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 || TLI.isCondCodeLegal(Result, LL.getValueType())))
2636 return DAG.getSetCC(N->getDebugLoc(), N0.getValueType(),
2641 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
2642 if (N0.getOpcode() == N1.getOpcode()) {
2643 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2644 if (Tmp.getNode()) return Tmp;
2647 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
2648 // fold (and (sra)) -> (and (srl)) when possible.
2649 if (!VT.isVector() &&
2650 SimplifyDemandedBits(SDValue(N, 0)))
2651 return SDValue(N, 0);
2653 // fold (zext_inreg (extload x)) -> (zextload x)
2654 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
2655 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2656 EVT MemVT = LN0->getMemoryVT();
2657 // If we zero all the possible extended bits, then we can turn this into
2658 // a zextload if we are running before legalize or the operation is legal.
2659 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2660 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2661 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2662 ((!LegalOperations && !LN0->isVolatile()) ||
2663 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2664 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
2665 LN0->getChain(), LN0->getBasePtr(),
2666 LN0->getPointerInfo(), MemVT,
2667 LN0->isVolatile(), LN0->isNonTemporal(),
2668 LN0->getAlignment());
2670 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2671 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2674 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
2675 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
2677 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2678 EVT MemVT = LN0->getMemoryVT();
2679 // If we zero all the possible extended bits, then we can turn this into
2680 // a zextload if we are running before legalize or the operation is legal.
2681 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2682 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2683 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2684 ((!LegalOperations && !LN0->isVolatile()) ||
2685 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2686 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
2688 LN0->getBasePtr(), LN0->getPointerInfo(),
2690 LN0->isVolatile(), LN0->isNonTemporal(),
2691 LN0->getAlignment());
2693 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2694 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2698 // fold (and (load x), 255) -> (zextload x, i8)
2699 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2700 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2701 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2702 (N0.getOpcode() == ISD::ANY_EXTEND &&
2703 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2704 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2705 LoadSDNode *LN0 = HasAnyExt
2706 ? cast<LoadSDNode>(N0.getOperand(0))
2707 : cast<LoadSDNode>(N0);
2708 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2709 LN0->isUnindexed() && N0.hasOneUse() && LN0->hasOneUse()) {
2710 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2711 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2712 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2713 EVT LoadedVT = LN0->getMemoryVT();
2715 if (ExtVT == LoadedVT &&
2716 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2717 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2720 DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy,
2721 LN0->getChain(), LN0->getBasePtr(),
2722 LN0->getPointerInfo(),
2723 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2724 LN0->getAlignment());
2726 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
2727 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2730 // Do not change the width of a volatile load.
2731 // Do not generate loads of non-round integer types since these can
2732 // be expensive (and would be wrong if the type is not byte sized).
2733 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
2734 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2735 EVT PtrType = LN0->getOperand(1).getValueType();
2737 unsigned Alignment = LN0->getAlignment();
2738 SDValue NewPtr = LN0->getBasePtr();
2740 // For big endian targets, we need to add an offset to the pointer
2741 // to load the correct bytes. For little endian systems, we merely
2742 // need to read fewer bytes from the same pointer.
2743 if (TLI.isBigEndian()) {
2744 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
2745 unsigned EVTStoreBytes = ExtVT.getStoreSize();
2746 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
2747 NewPtr = DAG.getNode(ISD::ADD, LN0->getDebugLoc(), PtrType,
2748 NewPtr, DAG.getConstant(PtrOff, PtrType));
2749 Alignment = MinAlign(Alignment, PtrOff);
2752 AddToWorkList(NewPtr.getNode());
2754 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2756 DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy,
2757 LN0->getChain(), NewPtr,
2758 LN0->getPointerInfo(),
2759 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2762 CombineTo(LN0, Load, Load.getValue(1));
2763 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2769 if (N0.getOpcode() == ISD::ADD && N1.getOpcode() == ISD::SRL &&
2770 VT.getSizeInBits() <= 64) {
2771 if (ConstantSDNode *ADDI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
2772 APInt ADDC = ADDI->getAPIntValue();
2773 if (!TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2774 // Look for (and (add x, c1), (lshr y, c2)). If C1 wasn't a legal
2775 // immediate for an add, but it is legal if its top c2 bits are set,
2776 // transform the ADD so the immediate doesn't need to be materialized
2778 if (ConstantSDNode *SRLI = dyn_cast<ConstantSDNode>(N1.getOperand(1))) {
2779 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
2780 SRLI->getZExtValue());
2781 if (DAG.MaskedValueIsZero(N0.getOperand(1), Mask)) {
2783 if (TLI.isLegalAddImmediate(ADDC.getSExtValue())) {
2785 DAG.getNode(ISD::ADD, N0.getDebugLoc(), VT,
2786 N0.getOperand(0), DAG.getConstant(ADDC, VT));
2787 CombineTo(N0.getNode(), NewAdd);
2788 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2799 /// MatchBSwapHWord - Match (a >> 8) | (a << 8) as (bswap a) >> 16
2801 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
2802 bool DemandHighBits) {
2803 if (!LegalOperations)
2806 EVT VT = N->getValueType(0);
2807 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
2809 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
2812 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
2813 bool LookPassAnd0 = false;
2814 bool LookPassAnd1 = false;
2815 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
2817 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
2819 if (N0.getOpcode() == ISD::AND) {
2820 if (!N0.getNode()->hasOneUse())
2822 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2823 if (!N01C || N01C->getZExtValue() != 0xFF00)
2825 N0 = N0.getOperand(0);
2826 LookPassAnd0 = true;
2829 if (N1.getOpcode() == ISD::AND) {
2830 if (!N1.getNode()->hasOneUse())
2832 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
2833 if (!N11C || N11C->getZExtValue() != 0xFF)
2835 N1 = N1.getOperand(0);
2836 LookPassAnd1 = true;
2839 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
2841 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
2843 if (!N0.getNode()->hasOneUse() ||
2844 !N1.getNode()->hasOneUse())
2847 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2848 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
2851 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
2854 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
2855 SDValue N00 = N0->getOperand(0);
2856 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
2857 if (!N00.getNode()->hasOneUse())
2859 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
2860 if (!N001C || N001C->getZExtValue() != 0xFF)
2862 N00 = N00.getOperand(0);
2863 LookPassAnd0 = true;
2866 SDValue N10 = N1->getOperand(0);
2867 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
2868 if (!N10.getNode()->hasOneUse())
2870 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
2871 if (!N101C || N101C->getZExtValue() != 0xFF00)
2873 N10 = N10.getOperand(0);
2874 LookPassAnd1 = true;
2880 // Make sure everything beyond the low halfword is zero since the SRL 16
2881 // will clear the top bits.
2882 unsigned OpSizeInBits = VT.getSizeInBits();
2883 if (DemandHighBits && OpSizeInBits > 16 &&
2884 (!LookPassAnd0 || !LookPassAnd1) &&
2885 !DAG.MaskedValueIsZero(N10, APInt::getHighBitsSet(OpSizeInBits, 16)))
2888 SDValue Res = DAG.getNode(ISD::BSWAP, N->getDebugLoc(), VT, N00);
2889 if (OpSizeInBits > 16)
2890 Res = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, Res,
2891 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
2895 /// isBSwapHWordElement - Return true if the specified node is an element
2896 /// that makes up a 32-bit packed halfword byteswap. i.e.
2897 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
2898 static bool isBSwapHWordElement(SDValue N, SmallVector<SDNode*,4> &Parts) {
2899 if (!N.getNode()->hasOneUse())
2902 unsigned Opc = N.getOpcode();
2903 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
2906 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2911 switch (N1C->getZExtValue()) {
2914 case 0xFF: Num = 0; break;
2915 case 0xFF00: Num = 1; break;
2916 case 0xFF0000: Num = 2; break;
2917 case 0xFF000000: Num = 3; break;
2920 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
2921 SDValue N0 = N.getOperand(0);
2922 if (Opc == ISD::AND) {
2923 if (Num == 0 || Num == 2) {
2925 // (x >> 8) & 0xff0000
2926 if (N0.getOpcode() != ISD::SRL)
2928 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2929 if (!C || C->getZExtValue() != 8)
2932 // (x << 8) & 0xff00
2933 // (x << 8) & 0xff000000
2934 if (N0.getOpcode() != ISD::SHL)
2936 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2937 if (!C || C->getZExtValue() != 8)
2940 } else if (Opc == ISD::SHL) {
2942 // (x & 0xff0000) << 8
2943 if (Num != 0 && Num != 2)
2945 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2946 if (!C || C->getZExtValue() != 8)
2948 } else { // Opc == ISD::SRL
2949 // (x & 0xff00) >> 8
2950 // (x & 0xff000000) >> 8
2951 if (Num != 1 && Num != 3)
2953 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2954 if (!C || C->getZExtValue() != 8)
2961 Parts[Num] = N0.getOperand(0).getNode();
2965 /// MatchBSwapHWord - Match a 32-bit packed halfword bswap. That is
2966 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
2967 /// => (rotl (bswap x), 16)
2968 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
2969 if (!LegalOperations)
2972 EVT VT = N->getValueType(0);
2975 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
2978 SmallVector<SDNode*,4> Parts(4, (SDNode*)0);
2980 // (or (or (and), (and)), (or (and), (and)))
2981 // (or (or (or (and), (and)), (and)), (and))
2982 if (N0.getOpcode() != ISD::OR)
2984 SDValue N00 = N0.getOperand(0);
2985 SDValue N01 = N0.getOperand(1);
2987 if (N1.getOpcode() == ISD::OR) {
2988 // (or (or (and), (and)), (or (and), (and)))
2989 SDValue N000 = N00.getOperand(0);
2990 if (!isBSwapHWordElement(N000, Parts))
2993 SDValue N001 = N00.getOperand(1);
2994 if (!isBSwapHWordElement(N001, Parts))
2996 SDValue N010 = N01.getOperand(0);
2997 if (!isBSwapHWordElement(N010, Parts))
2999 SDValue N011 = N01.getOperand(1);
3000 if (!isBSwapHWordElement(N011, Parts))
3003 // (or (or (or (and), (and)), (and)), (and))
3004 if (!isBSwapHWordElement(N1, Parts))
3006 if (!isBSwapHWordElement(N01, Parts))
3008 if (N00.getOpcode() != ISD::OR)
3010 SDValue N000 = N00.getOperand(0);
3011 if (!isBSwapHWordElement(N000, Parts))
3013 SDValue N001 = N00.getOperand(1);
3014 if (!isBSwapHWordElement(N001, Parts))
3018 // Make sure the parts are all coming from the same node.
3019 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
3022 SDValue BSwap = DAG.getNode(ISD::BSWAP, N->getDebugLoc(), VT,
3023 SDValue(Parts[0],0));
3025 // Result of the bswap should be rotated by 16. If it's not legal, than
3026 // do (x << 16) | (x >> 16).
3027 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
3028 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
3029 return DAG.getNode(ISD::ROTL, N->getDebugLoc(), VT, BSwap, ShAmt);
3030 if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
3031 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, BSwap, ShAmt);
3032 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT,
3033 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, BSwap, ShAmt),
3034 DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, BSwap, ShAmt));
3037 SDValue DAGCombiner::visitOR(SDNode *N) {
3038 SDValue N0 = N->getOperand(0);
3039 SDValue N1 = N->getOperand(1);
3040 SDValue LL, LR, RL, RR, CC0, CC1;
3041 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3042 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3043 EVT VT = N1.getValueType();
3046 if (VT.isVector()) {
3047 SDValue FoldedVOp = SimplifyVBinOp(N);
3048 if (FoldedVOp.getNode()) return FoldedVOp;
3050 // fold (or x, 0) -> x, vector edition
3051 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3053 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3056 // fold (or x, -1) -> -1, vector edition
3057 if (ISD::isBuildVectorAllOnes(N0.getNode()))
3059 if (ISD::isBuildVectorAllOnes(N1.getNode()))
3063 // fold (or x, undef) -> -1
3064 if (!LegalOperations &&
3065 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
3066 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
3067 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
3069 // fold (or c1, c2) -> c1|c2
3071 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
3072 // canonicalize constant to RHS
3074 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N1, N0);
3075 // fold (or x, 0) -> x
3076 if (N1C && N1C->isNullValue())
3078 // fold (or x, -1) -> -1
3079 if (N1C && N1C->isAllOnesValue())
3081 // fold (or x, c) -> c iff (x & ~c) == 0
3082 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
3085 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
3086 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
3087 if (BSwap.getNode() != 0)
3089 BSwap = MatchBSwapHWordLow(N, N0, N1);
3090 if (BSwap.getNode() != 0)
3094 SDValue ROR = ReassociateOps(ISD::OR, N->getDebugLoc(), N0, N1);
3095 if (ROR.getNode() != 0)
3097 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3098 // iff (c1 & c2) == 0.
3099 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3100 isa<ConstantSDNode>(N0.getOperand(1))) {
3101 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3102 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0)
3103 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
3104 DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
3105 N0.getOperand(0), N1),
3106 DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1));
3108 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3109 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3110 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3111 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3113 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3114 LL.getValueType().isInteger()) {
3115 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3116 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3117 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3118 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3119 SDValue ORNode = DAG.getNode(ISD::OR, LR.getDebugLoc(),
3120 LR.getValueType(), LL, RL);
3121 AddToWorkList(ORNode.getNode());
3122 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
3124 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3125 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3126 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3127 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3128 SDValue ANDNode = DAG.getNode(ISD::AND, LR.getDebugLoc(),
3129 LR.getValueType(), LL, RL);
3130 AddToWorkList(ANDNode.getNode());
3131 return DAG.getSetCC(N->getDebugLoc(), VT, ANDNode, LR, Op1);
3134 // canonicalize equivalent to ll == rl
3135 if (LL == RR && LR == RL) {
3136 Op1 = ISD::getSetCCSwappedOperands(Op1);
3139 if (LL == RL && LR == RR) {
3140 bool isInteger = LL.getValueType().isInteger();
3141 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3142 if (Result != ISD::SETCC_INVALID &&
3143 (!LegalOperations || TLI.isCondCodeLegal(Result, LL.getValueType())))
3144 return DAG.getSetCC(N->getDebugLoc(), N0.getValueType(),
3149 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3150 if (N0.getOpcode() == N1.getOpcode()) {
3151 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3152 if (Tmp.getNode()) return Tmp;
3155 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3156 if (N0.getOpcode() == ISD::AND &&
3157 N1.getOpcode() == ISD::AND &&
3158 N0.getOperand(1).getOpcode() == ISD::Constant &&
3159 N1.getOperand(1).getOpcode() == ISD::Constant &&
3160 // Don't increase # computations.
3161 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3162 // We can only do this xform if we know that bits from X that are set in C2
3163 // but not in C1 are already zero. Likewise for Y.
3164 const APInt &LHSMask =
3165 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3166 const APInt &RHSMask =
3167 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3169 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3170 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3171 SDValue X = DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
3172 N0.getOperand(0), N1.getOperand(0));
3173 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, X,
3174 DAG.getConstant(LHSMask | RHSMask, VT));
3178 // See if this is some rotate idiom.
3179 if (SDNode *Rot = MatchRotate(N0, N1, N->getDebugLoc()))
3180 return SDValue(Rot, 0);
3182 // Simplify the operands using demanded-bits information.
3183 if (!VT.isVector() &&
3184 SimplifyDemandedBits(SDValue(N, 0)))
3185 return SDValue(N, 0);
3190 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
3191 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3192 if (Op.getOpcode() == ISD::AND) {
3193 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3194 Mask = Op.getOperand(1);
3195 Op = Op.getOperand(0);
3201 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3209 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3210 // idioms for rotate, and if the target supports rotation instructions, generate
3212 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, DebugLoc DL) {
3213 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3214 EVT VT = LHS.getValueType();
3215 if (!TLI.isTypeLegal(VT)) return 0;
3217 // The target must have at least one rotate flavor.
3218 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3219 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3220 if (!HasROTL && !HasROTR) return 0;
3222 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3223 SDValue LHSShift; // The shift.
3224 SDValue LHSMask; // AND value if any.
3225 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3226 return 0; // Not part of a rotate.
3228 SDValue RHSShift; // The shift.
3229 SDValue RHSMask; // AND value if any.
3230 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3231 return 0; // Not part of a rotate.
3233 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3234 return 0; // Not shifting the same value.
3236 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3237 return 0; // Shifts must disagree.
3239 // Canonicalize shl to left side in a shl/srl pair.
3240 if (RHSShift.getOpcode() == ISD::SHL) {
3241 std::swap(LHS, RHS);
3242 std::swap(LHSShift, RHSShift);
3243 std::swap(LHSMask , RHSMask );
3246 unsigned OpSizeInBits = VT.getSizeInBits();
3247 SDValue LHSShiftArg = LHSShift.getOperand(0);
3248 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3249 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3251 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3252 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3253 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3254 RHSShiftAmt.getOpcode() == ISD::Constant) {
3255 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3256 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3257 if ((LShVal + RShVal) != OpSizeInBits)
3260 SDValue Rot = DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3261 LHSShiftArg, HasROTL ? LHSShiftAmt : RHSShiftAmt);
3263 // If there is an AND of either shifted operand, apply it to the result.
3264 if (LHSMask.getNode() || RHSMask.getNode()) {
3265 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3267 if (LHSMask.getNode()) {
3268 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3269 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3271 if (RHSMask.getNode()) {
3272 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3273 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3276 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3279 return Rot.getNode();
3282 // If there is a mask here, and we have a variable shift, we can't be sure
3283 // that we're masking out the right stuff.
3284 if (LHSMask.getNode() || RHSMask.getNode())
3287 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y)
3288 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotr x, (sub 32, y))
3289 if (RHSShiftAmt.getOpcode() == ISD::SUB &&
3290 LHSShiftAmt == RHSShiftAmt.getOperand(1)) {
3291 if (ConstantSDNode *SUBC =
3292 dyn_cast<ConstantSDNode>(RHSShiftAmt.getOperand(0))) {
3293 if (SUBC->getAPIntValue() == OpSizeInBits) {
3294 return DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT, LHSShiftArg,
3295 HasROTL ? LHSShiftAmt : RHSShiftAmt).getNode();
3300 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y)
3301 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotl x, (sub 32, y))
3302 if (LHSShiftAmt.getOpcode() == ISD::SUB &&
3303 RHSShiftAmt == LHSShiftAmt.getOperand(1)) {
3304 if (ConstantSDNode *SUBC =
3305 dyn_cast<ConstantSDNode>(LHSShiftAmt.getOperand(0))) {
3306 if (SUBC->getAPIntValue() == OpSizeInBits) {
3307 return DAG.getNode(HasROTR ? ISD::ROTR : ISD::ROTL, DL, VT, LHSShiftArg,
3308 HasROTR ? RHSShiftAmt : LHSShiftAmt).getNode();
3313 // Look for sign/zext/any-extended or truncate cases:
3314 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3315 LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3316 LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3317 LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3318 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND ||
3319 RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND ||
3320 RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND ||
3321 RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3322 SDValue LExtOp0 = LHSShiftAmt.getOperand(0);
3323 SDValue RExtOp0 = RHSShiftAmt.getOperand(0);
3324 if (RExtOp0.getOpcode() == ISD::SUB &&
3325 RExtOp0.getOperand(1) == LExtOp0) {
3326 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
3328 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
3329 // (rotr x, (sub 32, y))
3330 if (ConstantSDNode *SUBC =
3331 dyn_cast<ConstantSDNode>(RExtOp0.getOperand(0))) {
3332 if (SUBC->getAPIntValue() == OpSizeInBits) {
3333 return DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3335 HasROTL ? LHSShiftAmt : RHSShiftAmt).getNode();
3338 } else if (LExtOp0.getOpcode() == ISD::SUB &&
3339 RExtOp0 == LExtOp0.getOperand(1)) {
3340 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
3342 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
3343 // (rotl x, (sub 32, y))
3344 if (ConstantSDNode *SUBC =
3345 dyn_cast<ConstantSDNode>(LExtOp0.getOperand(0))) {
3346 if (SUBC->getAPIntValue() == OpSizeInBits) {
3347 return DAG.getNode(HasROTR ? ISD::ROTR : ISD::ROTL, DL, VT,
3349 HasROTR ? RHSShiftAmt : LHSShiftAmt).getNode();
3358 SDValue DAGCombiner::visitXOR(SDNode *N) {
3359 SDValue N0 = N->getOperand(0);
3360 SDValue N1 = N->getOperand(1);
3361 SDValue LHS, RHS, CC;
3362 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3363 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3364 EVT VT = N0.getValueType();
3367 if (VT.isVector()) {
3368 SDValue FoldedVOp = SimplifyVBinOp(N);
3369 if (FoldedVOp.getNode()) return FoldedVOp;
3371 // fold (xor x, 0) -> x, vector edition
3372 if (ISD::isBuildVectorAllZeros(N0.getNode()))
3374 if (ISD::isBuildVectorAllZeros(N1.getNode()))
3378 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3379 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3380 return DAG.getConstant(0, VT);
3381 // fold (xor x, undef) -> undef
3382 if (N0.getOpcode() == ISD::UNDEF)
3384 if (N1.getOpcode() == ISD::UNDEF)
3386 // fold (xor c1, c2) -> c1^c2
3388 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3389 // canonicalize constant to RHS
3391 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0);
3392 // fold (xor x, 0) -> x
3393 if (N1C && N1C->isNullValue())
3396 SDValue RXOR = ReassociateOps(ISD::XOR, N->getDebugLoc(), N0, N1);
3397 if (RXOR.getNode() != 0)
3400 // fold !(x cc y) -> (x !cc y)
3401 if (N1C && N1C->getAPIntValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3402 bool isInt = LHS.getValueType().isInteger();
3403 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3406 if (!LegalOperations || TLI.isCondCodeLegal(NotCC, LHS.getValueType())) {
3407 switch (N0.getOpcode()) {
3409 llvm_unreachable("Unhandled SetCC Equivalent!");
3411 return DAG.getSetCC(N->getDebugLoc(), VT, LHS, RHS, NotCC);
3412 case ISD::SELECT_CC:
3413 return DAG.getSelectCC(N->getDebugLoc(), LHS, RHS, N0.getOperand(2),
3414 N0.getOperand(3), NotCC);
3419 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3420 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3421 N0.getNode()->hasOneUse() &&
3422 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3423 SDValue V = N0.getOperand(0);
3424 V = DAG.getNode(ISD::XOR, N0.getDebugLoc(), V.getValueType(), V,
3425 DAG.getConstant(1, V.getValueType()));
3426 AddToWorkList(V.getNode());
3427 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, V);
3430 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3431 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3432 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3433 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3434 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3435 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3436 LHS = DAG.getNode(ISD::XOR, LHS.getDebugLoc(), VT, LHS, N1); // LHS = ~LHS
3437 RHS = DAG.getNode(ISD::XOR, RHS.getDebugLoc(), VT, RHS, N1); // RHS = ~RHS
3438 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3439 return DAG.getNode(NewOpcode, N->getDebugLoc(), VT, LHS, RHS);
3442 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3443 if (N1C && N1C->isAllOnesValue() &&
3444 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3445 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3446 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3447 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3448 LHS = DAG.getNode(ISD::XOR, LHS.getDebugLoc(), VT, LHS, N1); // LHS = ~LHS
3449 RHS = DAG.getNode(ISD::XOR, RHS.getDebugLoc(), VT, RHS, N1); // RHS = ~RHS
3450 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3451 return DAG.getNode(NewOpcode, N->getDebugLoc(), VT, LHS, RHS);
3454 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3455 if (N1C && N0.getOpcode() == ISD::XOR) {
3456 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3457 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3459 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N0.getOperand(1),
3460 DAG.getConstant(N1C->getAPIntValue() ^
3461 N00C->getAPIntValue(), VT));
3463 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N0.getOperand(0),
3464 DAG.getConstant(N1C->getAPIntValue() ^
3465 N01C->getAPIntValue(), VT));
3467 // fold (xor x, x) -> 0
3469 return tryFoldToZero(N->getDebugLoc(), TLI, VT, DAG, LegalOperations);
3471 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
3472 if (N0.getOpcode() == N1.getOpcode()) {
3473 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3474 if (Tmp.getNode()) return Tmp;
3477 // Simplify the expression using non-local knowledge.
3478 if (!VT.isVector() &&
3479 SimplifyDemandedBits(SDValue(N, 0)))
3480 return SDValue(N, 0);
3485 /// visitShiftByConstant - Handle transforms common to the three shifts, when
3486 /// the shift amount is a constant.
3487 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, unsigned Amt) {
3488 SDNode *LHS = N->getOperand(0).getNode();
3489 if (!LHS->hasOneUse()) return SDValue();
3491 // We want to pull some binops through shifts, so that we have (and (shift))
3492 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
3493 // thing happens with address calculations, so it's important to canonicalize
3495 bool HighBitSet = false; // Can we transform this if the high bit is set?
3497 switch (LHS->getOpcode()) {
3498 default: return SDValue();
3501 HighBitSet = false; // We can only transform sra if the high bit is clear.
3504 HighBitSet = true; // We can only transform sra if the high bit is set.
3507 if (N->getOpcode() != ISD::SHL)
3508 return SDValue(); // only shl(add) not sr[al](add).
3509 HighBitSet = false; // We can only transform sra if the high bit is clear.
3513 // We require the RHS of the binop to be a constant as well.
3514 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
3515 if (!BinOpCst) return SDValue();
3517 // FIXME: disable this unless the input to the binop is a shift by a constant.
3518 // If it is not a shift, it pessimizes some common cases like:
3520 // void foo(int *X, int i) { X[i & 1235] = 1; }
3521 // int bar(int *X, int i) { return X[i & 255]; }
3522 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
3523 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
3524 BinOpLHSVal->getOpcode() != ISD::SRA &&
3525 BinOpLHSVal->getOpcode() != ISD::SRL) ||
3526 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
3529 EVT VT = N->getValueType(0);
3531 // If this is a signed shift right, and the high bit is modified by the
3532 // logical operation, do not perform the transformation. The highBitSet
3533 // boolean indicates the value of the high bit of the constant which would
3534 // cause it to be modified for this operation.
3535 if (N->getOpcode() == ISD::SRA) {
3536 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
3537 if (BinOpRHSSignSet != HighBitSet)
3541 // Fold the constants, shifting the binop RHS by the shift amount.
3542 SDValue NewRHS = DAG.getNode(N->getOpcode(), LHS->getOperand(1).getDebugLoc(),
3544 LHS->getOperand(1), N->getOperand(1));
3546 // Create the new shift.
3547 SDValue NewShift = DAG.getNode(N->getOpcode(),
3548 LHS->getOperand(0).getDebugLoc(),
3549 VT, LHS->getOperand(0), N->getOperand(1));
3551 // Create the new binop.
3552 return DAG.getNode(LHS->getOpcode(), N->getDebugLoc(), VT, NewShift, NewRHS);
3555 SDValue DAGCombiner::visitSHL(SDNode *N) {
3556 SDValue N0 = N->getOperand(0);
3557 SDValue N1 = N->getOperand(1);
3558 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3559 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3560 EVT VT = N0.getValueType();
3561 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3563 // fold (shl c1, c2) -> c1<<c2
3565 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
3566 // fold (shl 0, x) -> 0
3567 if (N0C && N0C->isNullValue())
3569 // fold (shl x, c >= size(x)) -> undef
3570 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3571 return DAG.getUNDEF(VT);
3572 // fold (shl x, 0) -> x
3573 if (N1C && N1C->isNullValue())
3575 // fold (shl undef, x) -> 0
3576 if (N0.getOpcode() == ISD::UNDEF)
3577 return DAG.getConstant(0, VT);
3578 // if (shl x, c) is known to be zero, return 0
3579 if (DAG.MaskedValueIsZero(SDValue(N, 0),
3580 APInt::getAllOnesValue(OpSizeInBits)))
3581 return DAG.getConstant(0, VT);
3582 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
3583 if (N1.getOpcode() == ISD::TRUNCATE &&
3584 N1.getOperand(0).getOpcode() == ISD::AND &&
3585 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3586 SDValue N101 = N1.getOperand(0).getOperand(1);
3587 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3588 EVT TruncVT = N1.getValueType();
3589 SDValue N100 = N1.getOperand(0).getOperand(0);
3590 APInt TruncC = N101C->getAPIntValue();
3591 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
3592 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
3593 DAG.getNode(ISD::AND, N->getDebugLoc(), TruncVT,
3594 DAG.getNode(ISD::TRUNCATE,
3597 DAG.getConstant(TruncC, TruncVT)));
3601 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3602 return SDValue(N, 0);
3604 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
3605 if (N1C && N0.getOpcode() == ISD::SHL &&
3606 N0.getOperand(1).getOpcode() == ISD::Constant) {
3607 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3608 uint64_t c2 = N1C->getZExtValue();
3609 if (c1 + c2 >= OpSizeInBits)
3610 return DAG.getConstant(0, VT);
3611 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0.getOperand(0),
3612 DAG.getConstant(c1 + c2, N1.getValueType()));
3615 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
3616 // For this to be valid, the second form must not preserve any of the bits
3617 // that are shifted out by the inner shift in the first form. This means
3618 // the outer shift size must be >= the number of bits added by the ext.
3619 // As a corollary, we don't care what kind of ext it is.
3620 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
3621 N0.getOpcode() == ISD::ANY_EXTEND ||
3622 N0.getOpcode() == ISD::SIGN_EXTEND) &&
3623 N0.getOperand(0).getOpcode() == ISD::SHL &&
3624 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
3626 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
3627 uint64_t c2 = N1C->getZExtValue();
3628 EVT InnerShiftVT = N0.getOperand(0).getValueType();
3629 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
3630 if (c2 >= OpSizeInBits - InnerShiftSize) {
3631 if (c1 + c2 >= OpSizeInBits)
3632 return DAG.getConstant(0, VT);
3633 return DAG.getNode(ISD::SHL, N0->getDebugLoc(), VT,
3634 DAG.getNode(N0.getOpcode(), N0->getDebugLoc(), VT,
3635 N0.getOperand(0)->getOperand(0)),
3636 DAG.getConstant(c1 + c2, N1.getValueType()));
3640 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
3641 // (and (srl x, (sub c1, c2), MASK)
3642 // Only fold this if the inner shift has no other uses -- if it does, folding
3643 // this will increase the total number of instructions.
3644 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse() &&
3645 N0.getOperand(1).getOpcode() == ISD::Constant) {
3646 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3647 if (c1 < VT.getSizeInBits()) {
3648 uint64_t c2 = N1C->getZExtValue();
3649 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
3650 VT.getSizeInBits() - c1);
3653 Mask = Mask.shl(c2-c1);
3654 Shift = DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0.getOperand(0),
3655 DAG.getConstant(c2-c1, N1.getValueType()));
3657 Mask = Mask.lshr(c1-c2);
3658 Shift = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0),
3659 DAG.getConstant(c1-c2, N1.getValueType()));
3661 return DAG.getNode(ISD::AND, N0.getDebugLoc(), VT, Shift,
3662 DAG.getConstant(Mask, VT));
3665 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
3666 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
3667 SDValue HiBitsMask =
3668 DAG.getConstant(APInt::getHighBitsSet(VT.getSizeInBits(),
3669 VT.getSizeInBits() -
3670 N1C->getZExtValue()),
3672 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
3677 SDValue NewSHL = visitShiftByConstant(N, N1C->getZExtValue());
3678 if (NewSHL.getNode())
3685 SDValue DAGCombiner::visitSRA(SDNode *N) {
3686 SDValue N0 = N->getOperand(0);
3687 SDValue N1 = N->getOperand(1);
3688 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3689 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3690 EVT VT = N0.getValueType();
3691 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3693 // fold (sra c1, c2) -> (sra c1, c2)
3695 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
3696 // fold (sra 0, x) -> 0
3697 if (N0C && N0C->isNullValue())
3699 // fold (sra -1, x) -> -1
3700 if (N0C && N0C->isAllOnesValue())
3702 // fold (sra x, (setge c, size(x))) -> undef
3703 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3704 return DAG.getUNDEF(VT);
3705 // fold (sra x, 0) -> x
3706 if (N1C && N1C->isNullValue())
3708 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
3710 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
3711 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
3712 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
3714 ExtVT = EVT::getVectorVT(*DAG.getContext(),
3715 ExtVT, VT.getVectorNumElements());
3716 if ((!LegalOperations ||
3717 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
3718 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
3719 N0.getOperand(0), DAG.getValueType(ExtVT));
3722 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
3723 if (N1C && N0.getOpcode() == ISD::SRA) {
3724 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3725 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
3726 if (Sum >= OpSizeInBits) Sum = OpSizeInBits-1;
3727 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0.getOperand(0),
3728 DAG.getConstant(Sum, N1C->getValueType(0)));
3732 // fold (sra (shl X, m), (sub result_size, n))
3733 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
3734 // result_size - n != m.
3735 // If truncate is free for the target sext(shl) is likely to result in better
3737 if (N0.getOpcode() == ISD::SHL) {
3738 // Get the two constanst of the shifts, CN0 = m, CN = n.
3739 const ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3741 // Determine what the truncate's result bitsize and type would be.
3743 EVT::getIntegerVT(*DAG.getContext(),
3744 OpSizeInBits - N1C->getZExtValue());
3745 // Determine the residual right-shift amount.
3746 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
3748 // If the shift is not a no-op (in which case this should be just a sign
3749 // extend already), the truncated to type is legal, sign_extend is legal
3750 // on that type, and the truncate to that type is both legal and free,
3751 // perform the transform.
3752 if ((ShiftAmt > 0) &&
3753 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
3754 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
3755 TLI.isTruncateFree(VT, TruncVT)) {
3757 SDValue Amt = DAG.getConstant(ShiftAmt,
3758 getShiftAmountTy(N0.getOperand(0).getValueType()));
3759 SDValue Shift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), VT,
3760 N0.getOperand(0), Amt);
3761 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), TruncVT,
3763 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(),
3764 N->getValueType(0), Trunc);
3769 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
3770 if (N1.getOpcode() == ISD::TRUNCATE &&
3771 N1.getOperand(0).getOpcode() == ISD::AND &&
3772 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3773 SDValue N101 = N1.getOperand(0).getOperand(1);
3774 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3775 EVT TruncVT = N1.getValueType();
3776 SDValue N100 = N1.getOperand(0).getOperand(0);
3777 APInt TruncC = N101C->getAPIntValue();
3778 TruncC = TruncC.trunc(TruncVT.getScalarType().getSizeInBits());
3779 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0,
3780 DAG.getNode(ISD::AND, N->getDebugLoc(),
3782 DAG.getNode(ISD::TRUNCATE,
3785 DAG.getConstant(TruncC, TruncVT)));
3789 // fold (sra (trunc (sr x, c1)), c2) -> (trunc (sra x, c1+c2))
3790 // if c1 is equal to the number of bits the trunc removes
3791 if (N0.getOpcode() == ISD::TRUNCATE &&
3792 (N0.getOperand(0).getOpcode() == ISD::SRL ||
3793 N0.getOperand(0).getOpcode() == ISD::SRA) &&
3794 N0.getOperand(0).hasOneUse() &&
3795 N0.getOperand(0).getOperand(1).hasOneUse() &&
3796 N1C && isa<ConstantSDNode>(N0.getOperand(0).getOperand(1))) {
3797 EVT LargeVT = N0.getOperand(0).getValueType();
3798 ConstantSDNode *LargeShiftAmt =
3799 cast<ConstantSDNode>(N0.getOperand(0).getOperand(1));
3801 if (LargeVT.getScalarType().getSizeInBits() - OpSizeInBits ==
3802 LargeShiftAmt->getZExtValue()) {
3804 DAG.getConstant(LargeShiftAmt->getZExtValue() + N1C->getZExtValue(),
3805 getShiftAmountTy(N0.getOperand(0).getOperand(0).getValueType()));
3806 SDValue SRA = DAG.getNode(ISD::SRA, N->getDebugLoc(), LargeVT,
3807 N0.getOperand(0).getOperand(0), Amt);
3808 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, SRA);
3812 // Simplify, based on bits shifted out of the LHS.
3813 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3814 return SDValue(N, 0);
3817 // If the sign bit is known to be zero, switch this to a SRL.
3818 if (DAG.SignBitIsZero(N0))
3819 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0, N1);
3822 SDValue NewSRA = visitShiftByConstant(N, N1C->getZExtValue());
3823 if (NewSRA.getNode())
3830 SDValue DAGCombiner::visitSRL(SDNode *N) {
3831 SDValue N0 = N->getOperand(0);
3832 SDValue N1 = N->getOperand(1);
3833 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3834 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3835 EVT VT = N0.getValueType();
3836 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3838 // fold (srl c1, c2) -> c1 >>u c2
3840 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
3841 // fold (srl 0, x) -> 0
3842 if (N0C && N0C->isNullValue())
3844 // fold (srl x, c >= size(x)) -> undef
3845 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3846 return DAG.getUNDEF(VT);
3847 // fold (srl x, 0) -> x
3848 if (N1C && N1C->isNullValue())
3850 // if (srl x, c) is known to be zero, return 0
3851 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
3852 APInt::getAllOnesValue(OpSizeInBits)))
3853 return DAG.getConstant(0, VT);
3855 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
3856 if (N1C && N0.getOpcode() == ISD::SRL &&
3857 N0.getOperand(1).getOpcode() == ISD::Constant) {
3858 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3859 uint64_t c2 = N1C->getZExtValue();
3860 if (c1 + c2 >= OpSizeInBits)
3861 return DAG.getConstant(0, VT);
3862 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0),
3863 DAG.getConstant(c1 + c2, N1.getValueType()));
3866 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
3867 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
3868 N0.getOperand(0).getOpcode() == ISD::SRL &&
3869 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
3871 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
3872 uint64_t c2 = N1C->getZExtValue();
3873 EVT InnerShiftVT = N0.getOperand(0).getValueType();
3874 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
3875 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
3876 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
3877 if (c1 + OpSizeInBits == InnerShiftSize) {
3878 if (c1 + c2 >= InnerShiftSize)
3879 return DAG.getConstant(0, VT);
3880 return DAG.getNode(ISD::TRUNCATE, N0->getDebugLoc(), VT,
3881 DAG.getNode(ISD::SRL, N0->getDebugLoc(), InnerShiftVT,
3882 N0.getOperand(0)->getOperand(0),
3883 DAG.getConstant(c1 + c2, ShiftCountVT)));
3887 // fold (srl (shl x, c), c) -> (and x, cst2)
3888 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1 &&
3889 N0.getValueSizeInBits() <= 64) {
3890 uint64_t ShAmt = N1C->getZExtValue()+64-N0.getValueSizeInBits();
3891 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
3892 DAG.getConstant(~0ULL >> ShAmt, VT));
3896 // fold (srl (anyextend x), c) -> (anyextend (srl x, c))
3897 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
3898 // Shifting in all undef bits?
3899 EVT SmallVT = N0.getOperand(0).getValueType();
3900 if (N1C->getZExtValue() >= SmallVT.getSizeInBits())
3901 return DAG.getUNDEF(VT);
3903 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
3904 uint64_t ShiftAmt = N1C->getZExtValue();
3905 SDValue SmallShift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), SmallVT,
3907 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
3908 AddToWorkList(SmallShift.getNode());
3909 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, SmallShift);
3913 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
3914 // bit, which is unmodified by sra.
3915 if (N1C && N1C->getZExtValue() + 1 == VT.getSizeInBits()) {
3916 if (N0.getOpcode() == ISD::SRA)
3917 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0), N1);
3920 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
3921 if (N1C && N0.getOpcode() == ISD::CTLZ &&
3922 N1C->getAPIntValue() == Log2_32(VT.getSizeInBits())) {
3923 APInt KnownZero, KnownOne;
3924 DAG.ComputeMaskedBits(N0.getOperand(0), KnownZero, KnownOne);
3926 // If any of the input bits are KnownOne, then the input couldn't be all
3927 // zeros, thus the result of the srl will always be zero.
3928 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
3930 // If all of the bits input the to ctlz node are known to be zero, then
3931 // the result of the ctlz is "32" and the result of the shift is one.
3932 APInt UnknownBits = ~KnownZero;
3933 if (UnknownBits == 0) return DAG.getConstant(1, VT);
3935 // Otherwise, check to see if there is exactly one bit input to the ctlz.
3936 if ((UnknownBits & (UnknownBits - 1)) == 0) {
3937 // Okay, we know that only that the single bit specified by UnknownBits
3938 // could be set on input to the CTLZ node. If this bit is set, the SRL
3939 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
3940 // to an SRL/XOR pair, which is likely to simplify more.
3941 unsigned ShAmt = UnknownBits.countTrailingZeros();
3942 SDValue Op = N0.getOperand(0);
3945 Op = DAG.getNode(ISD::SRL, N0.getDebugLoc(), VT, Op,
3946 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
3947 AddToWorkList(Op.getNode());
3950 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
3951 Op, DAG.getConstant(1, VT));
3955 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
3956 if (N1.getOpcode() == ISD::TRUNCATE &&
3957 N1.getOperand(0).getOpcode() == ISD::AND &&
3958 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3959 SDValue N101 = N1.getOperand(0).getOperand(1);
3960 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3961 EVT TruncVT = N1.getValueType();
3962 SDValue N100 = N1.getOperand(0).getOperand(0);
3963 APInt TruncC = N101C->getAPIntValue();
3964 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
3965 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0,
3966 DAG.getNode(ISD::AND, N->getDebugLoc(),
3968 DAG.getNode(ISD::TRUNCATE,
3971 DAG.getConstant(TruncC, TruncVT)));
3975 // fold operands of srl based on knowledge that the low bits are not
3977 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3978 return SDValue(N, 0);
3981 SDValue NewSRL = visitShiftByConstant(N, N1C->getZExtValue());
3982 if (NewSRL.getNode())
3986 // Attempt to convert a srl of a load into a narrower zero-extending load.
3987 SDValue NarrowLoad = ReduceLoadWidth(N);
3988 if (NarrowLoad.getNode())
3991 // Here is a common situation. We want to optimize:
3994 // %b = and i32 %a, 2
3995 // %c = srl i32 %b, 1
3996 // brcond i32 %c ...
4002 // %c = setcc eq %b, 0
4005 // However when after the source operand of SRL is optimized into AND, the SRL
4006 // itself may not be optimized further. Look for it and add the BRCOND into
4008 if (N->hasOneUse()) {
4009 SDNode *Use = *N->use_begin();
4010 if (Use->getOpcode() == ISD::BRCOND)
4012 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
4013 // Also look pass the truncate.
4014 Use = *Use->use_begin();
4015 if (Use->getOpcode() == ISD::BRCOND)
4023 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
4024 SDValue N0 = N->getOperand(0);
4025 EVT VT = N->getValueType(0);
4027 // fold (ctlz c1) -> c2
4028 if (isa<ConstantSDNode>(N0))
4029 return DAG.getNode(ISD::CTLZ, N->getDebugLoc(), VT, N0);
4033 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
4034 SDValue N0 = N->getOperand(0);
4035 EVT VT = N->getValueType(0);
4037 // fold (ctlz_zero_undef c1) -> c2
4038 if (isa<ConstantSDNode>(N0))
4039 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, N->getDebugLoc(), VT, N0);
4043 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
4044 SDValue N0 = N->getOperand(0);
4045 EVT VT = N->getValueType(0);
4047 // fold (cttz c1) -> c2
4048 if (isa<ConstantSDNode>(N0))
4049 return DAG.getNode(ISD::CTTZ, N->getDebugLoc(), VT, N0);
4053 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
4054 SDValue N0 = N->getOperand(0);
4055 EVT VT = N->getValueType(0);
4057 // fold (cttz_zero_undef c1) -> c2
4058 if (isa<ConstantSDNode>(N0))
4059 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, N->getDebugLoc(), VT, N0);
4063 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
4064 SDValue N0 = N->getOperand(0);
4065 EVT VT = N->getValueType(0);
4067 // fold (ctpop c1) -> c2
4068 if (isa<ConstantSDNode>(N0))
4069 return DAG.getNode(ISD::CTPOP, N->getDebugLoc(), VT, N0);
4073 SDValue DAGCombiner::visitSELECT(SDNode *N) {
4074 SDValue N0 = N->getOperand(0);
4075 SDValue N1 = N->getOperand(1);
4076 SDValue N2 = N->getOperand(2);
4077 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
4078 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
4079 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
4080 EVT VT = N->getValueType(0);
4081 EVT VT0 = N0.getValueType();
4083 // fold (select C, X, X) -> X
4086 // fold (select true, X, Y) -> X
4087 if (N0C && !N0C->isNullValue())
4089 // fold (select false, X, Y) -> Y
4090 if (N0C && N0C->isNullValue())
4092 // fold (select C, 1, X) -> (or C, X)
4093 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4094 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N2);
4095 // fold (select C, 0, 1) -> (xor C, 1)
4096 if (VT.isInteger() &&
4099 TLI.getBooleanContents(false) ==
4100 TargetLowering::ZeroOrOneBooleanContent)) &&
4101 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4104 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT0,
4105 N0, DAG.getConstant(1, VT0));
4106 XORNode = DAG.getNode(ISD::XOR, N0.getDebugLoc(), VT0,
4107 N0, DAG.getConstant(1, VT0));
4108 AddToWorkList(XORNode.getNode());
4110 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, XORNode);
4111 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, XORNode);
4113 // fold (select C, 0, X) -> (and (not C), X)
4114 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4115 SDValue NOTNode = DAG.getNOT(N0.getDebugLoc(), N0, VT);
4116 AddToWorkList(NOTNode.getNode());
4117 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, NOTNode, N2);
4119 // fold (select C, X, 1) -> (or (not C), X)
4120 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4121 SDValue NOTNode = DAG.getNOT(N0.getDebugLoc(), N0, VT);
4122 AddToWorkList(NOTNode.getNode());
4123 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, NOTNode, N1);
4125 // fold (select C, X, 0) -> (and C, X)
4126 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4127 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, N1);
4128 // fold (select X, X, Y) -> (or X, Y)
4129 // fold (select X, 1, Y) -> (or X, Y)
4130 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4131 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N2);
4132 // fold (select X, Y, X) -> (and X, Y)
4133 // fold (select X, Y, 0) -> (and X, Y)
4134 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4135 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, N1);
4137 // If we can fold this based on the true/false value, do so.
4138 if (SimplifySelectOps(N, N1, N2))
4139 return SDValue(N, 0); // Don't revisit N.
4141 // fold selects based on a setcc into other things, such as min/max/abs
4142 if (N0.getOpcode() == ISD::SETCC) {
4144 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
4145 // having to say they don't support SELECT_CC on every type the DAG knows
4146 // about, since there is no way to mark an opcode illegal at all value types
4147 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other) &&
4148 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT))
4149 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT,
4150 N0.getOperand(0), N0.getOperand(1),
4151 N1, N2, N0.getOperand(2));
4152 return SimplifySelect(N->getDebugLoc(), N0, N1, N2);
4158 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
4159 SDValue N0 = N->getOperand(0);
4160 SDValue N1 = N->getOperand(1);
4161 SDValue N2 = N->getOperand(2);
4162 SDValue N3 = N->getOperand(3);
4163 SDValue N4 = N->getOperand(4);
4164 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
4166 // fold select_cc lhs, rhs, x, x, cc -> x
4170 // Determine if the condition we're dealing with is constant
4171 SDValue SCC = SimplifySetCC(TLI.getSetCCResultType(N0.getValueType()),
4172 N0, N1, CC, N->getDebugLoc(), false);
4173 if (SCC.getNode()) AddToWorkList(SCC.getNode());
4175 if (ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode())) {
4176 if (!SCCC->isNullValue())
4177 return N2; // cond always true -> true val
4179 return N3; // cond always false -> false val
4182 // Fold to a simpler select_cc
4183 if (SCC.getNode() && SCC.getOpcode() == ISD::SETCC)
4184 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), N2.getValueType(),
4185 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
4188 // If we can fold this based on the true/false value, do so.
4189 if (SimplifySelectOps(N, N2, N3))
4190 return SDValue(N, 0); // Don't revisit N.
4192 // fold select_cc into other things, such as min/max/abs
4193 return SimplifySelectCC(N->getDebugLoc(), N0, N1, N2, N3, CC);
4196 SDValue DAGCombiner::visitSETCC(SDNode *N) {
4197 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
4198 cast<CondCodeSDNode>(N->getOperand(2))->get(),
4202 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
4203 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
4204 // transformation. Returns true if extension are possible and the above
4205 // mentioned transformation is profitable.
4206 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
4208 SmallVector<SDNode*, 4> &ExtendNodes,
4209 const TargetLowering &TLI) {
4210 bool HasCopyToRegUses = false;
4211 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
4212 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
4213 UE = N0.getNode()->use_end();
4218 if (UI.getUse().getResNo() != N0.getResNo())
4220 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
4221 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
4222 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
4223 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
4224 // Sign bits will be lost after a zext.
4227 for (unsigned i = 0; i != 2; ++i) {
4228 SDValue UseOp = User->getOperand(i);
4231 if (!isa<ConstantSDNode>(UseOp))
4236 ExtendNodes.push_back(User);
4239 // If truncates aren't free and there are users we can't
4240 // extend, it isn't worthwhile.
4243 // Remember if this value is live-out.
4244 if (User->getOpcode() == ISD::CopyToReg)
4245 HasCopyToRegUses = true;
4248 if (HasCopyToRegUses) {
4249 bool BothLiveOut = false;
4250 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
4252 SDUse &Use = UI.getUse();
4253 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
4259 // Both unextended and extended values are live out. There had better be
4260 // a good reason for the transformation.
4261 return ExtendNodes.size();
4266 void DAGCombiner::ExtendSetCCUses(SmallVector<SDNode*, 4> SetCCs,
4267 SDValue Trunc, SDValue ExtLoad, DebugLoc DL,
4268 ISD::NodeType ExtType) {
4269 // Extend SetCC uses if necessary.
4270 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
4271 SDNode *SetCC = SetCCs[i];
4272 SmallVector<SDValue, 4> Ops;
4274 for (unsigned j = 0; j != 2; ++j) {
4275 SDValue SOp = SetCC->getOperand(j);
4277 Ops.push_back(ExtLoad);
4279 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
4282 Ops.push_back(SetCC->getOperand(2));
4283 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0),
4284 &Ops[0], Ops.size()));
4288 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
4289 SDValue N0 = N->getOperand(0);
4290 EVT VT = N->getValueType(0);
4292 // fold (sext c1) -> c1
4293 if (isa<ConstantSDNode>(N0))
4294 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N0);
4296 // fold (sext (sext x)) -> (sext x)
4297 // fold (sext (aext x)) -> (sext x)
4298 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4299 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT,
4302 if (N0.getOpcode() == ISD::TRUNCATE) {
4303 // fold (sext (truncate (load x))) -> (sext (smaller load x))
4304 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
4305 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4306 if (NarrowLoad.getNode()) {
4307 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4308 if (NarrowLoad.getNode() != N0.getNode()) {
4309 CombineTo(N0.getNode(), NarrowLoad);
4310 // CombineTo deleted the truncate, if needed, but not what's under it.
4313 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4316 // See if the value being truncated is already sign extended. If so, just
4317 // eliminate the trunc/sext pair.
4318 SDValue Op = N0.getOperand(0);
4319 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
4320 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
4321 unsigned DestBits = VT.getScalarType().getSizeInBits();
4322 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
4324 if (OpBits == DestBits) {
4325 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
4326 // bits, it is already ready.
4327 if (NumSignBits > DestBits-MidBits)
4329 } else if (OpBits < DestBits) {
4330 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
4331 // bits, just sext from i32.
4332 if (NumSignBits > OpBits-MidBits)
4333 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, Op);
4335 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
4336 // bits, just truncate to i32.
4337 if (NumSignBits > OpBits-MidBits)
4338 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4341 // fold (sext (truncate x)) -> (sextinreg x).
4342 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
4343 N0.getValueType())) {
4344 if (OpBits < DestBits)
4345 Op = DAG.getNode(ISD::ANY_EXTEND, N0.getDebugLoc(), VT, Op);
4346 else if (OpBits > DestBits)
4347 Op = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), VT, Op);
4348 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT, Op,
4349 DAG.getValueType(N0.getValueType()));
4353 // fold (sext (load x)) -> (sext (truncate (sextload x)))
4354 // None of the supported targets knows how to perform load and sign extend
4355 // on vectors in one instruction. We only perform this transformation on
4357 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4358 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4359 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) {
4360 bool DoXform = true;
4361 SmallVector<SDNode*, 4> SetCCs;
4362 if (!N0.hasOneUse())
4363 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
4365 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4366 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
4368 LN0->getBasePtr(), LN0->getPointerInfo(),
4370 LN0->isVolatile(), LN0->isNonTemporal(),
4371 LN0->getAlignment());
4372 CombineTo(N, ExtLoad);
4373 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4374 N0.getValueType(), ExtLoad);
4375 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4376 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4378 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4382 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
4383 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
4384 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
4385 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
4386 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4387 EVT MemVT = LN0->getMemoryVT();
4388 if ((!LegalOperations && !LN0->isVolatile()) ||
4389 TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) {
4390 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
4392 LN0->getBasePtr(), LN0->getPointerInfo(),
4394 LN0->isVolatile(), LN0->isNonTemporal(),
4395 LN0->getAlignment());
4396 CombineTo(N, ExtLoad);
4397 CombineTo(N0.getNode(),
4398 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4399 N0.getValueType(), ExtLoad),
4400 ExtLoad.getValue(1));
4401 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4405 // fold (sext (and/or/xor (load x), cst)) ->
4406 // (and/or/xor (sextload x), (sext cst))
4407 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
4408 N0.getOpcode() == ISD::XOR) &&
4409 isa<LoadSDNode>(N0.getOperand(0)) &&
4410 N0.getOperand(1).getOpcode() == ISD::Constant &&
4411 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()) &&
4412 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
4413 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
4414 if (LN0->getExtensionType() != ISD::ZEXTLOAD) {
4415 bool DoXform = true;
4416 SmallVector<SDNode*, 4> SetCCs;
4417 if (!N0.hasOneUse())
4418 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
4421 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, LN0->getDebugLoc(), VT,
4422 LN0->getChain(), LN0->getBasePtr(),
4423 LN0->getPointerInfo(),
4426 LN0->isNonTemporal(),
4427 LN0->getAlignment());
4428 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4429 Mask = Mask.sext(VT.getSizeInBits());
4430 SDValue And = DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
4431 ExtLoad, DAG.getConstant(Mask, VT));
4432 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
4433 N0.getOperand(0).getDebugLoc(),
4434 N0.getOperand(0).getValueType(), ExtLoad);
4436 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
4437 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4439 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4444 if (N0.getOpcode() == ISD::SETCC) {
4445 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
4446 // Only do this before legalize for now.
4447 if (VT.isVector() && !LegalOperations) {
4448 EVT N0VT = N0.getOperand(0).getValueType();
4449 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
4450 // of the same size as the compared operands. Only optimize sext(setcc())
4451 // if this is the case.
4452 EVT SVT = TLI.getSetCCResultType(N0VT);
4454 // We know that the # elements of the results is the same as the
4455 // # elements of the compare (and the # elements of the compare result
4456 // for that matter). Check to see that they are the same size. If so,
4457 // we know that the element size of the sext'd result matches the
4458 // element size of the compare operands.
4459 if (VT.getSizeInBits() == SVT.getSizeInBits())
4460 return DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4462 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4463 // If the desired elements are smaller or larger than the source
4464 // elements we can use a matching integer vector type and then
4465 // truncate/sign extend
4466 EVT MatchingElementType =
4467 EVT::getIntegerVT(*DAG.getContext(),
4468 N0VT.getScalarType().getSizeInBits());
4469 EVT MatchingVectorType =
4470 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4471 N0VT.getVectorNumElements());
4473 if (SVT == MatchingVectorType) {
4474 SDValue VsetCC = DAG.getSetCC(N->getDebugLoc(), MatchingVectorType,
4475 N0.getOperand(0), N0.getOperand(1),
4476 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4477 return DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT);
4481 // sext(setcc x, y, cc) -> (select_cc x, y, -1, 0, cc)
4482 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
4484 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
4486 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4487 NegOne, DAG.getConstant(0, VT),
4488 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4489 if (SCC.getNode()) return SCC;
4490 if (!LegalOperations ||
4491 TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(VT)))
4492 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4493 DAG.getSetCC(N->getDebugLoc(),
4494 TLI.getSetCCResultType(VT),
4495 N0.getOperand(0), N0.getOperand(1),
4496 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
4497 NegOne, DAG.getConstant(0, VT));
4500 // fold (sext x) -> (zext x) if the sign bit is known zero.
4501 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
4502 DAG.SignBitIsZero(N0))
4503 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
4508 // isTruncateOf - If N is a truncate of some other value, return true, record
4509 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
4510 // This function computes KnownZero to avoid a duplicated call to
4511 // ComputeMaskedBits in the caller.
4512 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
4515 if (N->getOpcode() == ISD::TRUNCATE) {
4516 Op = N->getOperand(0);
4517 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
4521 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
4522 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
4525 SDValue Op0 = N->getOperand(0);
4526 SDValue Op1 = N->getOperand(1);
4527 assert(Op0.getValueType() == Op1.getValueType());
4529 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
4530 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
4531 if (COp0 && COp0->isNullValue())
4533 else if (COp1 && COp1->isNullValue())
4538 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
4540 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
4546 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
4547 SDValue N0 = N->getOperand(0);
4548 EVT VT = N->getValueType(0);
4550 // fold (zext c1) -> c1
4551 if (isa<ConstantSDNode>(N0))
4552 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
4553 // fold (zext (zext x)) -> (zext x)
4554 // fold (zext (aext x)) -> (zext x)
4555 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4556 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT,
4559 // fold (zext (truncate x)) -> (zext x) or
4560 // (zext (truncate x)) -> (truncate x)
4561 // This is valid when the truncated bits of x are already zero.
4562 // FIXME: We should extend this to work for vectors too.
4565 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
4566 APInt TruncatedBits =
4567 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
4568 APInt(Op.getValueSizeInBits(), 0) :
4569 APInt::getBitsSet(Op.getValueSizeInBits(),
4570 N0.getValueSizeInBits(),
4571 std::min(Op.getValueSizeInBits(),
4572 VT.getSizeInBits()));
4573 if (TruncatedBits == (KnownZero & TruncatedBits)) {
4574 if (VT.bitsGT(Op.getValueType()))
4575 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, Op);
4576 if (VT.bitsLT(Op.getValueType()))
4577 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4583 // fold (zext (truncate (load x))) -> (zext (smaller load x))
4584 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
4585 if (N0.getOpcode() == ISD::TRUNCATE) {
4586 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4587 if (NarrowLoad.getNode()) {
4588 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4589 if (NarrowLoad.getNode() != N0.getNode()) {
4590 CombineTo(N0.getNode(), NarrowLoad);
4591 // CombineTo deleted the truncate, if needed, but not what's under it.
4594 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4598 // fold (zext (truncate x)) -> (and x, mask)
4599 if (N0.getOpcode() == ISD::TRUNCATE &&
4600 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
4602 // fold (zext (truncate (load x))) -> (zext (smaller load x))
4603 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
4604 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4605 if (NarrowLoad.getNode()) {
4606 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4607 if (NarrowLoad.getNode() != N0.getNode()) {
4608 CombineTo(N0.getNode(), NarrowLoad);
4609 // CombineTo deleted the truncate, if needed, but not what's under it.
4612 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4615 SDValue Op = N0.getOperand(0);
4616 if (Op.getValueType().bitsLT(VT)) {
4617 Op = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, Op);
4618 AddToWorkList(Op.getNode());
4619 } else if (Op.getValueType().bitsGT(VT)) {
4620 Op = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4621 AddToWorkList(Op.getNode());
4623 return DAG.getZeroExtendInReg(Op, N->getDebugLoc(),
4624 N0.getValueType().getScalarType());
4627 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
4628 // if either of the casts is not free.
4629 if (N0.getOpcode() == ISD::AND &&
4630 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
4631 N0.getOperand(1).getOpcode() == ISD::Constant &&
4632 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
4633 N0.getValueType()) ||
4634 !TLI.isZExtFree(N0.getValueType(), VT))) {
4635 SDValue X = N0.getOperand(0).getOperand(0);
4636 if (X.getValueType().bitsLT(VT)) {
4637 X = DAG.getNode(ISD::ANY_EXTEND, X.getDebugLoc(), VT, X);
4638 } else if (X.getValueType().bitsGT(VT)) {
4639 X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
4641 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4642 Mask = Mask.zext(VT.getSizeInBits());
4643 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4644 X, DAG.getConstant(Mask, VT));
4647 // fold (zext (load x)) -> (zext (truncate (zextload x)))
4648 // None of the supported targets knows how to perform load and vector_zext
4649 // on vectors in one instruction. We only perform this transformation on
4651 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4652 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4653 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
4654 bool DoXform = true;
4655 SmallVector<SDNode*, 4> SetCCs;
4656 if (!N0.hasOneUse())
4657 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
4659 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4660 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N->getDebugLoc(), VT,
4662 LN0->getBasePtr(), LN0->getPointerInfo(),
4664 LN0->isVolatile(), LN0->isNonTemporal(),
4665 LN0->getAlignment());
4666 CombineTo(N, ExtLoad);
4667 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4668 N0.getValueType(), ExtLoad);
4669 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4671 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4673 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4677 // fold (zext (and/or/xor (load x), cst)) ->
4678 // (and/or/xor (zextload x), (zext cst))
4679 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
4680 N0.getOpcode() == ISD::XOR) &&
4681 isa<LoadSDNode>(N0.getOperand(0)) &&
4682 N0.getOperand(1).getOpcode() == ISD::Constant &&
4683 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()) &&
4684 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
4685 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
4686 if (LN0->getExtensionType() != ISD::SEXTLOAD) {
4687 bool DoXform = true;
4688 SmallVector<SDNode*, 4> SetCCs;
4689 if (!N0.hasOneUse())
4690 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
4693 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), VT,
4694 LN0->getChain(), LN0->getBasePtr(),
4695 LN0->getPointerInfo(),
4698 LN0->isNonTemporal(),
4699 LN0->getAlignment());
4700 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4701 Mask = Mask.zext(VT.getSizeInBits());
4702 SDValue And = DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
4703 ExtLoad, DAG.getConstant(Mask, VT));
4704 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
4705 N0.getOperand(0).getDebugLoc(),
4706 N0.getOperand(0).getValueType(), ExtLoad);
4708 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
4709 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4711 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4716 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
4717 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
4718 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
4719 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
4720 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4721 EVT MemVT = LN0->getMemoryVT();
4722 if ((!LegalOperations && !LN0->isVolatile()) ||
4723 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) {
4724 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N->getDebugLoc(), VT,
4726 LN0->getBasePtr(), LN0->getPointerInfo(),
4728 LN0->isVolatile(), LN0->isNonTemporal(),
4729 LN0->getAlignment());
4730 CombineTo(N, ExtLoad);
4731 CombineTo(N0.getNode(),
4732 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), N0.getValueType(),
4734 ExtLoad.getValue(1));
4735 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4739 if (N0.getOpcode() == ISD::SETCC) {
4740 if (!LegalOperations && VT.isVector()) {
4741 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
4742 // Only do this before legalize for now.
4743 EVT N0VT = N0.getOperand(0).getValueType();
4744 EVT EltVT = VT.getVectorElementType();
4745 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
4746 DAG.getConstant(1, EltVT));
4747 if (VT.getSizeInBits() == N0VT.getSizeInBits())
4748 // We know that the # elements of the results is the same as the
4749 // # elements of the compare (and the # elements of the compare result
4750 // for that matter). Check to see that they are the same size. If so,
4751 // we know that the element size of the sext'd result matches the
4752 // element size of the compare operands.
4753 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4754 DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4756 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
4757 DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
4758 &OneOps[0], OneOps.size()));
4760 // If the desired elements are smaller or larger than the source
4761 // elements we can use a matching integer vector type and then
4762 // truncate/sign extend
4763 EVT MatchingElementType =
4764 EVT::getIntegerVT(*DAG.getContext(),
4765 N0VT.getScalarType().getSizeInBits());
4766 EVT MatchingVectorType =
4767 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4768 N0VT.getVectorNumElements());
4770 DAG.getSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
4772 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4773 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4774 DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT),
4775 DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
4776 &OneOps[0], OneOps.size()));
4779 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
4781 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4782 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
4783 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4784 if (SCC.getNode()) return SCC;
4787 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
4788 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
4789 isa<ConstantSDNode>(N0.getOperand(1)) &&
4790 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
4792 SDValue ShAmt = N0.getOperand(1);
4793 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
4794 if (N0.getOpcode() == ISD::SHL) {
4795 SDValue InnerZExt = N0.getOperand(0);
4796 // If the original shl may be shifting out bits, do not perform this
4798 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
4799 InnerZExt.getOperand(0).getValueType().getSizeInBits();
4800 if (ShAmtVal > KnownZeroBits)
4804 DebugLoc DL = N->getDebugLoc();
4806 // Ensure that the shift amount is wide enough for the shifted value.
4807 if (VT.getSizeInBits() >= 256)
4808 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
4810 return DAG.getNode(N0.getOpcode(), DL, VT,
4811 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
4818 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
4819 SDValue N0 = N->getOperand(0);
4820 EVT VT = N->getValueType(0);
4822 // fold (aext c1) -> c1
4823 if (isa<ConstantSDNode>(N0))
4824 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, N0);
4825 // fold (aext (aext x)) -> (aext x)
4826 // fold (aext (zext x)) -> (zext x)
4827 // fold (aext (sext x)) -> (sext x)
4828 if (N0.getOpcode() == ISD::ANY_EXTEND ||
4829 N0.getOpcode() == ISD::ZERO_EXTEND ||
4830 N0.getOpcode() == ISD::SIGN_EXTEND)
4831 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, N0.getOperand(0));
4833 // fold (aext (truncate (load x))) -> (aext (smaller load x))
4834 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
4835 if (N0.getOpcode() == ISD::TRUNCATE) {
4836 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4837 if (NarrowLoad.getNode()) {
4838 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4839 if (NarrowLoad.getNode() != N0.getNode()) {
4840 CombineTo(N0.getNode(), NarrowLoad);
4841 // CombineTo deleted the truncate, if needed, but not what's under it.
4844 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4848 // fold (aext (truncate x))
4849 if (N0.getOpcode() == ISD::TRUNCATE) {
4850 SDValue TruncOp = N0.getOperand(0);
4851 if (TruncOp.getValueType() == VT)
4852 return TruncOp; // x iff x size == zext size.
4853 if (TruncOp.getValueType().bitsGT(VT))
4854 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, TruncOp);
4855 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, TruncOp);
4858 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
4859 // if the trunc is not free.
4860 if (N0.getOpcode() == ISD::AND &&
4861 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
4862 N0.getOperand(1).getOpcode() == ISD::Constant &&
4863 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
4864 N0.getValueType())) {
4865 SDValue X = N0.getOperand(0).getOperand(0);
4866 if (X.getValueType().bitsLT(VT)) {
4867 X = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, X);
4868 } else if (X.getValueType().bitsGT(VT)) {
4869 X = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, X);
4871 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4872 Mask = Mask.zext(VT.getSizeInBits());
4873 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4874 X, DAG.getConstant(Mask, VT));
4877 // fold (aext (load x)) -> (aext (truncate (extload x)))
4878 // None of the supported targets knows how to perform load and any_ext
4879 // on vectors in one instruction. We only perform this transformation on
4881 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4882 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4883 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
4884 bool DoXform = true;
4885 SmallVector<SDNode*, 4> SetCCs;
4886 if (!N0.hasOneUse())
4887 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
4889 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4890 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
4892 LN0->getBasePtr(), LN0->getPointerInfo(),
4894 LN0->isVolatile(), LN0->isNonTemporal(),
4895 LN0->getAlignment());
4896 CombineTo(N, ExtLoad);
4897 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4898 N0.getValueType(), ExtLoad);
4899 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4900 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4902 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4906 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
4907 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
4908 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
4909 if (N0.getOpcode() == ISD::LOAD &&
4910 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
4912 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4913 EVT MemVT = LN0->getMemoryVT();
4914 SDValue ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), N->getDebugLoc(),
4915 VT, LN0->getChain(), LN0->getBasePtr(),
4916 LN0->getPointerInfo(), MemVT,
4917 LN0->isVolatile(), LN0->isNonTemporal(),
4918 LN0->getAlignment());
4919 CombineTo(N, ExtLoad);
4920 CombineTo(N0.getNode(),
4921 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4922 N0.getValueType(), ExtLoad),
4923 ExtLoad.getValue(1));
4924 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4927 if (N0.getOpcode() == ISD::SETCC) {
4928 // aext(setcc) -> sext_in_reg(vsetcc) for vectors.
4929 // Only do this before legalize for now.
4930 if (VT.isVector() && !LegalOperations) {
4931 EVT N0VT = N0.getOperand(0).getValueType();
4932 // We know that the # elements of the results is the same as the
4933 // # elements of the compare (and the # elements of the compare result
4934 // for that matter). Check to see that they are the same size. If so,
4935 // we know that the element size of the sext'd result matches the
4936 // element size of the compare operands.
4937 if (VT.getSizeInBits() == N0VT.getSizeInBits())
4938 return DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4940 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4941 // If the desired elements are smaller or larger than the source
4942 // elements we can use a matching integer vector type and then
4943 // truncate/sign extend
4945 EVT MatchingElementType =
4946 EVT::getIntegerVT(*DAG.getContext(),
4947 N0VT.getScalarType().getSizeInBits());
4948 EVT MatchingVectorType =
4949 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4950 N0VT.getVectorNumElements());
4952 DAG.getSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
4954 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4955 return DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT);
4959 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
4961 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4962 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
4963 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4971 /// GetDemandedBits - See if the specified operand can be simplified with the
4972 /// knowledge that only the bits specified by Mask are used. If so, return the
4973 /// simpler operand, otherwise return a null SDValue.
4974 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
4975 switch (V.getOpcode()) {
4977 case ISD::Constant: {
4978 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
4979 assert(CV != 0 && "Const value should be ConstSDNode.");
4980 const APInt &CVal = CV->getAPIntValue();
4981 APInt NewVal = CVal & Mask;
4982 if (NewVal != CVal) {
4983 return DAG.getConstant(NewVal, V.getValueType());
4989 // If the LHS or RHS don't contribute bits to the or, drop them.
4990 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
4991 return V.getOperand(1);
4992 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
4993 return V.getOperand(0);
4996 // Only look at single-use SRLs.
4997 if (!V.getNode()->hasOneUse())
4999 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
5000 // See if we can recursively simplify the LHS.
5001 unsigned Amt = RHSC->getZExtValue();
5003 // Watch out for shift count overflow though.
5004 if (Amt >= Mask.getBitWidth()) break;
5005 APInt NewMask = Mask << Amt;
5006 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
5007 if (SimplifyLHS.getNode())
5008 return DAG.getNode(ISD::SRL, V.getDebugLoc(), V.getValueType(),
5009 SimplifyLHS, V.getOperand(1));
5015 /// ReduceLoadWidth - If the result of a wider load is shifted to right of N
5016 /// bits and then truncated to a narrower type and where N is a multiple
5017 /// of number of bits of the narrower type, transform it to a narrower load
5018 /// from address + N / num of bits of new type. If the result is to be
5019 /// extended, also fold the extension to form a extending load.
5020 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
5021 unsigned Opc = N->getOpcode();
5023 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
5024 SDValue N0 = N->getOperand(0);
5025 EVT VT = N->getValueType(0);
5028 // This transformation isn't valid for vector loads.
5032 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
5034 if (Opc == ISD::SIGN_EXTEND_INREG) {
5035 ExtType = ISD::SEXTLOAD;
5036 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
5037 } else if (Opc == ISD::SRL) {
5038 // Another special-case: SRL is basically zero-extending a narrower value.
5039 ExtType = ISD::ZEXTLOAD;
5041 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
5042 if (!N01) return SDValue();
5043 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
5044 VT.getSizeInBits() - N01->getZExtValue());
5046 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, ExtVT))
5049 unsigned EVTBits = ExtVT.getSizeInBits();
5051 // Do not generate loads of non-round integer types since these can
5052 // be expensive (and would be wrong if the type is not byte sized).
5053 if (!ExtVT.isRound())
5057 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
5058 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5059 ShAmt = N01->getZExtValue();
5060 // Is the shift amount a multiple of size of VT?
5061 if ((ShAmt & (EVTBits-1)) == 0) {
5062 N0 = N0.getOperand(0);
5063 // Is the load width a multiple of size of VT?
5064 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
5068 // At this point, we must have a load or else we can't do the transform.
5069 if (!isa<LoadSDNode>(N0)) return SDValue();
5071 // If the shift amount is larger than the input type then we're not
5072 // accessing any of the loaded bytes. If the load was a zextload/extload
5073 // then the result of the shift+trunc is zero/undef (handled elsewhere).
5074 // If the load was a sextload then the result is a splat of the sign bit
5075 // of the extended byte. This is not worth optimizing for.
5076 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
5081 // If the load is shifted left (and the result isn't shifted back right),
5082 // we can fold the truncate through the shift.
5083 unsigned ShLeftAmt = 0;
5084 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
5085 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
5086 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
5087 ShLeftAmt = N01->getZExtValue();
5088 N0 = N0.getOperand(0);
5092 // If we haven't found a load, we can't narrow it. Don't transform one with
5093 // multiple uses, this would require adding a new load.
5094 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse() ||
5095 // Don't change the width of a volatile load.
5096 cast<LoadSDNode>(N0)->isVolatile())
5099 // Verify that we are actually reducing a load width here.
5100 if (cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits() < EVTBits)
5103 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5104 EVT PtrType = N0.getOperand(1).getValueType();
5106 if (PtrType == MVT::Untyped || PtrType.isExtended())
5107 // It's not possible to generate a constant of extended or untyped type.
5110 // For big endian targets, we need to adjust the offset to the pointer to
5111 // load the correct bytes.
5112 if (TLI.isBigEndian()) {
5113 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
5114 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
5115 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
5118 uint64_t PtrOff = ShAmt / 8;
5119 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
5120 SDValue NewPtr = DAG.getNode(ISD::ADD, LN0->getDebugLoc(),
5121 PtrType, LN0->getBasePtr(),
5122 DAG.getConstant(PtrOff, PtrType));
5123 AddToWorkList(NewPtr.getNode());
5126 if (ExtType == ISD::NON_EXTLOAD)
5127 Load = DAG.getLoad(VT, N0.getDebugLoc(), LN0->getChain(), NewPtr,
5128 LN0->getPointerInfo().getWithOffset(PtrOff),
5129 LN0->isVolatile(), LN0->isNonTemporal(),
5130 LN0->isInvariant(), NewAlign);
5132 Load = DAG.getExtLoad(ExtType, N0.getDebugLoc(), VT, LN0->getChain(),NewPtr,
5133 LN0->getPointerInfo().getWithOffset(PtrOff),
5134 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
5137 // Replace the old load's chain with the new load's chain.
5138 WorkListRemover DeadNodes(*this);
5139 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1));
5141 // Shift the result left, if we've swallowed a left shift.
5142 SDValue Result = Load;
5143 if (ShLeftAmt != 0) {
5144 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
5145 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
5147 Result = DAG.getNode(ISD::SHL, N0.getDebugLoc(), VT,
5148 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
5151 // Return the new loaded value.
5155 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
5156 SDValue N0 = N->getOperand(0);
5157 SDValue N1 = N->getOperand(1);
5158 EVT VT = N->getValueType(0);
5159 EVT EVT = cast<VTSDNode>(N1)->getVT();
5160 unsigned VTBits = VT.getScalarType().getSizeInBits();
5161 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
5163 // fold (sext_in_reg c1) -> c1
5164 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
5165 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT, N0, N1);
5167 // If the input is already sign extended, just drop the extension.
5168 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
5171 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
5172 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
5173 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT())) {
5174 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
5175 N0.getOperand(0), N1);
5178 // fold (sext_in_reg (sext x)) -> (sext x)
5179 // fold (sext_in_reg (aext x)) -> (sext x)
5180 // if x is small enough.
5181 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
5182 SDValue N00 = N0.getOperand(0);
5183 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
5184 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
5185 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N00, N1);
5188 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
5189 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
5190 return DAG.getZeroExtendInReg(N0, N->getDebugLoc(), EVT);
5192 // fold operands of sext_in_reg based on knowledge that the top bits are not
5194 if (SimplifyDemandedBits(SDValue(N, 0)))
5195 return SDValue(N, 0);
5197 // fold (sext_in_reg (load x)) -> (smaller sextload x)
5198 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
5199 SDValue NarrowLoad = ReduceLoadWidth(N);
5200 if (NarrowLoad.getNode())
5203 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
5204 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
5205 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
5206 if (N0.getOpcode() == ISD::SRL) {
5207 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
5208 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
5209 // We can turn this into an SRA iff the input to the SRL is already sign
5211 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
5212 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
5213 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT,
5214 N0.getOperand(0), N0.getOperand(1));
5218 // fold (sext_inreg (extload x)) -> (sextload x)
5219 if (ISD::isEXTLoad(N0.getNode()) &&
5220 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5221 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5222 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5223 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5224 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5225 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
5227 LN0->getBasePtr(), LN0->getPointerInfo(),
5229 LN0->isVolatile(), LN0->isNonTemporal(),
5230 LN0->getAlignment());
5231 CombineTo(N, ExtLoad);
5232 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5233 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5235 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
5236 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5238 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5239 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5240 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5241 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5242 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
5244 LN0->getBasePtr(), LN0->getPointerInfo(),
5246 LN0->isVolatile(), LN0->isNonTemporal(),
5247 LN0->getAlignment());
5248 CombineTo(N, ExtLoad);
5249 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5250 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5253 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
5254 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
5255 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
5256 N0.getOperand(1), false);
5257 if (BSwap.getNode() != 0)
5258 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
5265 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
5266 SDValue N0 = N->getOperand(0);
5267 EVT VT = N->getValueType(0);
5268 bool isLE = TLI.isLittleEndian();
5271 if (N0.getValueType() == N->getValueType(0))
5273 // fold (truncate c1) -> c1
5274 if (isa<ConstantSDNode>(N0))
5275 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0);
5276 // fold (truncate (truncate x)) -> (truncate x)
5277 if (N0.getOpcode() == ISD::TRUNCATE)
5278 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0.getOperand(0));
5279 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
5280 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
5281 N0.getOpcode() == ISD::SIGN_EXTEND ||
5282 N0.getOpcode() == ISD::ANY_EXTEND) {
5283 if (N0.getOperand(0).getValueType().bitsLT(VT))
5284 // if the source is smaller than the dest, we still need an extend
5285 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
5287 if (N0.getOperand(0).getValueType().bitsGT(VT))
5288 // if the source is larger than the dest, than we just need the truncate
5289 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0.getOperand(0));
5290 // if the source and dest are the same type, we can drop both the extend
5291 // and the truncate.
5292 return N0.getOperand(0);
5295 // Fold extract-and-trunc into a narrow extract. For example:
5296 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
5297 // i32 y = TRUNCATE(i64 x)
5299 // v16i8 b = BITCAST (v2i64 val)
5300 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
5302 // Note: We only run this optimization after type legalization (which often
5303 // creates this pattern) and before operation legalization after which
5304 // we need to be more careful about the vector instructions that we generate.
5305 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5306 LegalTypes && !LegalOperations && N0->hasOneUse()) {
5308 EVT VecTy = N0.getOperand(0).getValueType();
5309 EVT ExTy = N0.getValueType();
5310 EVT TrTy = N->getValueType(0);
5312 unsigned NumElem = VecTy.getVectorNumElements();
5313 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
5315 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
5316 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
5318 SDValue EltNo = N0->getOperand(1);
5319 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
5320 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5321 EVT IndexTy = N0->getOperand(1).getValueType();
5322 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
5324 SDValue V = DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
5325 NVT, N0.getOperand(0));
5327 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
5328 N->getDebugLoc(), TrTy, V,
5329 DAG.getConstant(Index, IndexTy));
5333 // See if we can simplify the input to this truncate through knowledge that
5334 // only the low bits are being used.
5335 // For example "trunc (or (shl x, 8), y)" // -> trunc y
5336 // Currently we only perform this optimization on scalars because vectors
5337 // may have different active low bits.
5338 if (!VT.isVector()) {
5340 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
5341 VT.getSizeInBits()));
5342 if (Shorter.getNode())
5343 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Shorter);
5345 // fold (truncate (load x)) -> (smaller load x)
5346 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
5347 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
5348 SDValue Reduced = ReduceLoadWidth(N);
5349 if (Reduced.getNode())
5352 // fold (trunc (concat ... x ...)) -> (concat ..., (trunc x), ...)),
5353 // where ... are all 'undef'.
5354 if (N0.getOpcode() == ISD::CONCAT_VECTORS && !LegalTypes) {
5355 SmallVector<EVT, 8> VTs;
5358 unsigned NumDefs = 0;
5360 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
5361 SDValue X = N0.getOperand(i);
5362 if (X.getOpcode() != ISD::UNDEF) {
5367 // Stop if more than one members are non-undef.
5370 VTs.push_back(EVT::getVectorVT(*DAG.getContext(),
5371 VT.getVectorElementType(),
5372 X.getValueType().getVectorNumElements()));
5376 return DAG.getUNDEF(VT);
5379 assert(V.getNode() && "The single defined operand is empty!");
5380 SmallVector<SDValue, 8> Opnds;
5381 for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
5383 Opnds.push_back(DAG.getUNDEF(VTs[i]));
5386 SDValue NV = DAG.getNode(ISD::TRUNCATE, V.getDebugLoc(), VTs[i], V);
5387 AddToWorkList(NV.getNode());
5388 Opnds.push_back(NV);
5390 return DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
5391 &Opnds[0], Opnds.size());
5395 // Simplify the operands using demanded-bits information.
5396 if (!VT.isVector() &&
5397 SimplifyDemandedBits(SDValue(N, 0)))
5398 return SDValue(N, 0);
5403 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
5404 SDValue Elt = N->getOperand(i);
5405 if (Elt.getOpcode() != ISD::MERGE_VALUES)
5406 return Elt.getNode();
5407 return Elt.getOperand(Elt.getResNo()).getNode();
5410 /// CombineConsecutiveLoads - build_pair (load, load) -> load
5411 /// if load locations are consecutive.
5412 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
5413 assert(N->getOpcode() == ISD::BUILD_PAIR);
5415 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
5416 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
5417 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
5418 LD1->getPointerInfo().getAddrSpace() !=
5419 LD2->getPointerInfo().getAddrSpace())
5421 EVT LD1VT = LD1->getValueType(0);
5423 if (ISD::isNON_EXTLoad(LD2) &&
5425 // If both are volatile this would reduce the number of volatile loads.
5426 // If one is volatile it might be ok, but play conservative and bail out.
5427 !LD1->isVolatile() &&
5428 !LD2->isVolatile() &&
5429 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
5430 unsigned Align = LD1->getAlignment();
5431 unsigned NewAlign = TLI.getDataLayout()->
5432 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
5434 if (NewAlign <= Align &&
5435 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
5436 return DAG.getLoad(VT, N->getDebugLoc(), LD1->getChain(),
5437 LD1->getBasePtr(), LD1->getPointerInfo(),
5438 false, false, false, Align);
5444 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
5445 SDValue N0 = N->getOperand(0);
5446 EVT VT = N->getValueType(0);
5448 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
5449 // Only do this before legalize, since afterward the target may be depending
5450 // on the bitconvert.
5451 // First check to see if this is all constant.
5453 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
5455 bool isSimple = true;
5456 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i)
5457 if (N0.getOperand(i).getOpcode() != ISD::UNDEF &&
5458 N0.getOperand(i).getOpcode() != ISD::Constant &&
5459 N0.getOperand(i).getOpcode() != ISD::ConstantFP) {
5464 EVT DestEltVT = N->getValueType(0).getVectorElementType();
5465 assert(!DestEltVT.isVector() &&
5466 "Element type of vector ValueType must not be vector!");
5468 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
5471 // If the input is a constant, let getNode fold it.
5472 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
5473 SDValue Res = DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, N0);
5474 if (Res.getNode() != N) {
5475 if (!LegalOperations ||
5476 TLI.isOperationLegal(Res.getNode()->getOpcode(), VT))
5479 // Folding it resulted in an illegal node, and it's too late to
5480 // do that. Clean up the old node and forego the transformation.
5481 // Ideally this won't happen very often, because instcombine
5482 // and the earlier dagcombine runs (where illegal nodes are
5483 // permitted) should have folded most of them already.
5484 DAG.DeleteNode(Res.getNode());
5488 // (conv (conv x, t1), t2) -> (conv x, t2)
5489 if (N0.getOpcode() == ISD::BITCAST)
5490 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT,
5493 // fold (conv (load x)) -> (load (conv*)x)
5494 // If the resultant load doesn't need a higher alignment than the original!
5495 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
5496 // Do not change the width of a volatile load.
5497 !cast<LoadSDNode>(N0)->isVolatile() &&
5498 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT))) {
5499 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5500 unsigned Align = TLI.getDataLayout()->
5501 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
5502 unsigned OrigAlign = LN0->getAlignment();
5504 if (Align <= OrigAlign) {
5505 SDValue Load = DAG.getLoad(VT, N->getDebugLoc(), LN0->getChain(),
5506 LN0->getBasePtr(), LN0->getPointerInfo(),
5507 LN0->isVolatile(), LN0->isNonTemporal(),
5508 LN0->isInvariant(), OrigAlign);
5510 CombineTo(N0.getNode(),
5511 DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5512 N0.getValueType(), Load),
5518 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
5519 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
5520 // This often reduces constant pool loads.
5521 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(VT)) ||
5522 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(VT))) &&
5523 N0.getNode()->hasOneUse() && VT.isInteger() &&
5524 !VT.isVector() && !N0.getValueType().isVector()) {
5525 SDValue NewConv = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(), VT,
5527 AddToWorkList(NewConv.getNode());
5529 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
5530 if (N0.getOpcode() == ISD::FNEG)
5531 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
5532 NewConv, DAG.getConstant(SignBit, VT));
5533 assert(N0.getOpcode() == ISD::FABS);
5534 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
5535 NewConv, DAG.getConstant(~SignBit, VT));
5538 // fold (bitconvert (fcopysign cst, x)) ->
5539 // (or (and (bitconvert x), sign), (and cst, (not sign)))
5540 // Note that we don't handle (copysign x, cst) because this can always be
5541 // folded to an fneg or fabs.
5542 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
5543 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
5544 VT.isInteger() && !VT.isVector()) {
5545 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
5546 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
5547 if (isTypeLegal(IntXVT)) {
5548 SDValue X = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5549 IntXVT, N0.getOperand(1));
5550 AddToWorkList(X.getNode());
5552 // If X has a different width than the result/lhs, sext it or truncate it.
5553 unsigned VTWidth = VT.getSizeInBits();
5554 if (OrigXWidth < VTWidth) {
5555 X = DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, X);
5556 AddToWorkList(X.getNode());
5557 } else if (OrigXWidth > VTWidth) {
5558 // To get the sign bit in the right place, we have to shift it right
5559 // before truncating.
5560 X = DAG.getNode(ISD::SRL, X.getDebugLoc(),
5561 X.getValueType(), X,
5562 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
5563 AddToWorkList(X.getNode());
5564 X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
5565 AddToWorkList(X.getNode());
5568 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
5569 X = DAG.getNode(ISD::AND, X.getDebugLoc(), VT,
5570 X, DAG.getConstant(SignBit, VT));
5571 AddToWorkList(X.getNode());
5573 SDValue Cst = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5574 VT, N0.getOperand(0));
5575 Cst = DAG.getNode(ISD::AND, Cst.getDebugLoc(), VT,
5576 Cst, DAG.getConstant(~SignBit, VT));
5577 AddToWorkList(Cst.getNode());
5579 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, X, Cst);
5583 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
5584 if (N0.getOpcode() == ISD::BUILD_PAIR) {
5585 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
5586 if (CombineLD.getNode())
5593 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
5594 EVT VT = N->getValueType(0);
5595 return CombineConsecutiveLoads(N, VT);
5598 /// ConstantFoldBITCASTofBUILD_VECTOR - We know that BV is a build_vector
5599 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
5600 /// destination element value type.
5601 SDValue DAGCombiner::
5602 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
5603 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
5605 // If this is already the right type, we're done.
5606 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
5608 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
5609 unsigned DstBitSize = DstEltVT.getSizeInBits();
5611 // If this is a conversion of N elements of one type to N elements of another
5612 // type, convert each element. This handles FP<->INT cases.
5613 if (SrcBitSize == DstBitSize) {
5614 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
5615 BV->getValueType(0).getVectorNumElements());
5617 // Due to the FP element handling below calling this routine recursively,
5618 // we can end up with a scalar-to-vector node here.
5619 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
5620 return DAG.getNode(ISD::SCALAR_TO_VECTOR, BV->getDebugLoc(), VT,
5621 DAG.getNode(ISD::BITCAST, BV->getDebugLoc(),
5622 DstEltVT, BV->getOperand(0)));
5624 SmallVector<SDValue, 8> Ops;
5625 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
5626 SDValue Op = BV->getOperand(i);
5627 // If the vector element type is not legal, the BUILD_VECTOR operands
5628 // are promoted and implicitly truncated. Make that explicit here.
5629 if (Op.getValueType() != SrcEltVT)
5630 Op = DAG.getNode(ISD::TRUNCATE, BV->getDebugLoc(), SrcEltVT, Op);
5631 Ops.push_back(DAG.getNode(ISD::BITCAST, BV->getDebugLoc(),
5633 AddToWorkList(Ops.back().getNode());
5635 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5636 &Ops[0], Ops.size());
5639 // Otherwise, we're growing or shrinking the elements. To avoid having to
5640 // handle annoying details of growing/shrinking FP values, we convert them to
5642 if (SrcEltVT.isFloatingPoint()) {
5643 // Convert the input float vector to a int vector where the elements are the
5645 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
5646 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
5647 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
5651 // Now we know the input is an integer vector. If the output is a FP type,
5652 // convert to integer first, then to FP of the right size.
5653 if (DstEltVT.isFloatingPoint()) {
5654 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
5655 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
5656 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
5658 // Next, convert to FP elements of the same size.
5659 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
5662 // Okay, we know the src/dst types are both integers of differing types.
5663 // Handling growing first.
5664 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
5665 if (SrcBitSize < DstBitSize) {
5666 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
5668 SmallVector<SDValue, 8> Ops;
5669 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
5670 i += NumInputsPerOutput) {
5671 bool isLE = TLI.isLittleEndian();
5672 APInt NewBits = APInt(DstBitSize, 0);
5673 bool EltIsUndef = true;
5674 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
5675 // Shift the previously computed bits over.
5676 NewBits <<= SrcBitSize;
5677 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
5678 if (Op.getOpcode() == ISD::UNDEF) continue;
5681 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
5682 zextOrTrunc(SrcBitSize).zext(DstBitSize);
5686 Ops.push_back(DAG.getUNDEF(DstEltVT));
5688 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
5691 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
5692 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5693 &Ops[0], Ops.size());
5696 // Finally, this must be the case where we are shrinking elements: each input
5697 // turns into multiple outputs.
5698 bool isS2V = ISD::isScalarToVector(BV);
5699 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
5700 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
5701 NumOutputsPerInput*BV->getNumOperands());
5702 SmallVector<SDValue, 8> Ops;
5704 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
5705 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
5706 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
5707 Ops.push_back(DAG.getUNDEF(DstEltVT));
5711 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
5712 getAPIntValue().zextOrTrunc(SrcBitSize);
5714 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
5715 APInt ThisVal = OpVal.trunc(DstBitSize);
5716 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
5717 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
5718 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
5719 return DAG.getNode(ISD::SCALAR_TO_VECTOR, BV->getDebugLoc(), VT,
5721 OpVal = OpVal.lshr(DstBitSize);
5724 // For big endian targets, swap the order of the pieces of each element.
5725 if (TLI.isBigEndian())
5726 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
5729 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5730 &Ops[0], Ops.size());
5733 SDValue DAGCombiner::visitFADD(SDNode *N) {
5734 SDValue N0 = N->getOperand(0);
5735 SDValue N1 = N->getOperand(1);
5736 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5737 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5738 EVT VT = N->getValueType(0);
5741 if (VT.isVector()) {
5742 SDValue FoldedVOp = SimplifyVBinOp(N);
5743 if (FoldedVOp.getNode()) return FoldedVOp;
5746 // fold (fadd c1, c2) -> c1 + c2
5748 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N1);
5749 // canonicalize constant to RHS
5750 if (N0CFP && !N1CFP)
5751 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N1, N0);
5752 // fold (fadd A, 0) -> A
5753 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
5754 N1CFP->getValueAPF().isZero())
5756 // fold (fadd A, (fneg B)) -> (fsub A, B)
5757 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
5758 isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
5759 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N0,
5760 GetNegatedExpression(N1, DAG, LegalOperations));
5761 // fold (fadd (fneg A), B) -> (fsub B, A)
5762 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
5763 isNegatibleForFree(N0, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
5764 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N1,
5765 GetNegatedExpression(N0, DAG, LegalOperations));
5767 // If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
5768 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
5769 N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
5770 isa<ConstantFPSDNode>(N0.getOperand(1)))
5771 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0.getOperand(0),
5772 DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5773 N0.getOperand(1), N1));
5775 // If allow, fold (fadd (fneg x), x) -> 0.0
5776 if (DAG.getTarget().Options.UnsafeFPMath &&
5777 N0.getOpcode() == ISD::FNEG && N0.getOperand(0) == N1) {
5778 return DAG.getConstantFP(0.0, VT);
5781 // If allow, fold (fadd x, (fneg x)) -> 0.0
5782 if (DAG.getTarget().Options.UnsafeFPMath &&
5783 N1.getOpcode() == ISD::FNEG && N1.getOperand(0) == N0) {
5784 return DAG.getConstantFP(0.0, VT);
5787 // In unsafe math mode, we can fold chains of FADD's of the same value
5788 // into multiplications. This transform is not safe in general because
5789 // we are reducing the number of rounding steps.
5790 if (DAG.getTarget().Options.UnsafeFPMath &&
5791 TLI.isOperationLegalOrCustom(ISD::FMUL, VT) &&
5793 if (N0.getOpcode() == ISD::FMUL) {
5794 ConstantFPSDNode *CFP00 = dyn_cast<ConstantFPSDNode>(N0.getOperand(0));
5795 ConstantFPSDNode *CFP01 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
5797 // (fadd (fmul c, x), x) -> (fmul c+1, x)
5798 if (CFP00 && !CFP01 && N0.getOperand(1) == N1) {
5799 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5801 DAG.getConstantFP(1.0, VT));
5802 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5806 // (fadd (fmul x, c), x) -> (fmul c+1, x)
5807 if (CFP01 && !CFP00 && N0.getOperand(0) == 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 (fadd x, x), x) -> (fmul 3.0, x)
5816 if (!CFP00 && !CFP01 && N0.getOperand(0) == N0.getOperand(1) &&
5817 N0.getOperand(0) == N1) {
5818 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5819 N1, DAG.getConstantFP(3.0, VT));
5822 // (fadd (fmul c, x), (fadd x, x)) -> (fmul c+2, x)
5823 if (CFP00 && !CFP01 && N1.getOpcode() == ISD::FADD &&
5824 N1.getOperand(0) == N1.getOperand(1) &&
5825 N0.getOperand(1) == N1.getOperand(0)) {
5826 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5828 DAG.getConstantFP(2.0, VT));
5829 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5830 N0.getOperand(1), NewCFP);
5833 // (fadd (fmul x, c), (fadd x, x)) -> (fmul c+2, x)
5834 if (CFP01 && !CFP00 && N1.getOpcode() == ISD::FADD &&
5835 N1.getOperand(0) == N1.getOperand(1) &&
5836 N0.getOperand(0) == N1.getOperand(0)) {
5837 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5839 DAG.getConstantFP(2.0, VT));
5840 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5841 N0.getOperand(0), NewCFP);
5845 if (N1.getOpcode() == ISD::FMUL) {
5846 ConstantFPSDNode *CFP10 = dyn_cast<ConstantFPSDNode>(N1.getOperand(0));
5847 ConstantFPSDNode *CFP11 = dyn_cast<ConstantFPSDNode>(N1.getOperand(1));
5849 // (fadd x, (fmul c, x)) -> (fmul c+1, x)
5850 if (CFP10 && !CFP11 && N1.getOperand(1) == N0) {
5851 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5853 DAG.getConstantFP(1.0, VT));
5854 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5858 // (fadd x, (fmul x, c)) -> (fmul c+1, x)
5859 if (CFP11 && !CFP10 && N1.getOperand(0) == 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, (fadd x, x)) -> (fmul 3.0, x)
5868 if (!CFP10 && !CFP11 && N1.getOperand(0) == N1.getOperand(1) &&
5869 N1.getOperand(0) == N0) {
5870 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5871 N0, DAG.getConstantFP(3.0, VT));
5874 // (fadd (fadd x, x), (fmul c, x)) -> (fmul c+2, x)
5875 if (CFP10 && !CFP11 && N1.getOpcode() == ISD::FADD &&
5876 N1.getOperand(0) == N1.getOperand(1) &&
5877 N0.getOperand(1) == N1.getOperand(0)) {
5878 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5880 DAG.getConstantFP(2.0, VT));
5881 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5882 N0.getOperand(1), NewCFP);
5885 // (fadd (fadd x, x), (fmul x, c)) -> (fmul c+2, x)
5886 if (CFP11 && !CFP10 && N1.getOpcode() == ISD::FADD &&
5887 N1.getOperand(0) == N1.getOperand(1) &&
5888 N0.getOperand(0) == N1.getOperand(0)) {
5889 SDValue NewCFP = DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5891 DAG.getConstantFP(2.0, VT));
5892 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5893 N0.getOperand(0), NewCFP);
5897 // (fadd (fadd x, x), (fadd x, x)) -> (fmul 4.0, x)
5898 if (N0.getOpcode() == ISD::FADD && N1.getOpcode() == ISD::FADD &&
5899 N0.getOperand(0) == N0.getOperand(1) &&
5900 N1.getOperand(0) == N1.getOperand(1) &&
5901 N0.getOperand(0) == N1.getOperand(0)) {
5902 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5904 DAG.getConstantFP(4.0, VT));
5908 // FADD -> FMA combines:
5909 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
5910 DAG.getTarget().Options.UnsafeFPMath) &&
5911 DAG.getTarget().getTargetLowering()->isFMAFasterThanMulAndAdd(VT) &&
5912 TLI.isOperationLegalOrCustom(ISD::FMA, VT)) {
5914 // fold (fadd (fmul x, y), z) -> (fma x, y, z)
5915 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse()) {
5916 return DAG.getNode(ISD::FMA, N->getDebugLoc(), VT,
5917 N0.getOperand(0), N0.getOperand(1), N1);
5920 // fold (fadd x, (fmul y, z)) -> (fma y, z, x)
5921 // Note: Commutes FADD operands.
5922 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse()) {
5923 return DAG.getNode(ISD::FMA, N->getDebugLoc(), VT,
5924 N1.getOperand(0), N1.getOperand(1), N0);
5931 SDValue DAGCombiner::visitFSUB(SDNode *N) {
5932 SDValue N0 = N->getOperand(0);
5933 SDValue N1 = N->getOperand(1);
5934 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5935 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5936 EVT VT = N->getValueType(0);
5937 DebugLoc dl = N->getDebugLoc();
5940 if (VT.isVector()) {
5941 SDValue FoldedVOp = SimplifyVBinOp(N);
5942 if (FoldedVOp.getNode()) return FoldedVOp;
5945 // fold (fsub c1, c2) -> c1-c2
5947 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N0, N1);
5948 // fold (fsub A, 0) -> A
5949 if (DAG.getTarget().Options.UnsafeFPMath &&
5950 N1CFP && N1CFP->getValueAPF().isZero())
5952 // fold (fsub 0, B) -> -B
5953 if (DAG.getTarget().Options.UnsafeFPMath &&
5954 N0CFP && N0CFP->getValueAPF().isZero()) {
5955 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
5956 return GetNegatedExpression(N1, DAG, LegalOperations);
5957 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
5958 return DAG.getNode(ISD::FNEG, dl, VT, N1);
5960 // fold (fsub A, (fneg B)) -> (fadd A, B)
5961 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
5962 return DAG.getNode(ISD::FADD, dl, VT, N0,
5963 GetNegatedExpression(N1, DAG, LegalOperations));
5965 // If 'unsafe math' is enabled, fold
5966 // (fsub x, x) -> 0.0 &
5967 // (fsub x, (fadd x, y)) -> (fneg y) &
5968 // (fsub x, (fadd y, x)) -> (fneg y)
5969 if (DAG.getTarget().Options.UnsafeFPMath) {
5971 return DAG.getConstantFP(0.0f, VT);
5973 if (N1.getOpcode() == ISD::FADD) {
5974 SDValue N10 = N1->getOperand(0);
5975 SDValue N11 = N1->getOperand(1);
5977 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI,
5978 &DAG.getTarget().Options))
5979 return GetNegatedExpression(N11, DAG, LegalOperations);
5980 else if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI,
5981 &DAG.getTarget().Options))
5982 return GetNegatedExpression(N10, DAG, LegalOperations);
5986 // FSUB -> FMA combines:
5987 if ((DAG.getTarget().Options.AllowFPOpFusion == FPOpFusion::Fast ||
5988 DAG.getTarget().Options.UnsafeFPMath) &&
5989 DAG.getTarget().getTargetLowering()->isFMAFasterThanMulAndAdd(VT) &&
5990 TLI.isOperationLegalOrCustom(ISD::FMA, VT)) {
5992 // fold (fsub (fmul x, y), z) -> (fma x, y, (fneg z))
5993 if (N0.getOpcode() == ISD::FMUL && N0->hasOneUse()) {
5994 return DAG.getNode(ISD::FMA, dl, VT,
5995 N0.getOperand(0), N0.getOperand(1),
5996 DAG.getNode(ISD::FNEG, dl, VT, N1));
5999 // fold (fsub x, (fmul y, z)) -> (fma (fneg y), z, x)
6000 // Note: Commutes FSUB operands.
6001 if (N1.getOpcode() == ISD::FMUL && N1->hasOneUse()) {
6002 return DAG.getNode(ISD::FMA, dl, VT,
6003 DAG.getNode(ISD::FNEG, dl, VT,
6005 N1.getOperand(1), N0);
6008 // fold (fsub (-(fmul, x, y)), z) -> (fma (fneg x), y, (fneg z))
6009 if (N0.getOpcode() == ISD::FNEG &&
6010 N0.getOperand(0).getOpcode() == ISD::FMUL &&
6011 N0->hasOneUse() && N0.getOperand(0).hasOneUse()) {
6012 SDValue N00 = N0.getOperand(0).getOperand(0);
6013 SDValue N01 = N0.getOperand(0).getOperand(1);
6014 return DAG.getNode(ISD::FMA, dl, VT,
6015 DAG.getNode(ISD::FNEG, dl, VT, N00), N01,
6016 DAG.getNode(ISD::FNEG, dl, VT, N1));
6023 SDValue DAGCombiner::visitFMUL(SDNode *N) {
6024 SDValue N0 = N->getOperand(0);
6025 SDValue N1 = N->getOperand(1);
6026 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6027 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6028 EVT VT = N->getValueType(0);
6029 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6032 if (VT.isVector()) {
6033 SDValue FoldedVOp = SimplifyVBinOp(N);
6034 if (FoldedVOp.getNode()) return FoldedVOp;
6037 // fold (fmul c1, c2) -> c1*c2
6039 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0, N1);
6040 // canonicalize constant to RHS
6041 if (N0CFP && !N1CFP)
6042 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N1, N0);
6043 // fold (fmul A, 0) -> 0
6044 if (DAG.getTarget().Options.UnsafeFPMath &&
6045 N1CFP && N1CFP->getValueAPF().isZero())
6047 // fold (fmul A, 0) -> 0, vector edition.
6048 if (DAG.getTarget().Options.UnsafeFPMath &&
6049 ISD::isBuildVectorAllZeros(N1.getNode()))
6051 // fold (fmul A, 1.0) -> A
6052 if (N1CFP && N1CFP->isExactlyValue(1.0))
6054 // fold (fmul X, 2.0) -> (fadd X, X)
6055 if (N1CFP && N1CFP->isExactlyValue(+2.0))
6056 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N0);
6057 // fold (fmul X, -1.0) -> (fneg X)
6058 if (N1CFP && N1CFP->isExactlyValue(-1.0))
6059 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6060 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT, N0);
6062 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
6063 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
6064 &DAG.getTarget().Options)) {
6065 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
6066 &DAG.getTarget().Options)) {
6067 // Both can be negated for free, check to see if at least one is cheaper
6069 if (LHSNeg == 2 || RHSNeg == 2)
6070 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
6071 GetNegatedExpression(N0, DAG, LegalOperations),
6072 GetNegatedExpression(N1, DAG, LegalOperations));
6076 // If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
6077 if (DAG.getTarget().Options.UnsafeFPMath &&
6078 N1CFP && N0.getOpcode() == ISD::FMUL &&
6079 N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
6080 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0.getOperand(0),
6081 DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
6082 N0.getOperand(1), N1));
6087 SDValue DAGCombiner::visitFMA(SDNode *N) {
6088 SDValue N0 = N->getOperand(0);
6089 SDValue N1 = N->getOperand(1);
6090 SDValue N2 = N->getOperand(2);
6091 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6092 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6093 EVT VT = N->getValueType(0);
6094 DebugLoc dl = N->getDebugLoc();
6096 if (DAG.getTarget().Options.UnsafeFPMath) {
6097 if (N0CFP && N0CFP->isZero())
6099 if (N1CFP && N1CFP->isZero())
6102 if (N0CFP && N0CFP->isExactlyValue(1.0))
6103 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N1, N2);
6104 if (N1CFP && N1CFP->isExactlyValue(1.0))
6105 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N2);
6107 // Canonicalize (fma c, x, y) -> (fma x, c, y)
6108 if (N0CFP && !N1CFP)
6109 return DAG.getNode(ISD::FMA, N->getDebugLoc(), VT, N1, N0, N2);
6111 // (fma x, c1, (fmul x, c2)) -> (fmul x, c1+c2)
6112 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6113 N2.getOpcode() == ISD::FMUL &&
6114 N0 == N2.getOperand(0) &&
6115 N2.getOperand(1).getOpcode() == ISD::ConstantFP) {
6116 return DAG.getNode(ISD::FMUL, dl, VT, N0,
6117 DAG.getNode(ISD::FADD, dl, VT, N1, N2.getOperand(1)));
6121 // (fma (fmul x, c1), c2, y) -> (fma x, c1*c2, y)
6122 if (DAG.getTarget().Options.UnsafeFPMath &&
6123 N0.getOpcode() == ISD::FMUL && N1CFP &&
6124 N0.getOperand(1).getOpcode() == ISD::ConstantFP) {
6125 return DAG.getNode(ISD::FMA, dl, VT,
6127 DAG.getNode(ISD::FMUL, dl, VT, N1, N0.getOperand(1)),
6131 // (fma x, 1, y) -> (fadd x, y)
6132 // (fma x, -1, y) -> (fadd (fneg x), y)
6134 if (N1CFP->isExactlyValue(1.0))
6135 return DAG.getNode(ISD::FADD, dl, VT, N0, N2);
6137 if (N1CFP->isExactlyValue(-1.0) &&
6138 (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))) {
6139 SDValue RHSNeg = DAG.getNode(ISD::FNEG, dl, VT, N0);
6140 AddToWorkList(RHSNeg.getNode());
6141 return DAG.getNode(ISD::FADD, dl, VT, N2, RHSNeg);
6145 // (fma x, c, x) -> (fmul x, (c+1))
6146 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP && N0 == N2) {
6147 return DAG.getNode(ISD::FMUL, dl, VT,
6149 DAG.getNode(ISD::FADD, dl, VT,
6150 N1, DAG.getConstantFP(1.0, VT)));
6153 // (fma x, c, (fneg x)) -> (fmul x, (c-1))
6154 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
6155 N2.getOpcode() == ISD::FNEG && N2.getOperand(0) == N0) {
6156 return DAG.getNode(ISD::FMUL, dl, VT,
6158 DAG.getNode(ISD::FADD, dl, VT,
6159 N1, DAG.getConstantFP(-1.0, VT)));
6166 SDValue DAGCombiner::visitFDIV(SDNode *N) {
6167 SDValue N0 = N->getOperand(0);
6168 SDValue N1 = N->getOperand(1);
6169 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6170 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6171 EVT VT = N->getValueType(0);
6172 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
6175 if (VT.isVector()) {
6176 SDValue FoldedVOp = SimplifyVBinOp(N);
6177 if (FoldedVOp.getNode()) return FoldedVOp;
6180 // fold (fdiv c1, c2) -> c1/c2
6182 return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT, N0, N1);
6184 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
6185 if (N1CFP && DAG.getTarget().Options.UnsafeFPMath) {
6186 // Compute the reciprocal 1.0 / c2.
6187 APFloat N1APF = N1CFP->getValueAPF();
6188 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
6189 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
6190 // Only do the transform if the reciprocal is a legal fp immediate that
6191 // isn't too nasty (eg NaN, denormal, ...).
6192 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
6193 (!LegalOperations ||
6194 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
6195 // backend)... we should handle this gracefully after Legalize.
6196 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
6197 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
6198 TLI.isFPImmLegal(Recip, VT)))
6199 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0,
6200 DAG.getConstantFP(Recip, VT));
6203 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
6204 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
6205 &DAG.getTarget().Options)) {
6206 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
6207 &DAG.getTarget().Options)) {
6208 // Both can be negated for free, check to see if at least one is cheaper
6210 if (LHSNeg == 2 || RHSNeg == 2)
6211 return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT,
6212 GetNegatedExpression(N0, DAG, LegalOperations),
6213 GetNegatedExpression(N1, DAG, LegalOperations));
6220 SDValue DAGCombiner::visitFREM(SDNode *N) {
6221 SDValue N0 = N->getOperand(0);
6222 SDValue N1 = N->getOperand(1);
6223 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6224 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6225 EVT VT = N->getValueType(0);
6227 // fold (frem c1, c2) -> fmod(c1,c2)
6229 return DAG.getNode(ISD::FREM, N->getDebugLoc(), VT, N0, N1);
6234 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
6235 SDValue N0 = N->getOperand(0);
6236 SDValue N1 = N->getOperand(1);
6237 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6238 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
6239 EVT VT = N->getValueType(0);
6241 if (N0CFP && N1CFP) // Constant fold
6242 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT, N0, N1);
6245 const APFloat& V = N1CFP->getValueAPF();
6246 // copysign(x, c1) -> fabs(x) iff ispos(c1)
6247 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
6248 if (!V.isNegative()) {
6249 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
6250 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
6252 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
6253 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT,
6254 DAG.getNode(ISD::FABS, N0.getDebugLoc(), VT, N0));
6258 // copysign(fabs(x), y) -> copysign(x, y)
6259 // copysign(fneg(x), y) -> copysign(x, y)
6260 // copysign(copysign(x,z), y) -> copysign(x, y)
6261 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
6262 N0.getOpcode() == ISD::FCOPYSIGN)
6263 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
6264 N0.getOperand(0), N1);
6266 // copysign(x, abs(y)) -> abs(x)
6267 if (N1.getOpcode() == ISD::FABS)
6268 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
6270 // copysign(x, copysign(y,z)) -> copysign(x, z)
6271 if (N1.getOpcode() == ISD::FCOPYSIGN)
6272 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
6273 N0, N1.getOperand(1));
6275 // copysign(x, fp_extend(y)) -> copysign(x, y)
6276 // copysign(x, fp_round(y)) -> copysign(x, y)
6277 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
6278 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
6279 N0, N1.getOperand(0));
6284 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
6285 SDValue N0 = N->getOperand(0);
6286 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
6287 EVT VT = N->getValueType(0);
6288 EVT OpVT = N0.getValueType();
6290 // fold (sint_to_fp c1) -> c1fp
6292 // ...but only if the target supports immediate floating-point values
6293 (!LegalOperations ||
6294 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
6295 return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
6297 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
6298 // but UINT_TO_FP is legal on this target, try to convert.
6299 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
6300 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
6301 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
6302 if (DAG.SignBitIsZero(N0))
6303 return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
6306 // The next optimizations are desireable only if SELECT_CC can be lowered.
6307 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
6308 // having to say they don't support SELECT_CC on every type the DAG knows
6309 // about, since there is no way to mark an opcode illegal at all value types
6310 // (See also visitSELECT)
6311 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other)) {
6312 // fold (sint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
6313 if (N0.getOpcode() == ISD::SETCC && N0.getValueType() == MVT::i1 &&
6315 (!LegalOperations ||
6316 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
6318 { N0.getOperand(0), N0.getOperand(1),
6319 DAG.getConstantFP(-1.0, VT) , DAG.getConstantFP(0.0, VT),
6321 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT, Ops, 5);
6324 // fold (sint_to_fp (zext (setcc x, y, cc))) ->
6325 // (select_cc x, y, 1.0, 0.0,, cc)
6326 if (N0.getOpcode() == ISD::ZERO_EXTEND &&
6327 N0.getOperand(0).getOpcode() == ISD::SETCC &&!VT.isVector() &&
6328 (!LegalOperations ||
6329 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
6331 { N0.getOperand(0).getOperand(0), N0.getOperand(0).getOperand(1),
6332 DAG.getConstantFP(1.0, VT) , DAG.getConstantFP(0.0, VT),
6333 N0.getOperand(0).getOperand(2) };
6334 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT, Ops, 5);
6341 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
6342 SDValue N0 = N->getOperand(0);
6343 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
6344 EVT VT = N->getValueType(0);
6345 EVT OpVT = N0.getValueType();
6347 // fold (uint_to_fp c1) -> c1fp
6349 // ...but only if the target supports immediate floating-point values
6350 (!LegalOperations ||
6351 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
6352 return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
6354 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
6355 // but SINT_TO_FP is legal on this target, try to convert.
6356 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
6357 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
6358 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
6359 if (DAG.SignBitIsZero(N0))
6360 return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
6363 // The next optimizations are desireable only if SELECT_CC can be lowered.
6364 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
6365 // having to say they don't support SELECT_CC on every type the DAG knows
6366 // about, since there is no way to mark an opcode illegal at all value types
6367 // (See also visitSELECT)
6368 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other)) {
6369 // fold (uint_to_fp (setcc x, y, cc)) -> (select_cc x, y, -1.0, 0.0,, cc)
6371 if (N0.getOpcode() == ISD::SETCC && !VT.isVector() &&
6372 (!LegalOperations ||
6373 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT))) {
6375 { N0.getOperand(0), N0.getOperand(1),
6376 DAG.getConstantFP(1.0, VT), DAG.getConstantFP(0.0, VT),
6378 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT, Ops, 5);
6385 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
6386 SDValue N0 = N->getOperand(0);
6387 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6388 EVT VT = N->getValueType(0);
6390 // fold (fp_to_sint c1fp) -> c1
6392 return DAG.getNode(ISD::FP_TO_SINT, N->getDebugLoc(), VT, N0);
6397 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
6398 SDValue N0 = N->getOperand(0);
6399 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6400 EVT VT = N->getValueType(0);
6402 // fold (fp_to_uint c1fp) -> c1
6404 return DAG.getNode(ISD::FP_TO_UINT, N->getDebugLoc(), VT, N0);
6409 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
6410 SDValue N0 = N->getOperand(0);
6411 SDValue N1 = N->getOperand(1);
6412 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6413 EVT VT = N->getValueType(0);
6415 // fold (fp_round c1fp) -> c1fp
6417 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT, N0, N1);
6419 // fold (fp_round (fp_extend x)) -> x
6420 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
6421 return N0.getOperand(0);
6423 // fold (fp_round (fp_round x)) -> (fp_round x)
6424 if (N0.getOpcode() == ISD::FP_ROUND) {
6425 // This is a value preserving truncation if both round's are.
6426 bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
6427 N0.getNode()->getConstantOperandVal(1) == 1;
6428 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT, N0.getOperand(0),
6429 DAG.getIntPtrConstant(IsTrunc));
6432 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
6433 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
6434 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(), VT,
6435 N0.getOperand(0), N1);
6436 AddToWorkList(Tmp.getNode());
6437 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
6438 Tmp, N0.getOperand(1));
6444 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
6445 SDValue N0 = N->getOperand(0);
6446 EVT VT = N->getValueType(0);
6447 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
6448 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6450 // fold (fp_round_inreg c1fp) -> c1fp
6451 if (N0CFP && isTypeLegal(EVT)) {
6452 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
6453 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, Round);
6459 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
6460 SDValue N0 = N->getOperand(0);
6461 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6462 EVT VT = N->getValueType(0);
6464 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
6465 if (N->hasOneUse() &&
6466 N->use_begin()->getOpcode() == ISD::FP_ROUND)
6469 // fold (fp_extend c1fp) -> c1fp
6471 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, N0);
6473 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
6475 if (N0.getOpcode() == ISD::FP_ROUND
6476 && N0.getNode()->getConstantOperandVal(1) == 1) {
6477 SDValue In = N0.getOperand(0);
6478 if (In.getValueType() == VT) return In;
6479 if (VT.bitsLT(In.getValueType()))
6480 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT,
6481 In, N0.getOperand(1));
6482 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, In);
6485 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
6486 if (ISD::isNON_EXTLoad(N0.getNode()) && N0.hasOneUse() &&
6487 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6488 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
6489 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6490 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
6492 LN0->getBasePtr(), LN0->getPointerInfo(),
6494 LN0->isVolatile(), LN0->isNonTemporal(),
6495 LN0->getAlignment());
6496 CombineTo(N, ExtLoad);
6497 CombineTo(N0.getNode(),
6498 DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(),
6499 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
6500 ExtLoad.getValue(1));
6501 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6507 SDValue DAGCombiner::visitFNEG(SDNode *N) {
6508 SDValue N0 = N->getOperand(0);
6509 EVT VT = N->getValueType(0);
6511 if (VT.isVector()) {
6512 SDValue FoldedVOp = SimplifyVUnaryOp(N);
6513 if (FoldedVOp.getNode()) return FoldedVOp;
6516 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
6517 &DAG.getTarget().Options))
6518 return GetNegatedExpression(N0, DAG, LegalOperations);
6520 // Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading
6521 // constant pool values.
6522 if (!TLI.isFNegFree(VT) && N0.getOpcode() == ISD::BITCAST &&
6524 N0.getNode()->hasOneUse() &&
6525 N0.getOperand(0).getValueType().isInteger()) {
6526 SDValue Int = N0.getOperand(0);
6527 EVT IntVT = Int.getValueType();
6528 if (IntVT.isInteger() && !IntVT.isVector()) {
6529 Int = DAG.getNode(ISD::XOR, N0.getDebugLoc(), IntVT, Int,
6530 DAG.getConstant(APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
6531 AddToWorkList(Int.getNode());
6532 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
6537 // (fneg (fmul c, x)) -> (fmul -c, x)
6538 if (N0.getOpcode() == ISD::FMUL) {
6539 ConstantFPSDNode *CFP1 = dyn_cast<ConstantFPSDNode>(N0.getOperand(1));
6541 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
6543 DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT,
6551 SDValue DAGCombiner::visitFCEIL(SDNode *N) {
6552 SDValue N0 = N->getOperand(0);
6553 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6554 EVT VT = N->getValueType(0);
6556 // fold (fceil c1) -> fceil(c1)
6558 return DAG.getNode(ISD::FCEIL, N->getDebugLoc(), VT, N0);
6563 SDValue DAGCombiner::visitFTRUNC(SDNode *N) {
6564 SDValue N0 = N->getOperand(0);
6565 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6566 EVT VT = N->getValueType(0);
6568 // fold (ftrunc c1) -> ftrunc(c1)
6570 return DAG.getNode(ISD::FTRUNC, N->getDebugLoc(), VT, N0);
6575 SDValue DAGCombiner::visitFFLOOR(SDNode *N) {
6576 SDValue N0 = N->getOperand(0);
6577 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6578 EVT VT = N->getValueType(0);
6580 // fold (ffloor c1) -> ffloor(c1)
6582 return DAG.getNode(ISD::FFLOOR, N->getDebugLoc(), VT, N0);
6587 SDValue DAGCombiner::visitFABS(SDNode *N) {
6588 SDValue N0 = N->getOperand(0);
6589 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6590 EVT VT = N->getValueType(0);
6592 if (VT.isVector()) {
6593 SDValue FoldedVOp = SimplifyVUnaryOp(N);
6594 if (FoldedVOp.getNode()) return FoldedVOp;
6597 // fold (fabs c1) -> fabs(c1)
6599 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
6600 // fold (fabs (fabs x)) -> (fabs x)
6601 if (N0.getOpcode() == ISD::FABS)
6602 return N->getOperand(0);
6603 // fold (fabs (fneg x)) -> (fabs x)
6604 // fold (fabs (fcopysign x, y)) -> (fabs x)
6605 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
6606 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0.getOperand(0));
6608 // Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading
6609 // constant pool values.
6610 if (!TLI.isFAbsFree(VT) &&
6611 N0.getOpcode() == ISD::BITCAST && N0.getNode()->hasOneUse() &&
6612 N0.getOperand(0).getValueType().isInteger() &&
6613 !N0.getOperand(0).getValueType().isVector()) {
6614 SDValue Int = N0.getOperand(0);
6615 EVT IntVT = Int.getValueType();
6616 if (IntVT.isInteger() && !IntVT.isVector()) {
6617 Int = DAG.getNode(ISD::AND, N0.getDebugLoc(), IntVT, Int,
6618 DAG.getConstant(~APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
6619 AddToWorkList(Int.getNode());
6620 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
6621 N->getValueType(0), Int);
6628 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
6629 SDValue Chain = N->getOperand(0);
6630 SDValue N1 = N->getOperand(1);
6631 SDValue N2 = N->getOperand(2);
6633 // If N is a constant we could fold this into a fallthrough or unconditional
6634 // branch. However that doesn't happen very often in normal code, because
6635 // Instcombine/SimplifyCFG should have handled the available opportunities.
6636 // If we did this folding here, it would be necessary to update the
6637 // MachineBasicBlock CFG, which is awkward.
6639 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
6641 if (N1.getOpcode() == ISD::SETCC &&
6642 TLI.isOperationLegalOrCustom(ISD::BR_CC, MVT::Other)) {
6643 return DAG.getNode(ISD::BR_CC, N->getDebugLoc(), MVT::Other,
6644 Chain, N1.getOperand(2),
6645 N1.getOperand(0), N1.getOperand(1), N2);
6648 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
6649 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
6650 (N1.getOperand(0).hasOneUse() &&
6651 N1.getOperand(0).getOpcode() == ISD::SRL))) {
6653 if (N1.getOpcode() == ISD::TRUNCATE) {
6654 // Look pass the truncate.
6655 Trunc = N1.getNode();
6656 N1 = N1.getOperand(0);
6659 // Match this pattern so that we can generate simpler code:
6662 // %b = and i32 %a, 2
6663 // %c = srl i32 %b, 1
6664 // brcond i32 %c ...
6669 // %b = and i32 %a, 2
6670 // %c = setcc eq %b, 0
6673 // This applies only when the AND constant value has one bit set and the
6674 // SRL constant is equal to the log2 of the AND constant. The back-end is
6675 // smart enough to convert the result into a TEST/JMP sequence.
6676 SDValue Op0 = N1.getOperand(0);
6677 SDValue Op1 = N1.getOperand(1);
6679 if (Op0.getOpcode() == ISD::AND &&
6680 Op1.getOpcode() == ISD::Constant) {
6681 SDValue AndOp1 = Op0.getOperand(1);
6683 if (AndOp1.getOpcode() == ISD::Constant) {
6684 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
6686 if (AndConst.isPowerOf2() &&
6687 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
6689 DAG.getSetCC(N->getDebugLoc(),
6690 TLI.getSetCCResultType(Op0.getValueType()),
6691 Op0, DAG.getConstant(0, Op0.getValueType()),
6694 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
6695 MVT::Other, Chain, SetCC, N2);
6696 // Don't add the new BRCond into the worklist or else SimplifySelectCC
6697 // will convert it back to (X & C1) >> C2.
6698 CombineTo(N, NewBRCond, false);
6699 // Truncate is dead.
6701 removeFromWorkList(Trunc);
6702 DAG.DeleteNode(Trunc);
6704 // Replace the uses of SRL with SETCC
6705 WorkListRemover DeadNodes(*this);
6706 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
6707 removeFromWorkList(N1.getNode());
6708 DAG.DeleteNode(N1.getNode());
6709 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6715 // Restore N1 if the above transformation doesn't match.
6716 N1 = N->getOperand(1);
6719 // Transform br(xor(x, y)) -> br(x != y)
6720 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
6721 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
6722 SDNode *TheXor = N1.getNode();
6723 SDValue Op0 = TheXor->getOperand(0);
6724 SDValue Op1 = TheXor->getOperand(1);
6725 if (Op0.getOpcode() == Op1.getOpcode()) {
6726 // Avoid missing important xor optimizations.
6727 SDValue Tmp = visitXOR(TheXor);
6728 if (Tmp.getNode() && Tmp.getNode() != TheXor) {
6729 DEBUG(dbgs() << "\nReplacing.8 ";
6731 dbgs() << "\nWith: ";
6732 Tmp.getNode()->dump(&DAG);
6734 WorkListRemover DeadNodes(*this);
6735 DAG.ReplaceAllUsesOfValueWith(N1, Tmp);
6736 removeFromWorkList(TheXor);
6737 DAG.DeleteNode(TheXor);
6738 return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
6739 MVT::Other, Chain, Tmp, N2);
6743 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
6745 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
6746 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
6747 Op0.getOpcode() == ISD::XOR) {
6748 TheXor = Op0.getNode();
6752 EVT SetCCVT = N1.getValueType();
6754 SetCCVT = TLI.getSetCCResultType(SetCCVT);
6755 SDValue SetCC = DAG.getSetCC(TheXor->getDebugLoc(),
6758 Equal ? ISD::SETEQ : ISD::SETNE);
6759 // Replace the uses of XOR with SETCC
6760 WorkListRemover DeadNodes(*this);
6761 DAG.ReplaceAllUsesOfValueWith(N1, SetCC);
6762 removeFromWorkList(N1.getNode());
6763 DAG.DeleteNode(N1.getNode());
6764 return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
6765 MVT::Other, Chain, SetCC, N2);
6772 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
6774 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
6775 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
6776 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
6778 // If N is a constant we could fold this into a fallthrough or unconditional
6779 // branch. However that doesn't happen very often in normal code, because
6780 // Instcombine/SimplifyCFG should have handled the available opportunities.
6781 // If we did this folding here, it would be necessary to update the
6782 // MachineBasicBlock CFG, which is awkward.
6784 // Use SimplifySetCC to simplify SETCC's.
6785 SDValue Simp = SimplifySetCC(TLI.getSetCCResultType(CondLHS.getValueType()),
6786 CondLHS, CondRHS, CC->get(), N->getDebugLoc(),
6788 if (Simp.getNode()) AddToWorkList(Simp.getNode());
6790 // fold to a simpler setcc
6791 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
6792 return DAG.getNode(ISD::BR_CC, N->getDebugLoc(), MVT::Other,
6793 N->getOperand(0), Simp.getOperand(2),
6794 Simp.getOperand(0), Simp.getOperand(1),
6800 /// canFoldInAddressingMode - Return true if 'Use' is a load or a store that
6801 /// uses N as its base pointer and that N may be folded in the load / store
6802 /// addressing mode.
6803 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
6805 const TargetLowering &TLI) {
6807 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
6808 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
6810 VT = Use->getValueType(0);
6811 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
6812 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
6814 VT = ST->getValue().getValueType();
6819 if (N->getOpcode() == ISD::ADD) {
6820 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
6823 AM.BaseOffs = Offset->getSExtValue();
6827 } else if (N->getOpcode() == ISD::SUB) {
6828 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
6831 AM.BaseOffs = -Offset->getSExtValue();
6838 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
6841 /// CombineToPreIndexedLoadStore - Try turning a load / store into a
6842 /// pre-indexed load / store when the base pointer is an add or subtract
6843 /// and it has other uses besides the load / store. After the
6844 /// transformation, the new indexed load / store has effectively folded
6845 /// the add / subtract in and all of its other uses are redirected to the
6846 /// new load / store.
6847 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
6848 if (Level < AfterLegalizeDAG)
6854 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6855 if (LD->isIndexed())
6857 VT = LD->getMemoryVT();
6858 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
6859 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
6861 Ptr = LD->getBasePtr();
6862 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6863 if (ST->isIndexed())
6865 VT = ST->getMemoryVT();
6866 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
6867 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
6869 Ptr = ST->getBasePtr();
6875 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
6876 // out. There is no reason to make this a preinc/predec.
6877 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
6878 Ptr.getNode()->hasOneUse())
6881 // Ask the target to do addressing mode selection.
6884 ISD::MemIndexedMode AM = ISD::UNINDEXED;
6885 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
6887 // Don't create a indexed load / store with zero offset.
6888 if (isa<ConstantSDNode>(Offset) &&
6889 cast<ConstantSDNode>(Offset)->isNullValue())
6892 // Try turning it into a pre-indexed load / store except when:
6893 // 1) The new base ptr is a frame index.
6894 // 2) If N is a store and the new base ptr is either the same as or is a
6895 // predecessor of the value being stored.
6896 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
6897 // that would create a cycle.
6898 // 4) All uses are load / store ops that use it as old base ptr.
6900 // Check #1. Preinc'ing a frame index would require copying the stack pointer
6901 // (plus the implicit offset) to a register to preinc anyway.
6902 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
6907 SDValue Val = cast<StoreSDNode>(N)->getValue();
6908 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
6912 // Now check for #3 and #4.
6913 bool RealUse = false;
6915 // Caches for hasPredecessorHelper
6916 SmallPtrSet<const SDNode *, 32> Visited;
6917 SmallVector<const SDNode *, 16> Worklist;
6919 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
6920 E = Ptr.getNode()->use_end(); I != E; ++I) {
6924 if (N->hasPredecessorHelper(Use, Visited, Worklist))
6927 // If Ptr may be folded in addressing mode of other use, then it's
6928 // not profitable to do this transformation.
6929 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
6938 Result = DAG.getIndexedLoad(SDValue(N,0), N->getDebugLoc(),
6939 BasePtr, Offset, AM);
6941 Result = DAG.getIndexedStore(SDValue(N,0), N->getDebugLoc(),
6942 BasePtr, Offset, AM);
6945 DEBUG(dbgs() << "\nReplacing.4 ";
6947 dbgs() << "\nWith: ";
6948 Result.getNode()->dump(&DAG);
6950 WorkListRemover DeadNodes(*this);
6952 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
6953 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
6955 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
6958 // Finally, since the node is now dead, remove it from the graph.
6961 // Replace the uses of Ptr with uses of the updated base value.
6962 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0));
6963 removeFromWorkList(Ptr.getNode());
6964 DAG.DeleteNode(Ptr.getNode());
6969 /// CombineToPostIndexedLoadStore - Try to combine a load / store with a
6970 /// add / sub of the base pointer node into a post-indexed load / store.
6971 /// The transformation folded the add / subtract into the new indexed
6972 /// load / store effectively and all of its uses are redirected to the
6973 /// new load / store.
6974 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
6975 if (Level < AfterLegalizeDAG)
6981 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6982 if (LD->isIndexed())
6984 VT = LD->getMemoryVT();
6985 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
6986 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
6988 Ptr = LD->getBasePtr();
6989 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6990 if (ST->isIndexed())
6992 VT = ST->getMemoryVT();
6993 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
6994 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
6996 Ptr = ST->getBasePtr();
7002 if (Ptr.getNode()->hasOneUse())
7005 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
7006 E = Ptr.getNode()->use_end(); I != E; ++I) {
7009 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
7014 ISD::MemIndexedMode AM = ISD::UNINDEXED;
7015 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
7016 // Don't create a indexed load / store with zero offset.
7017 if (isa<ConstantSDNode>(Offset) &&
7018 cast<ConstantSDNode>(Offset)->isNullValue())
7021 // Try turning it into a post-indexed load / store except when
7022 // 1) All uses are load / store ops that use it as base ptr (and
7023 // it may be folded as addressing mmode).
7024 // 2) Op must be independent of N, i.e. Op is neither a predecessor
7025 // nor a successor of N. Otherwise, if Op is folded that would
7028 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
7032 bool TryNext = false;
7033 for (SDNode::use_iterator II = BasePtr.getNode()->use_begin(),
7034 EE = BasePtr.getNode()->use_end(); II != EE; ++II) {
7036 if (Use == Ptr.getNode())
7039 // If all the uses are load / store addresses, then don't do the
7041 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
7042 bool RealUse = false;
7043 for (SDNode::use_iterator III = Use->use_begin(),
7044 EEE = Use->use_end(); III != EEE; ++III) {
7045 SDNode *UseUse = *III;
7046 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
7061 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
7062 SDValue Result = isLoad
7063 ? DAG.getIndexedLoad(SDValue(N,0), N->getDebugLoc(),
7064 BasePtr, Offset, AM)
7065 : DAG.getIndexedStore(SDValue(N,0), N->getDebugLoc(),
7066 BasePtr, Offset, AM);
7069 DEBUG(dbgs() << "\nReplacing.5 ";
7071 dbgs() << "\nWith: ";
7072 Result.getNode()->dump(&DAG);
7074 WorkListRemover DeadNodes(*this);
7076 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0));
7077 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2));
7079 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1));
7082 // Finally, since the node is now dead, remove it from the graph.
7085 // Replace the uses of Use with uses of the updated base value.
7086 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
7087 Result.getValue(isLoad ? 1 : 0));
7088 removeFromWorkList(Op);
7098 SDValue DAGCombiner::visitLOAD(SDNode *N) {
7099 LoadSDNode *LD = cast<LoadSDNode>(N);
7100 SDValue Chain = LD->getChain();
7101 SDValue Ptr = LD->getBasePtr();
7103 // If load is not volatile and there are no uses of the loaded value (and
7104 // the updated indexed value in case of indexed loads), change uses of the
7105 // chain value into uses of the chain input (i.e. delete the dead load).
7106 if (!LD->isVolatile()) {
7107 if (N->getValueType(1) == MVT::Other) {
7109 if (!N->hasAnyUseOfValue(0)) {
7110 // It's not safe to use the two value CombineTo variant here. e.g.
7111 // v1, chain2 = load chain1, loc
7112 // v2, chain3 = load chain2, loc
7114 // Now we replace use of chain2 with chain1. This makes the second load
7115 // isomorphic to the one we are deleting, and thus makes this load live.
7116 DEBUG(dbgs() << "\nReplacing.6 ";
7118 dbgs() << "\nWith chain: ";
7119 Chain.getNode()->dump(&DAG);
7121 WorkListRemover DeadNodes(*this);
7122 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain);
7124 if (N->use_empty()) {
7125 removeFromWorkList(N);
7129 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7133 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
7134 if (!N->hasAnyUseOfValue(0) && !N->hasAnyUseOfValue(1)) {
7135 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
7136 DEBUG(dbgs() << "\nReplacing.7 ";
7138 dbgs() << "\nWith: ";
7139 Undef.getNode()->dump(&DAG);
7140 dbgs() << " and 2 other values\n");
7141 WorkListRemover DeadNodes(*this);
7142 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef);
7143 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1),
7144 DAG.getUNDEF(N->getValueType(1)));
7145 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain);
7146 removeFromWorkList(N);
7148 return SDValue(N, 0); // Return N so it doesn't get rechecked!
7153 // If this load is directly stored, replace the load value with the stored
7155 // TODO: Handle store large -> read small portion.
7156 // TODO: Handle TRUNCSTORE/LOADEXT
7157 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
7158 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
7159 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
7160 if (PrevST->getBasePtr() == Ptr &&
7161 PrevST->getValue().getValueType() == N->getValueType(0))
7162 return CombineTo(N, Chain.getOperand(1), Chain);
7166 // Try to infer better alignment information than the load already has.
7167 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
7168 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
7169 if (Align > LD->getAlignment())
7170 return DAG.getExtLoad(LD->getExtensionType(), N->getDebugLoc(),
7171 LD->getValueType(0),
7172 Chain, Ptr, LD->getPointerInfo(),
7174 LD->isVolatile(), LD->isNonTemporal(), Align);
7179 // Walk up chain skipping non-aliasing memory nodes.
7180 SDValue BetterChain = FindBetterChain(N, Chain);
7182 // If there is a better chain.
7183 if (Chain != BetterChain) {
7186 // Replace the chain to void dependency.
7187 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
7188 ReplLoad = DAG.getLoad(N->getValueType(0), LD->getDebugLoc(),
7189 BetterChain, Ptr, LD->getPointerInfo(),
7190 LD->isVolatile(), LD->isNonTemporal(),
7191 LD->isInvariant(), LD->getAlignment());
7193 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), LD->getDebugLoc(),
7194 LD->getValueType(0),
7195 BetterChain, Ptr, LD->getPointerInfo(),
7198 LD->isNonTemporal(),
7199 LD->getAlignment());
7202 // Create token factor to keep old chain connected.
7203 SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
7204 MVT::Other, Chain, ReplLoad.getValue(1));
7206 // Make sure the new and old chains are cleaned up.
7207 AddToWorkList(Token.getNode());
7209 // Replace uses with load result and token factor. Don't add users
7211 return CombineTo(N, ReplLoad.getValue(0), Token, false);
7215 // Try transforming N to an indexed load.
7216 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
7217 return SDValue(N, 0);
7222 /// CheckForMaskedLoad - Check to see if V is (and load (ptr), imm), where the
7223 /// load is having specific bytes cleared out. If so, return the byte size
7224 /// being masked out and the shift amount.
7225 static std::pair<unsigned, unsigned>
7226 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
7227 std::pair<unsigned, unsigned> Result(0, 0);
7229 // Check for the structure we're looking for.
7230 if (V->getOpcode() != ISD::AND ||
7231 !isa<ConstantSDNode>(V->getOperand(1)) ||
7232 !ISD::isNormalLoad(V->getOperand(0).getNode()))
7235 // Check the chain and pointer.
7236 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
7237 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
7239 // The store should be chained directly to the load or be an operand of a
7241 if (LD == Chain.getNode())
7243 else if (Chain->getOpcode() != ISD::TokenFactor)
7244 return Result; // Fail.
7247 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
7248 if (Chain->getOperand(i).getNode() == LD) {
7252 if (!isOk) return Result;
7255 // This only handles simple types.
7256 if (V.getValueType() != MVT::i16 &&
7257 V.getValueType() != MVT::i32 &&
7258 V.getValueType() != MVT::i64)
7261 // Check the constant mask. Invert it so that the bits being masked out are
7262 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
7263 // follow the sign bit for uniformity.
7264 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
7265 unsigned NotMaskLZ = CountLeadingZeros_64(NotMask);
7266 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
7267 unsigned NotMaskTZ = CountTrailingZeros_64(NotMask);
7268 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
7269 if (NotMaskLZ == 64) return Result; // All zero mask.
7271 // See if we have a continuous run of bits. If so, we have 0*1+0*
7272 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
7275 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
7276 if (V.getValueType() != MVT::i64 && NotMaskLZ)
7277 NotMaskLZ -= 64-V.getValueSizeInBits();
7279 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
7280 switch (MaskedBytes) {
7284 default: return Result; // All one mask, or 5-byte mask.
7287 // Verify that the first bit starts at a multiple of mask so that the access
7288 // is aligned the same as the access width.
7289 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
7291 Result.first = MaskedBytes;
7292 Result.second = NotMaskTZ/8;
7297 /// ShrinkLoadReplaceStoreWithStore - Check to see if IVal is something that
7298 /// provides a value as specified by MaskInfo. If so, replace the specified
7299 /// store with a narrower store of truncated IVal.
7301 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
7302 SDValue IVal, StoreSDNode *St,
7304 unsigned NumBytes = MaskInfo.first;
7305 unsigned ByteShift = MaskInfo.second;
7306 SelectionDAG &DAG = DC->getDAG();
7308 // Check to see if IVal is all zeros in the part being masked in by the 'or'
7309 // that uses this. If not, this is not a replacement.
7310 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
7311 ByteShift*8, (ByteShift+NumBytes)*8);
7312 if (!DAG.MaskedValueIsZero(IVal, Mask)) return 0;
7314 // Check that it is legal on the target to do this. It is legal if the new
7315 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
7317 MVT VT = MVT::getIntegerVT(NumBytes*8);
7318 if (!DC->isTypeLegal(VT))
7321 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
7322 // shifted by ByteShift and truncated down to NumBytes.
7324 IVal = DAG.getNode(ISD::SRL, IVal->getDebugLoc(), IVal.getValueType(), IVal,
7325 DAG.getConstant(ByteShift*8,
7326 DC->getShiftAmountTy(IVal.getValueType())));
7328 // Figure out the offset for the store and the alignment of the access.
7330 unsigned NewAlign = St->getAlignment();
7332 if (DAG.getTargetLoweringInfo().isLittleEndian())
7333 StOffset = ByteShift;
7335 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
7337 SDValue Ptr = St->getBasePtr();
7339 Ptr = DAG.getNode(ISD::ADD, IVal->getDebugLoc(), Ptr.getValueType(),
7340 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
7341 NewAlign = MinAlign(NewAlign, StOffset);
7344 // Truncate down to the new size.
7345 IVal = DAG.getNode(ISD::TRUNCATE, IVal->getDebugLoc(), VT, IVal);
7348 return DAG.getStore(St->getChain(), St->getDebugLoc(), IVal, Ptr,
7349 St->getPointerInfo().getWithOffset(StOffset),
7350 false, false, NewAlign).getNode();
7354 /// ReduceLoadOpStoreWidth - Look for sequence of load / op / store where op is
7355 /// one of 'or', 'xor', and 'and' of immediates. If 'op' is only touching some
7356 /// of the loaded bits, try narrowing the load and store if it would end up
7357 /// being a win for performance or code size.
7358 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
7359 StoreSDNode *ST = cast<StoreSDNode>(N);
7360 if (ST->isVolatile())
7363 SDValue Chain = ST->getChain();
7364 SDValue Value = ST->getValue();
7365 SDValue Ptr = ST->getBasePtr();
7366 EVT VT = Value.getValueType();
7368 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
7371 unsigned Opc = Value.getOpcode();
7373 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
7374 // is a byte mask indicating a consecutive number of bytes, check to see if
7375 // Y is known to provide just those bytes. If so, we try to replace the
7376 // load + replace + store sequence with a single (narrower) store, which makes
7378 if (Opc == ISD::OR) {
7379 std::pair<unsigned, unsigned> MaskedLoad;
7380 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
7381 if (MaskedLoad.first)
7382 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
7383 Value.getOperand(1), ST,this))
7384 return SDValue(NewST, 0);
7386 // Or is commutative, so try swapping X and Y.
7387 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
7388 if (MaskedLoad.first)
7389 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
7390 Value.getOperand(0), ST,this))
7391 return SDValue(NewST, 0);
7394 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
7395 Value.getOperand(1).getOpcode() != ISD::Constant)
7398 SDValue N0 = Value.getOperand(0);
7399 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
7400 Chain == SDValue(N0.getNode(), 1)) {
7401 LoadSDNode *LD = cast<LoadSDNode>(N0);
7402 if (LD->getBasePtr() != Ptr ||
7403 LD->getPointerInfo().getAddrSpace() !=
7404 ST->getPointerInfo().getAddrSpace())
7407 // Find the type to narrow it the load / op / store to.
7408 SDValue N1 = Value.getOperand(1);
7409 unsigned BitWidth = N1.getValueSizeInBits();
7410 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
7411 if (Opc == ISD::AND)
7412 Imm ^= APInt::getAllOnesValue(BitWidth);
7413 if (Imm == 0 || Imm.isAllOnesValue())
7415 unsigned ShAmt = Imm.countTrailingZeros();
7416 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
7417 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
7418 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
7419 while (NewBW < BitWidth &&
7420 !(TLI.isOperationLegalOrCustom(Opc, NewVT) &&
7421 TLI.isNarrowingProfitable(VT, NewVT))) {
7422 NewBW = NextPowerOf2(NewBW);
7423 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
7425 if (NewBW >= BitWidth)
7428 // If the lsb changed does not start at the type bitwidth boundary,
7429 // start at the previous one.
7431 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
7432 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt, ShAmt + NewBW);
7433 if ((Imm & Mask) == Imm) {
7434 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
7435 if (Opc == ISD::AND)
7436 NewImm ^= APInt::getAllOnesValue(NewBW);
7437 uint64_t PtrOff = ShAmt / 8;
7438 // For big endian targets, we need to adjust the offset to the pointer to
7439 // load the correct bytes.
7440 if (TLI.isBigEndian())
7441 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
7443 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
7444 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
7445 if (NewAlign < TLI.getDataLayout()->getABITypeAlignment(NewVTTy))
7448 SDValue NewPtr = DAG.getNode(ISD::ADD, LD->getDebugLoc(),
7449 Ptr.getValueType(), Ptr,
7450 DAG.getConstant(PtrOff, Ptr.getValueType()));
7451 SDValue NewLD = DAG.getLoad(NewVT, N0.getDebugLoc(),
7452 LD->getChain(), NewPtr,
7453 LD->getPointerInfo().getWithOffset(PtrOff),
7454 LD->isVolatile(), LD->isNonTemporal(),
7455 LD->isInvariant(), NewAlign);
7456 SDValue NewVal = DAG.getNode(Opc, Value.getDebugLoc(), NewVT, NewLD,
7457 DAG.getConstant(NewImm, NewVT));
7458 SDValue NewST = DAG.getStore(Chain, N->getDebugLoc(),
7460 ST->getPointerInfo().getWithOffset(PtrOff),
7461 false, false, NewAlign);
7463 AddToWorkList(NewPtr.getNode());
7464 AddToWorkList(NewLD.getNode());
7465 AddToWorkList(NewVal.getNode());
7466 WorkListRemover DeadNodes(*this);
7467 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1));
7476 /// TransformFPLoadStorePair - For a given floating point load / store pair,
7477 /// if the load value isn't used by any other operations, then consider
7478 /// transforming the pair to integer load / store operations if the target
7479 /// deems the transformation profitable.
7480 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
7481 StoreSDNode *ST = cast<StoreSDNode>(N);
7482 SDValue Chain = ST->getChain();
7483 SDValue Value = ST->getValue();
7484 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
7485 Value.hasOneUse() &&
7486 Chain == SDValue(Value.getNode(), 1)) {
7487 LoadSDNode *LD = cast<LoadSDNode>(Value);
7488 EVT VT = LD->getMemoryVT();
7489 if (!VT.isFloatingPoint() ||
7490 VT != ST->getMemoryVT() ||
7491 LD->isNonTemporal() ||
7492 ST->isNonTemporal() ||
7493 LD->getPointerInfo().getAddrSpace() != 0 ||
7494 ST->getPointerInfo().getAddrSpace() != 0)
7497 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
7498 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
7499 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
7500 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
7501 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
7504 unsigned LDAlign = LD->getAlignment();
7505 unsigned STAlign = ST->getAlignment();
7506 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
7507 unsigned ABIAlign = TLI.getDataLayout()->getABITypeAlignment(IntVTTy);
7508 if (LDAlign < ABIAlign || STAlign < ABIAlign)
7511 SDValue NewLD = DAG.getLoad(IntVT, Value.getDebugLoc(),
7512 LD->getChain(), LD->getBasePtr(),
7513 LD->getPointerInfo(),
7514 false, false, false, LDAlign);
7516 SDValue NewST = DAG.getStore(NewLD.getValue(1), N->getDebugLoc(),
7517 NewLD, ST->getBasePtr(),
7518 ST->getPointerInfo(),
7519 false, false, STAlign);
7521 AddToWorkList(NewLD.getNode());
7522 AddToWorkList(NewST.getNode());
7523 WorkListRemover DeadNodes(*this);
7524 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1));
7532 /// Returns the base pointer and an integer offset from that object.
7533 static std::pair<SDValue, int64_t> GetPointerBaseAndOffset(SDValue Ptr) {
7534 if (Ptr->getOpcode() == ISD::ADD && isa<ConstantSDNode>(Ptr->getOperand(1))) {
7535 int64_t Offset = cast<ConstantSDNode>(Ptr->getOperand(1))->getSExtValue();
7536 SDValue Base = Ptr->getOperand(0);
7537 return std::make_pair(Base, Offset);
7540 return std::make_pair(Ptr, 0);
7543 /// Holds a pointer to an LSBaseSDNode as well as information on where it
7544 /// is located in a sequence of memory operations connected by a chain.
7546 MemOpLink (LSBaseSDNode *N, int64_t Offset, unsigned Seq):
7547 MemNode(N), OffsetFromBase(Offset), SequenceNum(Seq) { }
7548 // Ptr to the mem node.
7549 LSBaseSDNode *MemNode;
7550 // Offset from the base ptr.
7551 int64_t OffsetFromBase;
7552 // What is the sequence number of this mem node.
7553 // Lowest mem operand in the DAG starts at zero.
7554 unsigned SequenceNum;
7557 /// Sorts store nodes in a link according to their offset from a shared
7559 struct ConsecutiveMemoryChainSorter {
7560 bool operator()(MemOpLink LHS, MemOpLink RHS) {
7561 return LHS.OffsetFromBase < RHS.OffsetFromBase;
7565 bool DAGCombiner::MergeConsecutiveStores(StoreSDNode* St) {
7566 EVT MemVT = St->getMemoryVT();
7567 int64_t ElementSizeBytes = MemVT.getSizeInBits()/8;
7569 // Don't merge vectors into wider inputs.
7570 if (MemVT.isVector() || !MemVT.isSimple())
7573 // Perform an early exit check. Do not bother looking at stored values that
7574 // are not constants or loads.
7575 SDValue StoredVal = St->getValue();
7576 bool IsLoadSrc = isa<LoadSDNode>(StoredVal);
7577 if (!isa<ConstantSDNode>(StoredVal) && !isa<ConstantFPSDNode>(StoredVal) &&
7581 // Only look at ends of store sequences.
7582 SDValue Chain = SDValue(St, 1);
7583 if (Chain->hasOneUse() && Chain->use_begin()->getOpcode() == ISD::STORE)
7586 // This holds the base pointer and the offset in bytes from the base pointer.
7587 std::pair<SDValue, int64_t> BasePtr =
7588 GetPointerBaseAndOffset(St->getBasePtr());
7590 // We must have a base and an offset.
7591 if (!BasePtr.first.getNode())
7594 // Do not handle stores to undef base pointers.
7595 if (BasePtr.first.getOpcode() == ISD::UNDEF)
7598 // Save the LoadSDNodes that we find in the chain.
7599 // We need to make sure that these nodes do not interfere with
7600 // any of the store nodes.
7601 SmallVector<LSBaseSDNode*, 8> AliasLoadNodes;
7603 // Save the StoreSDNodes that we find in the chain.
7604 SmallVector<MemOpLink, 8> StoreNodes;
7606 // Walk up the chain and look for nodes with offsets from the same
7607 // base pointer. Stop when reaching an instruction with a different kind
7608 // or instruction which has a different base pointer.
7610 StoreSDNode *Index = St;
7612 // If the chain has more than one use, then we can't reorder the mem ops.
7613 if (Index != St && !SDValue(Index, 1)->hasOneUse())
7616 // Find the base pointer and offset for this memory node.
7617 std::pair<SDValue, int64_t> Ptr =
7618 GetPointerBaseAndOffset(Index->getBasePtr());
7620 // Check that the base pointer is the same as the original one.
7621 if (Ptr.first.getNode() != BasePtr.first.getNode())
7624 // Check that the alignment is the same.
7625 if (Index->getAlignment() != St->getAlignment())
7628 // The memory operands must not be volatile.
7629 if (Index->isVolatile() || Index->isIndexed())
7633 if (StoreSDNode *St = dyn_cast<StoreSDNode>(Index))
7634 if (St->isTruncatingStore())
7637 // The stored memory type must be the same.
7638 if (Index->getMemoryVT() != MemVT)
7641 // We do not allow unaligned stores because we want to prevent overriding
7643 if (Index->getAlignment()*8 != MemVT.getSizeInBits())
7646 // We found a potential memory operand to merge.
7647 StoreNodes.push_back(MemOpLink(Index, Ptr.second, Seq++));
7649 // Find the next memory operand in the chain. If the next operand in the
7650 // chain is a store then move up and continue the scan with the next
7651 // memory operand. If the next operand is a load save it and use alias
7652 // information to check if it interferes with anything.
7653 SDNode *NextInChain = Index->getChain().getNode();
7655 if (StoreSDNode *STn = dyn_cast<StoreSDNode>(NextInChain)) {
7656 // We found a store node. Use it for the next iteration.
7659 } else if (LoadSDNode *Ldn = dyn_cast<LoadSDNode>(NextInChain)) {
7660 // Save the load node for later. Continue the scan.
7661 AliasLoadNodes.push_back(Ldn);
7662 NextInChain = Ldn->getChain().getNode();
7671 // Check if there is anything to merge.
7672 if (StoreNodes.size() < 2)
7675 // Sort the memory operands according to their distance from the base pointer.
7676 std::sort(StoreNodes.begin(), StoreNodes.end(),
7677 ConsecutiveMemoryChainSorter());
7679 // Scan the memory operations on the chain and find the first non-consecutive
7680 // store memory address.
7681 unsigned LastConsecutiveStore = 0;
7682 int64_t StartAddress = StoreNodes[0].OffsetFromBase;
7683 for (unsigned i = 0, e = StoreNodes.size(); i < e; ++i) {
7685 // Check that the addresses are consecutive starting from the second
7686 // element in the list of stores.
7688 int64_t CurrAddress = StoreNodes[i].OffsetFromBase;
7689 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
7694 // Check if this store interferes with any of the loads that we found.
7695 for (unsigned ld = 0, lde = AliasLoadNodes.size(); ld < lde; ++ld)
7696 if (isAlias(AliasLoadNodes[ld], StoreNodes[i].MemNode)) {
7700 // We found a load that alias with this store. Stop the sequence.
7704 // Mark this node as useful.
7705 LastConsecutiveStore = i;
7708 // The node with the lowest store address.
7709 LSBaseSDNode *FirstInChain = StoreNodes[0].MemNode;
7711 // Store the constants into memory as one consecutive store.
7713 unsigned LastLegalType = 0;
7714 unsigned LastLegalVectorType = 0;
7715 bool NonZero = false;
7716 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
7717 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
7718 SDValue StoredVal = St->getValue();
7720 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(StoredVal)) {
7721 NonZero |= !C->isNullValue();
7722 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(StoredVal)) {
7723 NonZero |= !C->getConstantFPValue()->isNullValue();
7729 // Find a legal type for the constant store.
7730 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
7731 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
7732 if (TLI.isTypeLegal(StoreTy))
7733 LastLegalType = i+1;
7735 // Find a legal type for the vector store.
7736 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
7737 if (TLI.isTypeLegal(Ty))
7738 LastLegalVectorType = i + 1;
7741 // We only use vectors if the constant is known to be zero.
7743 LastLegalVectorType = 0;
7745 // Check if we found a legal integer type to store.
7746 if (LastLegalType == 0 && LastLegalVectorType == 0)
7749 bool UseVector = LastLegalVectorType > LastLegalType;
7750 unsigned NumElem = UseVector ? LastLegalVectorType : LastLegalType;
7752 // Make sure we have something to merge.
7756 unsigned EarliestNodeUsed = 0;
7757 for (unsigned i=0; i < NumElem; ++i) {
7758 // Find a chain for the new wide-store operand. Notice that some
7759 // of the store nodes that we found may not be selected for inclusion
7760 // in the wide store. The chain we use needs to be the chain of the
7761 // earliest store node which is *used* and replaced by the wide store.
7762 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
7763 EarliestNodeUsed = i;
7766 // The earliest Node in the DAG.
7767 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
7768 DebugLoc DL = StoreNodes[0].MemNode->getDebugLoc();
7772 // Find a legal type for the vector store.
7773 EVT Ty = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
7774 assert(TLI.isTypeLegal(Ty) && "Illegal vector store");
7775 StoredVal = DAG.getConstant(0, Ty);
7777 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
7778 APInt StoreInt(StoreBW, 0);
7780 // Construct a single integer constant which is made of the smaller
7782 bool IsLE = TLI.isLittleEndian();
7783 for (unsigned i = 0; i < NumElem ; ++i) {
7784 unsigned Idx = IsLE ?(NumElem - 1 - i) : i;
7785 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[Idx].MemNode);
7786 SDValue Val = St->getValue();
7787 StoreInt<<=ElementSizeBytes*8;
7788 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Val)) {
7789 StoreInt|=C->getAPIntValue().zext(StoreBW);
7790 } else if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Val)) {
7791 StoreInt|= C->getValueAPF().bitcastToAPInt().zext(StoreBW);
7793 assert(false && "Invalid constant element type");
7797 // Create the new Load and Store operations.
7798 EVT StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
7799 StoredVal = DAG.getConstant(StoreInt, StoreTy);
7802 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), DL, StoredVal,
7803 FirstInChain->getBasePtr(),
7804 FirstInChain->getPointerInfo(),
7806 FirstInChain->getAlignment());
7808 // Replace the first store with the new store
7809 CombineTo(EarliestOp, NewStore);
7810 // Erase all other stores.
7811 for (unsigned i = 0; i < NumElem ; ++i) {
7812 if (StoreNodes[i].MemNode == EarliestOp)
7814 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
7815 // ReplaceAllUsesWith will replace all uses that existed when it was
7816 // called, but graph optimizations may cause new ones to appear. For
7817 // example, the case in pr14333 looks like
7819 // St's chain -> St -> another store -> X
7821 // And the only difference from St to the other store is the chain.
7822 // When we change it's chain to be St's chain they become identical,
7823 // get CSEed and the net result is that X is now a use of St.
7824 // Since we know that St is redundant, just iterate.
7825 while (!St->use_empty())
7826 DAG.ReplaceAllUsesWith(SDValue(St, 0), St->getChain());
7827 removeFromWorkList(St);
7834 // Below we handle the case of multiple consecutive stores that
7835 // come from multiple consecutive loads. We merge them into a single
7836 // wide load and a single wide store.
7838 // Look for load nodes which are used by the stored values.
7839 SmallVector<MemOpLink, 8> LoadNodes;
7841 // Find acceptable loads. Loads need to have the same chain (token factor),
7842 // must not be zext, volatile, indexed, and they must be consecutive.
7844 for (unsigned i=0; i<LastConsecutiveStore+1; ++i) {
7845 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
7846 LoadSDNode *Ld = dyn_cast<LoadSDNode>(St->getValue());
7849 // Loads must only have one use.
7850 if (!Ld->hasNUsesOfValue(1, 0))
7853 // Check that the alignment is the same as the stores.
7854 if (Ld->getAlignment() != St->getAlignment())
7857 // The memory operands must not be volatile.
7858 if (Ld->isVolatile() || Ld->isIndexed())
7861 // We do not accept ext loads.
7862 if (Ld->getExtensionType() != ISD::NON_EXTLOAD)
7865 // The stored memory type must be the same.
7866 if (Ld->getMemoryVT() != MemVT)
7869 std::pair<SDValue, int64_t> LdPtr =
7870 GetPointerBaseAndOffset(Ld->getBasePtr());
7872 // If this is not the first ptr that we check.
7873 if (LdBasePtr.getNode()) {
7874 // The base ptr must be the same.
7875 if (LdPtr.first != LdBasePtr)
7878 // Check that all other base pointers are the same as this one.
7879 LdBasePtr = LdPtr.first;
7882 // We found a potential memory operand to merge.
7883 LoadNodes.push_back(MemOpLink(Ld, LdPtr.second, 0));
7886 if (LoadNodes.size() < 2)
7889 // Scan the memory operations on the chain and find the first non-consecutive
7890 // load memory address. These variables hold the index in the store node
7892 unsigned LastConsecutiveLoad = 0;
7893 // This variable refers to the size and not index in the array.
7894 unsigned LastLegalVectorType = 0;
7895 unsigned LastLegalIntegerType = 0;
7896 StartAddress = LoadNodes[0].OffsetFromBase;
7897 SDValue FirstChain = LoadNodes[0].MemNode->getChain();
7898 for (unsigned i = 1; i < LoadNodes.size(); ++i) {
7899 // All loads much share the same chain.
7900 if (LoadNodes[i].MemNode->getChain() != FirstChain)
7903 int64_t CurrAddress = LoadNodes[i].OffsetFromBase;
7904 if (CurrAddress - StartAddress != (ElementSizeBytes * i))
7906 LastConsecutiveLoad = i;
7908 // Find a legal type for the vector store.
7909 EVT StoreTy = EVT::getVectorVT(*DAG.getContext(), MemVT, i+1);
7910 if (TLI.isTypeLegal(StoreTy))
7911 LastLegalVectorType = i + 1;
7913 // Find a legal type for the integer store.
7914 unsigned StoreBW = (i+1) * ElementSizeBytes * 8;
7915 StoreTy = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
7916 if (TLI.isTypeLegal(StoreTy))
7917 LastLegalIntegerType = i + 1;
7920 // Only use vector types if the vector type is larger than the integer type.
7921 // If they are the same, use integers.
7922 bool UseVectorTy = LastLegalVectorType > LastLegalIntegerType;
7923 unsigned LastLegalType = std::max(LastLegalVectorType, LastLegalIntegerType);
7925 // We add +1 here because the LastXXX variables refer to location while
7926 // the NumElem refers to array/index size.
7927 unsigned NumElem = std::min(LastConsecutiveStore, LastConsecutiveLoad) + 1;
7928 NumElem = std::min(LastLegalType, NumElem);
7933 // The earliest Node in the DAG.
7934 unsigned EarliestNodeUsed = 0;
7935 LSBaseSDNode *EarliestOp = StoreNodes[EarliestNodeUsed].MemNode;
7936 for (unsigned i=1; i<NumElem; ++i) {
7937 // Find a chain for the new wide-store operand. Notice that some
7938 // of the store nodes that we found may not be selected for inclusion
7939 // in the wide store. The chain we use needs to be the chain of the
7940 // earliest store node which is *used* and replaced by the wide store.
7941 if (StoreNodes[i].SequenceNum > StoreNodes[EarliestNodeUsed].SequenceNum)
7942 EarliestNodeUsed = i;
7945 // Find if it is better to use vectors or integers to load and store
7949 JointMemOpVT = EVT::getVectorVT(*DAG.getContext(), MemVT, NumElem);
7951 unsigned StoreBW = NumElem * ElementSizeBytes * 8;
7952 JointMemOpVT = EVT::getIntegerVT(*DAG.getContext(), StoreBW);
7955 DebugLoc LoadDL = LoadNodes[0].MemNode->getDebugLoc();
7956 DebugLoc StoreDL = StoreNodes[0].MemNode->getDebugLoc();
7958 LoadSDNode *FirstLoad = cast<LoadSDNode>(LoadNodes[0].MemNode);
7959 SDValue NewLoad = DAG.getLoad(JointMemOpVT, LoadDL,
7960 FirstLoad->getChain(),
7961 FirstLoad->getBasePtr(),
7962 FirstLoad->getPointerInfo(),
7963 false, false, false,
7964 FirstLoad->getAlignment());
7966 SDValue NewStore = DAG.getStore(EarliestOp->getChain(), StoreDL, NewLoad,
7967 FirstInChain->getBasePtr(),
7968 FirstInChain->getPointerInfo(), false, false,
7969 FirstInChain->getAlignment());
7971 // Replace one of the loads with the new load.
7972 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[0].MemNode);
7973 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1),
7974 SDValue(NewLoad.getNode(), 1));
7976 // Remove the rest of the load chains.
7977 for (unsigned i = 1; i < NumElem ; ++i) {
7978 // Replace all chain users of the old load nodes with the chain of the new
7980 LoadSDNode *Ld = cast<LoadSDNode>(LoadNodes[i].MemNode);
7981 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Ld->getChain());
7984 // Replace the first store with the new store.
7985 CombineTo(EarliestOp, NewStore);
7986 // Erase all other stores.
7987 for (unsigned i = 0; i < NumElem ; ++i) {
7988 // Remove all Store nodes.
7989 if (StoreNodes[i].MemNode == EarliestOp)
7991 StoreSDNode *St = cast<StoreSDNode>(StoreNodes[i].MemNode);
7992 DAG.ReplaceAllUsesOfValueWith(SDValue(St, 0), St->getChain());
7993 removeFromWorkList(St);
8000 SDValue DAGCombiner::visitSTORE(SDNode *N) {
8001 StoreSDNode *ST = cast<StoreSDNode>(N);
8002 SDValue Chain = ST->getChain();
8003 SDValue Value = ST->getValue();
8004 SDValue Ptr = ST->getBasePtr();
8006 // If this is a store of a bit convert, store the input value if the
8007 // resultant store does not need a higher alignment than the original.
8008 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
8009 ST->isUnindexed()) {
8010 unsigned OrigAlign = ST->getAlignment();
8011 EVT SVT = Value.getOperand(0).getValueType();
8012 unsigned Align = TLI.getDataLayout()->
8013 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
8014 if (Align <= OrigAlign &&
8015 ((!LegalOperations && !ST->isVolatile()) ||
8016 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
8017 return DAG.getStore(Chain, N->getDebugLoc(), Value.getOperand(0),
8018 Ptr, ST->getPointerInfo(), ST->isVolatile(),
8019 ST->isNonTemporal(), OrigAlign);
8022 // Turn 'store undef, Ptr' -> nothing.
8023 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
8026 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
8027 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
8028 // NOTE: If the original store is volatile, this transform must not increase
8029 // the number of stores. For example, on x86-32 an f64 can be stored in one
8030 // processor operation but an i64 (which is not legal) requires two. So the
8031 // transform should not be done in this case.
8032 if (Value.getOpcode() != ISD::TargetConstantFP) {
8034 switch (CFP->getValueType(0).getSimpleVT().SimpleTy) {
8035 default: llvm_unreachable("Unknown FP type");
8036 case MVT::f16: // We don't do this for these yet.
8042 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
8043 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
8044 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
8045 bitcastToAPInt().getZExtValue(), MVT::i32);
8046 return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
8047 Ptr, ST->getPointerInfo(), ST->isVolatile(),
8048 ST->isNonTemporal(), ST->getAlignment());
8052 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
8053 !ST->isVolatile()) ||
8054 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
8055 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
8056 getZExtValue(), MVT::i64);
8057 return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
8058 Ptr, ST->getPointerInfo(), ST->isVolatile(),
8059 ST->isNonTemporal(), ST->getAlignment());
8062 if (!ST->isVolatile() &&
8063 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
8064 // Many FP stores are not made apparent until after legalize, e.g. for
8065 // argument passing. Since this is so common, custom legalize the
8066 // 64-bit integer store into two 32-bit stores.
8067 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
8068 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
8069 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
8070 if (TLI.isBigEndian()) std::swap(Lo, Hi);
8072 unsigned Alignment = ST->getAlignment();
8073 bool isVolatile = ST->isVolatile();
8074 bool isNonTemporal = ST->isNonTemporal();
8076 SDValue St0 = DAG.getStore(Chain, ST->getDebugLoc(), Lo,
8077 Ptr, ST->getPointerInfo(),
8078 isVolatile, isNonTemporal,
8079 ST->getAlignment());
8080 Ptr = DAG.getNode(ISD::ADD, N->getDebugLoc(), Ptr.getValueType(), Ptr,
8081 DAG.getConstant(4, Ptr.getValueType()));
8082 Alignment = MinAlign(Alignment, 4U);
8083 SDValue St1 = DAG.getStore(Chain, ST->getDebugLoc(), Hi,
8084 Ptr, ST->getPointerInfo().getWithOffset(4),
8085 isVolatile, isNonTemporal,
8087 return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
8096 // Try to infer better alignment information than the store already has.
8097 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
8098 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
8099 if (Align > ST->getAlignment())
8100 return DAG.getTruncStore(Chain, N->getDebugLoc(), Value,
8101 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
8102 ST->isVolatile(), ST->isNonTemporal(), Align);
8106 // Try transforming a pair floating point load / store ops to integer
8107 // load / store ops.
8108 SDValue NewST = TransformFPLoadStorePair(N);
8109 if (NewST.getNode())
8113 // Walk up chain skipping non-aliasing memory nodes.
8114 SDValue BetterChain = FindBetterChain(N, Chain);
8116 // If there is a better chain.
8117 if (Chain != BetterChain) {
8120 // Replace the chain to avoid dependency.
8121 if (ST->isTruncatingStore()) {
8122 ReplStore = DAG.getTruncStore(BetterChain, N->getDebugLoc(), Value, Ptr,
8123 ST->getPointerInfo(),
8124 ST->getMemoryVT(), ST->isVolatile(),
8125 ST->isNonTemporal(), ST->getAlignment());
8127 ReplStore = DAG.getStore(BetterChain, N->getDebugLoc(), Value, Ptr,
8128 ST->getPointerInfo(),
8129 ST->isVolatile(), ST->isNonTemporal(),
8130 ST->getAlignment());
8133 // Create token to keep both nodes around.
8134 SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
8135 MVT::Other, Chain, ReplStore);
8137 // Make sure the new and old chains are cleaned up.
8138 AddToWorkList(Token.getNode());
8140 // Don't add users to work list.
8141 return CombineTo(N, Token, false);
8145 // Try transforming N to an indexed store.
8146 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
8147 return SDValue(N, 0);
8149 // FIXME: is there such a thing as a truncating indexed store?
8150 if (ST->isTruncatingStore() && ST->isUnindexed() &&
8151 Value.getValueType().isInteger()) {
8152 // See if we can simplify the input to this truncstore with knowledge that
8153 // only the low bits are being used. For example:
8154 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
8156 GetDemandedBits(Value,
8157 APInt::getLowBitsSet(
8158 Value.getValueType().getScalarType().getSizeInBits(),
8159 ST->getMemoryVT().getScalarType().getSizeInBits()));
8160 AddToWorkList(Value.getNode());
8161 if (Shorter.getNode())
8162 return DAG.getTruncStore(Chain, N->getDebugLoc(), Shorter,
8163 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
8164 ST->isVolatile(), ST->isNonTemporal(),
8165 ST->getAlignment());
8167 // Otherwise, see if we can simplify the operation with
8168 // SimplifyDemandedBits, which only works if the value has a single use.
8169 if (SimplifyDemandedBits(Value,
8170 APInt::getLowBitsSet(
8171 Value.getValueType().getScalarType().getSizeInBits(),
8172 ST->getMemoryVT().getScalarType().getSizeInBits())))
8173 return SDValue(N, 0);
8176 // If this is a load followed by a store to the same location, then the store
8178 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
8179 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
8180 ST->isUnindexed() && !ST->isVolatile() &&
8181 // There can't be any side effects between the load and store, such as
8183 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
8184 // The store is dead, remove it.
8189 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
8190 // truncating store. We can do this even if this is already a truncstore.
8191 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
8192 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
8193 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
8194 ST->getMemoryVT())) {
8195 return DAG.getTruncStore(Chain, N->getDebugLoc(), Value.getOperand(0),
8196 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
8197 ST->isVolatile(), ST->isNonTemporal(),
8198 ST->getAlignment());
8201 // Only perform this optimization before the types are legal, because we
8202 // don't want to perform this optimization on every DAGCombine invocation.
8204 bool EverChanged = false;
8207 // There can be multiple store sequences on the same chain.
8208 // Keep trying to merge store sequences until we are unable to do so
8209 // or until we merge the last store on the chain.
8210 bool Changed = MergeConsecutiveStores(ST);
8211 EverChanged |= Changed;
8212 if (!Changed) break;
8213 } while (ST->getOpcode() != ISD::DELETED_NODE);
8216 return SDValue(N, 0);
8219 return ReduceLoadOpStoreWidth(N);
8222 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
8223 SDValue InVec = N->getOperand(0);
8224 SDValue InVal = N->getOperand(1);
8225 SDValue EltNo = N->getOperand(2);
8226 DebugLoc dl = N->getDebugLoc();
8228 // If the inserted element is an UNDEF, just use the input vector.
8229 if (InVal.getOpcode() == ISD::UNDEF)
8232 EVT VT = InVec.getValueType();
8234 // If we can't generate a legal BUILD_VECTOR, exit
8235 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
8238 // Check that we know which element is being inserted
8239 if (!isa<ConstantSDNode>(EltNo))
8241 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
8243 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
8244 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
8246 SmallVector<SDValue, 8> Ops;
8247 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
8248 Ops.append(InVec.getNode()->op_begin(),
8249 InVec.getNode()->op_end());
8250 } else if (InVec.getOpcode() == ISD::UNDEF) {
8251 unsigned NElts = VT.getVectorNumElements();
8252 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
8257 // Insert the element
8258 if (Elt < Ops.size()) {
8259 // All the operands of BUILD_VECTOR must have the same type;
8260 // we enforce that here.
8261 EVT OpVT = Ops[0].getValueType();
8262 if (InVal.getValueType() != OpVT)
8263 InVal = OpVT.bitsGT(InVal.getValueType()) ?
8264 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
8265 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
8269 // Return the new vector
8270 return DAG.getNode(ISD::BUILD_VECTOR, dl,
8271 VT, &Ops[0], Ops.size());
8274 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
8275 // (vextract (scalar_to_vector val, 0) -> val
8276 SDValue InVec = N->getOperand(0);
8277 EVT VT = InVec.getValueType();
8278 EVT NVT = N->getValueType(0);
8280 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
8281 // Check if the result type doesn't match the inserted element type. A
8282 // SCALAR_TO_VECTOR may truncate the inserted element and the
8283 // EXTRACT_VECTOR_ELT may widen the extracted vector.
8284 SDValue InOp = InVec.getOperand(0);
8285 if (InOp.getValueType() != NVT) {
8286 assert(InOp.getValueType().isInteger() && NVT.isInteger());
8287 return DAG.getSExtOrTrunc(InOp, InVec.getDebugLoc(), NVT);
8292 SDValue EltNo = N->getOperand(1);
8293 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
8295 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
8296 // We only perform this optimization before the op legalization phase because
8297 // we may introduce new vector instructions which are not backed by TD
8298 // patterns. For example on AVX, extracting elements from a wide vector
8299 // without using extract_subvector.
8300 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
8301 && ConstEltNo && !LegalOperations) {
8302 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
8303 int NumElem = VT.getVectorNumElements();
8304 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
8305 // Find the new index to extract from.
8306 int OrigElt = SVOp->getMaskElt(Elt);
8308 // Extracting an undef index is undef.
8310 return DAG.getUNDEF(NVT);
8312 // Select the right vector half to extract from.
8313 if (OrigElt < NumElem) {
8314 InVec = InVec->getOperand(0);
8316 InVec = InVec->getOperand(1);
8320 EVT IndexTy = N->getOperand(1).getValueType();
8321 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, N->getDebugLoc(), NVT,
8322 InVec, DAG.getConstant(OrigElt, IndexTy));
8325 // Perform only after legalization to ensure build_vector / vector_shuffle
8326 // optimizations have already been done.
8327 if (!LegalOperations) return SDValue();
8329 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
8330 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
8331 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
8334 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
8335 bool NewLoad = false;
8336 bool BCNumEltsChanged = false;
8337 EVT ExtVT = VT.getVectorElementType();
8340 // If the result of load has to be truncated, then it's not necessarily
8342 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
8345 if (InVec.getOpcode() == ISD::BITCAST) {
8346 // Don't duplicate a load with other uses.
8347 if (!InVec.hasOneUse())
8350 EVT BCVT = InVec.getOperand(0).getValueType();
8351 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
8353 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
8354 BCNumEltsChanged = true;
8355 InVec = InVec.getOperand(0);
8356 ExtVT = BCVT.getVectorElementType();
8360 LoadSDNode *LN0 = NULL;
8361 const ShuffleVectorSDNode *SVN = NULL;
8362 if (ISD::isNormalLoad(InVec.getNode())) {
8363 LN0 = cast<LoadSDNode>(InVec);
8364 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
8365 InVec.getOperand(0).getValueType() == ExtVT &&
8366 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
8367 // Don't duplicate a load with other uses.
8368 if (!InVec.hasOneUse())
8371 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
8372 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
8373 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
8375 // (load $addr+1*size)
8377 // Don't duplicate a load with other uses.
8378 if (!InVec.hasOneUse())
8381 // If the bit convert changed the number of elements, it is unsafe
8382 // to examine the mask.
8383 if (BCNumEltsChanged)
8386 // Select the input vector, guarding against out of range extract vector.
8387 unsigned NumElems = VT.getVectorNumElements();
8388 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
8389 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
8391 if (InVec.getOpcode() == ISD::BITCAST) {
8392 // Don't duplicate a load with other uses.
8393 if (!InVec.hasOneUse())
8396 InVec = InVec.getOperand(0);
8398 if (ISD::isNormalLoad(InVec.getNode())) {
8399 LN0 = cast<LoadSDNode>(InVec);
8400 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
8404 // Make sure we found a non-volatile load and the extractelement is
8406 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
8409 // If Idx was -1 above, Elt is going to be -1, so just return undef.
8411 return DAG.getUNDEF(LVT);
8413 unsigned Align = LN0->getAlignment();
8415 // Check the resultant load doesn't need a higher alignment than the
8419 ->getABITypeAlignment(LVT.getTypeForEVT(*DAG.getContext()));
8421 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, LVT))
8427 SDValue NewPtr = LN0->getBasePtr();
8428 unsigned PtrOff = 0;
8431 PtrOff = LVT.getSizeInBits() * Elt / 8;
8432 EVT PtrType = NewPtr.getValueType();
8433 if (TLI.isBigEndian())
8434 PtrOff = VT.getSizeInBits() / 8 - PtrOff;
8435 NewPtr = DAG.getNode(ISD::ADD, N->getDebugLoc(), PtrType, NewPtr,
8436 DAG.getConstant(PtrOff, PtrType));
8439 // The replacement we need to do here is a little tricky: we need to
8440 // replace an extractelement of a load with a load.
8441 // Use ReplaceAllUsesOfValuesWith to do the replacement.
8442 // Note that this replacement assumes that the extractvalue is the only
8443 // use of the load; that's okay because we don't want to perform this
8444 // transformation in other cases anyway.
8447 if (NVT.bitsGT(LVT)) {
8448 // If the result type of vextract is wider than the load, then issue an
8449 // extending load instead.
8450 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, LVT)
8451 ? ISD::ZEXTLOAD : ISD::EXTLOAD;
8452 Load = DAG.getExtLoad(ExtType, N->getDebugLoc(), NVT, LN0->getChain(),
8453 NewPtr, LN0->getPointerInfo().getWithOffset(PtrOff),
8454 LVT, LN0->isVolatile(), LN0->isNonTemporal(),Align);
8455 Chain = Load.getValue(1);
8457 Load = DAG.getLoad(LVT, N->getDebugLoc(), LN0->getChain(), NewPtr,
8458 LN0->getPointerInfo().getWithOffset(PtrOff),
8459 LN0->isVolatile(), LN0->isNonTemporal(),
8460 LN0->isInvariant(), Align);
8461 Chain = Load.getValue(1);
8462 if (NVT.bitsLT(LVT))
8463 Load = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), NVT, Load);
8465 Load = DAG.getNode(ISD::BITCAST, N->getDebugLoc(), NVT, Load);
8467 WorkListRemover DeadNodes(*this);
8468 SDValue From[] = { SDValue(N, 0), SDValue(LN0,1) };
8469 SDValue To[] = { Load, Chain };
8470 DAG.ReplaceAllUsesOfValuesWith(From, To, 2);
8471 // Since we're explcitly calling ReplaceAllUses, add the new node to the
8472 // worklist explicitly as well.
8473 AddToWorkList(Load.getNode());
8474 AddUsersToWorkList(Load.getNode()); // Add users too
8475 // Make sure to revisit this node to clean it up; it will usually be dead.
8477 return SDValue(N, 0);
8483 // Simplify (build_vec (ext )) to (bitcast (build_vec ))
8484 SDValue DAGCombiner::reduceBuildVecExtToExtBuildVec(SDNode *N) {
8485 // We perform this optimization post type-legalization because
8486 // the type-legalizer often scalarizes integer-promoted vectors.
8487 // Performing this optimization before may create bit-casts which
8488 // will be type-legalized to complex code sequences.
8489 // We perform this optimization only before the operation legalizer because we
8490 // may introduce illegal operations.
8491 if (Level != AfterLegalizeVectorOps && Level != AfterLegalizeTypes)
8494 unsigned NumInScalars = N->getNumOperands();
8495 DebugLoc dl = N->getDebugLoc();
8496 EVT VT = N->getValueType(0);
8498 // Check to see if this is a BUILD_VECTOR of a bunch of values
8499 // which come from any_extend or zero_extend nodes. If so, we can create
8500 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
8501 // optimizations. We do not handle sign-extend because we can't fill the sign
8503 EVT SourceType = MVT::Other;
8504 bool AllAnyExt = true;
8506 for (unsigned i = 0; i != NumInScalars; ++i) {
8507 SDValue In = N->getOperand(i);
8508 // Ignore undef inputs.
8509 if (In.getOpcode() == ISD::UNDEF) continue;
8511 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
8512 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
8514 // Abort if the element is not an extension.
8515 if (!ZeroExt && !AnyExt) {
8516 SourceType = MVT::Other;
8520 // The input is a ZeroExt or AnyExt. Check the original type.
8521 EVT InTy = In.getOperand(0).getValueType();
8523 // Check that all of the widened source types are the same.
8524 if (SourceType == MVT::Other)
8527 else if (InTy != SourceType) {
8528 // Multiple income types. Abort.
8529 SourceType = MVT::Other;
8533 // Check if all of the extends are ANY_EXTENDs.
8534 AllAnyExt &= AnyExt;
8537 // In order to have valid types, all of the inputs must be extended from the
8538 // same source type and all of the inputs must be any or zero extend.
8539 // Scalar sizes must be a power of two.
8540 EVT OutScalarTy = VT.getScalarType();
8541 bool ValidTypes = SourceType != MVT::Other &&
8542 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
8543 isPowerOf2_32(SourceType.getSizeInBits());
8545 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
8546 // turn into a single shuffle instruction.
8550 bool isLE = TLI.isLittleEndian();
8551 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
8552 assert(ElemRatio > 1 && "Invalid element size ratio");
8553 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
8554 DAG.getConstant(0, SourceType);
8556 unsigned NewBVElems = ElemRatio * VT.getVectorNumElements();
8557 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
8559 // Populate the new build_vector
8560 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
8561 SDValue Cast = N->getOperand(i);
8562 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
8563 Cast.getOpcode() == ISD::ZERO_EXTEND ||
8564 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
8566 if (Cast.getOpcode() == ISD::UNDEF)
8567 In = DAG.getUNDEF(SourceType);
8569 In = Cast->getOperand(0);
8570 unsigned Index = isLE ? (i * ElemRatio) :
8571 (i * ElemRatio + (ElemRatio - 1));
8573 assert(Index < Ops.size() && "Invalid index");
8577 // The type of the new BUILD_VECTOR node.
8578 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
8579 assert(VecVT.getSizeInBits() == VT.getSizeInBits() &&
8580 "Invalid vector size");
8581 // Check if the new vector type is legal.
8582 if (!isTypeLegal(VecVT)) return SDValue();
8584 // Make the new BUILD_VECTOR.
8585 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], Ops.size());
8587 // The new BUILD_VECTOR node has the potential to be further optimized.
8588 AddToWorkList(BV.getNode());
8589 // Bitcast to the desired type.
8590 return DAG.getNode(ISD::BITCAST, dl, VT, BV);
8593 SDValue DAGCombiner::reduceBuildVecConvertToConvertBuildVec(SDNode *N) {
8594 EVT VT = N->getValueType(0);
8596 unsigned NumInScalars = N->getNumOperands();
8597 DebugLoc dl = N->getDebugLoc();
8599 EVT SrcVT = MVT::Other;
8600 unsigned Opcode = ISD::DELETED_NODE;
8601 unsigned NumDefs = 0;
8603 for (unsigned i = 0; i != NumInScalars; ++i) {
8604 SDValue In = N->getOperand(i);
8605 unsigned Opc = In.getOpcode();
8607 if (Opc == ISD::UNDEF)
8610 // If all scalar values are floats and converted from integers.
8611 if (Opcode == ISD::DELETED_NODE &&
8612 (Opc == ISD::UINT_TO_FP || Opc == ISD::SINT_TO_FP)) {
8614 // If not supported by target, bail out.
8615 if (TLI.getOperationAction(Opcode, VT) != TargetLowering::Legal &&
8616 TLI.getOperationAction(Opcode, VT) != TargetLowering::Custom)
8622 EVT InVT = In.getOperand(0).getValueType();
8624 // If all scalar values are typed differently, bail out. It's chosen to
8625 // simplify BUILD_VECTOR of integer types.
8626 if (SrcVT == MVT::Other)
8633 // If the vector has just one element defined, it's not worth to fold it into
8634 // a vectorized one.
8638 assert((Opcode == ISD::UINT_TO_FP || Opcode == ISD::SINT_TO_FP)
8639 && "Should only handle conversion from integer to float.");
8640 assert(SrcVT != MVT::Other && "Cannot determine source type!");
8642 EVT NVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumInScalars);
8643 SmallVector<SDValue, 8> Opnds;
8644 for (unsigned i = 0; i != NumInScalars; ++i) {
8645 SDValue In = N->getOperand(i);
8647 if (In.getOpcode() == ISD::UNDEF)
8648 Opnds.push_back(DAG.getUNDEF(SrcVT));
8650 Opnds.push_back(In.getOperand(0));
8652 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NVT,
8653 &Opnds[0], Opnds.size());
8654 AddToWorkList(BV.getNode());
8656 return DAG.getNode(Opcode, dl, VT, BV);
8659 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
8660 unsigned NumInScalars = N->getNumOperands();
8661 DebugLoc dl = N->getDebugLoc();
8662 EVT VT = N->getValueType(0);
8664 // A vector built entirely of undefs is undef.
8665 if (ISD::allOperandsUndef(N))
8666 return DAG.getUNDEF(VT);
8668 SDValue V = reduceBuildVecExtToExtBuildVec(N);
8672 V = reduceBuildVecConvertToConvertBuildVec(N);
8676 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
8677 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
8678 // at most two distinct vectors, turn this into a shuffle node.
8680 // May only combine to shuffle after legalize if shuffle is legal.
8681 if (LegalOperations &&
8682 !TLI.isOperationLegalOrCustom(ISD::VECTOR_SHUFFLE, VT))
8685 SDValue VecIn1, VecIn2;
8686 for (unsigned i = 0; i != NumInScalars; ++i) {
8687 // Ignore undef inputs.
8688 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
8690 // If this input is something other than a EXTRACT_VECTOR_ELT with a
8691 // constant index, bail out.
8692 if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
8693 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
8694 VecIn1 = VecIn2 = SDValue(0, 0);
8698 // We allow up to two distinct input vectors.
8699 SDValue ExtractedFromVec = N->getOperand(i).getOperand(0);
8700 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
8703 if (VecIn1.getNode() == 0) {
8704 VecIn1 = ExtractedFromVec;
8705 } else if (VecIn2.getNode() == 0) {
8706 VecIn2 = ExtractedFromVec;
8709 VecIn1 = VecIn2 = SDValue(0, 0);
8714 // If everything is good, we can make a shuffle operation.
8715 if (VecIn1.getNode()) {
8716 SmallVector<int, 8> Mask;
8717 for (unsigned i = 0; i != NumInScalars; ++i) {
8718 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
8723 // If extracting from the first vector, just use the index directly.
8724 SDValue Extract = N->getOperand(i);
8725 SDValue ExtVal = Extract.getOperand(1);
8726 if (Extract.getOperand(0) == VecIn1) {
8727 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
8728 if (ExtIndex > VT.getVectorNumElements())
8731 Mask.push_back(ExtIndex);
8735 // Otherwise, use InIdx + VecSize
8736 unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue();
8737 Mask.push_back(Idx+NumInScalars);
8740 // We can't generate a shuffle node with mismatched input and output types.
8741 // Attempt to transform a single input vector to the correct type.
8742 if ((VT != VecIn1.getValueType())) {
8743 // We don't support shuffeling between TWO values of different types.
8744 if (VecIn2.getNode() != 0)
8747 // We only support widening of vectors which are half the size of the
8748 // output registers. For example XMM->YMM widening on X86 with AVX.
8749 if (VecIn1.getValueType().getSizeInBits()*2 != VT.getSizeInBits())
8752 // If the input vector type has a different base type to the output
8753 // vector type, bail out.
8754 if (VecIn1.getValueType().getVectorElementType() !=
8755 VT.getVectorElementType())
8758 // Widen the input vector by adding undef values.
8759 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
8760 VecIn1, DAG.getUNDEF(VecIn1.getValueType()));
8763 // If VecIn2 is unused then change it to undef.
8764 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
8766 // Check that we were able to transform all incoming values to the same
8768 if (VecIn2.getValueType() != VecIn1.getValueType() ||
8769 VecIn1.getValueType() != VT)
8772 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
8773 if (!isTypeLegal(VT))
8776 // Return the new VECTOR_SHUFFLE node.
8780 return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], &Mask[0]);
8786 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
8787 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
8788 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
8789 // inputs come from at most two distinct vectors, turn this into a shuffle
8792 // If we only have one input vector, we don't need to do any concatenation.
8793 if (N->getNumOperands() == 1)
8794 return N->getOperand(0);
8796 // Check if all of the operands are undefs.
8797 if (ISD::allOperandsUndef(N))
8798 return DAG.getUNDEF(N->getValueType(0));
8803 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
8804 EVT NVT = N->getValueType(0);
8805 SDValue V = N->getOperand(0);
8807 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
8808 // Handle only simple case where vector being inserted and vector
8809 // being extracted are of same type, and are half size of larger vectors.
8810 EVT BigVT = V->getOperand(0).getValueType();
8811 EVT SmallVT = V->getOperand(1).getValueType();
8812 if (NVT != SmallVT || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
8815 // Only handle cases where both indexes are constants with the same type.
8816 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
8817 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
8819 if (InsIdx && ExtIdx &&
8820 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
8821 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
8823 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
8825 // indices are equal => V1
8826 // otherwise => (extract_subvec V1, ExtIdx)
8827 if (InsIdx->getZExtValue() == ExtIdx->getZExtValue())
8828 return V->getOperand(1);
8829 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, N->getDebugLoc(), NVT,
8830 V->getOperand(0), N->getOperand(1));
8834 if (V->getOpcode() == ISD::CONCAT_VECTORS) {
8836 // (extract_subvec (concat V1, V2, ...), i)
8839 // Only operand 0 is checked as 'concat' assumes all inputs of the same type.
8840 if (V->getOperand(0).getValueType() != NVT)
8842 unsigned Idx = dyn_cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
8843 unsigned NumElems = NVT.getVectorNumElements();
8844 assert((Idx % NumElems) == 0 &&
8845 "IDX in concat is not a multiple of the result vector length.");
8846 return V->getOperand(Idx / NumElems);
8852 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
8853 EVT VT = N->getValueType(0);
8854 unsigned NumElts = VT.getVectorNumElements();
8856 SDValue N0 = N->getOperand(0);
8857 SDValue N1 = N->getOperand(1);
8859 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
8861 // Canonicalize shuffle undef, undef -> undef
8862 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
8863 return DAG.getUNDEF(VT);
8865 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8867 // Canonicalize shuffle v, v -> v, undef
8869 SmallVector<int, 8> NewMask;
8870 for (unsigned i = 0; i != NumElts; ++i) {
8871 int Idx = SVN->getMaskElt(i);
8872 if (Idx >= (int)NumElts) Idx -= NumElts;
8873 NewMask.push_back(Idx);
8875 return DAG.getVectorShuffle(VT, N->getDebugLoc(), N0, DAG.getUNDEF(VT),
8879 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
8880 if (N0.getOpcode() == ISD::UNDEF) {
8881 SmallVector<int, 8> NewMask;
8882 for (unsigned i = 0; i != NumElts; ++i) {
8883 int Idx = SVN->getMaskElt(i);
8885 if (Idx < (int)NumElts)
8890 NewMask.push_back(Idx);
8892 return DAG.getVectorShuffle(VT, N->getDebugLoc(), N1, DAG.getUNDEF(VT),
8896 // Remove references to rhs if it is undef
8897 if (N1.getOpcode() == ISD::UNDEF) {
8898 bool Changed = false;
8899 SmallVector<int, 8> NewMask;
8900 for (unsigned i = 0; i != NumElts; ++i) {
8901 int Idx = SVN->getMaskElt(i);
8902 if (Idx >= (int)NumElts) {
8906 NewMask.push_back(Idx);
8909 return DAG.getVectorShuffle(VT, N->getDebugLoc(), N0, N1, &NewMask[0]);
8912 // If it is a splat, check if the argument vector is another splat or a
8913 // build_vector with all scalar elements the same.
8914 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
8915 SDNode *V = N0.getNode();
8917 // If this is a bit convert that changes the element type of the vector but
8918 // not the number of vector elements, look through it. Be careful not to
8919 // look though conversions that change things like v4f32 to v2f64.
8920 if (V->getOpcode() == ISD::BITCAST) {
8921 SDValue ConvInput = V->getOperand(0);
8922 if (ConvInput.getValueType().isVector() &&
8923 ConvInput.getValueType().getVectorNumElements() == NumElts)
8924 V = ConvInput.getNode();
8927 if (V->getOpcode() == ISD::BUILD_VECTOR) {
8928 assert(V->getNumOperands() == NumElts &&
8929 "BUILD_VECTOR has wrong number of operands");
8931 bool AllSame = true;
8932 for (unsigned i = 0; i != NumElts; ++i) {
8933 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
8934 Base = V->getOperand(i);
8938 // Splat of <u, u, u, u>, return <u, u, u, u>
8939 if (!Base.getNode())
8941 for (unsigned i = 0; i != NumElts; ++i) {
8942 if (V->getOperand(i) != Base) {
8947 // Splat of <x, x, x, x>, return <x, x, x, x>
8953 // If this shuffle node is simply a swizzle of another shuffle node,
8954 // and it reverses the swizzle of the previous shuffle then we can
8955 // optimize shuffle(shuffle(x, undef), undef) -> x.
8956 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
8957 N1.getOpcode() == ISD::UNDEF) {
8959 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
8961 // Shuffle nodes can only reverse shuffles with a single non-undef value.
8962 if (N0.getOperand(1).getOpcode() != ISD::UNDEF)
8965 // The incoming shuffle must be of the same type as the result of the
8967 assert(OtherSV->getOperand(0).getValueType() == VT &&
8968 "Shuffle types don't match");
8970 for (unsigned i = 0; i != NumElts; ++i) {
8971 int Idx = SVN->getMaskElt(i);
8972 assert(Idx < (int)NumElts && "Index references undef operand");
8973 // Next, this index comes from the first value, which is the incoming
8974 // shuffle. Adopt the incoming index.
8976 Idx = OtherSV->getMaskElt(Idx);
8978 // The combined shuffle must map each index to itself.
8979 if (Idx >= 0 && (unsigned)Idx != i)
8983 return OtherSV->getOperand(0);
8989 SDValue DAGCombiner::visitMEMBARRIER(SDNode* N) {
8990 if (!TLI.getShouldFoldAtomicFences())
8993 SDValue atomic = N->getOperand(0);
8994 switch (atomic.getOpcode()) {
8995 case ISD::ATOMIC_CMP_SWAP:
8996 case ISD::ATOMIC_SWAP:
8997 case ISD::ATOMIC_LOAD_ADD:
8998 case ISD::ATOMIC_LOAD_SUB:
8999 case ISD::ATOMIC_LOAD_AND:
9000 case ISD::ATOMIC_LOAD_OR:
9001 case ISD::ATOMIC_LOAD_XOR:
9002 case ISD::ATOMIC_LOAD_NAND:
9003 case ISD::ATOMIC_LOAD_MIN:
9004 case ISD::ATOMIC_LOAD_MAX:
9005 case ISD::ATOMIC_LOAD_UMIN:
9006 case ISD::ATOMIC_LOAD_UMAX:
9012 SDValue fence = atomic.getOperand(0);
9013 if (fence.getOpcode() != ISD::MEMBARRIER)
9016 switch (atomic.getOpcode()) {
9017 case ISD::ATOMIC_CMP_SWAP:
9018 return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
9019 fence.getOperand(0),
9020 atomic.getOperand(1), atomic.getOperand(2),
9021 atomic.getOperand(3)), atomic.getResNo());
9022 case ISD::ATOMIC_SWAP:
9023 case ISD::ATOMIC_LOAD_ADD:
9024 case ISD::ATOMIC_LOAD_SUB:
9025 case ISD::ATOMIC_LOAD_AND:
9026 case ISD::ATOMIC_LOAD_OR:
9027 case ISD::ATOMIC_LOAD_XOR:
9028 case ISD::ATOMIC_LOAD_NAND:
9029 case ISD::ATOMIC_LOAD_MIN:
9030 case ISD::ATOMIC_LOAD_MAX:
9031 case ISD::ATOMIC_LOAD_UMIN:
9032 case ISD::ATOMIC_LOAD_UMAX:
9033 return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
9034 fence.getOperand(0),
9035 atomic.getOperand(1), atomic.getOperand(2)),
9042 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
9043 /// an AND to a vector_shuffle with the destination vector and a zero vector.
9044 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
9045 /// vector_shuffle V, Zero, <0, 4, 2, 4>
9046 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
9047 EVT VT = N->getValueType(0);
9048 DebugLoc dl = N->getDebugLoc();
9049 SDValue LHS = N->getOperand(0);
9050 SDValue RHS = N->getOperand(1);
9051 if (N->getOpcode() == ISD::AND) {
9052 if (RHS.getOpcode() == ISD::BITCAST)
9053 RHS = RHS.getOperand(0);
9054 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
9055 SmallVector<int, 8> Indices;
9056 unsigned NumElts = RHS.getNumOperands();
9057 for (unsigned i = 0; i != NumElts; ++i) {
9058 SDValue Elt = RHS.getOperand(i);
9059 if (!isa<ConstantSDNode>(Elt))
9062 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
9063 Indices.push_back(i);
9064 else if (cast<ConstantSDNode>(Elt)->isNullValue())
9065 Indices.push_back(NumElts);
9070 // Let's see if the target supports this vector_shuffle.
9071 EVT RVT = RHS.getValueType();
9072 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
9075 // Return the new VECTOR_SHUFFLE node.
9076 EVT EltVT = RVT.getVectorElementType();
9077 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
9078 DAG.getConstant(0, EltVT));
9079 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
9080 RVT, &ZeroOps[0], ZeroOps.size());
9081 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
9082 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
9083 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
9090 /// SimplifyVBinOp - Visit a binary vector operation, like ADD.
9091 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
9092 // After legalize, the target may be depending on adds and other
9093 // binary ops to provide legal ways to construct constants or other
9094 // things. Simplifying them may result in a loss of legality.
9095 if (LegalOperations) return SDValue();
9097 assert(N->getValueType(0).isVector() &&
9098 "SimplifyVBinOp only works on vectors!");
9100 SDValue LHS = N->getOperand(0);
9101 SDValue RHS = N->getOperand(1);
9102 SDValue Shuffle = XformToShuffleWithZero(N);
9103 if (Shuffle.getNode()) return Shuffle;
9105 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
9107 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
9108 RHS.getOpcode() == ISD::BUILD_VECTOR) {
9109 SmallVector<SDValue, 8> Ops;
9110 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
9111 SDValue LHSOp = LHS.getOperand(i);
9112 SDValue RHSOp = RHS.getOperand(i);
9113 // If these two elements can't be folded, bail out.
9114 if ((LHSOp.getOpcode() != ISD::UNDEF &&
9115 LHSOp.getOpcode() != ISD::Constant &&
9116 LHSOp.getOpcode() != ISD::ConstantFP) ||
9117 (RHSOp.getOpcode() != ISD::UNDEF &&
9118 RHSOp.getOpcode() != ISD::Constant &&
9119 RHSOp.getOpcode() != ISD::ConstantFP))
9122 // Can't fold divide by zero.
9123 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
9124 N->getOpcode() == ISD::FDIV) {
9125 if ((RHSOp.getOpcode() == ISD::Constant &&
9126 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
9127 (RHSOp.getOpcode() == ISD::ConstantFP &&
9128 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
9132 EVT VT = LHSOp.getValueType();
9133 EVT RVT = RHSOp.getValueType();
9135 // Integer BUILD_VECTOR operands may have types larger than the element
9136 // size (e.g., when the element type is not legal). Prior to type
9137 // legalization, the types may not match between the two BUILD_VECTORS.
9138 // Truncate one of the operands to make them match.
9139 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
9140 RHSOp = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, RHSOp);
9142 LHSOp = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), RVT, LHSOp);
9146 SDValue FoldOp = DAG.getNode(N->getOpcode(), LHS.getDebugLoc(), VT,
9148 if (FoldOp.getOpcode() != ISD::UNDEF &&
9149 FoldOp.getOpcode() != ISD::Constant &&
9150 FoldOp.getOpcode() != ISD::ConstantFP)
9152 Ops.push_back(FoldOp);
9153 AddToWorkList(FoldOp.getNode());
9156 if (Ops.size() == LHS.getNumOperands())
9157 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
9158 LHS.getValueType(), &Ops[0], Ops.size());
9164 /// SimplifyVUnaryOp - Visit a binary vector operation, like FABS/FNEG.
9165 SDValue DAGCombiner::SimplifyVUnaryOp(SDNode *N) {
9166 // After legalize, the target may be depending on adds and other
9167 // binary ops to provide legal ways to construct constants or other
9168 // things. Simplifying them may result in a loss of legality.
9169 if (LegalOperations) return SDValue();
9171 assert(N->getValueType(0).isVector() &&
9172 "SimplifyVUnaryOp only works on vectors!");
9174 SDValue N0 = N->getOperand(0);
9176 if (N0.getOpcode() != ISD::BUILD_VECTOR)
9179 // Operand is a BUILD_VECTOR node, see if we can constant fold it.
9180 SmallVector<SDValue, 8> Ops;
9181 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i) {
9182 SDValue Op = N0.getOperand(i);
9183 if (Op.getOpcode() != ISD::UNDEF &&
9184 Op.getOpcode() != ISD::ConstantFP)
9186 EVT EltVT = Op.getValueType();
9187 SDValue FoldOp = DAG.getNode(N->getOpcode(), N0.getDebugLoc(), EltVT, Op);
9188 if (FoldOp.getOpcode() != ISD::UNDEF &&
9189 FoldOp.getOpcode() != ISD::ConstantFP)
9191 Ops.push_back(FoldOp);
9192 AddToWorkList(FoldOp.getNode());
9195 if (Ops.size() != N0.getNumOperands())
9198 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
9199 N0.getValueType(), &Ops[0], Ops.size());
9202 SDValue DAGCombiner::SimplifySelect(DebugLoc DL, SDValue N0,
9203 SDValue N1, SDValue N2){
9204 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
9206 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
9207 cast<CondCodeSDNode>(N0.getOperand(2))->get());
9209 // If we got a simplified select_cc node back from SimplifySelectCC, then
9210 // break it down into a new SETCC node, and a new SELECT node, and then return
9211 // the SELECT node, since we were called with a SELECT node.
9212 if (SCC.getNode()) {
9213 // Check to see if we got a select_cc back (to turn into setcc/select).
9214 // Otherwise, just return whatever node we got back, like fabs.
9215 if (SCC.getOpcode() == ISD::SELECT_CC) {
9216 SDValue SETCC = DAG.getNode(ISD::SETCC, N0.getDebugLoc(),
9218 SCC.getOperand(0), SCC.getOperand(1),
9220 AddToWorkList(SETCC.getNode());
9221 return DAG.getNode(ISD::SELECT, SCC.getDebugLoc(), SCC.getValueType(),
9222 SCC.getOperand(2), SCC.getOperand(3), SETCC);
9230 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
9231 /// are the two values being selected between, see if we can simplify the
9232 /// select. Callers of this should assume that TheSelect is deleted if this
9233 /// returns true. As such, they should return the appropriate thing (e.g. the
9234 /// node) back to the top-level of the DAG combiner loop to avoid it being
9236 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
9239 // Cannot simplify select with vector condition
9240 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
9242 // If this is a select from two identical things, try to pull the operation
9243 // through the select.
9244 if (LHS.getOpcode() != RHS.getOpcode() ||
9245 !LHS.hasOneUse() || !RHS.hasOneUse())
9248 // If this is a load and the token chain is identical, replace the select
9249 // of two loads with a load through a select of the address to load from.
9250 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
9251 // constants have been dropped into the constant pool.
9252 if (LHS.getOpcode() == ISD::LOAD) {
9253 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
9254 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
9256 // Token chains must be identical.
9257 if (LHS.getOperand(0) != RHS.getOperand(0) ||
9258 // Do not let this transformation reduce the number of volatile loads.
9259 LLD->isVolatile() || RLD->isVolatile() ||
9260 // If this is an EXTLOAD, the VT's must match.
9261 LLD->getMemoryVT() != RLD->getMemoryVT() ||
9262 // If this is an EXTLOAD, the kind of extension must match.
9263 (LLD->getExtensionType() != RLD->getExtensionType() &&
9264 // The only exception is if one of the extensions is anyext.
9265 LLD->getExtensionType() != ISD::EXTLOAD &&
9266 RLD->getExtensionType() != ISD::EXTLOAD) ||
9267 // FIXME: this discards src value information. This is
9268 // over-conservative. It would be beneficial to be able to remember
9269 // both potential memory locations. Since we are discarding
9270 // src value info, don't do the transformation if the memory
9271 // locations are not in the default address space.
9272 LLD->getPointerInfo().getAddrSpace() != 0 ||
9273 RLD->getPointerInfo().getAddrSpace() != 0)
9276 // Check that the select condition doesn't reach either load. If so,
9277 // folding this will induce a cycle into the DAG. If not, this is safe to
9278 // xform, so create a select of the addresses.
9280 if (TheSelect->getOpcode() == ISD::SELECT) {
9281 SDNode *CondNode = TheSelect->getOperand(0).getNode();
9282 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
9283 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
9285 // The loads must not depend on one another.
9286 if (LLD->isPredecessorOf(RLD) ||
9287 RLD->isPredecessorOf(LLD))
9289 Addr = DAG.getNode(ISD::SELECT, TheSelect->getDebugLoc(),
9290 LLD->getBasePtr().getValueType(),
9291 TheSelect->getOperand(0), LLD->getBasePtr(),
9293 } else { // Otherwise SELECT_CC
9294 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
9295 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
9297 if ((LLD->hasAnyUseOfValue(1) &&
9298 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
9299 (RLD->hasAnyUseOfValue(1) &&
9300 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
9303 Addr = DAG.getNode(ISD::SELECT_CC, TheSelect->getDebugLoc(),
9304 LLD->getBasePtr().getValueType(),
9305 TheSelect->getOperand(0),
9306 TheSelect->getOperand(1),
9307 LLD->getBasePtr(), RLD->getBasePtr(),
9308 TheSelect->getOperand(4));
9312 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
9313 Load = DAG.getLoad(TheSelect->getValueType(0),
9314 TheSelect->getDebugLoc(),
9315 // FIXME: Discards pointer info.
9316 LLD->getChain(), Addr, MachinePointerInfo(),
9317 LLD->isVolatile(), LLD->isNonTemporal(),
9318 LLD->isInvariant(), LLD->getAlignment());
9320 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
9321 RLD->getExtensionType() : LLD->getExtensionType(),
9322 TheSelect->getDebugLoc(),
9323 TheSelect->getValueType(0),
9324 // FIXME: Discards pointer info.
9325 LLD->getChain(), Addr, MachinePointerInfo(),
9326 LLD->getMemoryVT(), LLD->isVolatile(),
9327 LLD->isNonTemporal(), LLD->getAlignment());
9330 // Users of the select now use the result of the load.
9331 CombineTo(TheSelect, Load);
9333 // Users of the old loads now use the new load's chain. We know the
9334 // old-load value is dead now.
9335 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
9336 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
9343 /// SimplifySelectCC - Simplify an expression of the form (N0 cond N1) ? N2 : N3
9344 /// where 'cond' is the comparison specified by CC.
9345 SDValue DAGCombiner::SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1,
9346 SDValue N2, SDValue N3,
9347 ISD::CondCode CC, bool NotExtCompare) {
9348 // (x ? y : y) -> y.
9349 if (N2 == N3) return N2;
9351 EVT VT = N2.getValueType();
9352 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
9353 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
9354 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
9356 // Determine if the condition we're dealing with is constant
9357 SDValue SCC = SimplifySetCC(TLI.getSetCCResultType(N0.getValueType()),
9358 N0, N1, CC, DL, false);
9359 if (SCC.getNode()) AddToWorkList(SCC.getNode());
9360 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
9362 // fold select_cc true, x, y -> x
9363 if (SCCC && !SCCC->isNullValue())
9365 // fold select_cc false, x, y -> y
9366 if (SCCC && SCCC->isNullValue())
9369 // Check to see if we can simplify the select into an fabs node
9370 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
9371 // Allow either -0.0 or 0.0
9372 if (CFP->getValueAPF().isZero()) {
9373 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
9374 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
9375 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
9376 N2 == N3.getOperand(0))
9377 return DAG.getNode(ISD::FABS, DL, VT, N0);
9379 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
9380 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
9381 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
9382 N2.getOperand(0) == N3)
9383 return DAG.getNode(ISD::FABS, DL, VT, N3);
9387 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
9388 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
9389 // in it. This is a win when the constant is not otherwise available because
9390 // it replaces two constant pool loads with one. We only do this if the FP
9391 // type is known to be legal, because if it isn't, then we are before legalize
9392 // types an we want the other legalization to happen first (e.g. to avoid
9393 // messing with soft float) and if the ConstantFP is not legal, because if
9394 // it is legal, we may not need to store the FP constant in a constant pool.
9395 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
9396 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
9397 if (TLI.isTypeLegal(N2.getValueType()) &&
9398 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
9399 TargetLowering::Legal) &&
9400 // If both constants have multiple uses, then we won't need to do an
9401 // extra load, they are likely around in registers for other users.
9402 (TV->hasOneUse() || FV->hasOneUse())) {
9403 Constant *Elts[] = {
9404 const_cast<ConstantFP*>(FV->getConstantFPValue()),
9405 const_cast<ConstantFP*>(TV->getConstantFPValue())
9407 Type *FPTy = Elts[0]->getType();
9408 const DataLayout &TD = *TLI.getDataLayout();
9410 // Create a ConstantArray of the two constants.
9411 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
9412 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
9413 TD.getPrefTypeAlignment(FPTy));
9414 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
9416 // Get the offsets to the 0 and 1 element of the array so that we can
9417 // select between them.
9418 SDValue Zero = DAG.getIntPtrConstant(0);
9419 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
9420 SDValue One = DAG.getIntPtrConstant(EltSize);
9422 SDValue Cond = DAG.getSetCC(DL,
9423 TLI.getSetCCResultType(N0.getValueType()),
9425 AddToWorkList(Cond.getNode());
9426 SDValue CstOffset = DAG.getNode(ISD::SELECT, DL, Zero.getValueType(),
9428 AddToWorkList(CstOffset.getNode());
9429 CPIdx = DAG.getNode(ISD::ADD, DL, TLI.getPointerTy(), CPIdx,
9431 AddToWorkList(CPIdx.getNode());
9432 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
9433 MachinePointerInfo::getConstantPool(), false,
9434 false, false, Alignment);
9439 // Check to see if we can perform the "gzip trick", transforming
9440 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
9441 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
9442 (N1C->isNullValue() || // (a < 0) ? b : 0
9443 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
9444 EVT XType = N0.getValueType();
9445 EVT AType = N2.getValueType();
9446 if (XType.bitsGE(AType)) {
9447 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
9448 // single-bit constant.
9449 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
9450 unsigned ShCtV = N2C->getAPIntValue().logBase2();
9451 ShCtV = XType.getSizeInBits()-ShCtV-1;
9452 SDValue ShCt = DAG.getConstant(ShCtV,
9453 getShiftAmountTy(N0.getValueType()));
9454 SDValue Shift = DAG.getNode(ISD::SRL, N0.getDebugLoc(),
9456 AddToWorkList(Shift.getNode());
9458 if (XType.bitsGT(AType)) {
9459 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
9460 AddToWorkList(Shift.getNode());
9463 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
9466 SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(),
9468 DAG.getConstant(XType.getSizeInBits()-1,
9469 getShiftAmountTy(N0.getValueType())));
9470 AddToWorkList(Shift.getNode());
9472 if (XType.bitsGT(AType)) {
9473 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
9474 AddToWorkList(Shift.getNode());
9477 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
9481 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
9482 // where y is has a single bit set.
9483 // A plaintext description would be, we can turn the SELECT_CC into an AND
9484 // when the condition can be materialized as an all-ones register. Any
9485 // single bit-test can be materialized as an all-ones register with
9486 // shift-left and shift-right-arith.
9487 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
9488 N0->getValueType(0) == VT &&
9489 N1C && N1C->isNullValue() &&
9490 N2C && N2C->isNullValue()) {
9491 SDValue AndLHS = N0->getOperand(0);
9492 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
9493 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
9494 // Shift the tested bit over the sign bit.
9495 APInt AndMask = ConstAndRHS->getAPIntValue();
9497 DAG.getConstant(AndMask.countLeadingZeros(),
9498 getShiftAmountTy(AndLHS.getValueType()));
9499 SDValue Shl = DAG.getNode(ISD::SHL, N0.getDebugLoc(), VT, AndLHS, ShlAmt);
9501 // Now arithmetic right shift it all the way over, so the result is either
9502 // all-ones, or zero.
9504 DAG.getConstant(AndMask.getBitWidth()-1,
9505 getShiftAmountTy(Shl.getValueType()));
9506 SDValue Shr = DAG.getNode(ISD::SRA, N0.getDebugLoc(), VT, Shl, ShrAmt);
9508 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
9512 // fold select C, 16, 0 -> shl C, 4
9513 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
9514 TLI.getBooleanContents(N0.getValueType().isVector()) ==
9515 TargetLowering::ZeroOrOneBooleanContent) {
9517 // If the caller doesn't want us to simplify this into a zext of a compare,
9519 if (NotExtCompare && N2C->getAPIntValue() == 1)
9522 // Get a SetCC of the condition
9523 // NOTE: Don't create a SETCC if it's not legal on this target.
9524 if (!LegalOperations ||
9525 TLI.isOperationLegal(ISD::SETCC,
9526 LegalTypes ? TLI.getSetCCResultType(N0.getValueType()) : MVT::i1)) {
9528 // cast from setcc result type to select result type
9530 SCC = DAG.getSetCC(DL, TLI.getSetCCResultType(N0.getValueType()),
9532 if (N2.getValueType().bitsLT(SCC.getValueType()))
9533 Temp = DAG.getZeroExtendInReg(SCC, N2.getDebugLoc(),
9536 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getDebugLoc(),
9537 N2.getValueType(), SCC);
9539 SCC = DAG.getSetCC(N0.getDebugLoc(), MVT::i1, N0, N1, CC);
9540 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getDebugLoc(),
9541 N2.getValueType(), SCC);
9544 AddToWorkList(SCC.getNode());
9545 AddToWorkList(Temp.getNode());
9547 if (N2C->getAPIntValue() == 1)
9550 // shl setcc result by log2 n2c
9551 return DAG.getNode(ISD::SHL, DL, N2.getValueType(), Temp,
9552 DAG.getConstant(N2C->getAPIntValue().logBase2(),
9553 getShiftAmountTy(Temp.getValueType())));
9557 // Check to see if this is the equivalent of setcc
9558 // FIXME: Turn all of these into setcc if setcc if setcc is legal
9559 // otherwise, go ahead with the folds.
9560 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
9561 EVT XType = N0.getValueType();
9562 if (!LegalOperations ||
9563 TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(XType))) {
9564 SDValue Res = DAG.getSetCC(DL, TLI.getSetCCResultType(XType), N0, N1, CC);
9565 if (Res.getValueType() != VT)
9566 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
9570 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
9571 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
9572 (!LegalOperations ||
9573 TLI.isOperationLegal(ISD::CTLZ, XType))) {
9574 SDValue Ctlz = DAG.getNode(ISD::CTLZ, N0.getDebugLoc(), XType, N0);
9575 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
9576 DAG.getConstant(Log2_32(XType.getSizeInBits()),
9577 getShiftAmountTy(Ctlz.getValueType())));
9579 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
9580 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
9581 SDValue NegN0 = DAG.getNode(ISD::SUB, N0.getDebugLoc(),
9582 XType, DAG.getConstant(0, XType), N0);
9583 SDValue NotN0 = DAG.getNOT(N0.getDebugLoc(), N0, XType);
9584 return DAG.getNode(ISD::SRL, DL, XType,
9585 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
9586 DAG.getConstant(XType.getSizeInBits()-1,
9587 getShiftAmountTy(XType)));
9589 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
9590 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
9591 SDValue Sign = DAG.getNode(ISD::SRL, N0.getDebugLoc(), XType, N0,
9592 DAG.getConstant(XType.getSizeInBits()-1,
9593 getShiftAmountTy(N0.getValueType())));
9594 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
9598 // Check to see if this is an integer abs.
9599 // select_cc setg[te] X, 0, X, -X ->
9600 // select_cc setgt X, -1, X, -X ->
9601 // select_cc setl[te] X, 0, -X, X ->
9602 // select_cc setlt X, 1, -X, X ->
9603 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
9605 ConstantSDNode *SubC = NULL;
9606 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
9607 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
9608 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
9609 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
9610 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
9611 (N1C->isOne() && CC == ISD::SETLT)) &&
9612 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
9613 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
9615 EVT XType = N0.getValueType();
9616 if (SubC && SubC->isNullValue() && XType.isInteger()) {
9617 SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(), XType,
9619 DAG.getConstant(XType.getSizeInBits()-1,
9620 getShiftAmountTy(N0.getValueType())));
9621 SDValue Add = DAG.getNode(ISD::ADD, N0.getDebugLoc(),
9623 AddToWorkList(Shift.getNode());
9624 AddToWorkList(Add.getNode());
9625 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
9632 /// SimplifySetCC - This is a stub for TargetLowering::SimplifySetCC.
9633 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
9634 SDValue N1, ISD::CondCode Cond,
9635 DebugLoc DL, bool foldBooleans) {
9636 TargetLowering::DAGCombinerInfo
9637 DagCombineInfo(DAG, !LegalTypes, !LegalOperations, false, this);
9638 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
9641 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
9642 /// return a DAG expression to select that will generate the same value by
9643 /// multiplying by a magic number. See:
9644 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
9645 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
9646 std::vector<SDNode*> Built;
9647 SDValue S = TLI.BuildSDIV(N, DAG, LegalOperations, &Built);
9649 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
9655 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
9656 /// return a DAG expression to select that will generate the same value by
9657 /// multiplying by a magic number. See:
9658 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
9659 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
9660 std::vector<SDNode*> Built;
9661 SDValue S = TLI.BuildUDIV(N, DAG, LegalOperations, &Built);
9663 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
9669 /// FindBaseOffset - Return true if base is a frame index, which is known not
9670 // to alias with anything but itself. Provides base object and offset as
9672 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
9673 const GlobalValue *&GV, const void *&CV) {
9674 // Assume it is a primitive operation.
9675 Base = Ptr; Offset = 0; GV = 0; CV = 0;
9677 // If it's an adding a simple constant then integrate the offset.
9678 if (Base.getOpcode() == ISD::ADD) {
9679 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
9680 Base = Base.getOperand(0);
9681 Offset += C->getZExtValue();
9685 // Return the underlying GlobalValue, and update the Offset. Return false
9686 // for GlobalAddressSDNode since the same GlobalAddress may be represented
9687 // by multiple nodes with different offsets.
9688 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
9689 GV = G->getGlobal();
9690 Offset += G->getOffset();
9694 // Return the underlying Constant value, and update the Offset. Return false
9695 // for ConstantSDNodes since the same constant pool entry may be represented
9696 // by multiple nodes with different offsets.
9697 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
9698 CV = C->isMachineConstantPoolEntry() ? (const void *)C->getMachineCPVal()
9699 : (const void *)C->getConstVal();
9700 Offset += C->getOffset();
9703 // If it's any of the following then it can't alias with anything but itself.
9704 return isa<FrameIndexSDNode>(Base);
9707 /// isAlias - Return true if there is any possibility that the two addresses
9709 bool DAGCombiner::isAlias(SDValue Ptr1, int64_t Size1,
9710 const Value *SrcValue1, int SrcValueOffset1,
9711 unsigned SrcValueAlign1,
9712 const MDNode *TBAAInfo1,
9713 SDValue Ptr2, int64_t Size2,
9714 const Value *SrcValue2, int SrcValueOffset2,
9715 unsigned SrcValueAlign2,
9716 const MDNode *TBAAInfo2) const {
9717 // If they are the same then they must be aliases.
9718 if (Ptr1 == Ptr2) return true;
9720 // Gather base node and offset information.
9721 SDValue Base1, Base2;
9722 int64_t Offset1, Offset2;
9723 const GlobalValue *GV1, *GV2;
9724 const void *CV1, *CV2;
9725 bool isFrameIndex1 = FindBaseOffset(Ptr1, Base1, Offset1, GV1, CV1);
9726 bool isFrameIndex2 = FindBaseOffset(Ptr2, Base2, Offset2, GV2, CV2);
9728 // If they have a same base address then check to see if they overlap.
9729 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
9730 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
9732 // It is possible for different frame indices to alias each other, mostly
9733 // when tail call optimization reuses return address slots for arguments.
9734 // To catch this case, look up the actual index of frame indices to compute
9735 // the real alias relationship.
9736 if (isFrameIndex1 && isFrameIndex2) {
9737 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
9738 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
9739 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
9740 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
9743 // Otherwise, if we know what the bases are, and they aren't identical, then
9744 // we know they cannot alias.
9745 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
9748 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
9749 // compared to the size and offset of the access, we may be able to prove they
9750 // do not alias. This check is conservative for now to catch cases created by
9751 // splitting vector types.
9752 if ((SrcValueAlign1 == SrcValueAlign2) &&
9753 (SrcValueOffset1 != SrcValueOffset2) &&
9754 (Size1 == Size2) && (SrcValueAlign1 > Size1)) {
9755 int64_t OffAlign1 = SrcValueOffset1 % SrcValueAlign1;
9756 int64_t OffAlign2 = SrcValueOffset2 % SrcValueAlign1;
9758 // There is no overlap between these relatively aligned accesses of similar
9759 // size, return no alias.
9760 if ((OffAlign1 + Size1) <= OffAlign2 || (OffAlign2 + Size2) <= OffAlign1)
9764 if (CombinerGlobalAA) {
9765 // Use alias analysis information.
9766 int64_t MinOffset = std::min(SrcValueOffset1, SrcValueOffset2);
9767 int64_t Overlap1 = Size1 + SrcValueOffset1 - MinOffset;
9768 int64_t Overlap2 = Size2 + SrcValueOffset2 - MinOffset;
9769 AliasAnalysis::AliasResult AAResult =
9770 AA.alias(AliasAnalysis::Location(SrcValue1, Overlap1, TBAAInfo1),
9771 AliasAnalysis::Location(SrcValue2, Overlap2, TBAAInfo2));
9772 if (AAResult == AliasAnalysis::NoAlias)
9776 // Otherwise we have to assume they alias.
9780 bool DAGCombiner::isAlias(LSBaseSDNode *Op0, LSBaseSDNode *Op1) {
9782 int64_t Size0, Size1;
9783 const Value *SrcValue0, *SrcValue1;
9784 int SrcValueOffset0, SrcValueOffset1;
9785 unsigned SrcValueAlign0, SrcValueAlign1;
9786 const MDNode *SrcTBAAInfo0, *SrcTBAAInfo1;
9787 FindAliasInfo(Op0, Ptr0, Size0, SrcValue0, SrcValueOffset0,
9788 SrcValueAlign0, SrcTBAAInfo0);
9789 FindAliasInfo(Op1, Ptr1, Size1, SrcValue1, SrcValueOffset1,
9790 SrcValueAlign1, SrcTBAAInfo1);
9791 return isAlias(Ptr0, Size0, SrcValue0, SrcValueOffset0,
9792 SrcValueAlign0, SrcTBAAInfo0,
9793 Ptr1, Size1, SrcValue1, SrcValueOffset1,
9794 SrcValueAlign1, SrcTBAAInfo1);
9797 /// FindAliasInfo - Extracts the relevant alias information from the memory
9798 /// node. Returns true if the operand was a load.
9799 bool DAGCombiner::FindAliasInfo(SDNode *N,
9800 SDValue &Ptr, int64_t &Size,
9801 const Value *&SrcValue,
9802 int &SrcValueOffset,
9803 unsigned &SrcValueAlign,
9804 const MDNode *&TBAAInfo) const {
9805 LSBaseSDNode *LS = cast<LSBaseSDNode>(N);
9807 Ptr = LS->getBasePtr();
9808 Size = LS->getMemoryVT().getSizeInBits() >> 3;
9809 SrcValue = LS->getSrcValue();
9810 SrcValueOffset = LS->getSrcValueOffset();
9811 SrcValueAlign = LS->getOriginalAlignment();
9812 TBAAInfo = LS->getTBAAInfo();
9813 return isa<LoadSDNode>(LS);
9816 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
9817 /// looking for aliasing nodes and adding them to the Aliases vector.
9818 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
9819 SmallVector<SDValue, 8> &Aliases) {
9820 SmallVector<SDValue, 8> Chains; // List of chains to visit.
9821 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
9823 // Get alias information for node.
9826 const Value *SrcValue;
9828 unsigned SrcValueAlign;
9829 const MDNode *SrcTBAAInfo;
9830 bool IsLoad = FindAliasInfo(N, Ptr, Size, SrcValue, SrcValueOffset,
9831 SrcValueAlign, SrcTBAAInfo);
9834 Chains.push_back(OriginalChain);
9837 // Look at each chain and determine if it is an alias. If so, add it to the
9838 // aliases list. If not, then continue up the chain looking for the next
9840 while (!Chains.empty()) {
9841 SDValue Chain = Chains.back();
9844 // For TokenFactor nodes, look at each operand and only continue up the
9845 // chain until we find two aliases. If we've seen two aliases, assume we'll
9846 // find more and revert to original chain since the xform is unlikely to be
9849 // FIXME: The depth check could be made to return the last non-aliasing
9850 // chain we found before we hit a tokenfactor rather than the original
9852 if (Depth > 6 || Aliases.size() == 2) {
9854 Aliases.push_back(OriginalChain);
9858 // Don't bother if we've been before.
9859 if (!Visited.insert(Chain.getNode()))
9862 switch (Chain.getOpcode()) {
9863 case ISD::EntryToken:
9864 // Entry token is ideal chain operand, but handled in FindBetterChain.
9869 // Get alias information for Chain.
9872 const Value *OpSrcValue;
9873 int OpSrcValueOffset;
9874 unsigned OpSrcValueAlign;
9875 const MDNode *OpSrcTBAAInfo;
9876 bool IsOpLoad = FindAliasInfo(Chain.getNode(), OpPtr, OpSize,
9877 OpSrcValue, OpSrcValueOffset,
9881 // If chain is alias then stop here.
9882 if (!(IsLoad && IsOpLoad) &&
9883 isAlias(Ptr, Size, SrcValue, SrcValueOffset, SrcValueAlign,
9885 OpPtr, OpSize, OpSrcValue, OpSrcValueOffset,
9886 OpSrcValueAlign, OpSrcTBAAInfo)) {
9887 Aliases.push_back(Chain);
9889 // Look further up the chain.
9890 Chains.push_back(Chain.getOperand(0));
9896 case ISD::TokenFactor:
9897 // We have to check each of the operands of the token factor for "small"
9898 // token factors, so we queue them up. Adding the operands to the queue
9899 // (stack) in reverse order maintains the original order and increases the
9900 // likelihood that getNode will find a matching token factor (CSE.)
9901 if (Chain.getNumOperands() > 16) {
9902 Aliases.push_back(Chain);
9905 for (unsigned n = Chain.getNumOperands(); n;)
9906 Chains.push_back(Chain.getOperand(--n));
9911 // For all other instructions we will just have to take what we can get.
9912 Aliases.push_back(Chain);
9918 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
9919 /// for a better chain (aliasing node.)
9920 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
9921 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
9923 // Accumulate all the aliases to this node.
9924 GatherAllAliases(N, OldChain, Aliases);
9926 // If no operands then chain to entry token.
9927 if (Aliases.size() == 0)
9928 return DAG.getEntryNode();
9930 // If a single operand then chain to it. We don't need to revisit it.
9931 if (Aliases.size() == 1)
9934 // Construct a custom tailored token factor.
9935 return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
9936 &Aliases[0], Aliases.size());
9939 // SelectionDAG::Combine - This is the entry point for the file.
9941 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
9942 CodeGenOpt::Level OptLevel) {
9943 /// run - This is the main entry point to this class.
9945 DAGCombiner(*this, AA, OptLevel).Run(Level);