (LHSMask.getNode()) {
3271 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3272 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3274 if (RHSMask.getNode()) {
3275 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3276 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3279 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3282 return Rot.getNode();
3285 // If there is a mask here, and we have a variable shift, we can't be sure
3286 // that we're masking out the right stuff.
3287 if (LHSMask.getNode() || RHSMask.getNode())
3290 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y)
3291 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotr x, (sub 32, y))
3292 if (RHSShiftAmt.getOpcode() == ISD::SUB &&
3293 LHSShiftAmt == RHSShiftAmt.getOperand(1)) {
3294 if (ConstantSDNode *SUBC =
3295 dyn_cast<ConstantSDNode>(RHSShiftAmt.getOperand(0))) {
3296 if (SUBC->getAPIntValue() == OpSizeInBits) {
3297 return DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT, LHSShiftArg,
3298 HasROTL ? LHSShiftAmt : RHSShiftAmt).getNode();
3303 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y)
3304 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotl x, (sub 32, y))
3305 if (LHSShiftAmt.getOpcode() == ISD::SUB &&
3306 RHSShiftAmt == LHSShiftAmt.getOperand(1)) {
3307 if (ConstantSDNode *SUBC =
3308 dyn_cast<ConstantSDNode>(LHSShiftAmt.getOperand(0))) {
3309 if (SUBC->getAPIntValue() == OpSizeInBits) {
3310 return DAG.getNode(HasROTR ? ISD::ROTR : ISD::ROTL, DL, VT, LHSShiftArg,
3311 HasROTR ? RHSShiftAmt : LHSShiftAmt).getNode();
3316 // Look for sign/zext/any-extended or truncate cases:
3317 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3318 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3319 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3320 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3321 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3322 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3323 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3324 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3325 SDValue LExtOp0 = LHSShiftAmt.getOperand(0);
3326 SDValue RExtOp0 = RHSShiftAmt.getOperand(0);
3327 if (RExtOp0.getOpcode() == ISD::SUB &&
3328 RExtOp0.getOperand(1) == LExtOp0) {
3329 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
3331 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
3332 // (rotr x, (sub 32, y))
3333 if (ConstantSDNode *SUBC =
3334 dyn_cast<ConstantSDNode>(RExtOp0.getOperand(0))) {
3335 if (SUBC->getAPIntValue() == OpSizeInBits) {
3336 return DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3338 HasROTL ? LHSShiftAmt : RHSShiftAmt).getNode();
3341 } else if (LExtOp0.getOpcode() == ISD::SUB &&
3342 RExtOp0 == LExtOp0.getOperand(1)) {
3343 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
3345 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
3346 // (rotl x, (sub 32, y))
3347 if (ConstantSDNode *SUBC =
3348 dyn_cast<ConstantSDNode>(LExtOp0.getOperand(0))) {
3349 if (SUBC->getAPIntValue() == OpSizeInBits) {
3350 return DAG.getNode(HasROTR ? ISD::ROTR : ISD::ROTL, DL, VT,
3352 HasROTR ? RHSShiftAmt : LHSShiftAmt).getNode();
3361 SDValue DAGCombiner::visitXOR(SDNode *N) {
3362 SDValue N0 = N->getOperand(0);
3363 SDValue N1 = N->getOperand(1);
3364 SDValue LHS, RHS, CC;
3365 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3366 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3367 EVT VT = N0.getValueType();
3370 if (VT.isVector()) {
3371 SDValue FoldedVOp = SimplifyVBinOp(N);
3372 if (FoldedVOp.getNode()) return FoldedVOp;
3374 // fold (xor x, 0) -> x, vector edition
3375 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3377 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3381 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3382 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3383 return DAG.getConstant(0, VT);
3384 // fold (xor x, undef) -> undef
3385 if (N0.getOpcode() == ISD::UNDEF)
3387 if (N1.getOpcode() == ISD::UNDEF)
3389 // fold (xor c1, c2) -> c1^c2
3391 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3392 // canonicalize constant to RHS
3394 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0);
3395 // fold (xor x, 0) -> x
3396 if (N1C && N1C->isNullValue())
3399 SDValue RXOR = ReassociateOps(ISD::XOR, N->getDebugLoc(), N0, N1);
3400 if (RXOR.getNode() != 0)
3403 // fold !(x cc y) -> (x !cc y)
3404 if (N1C && N1C->getAPIntValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3405 bool isInt = LHS.getValueType().isInteger();
3406 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3409 if (!LegalOperations ||
3410 TLI.isCondCodeLegal(NotCC, LHS.getSimpleValueType())) {
3411 switch (N0.getOpcode()) {
3413 llvm_unreachable("Unhandled SetCC Equivalent!");
3415 return DAG.getSetCC(N->getDebugLoc(), VT, LHS, RHS, NotCC);
3416 case ISD::SELECT_CC:
3417 return DAG.getSelectCC(N->getDebugLoc(), LHS, RHS, N0.getOperand(2),
3418 N0.getOperand(3), NotCC);
3423 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3424 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3425 N0.getNode()->hasOneUse() &&
3426 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3427 SDValue V = N0.getOperand(0);
3428 V = DAG.getNode(ISD::XOR, N0.getDebugLoc(), V.getValueType(), V,
3429 DAG.getConstant(1, V.getValueType()));
3430 AddToWorkList(V.getNode());
3431 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, V);
3434 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3435 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3436 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3437 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3438 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3439 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3440 LHS = DAG.getNode(ISD::XOR, LHS.getDebugLoc(), VT, LHS, N1); // LHS = ~LHS
3441 RHS = DAG.getNode(ISD::XOR, RHS.getDebugLoc(), VT, RHS, N1); // RHS = ~RHS
3442 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3443 return DAG.getNode(NewOpcode, N->getDebugLoc(), VT, LHS, RHS);
3446 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3447 if (N1C && N1C->isAllOnesValue() &&
3448 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3449 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3450 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3451 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3452 LHS = DAG.getNode(ISD::XOR, LHS.getDebugLoc(), VT, LHS, N1); // LHS = ~LHS
3453 RHS = DAG.getNode(ISD::XOR, RHS.getDebugLoc(), VT, RHS, N1); // RHS = ~RHS
3454 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3455 return DAG.getNode(NewOpcode, N->getDebugLoc(), VT, LHS, RHS);
3458 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3459 if (N1C && N0.getOpcode() == ISD::XOR) {
3460 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3461 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3463 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N0.getOperand(1),
3464 DAG.getConstant(N1C->getAPIntValue() ^
3465 N00C->getAPIntValue(), VT));
3467 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N0.getOperand(0),
3468 DAG.getConstant(N1C->getAPIntValue() ^
3469 N01C->getAPIntValue(), VT));
3471 // fold (xor x, x) -> 0
3473 return tryFoldToZero(N->getDebugLoc(), TLI, VT, DAG, LegalOperations);
3475 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
3476 if (N0.getOpcode() == N1.getOpcode()) {
3477 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3478 if (Tmp.getNode()) return Tmp;
3481 // Simplify the expression using non-local knowledge.
3482 if (!VT.isVector() &&
3483 SimplifyDemandedBits(SDValue(N, 0)))
3484 return SDValue(N, 0);
3489 /// visitShiftByConstant - Handle transforms common to the three shifts, when
3490 /// the shift amount is a constant.
3491 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, unsigned Amt) {
3492 SDNode *LHS = N->getOperand(0).getNode();
3493 if (!LHS->hasOneUse()) return SDValue();
3495 // We want to pull some binops through shifts, so that we have (and (shift))
3496 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
3497 // thing happens with address calculations, so it's important to canonicalize
3499 bool HighBitSet = false; // Can we transform this if the high bit is set?
3501 switch (LHS->getOpcode()) {
3502 default: return SDValue();
3505 HighBitSet = false; // We can only transform sra if the high bit is clear.
3508 HighBitSet = true; // We can only transform sra if the high bit is set.
3511 if (N->getOpcode() != ISD::SHL)
3512 return SDValue(); // only shl(add) not sr[al](add).
3513 HighBitSet = false; // We can only transform sra if the high bit is clear.
3517 // We require the RHS of the binop to be a constant as well.
3518 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
3519 if (!BinOpCst) return SDValue();
3521 // FIXME: disable this unless the input to the binop is a shift by a constant.
3522 // If it is not a shift, it pessimizes some common cases like:
3524 // void foo(int *X, int i) { X[i & 1235] = 1; }
3525 // int bar(int *X, int i) { return X[i & 255]; }
3526 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
3527 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
3528 BinOpLHSVal->getOpcode() != ISD::SRA &&
3529 BinOpLHSVal->getOpcode() != ISD::SRL) ||
3530 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
3533 EVT VT = N->getValueType(0);
3535 // If this is a signed shift right, and the high bit is modified by the
3536 // logical operation, do not perform the transformation. The highBitSet
3537 // boolean indicates the value of the high bit of the constant which would
3538 // cause it to be modified for this operation.
3539 if (N->getOpcode() == ISD::SRA) {
3540 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
3541 if (BinOpRHSSignSet != HighBitSet)
3545 // Fold the constants, shifting the binop RHS by the shift amount.
3546 SDValue NewRHS = DAG.getNode(N->getOpcode(), LHS->getOperand(1).getDebugLoc(),
3548 LHS->getOperand(1), N->getOperand(1));
3550 // Create the new shift.
3551 SDValue NewShift = DAG.getNode(N->getOpcode(),
3552 LHS->getOperand(0).getDebugLoc(),
3553 VT, LHS->getOperand(0), N->getOperand(1));
3555 // Create the new binop.
3556 return DAG.getNode(LHS->getOpcode(), N->getDebugLoc(), VT, NewShift, NewRHS);
3559 SDValue DAGCombiner::visitSHL(SDNode *N) {
3560 SDValue N0 = N->getOperand(0);
3561 SDValue N1 = N->getOperand(1);
3562 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3563 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3564 EVT VT = N0.getValueType();
3565 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3567 // fold (shl c1, c2) -> c1<<c2
3569 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
3570 // fold (shl 0, x) -> 0
3571 if (N0C && N0C->isNullValue())
3573 // fold (shl x, c >= size(x)) -> undef
3574 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3575 return DAG.getUNDEF(VT);
3576 // fold (shl x, 0) -> x
3577 if (N1C && N1C->isNullValue())
3579 // fold (shl undef, x) -> 0
3580 if (N0.getOpcode() == ISD::UNDEF)
3581 return DAG.getConstant(0, VT);
3582 // if (shl x, c) is known to be zero, return 0
3583 if (DAG.MaskedValueIsZero(SDValue(N, 0),
3584 APInt::getAllOnesValue(OpSizeInBits)))
3585 return DAG.getConstant(0, VT);
3586 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
3587 if (N1.getOpcode() == ISD::TRUNCATE &&
3588 N1.getOperand(0).getOpcode() == ISD::AND &&
3589 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3590 SDValue N101 = N1.getOperand(0).getOperand(1);
3591 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3592 EVT TruncVT = N1.getValueType();
3593 SDValue N100 = N1.getOperand(0).getOperand(0);
3594 APInt TruncC = N101C->getAPIntValue();
3595 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
3596 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
3597 DAG.getNode(ISD::AND, N->getDebugLoc(), TruncVT,
3598 DAG.getNode(ISD::TRUNCATE,
3601 DAG.getConstant(TruncC, TruncVT)));
3605 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3606 return SDValue(N, 0);
3608 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
3609 if (N1C && N0.getOpcode() == ISD::SHL &&
3610 N0.getOperand(1).getOpcode() == ISD::Constant) {
3611 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3612 uint64_t c2 = N1C->getZExtValue();
3613 if (c1 + c2 >= OpSizeInBits)
3614 return DAG.getConstant(0, VT);
3615 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0.getOperand(0),
3616 DAG.getConstant(c1 + c2, N1.getValueType()));
3619 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
3620 // For this to be valid, the second form must not preserve any of the bits
3621 // that are shifted out by the inner shift in the first form. This means
3622 // the outer shift size must be >= the number of bits added by the ext.
3623 // As a corollary, we don't care what kind of ext it is.
3624 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
3625 N0.getOpcode() == ISD::ANY_EXTEND ||
3626 N0.getOpcode() == ISD::SIGN_EXTEND) &&
3627 N0.getOperand(0).getOpcode() == ISD::SHL &&
3628 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
3630 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
3631 uint64_t c2 = N1C->getZExtValue();
3632 EVT InnerShiftVT = N0.getOperand(0).getValueType();
3633 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
3634 if (c2 >= OpSizeInBits - InnerShiftSize) {
3635 if (c1 + c2 >= OpSizeInBits)
3636 return DAG.getConstant(0, VT);
3637 return DAG.getNode(ISD::SHL, N0->getDebugLoc(), VT,
3638 DAG.getNode(N0.getOpcode(), N0->getDebugLoc(), VT,
3639 N0.getOperand(0)->getOperand(0)),
3640 DAG.getConstant(c1 + c2, N1.getValueType()));
3644 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
3645 // (and (srl x, (sub c1, c2), MASK)
3646 // Only fold this if the inner shift has no other uses -- if it does, folding
3647 // this will increase the total number of instructions.
3648 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse() &&
3649 N0.getOperand(1).getOpcode() == ISD::Constant) {
3650 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3651 if (c1 < VT.getSizeInBits()) {
3652 uint64_t c2 = N1C->getZExtValue();
3653 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
3654 VT.getSizeInBits() - c1);
3657 Mask = Mask.shl(c2-c1);
3658 Shift = DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0.getOperand(0),
3659 DAG.getConstant(c2-c1, N1.getValueType()));
3661 Mask = Mask.lshr(c1-c2);
3662 Shift = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0),
3663 DAG.getConstant(c1-c2, N1.getValueType()));
3665 return DAG.getNode(ISD::AND, N0.getDebugLoc(), VT, Shift,
3666 DAG.getConstant(Mask, VT));
3669 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
3670 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
3671 SDValue HiBitsMask =
3672 DAG.getConstant(APInt::getHighBitsSet(VT.getSizeInBits(),
3673 VT.getSizeInBits() -
3674 N1C->getZExtValue()),
3676 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
3681 SDValue NewSHL = visitShiftByConstant(N, N1C->getZExtValue());
3682 if (NewSHL.getNode())
3689 SDValue DAGCombiner::visitSRA(SDNode *N) {
3690 SDValue N0 = N->getOperand(0);
3691 SDValue N1 = N->getOperand(1);
3692 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3693 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3694 EVT VT = N0.getValueType();
3695 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3697 // fold (sra c1, c2) -> (sra c1, c2)
3699 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
3700 // fold (sra 0, x) -> 0
3701 if (N0C && N0C->isNullValue())
3703 // fold (sra -1, x) -> -1
3704 if (N0C && N0C->isAllOnesValue())
3706 // fold (sra x, (setge c, size(x))) -> undef
3707 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3708 return DAG.getUNDEF(VT);
3709 // fold (sra x, 0) -> x
3710 if (N1C && N1C->isNullValue())
3712 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
3714 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
3715 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
3716 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
3718 ExtVT = EVT::getVectorVT(*DAG.getContext(),
3719 ExtVT, VT.getVectorNumElements());
3720 if ((!LegalOperations ||
3721 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
3722 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
3723 N0.getOperand(0), DAG.getValueType(ExtVT));
3726 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
3727 if (N1C && N0.getOpcode() == ISD::SRA) {
3728 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3729 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
3730 if (Sum >= OpSizeInBits) Sum = OpSizeInBits-1;
3731 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0.getOperand(0),
3732 DAG.getConstant(Sum, N1C->getValueType(0)));
3736 // fold (sra (shl X, m), (sub result_size, n))
3737 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
3738 // result_size - n != m.
3739 // If truncate is free for the target sext(shl) is likely to result in better
3741 if (N0.getOpcode() == ISD::SHL) {
3742 // Get the two constanst of the shifts, CN0 = m, CN = n.
3743 const ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3745 // Determine what the truncate's result bitsize and type would be.
3747 EVT::getIntegerVT(*DAG.getContext(),
3748 OpSizeInBits - N1C->getZExtValue());
3749 // Determine the residual right-shift amount.
3750 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
3752 // If the shift is not a no-op (in which case this should be just a sign
3753 // extend already), the truncated to type is legal, sign_extend is legal
3754 // on that type, and the truncate to that type is both legal and free,
3755 // perform the transform.
3756 if ((ShiftAmt > 0) &&
3757 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
3758 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
3759 TLI.isTruncateFree(VT, TruncVT)) {
3761 SDValue Amt = DAG.getConstant(ShiftAmt,
3762 getShiftAmountTy(N0.getOperand(0).getValueType()));
3763 SDValue Shift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), VT,
3764 N0.getOperand(0), Amt);
3765 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), TruncVT,
3767 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(),
3768 N->getValueType(0), Trunc);
3773 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
3774 if (N1.getOpcode() == ISD::TRUNCATE &&
3775 N1.getOperand(0).getOpcode() == ISD::AND &&
3776 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3777 SDValue N101 = N1.getOperand(0).getOperand(1);
3778 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3779 EVT TruncVT = N1.getValueType();
3780 SDValue N100 = N1.getOperand(0).getOperand(0);
3781 APInt TruncC = N101C->getAPIntValue();
3782 TruncC = TruncC.trunc(TruncVT.getScalarType().getSizeInBits());
3783 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0,
3784 DAG.getNode(ISD::AND, N->getDebugLoc(),
3786 DAG.getNode(ISD::TRUNCATE,
3789 DAG.getConstant(TruncC, TruncVT)));
3793 // fold (sra (trunc (sr x, c1)), c2) -> (trunc (sra x, c1+c2))
3794 // if c1 is equal to the number of bits the trunc removes
3795 if (N0.getOpcode() == ISD::TRUNCATE &&
3796 (N0.getOperand(0).getOpcode() == ISD::SRL ||
3797 N0.getOperand(0).getOpcode() == ISD::SRA) &&
3798 N0.getOperand(0).hasOneUse() &&
3799 N0.getOperand(0).getOperand(1).hasOneUse() &&
3800 N1C && isa<ConstantSDNode>(N0.getOperand(0).getOperand(1))) {
3801 EVT LargeVT = N0.getOperand(0).getValueType();
3802 ConstantSDNode *LargeShiftAmt =
3803 cast<ConstantSDNode>(N0.getOperand(0).getOperand(1));
3805 if (LargeVT.getScalarType().getSizeInBits() - OpSizeInBits ==
3806 LargeShiftAmt->getZExtValue()) {
3808 DAG.getConstant(LargeShiftAmt->getZExtValue() + N1C->getZExtValue(),
3809 getShiftAmountTy(N0.getOperand(0).getOperand(0).getValueType()));
3810 SDValue SRA = DAG.getNode(ISD::SRA, N->getDebugLoc(), LargeVT,
3811 N0.getOperand(0).getOperand(0), Amt);
3812 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, SRA);
3816 // Simplify, based on bits shifted out of the LHS.
3817 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3818 return SDValue(N, 0);
3821 // If the sign bit is known to be zero, switch this to a SRL.
3822 if (DAG.SignBitIsZero(N0))
3823 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0, N1);
3826 SDValue NewSRA = visitShiftByConstant(N, N1C->getZExtValue());
3827 if (NewSRA.getNode())
3834 SDValue DAGCombiner::visitSRL(SDNode *N) {
3835 SDValue N0 = N->getOperand(0);
3836 SDValue N1 = N->getOperand(1);
3837 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3838 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3839 EVT VT = N0.getValueType();
3840 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3842 // fold (srl c1, c2) -> c1 >>u c2
3844 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
3845 // fold (srl 0, x) -> 0
3846 if (N0C && N0C->isNullValue())
3848 // fold (srl x, c >= size(x)) -> undef
3849 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3850 return DAG.getUNDEF(VT);
3851 // fold (srl x, 0) -> x
3852 if (N1C && N1C->isNullValue())
3854 // if (srl x, c) is known to be zero, return 0
3855 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
3856 APInt::getAllOnesValue(OpSizeInBits)))
3857 return DAG.getConstant(0, VT);
3859 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
3860 if (N1C && N0.getOpcode() == ISD::SRL &&
3861 N0.getOperand(1).getOpcode() == ISD::Constant) {
3862 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3863 uint64_t c2 = N1C->getZExtValue();
3864 if (c1 + c2 >= OpSizeInBits)
3865 return DAG.getConstant(0, VT);
3866 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0),
3867 DAG.getConstant(c1 + c2, N1.getValueType()));
3870 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
3871 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
3872 N0.getOperand(0).getOpcode() == ISD::SRL &&
3873 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
3875 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
3876 uint64_t c2 = N1C->getZExtValue();
3877 EVT InnerShiftVT = N0.getOperand(0).getValueType();
3878 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
3879 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
3880 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
3881 if (c1 + OpSizeInBits == InnerShiftSize) {
3882 if (c1 + c2 >= InnerShiftSize)
3883 return DAG.getConstant(0, VT);
3884 return DAG.getNode(ISD::TRUNCATE, N0->getDebugLoc(), VT,
3885 DAG.getNode(ISD::SRL, N0->getDebugLoc(), InnerShiftVT,
3886 N0.getOperand(0)->getOperand(0),
3887 DAG.getConstant(c1 + c2, ShiftCountVT)));
3891 // fold (srl (shl x, c), c) -> (and x, cst2)
3892 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1 &&
3893 N0.getValueSizeInBits() <= 64) {
3894 uint64_t ShAmt = N1C->getZExtValue()+64-N0.getValueSizeInBits();
3895 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
3896 DAG.getConstant(~0ULL >> ShAmt, VT));
3900 // fold (srl (anyextend x), c) -> (anyextend (srl x, c))
3901 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
3902 // Shifting in all undef bits?
3903 EVT SmallVT = N0.getOperand(0).getValueType();
3904 if (N1C->getZExtValue() >= SmallVT.getSizeInBits())
3905 return DAG.getUNDEF(VT);
3907 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
3908 uint64_t ShiftAmt = N1C->getZExtValue();
3909 SDValue SmallShift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), SmallVT,
3911 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
3912 AddToWorkList(SmallShift.getNode());
3913 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, SmallShift);
3917 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
3918 // bit, which is unmodified by sra.
3919 if (N1C && N1C->getZExtValue() + 1 == VT.getSizeInBits()) {
3920 if (N0.getOpcode() == ISD::SRA)
3921 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0), N1);
3924 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
3925 if (N1C && N0.getOpcode() == ISD::CTLZ &&
3926 N1C->getAPIntValue() == Log2_32(VT.getSizeInBits())) {
3927 APInt KnownZero, KnownOne;
3928 DAG.ComputeMaskedBits(N0.getOperand(0), KnownZero, KnownOne);
3930 // If any of the input bits are KnownOne, then the input couldn't be all
3931 // zeros, thus the result of the srl will always be zero.
3932 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
3934 // If all of the bits input the to ctlz node are known to be zero, then
3935 // the result of the ctlz is "32" and the result of the shift is one.
3936 APInt UnknownBits = ~KnownZero;
3937 if (UnknownBits == 0) return DAG.getConstant(1, VT);
3939 // Otherwise, check to see if there is exactly one bit input to the ctlz.
3940 if ((UnknownBits & (UnknownBits - 1)) == 0) {
3941 // Okay, we know that only that the single bit specified by UnknownBits
3942 // could be set on input to the CTLZ node. If this bit is set, the SRL
3943 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
3944 // to an SRL/XOR pair, which is likely to simplify more.
3945 unsigned ShAmt = UnknownBits.countTrailingZeros();
3946 SDValue Op = N0.getOperand(0);
3949 Op = DAG.getNode(ISD::SRL, N0.getDebugLoc(), VT, Op,
3950 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
3951 AddToWorkList(Op.getNode());
3954 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
3955 Op, DAG.getConstant(1, VT));
3959 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
3960 if (N1.getOpcode() == ISD::TRUNCATE &&
3961 N1.getOperand(0).getOpcode() == ISD::AND &&
3962 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3963 SDValue N101 = N1.getOperand(0).getOperand(1);
3964 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3965 EVT TruncVT = N1.getValueType();
3966 SDValue N100 = N1.getOperand(0).getOperand(0);
3967 APInt TruncC = N101C->getAPIntValue();
3968 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
3969 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0,
3970 DAG.getNode(ISD::AND, N->getDebugLoc(),
3972 DAG.getNode(ISD::TRUNCATE,
3975 DAG.getConstant(TruncC, TruncVT)));
3979 // fold operands of srl based on knowledge that the low bits are not
3981 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3982 return SDValue(N, 0);
3985 SDValue NewSRL = visitShiftByConstant(N, N1C->getZExtValue());
3986 if (NewSRL.getNode())
3990 // Attempt to convert a srl of a load into a narrower zero-extending load.
3991 SDValue NarrowLoad = ReduceLoadWidth(N);
3992 if (NarrowLoad.getNode())
3995 // Here is a common situation. We want to optimize:
3998 // %b = and i32 %a, 2
3999 // %c = srl i32 %b, 1
4000 // brcond i32 %c ...
4006 // %c = setcc eq %b, 0
4009 // However when after the source operand of SRL is optimized into AND, the SRL
4010 // itself may not be optimized further. Look for it and add the BRCOND into
4012 if (N->hasOneUse()) {
4013 SDNode *Use = *N->use_begin();
4014 if (Use->getOpcode() == ISD::BRCOND)
4016 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
4017 // Also look pass the truncate.
4018 Use = *Use->use_begin();
4019 if (Use->getOpcode() == ISD::BRCOND)
4027 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
4028 SDValue N0 = N->getOperand(0);
4029 EVT VT = N->getValueType(0);
4031 // fold (ctlz c1) -> c2
4032 if (isa<ConstantSDNode>(N0))
4033 return DAG.getNode(ISD::CTLZ, N->getDebugLoc(), VT, N0);
4037 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4038 SDValue N0 = N->getOperand(0);
4039 EVT VT = N->getValueType(0);
4041 // fold (ctlz_zero_undef c1) -> c2
4042 if (isa<ConstantSDNode>(N0))
4043 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, N->getDebugLoc(), VT, N0);
4047 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4048 SDValue N0 = N->getOperand(0);
4049 EVT VT = N->getValueType(0);
4051 // fold (cttz c1) -> c2
4052 if (isa<ConstantSDNode>(N0))
4053 return DAG.getNode(ISD::CTTZ, N->getDebugLoc(), VT, N0);
4057 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4058 SDValue N0 = N->getOperand(0);
4059 EVT VT = N->getValueType(0);
4061 // fold (cttz_zero_undef c1) -> c2
4062 if (isa<ConstantSDNode>(N0))
4063 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, N->getDebugLoc(), VT, N0);
4067 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4068 SDValue N0 = N->getOperand(0);
4069 EVT VT = N->getValueType(0);
4071 // fold (ctpop c1) -> c2
4072 if (isa<ConstantSDNode>(N0))
4073 return DAG.getNode(ISD::CTPOP, N->getDebugLoc(), VT, N0);
4077 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4078 SDValue N0 = N->getOperand(0);
4079 SDValue N1 = N->getOperand(1);
4080 SDValue N2 = N->getOperand(2);
4081 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4082 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4083 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
4084 EVT VT = N->getValueType(0);
4085 EVT VT0 = N0.getValueType();
4087 // fold (select C, X, X) -> X
4090 // fold (select true, X, Y) -> X
4091 if (N0C && !N0C->isNullValue())
4093 // fold (select false, X, Y) -> Y
4094 if (N0C && N0C->isNullValue())
4096 // fold (select C, 1, X) -> (or C, X)
4097 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4098 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N2);
4099 // fold (select C, 0, 1) -> (xor C, 1)
4100 if (VT.isInteger() &&
4103 TLI.getBooleanContents(false) ==
4104 TargetLowering::ZeroOrOneBooleanContent)) &&
4105 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4108 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT0,
4109 N0, DAG.getConstant(1, VT0));
4110 XORNode = DAG.getNode(ISD::XOR, N0.getDebugLoc(), VT0,
4111 N0, DAG.getConstant(1, VT0));
4112 AddToWorkList(XORNode.getNode());
4114 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, XORNode);
4115 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, XORNode);
4117 // fold (select C, 0, X) -> (and (not C), X)
4118 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4119 SDValue NOTNode = DAG.getNOT(N0.getDebugLoc(), N0, VT);
4120 AddToWorkList(NOTNode.getNode());
4121 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, NOTNode, N2);
4123 // fold (select C, X, 1) -> (or (not C), X)
4124 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4125 SDValue NOTNode = DAG.getNOT(N0.getDebugLoc(), N0, VT);
4126 AddToWorkList(NOTNode.getNode());
4127 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, NOTNode, N1);
4129 // fold (select C, X, 0) -> (and C, X)
4130 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4131 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, N1);
4132 // fold (select X, X, Y) -> (or X, Y)
4133 // fold (select X, 1, Y) -> (or X, Y)
4134 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4135 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N2);
4136 // fold (select X, Y, X) -> (and X, Y)
4137 // fold (select X, Y, 0) -> (and X, Y)
4138 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4139 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, N1);
4141 // If we can fold this based on the true/false value, do so.
4142 if (SimplifySelectOps(N, N1, N2))
4143 return SDValue(N, 0); // Don't revisit N.
4145 // fold selects based on a setcc into other things, such as min/max/abs
4146 if (N0.getOpcode() == ISD::SETCC) {
4148 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
4149 // having to say they don't support SELECT_CC on every type the DAG knows
4150 // about, since there is no way to mark an opcode illegal at all value types
4151 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other) &&
4152 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT))
4153 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT,
4154 N0.getOperand(0), N0.getOperand(1),
4155 N1, N2, N0.getOperand(2));
4156 return SimplifySelect(N->getDebugLoc(), N0, N1, N2);
4162 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
4163 SDValue N0 = N->getOperand(0);
4164 SDValue N1 = N->getOperand(1);
4165 SDValue N2 = N->getOperand(2);
4166 SDValue N3 = N->getOperand(3);
4167 SDValue N4 = N->getOperand(4);
4168 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
4170 // fold select_cc lhs, rhs, x, x, cc -> x
4174 // Determine if the condition we're dealing with is constant
4175 SDValue SCC = SimplifySetCC(TLI.getSetCCResultType(N0.getValueType()),
4176 N0, N1, CC, N->getDebugLoc(), false);
4177 if (SCC.getNode()) AddToWorkList(SCC.getNode());
4179 if (ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode())) {
4180 if (!SCCC->isNullValue())
4181 return N2; // cond always true -> true val
4183 return N3; // cond always false -> false val
4186 // Fold to a simpler select_cc
4187 if (SCC.getNode() && SCC.getOpcode() == ISD::SETCC)
4188 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), N2.getValueType(),
4189 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
4192 // If we can fold this based on the true/false value, do so.
4193 if (SimplifySelectOps(N, N2, N3))
4194 return SDValue(N, 0); // Don't revisit N.
4196 // fold select_cc into other things, such as min/max/abs
4197 return SimplifySelectCC(N->getDebugLoc(), N0, N1, N2, N3, CC);
4200 SDValue DAGCombiner::visitSETCC(SDNode *N) {
4201 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
4202 cast<CondCodeSDNode>(N->getOperand(2))->get(),
4206 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
4207 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
4208 // transformation. Returns true if extension are possible and the above
4209 // mentioned transformation is profitable.
4210 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
4212 SmallVector<SDNode*, 4> &ExtendNodes,
4213 const TargetLowering &TLI) {
4214 bool HasCopyToRegUses = false;
4215 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
4216 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
4217 UE = N0.getNode()->use_end();
4222 if (UI.getUse().getResNo() != N0.getResNo())
4224 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
4225 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
4226 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
4227 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
4228 // Sign bits will be lost after a zext.
4231 for (unsigned i = 0; i != 2; ++i) {
4232 SDValue UseOp = User->getOperand(i);
4235 if (!isa<ConstantSDNode>(UseOp))
4240 ExtendNodes.push_back(User);
4243 // If truncates aren't free and there are users we can't
4244 // extend, it isn't worthwhile.
4247 // Remember if this value is live-out.
4248 if (User->getOpcode() == ISD::CopyToReg)
4249 HasCopyToRegUses = true;
4252 if (HasCopyToRegUses) {
4253 bool BothLiveOut = false;
4254 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
4256 SDUse &Use = UI.getUse();
4257 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
4263 // Both unextended and extended values are live out. There had better be
4264 // a good reason for the transformation.
4265 return ExtendNodes.size();
4270 void DAGCombiner::ExtendSetCCUses(SmallVector<SDNode*, 4> SetCCs,
4271 SDValue Trunc, SDValue ExtLoad, DebugLoc DL,
4272 ISD::NodeType ExtType) {
4273 // Extend SetCC uses if necessary.
4274 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
4275 SDNode *SetCC = SetCCs[i];
4276 SmallVector<SDValue, 4> Ops;
4278 for (unsigned j = 0; j != 2; ++j) {
4279 SDValue SOp = SetCC->getOperand(j);
4281 Ops.push_back(ExtLoad);
4283 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
4286 Ops.push_back(SetCC->getOperand(2));
4287 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0),
4288 &Ops[0], Ops.size()));
4292 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
4293 SDValue N0 = N->getOperand(0);
4294 EVT VT = N->getValueType(0);
4296 // fold (sext c1) -> c1
4297 if (isa<ConstantSDNode>(N0))
4298 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N0);
4300 // fold (sext (sext x)) -> (sext x)
4301 // fold (sext (aext x)) -> (sext x)
4302 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4303 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT,
4306 if (N0.getOpcode() == ISD::TRUNCATE) {
4307 // fold (sext (truncate (load x))) -> (sext (smaller load x))
4308 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
4309 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4310 if (NarrowLoad.getNode()) {
4311 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4312 if (NarrowLoad.getNode() != N0.getNode()) {
4313 CombineTo(N0.getNode(), NarrowLoad);
4314 // CombineTo deleted the truncate, if needed, but not what's under it.
4317 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4320 // See if the value being truncated is already sign extended. If so, just
4321 // eliminate the trunc/sext pair.
4322 SDValue Op = N0.getOperand(0);
4323 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
4324 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
4325 unsigned DestBits = VT.getScalarType().getSizeInBits();
4326 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
4328 if (OpBits == DestBits) {
4329 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
4330 // bits, it is already ready.
4331 if (NumSignBits > DestBits-MidBits)
4333 } else if (OpBits < DestBits) {
4334 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
4335 // bits, just sext from i32.
4336 if (NumSignBits > OpBits-MidBits)
4337 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, Op);
4339 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
4340 // bits, just truncate to i32.
4341 if (NumSignBits > OpBits-MidBits)
4342 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4345 // fold (sext (truncate x)) -> (sextinreg x).
4346 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
4347 N0.getValueType())) {
4348 if (OpBits < DestBits)
4349 Op = DAG.getNode(ISD::ANY_EXTEND, N0.getDebugLoc(), VT, Op);
4350 else if (OpBits > DestBits)
4351 Op = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), VT, Op);
4352 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT, Op,
4353 DAG.getValueType(N0.getValueType()));
4357 // fold (sext (load x)) -> (sext (truncate (sextload x)))
4358 // None of the supported targets knows how to perform load and sign extend
4359 // on vectors in one instruction. We only perform this transformation on
4361 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4362 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4363 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) {
4364 bool DoXform = true;
4365 SmallVector<SDNode*, 4> SetCCs;
4366 if (!N0.hasOneUse())
4367 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
4369 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4370 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
4372 LN0->getBasePtr(), LN0->getPointerInfo(),
4374 LN0->isVolatile(), LN0->isNonTemporal(),
4375 LN0->getAlignment());
4376 CombineTo(N, ExtLoad);
4377 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4378 N0.getValueType(), ExtLoad);
4379 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4380 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4382 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4386 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
4387 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
4388 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
4389 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
4390 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4391 EVT MemVT = LN0->getMemoryVT();
4392 if ((!LegalOperations && !LN0->isVolatile()) ||
4393 TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) {
4394 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
4396 LN0->getBasePtr(), LN0->getPointerInfo(),
4398 LN0->isVolatile(), LN0->isNonTemporal(),
4399 LN0->getAlignment());
4400 CombineTo(N, ExtLoad);
4401 CombineTo(N0.getNode(),
4402 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4403 N0.getValueType(), ExtLoad),
4404 ExtLoad.getValue(1));
4405 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4409 // fold (sext (and/or/xor (load x), cst)) ->
4410 // (and/or/xor (sextload x), (sext cst))
4411 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
4412 N0.getOpcode() == ISD::XOR) &&
4413 isa<LoadSDNode>(N0.getOperand(0)) &&
4414 N0.getOperand(1).getOpcode() == ISD::Constant &&
4415 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()) &&
4416 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
4417 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
4418 if (LN0->getExtensionType() != ISD::ZEXTLOAD) {
4419 bool DoXform = true;
4420 SmallVector<SDNode*, 4> SetCCs;
4421 if (!N0.hasOneUse())
4422 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
4425 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, LN0->getDebugLoc(), VT,
4426 LN0->getChain(), LN0->getBasePtr(),
4427 LN0->getPointerInfo(),
4430 LN0->isNonTemporal(),
4431 LN0->getAlignment());
4432 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4433 Mask = Mask.sext(VT.getSizeInBits());
4434 SDValue And = DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
4435 ExtLoad, DAG.getConstant(Mask, VT));
4436 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
4437 N0.getOperand(0).getDebugLoc(),
4438 N0.getOperand(0).getValueType(), ExtLoad);
4440 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
4441 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4443 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4448 if (N0.getOpcode() == ISD::SETCC) {
4449 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
4450 // Only do this before legalize for now.
4451 if (VT.isVector() && !LegalOperations) {
4452 EVT N0VT = N0.getOperand(0).getValueType();
4453 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
4454 // of the same size as the compared operands. Only optimize sext(setcc())
4455 // if this is the case.
4456 EVT SVT = TLI.getSetCCResultType(N0VT);
4458 // We know that the # elements of the results is the same as the
4459 // # elements of the compare (and the # elements of the compare result
4460 // for that matter). Check to see that they are the same size. If so,
4461 // we know that the element size of the sext'd result matches the
4462 // element size of the compare operands.
4463 if (VT.getSizeInBits() == SVT.getSizeInBits())
4464 return DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4466 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4467 // If the desired elements are smaller or larger than the source
4468 // elements we can use a matching integer vector type and then
4469 // truncate/sign extend
4470 EVT MatchingElementType =
4471 EVT::getIntegerVT(*DAG.getContext(),
4472 N0VT.getScalarType().getSizeInBits());
4473 EVT MatchingVectorType =
4474 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4475 N0VT.getVectorNumElements());
4477 if (SVT == MatchingVectorType) {
4478 SDValue VsetCC = DAG.getSetCC(N->getDebugLoc(), MatchingVectorType,
4479 N0.getOperand(0), N0.getOperand(1),
4480 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4481 return DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT);
4485 // sext(setcc x, y, cc) -> (select_cc x, y, -1, 0, cc)
4486 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
4488 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
4490 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4491 NegOne, DAG.getConstant(0, VT),
4492 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4493 if (SCC.getNode()) return SCC;
4494 if (!LegalOperations ||
4495 TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(VT)))
4496 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4497 DAG.getSetCC(N->getDebugLoc(),
4498 TLI.getSetCCResultType(VT),
4499 N0.getOperand(0), N0.getOperand(1),
4500 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
4501 NegOne, DAG.getConstant(0, VT));
4504 // fold (sext x) -> (zext x) if the sign bit is known zero.
4505 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
4506 DAG.SignBitIsZero(N0))
4507 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
4512 // isTruncateOf - If N is a truncate of some other value, return true, record
4513 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
4514 // This function computes KnownZero to avoid a duplicated call to
4515 // ComputeMaskedBits in the caller.
4516 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
4519 if (N->getOpcode() == ISD::TRUNCATE) {
4520 Op = N->getOperand(0);
4521 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
4525 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
4526 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
4529 SDValue Op0 = N->getOperand(0);
4530 SDValue Op1 = N->getOperand(1);
4531 assert(Op0.getValueType() == Op1.getValueType());
4533 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
4534 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
4535 if (COp0 && COp0->isNullValue())
4537 else if (COp1 && COp1->isNullValue())
4542 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
4544 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
4550 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
4551 SDValue N0 = N->getOperand(0);
4552 EVT VT = N->getValueType(0);
4554 // fold (zext c1) -> c1
4555 if (isa<ConstantSDNode>(N0))
4556 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
4557 // fold (zext (zext x)) -> (zext x)
4558 // fold (zext (aext x)) -> (zext x)
4559 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4560 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT,
4563 // fold (zext (truncate x)) -> (zext x) or
4564 // (zext (truncate x)) -> (truncate x)
4565 // This is valid when the truncated bits of x are already zero.
4566 // FIXME: We should extend this to work for vectors too.
4569 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
4570 APInt TruncatedBits =
4571 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
4572 APInt(Op.getValueSizeInBits(), 0) :
4573 APInt::getBitsSet(Op.getValueSizeInBits(),
4574 N0.getValueSizeInBits(),
4575 std::min(Op.getValueSizeInBits(),
4576 VT.getSizeInBits()));
4577 if (TruncatedBits == (KnownZero & TruncatedBits)) {
4578 if (VT.bitsGT(Op.getValueType()))
4579 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, Op);
4580 if (VT.bitsLT(Op.getValueType()))
4581 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4587 // fold (zext (truncate (load x))) -> (zext (smaller load x))
4588 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
4589 if (N0.getOpcode() == ISD::TRUNCATE) {
4590 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4591 if (NarrowLoad.getNode()) {
4592 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4593 if (NarrowLoad.getNode() != N0.getNode()) {
4594 CombineTo(N0.getNode(), NarrowLoad);
4595 // CombineTo deleted the truncate, if needed, but not what's under it.
4598 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4602 // fold (zext (truncate x)) -> (and x, mask)
4603 if (N0.getOpcode() == ISD::TRUNCATE &&
4604 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
4606 // fold (zext (truncate (load x))) -> (zext (smaller load x))
4607 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
4608 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4609 if (NarrowLoad.getNode()) {
4610 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4611 if (NarrowLoad.getNode() != N0.getNode()) {
4612 CombineTo(N0.getNode(), NarrowLoad);
4613 // CombineTo deleted the truncate, if needed, but not what's under it.
4616 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4619 SDValue Op = N0.getOperand(0);
4620 if (Op.getValueType().bitsLT(VT)) {
4621 Op = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, Op);
4622 AddToWorkList(Op.getNode());
4623 } else if (Op.getValueType().bitsGT(VT)) {
4624 Op = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4625 AddToWorkList(Op.getNode());
4627 return DAG.getZeroExtendInReg(Op, N->getDebugLoc(),
4628 N0.getValueType().getScalarType());
4631 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
4632 // if either of the casts is not free.
4633 if (N0.getOpcode() == ISD::AND &&
4634 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
4635 N0.getOperand(1).getOpcode() == ISD::Constant &&
4636 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
4637 N0.getValueType()) ||
4638 !TLI.isZExtFree(N0.getValueType(), VT))) {
4639 SDValue X = N0.getOperand(0).getOperand(0);
4640 if (X.getValueType().bitsLT(VT)) {
4641 X = DAG.getNode(ISD::ANY_EXTEND, X.getDebugLoc(), VT, X);
4642 } else if (X.getValueType().bitsGT(VT)) {
4643 X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
4645 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4646 Mask = Mask.zext(VT.getSizeInBits());
4647 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4648 X, DAG.getConstant(Mask, VT));
4651 // fold (zext (load x)) -> (zext (truncate (zextload x)))
4652 // None of the supported targets knows how to perform load and vector_zext
4653 // on vectors in one instruction. We only perform this transformation on
4655 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4656 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4657 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
4658 bool DoXform = true;
4659 SmallVector<SDNode*, 4> SetCCs;
4660 if (!N0.hasOneUse())
4661 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
4663 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4664 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N->getDebugLoc(), VT,
4666 LN0->getBasePtr(), LN0->getPointerInfo(),
4668 LN0->isVolatile(), LN0->isNonTemporal(),
4669 LN0->getAlignment());
4670 CombineTo(N, ExtLoad);
4671 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4672 N0.getValueType(), ExtLoad);
4673 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4675 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4677 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4681 // fold (zext (and/or/xor (load x), cst)) ->
4682 // (and/or/xor (zextload x), (zext cst))
4683 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
4684 N0.getOpcode() == ISD::XOR) &&
4685 isa<LoadSDNode>(N0.getOperand(0)) &&
4686 N0.getOperand(1).getOpcode() == ISD::Constant &&
4687 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()) &&
4688 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
4689 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
4690 if (LN0->getExtensionType() != ISD::SEXTLOAD) {
4691 bool DoXform = true;
4692 SmallVector<SDNode*, 4> SetCCs;
4693 if (!N0.hasOneUse())
4694 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
4697 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), VT,
4698 LN0->getChain(), LN0->getBasePtr(),
4699 LN0->getPointerInfo(),
4702 LN0->isNonTemporal(),
4703 LN0->getAlignment());
4704 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4705 Mask = Mask.zext(VT.getSizeInBits());
4706 SDValue And = DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
4707 ExtLoad, DAG.getConstant(Mask, VT));
4708 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
4709 N0.getOperand(0).getDebugLoc(),
4710 N0.getOperand(0).getValueType(), ExtLoad);
4712 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
4713 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4715 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4720 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
4721 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
4722 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
4723 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
4724 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4725 EVT MemVT = LN0->getMemoryVT();
4726 if ((!LegalOperations && !LN0->isVolatile()) ||
4727 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) {
4728 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N->getDebugLoc(), VT,
4730 LN0->getBasePtr(), LN0->getPointerInfo(),
4732 LN0->isVolatile(), LN0->isNonTemporal(),
4733 LN0->getAlignment());
4734 CombineTo(N, ExtLoad);
4735 CombineTo(N0.getNode(),
4736 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), N0.getValueType(),
4738 ExtLoad.getValue(1));
4739 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4743 if (N0.getOpcode() == ISD::SETCC) {
4744 if (!LegalOperations && VT.isVector()) {
4745 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
4746 // Only do this before legalize for now.
4747 EVT N0VT = N0.getOperand(0).getValueType();
4748 EVT EltVT = VT.getVectorElementType();
4749 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
4750 DAG.getConstant(1, EltVT));
4751 if (VT.getSizeInBits() == N0VT.getSizeInBits())
4752 // We know that the # elements of the results is the same as the
4753 // # elements of the compare (and the # elements of the compare result
4754 // for that matter). Check to see that they are the same size. If so,
4755 // we know that the element size of the sext'd result matches the
4756 // element size of the compare operands.
4757 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4758 DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4760 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
4761 DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
4762 &OneOps[0], OneOps.size()));
4764 // If the desired elements are smaller or larger than the source
4765 // elements we can use a matching integer vector type and then
4766 // truncate/sign extend
4767 EVT MatchingElementType =
4768 EVT::getIntegerVT(*DAG.getContext(),
4769 N0VT.getScalarType().getSizeInBits());
4770 EVT MatchingVectorType =
4771 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4772 N0VT.getVectorNumElements());
4774 DAG.getSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
4776 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4777 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4778 DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT),
4779 DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
4780 &OneOps[0], OneOps.size()));
4783 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
4785 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4786 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
4787 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4788 if (SCC.getNode()) return SCC;
4791 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
4792 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
4793 isa<ConstantSDNode>(N0.getOperand(1)) &&
4794 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
4796 SDValue ShAmt = N0.getOperand(1);
4797 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
4798 if (N0.getOpcode() == ISD::SHL) {
4799 SDValue InnerZExt = N0.getOperand(0);
4800 // If the original shl may be shifting out bits, do not perform this
4802 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
4803 InnerZExt.getOperand(0).getValueType().getSizeInBits();
4804 if (ShAmtVal > KnownZeroBits)
4808 DebugLoc DL = N->getDebugLoc();
4810 // Ensure that the shift amount is wide enough for the shifted value.
4811 if (VT.getSizeInBits() >= 256)
4812 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
4814 return DAG.getNode(N0.getOpcode(), DL, VT,
4815 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
4822 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
4823 SDValue N0 = N->getOperand(0);
4824 EVT VT = N->getValueType(0);
4826 // fold (aext c1) -> c1
4827 if (isa<ConstantSDNode>(N0))
4828 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, N0);
4829 // fold (aext (aext x)) -> (aext x)
4830 // fold (aext (zext x)) -> (zext x)
4831 // fold (aext (sext x)) -> (sext x)
4832 if (N0.getOpcode() == ISD::ANY_EXTEND ||
4833 N0.getOpcode() == ISD::ZERO_EXTEND ||
4834 N0.getOpcode() == ISD::SIGN_EXTEND)
4835 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, N0.getOperand(0));
4837 // fold (aext (truncate (load x))) -> (aext (smaller load x))
4838 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
4839 if (N0.getOpcode() == ISD::TRUNCATE) {
4840 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4841 if (NarrowLoad.getNode()) {
4842 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4843 if (NarrowLoad.getNode() != N0.getNode()) {
4844 CombineTo(N0.getNode(), NarrowLoad);
4845 // CombineTo deleted the truncate, if needed, but not what's under it.
4848 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4852 // fold (aext (truncate x))
4853 if (N0.getOpcode() == ISD::TRUNCATE) {
4854 SDValue TruncOp = N0.getOperand(0);
4855 if (TruncOp.getValueType() == VT)
4856 return TruncOp; // x iff x size == zext size.
4857 if (TruncOp.getValueType().bitsGT(VT))
4858 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, TruncOp);
4859 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, TruncOp);
4862 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
4863 // if the trunc is not free.
4864 if (N0.getOpcode() == ISD::AND &&
4865 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
4866 N0.getOperand(1).getOpcode() == ISD::Constant &&
4867 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
4868 N0.getValueType())) {
4869 SDValue X = N0.getOperand(0).getOperand(0);
4870 if (X.getValueType().bitsLT(VT)) {
4871 X = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, X);
4872 } else if (X.getValueType().bitsGT(VT)) {
4873 X = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, X);
4875 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4876 Mask = Mask.zext(VT.getSizeInBits());
4877 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4878 X, DAG.getConstant(Mask, VT));
4881 // fold (aext (load x)) -> (aext (truncate (extload x)))
4882 // None of the supported targets knows how to perform load and any_ext
4883 // on vectors in one instruction. We only perform this transformation on
4885 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4886 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4887 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
4888 bool DoXform = true;
4889 SmallVector<SDNode*, 4> SetCCs;
4890 if (!N0.hasOneUse())
4891 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
4893 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4894 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
4896 LN0->getBasePtr(), LN0->getPointerInfo(),
4898 LN0->isVolatile(), LN0->isNonTemporal(),
4899 LN0->getAlignment());
4900 CombineTo(N, ExtLoad);
4901 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4902 N0.getValueType(), ExtLoad);
4903 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4904 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4906 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4910 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
4911 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
4912 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
4913 if (N0.getOpcode() == ISD::LOAD &&
4914 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
4916 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4917 EVT MemVT = LN0->getMemoryVT();
4918 SDValue ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), N->getDebugLoc(),
4919 VT, LN0->getChain(), LN0->getBasePtr(),
4920 LN0->getPointerInfo(), MemVT,
4921 LN0->isVolatile(), LN0->isNonTemporal(),
4922 LN0->getAlignment());
4923 CombineTo(N, ExtLoad);
4924 CombineTo(N0.getNode(),
4925 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4926 N0.getValueType(), ExtLoad),
4927 ExtLoad.getValue(1));
4928 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4931 if (N0.getOpcode() == ISD::SETCC) {
4932 // aext(setcc) -> sext_in_reg(vsetcc) for vectors.
4933 // Only do this before legalize for now.
4934 if (VT.isVector() && !LegalOperations) {
4935 EVT N0VT = N0.getOperand(0).getValueType();
4936 // We know that the # elements of the results is the same as the
4937 // # elements of the compare (and the # elements of the compare result
4938 // for that matter). Check to see that they are the same size. If so,
4939 // we know that the element size of the sext'd result matches the
4940 // element size of the compare operands.
4941 if (VT.getSizeInBits() == N0VT.getSizeInBits())
4942 return DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4944 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4945 // If the desired elements are smaller or larger than the source
4946 // elements we can use a matching integer vector type and then
4947 // truncate/sign extend
4949 EVT MatchingElementType =
4950 EVT::getIntegerVT(*DAG.getContext(),
4951 N0VT.getScalarType().getSizeInBits());
4952 EVT MatchingVectorType =
4953 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4954 N0VT.getVectorNumElements());
4956 DAG.getSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
4958 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4959 return DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT);
4963 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
4965 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4966 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
4967 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4975 /// GetDemandedBits - See if the specified operand can be simplified with the
4976 /// knowledge that only the bits specified by Mask are used. If so, return the
4977 /// simpler operand, otherwise return a null SDValue.
4978 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
4979 switch (V.getOpcode()) {
4981 case ISD::Constant: {
4982 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
4983 assert(CV != 0 && "Const value should be ConstSDNode.");
4984 const APInt &CVal = CV->getAPIntValue();
4985 APInt NewVal = CVal & Mask;
4986 if (NewVal != CVal) {
4987 return DAG.getConstant(NewVal, V.getValueType());
4993 // If the LHS or RHS don't contribute bits to the or, drop them.
4994 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
4995 return V.getOperand(1);
4996 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
4997 return V.getOperand(0);
5000 // Only look at single-use SRLs.
5001 if (!V.getNode()->hasOneUse())
5003 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
5004 // See if we can recursively simplify the LHS.
5005 unsigned Amt = RHSC->getZExtValue();
5007 // Watch out for shift count overflow though.
5008 if (Amt >= Mask.getBitWidth()) break;
5009 APInt NewMask = Mask << Amt;
5010 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
5011 if (SimplifyLHS.getNode())
5012 return DAG.getNode(ISD::SRL, V.getDebugLoc(), V.getValueType(),
5013 SimplifyLHS, V.getOperand(1));
5019 /// ReduceLoadWidth - If the result of a wider load is shifted to right of N
5020 /// bits and then truncated to a narrower type and where N is a multiple
5021 /// of number of bits of the narrower type, transform it to a narrower load
5022 /// from address + N / num of bits of new type. If the result is to be
5023 /// extended, also fold the extension to form a extending load.
5024 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
5025 unsigned Opc = N->getOpcode();
5027 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
5028 SDValue N0 = N->getOperand(0);
5029 EVT VT = N->getValueType(0);
5032 // This transformation isn't valid for vector loads.
5036 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
5038 if (Opc == ISD::SIGN_EXTEND_INREG) {
5039 ExtType = ISD::SEXTLOAD;
5040 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
5041 } else if (Opc == ISD::SRL) {
5042 // Another special-case: SRL is basically zero-extending a narrower value.
5043 ExtType = ISD::ZEXTLOAD;
5045 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
5046 if (!N01) return SDValue();
5047 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
5048 VT.getSizeInBits() - N01->getZExtValue());
5050 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, ExtVT))
5053 unsigned EVTBits = ExtVT.getSizeInBits();
5055 // Do not generate loads of non-round integer types since these can
5056 // be expensive (and would be wrong if the type is not byte sized).
5057 if (!ExtVT.isRound())
5061 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
5062 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5063 ShAmt = N01->getZExtValue();
5064 // Is the shift amount a multiple of size of VT?
5065 if ((ShAmt & (EVTBits-1)) == 0) {
5066 N0 = N0.getOperand(0);
5067 // Is the load width a multiple of size of VT?
5068 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
5072 // At this point, we must have a load or else we can't do the transform.
5073 if (!isa<LoadSDNode>(N0)) return SDValue();
5075 // Because a SRL must be assumed to *need* to zero-extend the high bits
5076 // (as opposed to anyext the high bits), we can't combine the zextload
5077 // lowering of SRL and an sextload.
5078 if (cast<LoadSDNode>(N0)->getExtensionType() == ISD::SEXTLOAD)
5081 // If the shift amount is larger than the input type then we're not
5082 // accessing any of the loaded bytes. If the load was a zextload/extload
5083 // then the result of the shift+trunc is zero/undef (handled elsewhere).
5084 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
5089 // If the load is shifted left (and the result isn't shifted back right),
5090 // we can fold the truncate through the shift.
5091 unsigned ShLeftAmt = 0;
5092 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
5093 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
5094 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5095 ShLeftAmt = N01->getZExtValue();
5096 N0 = N0.getOperand(0);
5100 // If we haven't found a load, we can't narrow it. Don't transform one with
5101 // multiple uses, this would require adding a new load.
5102 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse() ||
5103 // Don't change the width of a volatile load.
5104 cast<LoadSDNode>(N0)->isVolatile())
5107 // Verify that we are actually reducing a load width here.
5108 if (cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits() < EVTBits)
5111 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5112 EVT PtrType = N0.getOperand(1).getValueType();
5114 if (PtrType == MVT::Untyped || PtrType.isExtended())
5115 // It's not possible to generate a constant of extended or untyped type.
5118 // For big endian targets, we need to adjust the offset to the pointer to
5119 // load the correct bytes.
5120 if (TLI.isBigEndian()) {
5121 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
5122 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
5123 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
5126 uint64_t PtrOff = ShAmt / 8;
5127 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
5128 SDValue NewPtr = DAG.getNode(ISD::ADD, LN0->getDebugLoc(),
5129 PtrType, LN0->getBasePtr(),
5130 DAG.getConstant(PtrOff, PtrType));
5131 AddToWorkList(NewPtr.getNode());
5134 if (ExtType == ISD::NON_EXTLOAD)
5135 Load = DAG.getLoad(VT, N0.getDebugLoc(), LN0->getChain(), NewPtr,
5136 LN0->getPointerInfo().getWithOffset(PtrOff),
5137 LN0->isVolatile(), LN0->isNonTemporal(),
5138 LN0->isInvariant(), NewAlign);
5140 Load = DAG.getExtLoad(ExtType, N0.getDebugLoc(), VT, LN0->getChain(),NewPtr,
5141 LN0->getPointerInfo().getWithOffset(PtrOff),
5142 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
5145 // Replace the old load's chain with the new load's chain.
5146 WorkListRemover DeadNodes(*this);
5147 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
5149 // Shift the result left, if we've swallowed a left shift.
5150 SDValue Result = Load;
5151 if (ShLeftAmt != 0) {
5152 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
5153 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
5155 Result = DAG.getNode(ISD::SHL, N0.getDebugLoc(), VT,
5156 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
5159 // Return the new loaded value.
5163 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
5164 SDValue N0 = N->getOperand(0);
5165 SDValue N1 = N->getOperand(1);
5166 EVT VT = N->getValueType(0);
5167 EVT EVT = cast<VTSDNode>(N1)->getVT();
5168 unsigned VTBits = VT.getScalarType().getSizeInBits();
5169 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
5171 // fold (sext_in_reg c1) -> c1
5172 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
5173 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT, N0, N1);
5175 // If the input is already sign extended, just drop the extension.
5176 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
5179 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
5180 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
5181 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT())) {
5182 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
5183 N0.getOperand(0), N1);
5186 // fold (sext_in_reg (sext x)) -> (sext x)
5187 // fold (sext_in_reg (aext x)) -> (sext x)
5188 // if x is small enough.
5189 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
5190 SDValue N00 = N0.getOperand(0);
5191 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
5192 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
5193 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N00, N1);
5196 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
5197 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
5198 return DAG.getZeroExtendInReg(N0, N->getDebugLoc(), EVT);
5200 // fold operands of sext_in_reg based on knowledge that the top bits are not
5202 if (SimplifyDemandedBits(SDValue(N, 0)))
5203 return SDValue(N, 0);
5205 // fold (sext_in_reg (load x)) -> (smaller sextload x)
5206 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
5207 SDValue NarrowLoad = ReduceLoadWidth(N);
5208 if (NarrowLoad.getNode())
5211 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
5212 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
5213 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
5214 if (N0.getOpcode() == ISD::SRL) {
5215 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
5216 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
5217 // We can turn this into an SRA iff the input to the SRL is already sign
5219 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
5220 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
5221 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT,
5222 N0.getOperand(0), N0.getOperand(1));
5226 // fold (sext_inreg (extload x)) -> (sextload x)
5227 if (ISD::isEXTLoad(N0.getNode()) &&
5228 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5229 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5230 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5231 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5232 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5233 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
5235 LN0->getBasePtr(), LN0->getPointerInfo(),
5237 LN0->isVolatile(), LN0->isNonTemporal(),
5238 LN0->getAlignment());
5239 CombineTo(N, ExtLoad);
5240 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5241 AddToWorkList(ExtLoad.getNode());
5242 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5244 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
5245 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5247 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5248 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5249 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5250 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5251 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
5253 LN0->getBasePtr(), LN0->getPointerInfo(),
5255 LN0->isVolatile(), LN0->isNonTemporal(),
5256 LN0->getAlignment());
5257 CombineTo(N, ExtLoad);
5258 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5259 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5262 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
5263 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
5264 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
5265 N0.getOperand(1), false);
5266 if (BSwap.getNode() != 0)
5267 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
5274 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
5275 SDValue N0 = N->getOperand(0);
5276 EVT VT = N->getValueType(0);
5277 bool isLE = TLI.isLittleEndian();
5280 if (N0.getValueType() == N->getValueType(0))
5282 // fold (truncate c1) -> c1
5283 if (isa<ConstantSDNode>(N0))
5284 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0);
5285 // fold (truncate (truncate x)) -> (truncate x)
5286 if (N0.getOpcode() == ISD::TRUNCATE)
5287 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0.getOperand(0));
5288 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
5289 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
5290 N0.getOpcode() == ISD::SIGN_EXTEND ||
5291 N0.getOpcode() == ISD::ANY_EXTEND) {
5292 if (N0.getOperand(0).getValueType().bitsLT(VT))
5293 // if the source is smaller than the dest, we still need an extend
5294 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
5296 if (N0.getOperand(0).getValueType().bitsGT(VT))
5297 // if the source is larger than the dest, than we just need the truncate
5298 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0.getOperand(0));
5299 // if the source and dest are the same type, we can drop both the extend
5300 // and the truncate.
5301 return N0.getOperand(0);
5304 // Fold extract-and-trunc into a narrow extract. For example:
5305 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
5306 // i32 y = TRUNCATE(i64 x)
5308 // v16i8 b = BITCAST (v2i64 val)
5309 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
5311 // Note: We only run this optimization after type legalization (which often
5312 // creates this pattern) and before operation legalization after which
5313 // we need to be more careful about the vector instructions that we generate.
5314 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5315 LegalTypes && !LegalOperations && N0->hasOneUse()) {
5317 EVT VecTy = N0.getOperand(0).getValueType();
5318 EVT ExTy = N0.getValueType();
5319 EVT TrTy = N->getValueType(0);
5321 unsigned NumElem = VecTy.getVectorNumElements();
5322 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
5324 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
5325 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
5327 SDValue EltNo = N0->getOperand(1);
5328 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
5329 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5330 EVT IndexTy = N0->getOperand(1).getValueType();
5331 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
5333 SDValue V = DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
5334 NVT, N0.getOperand(0));
5336 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
5337 N->getDebugLoc(), TrTy, V,
5338 DAG.getConstant(Index, IndexTy));
5342 // See if we can simplify the input to this truncate through knowledge that
5343 // only the low bits are being used.
5344 // For example "trunc (or (shl x, 8), y)" // -> trunc y
5345 // Currently we only perform this optimization on scalars because vectors
5346 // may have different active low bits.
5347 if (!VT.isVector()) {
5349 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
5350 VT.getSizeInBits()));
5351 if (Shorter.getNode())
5352 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Shorter);
5354 // fold (truncate (load x)) -> (smaller load x)
5355 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
5356 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
5357 SDValue Reduced = ReduceLoadWidth(N);
5358 if (Reduced.getNode())
5361 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
5362 // where ... are all 'undef'.
5363 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
5364 SmallVector<EVT, 8> VTs;
5367 unsigned NumDefs = 0;
5369 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
5370 SDValue X = N0.getOperand(i);
5371 if (X.getOpcode() != ISD::UNDEF) {
5376 // Stop if more than one members are non-undef.
5379 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
5380 VT.getVectorElementType(),
5381 X.getValueType().getVectorNumElements()));
5385 return DAG.getUNDEF(VT);
5388 assert(V.getNode() && "The single defined operand is empty!");
5389 SmallVector<SDValue, 8> Opnds;
5390 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
5392 Opnds.push_back(DAG.getUNDEF(VTs[i]));
5395 SDValue NV = DAG.getNode(ISD::TRUNCATE, V.getDebugLoc(), VTs[i], V);
5396 AddToWorkList(NV.getNode());
5397 Opnds.push_back(NV);
5399 return DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
5400 &Opnds[0], Opnds.size());
5404 // Simplify the operands using demanded-bits information.
5405 if (!VT.isVector() &&
5406 SimplifyDemandedBits(SDValue(N, 0)))
5407 return SDValue(N, 0);
5412 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
5413 SDValue Elt = N->getOperand(i);
5414 if (Elt.getOpcode() != ISD::MERGE_VALUES)
5415 return Elt.getNode();
5416 return Elt.getOperand(Elt.getResNo()).getNode();
5419 /// CombineConsecutiveLoads - build_pair (load, load) -> load
5420 /// if load locations are consecutive.
5421 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
5422 assert(N->getOpcode() == ISD::BUILD_PAIR);
5424 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
5425 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
5426 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
5427 LD1->getPointerInfo().getAddrSpace() !=
5428 LD2->getPointerInfo().getAddrSpace())
5430 EVT LD1VT = LD1->getValueType(0);
5432 if (ISD::isNON_EXTLoad(LD2) &&
5434 // If both are volatile this would reduce the number of volatile loads.
5435 // If one is volatile it might be ok, but play conservative and bail out.
5436 !LD1->isVolatile() &&
5437 !LD2->isVolatile() &&
5438 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
5439 unsigned Align = LD1->getAlignment();
5440 unsigned NewAlign = TLI.getDataLayout()->
5441 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
5443 if (NewAlign <= Align &&
5444 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
5445 return DAG.getLoad(VT, N->getDebugLoc(), LD1->getChain(),
5446 LD1->getBasePtr(), LD1->getPointerInfo(),
5447 false, false, false, Align);
5453 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
5454 SDValue N0 = N->getOperand(0);
5455 EVT VT = N->getValueType(0);
5457 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
5458 // Only do this before legalize, since afterward the target may be depending
5459 // on the bitconvert.
5460 // First check to see if this is all constant.
5462 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
5464 bool isSimple = true;
5465 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i)
5466 if (N0.getOperand(i).getOpcode() != ISD::UNDEF &&
5467 N0.getOperand(i).getOpcode() != ISD::Constant &&
5468 N0.getOperand(i).getOpcode() != ISD::ConstantFP) {
5473 EVT DestEltVT = N->getValueType(0).getVectorElementType();
5474 assert(!DestEltVT.isVector() &&
5475 "Element type of vector ValueType must not be vector!");
5477 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
5480 // If the input is a constant, let getNode fold it.
5481 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
5482 SDValue Res = DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, N0);
5483 if (Res.getNode() != N) {
5484 if (!LegalOperations ||
5485 TLI.isOperationLegal(Res.getNode()->getOpcode(), VT))
5488 // Folding it resulted in an illegal node, and it's too late to
5489 // do that. Clean up the old node and forego the transformation.
5490 // Ideally this won't happen very often, because instcombine
5491 // and the earlier dagcombine runs (where illegal nodes are
5492 // permitted) should have folded most of them already.
5493 DAG.DeleteNode(Res.getNode());
5497 // (conv (conv x, t1), t2) -> (conv x, t2)
5498 if (N0.getOpcode() == ISD::BITCAST)
5499 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT,
5502 // fold (conv (load x)) -> (load (conv*)x)
5503 // If the resultant load doesn't need a higher alignment than the original!
5504 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
5505 // Do not change the width of a volatile load.
5506 !cast<LoadSDNode>(N0)->isVolatile() &&
5507 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT))) {
5508 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5509 unsigned Align = TLI.getDataLayout()->
5510 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
5511 unsigned OrigAlign = LN0->getAlignment();
5513 if (Align <= OrigAlign) {
5514 SDValue Load = DAG.getLoad(VT, N->getDebugLoc(), LN0->getChain(),
5515 LN0->getBasePtr(), LN0->getPointerInfo(),
5516 LN0->isVolatile(), LN0->isNonTemporal(),
5517 LN0->isInvariant(), OrigAlign);
5519 CombineTo(N0.getNode(),
5520 DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5521 N0.getValueType(), Load),
5527 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
5528 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
5529 // This often reduces constant pool loads.
5530 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(VT)) ||
5531 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(VT))) &&
5532 N0.getNode()->hasOneUse() && VT.isInteger() &&
5533 !VT.isVector() && !N0.getValueType().isVector()) {
5534 SDValue NewConv = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(), VT,
5536 AddToWorkList(NewConv.getNode());
5538 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
5539 if (N0.getOpcode() == ISD::FNEG)
5540 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
5541 NewConv, DAG.getConstant(SignBit, VT));
5542 assert(N0.getOpcode() == ISD::FABS);
5543 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
5544 NewConv, DAG.getConstant(~SignBit, VT));
5547 // fold (bitconvert (fcopysign cst, x)) ->
5548 // (or (and (bitconvert x), sign), (and cst, (not sign)))
5549 // Note that we don't handle (copysign x, cst) because this can always be
5550 // folded to an fneg or fabs.
5551 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
5552 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
5553 VT.isInteger() && !VT.isVector()) {
5554 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
5555 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
5556 if (isTypeLegal(IntXVT)) {
5557 SDValue X = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5558 IntXVT, N0.getOperand(1));
5559 AddToWorkList(X.getNode());
5561 // If X has a different width than the result/lhs, sext it or truncate it.
5562 unsigned VTWidth = VT.getSizeInBits();
5563 if (OrigXWidth < VTWidth) {
5564 X = DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, X);
5565 AddToWorkList(X.getNode());
5566 } else if (OrigXWidth > VTWidth) {
5567 // To get the sign bit in the right place, we have to shift it right
5568 // before truncating.
5569 X = DAG.getNode(ISD::SRL, X.getDebugLoc(),
5570 X.getValueType(), X,
5571 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
5572 AddToWorkList(X.getNode());
5573 X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
5574 AddToWorkList(X.getNode());
5577 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
5578 X = DAG.getNode(ISD::AND, X.getDebugLoc(), VT,
5579 X, DAG.getConstant(SignBit, VT));
5580 AddToWorkList(X.getNode());
5582 SDValue Cst = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5583 VT, N0.getOperand(0));
5584 Cst = DAG.getNode(ISD::AND, Cst.getDebugLoc(), VT,
5585 Cst, DAG.getConstant(~SignBit, VT));
5586 AddToWorkList(Cst.getNode());
5588 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, X, Cst);
5592 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
5593 if (N0.getOpcode() == ISD::BUILD_PAIR) {
5594 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
5595 if (CombineLD.getNode())
5602 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
5603 EVT VT = N->getValueType(0);
5604 return CombineConsecutiveLoads(N, VT);
5607 /// ConstantFoldBITCASTofBUILD_VECTOR - We know that BV is a build_vector
5608 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
5609 /// destination element value type.
5610 SDValue DAGCombiner::
5611 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
5612 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
5614 // If this is already the right type, we're done.
5615 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
5617 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
5618 unsigned DstBitSize = DstEltVT.getSizeInBits();
5620 // If this is a conversion of N elements of one type to N elements of another
5621 // type, convert each element. This handles FP<->INT cases.
5622 if (SrcBitSize == DstBitSize) {
5623 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
5624 BV->getValueType(0).getVectorNumElements());
5626 // Due to the FP element handling below calling this routine recursively,
5627 // we can end up with a scalar-to-vector node here.
5628 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
5629 return DAG.getNode(ISD::SCALAR_TO_VECTOR, BV->getDebugLoc(), VT,
5630 DAG.getNode(ISD::BITCAST, BV->getDebugLoc(),
5631 DstEltVT, BV->getOperand(0)));
5633 SmallVector<SDValue, 8> Ops;
5634 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
5635 SDValue Op = BV->getOperand(i);
5636 // If the vector element type is not legal, the BUILD_VECTOR operands
5637 // are promoted and implicitly truncated. Make that explicit here.
5638 if (Op.getValueType() != SrcEltVT)
5639 Op = DAG.getNode(ISD::TRUNCATE, BV->getDebugLoc(), SrcEltVT, Op);
5640 Ops.push_back(DAG.getNode(ISD::BITCAST, BV->getDebugLoc(),
5642 AddToWorkList(Ops.back().getNode());
5644 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5645 &Ops[0], Ops.size());
5648 // Otherwise, we're growing or shrinking the elements. To avoid having to
5649 // handle annoying details of growing/shrinking FP values, we convert them to
5651 if (SrcEltVT.isFloatingPoint()) {
5652 // Convert the input float vector to a int vector where the elements are the
5654 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
5655 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
5656 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
5660 // Now we know the input is an integer vector. If the output is a FP type,
5661 // convert to integer first, then to FP of the right size.
5662 if (DstEltVT.isFloatingPoint()) {
5663 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
5664 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
5665 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
5667 // Next, convert to FP elements of the same size.
5668 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
5671 // Okay, we know the src/dst types are both integers of differing types.
5672 // Handling growing first.
5673 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
5674 if (SrcBitSize < DstBitSize) {
5675 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
5677 SmallVector<SDValue, 8> Ops;
5678 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
5679 i += NumInputsPerOutput) {
5680 bool isLE = TLI.isLittleEndian();
5681 APInt NewBits = APInt(DstBitSize, 0);
5682 bool EltIsUndef = true;
5683 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
5684 // Shift the previously computed bits over.
5685 NewBits <<= SrcBitSize;
5686 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
5687 if (Op.getOpcode() == ISD::UNDEF) continue;
5690 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
5691 zextOrTrunc(SrcBitSize).zext(DstBitSize);
5695 Ops.push_back(DAG.getUNDEF(DstEltVT));
5697 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
5700 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
5701 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5702 &Ops[0], Ops.size());
5705 // Finally, this must be the case where we are shrinking elements: each input
5706 // turns into multiple outputs.
5707 bool isS2V = ISD::isScalarToVector(BV);
5708 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
5709 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
5710 NumOutputsPerInput*BV->getNumOperands());
5711 SmallVector<SDValue, 8> Ops;
5713 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
5714 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
5715 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
5716 Ops.push_back(DAG.getUNDEF(DstEltVT));
5720 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
5721 getAPIntValue().zextOrTrunc(SrcBitSize);
5723 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
5724 APInt ThisVal = OpVal.trunc(DstBitSize);
5725 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
5726 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
5727 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
5728 return DAG.getNode(ISD::SCALAR_TO_VECTOR, BV->getDebugLoc(), VT,
5730 OpVal = OpVal.lshr(DstBitSize);
5733 // For big endian targets, swap the order of the pieces of each element.
5734 if (TLI.isBigEndian())
5735 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
5738 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5739 &Ops[0], Ops.size());
5742 SDValue DAGCombiner::visitFADD(SDNode *N) {
5743 SDValue N0 = N->getOperand(0);
5744 SDValue N1 = N->getOperand(1);
5745 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5746 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5747 EVT VT = N->getValueType(0);
5750 if (VT.isVector()) {
5751 SDValue FoldedVOp = SimplifyVBinOp(N);
5752 if (FoldedVOp.getNode()) return FoldedVOp;
5755 // fold (fadd c1, c2) -> c1 + c2
5757 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N1);
5758 // canonicalize constant to RHS
5759 if (N0CFP && !N1CFP)
5760 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N1, N0);
5761 // fold (fadd A, 0) -> A
5762 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
5763 N1CFP->getValueAPF().isZero())
5765 // fold (fadd A, (fneg B)) -> (fsub A, B)
5766 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
5767 isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
5768 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N0,
5769 GetNegatedExpression(N1, DAG, LegalOperations));
5770 // fold (fadd (fneg A), B) -> (fsub B, A)
5771 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
5772 isNegatibleForFree(N0, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
5773 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N1,
5774 GetNegatedExpression(N0, DAG, LegalOperations));
5776 // If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
5777 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
5778 N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
5779 isa<ConstantFPSDNode>(N0.getOperand(1)))
5780 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0.getOperand(0),
5781 DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5782 N0.getOperand(1), N1));
5784 // If allow, fold (fadd (fneg x), x) -> 0.0
5785 if (DAG.getTarget().Options.UnsafeFPMath &&
5786 N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1) {
5787 return DAG.getConstantFP(0.0, VT);
5790 // If allow, fold (fadd x, (fneg x)) -> 0.0
5791 if (DAG.getTarget().Options.UnsafeFPMath &&
5792 N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0) {
5793 return DAG.getConstantFP(0.0, VT);
5796 // In unsafe math mode, we can fold chains of FADD's of the same value
5797 // into multiplications. This transform is not safe in general because
5798 // we are reducing the number of rounding steps.
5799 if (DAG.getTarget().Options.UnsafeFPMath &&
5800 TLI.isOperationLegalOrCustom(ISD::FMUL, VT) &&
5802 if (N0.getOpcode() == ISD::FMUL) {
5803 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
5804 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
5806 // (fadd (fmul c, x), x) -> (fmul c+1, x)
5807 if (CFP00 && !CFP01 && N0.getOperand(1) == N1) {
5808 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5810 DAG.getConstantFP(1.0, VT));
5811 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5815 // (fadd (fmul x, c), x) -> (fmul c+1, x)
5816 if (CFP01 && !CFP00 && N0.getOperand(0) == N1) {
5817 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5819 DAG.getConstantFP(1.0, VT));
5820 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5824 // (fadd (fadd x, x), x) -> (fmul 3.0, x)
5825 if (!CFP00 && !CFP01 && N0.getOperand(0) == N0.getOperand(1) &&
5826 N0.getOperand(0) == N1) {
5827 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5828 N1, DAG.getConstantFP(3.0, VT));
5831 // (fadd (fmul c, x), (fadd x, x)) -> (fmul c+2, x)
5832 if (CFP00 && !CFP01 && N1.getOpcode() == ISD::FADD &&
5833 N1.getOperand(0) == N1.getOperand(1) &&
5834 N0.getOperand(1) == N1.getOperand(0)) {
5835 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5837 DAG.getConstantFP(2.0, VT));
5838 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5839 N0.getOperand(1), NewCFP);
5842 // (fadd (fmul x, c), (fadd x, x)) -> (fmul c+2, x)
5843 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
5844 N1.getOperand(0) == N1.getOperand(1) &&
5845 N0.getOperand(0) == N1.getOperand(0)) {
5846 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5848 DAG.getConstantFP(2.0, VT));
5849 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5850 N0.getOperand(0), NewCFP);
5854 if (N1.getOpcode() == ISD::FMUL) {
5855 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
5856 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
5858 // (fadd x, (fmul c, x)) -> (fmul c+1, x)
5859 if (CFP10 && !CFP11 && N1.getOperand(1) == N0) {
5860 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5862 DAG.getConstantFP(1.0, VT));
5863 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5867 // (fadd x, (fmul x, c)) -> (fmul c+1, x)
5868 if (CFP11 && !CFP10 && N1.getOperand(0) == N0) {
5869 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5871 DAG.getConstantFP(1.0, VT));
5872 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5876 // (fadd x, (fadd x, x)) -> (fmul 3.0, x)
5877 if (!CFP10 && !CFP11 && N1.getOperand(0) == N1.getOperand(1) &&
5878 N1.getOperand(0) == N0) {
5879 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5880 N0, DAG.getConstantFP(3.0, VT));
5883 // (fadd (fadd x, x), (fmul c, x)) -> (fmul c+2, x)
5884 if (CFP10 && !CFP11 && N1.getOpcode() == ISD::FADD &&
5885 N1.getOperand(0) == N1.getOperand(1) &&
5886 N0.getOperand(1) == N1.getOperand(0)) {
5887 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5889 DAG.getConstantFP(2.0, VT));
5890 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5891 N0.getOperand(1), NewCFP);
5894 // (fadd (fadd x, x), (fmul x, c)) -> (fmul c+2, x)
5895 if (CFP11 && !CFP10 && N1.getOpcode() == ISD::FADD &&
5896 N1.getOperand(0) == N1.getOperand(1) &&
5897 N0.getOperand(0) == N1.getOperand(0)) {
5898 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5900 DAG.getConstantFP(2.0, VT));
5901 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5902 N0.getOperand(0), NewCFP);
5906 // (fadd (fadd x, x), (fadd x, x)) -> (fmul 4.0, x)
5907 if (N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
5908 N0.getOperand(0) == N0.getOperand(1) &&
5909 N1.getOperand(0) == N1.getOperand(1) &&
5910 N0.getOperand(0) == N1.getOperand(0)) {
5911 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5913 DAG.getConstantFP(4.0, VT));
5917 // FADD -> FMA combines:
5918 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
5919 DAG.getTarget().Options.UnsafeFPMath) &&
5920 DAG.getTarget().getTargetLowering()->isFMAFasterThanMulAndAdd(VT) &&
5921 TLI.isOperationLegalOrCustom(ISD::FMA, VT)) {
5923 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
5924 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse()) {
5925 return DAG.getNode(ISD::FMA, N->getDebugLoc(), VT,
5926 N0.getOperand(0), N0.getOperand(1), N1);
5929 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
5930 // Note: Commutes FADD operands.
5931 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse()) {
5932 return DAG.getNode(ISD::FMA, N->getDebugLoc(), VT,
5933 N1.getOperand(0), N1.getOperand(1), N0);
5940 SDValue DAGCombiner::visitFSUB(SDNode *N) {
5941 SDValue N0 = N->getOperand(0);
5942 SDValue N1 = N->getOperand(1);
5943 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5944 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5945 EVT VT = N->getValueType(0);
5946 DebugLoc dl = N->getDebugLoc();
5949 if (VT.isVector()) {
5950 SDValue FoldedVOp = SimplifyVBinOp(N);
5951 if (FoldedVOp.getNode()) return FoldedVOp;
5954 // fold (fsub c1, c2) -> c1-c2
5956 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N0, N1);
5957 // fold (fsub A, 0) -> A
5958 if (DAG.getTarget().Options.UnsafeFPMath &&
5959 N1CFP && N1CFP->getValueAPF().isZero())
5961 // fold (fsub 0, B) -> -B
5962 if (DAG.getTarget().Options.UnsafeFPMath &&
5963 N0CFP && N0CFP->getValueAPF().isZero()) {
5964 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
5965 return GetNegatedExpression(N1, DAG, LegalOperations);
5966 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
5967 return DAG.getNode(ISD::FNEG, dl, VT, N1);
5969 // fold (fsub A, (fneg B)) -> (fadd A, B)
5970 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
5971 return DAG.getNode(ISD::FADD, dl, VT, N0,
5972 GetNegatedExpression(N1, DAG, LegalOperations));
5974 // If 'unsafe math' is enabled, fold
5975 // (fsub x, x) -> 0.0 &
5976 // (fsub x, (fadd x, y)) -> (fneg y) &
5977 // (fsub x, (fadd y, x)) -> (fneg y)
5978 if (DAG.getTarget().Options.UnsafeFPMath) {
5980 return DAG.getConstantFP(0.0f, VT);
5982 if (N1.getOpcode() == ISD::FADD) {
5983 SDValue N10 = N1->getOperand(0);
5984 SDValue N11 = N1->getOperand(1);
5986 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI,
5987 &DAG.getTarget().Options))
5988 return GetNegatedExpression(N11, DAG, LegalOperations);
5989 else if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI,
5990 &DAG.getTarget().Options))
5991 return GetNegatedExpression(N10, DAG, LegalOperations);
5995 // FSUB -> FMA combines:
5996 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
5997 DAG.getTarget().Options.UnsafeFPMath) &&
5998 DAG.getTarget().getTargetLowering()->isFMAFasterThanMulAndAdd(VT) &&
5999 TLI.isOperationLegalOrCustom(ISD::FMA, VT)) {
6001 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
6002 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse()) {
6003 return DAG.getNode(ISD::FMA, dl, VT,
6004 N0.getOperand(0), N0.getOperand(1),
6005 DAG.getNode(ISD::FNEG, dl, VT, N1));
6008 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
6009 // Note: Commutes FSUB operands.
6010 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse()) {
6011 return DAG.getNode(ISD::FMA, dl, VT,
6012 DAG.getNode(ISD::FNEG, dl, VT,
6014 N1.getOperand(1), N0);
6017 // fold (fsub (-(fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
6018 if (N0.getOpcode() == ISD::FNEG &&
6019 N0.getOperand(0).getOpcode() == ISD::FMUL &&
6020 N0->hasOneUse() && N0.getOperand(0).hasOneUse()) {
6021 SDValue N00 = N0.getOperand(0).getOperand(0);
6022 SDValue N01 = N0.getOperand(0).getOperand(1);
6023 return DAG.getNode(ISD::FMA, dl, VT,
6024 DAG.getNode(ISD::FNEG, dl, VT, N00), N01,
6025 DAG.getNode(ISD::FNEG, dl, VT, N1));
6032 SDValue DAGCombiner::visitFMUL(SDNode *N) {
6033 SDValue N0 = N->getOperand(0);
6034 SDValue N1 = N->getOperand(1);
6035 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6036 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6037 EVT VT = N->getValueType(0);
6038 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6041 if (VT.isVector()) {
6042 SDValue FoldedVOp = SimplifyVBinOp(N);
6043 if (FoldedVOp.getNode()) return FoldedVOp;
6046 // fold (fmul c1, c2) -> c1*c2
6048 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0, N1);
6049 // canonicalize constant to RHS
6050 if (N0CFP && !N1CFP)
6051 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N1, N0);
6052 // fold (fmul A, 0) -> 0
6053 if (DAG.getTarget().Options.UnsafeFPMath &&
6054 N1CFP && N1CFP->getValueAPF().isZero())
6056 // fold (fmul A, 0) -> 0, vector edition.
6057 if (DAG.getTarget().Options.UnsafeFPMath &&
6058 ISD::isBuildVectorAllZeros(N1.getNode()))
6060 // fold (fmul A, 1.0) -> A
6061 if (N1CFP && N1CFP->isExactlyValue(1.0))
6063 // fold (fmul X, 2.0) -> (fadd X, X)
6064 if (N1CFP && N1CFP->isExactlyValue(+2.0))
6065 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N0);
6066 // fold (fmul X, -1.0) -> (fneg X)
6067 if (N1CFP && N1CFP->isExactlyValue(-1.0))
6068 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6069 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT, N0);
6071 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
6072 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
6073 &DAG.getTarget().Options)) {
6074 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
6075 &DAG.getTarget().Options)) {
6076 // Both can be negated for free, check to see if at least one is cheaper
6078 if (LHSNeg == 2 || RHSNeg == 2)
6079 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
6080 GetNegatedExpression(N0, DAG, LegalOperations),
6081 GetNegatedExpression(N1, DAG, LegalOperations));
6085 // If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
6086 if (DAG.getTarget().Options.UnsafeFPMath &&
6087 N1CFP && N0.getOpcode() == ISD::FMUL &&
6088 N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
6089 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0.getOperand(0),
6090 DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
6091 N0.getOperand(1), N1));
6096 SDValue DAGCombiner::visitFMA(SDNode *N) {
6097 SDValue N0 = N->getOperand(0);
6098 SDValue N1 = N->getOperand(1);
6099 SDValue N2 = N->getOperand(2);
6100 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6101 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6102 EVT VT = N->getValueType(0);
6103 DebugLoc dl = N->getDebugLoc();
6105 if (DAG.getTarget().Options.UnsafeFPMath) {
6106 if (N0CFP && N0CFP->isZero())
6108 if (N1CFP && N1CFP->isZero())
6111 if (N0CFP && N0CFP->isExactlyValue(1.0))
6112 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N1, N2);
6113 if (N1CFP && N1CFP->isExactlyValue(1.0))
6114 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N2);
6116 // Canonicalize (fma c, x, y) -> (fma x, c, y)
6117 if (N0CFP && !N1CFP)
6118 return DAG.getNode(ISD::FMA, N->getDebugLoc(), VT, N1, N0, N2);
6120 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
6121 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6122 N2.getOpcode() == ISD::FMUL &&
6123 N0 == N2.getOperand(0) &&
6124 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
6125 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6126 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
6130 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
6131 if (DAG.getTarget().Options.UnsafeFPMath &&
6132 N0.getOpcode() == ISD::FMUL && N1CFP &&
6133 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
6134 return DAG.getNode(ISD::FMA, dl, VT,
6136 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
6140 // (fma x, 1, y) -> (fadd x, y)
6141 // (fma x, -1, y) -> (fadd (fneg x), y)
6143 if (N1CFP->isExactlyValue(1.0))
6144 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
6146 if (N1CFP->isExactlyValue(-1.0) &&
6147 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
6148 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
6149 AddToWorkList(RHSNeg.getNode());
6150 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
6154 // (fma x, c, x) -> (fmul x, (c+1))
6155 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP && N0 == N2) {
6156 return DAG.getNode(ISD::FMUL, dl, VT,
6158 DAG.getNode(ISD::FADD, dl, VT,
6159 N1, DAG.getConstantFP(1.0, VT)));
6162 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
6163 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6164 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0) {
6165 return DAG.getNode(ISD::FMUL, dl, VT,
6167 DAG.getNode(ISD::FADD, dl, VT,
6168 N1, DAG.getConstantFP(-1.0, VT)));
6175 SDValue DAGCombiner::visitFDIV(SDNode *N) {
6176 SDValue N0 = N->getOperand(0);
6177 SDValue N1 = N->getOperand(1);
6178 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6179 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6180 EVT VT = N->getValueType(0);
6181 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6184 if (VT.isVector()) {
6185 SDValue FoldedVOp = SimplifyVBinOp(N);
6186 if (FoldedVOp.getNode()) return FoldedVOp;
6189 // fold (fdiv c1, c2) -> c1/c2
6191 return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT, N0, N1);
6193 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
6194 if (N1CFP && DAG.getTarget().Options.UnsafeFPMath) {
6195 // Compute the reciprocal 1.0 / c2.
6196 APFloat N1APF = N1CFP->getValueAPF();
6197 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
6198 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
6199 // Only do the transform if the reciprocal is a legal fp immediate that
6200 // isn't too nasty (eg NaN, denormal, ...).
6201 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
6202 (!LegalOperations ||
6203 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
6204 // backend)... we should handle this gracefully after Legalize.
6205 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
6206 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
6207 TLI.isFPImmLegal(Recip, VT)))
6208 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0,
6209 DAG.getConstantFP(Recip, VT));
6212 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
6213 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
6214 &DAG.getTarget().Options)) {
6215 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
6216 &DAG.getTarget().Options)) {
6217 // Both can be negated for free, check to see if at least one is cheaper
6219 if (LHSNeg == 2 || RHSNeg == 2)
6220 return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT,
6221 GetNegatedExpression(N0, DAG, LegalOperations),
6222 GetNegatedExpression(N1, DAG, LegalOperations));
6229 SDValue DAGCombiner::visitFREM(SDNode *N) {
6230 SDValue N0 = N->getOperand(0);
6231 SDValue N1 = N->getOperand(1);
6232 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6233 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6234 EVT VT = N->getValueType(0);
6236 // fold (frem c1, c2) -> fmod(c1,c2)
6238 return DAG.getNode(ISD::FREM, N->getDebugLoc(), VT, N0, N1);
6243 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
6244 SDValue N0 = N->getOperand(0);
6245 SDValue N1 = N->getOperand(1);
6246 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6247 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6248 EVT VT = N->getValueType(0);
6250 if (N0CFP && N1CFP) // Constant fold
6251 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT, N0, N1);
6254 const APFloat& V = N1CFP->getValueAPF();
6255 // copysign(x, c1) -> fabs(x) iff ispos(c1)
6256 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
6257 if (!V.isNegative()) {
6258 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
6259 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
6261 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6262 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT,
6263 DAG.getNode(ISD::FABS, N0.getDebugLoc(), VT, N0));
6267 // copysign(fabs(x), y) -> copysign(x, y)
6268 // copysign(fneg(x), y) -> copysign(x, y)
6269 // copysign(copysign(x,z), y) -> copysign(x, y)
6270 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
6271 N0.getOpcode() == ISD::FCOPYSIGN)
6272 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
6273 N0.getOperand(0), N1);
6275 // copysign(x, abs(y)) -> abs(x)
6276 if (N1.getOpcode() == ISD::FABS)
6277 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
6279 // copysign(x, copysign(y,z)) -> copysign(x, z)
6280 if (N1.getOpcode() == ISD::FCOPYSIGN)
6281 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
6282 N0, N1.getOperand(1));
6284 // copysign(x, fp_extend(y)) -> copysign(x, y)
6285 // copysign(x, fp_round(y)) -> copysign(x, y)
6286 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
6287 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
6288 N0, N1.getOperand(0));
6293 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
6294 SDValue N0 = N->getOperand(0);
6295 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
6296 EVT VT = N->getValueType(0);
6297 EVT OpVT = N0.getValueType();
6299 // fold (sint_to_fp c1) -> c1fp
6301 // ...but only if the target supports immediate floating-point values
6302 (!LegalOperations ||
6303 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
6304 return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
6306 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
6307 // but UINT_TO_FP is legal on this target, try to convert.
6308 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
6309 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
6310 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
6311 if (DAG.SignBitIsZero(N0))
6312 return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
6315 // The next optimizations are desireable only if SELECT_CC can be lowered.
6316 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
6317 // having to say they don't support SELECT_CC on every type the DAG knows
6318 // about, since there is no way to mark an opcode illegal at all value types
6319 // (See also visitSELECT)
6320 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other)) {
6321 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
6322 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
6324 (!LegalOperations ||
6325 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
6327 { N0.getOperand(0), N0.getOperand(1),
6328 DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT),
6330 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT, Ops, 5);
6333 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
6334 // (select_cc x, y, 1.0, 0.0,, cc)
6335 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
6336 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
6337 (!LegalOperations ||
6338 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
6340 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
6341 DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT),
6342 N0.getOperand(0).getOperand(2) };
6343 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT, Ops, 5);
6350 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
6351 SDValue N0 = N->getOperand(0);
6352 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
6353 EVT VT = N->getValueType(0);
6354 EVT OpVT = N0.getValueType();
6356 // fold (uint_to_fp c1) -> c1fp
6358 // ...but only if the target supports immediate floating-point values
6359 (!LegalOperations ||
6360 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
6361 return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
6363 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
6364 // but SINT_TO_FP is legal on this target, try to convert.
6365 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
6366 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
6367 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
6368 if (DAG.SignBitIsZero(N0))
6369 return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
6372 // The next optimizations are desireable only if SELECT_CC can be lowered.
6373 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
6374 // having to say they don't support SELECT_CC on every type the DAG knows
6375 // about, since there is no way to mark an opcode illegal at all value types
6376 // (See also visitSELECT)
6377 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other)) {
6378 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
6380 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
6381 (!LegalOperations ||
6382 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
6384 { N0.getOperand(0), N0.getOperand(1),
6385 DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT),
6387 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT, Ops, 5);
6394 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
6395 SDValue N0 = N->getOperand(0);
6396 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6397 EVT VT = N->getValueType(0);
6399 // fold (fp_to_sint c1fp) -> c1
6401 return DAG.getNode(ISD::FP_TO_SINT, N->getDebugLoc(), VT, N0);
6406 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
6407 SDValue N0 = N->getOperand(0);
6408 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6409 EVT VT = N->getValueType(0);
6411 // fold (fp_to_uint c1fp) -> c1
6413 return DAG.getNode(ISD::FP_TO_UINT, N->getDebugLoc(), VT, N0);
6418 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
6419 SDValue N0 = N->getOperand(0);
6420 SDValue N1 = N->getOperand(1);
6421 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6422 EVT VT = N->getValueType(0);
6424 // fold (fp_round c1fp) -> c1fp
6426 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT, N0, N1);
6428 // fold (fp_round (fp_extend x)) -> x
6429 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
6430 return N0.getOperand(0);
6432 // fold (fp_round (fp_round x)) -> (fp_round x)
6433 if (N0.getOpcode() == ISD::FP_ROUND) {
6434 // This is a value preserving truncation if both round's are.
6435 bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
6436 N0.getNode()->getConstantOperandVal(1) == 1;
6437 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT, N0.getOperand(0),
6438 DAG.getIntPtrConstant(IsTrunc));
6441 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
6442 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
6443 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(), VT,
6444 N0.getOperand(0), N1);
6445 AddToWorkList(Tmp.getNode());
6446 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
6447 Tmp, N0.getOperand(1));
6453 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
6454 SDValue N0 = N->getOperand(0);
6455 EVT VT = N->getValueType(0);
6456 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
6457 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6459 // fold (fp_round_inreg c1fp) -> c1fp
6460 if (N0CFP && isTypeLegal(EVT)) {
6461 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
6462 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, Round);
6468 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
6469 SDValue N0 = N->getOperand(0);
6470 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6471 EVT VT = N->getValueType(0);
6473 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
6474 if (N->hasOneUse() &&
6475 N->use_begin()->getOpcode() == ISD::FP_ROUND)
6478 // fold (fp_extend c1fp) -> c1fp
6480 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, N0);
6482 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
6484 if (N0.getOpcode() == ISD::FP_ROUND
6485 && N0.getNode()->getConstantOperandVal(1) == 1) {
6486 SDValue In = N0.getOperand(0);
6487 if (In.getValueType() == VT) return In;
6488 if (VT.bitsLT(In.getValueType()))
6489 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT,
6490 In, N0.getOperand(1));
6491 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, In);
6494 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
6495 if (ISD::isNON_EXTLoad(N0.getNode()) && N0.hasOneUse() &&
6496 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6497 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
6498 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6499 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
6501 LN0->getBasePtr(), LN0->getPointerInfo(),
6503 LN0->isVolatile(), LN0->isNonTemporal(),
6504 LN0->getAlignment());
6505 CombineTo(N, ExtLoad);
6506 CombineTo(N0.getNode(),
6507 DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(),
6508 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
6509 ExtLoad.getValue(1));
6510 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6516 SDValue DAGCombiner::visitFNEG(SDNode *N) {
6517 SDValue N0 = N->getOperand(0);
6518 EVT VT = N->getValueType(0);
6520 if (VT.isVector()) {
6521 SDValue FoldedVOp = SimplifyVUnaryOp(N);
6522 if (FoldedVOp.getNode()) return FoldedVOp;
6525 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
6526 &DAG.getTarget().Options))
6527 return GetNegatedExpression(N0, DAG, LegalOperations);
6529 // Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading
6530 // constant pool values.
6531 if (!TLI.isFNegFree(VT) && N0.getOpcode() == ISD::BITCAST &&
6533 N0.getNode()->hasOneUse() &&
6534 N0.getOperand(0).getValueType().isInteger()) {
6535 SDValue Int = N0.getOperand(0);
6536 EVT IntVT = Int.getValueType();
6537 if (IntVT.isInteger() && !IntVT.isVector()) {
6538 Int = DAG.getNode(ISD::XOR, N0.getDebugLoc(), IntVT, Int,
6539 DAG.getConstant(APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
6540 AddToWorkList(Int.getNode());
6541 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
6546 // (fneg (fmul c, x)) -> (fmul -c, x)
6547 if (N0.getOpcode() == ISD::FMUL) {
6548 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
6550 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
6552 DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT,
6560 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
6561 SDValue N0 = N->getOperand(0);
6562 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6563 EVT VT = N->getValueType(0);
6565 // fold (fceil c1) -> fceil(c1)
6567 return DAG.getNode(ISD::FCEIL, N->getDebugLoc(), VT, N0);
6572 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
6573 SDValue N0 = N->getOperand(0);
6574 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6575 EVT VT = N->getValueType(0);
6577 // fold (ftrunc c1) -> ftrunc(c1)
6579 return DAG.getNode(ISD::FTRUNC, N->getDebugLoc(), VT, N0);
6584 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
6585 SDValue N0 = N->getOperand(0);
6586 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6587 EVT VT = N->getValueType(0);
6589 // fold (ffloor c1) -> ffloor(c1)
6591 return DAG.getNode(ISD::FFLOOR, N->getDebugLoc(), VT, N0);
6596 SDValue DAGCombiner::visitFABS(SDNode *N) {
6597 SDValue N0 = N->getOperand(0);
6598 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6599 EVT VT = N->getValueType(0);
6601 if (VT.isVector()) {
6602 SDValue FoldedVOp = SimplifyVUnaryOp(N);
6603 if (FoldedVOp.getNode()) return FoldedVOp;
6606 // fold (fabs c1) -> fabs(c1)
6608 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
6609 // fold (fabs (fabs x)) -> (fabs x)
6610 if (N0.getOpcode() == ISD::FABS)
6611 return N->getOperand(0);
6612 // fold (fabs (fneg x)) -> (fabs x)
6613 // fold (fabs (fcopysign x, y)) -> (fabs x)
6614 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
6615 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0.getOperand(0));
6617 // Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading
6618 // constant pool values.
6619 if (!TLI.isFAbsFree(VT) &&
6620 N0.getOpcode() == ISD::BITCAST && N0.getNode()->hasOneUse() &&
6621 N0.getOperand(0).getValueType().isInteger() &&
6622 !N0.getOperand(0).getValueType().isVector()) {
6623 SDValue Int = N0.getOperand(0);
6624 EVT IntVT = Int.getValueType();
6625 if (IntVT.isInteger() && !IntVT.isVector()) {
6626 Int = DAG.getNode(ISD::AND, N0.getDebugLoc(), IntVT, Int,
6627 DAG.getConstant(~APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
6628 AddToWorkList(Int.getNode());
6629 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
6630 N->getValueType(0), Int);
6637 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
6638 SDValue Chain = N->getOperand(0);
6639 SDValue N1 = N->getOperand(1);
6640 SDValue N2 = N->getOperand(2);
6642 // If N is a constant we could fold this into a fallthrough or unconditional
6643 // branch. However that doesn't happen very often in normal code, because
6644 // Instcombine/SimplifyCFG should have handled the available opportunities.
6645 // If we did this folding here, it would be necessary to update the
6646 // MachineBasicBlock CFG, which is awkward.
6648 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
6650 if (N1.getOpcode() == ISD::SETCC &&
6651 TLI.isOperationLegalOrCustom(ISD::BR_CC, MVT::Other)) {
6652 return DAG.getNode(ISD::BR_CC, N->getDebugLoc(), MVT::Other,
6653 Chain, N1.getOperand(2),
6654 N1.getOperand(0), N1.getOperand(1), N2);
6657 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
6658 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
6659 (N1.getOperand(0).hasOneUse() &&
6660 N1.getOperand(0).getOpcode() == ISD::SRL))) {
6662 if (N1.getOpcode() == ISD::TRUNCATE) {
6663 // Look pass the truncate.
6664 Trunc = N1.getNode();
6665 N1 = N1.getOperand(0);
6668 // Match this pattern so that we can generate simpler code:
6671 // %b = and i32 %a, 2
6672 // %c = srl i32 %b, 1
6673 // brcond i32 %c ...
6678 // %b = and i32 %a, 2
6679 // %c = setcc eq %b, 0
6682 // This applies only when the AND constant value has one bit set and the
6683 // SRL constant is equal to the log2 of the AND constant. The back-end is
6684 // smart enough to convert the result into a TEST/JMP sequence.
6685 SDValue Op0 = N1.getOperand(0);
6686 SDValue Op1 = N1.getOperand(1);
6688 if (Op0.getOpcode() == ISD::AND &&
6689 Op1.getOpcode() == ISD::Constant) {
6690 SDValue AndOp1 = Op0.getOperand(1);
6692 if (AndOp1.getOpcode() == ISD::Constant) {
6693 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
6695 if (AndConst.isPowerOf2() &&
6696 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
6698 DAG.getSetCC(N->getDebugLoc(),
6699 TLI.getSetCCResultType(Op0.getValueType()),
6700 Op0, DAG.getConstant(0, Op0.getValueType()),
6703 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
6704 MVT::Other, Chain, SetCC, N2);
6705 // Don't add the new BRCond into the worklist or else SimplifySelectCC
6706 // will convert it back to (X & C1) >> C2.
6707 CombineTo(N, NewBRCond, false);
6708 // Truncate is dead.
6710 removeFromWorkList(Trunc);
6711 DAG.DeleteNode(Trunc);
6713 // Replace the uses of SRL with SETCC
6714 WorkListRemover DeadNodes(*this);
6715 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
6716 removeFromWorkList(N1.getNode());
6717 DAG.DeleteNode(N1.getNode());
6718 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6724 // Restore N1 if the above transformation doesn't match.
6725 N1 = N->getOperand(1);
6728 // Transform br(xor(x, y)) -> br(x != y)
6729 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
6730 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
6731 SDNode *TheXor = N1.getNode();
6732 SDValue Op0 = TheXor->getOperand(0);
6733 SDValue Op1 = TheXor->getOperand(1);
6734 if (Op0.getOpcode() == Op1.getOpcode()) {
6735 // Avoid missing important xor optimizations.
6736 SDValue Tmp = visitXOR(TheXor);
6737 if (Tmp.getNode() && Tmp.getNode() != TheXor) {
6738 DEBUG(dbgs() << "\nReplacing.8 ";
6740 dbgs() << "\nWith: ";
6741 Tmp.getNode()->dump(&DAG);
6743 WorkListRemover DeadNodes(*this);
6744 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
6745 removeFromWorkList(TheXor);
6746 DAG.DeleteNode(TheXor);
6747 return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
6748 MVT::Other, Chain, Tmp, N2);
6752 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
6754 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
6755 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
6756 Op0.getOpcode() == ISD::XOR) {
6757 TheXor = Op0.getNode();
6761 EVT SetCCVT = N1.getValueType();
6763 SetCCVT = TLI.getSetCCResultType(SetCCVT);
6764 SDValue SetCC = DAG.getSetCC(TheXor->getDebugLoc(),
6767 Equal ? ISD::SETEQ : ISD::SETNE);
6768 // Replace the uses of XOR with SETCC
6769 WorkListRemover DeadNodes(*this);
6770 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
6771 removeFromWorkList(N1.getNode());
6772 DAG.DeleteNode(N1.getNode());
6773 return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
6774 MVT::Other, Chain, SetCC, N2);
6781 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
6783 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
6784 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
6785 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
6787 // If N is a constant we could fold this into a fallthrough or unconditional
6788 // branch. However that doesn't happen very often in normal code, because
6789 // Instcombine/SimplifyCFG should have handled the available opportunities.
6790 // If we did this folding here, it would be necessary to update the
6791 // MachineBasicBlock CFG, which is awkward.
6793 // Use SimplifySetCC to simplify SETCC's.
6794 SDValue Simp = SimplifySetCC(TLI.getSetCCResultType(CondLHS.getValueType()),
6795 CondLHS, CondRHS, CC->get(), N->getDebugLoc(),
6797 if (Simp.getNode()) AddToWorkList(Simp.getNode());
6799 // fold to a simpler setcc
6800 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
6801 return DAG.getNode(ISD::BR_CC, N->getDebugLoc(), MVT::Other,
6802 N->getOperand(0), Simp.getOperand(2),
6803 Simp.getOperand(0), Simp.getOperand(1),
6809 /// canFoldInAddressingMode - Return true if 'Use' is a load or a store that
6810 /// uses N as its base pointer and that N may be folded in the load / store
6811 /// addressing mode.
6812 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
6814 const TargetLowering &TLI) {
6816 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
6817 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
6819 VT = Use->getValueType(0);
6820 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
6821 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
6823 VT = ST->getValue().getValueType();
6828 if (N->getOpcode() == ISD::ADD) {
6829 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
6832 AM.BaseOffs = Offset->getSExtValue();
6836 } else if (N->getOpcode() == ISD::SUB) {
6837 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
6840 AM.BaseOffs = -Offset->getSExtValue();
6847 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
6850 /// CombineToPreIndexedLoadStore - Try turning a load / store into a
6851 /// pre-indexed load / store when the base pointer is an add or subtract
6852 /// and it has other uses besides the load / store. After the
6853 /// transformation, the new indexed load / store has effectively folded
6854 /// the add / subtract in and all of its other uses are redirected to the
6855 /// new load / store.
6856 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
6857 if (Level < AfterLegalizeDAG)
6863 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6864 if (LD->isIndexed())
6866 VT = LD->getMemoryVT();
6867 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
6868 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
6870 Ptr = LD->getBasePtr();
6871 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6872 if (ST->isIndexed())
6874 VT = ST->getMemoryVT();
6875 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
6876 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
6878 Ptr = ST->getBasePtr();
6884 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
6885 // out. There is no reason to make this a preinc/predec.
6886 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
6887 Ptr.getNode()->hasOneUse())
6890 // Ask the target to do addressing mode selection.
6893 ISD::MemIndexedMode AM = ISD::UNINDEXED;
6894 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
6896 // Don't create a indexed load / store with zero offset.
6897 if (isa<ConstantSDNode>(Offset) &&
6898 cast<ConstantSDNode>(Offset)->isNullValue())
6901 // Try turning it into a pre-indexed load / store except when:
6902 // 1) The new base ptr is a frame index.
6903 // 2) If N is a store and the new base ptr is either the same as or is a
6904 // predecessor of the value being stored.
6905 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
6906 // that would create a cycle.
6907 // 4) All uses are load / store ops that use it as old base ptr.
6909 // Check #1. Preinc'ing a frame index would require copying the stack pointer
6910 // (plus the implicit offset) to a register to preinc anyway.
6911 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
6916 SDValue Val = cast<StoreSDNode>(N)->getValue();
6917 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
6921 // Now check for #3 and #4.
6922 bool RealUse = false;
6924 // Caches for hasPredecessorHelper
6925 SmallPtrSet<const SDNode *, 32> Visited;
6926 SmallVector<const SDNode *, 16> Worklist;
6928 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
6929 E = Ptr.getNode()->use_end(); I != E; ++I) {
6933 if (N->hasPredecessorHelper(Use, Visited, Worklist))
6936 // If Ptr may be folded in addressing mode of other use, then it's
6937 // not profitable to do this transformation.
6938 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
6947 Result = DAG.getIndexedLoad(SDValue(N,0), N->getDebugLoc(),
6948 BasePtr, Offset, AM);
6950 Result = DAG.getIndexedStore(SDValue(N,0), N->getDebugLoc(),
6951 BasePtr, Offset, AM);
6954 DEBUG(dbgs() << "\nReplacing.4 ";
6956 dbgs() << "\nWith: ";
6957 Result.getNode()->dump(&DAG);
6959 WorkListRemover DeadNodes(*this);
6961 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
6962 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
6964 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
6967 // Finally, since the node is now dead, remove it from the graph.
6970 // Replace the uses of Ptr with uses of the updated base value.
6971 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
6972 removeFromWorkList(Ptr.getNode());
6973 DAG.DeleteNode(Ptr.getNode());
6978 /// CombineToPostIndexedLoadStore - Try to combine a load / store with a
6979 /// add / sub of the base pointer node into a post-indexed load / store.
6980 /// The transformation folded the add / subtract into the new indexed
6981 /// load / store effectively and all of its uses are redirected to the
6982 /// new load / store.
6983 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
6984 if (Level < AfterLegalizeDAG)
6990 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6991 if (LD->isIndexed())
6993 VT = LD->getMemoryVT();
6994 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
6995 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
6997 Ptr = LD->getBasePtr();
6998 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6999 if (ST->isIndexed())
7001 VT = ST->getMemoryVT();
7002 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
7003 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
7005 Ptr = ST->getBasePtr();
7011 if (Ptr.getNode()->hasOneUse())
7014 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
7015 E = Ptr.getNode()->use_end(); I != E; ++I) {
7018 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
7023 ISD::MemIndexedMode AM = ISD::UNINDEXED;
7024 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
7025 // Don't create a indexed load / store with zero offset.
7026 if (isa<ConstantSDNode>(Offset) &&
7027 cast<ConstantSDNode>(Offset)->isNullValue())
7030 // Try turning it into a post-indexed load / store except when
7031 // 1) All uses are load / store ops that use it as base ptr (and
7032 // it may be folded as addressing mmode).
7033 // 2) Op must be independent of N, i.e. Op is neither a predecessor
7034 // nor a successor of N. Otherwise, if Op is folded that would
7037 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
7041 bool TryNext = false;
7042 for (SDNode::use_iterator II = BasePtr.getNode()->use_begin(),
7043 EE = BasePtr.getNode()->use_end(); II != EE; ++II) {
7045 if (Use == Ptr.getNode())
7048 // If all the uses are load / store addresses, then don't do the
7050 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
7051 bool RealUse = false;
7052 for (SDNode::use_iterator III = Use->use_begin(),
7053 EEE = Use->use_end(); III != EEE; ++III) {
7054 SDNode *UseUse = *III;
7055 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
7070 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
7071 SDValue Result = isLoad
7072 ? DAG.getIndexedLoad(SDValue(N,0), N->getDebugLoc(),
7073 BasePtr, Offset, AM)
7074 : DAG.getIndexedStore(SDValue(N,0), N->getDebugLoc(),
7075 BasePtr, Offset, AM);
7078 DEBUG(dbgs() << "\nReplacing.5 ";
7080 dbgs() << "\nWith: ";
7081 Result.getNode()->dump(&DAG);
7083 WorkListRemover DeadNodes(*this);
7085 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
7086 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
7088 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
7091 // Finally, since the node is now dead, remove it from the graph.
7094 // Replace the uses of Use with uses of the updated base value.
7095 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
7096 Result.getValue(isLoad ? 1 : 0));
7097 removeFromWorkList(Op);
7107 SDValue DAGCombiner::visitLOAD(SDNode *N) {
7108 LoadSDNode *LD = cast<LoadSDNode>(N);
7109 SDValue Chain = LD->getChain();
7110 SDValue Ptr = LD->getBasePtr();
7112 // If load is not volatile and there are no uses of the loaded value (and
7113 // the updated indexed value in case of indexed loads), change uses of the
7114 // chain value into uses of the chain input (i.e. delete the dead load).
7115 if (!LD->isVolatile()) {
7116 if (N->getValueType(1) == MVT::Other) {
7118 if (!N->hasAnyUseOfValue(0)) {
7119 // It's not safe to use the two value CombineTo variant here. e.g.
7120 // v1, chain2 = load chain1, loc
7121 // v2, chain3 = load chain2, loc
7123 // Now we replace use of chain2 with chain1. This makes the second load
7124 // isomorphic to the one we are deleting, and thus makes this load live.
7125 DEBUG(dbgs() << "\nReplacing.6 ";
7127 dbgs() << "\nWith chain: ";
7128 Chain.getNode()->dump(&DAG);
7130 WorkListRemover DeadNodes(*this);
7131 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
7133 if (N->use_empty()) {
7134 removeFromWorkList(N);
7138 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7142 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
7143 if (!N->hasAnyUseOfValue(0) && !N->hasAnyUseOfValue(1)) {
7144 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
7145 DEBUG(dbgs() << "\nReplacing.7 ";
7147 dbgs() << "\nWith: ";
7148 Undef.getNode()->dump(&DAG);
7149 dbgs() << " and 2 other values\n");
7150 WorkListRemover DeadNodes(*this);
7151 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
7152 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1),
7153 DAG.getUNDEF(N->getValueType(1)));
7154 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
7155 removeFromWorkList(N);
7157 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7162 // If this load is directly stored, replace the load value with the stored
7164 // TODO: Handle store large -> read small portion.
7165 // TODO: Handle TRUNCSTORE/LOADEXT
7166 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
7167 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
7168 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
7169 if (PrevST->getBasePtr() == Ptr &&
7170 PrevST->getValue().getValueType() == N->getValueType(0))
7171 return CombineTo(N, Chain.getOperand(1), Chain);
7175 // Try to infer better alignment information than the load already has.
7176 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
7177 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
7178 if (Align > LD->getAlignment())
7179 return DAG.getExtLoad(LD->getExtensionType(), N->getDebugLoc(),
7180 LD->getValueType(0),
7181 Chain, Ptr, LD->getPointerInfo(),
7183 LD->isVolatile(), LD->isNonTemporal(), Align);
7188 // Walk up chain skipping non-aliasing memory nodes.
7189 SDValue BetterChain = FindBetterChain(N, Chain);
7191 // If there is a better chain.
7192 if (Chain != BetterChain) {
7195 // Replace the chain to void dependency.
7196 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
7197 ReplLoad = DAG.getLoad(N->getValueType(0), LD->getDebugLoc(),
7198 BetterChain, Ptr, LD->getPointerInfo(),
7199 LD->isVolatile(), LD->isNonTemporal(),
7200 LD->isInvariant(), LD->getAlignment());
7202 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), LD->getDebugLoc(),
7203 LD->getValueType(0),
7204 BetterChain, Ptr, LD->getPointerInfo(),
7207 LD->isNonTemporal(),
7208 LD->getAlignment());
7211 // Create token factor to keep old chain connected.
7212 SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
7213 MVT::Other, Chain, ReplLoad.getValue(1));
7215 // Make sure the new and old chains are cleaned up.
7216 AddToWorkList(Token.getNode());
7218 // Replace uses with load result and token factor. Don't add users
7220 return CombineTo(N, ReplLoad.getValue(0), Token, false);
7224 // Try transforming N to an indexed load.
7225 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
7226 return SDValue(N, 0);
7231 /// CheckForMaskedLoad - Check to see if V is (and load (ptr), imm), where the
7232 /// load is having specific bytes cleared out. If so, return the byte size
7233 /// being masked out and the shift amount.
7234 static std::pair<unsigned, unsigned>
7235 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
7236 std::pair<unsigned, unsigned> Result(0, 0);
7238 // Check for the structure we're looking for.
7239 if (V->getOpcode() != ISD::AND ||
7240 !isa<ConstantSDNode>(V->getOperand(1)) ||
7241 !ISD::isNormalLoad(V->getOperand(0).getNode()))
7244 // Check the chain and pointer.
7245 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
7246 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
7248 // The store should be chained directly to the load or be an operand of a
7250 if (LD == Chain.getNode())
7252 else if (Chain->getOpcode() != ISD::TokenFactor)
7253 return Result; // Fail.
7256 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
7257 if (Chain->getOperand(i).getNode() == LD) {
7261 if (!isOk) return Result;
7264 // This only handles simple types.
7265 if (V.getValueType() != MVT::i16 &&
7266 V.getValueType() != MVT::i32 &&
7267 V.getValueType() != MVT::i64)
7270 // Check the constant mask. Invert it so that the bits being masked out are
7271 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
7272 // follow the sign bit for uniformity.
7273 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
7274 unsigned NotMaskLZ = CountLeadingZeros_64(NotMask);
7275 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
7276 unsigned NotMaskTZ = CountTrailingZeros_64(NotMask);
7277 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
7278 if (NotMaskLZ == 64) return Result; // All zero mask.
7280 // See if we have a continuous run of bits. If so, we have 0*1+0*
7281 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
7284 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
7285 if (V.getValueType() != MVT::i64 && NotMaskLZ)
7286 NotMaskLZ -= 64-V.getValueSizeInBits();
7288 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
7289 switch (MaskedBytes) {
7293 default: return Result; // All one mask, or 5-byte mask.
7296 // Verify that the first bit starts at a multiple of mask so that the access
7297 // is aligned the same as the access width.
7298 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
7300 Result.first = MaskedBytes;
7301 Result.second = NotMaskTZ/8;
7306 /// ShrinkLoadReplaceStoreWithStore - Check to see if IVal is something that
7307 /// provides a value as specified by MaskInfo. If so, replace the specified
7308 /// store with a narrower store of truncated IVal.
7310 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
7311 SDValue IVal, StoreSDNode *St,
7313 unsigned NumBytes = MaskInfo.first;
7314 unsigned ByteShift = MaskInfo.second;
7315 SelectionDAG &DAG = DC->getDAG();
7317 // Check to see if IVal is all zeros in the part being masked in by the 'or'
7318 // that uses this. If not, this is not a replacement.
7319 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
7320 ByteShift*8, (ByteShift+NumBytes)*8);
7321 if (!DAG.MaskedValueIsZero(IVal, Mask)) return 0;
7323 // Check that it is legal on the target to do this. It is legal if the new
7324 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
7326 MVT VT = MVT::getIntegerVT(NumBytes*8);
7327 if (!DC->isTypeLegal(VT))
7330 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
7331 // shifted by ByteShift and truncated down to NumBytes.
7333 IVal = DAG.getNode(ISD::SRL, IVal->getDebugLoc(), IVal.getValueType(), IVal,
7334 DAG.getConstant(ByteShift*8,
7335 DC->getShiftAmountTy(IVal.getValueType())));
7337 // Figure out the offset for the store and the alignment of the access.
7339 unsigned NewAlign = St->getAlignment();
7341 if (DAG.getTargetLoweringInfo().isLittleEndian())
7342 StOffset = ByteShift;
7344 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
7346 SDValue Ptr = St->getBasePtr();
7348 Ptr = DAG.getNode(ISD::ADD, IVal->getDebugLoc(), Ptr.getValueType(),
7349 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
7350 NewAlign = MinAlign(NewAlign, StOffset);
7353 // Truncate down to the new size.
7354 IVal = DAG.getNode(ISD::TRUNCATE, IVal->getDebugLoc(), VT, IVal);
7357 return DAG.getStore(St->getChain(), St->getDebugLoc(), IVal, Ptr,
7358 St->getPointerInfo().getWithOffset(StOffset),
7359 false, false, NewAlign).getNode();
7363 /// ReduceLoadOpStoreWidth - Look for sequence of load / op / store where op is
7364 /// one of 'or', 'xor', and 'and' of immediates. If 'op' is only touching some
7365 /// of the loaded bits, try narrowing the load and store if it would end up
7366 /// being a win for performance or code size.
7367 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
7368 StoreSDNode *ST = cast<StoreSDNode>(N);
7369 if (ST->isVolatile())
7372 SDValue Chain = ST->getChain();
7373 SDValue Value = ST->getValue();
7374 SDValue Ptr = ST->getBasePtr();
7375 EVT VT = Value.getValueType();
7377 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
7380 unsigned Opc = Value.getOpcode();
7382 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
7383 // is a byte mask indicating a consecutive number of bytes, check to see if
7384 // Y is known to provide just those bytes. If so, we try to replace the
7385 // load + replace + store sequence with a single (narrower) store, which makes
7387 if (Opc == ISD::OR) {
7388 std::pair<unsigned, unsigned> MaskedLoad;
7389 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
7390 if (MaskedLoad.first)
7391 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
7392 Value.getOperand(1), ST,this))
7393 return SDValue(NewST, 0);
7395 // Or is commutative, so try swapping X and Y.
7396 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
7397 if (MaskedLoad.first)
7398 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
7399 Value.getOperand(0), ST,this))
7400 return SDValue(NewST, 0);
7403 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
7404 Value.getOperand(1).getOpcode() != ISD::Constant)
7407 SDValue N0 = Value.getOperand(0);
7408 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7409 Chain == SDValue(N0.getNode(), 1)) {
7410 LoadSDNode *LD = cast<LoadSDNode>(N0);
7411 if (LD->getBasePtr() != Ptr ||
7412 LD->getPointerInfo().getAddrSpace() !=
7413 ST->getPointerInfo().getAddrSpace())
7416 // Find the type to narrow it the load / op / store to.
7417 SDValue N1 = Value.getOperand(1);
7418 unsigned BitWidth = N1.getValueSizeInBits();
7419 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
7420 if (Opc == ISD::AND)
7421 Imm ^= APInt::getAllOnesValue(BitWidth);
7422 if (Imm == 0 || Imm.isAllOnesValue())
7424 unsigned ShAmt = Imm.countTrailingZeros();
7425 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
7426 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
7427 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
7428 while (NewBW < BitWidth &&
7429 !(TLI.isOperationLegalOrCustom(Opc, NewVT) &&
7430 TLI.isNarrowingProfitable(VT, NewVT))) {
7431 NewBW = NextPowerOf2(NewBW);
7432 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
7434 if (NewBW >= BitWidth)
7437 // If the lsb changed does not start at the type bitwidth boundary,
7438 // start at the previous one.
7440 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
7441 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt,
7442 std::min(BitWidth, ShAmt + NewBW));
7443 if ((Imm & Mask) == Imm) {
7444 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
7445 if (Opc == ISD::AND)
7446 NewImm ^= APInt::getAllOnesValue(NewBW);
7447 uint64_t PtrOff = ShAmt / 8;
7448 // For big endian targets, we need to adjust the offset to the pointer to
7449 // load the correct bytes.
7450 if (TLI.isBigEndian())
7451 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
7453 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
7454 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
7455 if (NewAlign < TLI.getDataLayout()->getABITypeAlignment(NewVTTy))
7458 SDValue NewPtr = DAG.getNode(ISD::ADD, LD->getDebugLoc(),
7459 Ptr.getValueType(), Ptr,
7460 DAG.getConstant(PtrOff, Ptr.getValueType()));
7461 SDValue NewLD = DAG.getLoad(NewVT, N0.getDebugLoc(),
7462 LD->getChain(), NewPtr,
7463 LD->getPointerInfo().getWithOffset(PtrOff),
7464 LD->isVolatile(), LD->isNonTemporal(),
7465 LD->isInvariant(), NewAlign);
7466 SDValue NewVal = DAG.getNode(Opc, Value.getDebugLoc(), NewVT, NewLD,
7467 DAG.getConstant(NewImm, NewVT));
7468 SDValue NewST = DAG.getStore(Chain, N->getDebugLoc(),
7470 ST->getPointerInfo().getWithOffset(PtrOff),
7471 false, false, NewAlign);
7473 AddToWorkList(NewPtr.getNode());
7474 AddToWorkList(NewLD.getNode());
7475 AddToWorkList(NewVal.getNode());
7476 WorkListRemover DeadNodes(*this);
7477 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
7486 /// TransformFPLoadStorePair - For a given floating point load / store pair,
7487 /// if the load value isn't used by any other operations, then consider
7488 /// transforming the pair to integer load / store operations if the target
7489 /// deems the transformation profitable.
7490 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
7491 StoreSDNode *ST = cast<StoreSDNode>(N);
7492 SDValue Chain = ST->getChain();
7493 SDValue Value = ST->getValue();
7494 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
7495 Value.hasOneUse() &&
7496 Chain == SDValue(Value.getNode(), 1)) {
7497 LoadSDNode *LD = cast<LoadSDNode>(Value);
7498 EVT VT = LD->getMemoryVT();
7499 if (!VT.isFloatingPoint() ||
7500 VT != ST->getMemoryVT() ||
7501 LD->isNonTemporal() ||
7502 ST->isNonTemporal() ||
7503 LD->getPointerInfo().getAddrSpace() != 0 ||
7504 ST->getPointerInfo().getAddrSpace() != 0)
7507 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
7508 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
7509 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
7510 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
7511 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
7514 unsigned LDAlign = LD->getAlignment();
7515 unsigned STAlign = ST->getAlignment();
7516 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
7517 unsigned ABIAlign = TLI.getDataLayout()->getABITypeAlignment(IntVTTy);
7518 if (LDAlign < ABIAlign || STAlign < ABIAlign)
7521 SDValue NewLD = DAG.getLoad(IntVT, Value.getDebugLoc(),
7522 LD->getChain(), LD->getBasePtr(),
7523 LD->getPointerInfo(),
7524 false, false, false, LDAlign);
7526 SDValue NewST = DAG.getStore(NewLD.getValue(1), N->getDebugLoc(),
7527 NewLD, ST->getBasePtr(),
7528 ST->getPointerInfo(),
7529 false, false, STAlign);
7531 AddToWorkList(NewLD.getNode());
7532 AddToWorkList(NewST.getNode());
7533 WorkListRemover DeadNodes(*this);
7534 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
7542 /// Returns the base pointer and an integer offset from that object.
7543 static std::pair<SDValue, int64_t> GetPointerBaseAndOffset(SDValue Ptr) {
7544 if (Ptr->getOpcode() == ISD::ADD && isa<ConstantSDNode>(Ptr->getOperand(1))) {
7545 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
7546 SDValue Base = Ptr->getOperand(0);
7547 return std::make_pair(Base, Offset);
7550 return std::make_pair(Ptr, 0);
7553 /// Holds a pointer to an LSBaseSDNode as well as information on where it
7554 /// is located in a sequence of memory operations connected by a chain.
7556 MemOpLink (LSBaseSDNode *N, int64_t Offset, unsigned Seq):
7557 MemNode(N), OffsetFromBase(Offset), SequenceNum(Seq) { }
7558 // Ptr to the mem node.
7559 LSBaseSDNode *MemNode;
7560 // Offset from the base ptr.
7561 int64_t OffsetFromBase;
7562 // What is the sequence number of this mem node.
7563 // Lowest mem operand in the DAG starts at zero.
7564 unsigned SequenceNum;
7567 /// Sorts store nodes in a link according to their offset from a shared
7569 struct ConsecutiveMemoryChainSorter {
7570 bool operator()(MemOpLink LHS, MemOpLink RHS) {
7571 return LHS.OffsetFromBase < RHS.OffsetFromBase;
7575 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
7576 EVT MemVT = St->getMemoryVT();
7577 int64_t ElementSizeBytes = MemVT.getSizeInBits()/8;
7579 // Don't merge vectors into wider inputs.
7580 if (MemVT.isVector() || !MemVT.isSimple())
7583 // Perform an early exit check. Do not bother looking at stored values that
7584 // are not constants or loads.
7585 SDValue StoredVal = St->getValue();
7586 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
7587 if (!isa<ConstantSDNode>(StoredVal) && !isa<ConstantFPSDNode>(StoredVal) &&
7591 // Only look at ends of store sequences.
7592 SDValue Chain = SDValue(St, 1);
7593 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
7596 // This holds the base pointer and the offset in bytes from the base pointer.
7597 std::pair<SDValue, int64_t> BasePtr =
7598 GetPointerBaseAndOffset(St->getBasePtr());
7600 // We must have a base and an offset.
7601 if (!BasePtr.first.getNode())
7604 // Do not handle stores to undef base pointers.
7605 if (BasePtr.first.getOpcode() == ISD::UNDEF)
7608 // Save the LoadSDNodes that we find in the chain.
7609 // We need to make sure that these nodes do not interfere with
7610 // any of the store nodes.
7611 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
7613 // Save the StoreSDNodes that we find in the chain.
7614 SmallVector<MemOpLink, 8> StoreNodes;
7616 // Walk up the chain and look for nodes with offsets from the same
7617 // base pointer. Stop when reaching an instruction with a different kind
7618 // or instruction which has a different base pointer.
7620 StoreSDNode *Index = St;
7622 // If the chain has more than one use, then we can't reorder the mem ops.
7623 if (Index != St && !SDValue(Index, 1)->hasOneUse())
7626 // Find the base pointer and offset for this memory node.
7627 std::pair<SDValue, int64_t> Ptr =
7628 GetPointerBaseAndOffset(Index->getBasePtr());
7630 // Check that the base pointer is the same as the original one.
7631 if (Ptr.first.getNode() != BasePtr.first.getNode())
7634 // Check that the alignment is the same.
7635 if (Index->getAlignment() != St->getAlignment())
7638 // The memory operands must not be volatile.
7639 if (Index->isVolatile() || Index->isIndexed())
7643 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
7644 if (St->isTruncatingStore())
7647 // The stored memory type must be the same.
7648 if (Index->getMemoryVT() != MemVT)
7651 // We do not allow unaligned stores because we want to prevent overriding
7653 if (Index->getAlignment()*8 != MemVT.getSizeInBits())
7656 // We found a potential memory operand to merge.
7657 StoreNodes.push_back(MemOpLink(Index, Ptr.second, Seq++));
7659 // Find the next memory operand in the chain. If the next operand in the
7660 // chain is a store then move up and continue the scan with the next
7661 // memory operand. If the next operand is a load save it and use alias
7662 // information to check if it interferes with anything.
7663 SDNode *NextInChain = Index->getChain().getNode();
7665 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
7666 // We found a store node. Use it for the next iteration.
7669 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
7670 // Save the load node for later. Continue the scan.
7671 AliasLoadNodes.push_back(Ldn);
7672 NextInChain = Ldn->getChain().getNode();
7681 // Check if there is anything to merge.
7682 if (StoreNodes.size() < 2)
7685 // Sort the memory operands according to their distance from the base pointer.
7686 std::sort(StoreNodes.begin(), StoreNodes.end(),
7687 ConsecutiveMemoryChainSorter());
7689 // Scan the memory operations on the chain and find the first non-consecutive
7690 // store memory address.
7691 unsigned LastConsecutiveStore = 0;
7692 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
7693 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) {
7695 // Check that the addresses are consecutive starting from the second
7696 // element in the list of stores.
7698 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
7699 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
7704 // Check if this store interferes with any of the loads that we found.
7705 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld)
7706 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) {
7710 // We found a load that alias with this store. Stop the sequence.
7714 // Mark this node as useful.
7715 LastConsecutiveStore = i;
7718 // The node with the lowest store address.
7719 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
7721 // Store the constants into memory as one consecutive store.
7723 unsigned LastLegalType = 0;
7724 unsigned LastLegalVectorType = 0;
7725 bool NonZero = false;
7726 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
7727 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
7728 SDValue StoredVal = St->getValue();
7730 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
7731 NonZero |= !C->isNullValue();
7732 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) {
7733 NonZero |= !C->getConstantFPValue()->isNullValue();
7739 // Find a legal type for the constant store.
7740 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
7741 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
7742 if (TLI.isTypeLegal(StoreTy))
7743 LastLegalType = i+1;
7745 // Find a legal type for the vector store.
7746 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
7747 if (TLI.isTypeLegal(Ty))
7748 LastLegalVectorType = i + 1;
7751 // We only use vectors if the constant is known to be zero.
7753 LastLegalVectorType = 0;
7755 // Check if we found a legal integer type to store.
7756 if (LastLegalType == 0 && LastLegalVectorType == 0)
7759 bool UseVector = LastLegalVectorType > LastLegalType;
7760 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType;
7762 // Make sure we have something to merge.
7766 unsigned EarliestNodeUsed = 0;
7767 for (unsigned i=0; i < NumElem; ++i) {
7768 // Find a chain for the new wide-store operand. Notice that some
7769 // of the store nodes that we found may not be selected for inclusion
7770 // in the wide store. The chain we use needs to be the chain of the
7771 // earliest store node which is *used* and replaced by the wide store.
7772 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
7773 EarliestNodeUsed = i;
7776 // The earliest Node in the DAG.
7777 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
7778 DebugLoc DL = StoreNodes[0].MemNode->getDebugLoc();
7782 // Find a legal type for the vector store.
7783 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
7784 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
7785 StoredVal = DAG.getConstant(0, Ty);
7787 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
7788 APInt StoreInt(StoreBW, 0);
7790 // Construct a single integer constant which is made of the smaller
7792 bool IsLE = TLI.isLittleEndian();
7793 for (unsigned i = 0; i < NumElem ; ++i) {
7794 unsigned Idx = IsLE ?(NumElem - 1 - i) : i;
7795 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
7796 SDValue Val = St->getValue();
7797 StoreInt<<=ElementSizeBytes*8;
7798 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
7799 StoreInt|=C->getAPIntValue().zext(StoreBW);
7800 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
7801 StoreInt|= C->getValueAPF().bitcastToAPInt().zext(StoreBW);
7803 assert(false && "Invalid constant element type");
7807 // Create the new Load and Store operations.
7808 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
7809 StoredVal = DAG.getConstant(StoreInt, StoreTy);
7812 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), DL, StoredVal,
7813 FirstInChain->getBasePtr(),
7814 FirstInChain->getPointerInfo(),
7816 FirstInChain->getAlignment());
7818 // Replace the first store with the new store
7819 CombineTo(EarliestOp, NewStore);
7820 // Erase all other stores.
7821 for (unsigned i = 0; i < NumElem ; ++i) {
7822 if (StoreNodes[i].MemNode == EarliestOp)
7824 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
7825 // ReplaceAllUsesWith will replace all uses that existed when it was
7826 // called, but graph optimizations may cause new ones to appear. For
7827 // example, the case in pr14333 looks like
7829 // St's chain -> St -> another store -> X
7831 // And the only difference from St to the other store is the chain.
7832 // When we change it's chain to be St's chain they become identical,
7833 // get CSEed and the net result is that X is now a use of St.
7834 // Since we know that St is redundant, just iterate.
7835 while (!St->use_empty())
7836 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain());
7837 removeFromWorkList(St);
7844 // Below we handle the case of multiple consecutive stores that
7845 // come from multiple consecutive loads. We merge them into a single
7846 // wide load and a single wide store.
7848 // Look for load nodes which are used by the stored values.
7849 SmallVector<MemOpLink, 8> LoadNodes;
7851 // Find acceptable loads. Loads need to have the same chain (token factor),
7852 // must not be zext, volatile, indexed, and they must be consecutive.
7854 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
7855 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
7856 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
7859 // Loads must only have one use.
7860 if (!Ld->hasNUsesOfValue(1, 0))
7863 // Check that the alignment is the same as the stores.
7864 if (Ld->getAlignment() != St->getAlignment())
7867 // The memory operands must not be volatile.
7868 if (Ld->isVolatile() || Ld->isIndexed())
7871 // We do not accept ext loads.
7872 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
7875 // The stored memory type must be the same.
7876 if (Ld->getMemoryVT() != MemVT)
7879 std::pair<SDValue, int64_t> LdPtr =
7880 GetPointerBaseAndOffset(Ld->getBasePtr());
7882 // If this is not the first ptr that we check.
7883 if (LdBasePtr.getNode()) {
7884 // The base ptr must be the same.
7885 if (LdPtr.first != LdBasePtr)
7888 // Check that all other base pointers are the same as this one.
7889 LdBasePtr = LdPtr.first;
7892 // We found a potential memory operand to merge.
7893 LoadNodes.push_back(MemOpLink(Ld, LdPtr.second, 0));
7896 if (LoadNodes.size() < 2)
7899 // Scan the memory operations on the chain and find the first non-consecutive
7900 // load memory address. These variables hold the index in the store node
7902 unsigned LastConsecutiveLoad = 0;
7903 // This variable refers to the size and not index in the array.
7904 unsigned LastLegalVectorType = 0;
7905 unsigned LastLegalIntegerType = 0;
7906 StartAddress = LoadNodes[0].OffsetFromBase;
7907 SDValue FirstChain = LoadNodes[0].MemNode->getChain();
7908 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
7909 // All loads much share the same chain.
7910 if (LoadNodes[i].MemNode->getChain() != FirstChain)
7913 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
7914 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
7916 LastConsecutiveLoad = i;
7918 // Find a legal type for the vector store.
7919 EVT StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
7920 if (TLI.isTypeLegal(StoreTy))
7921 LastLegalVectorType = i + 1;
7923 // Find a legal type for the integer store.
7924 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
7925 StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
7926 if (TLI.isTypeLegal(StoreTy))
7927 LastLegalIntegerType = i + 1;
7930 // Only use vector types if the vector type is larger than the integer type.
7931 // If they are the same, use integers.
7932 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType;
7933 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType);
7935 // We add +1 here because the LastXXX variables refer to location while
7936 // the NumElem refers to array/index size.
7937 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1;
7938 NumElem = std::min(LastLegalType, NumElem);
7943 // The earliest Node in the DAG.
7944 unsigned EarliestNodeUsed = 0;
7945 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
7946 for (unsigned i=1; i<NumElem; ++i) {
7947 // Find a chain for the new wide-store operand. Notice that some
7948 // of the store nodes that we found may not be selected for inclusion
7949 // in the wide store. The chain we use needs to be the chain of the
7950 // earliest store node which is *used* and replaced by the wide store.
7951 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
7952 EarliestNodeUsed = i;
7955 // Find if it is better to use vectors or integers to load and store
7959 JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
7961 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
7962 JointMemOpVT = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
7965 DebugLoc LoadDL = LoadNodes[0].MemNode->getDebugLoc();
7966 DebugLoc StoreDL = StoreNodes[0].MemNode->getDebugLoc();
7968 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
7969 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL,
7970 FirstLoad->getChain(),
7971 FirstLoad->getBasePtr(),
7972 FirstLoad->getPointerInfo(),
7973 false, false, false,
7974 FirstLoad->getAlignment());
7976 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), StoreDL, NewLoad,
7977 FirstInChain->getBasePtr(),
7978 FirstInChain->getPointerInfo(), false, false,
7979 FirstInChain->getAlignment());
7981 // Replace one of the loads with the new load.
7982 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode);
7983 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
7984 SDValue(NewLoad.getNode(), 1));
7986 // Remove the rest of the load chains.
7987 for (unsigned i = 1; i < NumElem ; ++i) {
7988 // Replace all chain users of the old load nodes with the chain of the new
7990 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
7991 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain());
7994 // Replace the first store with the new store.
7995 CombineTo(EarliestOp, NewStore);
7996 // Erase all other stores.
7997 for (unsigned i = 0; i < NumElem ; ++i) {
7998 // Remove all Store nodes.
7999 if (StoreNodes[i].MemNode == EarliestOp)
8001 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
8002 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain());
8003 removeFromWorkList(St);
8010 SDValue DAGCombiner::visitSTORE(SDNode *N) {
8011 StoreSDNode *ST = cast<StoreSDNode>(N);
8012 SDValue Chain = ST->getChain();
8013 SDValue Value = ST->getValue();
8014 SDValue Ptr = ST->getBasePtr();
8016 // If this is a store of a bit convert, store the input value if the
8017 // resultant store does not need a higher alignment than the original.
8018 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
8019 ST->isUnindexed()) {
8020 unsigned OrigAlign = ST->getAlignment();
8021 EVT SVT = Value.getOperand(0).getValueType();
8022 unsigned Align = TLI.getDataLayout()->
8023 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
8024 if (Align <= OrigAlign &&
8025 ((!LegalOperations && !ST->isVolatile()) ||
8026 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
8027 return DAG.getStore(Chain, N->getDebugLoc(), Value.getOperand(0),
8028 Ptr, ST->getPointerInfo(), ST->isVolatile(),
8029 ST->isNonTemporal(), OrigAlign);
8032 // Turn 'store undef, Ptr' -> nothing.
8033 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
8036 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
8037 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
8038 // NOTE: If the original store is volatile, this transform must not increase
8039 // the number of stores. For example, on x86-32 an f64 can be stored in one
8040 // processor operation but an i64 (which is not legal) requires two. So the
8041 // transform should not be done in this case.
8042 if (Value.getOpcode() != ISD::TargetConstantFP) {
8044 switch (CFP->getValueType(0).getSimpleVT().SimpleTy) {
8045 default: llvm_unreachable("Unknown FP type");
8046 case MVT::f16: // We don't do this for these yet.
8052 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
8053 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
8054 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
8055 bitcastToAPInt().getZExtValue(), MVT::i32);
8056 return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
8057 Ptr, ST->getPointerInfo(), ST->isVolatile(),
8058 ST->isNonTemporal(), ST->getAlignment());
8062 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
8063 !ST->isVolatile()) ||
8064 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
8065 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
8066 getZExtValue(), MVT::i64);
8067 return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
8068 Ptr, ST->getPointerInfo(), ST->isVolatile(),
8069 ST->isNonTemporal(), ST->getAlignment());
8072 if (!ST->isVolatile() &&
8073 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
8074 // Many FP stores are not made apparent until after legalize, e.g. for
8075 // argument passing. Since this is so common, custom legalize the
8076 // 64-bit integer store into two 32-bit stores.
8077 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
8078 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
8079 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
8080 if (TLI.isBigEndian()) std::swap(Lo, Hi);
8082 unsigned Alignment = ST->getAlignment();
8083 bool isVolatile = ST->isVolatile();
8084 bool isNonTemporal = ST->isNonTemporal();
8086 SDValue St0 = DAG.getStore(Chain, ST->getDebugLoc(), Lo,
8087 Ptr, ST->getPointerInfo(),
8088 isVolatile, isNonTemporal,
8089 ST->getAlignment());
8090 Ptr = DAG.getNode(ISD::ADD, N->getDebugLoc(), Ptr.getValueType(), Ptr,
8091 DAG.getConstant(4, Ptr.getValueType()));
8092 Alignment = MinAlign(Alignment, 4U);
8093 SDValue St1 = DAG.getStore(Chain, ST->getDebugLoc(), Hi,
8094 Ptr, ST->getPointerInfo().getWithOffset(4),
8095 isVolatile, isNonTemporal,
8097 return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
8106 // Try to infer better alignment information than the store already has.
8107 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
8108 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
8109 if (Align > ST->getAlignment())
8110 return DAG.getTruncStore(Chain, N->getDebugLoc(), Value,
8111 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
8112 ST->isVolatile(), ST->isNonTemporal(), Align);
8116 // Try transforming a pair floating point load / store ops to integer
8117 // load / store ops.
8118 SDValue NewST = TransformFPLoadStorePair(N);
8119 if (NewST.getNode())
8123 // Walk up chain skipping non-aliasing memory nodes.
8124 SDValue BetterChain = FindBetterChain(N, Chain);
8126 // If there is a better chain.
8127 if (Chain != BetterChain) {
8130 // Replace the chain to avoid dependency.
8131 if (ST->isTruncatingStore()) {
8132 ReplStore = DAG.getTruncStore(BetterChain, N->getDebugLoc(), Value, Ptr,
8133 ST->getPointerInfo(),
8134 ST->getMemoryVT(), ST->isVolatile(),
8135 ST->isNonTemporal(), ST->getAlignment());
8137 ReplStore = DAG.getStore(BetterChain, N->getDebugLoc(), Value, Ptr,
8138 ST->getPointerInfo(),
8139 ST->isVolatile(), ST->isNonTemporal(),
8140 ST->getAlignment());
8143 // Create token to keep both nodes around.
8144 SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
8145 MVT::Other, Chain, ReplStore);
8147 // Make sure the new and old chains are cleaned up.
8148 AddToWorkList(Token.getNode());
8150 // Don't add users to work list.
8151 return CombineTo(N, Token, false);
8155 // Try transforming N to an indexed store.
8156 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
8157 return SDValue(N, 0);
8159 // FIXME: is there such a thing as a truncating indexed store?
8160 if (ST->isTruncatingStore() && ST->isUnindexed() &&
8161 Value.getValueType().isInteger()) {
8162 // See if we can simplify the input to this truncstore with knowledge that
8163 // only the low bits are being used. For example:
8164 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
8166 GetDemandedBits(Value,
8167 APInt::getLowBitsSet(
8168 Value.getValueType().getScalarType().getSizeInBits(),
8169 ST->getMemoryVT().getScalarType().getSizeInBits()));
8170 AddToWorkList(Value.getNode());
8171 if (Shorter.getNode())
8172 return DAG.getTruncStore(Chain, N->getDebugLoc(), Shorter,
8173 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
8174 ST->isVolatile(), ST->isNonTemporal(),
8175 ST->getAlignment());
8177 // Otherwise, see if we can simplify the operation with
8178 // SimplifyDemandedBits, which only works if the value has a single use.
8179 if (SimplifyDemandedBits(Value,
8180 APInt::getLowBitsSet(
8181 Value.getValueType().getScalarType().getSizeInBits(),
8182 ST->getMemoryVT().getScalarType().getSizeInBits())))
8183 return SDValue(N, 0);
8186 // If this is a load followed by a store to the same location, then the store
8188 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
8189 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
8190 ST->isUnindexed() && !ST->isVolatile() &&
8191 // There can't be any side effects between the load and store, such as
8193 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
8194 // The store is dead, remove it.
8199 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
8200 // truncating store. We can do this even if this is already a truncstore.
8201 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
8202 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
8203 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
8204 ST->getMemoryVT())) {
8205 return DAG.getTruncStore(Chain, N->getDebugLoc(), Value.getOperand(0),
8206 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
8207 ST->isVolatile(), ST->isNonTemporal(),
8208 ST->getAlignment());
8211 // Only perform this optimization before the types are legal, because we
8212 // don't want to perform this optimization on every DAGCombine invocation.
8214 bool EverChanged = false;
8217 // There can be multiple store sequences on the same chain.
8218 // Keep trying to merge store sequences until we are unable to do so
8219 // or until we merge the last store on the chain.
8220 bool Changed = MergeConsecutiveStores(ST);
8221 EverChanged |= Changed;
8222 if (!Changed) break;
8223 } while (ST->getOpcode() != ISD::DELETED_NODE);
8226 return SDValue(N, 0);
8229 return ReduceLoadOpStoreWidth(N);
8232 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
8233 SDValue InVec = N->getOperand(0);
8234 SDValue InVal = N->getOperand(1);
8235 SDValue EltNo = N->getOperand(2);
8236 DebugLoc dl = N->getDebugLoc();
8238 // If the inserted element is an UNDEF, just use the input vector.
8239 if (InVal.getOpcode() == ISD::UNDEF)
8242 EVT VT = InVec.getValueType();
8244 // If we can't generate a legal BUILD_VECTOR, exit
8245 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
8248 // Check that we know which element is being inserted
8249 if (!isa<ConstantSDNode>(EltNo))
8251 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
8253 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
8254 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
8256 SmallVector<SDValue, 8> Ops;
8257 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
8258 Ops.append(InVec.getNode()->op_begin(),
8259 InVec.getNode()->op_end());
8260 } else if (InVec.getOpcode() == ISD::UNDEF) {
8261 unsigned NElts = VT.getVectorNumElements();
8262 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
8267 // Insert the element
8268 if (Elt < Ops.size()) {
8269 // All the operands of BUILD_VECTOR must have the same type;
8270 // we enforce that here.
8271 EVT OpVT = Ops[0].getValueType();
8272 if (InVal.getValueType() != OpVT)
8273 InVal = OpVT.bitsGT(InVal.getValueType()) ?
8274 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
8275 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
8279 // Return the new vector
8280 return DAG.getNode(ISD::BUILD_VECTOR, dl,
8281 VT, &Ops[0], Ops.size());
8284 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
8285 // (vextract (scalar_to_vector val, 0) -> val
8286 SDValue InVec = N->getOperand(0);
8287 EVT VT = InVec.getValueType();
8288 EVT NVT = N->getValueType(0);
8290 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
8291 // Check if the result type doesn't match the inserted element type. A
8292 // SCALAR_TO_VECTOR may truncate the inserted element and the
8293 // EXTRACT_VECTOR_ELT may widen the extracted vector.
8294 SDValue InOp = InVec.getOperand(0);
8295 if (InOp.getValueType() != NVT) {
8296 assert(InOp.getValueType().isInteger() && NVT.isInteger());
8297 return DAG.getSExtOrTrunc(InOp, InVec.getDebugLoc(), NVT);
8302 SDValue EltNo = N->getOperand(1);
8303 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
8305 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
8306 // We only perform this optimization before the op legalization phase because
8307 // we may introduce new vector instructions which are not backed by TD
8308 // patterns. For example on AVX, extracting elements from a wide vector
8309 // without using extract_subvector.
8310 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
8311 && ConstEltNo && !LegalOperations) {
8312 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
8313 int NumElem = VT.getVectorNumElements();
8314 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
8315 // Find the new index to extract from.
8316 int OrigElt = SVOp->getMaskElt(Elt);
8318 // Extracting an undef index is undef.
8320 return DAG.getUNDEF(NVT);
8322 // Select the right vector half to extract from.
8323 if (OrigElt < NumElem) {
8324 InVec = InVec->getOperand(0);
8326 InVec = InVec->getOperand(1);
8330 EVT IndexTy = N->getOperand(1).getValueType();
8331 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, N->getDebugLoc(), NVT,
8332 InVec, DAG.getConstant(OrigElt, IndexTy));
8335 // Perform only after legalization to ensure build_vector / vector_shuffle
8336 // optimizations have already been done.
8337 if (!LegalOperations) return SDValue();
8339 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
8340 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
8341 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
8344 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
8345 bool NewLoad = false;
8346 bool BCNumEltsChanged = false;
8347 EVT ExtVT = VT.getVectorElementType();
8350 // If the result of load has to be truncated, then it's not necessarily
8352 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
8355 if (InVec.getOpcode() == ISD::BITCAST) {
8356 // Don't duplicate a load with other uses.
8357 if (!InVec.hasOneUse())
8360 EVT BCVT = InVec.getOperand(0).getValueType();
8361 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
8363 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
8364 BCNumEltsChanged = true;
8365 InVec = InVec.getOperand(0);
8366 ExtVT = BCVT.getVectorElementType();
8370 LoadSDNode *LN0 = NULL;
8371 const ShuffleVectorSDNode *SVN = NULL;
8372 if (ISD::isNormalLoad(InVec.getNode())) {
8373 LN0 = cast<LoadSDNode>(InVec);
8374 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
8375 InVec.getOperand(0).getValueType() == ExtVT &&
8376 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
8377 // Don't duplicate a load with other uses.
8378 if (!InVec.hasOneUse())
8381 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
8382 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
8383 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
8385 // (load $addr+1*size)
8387 // Don't duplicate a load with other uses.
8388 if (!InVec.hasOneUse())
8391 // If the bit convert changed the number of elements, it is unsafe
8392 // to examine the mask.
8393 if (BCNumEltsChanged)
8396 // Select the input vector, guarding against out of range extract vector.
8397 unsigned NumElems = VT.getVectorNumElements();
8398 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
8399 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
8401 if (InVec.getOpcode() == ISD::BITCAST) {
8402 // Don't duplicate a load with other uses.
8403 if (!InVec.hasOneUse())
8406 InVec = InVec.getOperand(0);
8408 if (ISD::isNormalLoad(InVec.getNode())) {
8409 LN0 = cast<LoadSDNode>(InVec);
8410 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
8414 // Make sure we found a non-volatile load and the extractelement is
8416 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
8419 // If Idx was -1 above, Elt is going to be -1, so just return undef.
8421 return DAG.getUNDEF(LVT);
8423 unsigned Align = LN0->getAlignment();
8425 // Check the resultant load doesn't need a higher alignment than the
8429 ->getABITypeAlignment(LVT.getTypeForEVT(*DAG.getContext()));
8431 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, LVT))
8437 SDValue NewPtr = LN0->getBasePtr();
8438 unsigned PtrOff = 0;
8441 PtrOff = LVT.getSizeInBits() * Elt / 8;
8442 EVT PtrType = NewPtr.getValueType();
8443 if (TLI.isBigEndian())
8444 PtrOff = VT.getSizeInBits() / 8 - PtrOff;
8445 NewPtr = DAG.getNode(ISD::ADD, N->getDebugLoc(), PtrType, NewPtr,
8446 DAG.getConstant(PtrOff, PtrType));
8449 // The replacement we need to do here is a little tricky: we need to
8450 // replace an extractelement of a load with a load.
8451 // Use ReplaceAllUsesOfValuesWith to do the replacement.
8452 // Note that this replacement assumes that the extractvalue is the only
8453 // use of the load; that's okay because we don't want to perform this
8454 // transformation in other cases anyway.
8457 if (NVT.bitsGT(LVT)) {
8458 // If the result type of vextract is wider than the load, then issue an
8459 // extending load instead.
8460 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, LVT)
8461 ? ISD::ZEXTLOAD : ISD::EXTLOAD;
8462 Load = DAG.getExtLoad(ExtType, N->getDebugLoc(), NVT, LN0->getChain(),
8463 NewPtr, LN0->getPointerInfo().getWithOffset(PtrOff),
8464 LVT, LN0->isVolatile(), LN0->isNonTemporal(),Align);
8465 Chain = Load.getValue(1);
8467 Load = DAG.getLoad(LVT, N->getDebugLoc(), LN0->getChain(), NewPtr,
8468 LN0->getPointerInfo().getWithOffset(PtrOff),
8469 LN0->isVolatile(), LN0->isNonTemporal(),
8470 LN0->isInvariant(), Align);
8471 Chain = Load.getValue(1);
8472 if (NVT.bitsLT(LVT))
8473 Load = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), NVT, Load);
8475 Load = DAG.getNode(ISD::BITCAST, N->getDebugLoc(), NVT, Load);
8477 WorkListRemover DeadNodes(*this);
8478 SDValue From[] = { SDValue(N, 0), SDValue(LN0,1) };
8479 SDValue To[] = { Load, Chain };
8480 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
8481 // Since we're explcitly calling ReplaceAllUses, add the new node to the
8482 // worklist explicitly as well.
8483 AddToWorkList(Load.getNode());
8484 AddUsersToWorkList(Load.getNode()); // Add users too
8485 // Make sure to revisit this node to clean it up; it will usually be dead.
8487 return SDValue(N, 0);
8493 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
8494 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
8495 // We perform this optimization post type-legalization because
8496 // the type-legalizer often scalarizes integer-promoted vectors.
8497 // Performing this optimization before may create bit-casts which
8498 // will be type-legalized to complex code sequences.
8499 // We perform this optimization only before the operation legalizer because we
8500 // may introduce illegal operations.
8501 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
8504 unsigned NumInScalars = N->getNumOperands();
8505 DebugLoc dl = N->getDebugLoc();
8506 EVT VT = N->getValueType(0);
8508 // Check to see if this is a BUILD_VECTOR of a bunch of values
8509 // which come from any_extend or zero_extend nodes. If so, we can create
8510 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
8511 // optimizations. We do not handle sign-extend because we can't fill the sign
8513 EVT SourceType = MVT::Other;
8514 bool AllAnyExt = true;
8516 for (unsigned i = 0; i != NumInScalars; ++i) {
8517 SDValue In = N->getOperand(i);
8518 // Ignore undef inputs.
8519 if (In.getOpcode() == ISD::UNDEF) continue;
8521 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
8522 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
8524 // Abort if the element is not an extension.
8525 if (!ZeroExt && !AnyExt) {
8526 SourceType = MVT::Other;
8530 // The input is a ZeroExt or AnyExt. Check the original type.
8531 EVT InTy = In.getOperand(0).getValueType();
8533 // Check that all of the widened source types are the same.
8534 if (SourceType == MVT::Other)
8537 else if (InTy != SourceType) {
8538 // Multiple income types. Abort.
8539 SourceType = MVT::Other;
8543 // Check if all of the extends are ANY_EXTENDs.
8544 AllAnyExt &= AnyExt;
8547 // In order to have valid types, all of the inputs must be extended from the
8548 // same source type and all of the inputs must be any or zero extend.
8549 // Scalar sizes must be a power of two.
8550 EVT OutScalarTy = VT.getScalarType();
8551 bool ValidTypes = SourceType != MVT::Other &&
8552 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
8553 isPowerOf2_32(SourceType.getSizeInBits());
8555 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
8556 // turn into a single shuffle instruction.
8560 bool isLE = TLI.isLittleEndian();
8561 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
8562 assert(ElemRatio > 1 && "Invalid element size ratio");
8563 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
8564 DAG.getConstant(0, SourceType);
8566 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
8567 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
8569 // Populate the new build_vector
8570 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
8571 SDValue Cast = N->getOperand(i);
8572 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
8573 Cast.getOpcode() == ISD::ZERO_EXTEND ||
8574 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
8576 if (Cast.getOpcode() == ISD::UNDEF)
8577 In = DAG.getUNDEF(SourceType);
8579 In = Cast->getOperand(0);
8580 unsigned Index = isLE ? (i * ElemRatio) :
8581 (i * ElemRatio + (ElemRatio - 1));
8583 assert(Index < Ops.size() && "Invalid index");
8587 // The type of the new BUILD_VECTOR node.
8588 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
8589 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
8590 "Invalid vector size");
8591 // Check if the new vector type is legal.
8592 if (!isTypeLegal(VecVT)) return SDValue();
8594 // Make the new BUILD_VECTOR.
8595 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], Ops.size());
8597 // The new BUILD_VECTOR node has the potential to be further optimized.
8598 AddToWorkList(BV.getNode());
8599 // Bitcast to the desired type.
8600 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8603 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
8604 EVT VT = N->getValueType(0);
8606 unsigned NumInScalars = N->getNumOperands();
8607 DebugLoc dl = N->getDebugLoc();
8609 EVT SrcVT = MVT::Other;
8610 unsigned Opcode = ISD::DELETED_NODE;
8611 unsigned NumDefs = 0;
8613 for (unsigned i = 0; i != NumInScalars; ++i) {
8614 SDValue In = N->getOperand(i);
8615 unsigned Opc = In.getOpcode();
8617 if (Opc == ISD::UNDEF)
8620 // If all scalar values are floats and converted from integers.
8621 if (Opcode == ISD::DELETED_NODE &&
8622 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
8624 // If not supported by target, bail out.
8625 if (TLI.getOperationAction(Opcode, VT) != TargetLowering::Legal &&
8626 TLI.getOperationAction(Opcode, VT) != TargetLowering::Custom)
8632 EVT InVT = In.getOperand(0).getValueType();
8634 // If all scalar values are typed differently, bail out. It's chosen to
8635 // simplify BUILD_VECTOR of integer types.
8636 if (SrcVT == MVT::Other)
8643 // If the vector has just one element defined, it's not worth to fold it into
8644 // a vectorized one.
8648 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
8649 && "Should only handle conversion from integer to float.");
8650 assert(SrcVT != MVT::Other && "Cannot determine source type!");
8652 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
8653 SmallVector<SDValue, 8> Opnds;
8654 for (unsigned i = 0; i != NumInScalars; ++i) {
8655 SDValue In = N->getOperand(i);
8657 if (In.getOpcode() == ISD::UNDEF)
8658 Opnds.push_back(DAG.getUNDEF(SrcVT));
8660 Opnds.push_back(In.getOperand(0));
8662 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT,
8663 &Opnds[0], Opnds.size());
8664 AddToWorkList(BV.getNode());
8666 return DAG.getNode(Opcode, dl, VT, BV);
8669 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
8670 unsigned NumInScalars = N->getNumOperands();
8671 DebugLoc dl = N->getDebugLoc();
8672 EVT VT = N->getValueType(0);
8674 // A vector built entirely of undefs is undef.
8675 if (ISD::allOperandsUndef(N))
8676 return DAG.getUNDEF(VT);
8678 SDValue V = reduceBuildVecExtToExtBuildVec(N);
8682 V = reduceBuildVecConvertToConvertBuildVec(N);
8686 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
8687 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
8688 // at most two distinct vectors, turn this into a shuffle node.
8690 // May only combine to shuffle after legalize if shuffle is legal.
8691 if (LegalOperations &&
8692 !TLI.isOperationLegalOrCustom(ISD::VECTOR_SHUFFLE, VT))
8695 SDValue VecIn1, VecIn2;
8696 for (unsigned i = 0; i != NumInScalars; ++i) {
8697 // Ignore undef inputs.
8698 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
8700 // If this input is something other than a EXTRACT_VECTOR_ELT with a
8701 // constant index, bail out.
8702 if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
8703 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
8704 VecIn1 = VecIn2 = SDValue(0, 0);
8708 // We allow up to two distinct input vectors.
8709 SDValue ExtractedFromVec = N->getOperand(i).getOperand(0);
8710 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
8713 if (VecIn1.getNode() == 0) {
8714 VecIn1 = ExtractedFromVec;
8715 } else if (VecIn2.getNode() == 0) {
8716 VecIn2 = ExtractedFromVec;
8719 VecIn1 = VecIn2 = SDValue(0, 0);
8724 // If everything is good, we can make a shuffle operation.
8725 if (VecIn1.getNode()) {
8726 SmallVector<int, 8> Mask;
8727 for (unsigned i = 0; i != NumInScalars; ++i) {
8728 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
8733 // If extracting from the first vector, just use the index directly.
8734 SDValue Extract = N->getOperand(i);
8735 SDValue ExtVal = Extract.getOperand(1);
8736 if (Extract.getOperand(0) == VecIn1) {
8737 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
8738 if (ExtIndex > VT.getVectorNumElements())
8741 Mask.push_back(ExtIndex);
8745 // Otherwise, use InIdx + VecSize
8746 unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue();
8747 Mask.push_back(Idx+NumInScalars);
8750 // We can't generate a shuffle node with mismatched input and output types.
8751 // Attempt to transform a single input vector to the correct type.
8752 if ((VT != VecIn1.getValueType())) {
8753 // We don't support shuffeling between TWO values of different types.
8754 if (VecIn2.getNode() != 0)
8757 // We only support widening of vectors which are half the size of the
8758 // output registers. For example XMM->YMM widening on X86 with AVX.
8759 if (VecIn1.getValueType().getSizeInBits()*2 != VT.getSizeInBits())
8762 // If the input vector type has a different base type to the output
8763 // vector type, bail out.
8764 if (VecIn1.getValueType().getVectorElementType() !=
8765 VT.getVectorElementType())
8768 // Widen the input vector by adding undef values.
8769 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
8770 VecIn1, DAG.getUNDEF(VecIn1.getValueType()));
8773 // If VecIn2 is unused then change it to undef.
8774 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
8776 // Check that we were able to transform all incoming values to the same
8778 if (VecIn2.getValueType() != VecIn1.getValueType() ||
8779 VecIn1.getValueType() != VT)
8782 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
8783 if (!isTypeLegal(VT))
8786 // Return the new VECTOR_SHUFFLE node.
8790 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]);
8796 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
8797 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
8798 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
8799 // inputs come from at most two distinct vectors, turn this into a shuffle
8802 // If we only have one input vector, we don't need to do any concatenation.
8803 if (N->getNumOperands() == 1)
8804 return N->getOperand(0);
8806 // Check if all of the operands are undefs.
8807 if (ISD::allOperandsUndef(N))
8808 return DAG.getUNDEF(N->getValueType(0));
8813 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
8814 EVT NVT = N->getValueType(0);
8815 SDValue V = N->getOperand(0);
8817 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
8818 // Handle only simple case where vector being inserted and vector
8819 // being extracted are of same type, and are half size of larger vectors.
8820 EVT BigVT = V->getOperand(0).getValueType();
8821 EVT SmallVT = V->getOperand(1).getValueType();
8822 if (NVT != SmallVT || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
8825 // Only handle cases where both indexes are constants with the same type.
8826 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
8827 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
8829 if (InsIdx && ExtIdx &&
8830 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
8831 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
8833 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
8835 // indices are equal => V1
8836 // otherwise => (extract_subvec V1, ExtIdx)
8837 if (InsIdx->getZExtValue() == ExtIdx->getZExtValue())
8838 return V->getOperand(1);
8839 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, N->getDebugLoc(), NVT,
8840 V->getOperand(0), N->getOperand(1));
8844 if (V->getOpcode() == ISD::CONCAT_VECTORS) {
8846 // (extract_subvec (concat V1, V2, ...), i)
8849 // Only operand 0 is checked as 'concat' assumes all inputs of the same type.
8850 if (V->getOperand(0).getValueType() != NVT)
8852 unsigned Idx = dyn_cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8853 unsigned NumElems = NVT.getVectorNumElements();
8854 assert((Idx % NumElems) == 0 &&
8855 "IDX in concat is not a multiple of the result vector length.");
8856 return V->getOperand(Idx / NumElems);
8862 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
8863 EVT VT = N->getValueType(0);
8864 unsigned NumElts = VT.getVectorNumElements();
8866 SDValue N0 = N->getOperand(0);
8867 SDValue N1 = N->getOperand(1);
8869 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
8871 // Canonicalize shuffle undef, undef -> undef
8872 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
8873 return DAG.getUNDEF(VT);
8875 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8877 // Canonicalize shuffle v, v -> v, undef
8879 SmallVector<int, 8> NewMask;
8880 for (unsigned i = 0; i != NumElts; ++i) {
8881 int Idx = SVN->getMaskElt(i);
8882 if (Idx >= (int)NumElts) Idx -= NumElts;
8883 NewMask.push_back(Idx);
8885 return DAG.getVectorShuffle(VT, N->getDebugLoc(), N0, DAG.getUNDEF(VT),
8889 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
8890 if (N0.getOpcode() == ISD::UNDEF) {
8891 SmallVector<int, 8> NewMask;
8892 for (unsigned i = 0; i != NumElts; ++i) {
8893 int Idx = SVN->getMaskElt(i);
8895 if (Idx < (int)NumElts)
8900 NewMask.push_back(Idx);
8902 return DAG.getVectorShuffle(VT, N->getDebugLoc(), N1, DAG.getUNDEF(VT),
8906 // Remove references to rhs if it is undef
8907 if (N1.getOpcode() == ISD::UNDEF) {
8908 bool Changed = false;
8909 SmallVector<int, 8> NewMask;
8910 for (unsigned i = 0; i != NumElts; ++i) {
8911 int Idx = SVN->getMaskElt(i);
8912 if (Idx >= (int)NumElts) {
8916 NewMask.push_back(Idx);
8919 return DAG.getVectorShuffle(VT, N->getDebugLoc(), N0, N1, &NewMask[0]);
8922 // If it is a splat, check if the argument vector is another splat or a
8923 // build_vector with all scalar elements the same.
8924 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
8925 SDNode *V = N0.getNode();
8927 // If this is a bit convert that changes the element type of the vector but
8928 // not the number of vector elements, look through it. Be careful not to
8929 // look though conversions that change things like v4f32 to v2f64.
8930 if (V->getOpcode() == ISD::BITCAST) {
8931 SDValue ConvInput = V->getOperand(0);
8932 if (ConvInput.getValueType().isVector() &&
8933 ConvInput.getValueType().getVectorNumElements() == NumElts)
8934 V = ConvInput.getNode();
8937 if (V->getOpcode() == ISD::BUILD_VECTOR) {
8938 assert(V->getNumOperands() == NumElts &&
8939 "BUILD_VECTOR has wrong number of operands");
8941 bool AllSame = true;
8942 for (unsigned i = 0; i != NumElts; ++i) {
8943 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
8944 Base = V->getOperand(i);
8948 // Splat of <u, u, u, u>, return <u, u, u, u>
8949 if (!Base.getNode())
8951 for (unsigned i = 0; i != NumElts; ++i) {
8952 if (V->getOperand(i) != Base) {
8957 // Splat of <x, x, x, x>, return <x, x, x, x>
8963 // If this shuffle node is simply a swizzle of another shuffle node,
8964 // and it reverses the swizzle of the previous shuffle then we can
8965 // optimize shuffle(shuffle(x, undef), undef) -> x.
8966 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
8967 N1.getOpcode() == ISD::UNDEF) {
8969 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
8971 // Shuffle nodes can only reverse shuffles with a single non-undef value.
8972 if (N0.getOperand(1).getOpcode() != ISD::UNDEF)
8975 // The incoming shuffle must be of the same type as the result of the
8977 assert(OtherSV->getOperand(0).getValueType() == VT &&
8978 "Shuffle types don't match");
8980 for (unsigned i = 0; i != NumElts; ++i) {
8981 int Idx = SVN->getMaskElt(i);
8982 assert(Idx < (int)NumElts && "Index references undef operand");
8983 // Next, this index comes from the first value, which is the incoming
8984 // shuffle. Adopt the incoming index.
8986 Idx = OtherSV->getMaskElt(Idx);
8988 // The combined shuffle must map each index to itself.
8989 if (Idx >= 0 && (unsigned)Idx != i)
8993 return OtherSV->getOperand(0);
8999 SDValue DAGCombiner::visitMEMBARRIER(SDNode* N) {
9000 if (!TLI.getShouldFoldAtomicFences())
9003 SDValue atomic = N->getOperand(0);
9004 switch (atomic.getOpcode()) {
9005 case ISD::ATOMIC_CMP_SWAP:
9006 case ISD::ATOMIC_SWAP:
9007 case ISD::ATOMIC_LOAD_ADD:
9008 case ISD::ATOMIC_LOAD_SUB:
9009 case ISD::ATOMIC_LOAD_AND:
9010 case ISD::ATOMIC_LOAD_OR:
9011 case ISD::ATOMIC_LOAD_XOR:
9012 case ISD::ATOMIC_LOAD_NAND:
9013 case ISD::ATOMIC_LOAD_MIN:
9014 case ISD::ATOMIC_LOAD_MAX:
9015 case ISD::ATOMIC_LOAD_UMIN:
9016 case ISD::ATOMIC_LOAD_UMAX:
9022 SDValue fence = atomic.getOperand(0);
9023 if (fence.getOpcode() != ISD::MEMBARRIER)
9026 switch (atomic.getOpcode()) {
9027 case ISD::ATOMIC_CMP_SWAP:
9028 return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
9029 fence.getOperand(0),
9030 atomic.getOperand(1), atomic.getOperand(2),
9031 atomic.getOperand(3)), atomic.getResNo());
9032 case ISD::ATOMIC_SWAP:
9033 case ISD::ATOMIC_LOAD_ADD:
9034 case ISD::ATOMIC_LOAD_SUB:
9035 case ISD::ATOMIC_LOAD_AND:
9036 case ISD::ATOMIC_LOAD_OR:
9037 case ISD::ATOMIC_LOAD_XOR:
9038 case ISD::ATOMIC_LOAD_NAND:
9039 case ISD::ATOMIC_LOAD_MIN:
9040 case ISD::ATOMIC_LOAD_MAX:
9041 case ISD::ATOMIC_LOAD_UMIN:
9042 case ISD::ATOMIC_LOAD_UMAX:
9043 return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
9044 fence.getOperand(0),
9045 atomic.getOperand(1), atomic.getOperand(2)),
9052 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
9053 /// an AND to a vector_shuffle with the destination vector and a zero vector.
9054 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
9055 /// vector_shuffle V, Zero, <0, 4, 2, 4>
9056 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
9057 EVT VT = N->getValueType(0);
9058 DebugLoc dl = N->getDebugLoc();
9059 SDValue LHS = N->getOperand(0);
9060 SDValue RHS = N->getOperand(1);
9061 if (N->getOpcode() == ISD::AND) {
9062 if (RHS.getOpcode() == ISD::BITCAST)
9063 RHS = RHS.getOperand(0);
9064 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
9065 SmallVector<int, 8> Indices;
9066 unsigned NumElts = RHS.getNumOperands();
9067 for (unsigned i = 0; i != NumElts; ++i) {
9068 SDValue Elt = RHS.getOperand(i);
9069 if (!isa<ConstantSDNode>(Elt))
9072 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
9073 Indices.push_back(i);
9074 else if (cast<ConstantSDNode>(Elt)->isNullValue())
9075 Indices.push_back(NumElts);
9080 // Let's see if the target supports this vector_shuffle.
9081 EVT RVT = RHS.getValueType();
9082 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
9085 // Return the new VECTOR_SHUFFLE node.
9086 EVT EltVT = RVT.getVectorElementType();
9087 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
9088 DAG.getConstant(0, EltVT));
9089 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
9090 RVT, &ZeroOps[0], ZeroOps.size());
9091 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
9092 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
9093 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
9100 /// SimplifyVBinOp - Visit a binary vector operation, like ADD.
9101 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
9102 // After legalize, the target may be depending on adds and other
9103 // binary ops to provide legal ways to construct constants or other
9104 // things. Simplifying them may result in a loss of legality.
9105 if (LegalOperations) return SDValue();
9107 assert(N->getValueType(0).isVector() &&
9108 "SimplifyVBinOp only works on vectors!");
9110 SDValue LHS = N->getOperand(0);
9111 SDValue RHS = N->getOperand(1);
9112 SDValue Shuffle = XformToShuffleWithZero(N);
9113 if (Shuffle.getNode()) return Shuffle;
9115 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
9117 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
9118 RHS.getOpcode() == ISD::BUILD_VECTOR) {
9119 SmallVector<SDValue, 8> Ops;
9120 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
9121 SDValue LHSOp = LHS.getOperand(i);
9122 SDValue RHSOp = RHS.getOperand(i);
9123 // If these two elements can't be folded, bail out.
9124 if ((LHSOp.getOpcode() != ISD::UNDEF &&
9125 LHSOp.getOpcode() != ISD::Constant &&
9126 LHSOp.getOpcode() != ISD::ConstantFP) ||
9127 (RHSOp.getOpcode() != ISD::UNDEF &&
9128 RHSOp.getOpcode() != ISD::Constant &&
9129 RHSOp.getOpcode() != ISD::ConstantFP))
9132 // Can't fold divide by zero.
9133 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
9134 N->getOpcode() == ISD::FDIV) {
9135 if ((RHSOp.getOpcode() == ISD::Constant &&
9136 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
9137 (RHSOp.getOpcode() == ISD::ConstantFP &&
9138 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
9142 EVT VT = LHSOp.getValueType();
9143 EVT RVT = RHSOp.getValueType();
9145 // Integer BUILD_VECTOR operands may have types larger than the element
9146 // size (e.g., when the element type is not legal). Prior to type
9147 // legalization, the types may not match between the two BUILD_VECTORS.
9148 // Truncate one of the operands to make them match.
9149 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
9150 RHSOp = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, RHSOp);
9152 LHSOp = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), RVT, LHSOp);
9156 SDValue FoldOp = DAG.getNode(N->getOpcode(), LHS.getDebugLoc(), VT,
9158 if (FoldOp.getOpcode() != ISD::UNDEF &&
9159 FoldOp.getOpcode() != ISD::Constant &&
9160 FoldOp.getOpcode() != ISD::ConstantFP)
9162 Ops.push_back(FoldOp);
9163 AddToWorkList(FoldOp.getNode());
9166 if (Ops.size() == LHS.getNumOperands())
9167 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
9168 LHS.getValueType(), &Ops[0], Ops.size());
9174 /// SimplifyVUnaryOp - Visit a binary vector operation, like FABS/FNEG.
9175 SDValue DAGCombiner::SimplifyVUnaryOp(SDNode *N) {
9176 // After legalize, the target may be depending on adds and other
9177 // binary ops to provide legal ways to construct constants or other
9178 // things. Simplifying them may result in a loss of legality.
9179 if (LegalOperations) return SDValue();
9181 assert(N->getValueType(0).isVector() &&
9182 "SimplifyVUnaryOp only works on vectors!");
9184 SDValue N0 = N->getOperand(0);
9186 if (N0.getOpcode() != ISD::BUILD_VECTOR)
9189 // Operand is a BUILD_VECTOR node, see if we can constant fold it.
9190 SmallVector<SDValue, 8> Ops;
9191 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
9192 SDValue Op = N0.getOperand(i);
9193 if (Op.getOpcode() != ISD::UNDEF &&
9194 Op.getOpcode() != ISD::ConstantFP)
9196 EVT EltVT = Op.getValueType();
9197 SDValue FoldOp = DAG.getNode(N->getOpcode(), N0.getDebugLoc(), EltVT, Op);
9198 if (FoldOp.getOpcode() != ISD::UNDEF &&
9199 FoldOp.getOpcode() != ISD::ConstantFP)
9201 Ops.push_back(FoldOp);
9202 AddToWorkList(FoldOp.getNode());
9205 if (Ops.size() != N0.getNumOperands())
9208 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
9209 N0.getValueType(), &Ops[0], Ops.size());
9212 SDValue DAGCombiner::SimplifySelect(DebugLoc DL, SDValue N0,
9213 SDValue N1, SDValue N2){
9214 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
9216 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
9217 cast<CondCodeSDNode>(N0.getOperand(2))->get());
9219 // If we got a simplified select_cc node back from SimplifySelectCC, then
9220 // break it down into a new SETCC node, and a new SELECT node, and then return
9221 // the SELECT node, since we were called with a SELECT node.
9222 if (SCC.getNode()) {
9223 // Check to see if we got a select_cc back (to turn into setcc/select).
9224 // Otherwise, just return whatever node we got back, like fabs.
9225 if (SCC.getOpcode() == ISD::SELECT_CC) {
9226 SDValue SETCC = DAG.getNode(ISD::SETCC, N0.getDebugLoc(),
9228 SCC.getOperand(0), SCC.getOperand(1),
9230 AddToWorkList(SETCC.getNode());
9231 return DAG.getNode(ISD::SELECT, SCC.getDebugLoc(), SCC.getValueType(),
9232 SCC.getOperand(2), SCC.getOperand(3), SETCC);
9240 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
9241 /// are the two values being selected between, see if we can simplify the
9242 /// select. Callers of this should assume that TheSelect is deleted if this
9243 /// returns true. As such, they should return the appropriate thing (e.g. the
9244 /// node) back to the top-level of the DAG combiner loop to avoid it being
9246 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
9249 // Cannot simplify select with vector condition
9250 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
9252 // If this is a select from two identical things, try to pull the operation
9253 // through the select.
9254 if (LHS.getOpcode() != RHS.getOpcode() ||
9255 !LHS.hasOneUse() || !RHS.hasOneUse())
9258 // If this is a load and the token chain is identical, replace the select
9259 // of two loads with a load through a select of the address to load from.
9260 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
9261 // constants have been dropped into the constant pool.
9262 if (LHS.getOpcode() == ISD::LOAD) {
9263 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
9264 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
9266 // Token chains must be identical.
9267 if (LHS.getOperand(0) != RHS.getOperand(0) ||
9268 // Do not let this transformation reduce the number of volatile loads.
9269 LLD->isVolatile() || RLD->isVolatile() ||
9270 // If this is an EXTLOAD, the VT's must match.
9271 LLD->getMemoryVT() != RLD->getMemoryVT() ||
9272 // If this is an EXTLOAD, the kind of extension must match.
9273 (LLD->getExtensionType() != RLD->getExtensionType() &&
9274 // The only exception is if one of the extensions is anyext.
9275 LLD->getExtensionType() != ISD::EXTLOAD &&
9276 RLD->getExtensionType() != ISD::EXTLOAD) ||
9277 // FIXME: this discards src value information. This is
9278 // over-conservative. It would be beneficial to be able to remember
9279 // both potential memory locations. Since we are discarding
9280 // src value info, don't do the transformation if the memory
9281 // locations are not in the default address space.
9282 LLD->getPointerInfo().getAddrSpace() != 0 ||
9283 RLD->getPointerInfo().getAddrSpace() != 0)
9286 // Check that the select condition doesn't reach either load. If so,
9287 // folding this will induce a cycle into the DAG. If not, this is safe to
9288 // xform, so create a select of the addresses.
9290 if (TheSelect->getOpcode() == ISD::SELECT) {
9291 SDNode *CondNode = TheSelect->getOperand(0).getNode();
9292 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
9293 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
9295 // The loads must not depend on one another.
9296 if (LLD->isPredecessorOf(RLD) ||
9297 RLD->isPredecessorOf(LLD))
9299 Addr = DAG.getNode(ISD::SELECT, TheSelect->getDebugLoc(),
9300 LLD->getBasePtr().getValueType(),
9301 TheSelect->getOperand(0), LLD->getBasePtr(),
9303 } else { // Otherwise SELECT_CC
9304 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
9305 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
9307 if ((LLD->hasAnyUseOfValue(1) &&
9308 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
9309 (RLD->hasAnyUseOfValue(1) &&
9310 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
9313 Addr = DAG.getNode(ISD::SELECT_CC, TheSelect->getDebugLoc(),
9314 LLD->getBasePtr().getValueType(),
9315 TheSelect->getOperand(0),
9316 TheSelect->getOperand(1),
9317 LLD->getBasePtr(), RLD->getBasePtr(),
9318 TheSelect->getOperand(4));
9322 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
9323 Load = DAG.getLoad(TheSelect->getValueType(0),
9324 TheSelect->getDebugLoc(),
9325 // FIXME: Discards pointer info.
9326 LLD->getChain(), Addr, MachinePointerInfo(),
9327 LLD->isVolatile(), LLD->isNonTemporal(),
9328 LLD->isInvariant(), LLD->getAlignment());
9330 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
9331 RLD->getExtensionType() : LLD->getExtensionType(),
9332 TheSelect->getDebugLoc(),
9333 TheSelect->getValueType(0),
9334 // FIXME: Discards pointer info.
9335 LLD->getChain(), Addr, MachinePointerInfo(),
9336 LLD->getMemoryVT(), LLD->isVolatile(),
9337 LLD->isNonTemporal(), LLD->getAlignment());
9340 // Users of the select now use the result of the load.
9341 CombineTo(TheSelect, Load);
9343 // Users of the old loads now use the new load's chain. We know the
9344 // old-load value is dead now.
9345 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
9346 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
9353 /// SimplifySelectCC - Simplify an expression of the form (N0 cond N1) ? N2 : N3
9354 /// where 'cond' is the comparison specified by CC.
9355 SDValue DAGCombiner::SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1,
9356 SDValue N2, SDValue N3,
9357 ISD::CondCode CC, bool NotExtCompare) {
9358 // (x ? y : y) -> y.
9359 if (N2 == N3) return N2;
9361 EVT VT = N2.getValueType();
9362 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
9363 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
9364 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
9366 // Determine if the condition we're dealing with is constant
9367 SDValue SCC = SimplifySetCC(TLI.getSetCCResultType(N0.getValueType()),
9368 N0, N1, CC, DL, false);
9369 if (SCC.getNode()) AddToWorkList(SCC.getNode());
9370 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
9372 // fold select_cc true, x, y -> x
9373 if (SCCC && !SCCC->isNullValue())
9375 // fold select_cc false, x, y -> y
9376 if (SCCC && SCCC->isNullValue())
9379 // Check to see if we can simplify the select into an fabs node
9380 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
9381 // Allow either -0.0 or 0.0
9382 if (CFP->getValueAPF().isZero()) {
9383 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
9384 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
9385 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
9386 N2 == N3.getOperand(0))
9387 return DAG.getNode(ISD::FABS, DL, VT, N0);
9389 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
9390 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
9391 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
9392 N2.getOperand(0) == N3)
9393 return DAG.getNode(ISD::FABS, DL, VT, N3);
9397 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
9398 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
9399 // in it. This is a win when the constant is not otherwise available because
9400 // it replaces two constant pool loads with one. We only do this if the FP
9401 // type is known to be legal, because if it isn't, then we are before legalize
9402 // types an we want the other legalization to happen first (e.g. to avoid
9403 // messing with soft float) and if the ConstantFP is not legal, because if
9404 // it is legal, we may not need to store the FP constant in a constant pool.
9405 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
9406 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
9407 if (TLI.isTypeLegal(N2.getValueType()) &&
9408 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
9409 TargetLowering::Legal) &&
9410 // If both constants have multiple uses, then we won't need to do an
9411 // extra load, they are likely around in registers for other users.
9412 (TV->hasOneUse() || FV->hasOneUse())) {
9413 Constant *Elts[] = {
9414 const_cast<ConstantFP*>(FV->getConstantFPValue()),
9415 const_cast<ConstantFP*>(TV->getConstantFPValue())
9417 Type *FPTy = Elts[0]->getType();
9418 const DataLayout &TD = *TLI.getDataLayout();
9420 // Create a ConstantArray of the two constants.
9421 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
9422 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
9423 TD.getPrefTypeAlignment(FPTy));
9424 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
9426 // Get the offsets to the 0 and 1 element of the array so that we can
9427 // select between them.
9428 SDValue Zero = DAG.getIntPtrConstant(0);
9429 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
9430 SDValue One = DAG.getIntPtrConstant(EltSize);
9432 SDValue Cond = DAG.getSetCC(DL,
9433 TLI.getSetCCResultType(N0.getValueType()),
9435 AddToWorkList(Cond.getNode());
9436 SDValue CstOffset = DAG.getNode(ISD::SELECT, DL, Zero.getValueType(),
9438 AddToWorkList(CstOffset.getNode());
9439 CPIdx = DAG.getNode(ISD::ADD, DL, TLI.getPointerTy(), CPIdx,
9441 AddToWorkList(CPIdx.getNode());
9442 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
9443 MachinePointerInfo::getConstantPool(), false,
9444 false, false, Alignment);
9449 // Check to see if we can perform the "gzip trick", transforming
9450 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
9451 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
9452 (N1C->isNullValue() || // (a < 0) ? b : 0
9453 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
9454 EVT XType = N0.getValueType();
9455 EVT AType = N2.getValueType();
9456 if (XType.bitsGE(AType)) {
9457 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
9458 // single-bit constant.
9459 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
9460 unsigned ShCtV = N2C->getAPIntValue().logBase2();
9461 ShCtV = XType.getSizeInBits()-ShCtV-1;
9462 SDValue ShCt = DAG.getConstant(ShCtV,
9463 getShiftAmountTy(N0.getValueType()));
9464 SDValue Shift = DAG.getNode(ISD::SRL, N0.getDebugLoc(),
9466 AddToWorkList(Shift.getNode());
9468 if (XType.bitsGT(AType)) {
9469 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
9470 AddToWorkList(Shift.getNode());
9473 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
9476 SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(),
9478 DAG.getConstant(XType.getSizeInBits()-1,
9479 getShiftAmountTy(N0.getValueType())));
9480 AddToWorkList(Shift.getNode());
9482 if (XType.bitsGT(AType)) {
9483 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
9484 AddToWorkList(Shift.getNode());
9487 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
9491 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
9492 // where y is has a single bit set.
9493 // A plaintext description would be, we can turn the SELECT_CC into an AND
9494 // when the condition can be materialized as an all-ones register. Any
9495 // single bit-test can be materialized as an all-ones register with
9496 // shift-left and shift-right-arith.
9497 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
9498 N0->getValueType(0) == VT &&
9499 N1C && N1C->isNullValue() &&
9500 N2C && N2C->isNullValue()) {
9501 SDValue AndLHS = N0->getOperand(0);
9502 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
9503 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
9504 // Shift the tested bit over the sign bit.
9505 APInt AndMask = ConstAndRHS->getAPIntValue();
9507 DAG.getConstant(AndMask.countLeadingZeros(),
9508 getShiftAmountTy(AndLHS.getValueType()));
9509 SDValue Shl = DAG.getNode(ISD::SHL, N0.getDebugLoc(), VT, AndLHS, ShlAmt);
9511 // Now arithmetic right shift it all the way over, so the result is either
9512 // all-ones, or zero.
9514 DAG.getConstant(AndMask.getBitWidth()-1,
9515 getShiftAmountTy(Shl.getValueType()));
9516 SDValue Shr = DAG.getNode(ISD::SRA, N0.getDebugLoc(), VT, Shl, ShrAmt);
9518 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
9522 // fold select C, 16, 0 -> shl C, 4
9523 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
9524 TLI.getBooleanContents(N0.getValueType().isVector()) ==
9525 TargetLowering::ZeroOrOneBooleanContent) {
9527 // If the caller doesn't want us to simplify this into a zext of a compare,
9529 if (NotExtCompare && N2C->getAPIntValue() == 1)
9532 // Get a SetCC of the condition
9533 // NOTE: Don't create a SETCC if it's not legal on this target.
9534 if (!LegalOperations ||
9535 TLI.isOperationLegal(ISD::SETCC,
9536 LegalTypes ? TLI.getSetCCResultType(N0.getValueType()) : MVT::i1)) {
9538 // cast from setcc result type to select result type
9540 SCC = DAG.getSetCC(DL, TLI.getSetCCResultType(N0.getValueType()),
9542 if (N2.getValueType().bitsLT(SCC.getValueType()))
9543 Temp = DAG.getZeroExtendInReg(SCC, N2.getDebugLoc(),
9546 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getDebugLoc(),
9547 N2.getValueType(), SCC);
9549 SCC = DAG.getSetCC(N0.getDebugLoc(), MVT::i1, N0, N1, CC);
9550 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getDebugLoc(),
9551 N2.getValueType(), SCC);
9554 AddToWorkList(SCC.getNode());
9555 AddToWorkList(Temp.getNode());
9557 if (N2C->getAPIntValue() == 1)
9560 // shl setcc result by log2 n2c
9561 return DAG.getNode(ISD::SHL, DL, N2.getValueType(), Temp,
9562 DAG.getConstant(N2C->getAPIntValue().logBase2(),
9563 getShiftAmountTy(Temp.getValueType())));
9567 // Check to see if this is the equivalent of setcc
9568 // FIXME: Turn all of these into setcc if setcc if setcc is legal
9569 // otherwise, go ahead with the folds.
9570 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
9571 EVT XType = N0.getValueType();
9572 if (!LegalOperations ||
9573 TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(XType))) {
9574 SDValue Res = DAG.getSetCC(DL, TLI.getSetCCResultType(XType), N0, N1, CC);
9575 if (Res.getValueType() != VT)
9576 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
9580 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
9581 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
9582 (!LegalOperations ||
9583 TLI.isOperationLegal(ISD::CTLZ, XType))) {
9584 SDValue Ctlz = DAG.getNode(ISD::CTLZ, N0.getDebugLoc(), XType, N0);
9585 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
9586 DAG.getConstant(Log2_32(XType.getSizeInBits()),
9587 getShiftAmountTy(Ctlz.getValueType())));
9589 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
9590 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
9591 SDValue NegN0 = DAG.getNode(ISD::SUB, N0.getDebugLoc(),
9592 XType, DAG.getConstant(0, XType), N0);
9593 SDValue NotN0 = DAG.getNOT(N0.getDebugLoc(), N0, XType);
9594 return DAG.getNode(ISD::SRL, DL, XType,
9595 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
9596 DAG.getConstant(XType.getSizeInBits()-1,
9597 getShiftAmountTy(XType)));
9599 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
9600 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
9601 SDValue Sign = DAG.getNode(ISD::SRL, N0.getDebugLoc(), XType, N0,
9602 DAG.getConstant(XType.getSizeInBits()-1,
9603 getShiftAmountTy(N0.getValueType())));
9604 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
9608 // Check to see if this is an integer abs.
9609 // select_cc setg[te] X, 0, X, -X ->
9610 // select_cc setgt X, -1, X, -X ->
9611 // select_cc setl[te] X, 0, -X, X ->
9612 // select_cc setlt X, 1, -X, X ->
9613 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
9615 ConstantSDNode *SubC = NULL;
9616 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
9617 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
9618 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
9619 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
9620 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
9621 (N1C->isOne() && CC == ISD::SETLT)) &&
9622 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
9623 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
9625 EVT XType = N0.getValueType();
9626 if (SubC && SubC->isNullValue() && XType.isInteger()) {
9627 SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(), XType,
9629 DAG.getConstant(XType.getSizeInBits()-1,
9630 getShiftAmountTy(N0.getValueType())));
9631 SDValue Add = DAG.getNode(ISD::ADD, N0.getDebugLoc(),
9633 AddToWorkList(Shift.getNode());
9634 AddToWorkList(Add.getNode());
9635 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
9642 /// SimplifySetCC - This is a stub for TargetLowering::SimplifySetCC.
9643 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
9644 SDValue N1, ISD::CondCode Cond,
9645 DebugLoc DL, bool foldBooleans) {
9646 TargetLowering::DAGCombinerInfo
9647 DagCombineInfo(DAG, !LegalTypes, !LegalOperations, false, this);
9648 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
9651 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
9652 /// return a DAG expression to select that will generate the same value by
9653 /// multiplying by a magic number. See:
9654 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
9655 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
9656 std::vector<SDNode*> Built;
9657 SDValue S = TLI.BuildSDIV(N, DAG, LegalOperations, &Built);
9659 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
9665 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
9666 /// return a DAG expression to select that will generate the same value by
9667 /// multiplying by a magic number. See:
9668 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
9669 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
9670 std::vector<SDNode*> Built;
9671 SDValue S = TLI.BuildUDIV(N, DAG, LegalOperations, &Built);
9673 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
9679 /// FindBaseOffset - Return true if base is a frame index, which is known not
9680 // to alias with anything but itself. Provides base object and offset as
9682 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
9683 const GlobalValue *&GV, const void *&CV) {
9684 // Assume it is a primitive operation.
9685 Base = Ptr; Offset = 0; GV = 0; CV = 0;
9687 // If it's an adding a simple constant then integrate the offset.
9688 if (Base.getOpcode() == ISD::ADD) {
9689 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
9690 Base = Base.getOperand(0);
9691 Offset += C->getZExtValue();
9695 // Return the underlying GlobalValue, and update the Offset. Return false
9696 // for GlobalAddressSDNode since the same GlobalAddress may be represented
9697 // by multiple nodes with different offsets.
9698 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
9699 GV = G->getGlobal();
9700 Offset += G->getOffset();
9704 // Return the underlying Constant value, and update the Offset. Return false
9705 // for ConstantSDNodes since the same constant pool entry may be represented
9706 // by multiple nodes with different offsets.
9707 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
9708 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
9709 : (const void *)C->getConstVal();
9710 Offset += C->getOffset();
9713 // If it's any of the following then it can't alias with anything but itself.
9714 return isa<FrameIndexSDNode>(Base);
9717 /// isAlias - Return true if there is any possibility that the two addresses
9719 bool DAGCombiner::isAlias(SDValue Ptr1, int64_t Size1,
9720 const Value *SrcValue1, int SrcValueOffset1,
9721 unsigned SrcValueAlign1,
9722 const MDNode *TBAAInfo1,
9723 SDValue Ptr2, int64_t Size2,
9724 const Value *SrcValue2, int SrcValueOffset2,
9725 unsigned SrcValueAlign2,
9726 const MDNode *TBAAInfo2) const {
9727 // If they are the same then they must be aliases.
9728 if (Ptr1 == Ptr2) return true;
9730 // Gather base node and offset information.
9731 SDValue Base1, Base2;
9732 int64_t Offset1, Offset2;
9733 const GlobalValue *GV1, *GV2;
9734 const void *CV1, *CV2;
9735 bool isFrameIndex1 = FindBaseOffset(Ptr1, Base1, Offset1, GV1, CV1);
9736 bool isFrameIndex2 = FindBaseOffset(Ptr2, Base2, Offset2, GV2, CV2);
9738 // If they have a same base address then check to see if they overlap.
9739 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
9740 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
9742 // It is possible for different frame indices to alias each other, mostly
9743 // when tail call optimization reuses return address slots for arguments.
9744 // To catch this case, look up the actual index of frame indices to compute
9745 // the real alias relationship.
9746 if (isFrameIndex1 && isFrameIndex2) {
9747 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
9748 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
9749 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
9750 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
9753 // Otherwise, if we know what the bases are, and they aren't identical, then
9754 // we know they cannot alias.
9755 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
9758 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
9759 // compared to the size and offset of the access, we may be able to prove they
9760 // do not alias. This check is conservative for now to catch cases created by
9761 // splitting vector types.
9762 if ((SrcValueAlign1 == SrcValueAlign2) &&
9763 (SrcValueOffset1 != SrcValueOffset2) &&
9764 (Size1 == Size2) && (SrcValueAlign1 > Size1)) {
9765 int64_t OffAlign1 = SrcValueOffset1 % SrcValueAlign1;
9766 int64_t OffAlign2 = SrcValueOffset2 % SrcValueAlign1;
9768 // There is no overlap between these relatively aligned accesses of similar
9769 // size, return no alias.
9770 if ((OffAlign1 + Size1) <= OffAlign2 || (OffAlign2 + Size2) <= OffAlign1)
9774 if (CombinerGlobalAA) {
9775 // Use alias analysis information.
9776 int64_t MinOffset = std::min(SrcValueOffset1, SrcValueOffset2);
9777 int64_t Overlap1 = Size1 + SrcValueOffset1 - MinOffset;
9778 int64_t Overlap2 = Size2 + SrcValueOffset2 - MinOffset;
9779 AliasAnalysis::AliasResult AAResult =
9780 AA.alias(AliasAnalysis::Location(SrcValue1, Overlap1, TBAAInfo1),
9781 AliasAnalysis::Location(SrcValue2, Overlap2, TBAAInfo2));
9782 if (AAResult == AliasAnalysis::NoAlias)
9786 // Otherwise we have to assume they alias.
9790 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) {
9792 int64_t Size0, Size1;
9793 const Value *SrcValue0, *SrcValue1;
9794 int SrcValueOffset0, SrcValueOffset1;
9795 unsigned SrcValueAlign0, SrcValueAlign1;
9796 const MDNode *SrcTBAAInfo0, *SrcTBAAInfo1;
9797 FindAliasInfo(Op0, Ptr0, Size0, SrcValue0, SrcValueOffset0,
9798 SrcValueAlign0, SrcTBAAInfo0);
9799 FindAliasInfo(Op1, Ptr1, Size1, SrcValue1, SrcValueOffset1,
9800 SrcValueAlign1, SrcTBAAInfo1);
9801 return isAlias(Ptr0, Size0, SrcValue0, SrcValueOffset0,
9802 SrcValueAlign0, SrcTBAAInfo0,
9803 Ptr1, Size1, SrcValue1, SrcValueOffset1,
9804 SrcValueAlign1, SrcTBAAInfo1);
9807 /// FindAliasInfo - Extracts the relevant alias information from the memory
9808 /// node. Returns true if the operand was a load.
9809 bool DAGCombiner::FindAliasInfo(SDNode *N,
9810 SDValue &Ptr, int64_t &Size,
9811 const Value *&SrcValue,
9812 int &SrcValueOffset,
9813 unsigned &SrcValueAlign,
9814 const MDNode *&TBAAInfo) const {
9815 LSBaseSDNode *LS = cast<LSBaseSDNode>(N);
9817 Ptr = LS->getBasePtr();
9818 Size = LS->getMemoryVT().getSizeInBits() >> 3;
9819 SrcValue = LS->getSrcValue();
9820 SrcValueOffset = LS->getSrcValueOffset();
9821 SrcValueAlign = LS->getOriginalAlignment();
9822 TBAAInfo = LS->getTBAAInfo();
9823 return isa<LoadSDNode>(LS);
9826 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
9827 /// looking for aliasing nodes and adding them to the Aliases vector.
9828 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
9829 SmallVector<SDValue, 8> &Aliases) {
9830 SmallVector<SDValue, 8> Chains; // List of chains to visit.
9831 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
9833 // Get alias information for node.
9836 const Value *SrcValue;
9838 unsigned SrcValueAlign;
9839 const MDNode *SrcTBAAInfo;
9840 bool IsLoad = FindAliasInfo(N, Ptr, Size, SrcValue, SrcValueOffset,
9841 SrcValueAlign, SrcTBAAInfo);
9844 Chains.push_back(OriginalChain);
9847 // Look at each chain and determine if it is an alias. If so, add it to the
9848 // aliases list. If not, then continue up the chain looking for the next
9850 while (!Chains.empty()) {
9851 SDValue Chain = Chains.back();
9854 // For TokenFactor nodes, look at each operand and only continue up the
9855 // chain until we find two aliases. If we've seen two aliases, assume we'll
9856 // find more and revert to original chain since the xform is unlikely to be
9859 // FIXME: The depth check could be made to return the last non-aliasing
9860 // chain we found before we hit a tokenfactor rather than the original
9862 if (Depth > 6 || Aliases.size() == 2) {
9864 Aliases.push_back(OriginalChain);
9868 // Don't bother if we've been before.
9869 if (!Visited.insert(Chain.getNode()))
9872 switch (Chain.getOpcode()) {
9873 case ISD::EntryToken:
9874 // Entry token is ideal chain operand, but handled in FindBetterChain.
9879 // Get alias information for Chain.
9882 const Value *OpSrcValue;
9883 int OpSrcValueOffset;
9884 unsigned OpSrcValueAlign;
9885 const MDNode *OpSrcTBAAInfo;
9886 bool IsOpLoad = FindAliasInfo(Chain.getNode(), OpPtr, OpSize,
9887 OpSrcValue, OpSrcValueOffset,
9891 // If chain is alias then stop here.
9892 if (!(IsLoad && IsOpLoad) &&
9893 isAlias(Ptr, Size, SrcValue, SrcValueOffset, SrcValueAlign,
9895 OpPtr, OpSize, OpSrcValue, OpSrcValueOffset,
9896 OpSrcValueAlign, OpSrcTBAAInfo)) {
9897 Aliases.push_back(Chain);
9899 // Look further up the chain.
9900 Chains.push_back(Chain.getOperand(0));
9906 case ISD::TokenFactor:
9907 // We have to check each of the operands of the token factor for "small"
9908 // token factors, so we queue them up. Adding the operands to the queue
9909 // (stack) in reverse order maintains the original order and increases the
9910 // likelihood that getNode will find a matching token factor (CSE.)
9911 if (Chain.getNumOperands() > 16) {
9912 Aliases.push_back(Chain);
9915 for (unsigned n = Chain.getNumOperands(); n;)
9916 Chains.push_back(Chain.getOperand(--n));
9921 // For all other instructions we will just have to take what we can get.
9922 Aliases.push_back(Chain);
9928 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
9929 /// for a better chain (aliasing node.)
9930 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
9931 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
9933 // Accumulate all the aliases to this node.
9934 GatherAllAliases(N, OldChain, Aliases);
9936 // If no operands then chain to entry token.
9937 if (Aliases.size() == 0)
9938 return DAG.getEntryNode();
9940 // If a single operand then chain to it. We don't need to revisit it.
9941 if (Aliases.size() == 1)
9944 // Construct a custom tailored token factor.
9945 return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
9946 &Aliases[0], Aliases.size());
9949 // SelectionDAG::Combine - This is the entry point for the file.
9951 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
9952 CodeGenOpt::Level OptLevel) {
9953 /// run - This is the main entry point to this class.
9955 DAGCombiner(*this, AA, OptLevel).Run(Level);