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/DerivedTypes.h"
22 #include "llvm/LLVMContext.h"
23 #include "llvm/CodeGen/MachineFunction.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/Analysis/AliasAnalysis.h"
26 #include "llvm/Target/TargetData.h"
27 #include "llvm/Target/TargetLowering.h"
28 #include "llvm/Target/TargetMachine.h"
29 #include "llvm/Target/TargetOptions.h"
30 #include "llvm/ADT/SmallPtrSet.h"
31 #include "llvm/ADT/Statistic.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/ErrorHandling.h"
35 #include "llvm/Support/MathExtras.h"
36 #include "llvm/Support/raw_ostream.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 visitSHL(SDNode *N);
198 SDValue visitSRA(SDNode *N);
199 SDValue visitSRL(SDNode *N);
200 SDValue visitCTLZ(SDNode *N);
201 SDValue visitCTLZ_ZERO_UNDEF(SDNode *N);
202 SDValue visitCTTZ(SDNode *N);
203 SDValue visitCTTZ_ZERO_UNDEF(SDNode *N);
204 SDValue visitCTPOP(SDNode *N);
205 SDValue visitSELECT(SDNode *N);
206 SDValue visitSELECT_CC(SDNode *N);
207 SDValue visitSETCC(SDNode *N);
208 SDValue visitSIGN_EXTEND(SDNode *N);
209 SDValue visitZERO_EXTEND(SDNode *N);
210 SDValue visitANY_EXTEND(SDNode *N);
211 SDValue visitSIGN_EXTEND_INREG(SDNode *N);
212 SDValue visitTRUNCATE(SDNode *N);
213 SDValue visitBITCAST(SDNode *N);
214 SDValue visitBUILD_PAIR(SDNode *N);
215 SDValue visitFADD(SDNode *N);
216 SDValue visitFSUB(SDNode *N);
217 SDValue visitFMUL(SDNode *N);
218 SDValue visitFDIV(SDNode *N);
219 SDValue visitFREM(SDNode *N);
220 SDValue visitFCOPYSIGN(SDNode *N);
221 SDValue visitSINT_TO_FP(SDNode *N);
222 SDValue visitUINT_TO_FP(SDNode *N);
223 SDValue visitFP_TO_SINT(SDNode *N);
224 SDValue visitFP_TO_UINT(SDNode *N);
225 SDValue visitFP_ROUND(SDNode *N);
226 SDValue visitFP_ROUND_INREG(SDNode *N);
227 SDValue visitFP_EXTEND(SDNode *N);
228 SDValue visitFNEG(SDNode *N);
229 SDValue visitFABS(SDNode *N);
230 SDValue visitBRCOND(SDNode *N);
231 SDValue visitBR_CC(SDNode *N);
232 SDValue visitLOAD(SDNode *N);
233 SDValue visitSTORE(SDNode *N);
234 SDValue visitINSERT_VECTOR_ELT(SDNode *N);
235 SDValue visitEXTRACT_VECTOR_ELT(SDNode *N);
236 SDValue visitBUILD_VECTOR(SDNode *N);
237 SDValue visitCONCAT_VECTORS(SDNode *N);
238 SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
239 SDValue visitVECTOR_SHUFFLE(SDNode *N);
240 SDValue visitMEMBARRIER(SDNode *N);
242 SDValue XformToShuffleWithZero(SDNode *N);
243 SDValue ReassociateOps(unsigned Opc, DebugLoc DL, SDValue LHS, SDValue RHS);
245 SDValue visitShiftByConstant(SDNode *N, unsigned Amt);
247 bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
248 SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
249 SDValue SimplifySelect(DebugLoc DL, SDValue N0, SDValue N1, SDValue N2);
250 SDValue SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1, SDValue N2,
251 SDValue N3, ISD::CondCode CC,
252 bool NotExtCompare = false);
253 SDValue SimplifySetCC(EVT VT, SDValue N0, SDValue N1, ISD::CondCode Cond,
254 DebugLoc DL, bool foldBooleans = true);
255 SDValue SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
257 SDValue CombineConsecutiveLoads(SDNode *N, EVT VT);
258 SDValue ConstantFoldBITCASTofBUILD_VECTOR(SDNode *, EVT);
259 SDValue BuildSDIV(SDNode *N);
260 SDValue BuildUDIV(SDNode *N);
261 SDValue MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
262 bool DemandHighBits = true);
263 SDValue MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1);
264 SDNode *MatchRotate(SDValue LHS, SDValue RHS, DebugLoc DL);
265 SDValue ReduceLoadWidth(SDNode *N);
266 SDValue ReduceLoadOpStoreWidth(SDNode *N);
267 SDValue TransformFPLoadStorePair(SDNode *N);
269 SDValue GetDemandedBits(SDValue V, const APInt &Mask);
271 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
272 /// looking for aliasing nodes and adding them to the Aliases vector.
273 void GatherAllAliases(SDNode *N, SDValue OriginalChain,
274 SmallVector<SDValue, 8> &Aliases);
276 /// isAlias - Return true if there is any possibility that the two addresses
278 bool isAlias(SDValue Ptr1, int64_t Size1,
279 const Value *SrcValue1, int SrcValueOffset1,
280 unsigned SrcValueAlign1,
281 const MDNode *TBAAInfo1,
282 SDValue Ptr2, int64_t Size2,
283 const Value *SrcValue2, int SrcValueOffset2,
284 unsigned SrcValueAlign2,
285 const MDNode *TBAAInfo2) const;
287 /// FindAliasInfo - Extracts the relevant alias information from the memory
288 /// node. Returns true if the operand was a load.
289 bool FindAliasInfo(SDNode *N,
290 SDValue &Ptr, int64_t &Size,
291 const Value *&SrcValue, int &SrcValueOffset,
292 unsigned &SrcValueAlignment,
293 const MDNode *&TBAAInfo) const;
295 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes,
296 /// looking for a better chain (aliasing node.)
297 SDValue FindBetterChain(SDNode *N, SDValue Chain);
300 DAGCombiner(SelectionDAG &D, AliasAnalysis &A, CodeGenOpt::Level OL)
301 : DAG(D), TLI(D.getTargetLoweringInfo()), Level(BeforeLegalizeTypes),
302 OptLevel(OL), LegalOperations(false), LegalTypes(false), AA(A) {}
304 /// Run - runs the dag combiner on all nodes in the work list
305 void Run(CombineLevel AtLevel);
307 SelectionDAG &getDAG() const { return DAG; }
309 /// getShiftAmountTy - Returns a type large enough to hold any valid
310 /// shift amount - before type legalization these can be huge.
311 EVT getShiftAmountTy(EVT LHSTy) {
312 return LegalTypes ? TLI.getShiftAmountTy(LHSTy) : TLI.getPointerTy();
315 /// isTypeLegal - This method returns true if we are running before type
316 /// legalization or if the specified VT is legal.
317 bool isTypeLegal(const EVT &VT) {
318 if (!LegalTypes) return true;
319 return TLI.isTypeLegal(VT);
326 /// WorkListRemover - This class is a DAGUpdateListener that removes any deleted
327 /// nodes from the worklist.
328 class WorkListRemover : public SelectionDAG::DAGUpdateListener {
331 explicit WorkListRemover(DAGCombiner &dc) : DC(dc) {}
333 virtual void NodeDeleted(SDNode *N, SDNode *E) {
334 DC.removeFromWorkList(N);
337 virtual void NodeUpdated(SDNode *N) {
343 //===----------------------------------------------------------------------===//
344 // TargetLowering::DAGCombinerInfo implementation
345 //===----------------------------------------------------------------------===//
347 void TargetLowering::DAGCombinerInfo::AddToWorklist(SDNode *N) {
348 ((DAGCombiner*)DC)->AddToWorkList(N);
351 void TargetLowering::DAGCombinerInfo::RemoveFromWorklist(SDNode *N) {
352 ((DAGCombiner*)DC)->removeFromWorkList(N);
355 SDValue TargetLowering::DAGCombinerInfo::
356 CombineTo(SDNode *N, const std::vector<SDValue> &To, bool AddTo) {
357 return ((DAGCombiner*)DC)->CombineTo(N, &To[0], To.size(), AddTo);
360 SDValue TargetLowering::DAGCombinerInfo::
361 CombineTo(SDNode *N, SDValue Res, bool AddTo) {
362 return ((DAGCombiner*)DC)->CombineTo(N, Res, AddTo);
366 SDValue TargetLowering::DAGCombinerInfo::
367 CombineTo(SDNode *N, SDValue Res0, SDValue Res1, bool AddTo) {
368 return ((DAGCombiner*)DC)->CombineTo(N, Res0, Res1, AddTo);
371 void TargetLowering::DAGCombinerInfo::
372 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
373 return ((DAGCombiner*)DC)->CommitTargetLoweringOpt(TLO);
376 //===----------------------------------------------------------------------===//
378 //===----------------------------------------------------------------------===//
380 /// isNegatibleForFree - Return 1 if we can compute the negated form of the
381 /// specified expression for the same cost as the expression itself, or 2 if we
382 /// can compute the negated form more cheaply than the expression itself.
383 static char isNegatibleForFree(SDValue Op, bool LegalOperations,
384 const TargetLowering &TLI,
385 const TargetOptions *Options,
386 unsigned Depth = 0) {
387 // No compile time optimizations on this type.
388 if (Op.getValueType() == MVT::ppcf128)
391 // fneg is removable even if it has multiple uses.
392 if (Op.getOpcode() == ISD::FNEG) return 2;
394 // Don't allow anything with multiple uses.
395 if (!Op.hasOneUse()) return 0;
397 // Don't recurse exponentially.
398 if (Depth > 6) return 0;
400 switch (Op.getOpcode()) {
401 default: return false;
402 case ISD::ConstantFP:
403 // Don't invert constant FP values after legalize. The negated constant
404 // isn't necessarily legal.
405 return LegalOperations ? 0 : 1;
407 // FIXME: determine better conditions for this xform.
408 if (!Options->UnsafeFPMath) return 0;
410 // After operation legalization, it might not be legal to create new FSUBs.
411 if (LegalOperations &&
412 !TLI.isOperationLegalOrCustom(ISD::FSUB, Op.getValueType()))
415 // fold (fsub (fadd A, B)) -> (fsub (fneg A), B)
416 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
419 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
420 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
423 // We can't turn -(A-B) into B-A when we honor signed zeros.
424 if (!Options->UnsafeFPMath) return 0;
426 // fold (fneg (fsub A, B)) -> (fsub B, A)
431 if (Options->HonorSignDependentRoundingFPMath()) return 0;
433 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y))
434 if (char V = isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI,
438 return isNegatibleForFree(Op.getOperand(1), LegalOperations, TLI, Options,
444 return isNegatibleForFree(Op.getOperand(0), LegalOperations, TLI, Options,
449 /// GetNegatedExpression - If isNegatibleForFree returns true, this function
450 /// returns the newly negated expression.
451 static SDValue GetNegatedExpression(SDValue Op, SelectionDAG &DAG,
452 bool LegalOperations, unsigned Depth = 0) {
453 // fneg is removable even if it has multiple uses.
454 if (Op.getOpcode() == ISD::FNEG) return Op.getOperand(0);
456 // Don't allow anything with multiple uses.
457 assert(Op.hasOneUse() && "Unknown reuse!");
459 assert(Depth <= 6 && "GetNegatedExpression doesn't match isNegatibleForFree");
460 switch (Op.getOpcode()) {
461 default: llvm_unreachable("Unknown code");
462 case ISD::ConstantFP: {
463 APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF();
465 return DAG.getConstantFP(V, Op.getValueType());
468 // FIXME: determine better conditions for this xform.
469 assert(DAG.getTarget().Options.UnsafeFPMath);
471 // fold (fneg (fadd A, B)) -> (fsub (fneg A), B)
472 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
473 DAG.getTargetLoweringInfo(),
474 &DAG.getTarget().Options, Depth+1))
475 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
476 GetNegatedExpression(Op.getOperand(0), DAG,
477 LegalOperations, Depth+1),
479 // fold (fneg (fadd A, B)) -> (fsub (fneg B), A)
480 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
481 GetNegatedExpression(Op.getOperand(1), DAG,
482 LegalOperations, Depth+1),
485 // We can't turn -(A-B) into B-A when we honor signed zeros.
486 assert(DAG.getTarget().Options.UnsafeFPMath);
488 // fold (fneg (fsub 0, B)) -> B
489 if (ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(Op.getOperand(0)))
490 if (N0CFP->getValueAPF().isZero())
491 return Op.getOperand(1);
493 // fold (fneg (fsub A, B)) -> (fsub B, A)
494 return DAG.getNode(ISD::FSUB, Op.getDebugLoc(), Op.getValueType(),
495 Op.getOperand(1), Op.getOperand(0));
499 assert(!DAG.getTarget().Options.HonorSignDependentRoundingFPMath());
501 // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y)
502 if (isNegatibleForFree(Op.getOperand(0), LegalOperations,
503 DAG.getTargetLoweringInfo(),
504 &DAG.getTarget().Options, Depth+1))
505 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
506 GetNegatedExpression(Op.getOperand(0), DAG,
507 LegalOperations, Depth+1),
510 // fold (fneg (fmul X, Y)) -> (fmul X, (fneg Y))
511 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
513 GetNegatedExpression(Op.getOperand(1), DAG,
514 LegalOperations, Depth+1));
518 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), Op.getValueType(),
519 GetNegatedExpression(Op.getOperand(0), DAG,
520 LegalOperations, Depth+1));
522 return DAG.getNode(ISD::FP_ROUND, Op.getDebugLoc(), Op.getValueType(),
523 GetNegatedExpression(Op.getOperand(0), DAG,
524 LegalOperations, Depth+1),
530 // isSetCCEquivalent - Return true if this node is a setcc, or is a select_cc
531 // that selects between the values 1 and 0, making it equivalent to a setcc.
532 // Also, set the incoming LHS, RHS, and CC references to the appropriate
533 // nodes based on the type of node we are checking. This simplifies life a
534 // bit for the callers.
535 static bool isSetCCEquivalent(SDValue N, SDValue &LHS, SDValue &RHS,
537 if (N.getOpcode() == ISD::SETCC) {
538 LHS = N.getOperand(0);
539 RHS = N.getOperand(1);
540 CC = N.getOperand(2);
543 if (N.getOpcode() == ISD::SELECT_CC &&
544 N.getOperand(2).getOpcode() == ISD::Constant &&
545 N.getOperand(3).getOpcode() == ISD::Constant &&
546 cast<ConstantSDNode>(N.getOperand(2))->getAPIntValue() == 1 &&
547 cast<ConstantSDNode>(N.getOperand(3))->isNullValue()) {
548 LHS = N.getOperand(0);
549 RHS = N.getOperand(1);
550 CC = N.getOperand(4);
556 // isOneUseSetCC - Return true if this is a SetCC-equivalent operation with only
557 // one use. If this is true, it allows the users to invert the operation for
558 // free when it is profitable to do so.
559 static bool isOneUseSetCC(SDValue N) {
561 if (isSetCCEquivalent(N, N0, N1, N2) && N.getNode()->hasOneUse())
566 SDValue DAGCombiner::ReassociateOps(unsigned Opc, DebugLoc DL,
567 SDValue N0, SDValue N1) {
568 EVT VT = N0.getValueType();
569 if (N0.getOpcode() == Opc && isa<ConstantSDNode>(N0.getOperand(1))) {
570 if (isa<ConstantSDNode>(N1)) {
571 // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
573 DAG.FoldConstantArithmetic(Opc, VT,
574 cast<ConstantSDNode>(N0.getOperand(1)),
575 cast<ConstantSDNode>(N1));
576 return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
578 if (N0.hasOneUse()) {
579 // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one use
580 SDValue OpNode = DAG.getNode(Opc, N0.getDebugLoc(), VT,
581 N0.getOperand(0), N1);
582 AddToWorkList(OpNode.getNode());
583 return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
587 if (N1.getOpcode() == Opc && isa<ConstantSDNode>(N1.getOperand(1))) {
588 if (isa<ConstantSDNode>(N0)) {
589 // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
591 DAG.FoldConstantArithmetic(Opc, VT,
592 cast<ConstantSDNode>(N1.getOperand(1)),
593 cast<ConstantSDNode>(N0));
594 return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
596 if (N1.hasOneUse()) {
597 // reassoc. (op y, (op x, c1)) -> (op (op x, y), c1) iff x+c1 has one use
598 SDValue OpNode = DAG.getNode(Opc, N0.getDebugLoc(), VT,
599 N1.getOperand(0), N0);
600 AddToWorkList(OpNode.getNode());
601 return DAG.getNode(Opc, DL, VT, OpNode, N1.getOperand(1));
608 SDValue DAGCombiner::CombineTo(SDNode *N, const SDValue *To, unsigned NumTo,
610 assert(N->getNumValues() == NumTo && "Broken CombineTo call!");
612 DEBUG(dbgs() << "\nReplacing.1 ";
614 dbgs() << "\nWith: ";
615 To[0].getNode()->dump(&DAG);
616 dbgs() << " and " << NumTo-1 << " other values\n";
617 for (unsigned i = 0, e = NumTo; i != e; ++i)
618 assert((!To[i].getNode() ||
619 N->getValueType(i) == To[i].getValueType()) &&
620 "Cannot combine value to value of different type!"));
621 WorkListRemover DeadNodes(*this);
622 DAG.ReplaceAllUsesWith(N, To, &DeadNodes);
625 // Push the new nodes and any users onto the worklist
626 for (unsigned i = 0, e = NumTo; i != e; ++i) {
627 if (To[i].getNode()) {
628 AddToWorkList(To[i].getNode());
629 AddUsersToWorkList(To[i].getNode());
634 // Finally, if the node is now dead, remove it from the graph. The node
635 // may not be dead if the replacement process recursively simplified to
636 // something else needing this node.
637 if (N->use_empty()) {
638 // Nodes can be reintroduced into the worklist. Make sure we do not
639 // process a node that has been replaced.
640 removeFromWorkList(N);
642 // Finally, since the node is now dead, remove it from the graph.
645 return SDValue(N, 0);
649 CommitTargetLoweringOpt(const TargetLowering::TargetLoweringOpt &TLO) {
650 // Replace all uses. If any nodes become isomorphic to other nodes and
651 // are deleted, make sure to remove them from our worklist.
652 WorkListRemover DeadNodes(*this);
653 DAG.ReplaceAllUsesOfValueWith(TLO.Old, TLO.New, &DeadNodes);
655 // Push the new node and any (possibly new) users onto the worklist.
656 AddToWorkList(TLO.New.getNode());
657 AddUsersToWorkList(TLO.New.getNode());
659 // Finally, if the node is now dead, remove it from the graph. The node
660 // may not be dead if the replacement process recursively simplified to
661 // something else needing this node.
662 if (TLO.Old.getNode()->use_empty()) {
663 removeFromWorkList(TLO.Old.getNode());
665 // If the operands of this node are only used by the node, they will now
666 // be dead. Make sure to visit them first to delete dead nodes early.
667 for (unsigned i = 0, e = TLO.Old.getNode()->getNumOperands(); i != e; ++i)
668 if (TLO.Old.getNode()->getOperand(i).getNode()->hasOneUse())
669 AddToWorkList(TLO.Old.getNode()->getOperand(i).getNode());
671 DAG.DeleteNode(TLO.Old.getNode());
675 /// SimplifyDemandedBits - Check the specified integer node value to see if
676 /// it can be simplified or if things it uses can be simplified by bit
677 /// propagation. If so, return true.
678 bool DAGCombiner::SimplifyDemandedBits(SDValue Op, const APInt &Demanded) {
679 TargetLowering::TargetLoweringOpt TLO(DAG, LegalTypes, LegalOperations);
680 APInt KnownZero, KnownOne;
681 if (!TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
685 AddToWorkList(Op.getNode());
687 // Replace the old value with the new one.
689 DEBUG(dbgs() << "\nReplacing.2 ";
690 TLO.Old.getNode()->dump(&DAG);
691 dbgs() << "\nWith: ";
692 TLO.New.getNode()->dump(&DAG);
695 CommitTargetLoweringOpt(TLO);
699 void DAGCombiner::ReplaceLoadWithPromotedLoad(SDNode *Load, SDNode *ExtLoad) {
700 DebugLoc dl = Load->getDebugLoc();
701 EVT VT = Load->getValueType(0);
702 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, VT, SDValue(ExtLoad, 0));
704 DEBUG(dbgs() << "\nReplacing.9 ";
706 dbgs() << "\nWith: ";
707 Trunc.getNode()->dump(&DAG);
709 WorkListRemover DeadNodes(*this);
710 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 0), Trunc, &DeadNodes);
711 DAG.ReplaceAllUsesOfValueWith(SDValue(Load, 1), SDValue(ExtLoad, 1),
713 removeFromWorkList(Load);
714 DAG.DeleteNode(Load);
715 AddToWorkList(Trunc.getNode());
718 SDValue DAGCombiner::PromoteOperand(SDValue Op, EVT PVT, bool &Replace) {
720 DebugLoc dl = Op.getDebugLoc();
721 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Op)) {
722 EVT MemVT = LD->getMemoryVT();
723 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
724 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
726 : LD->getExtensionType();
728 return DAG.getExtLoad(ExtType, dl, PVT,
729 LD->getChain(), LD->getBasePtr(),
730 LD->getPointerInfo(),
731 MemVT, LD->isVolatile(),
732 LD->isNonTemporal(), LD->getAlignment());
735 unsigned Opc = Op.getOpcode();
738 case ISD::AssertSext:
739 return DAG.getNode(ISD::AssertSext, dl, PVT,
740 SExtPromoteOperand(Op.getOperand(0), PVT),
742 case ISD::AssertZext:
743 return DAG.getNode(ISD::AssertZext, dl, PVT,
744 ZExtPromoteOperand(Op.getOperand(0), PVT),
746 case ISD::Constant: {
748 Op.getValueType().isByteSized() ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
749 return DAG.getNode(ExtOpc, dl, PVT, Op);
753 if (!TLI.isOperationLegal(ISD::ANY_EXTEND, PVT))
755 return DAG.getNode(ISD::ANY_EXTEND, dl, PVT, Op);
758 SDValue DAGCombiner::SExtPromoteOperand(SDValue Op, EVT PVT) {
759 if (!TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, PVT))
761 EVT OldVT = Op.getValueType();
762 DebugLoc dl = Op.getDebugLoc();
763 bool Replace = false;
764 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
765 if (NewOp.getNode() == 0)
767 AddToWorkList(NewOp.getNode());
770 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
771 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, NewOp.getValueType(), NewOp,
772 DAG.getValueType(OldVT));
775 SDValue DAGCombiner::ZExtPromoteOperand(SDValue Op, EVT PVT) {
776 EVT OldVT = Op.getValueType();
777 DebugLoc dl = Op.getDebugLoc();
778 bool Replace = false;
779 SDValue NewOp = PromoteOperand(Op, PVT, Replace);
780 if (NewOp.getNode() == 0)
782 AddToWorkList(NewOp.getNode());
785 ReplaceLoadWithPromotedLoad(Op.getNode(), NewOp.getNode());
786 return DAG.getZeroExtendInReg(NewOp, dl, OldVT);
789 /// PromoteIntBinOp - Promote the specified integer binary operation if the
790 /// target indicates it is beneficial. e.g. On x86, it's usually better to
791 /// promote i16 operations to i32 since i16 instructions are longer.
792 SDValue DAGCombiner::PromoteIntBinOp(SDValue Op) {
793 if (!LegalOperations)
796 EVT VT = Op.getValueType();
797 if (VT.isVector() || !VT.isInteger())
800 // If operation type is 'undesirable', e.g. i16 on x86, consider
802 unsigned Opc = Op.getOpcode();
803 if (TLI.isTypeDesirableForOp(Opc, VT))
807 // Consult target whether it is a good idea to promote this operation and
808 // what's the right type to promote it to.
809 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
810 assert(PVT != VT && "Don't know what type to promote to!");
812 bool Replace0 = false;
813 SDValue N0 = Op.getOperand(0);
814 SDValue NN0 = PromoteOperand(N0, PVT, Replace0);
815 if (NN0.getNode() == 0)
818 bool Replace1 = false;
819 SDValue N1 = Op.getOperand(1);
824 NN1 = PromoteOperand(N1, PVT, Replace1);
825 if (NN1.getNode() == 0)
829 AddToWorkList(NN0.getNode());
831 AddToWorkList(NN1.getNode());
834 ReplaceLoadWithPromotedLoad(N0.getNode(), NN0.getNode());
836 ReplaceLoadWithPromotedLoad(N1.getNode(), NN1.getNode());
838 DEBUG(dbgs() << "\nPromoting ";
839 Op.getNode()->dump(&DAG));
840 DebugLoc dl = Op.getDebugLoc();
841 return DAG.getNode(ISD::TRUNCATE, dl, VT,
842 DAG.getNode(Opc, dl, PVT, NN0, NN1));
847 /// PromoteIntShiftOp - Promote the specified integer shift operation if the
848 /// target indicates it is beneficial. e.g. On x86, it's usually better to
849 /// promote i16 operations to i32 since i16 instructions are longer.
850 SDValue DAGCombiner::PromoteIntShiftOp(SDValue Op) {
851 if (!LegalOperations)
854 EVT VT = Op.getValueType();
855 if (VT.isVector() || !VT.isInteger())
858 // If operation type is 'undesirable', e.g. i16 on x86, consider
860 unsigned Opc = Op.getOpcode();
861 if (TLI.isTypeDesirableForOp(Opc, VT))
865 // Consult target whether it is a good idea to promote this operation and
866 // what's the right type to promote it to.
867 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
868 assert(PVT != VT && "Don't know what type to promote to!");
870 bool Replace = false;
871 SDValue N0 = Op.getOperand(0);
873 N0 = SExtPromoteOperand(Op.getOperand(0), PVT);
874 else if (Opc == ISD::SRL)
875 N0 = ZExtPromoteOperand(Op.getOperand(0), PVT);
877 N0 = PromoteOperand(N0, PVT, Replace);
878 if (N0.getNode() == 0)
881 AddToWorkList(N0.getNode());
883 ReplaceLoadWithPromotedLoad(Op.getOperand(0).getNode(), N0.getNode());
885 DEBUG(dbgs() << "\nPromoting ";
886 Op.getNode()->dump(&DAG));
887 DebugLoc dl = Op.getDebugLoc();
888 return DAG.getNode(ISD::TRUNCATE, dl, VT,
889 DAG.getNode(Opc, dl, PVT, N0, Op.getOperand(1)));
894 SDValue DAGCombiner::PromoteExtend(SDValue Op) {
895 if (!LegalOperations)
898 EVT VT = Op.getValueType();
899 if (VT.isVector() || !VT.isInteger())
902 // If operation type is 'undesirable', e.g. i16 on x86, consider
904 unsigned Opc = Op.getOpcode();
905 if (TLI.isTypeDesirableForOp(Opc, VT))
909 // Consult target whether it is a good idea to promote this operation and
910 // what's the right type to promote it to.
911 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
912 assert(PVT != VT && "Don't know what type to promote to!");
913 // fold (aext (aext x)) -> (aext x)
914 // fold (aext (zext x)) -> (zext x)
915 // fold (aext (sext x)) -> (sext x)
916 DEBUG(dbgs() << "\nPromoting ";
917 Op.getNode()->dump(&DAG));
918 return DAG.getNode(Op.getOpcode(), Op.getDebugLoc(), VT, Op.getOperand(0));
923 bool DAGCombiner::PromoteLoad(SDValue Op) {
924 if (!LegalOperations)
927 EVT VT = Op.getValueType();
928 if (VT.isVector() || !VT.isInteger())
931 // If operation type is 'undesirable', e.g. i16 on x86, consider
933 unsigned Opc = Op.getOpcode();
934 if (TLI.isTypeDesirableForOp(Opc, VT))
938 // Consult target whether it is a good idea to promote this operation and
939 // what's the right type to promote it to.
940 if (TLI.IsDesirableToPromoteOp(Op, PVT)) {
941 assert(PVT != VT && "Don't know what type to promote to!");
943 DebugLoc dl = Op.getDebugLoc();
944 SDNode *N = Op.getNode();
945 LoadSDNode *LD = cast<LoadSDNode>(N);
946 EVT MemVT = LD->getMemoryVT();
947 ISD::LoadExtType ExtType = ISD::isNON_EXTLoad(LD)
948 ? (TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT) ? ISD::ZEXTLOAD
950 : LD->getExtensionType();
951 SDValue NewLD = DAG.getExtLoad(ExtType, dl, PVT,
952 LD->getChain(), LD->getBasePtr(),
953 LD->getPointerInfo(),
954 MemVT, LD->isVolatile(),
955 LD->isNonTemporal(), LD->getAlignment());
956 SDValue Result = DAG.getNode(ISD::TRUNCATE, dl, VT, NewLD);
958 DEBUG(dbgs() << "\nPromoting ";
961 Result.getNode()->dump(&DAG);
963 WorkListRemover DeadNodes(*this);
964 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result, &DeadNodes);
965 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), NewLD.getValue(1), &DeadNodes);
966 removeFromWorkList(N);
968 AddToWorkList(Result.getNode());
975 //===----------------------------------------------------------------------===//
976 // Main DAG Combiner implementation
977 //===----------------------------------------------------------------------===//
979 void DAGCombiner::Run(CombineLevel AtLevel) {
980 // set the instance variables, so that the various visit routines may use it.
982 LegalOperations = Level >= AfterLegalizeVectorOps;
983 LegalTypes = Level >= AfterLegalizeTypes;
985 // Add all the dag nodes to the worklist.
986 for (SelectionDAG::allnodes_iterator I = DAG.allnodes_begin(),
987 E = DAG.allnodes_end(); I != E; ++I)
990 // Create a dummy node (which is not added to allnodes), that adds a reference
991 // to the root node, preventing it from being deleted, and tracking any
992 // changes of the root.
993 HandleSDNode Dummy(DAG.getRoot());
995 // The root of the dag may dangle to deleted nodes until the dag combiner is
996 // done. Set it to null to avoid confusion.
997 DAG.setRoot(SDValue());
999 // while the worklist isn't empty, find a node and
1000 // try and combine it.
1001 while (!WorkListContents.empty()) {
1003 // The WorkListOrder holds the SDNodes in order, but it may contain duplicates.
1004 // In order to avoid a linear scan, we use a set (O(log N)) to hold what the
1005 // worklist *should* contain, and check the node we want to visit is should
1006 // actually be visited.
1008 N = WorkListOrder.pop_back_val();
1009 } while (!WorkListContents.erase(N));
1011 // If N has no uses, it is dead. Make sure to revisit all N's operands once
1012 // N is deleted from the DAG, since they too may now be dead or may have a
1013 // reduced number of uses, allowing other xforms.
1014 if (N->use_empty() && N != &Dummy) {
1015 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1016 AddToWorkList(N->getOperand(i).getNode());
1022 SDValue RV = combine(N);
1024 if (RV.getNode() == 0)
1029 // If we get back the same node we passed in, rather than a new node or
1030 // zero, we know that the node must have defined multiple values and
1031 // CombineTo was used. Since CombineTo takes care of the worklist
1032 // mechanics for us, we have no work to do in this case.
1033 if (RV.getNode() == N)
1036 assert(N->getOpcode() != ISD::DELETED_NODE &&
1037 RV.getNode()->getOpcode() != ISD::DELETED_NODE &&
1038 "Node was deleted but visit returned new node!");
1040 DEBUG(dbgs() << "\nReplacing.3 ";
1042 dbgs() << "\nWith: ";
1043 RV.getNode()->dump(&DAG);
1046 // Transfer debug value.
1047 DAG.TransferDbgValues(SDValue(N, 0), RV);
1048 WorkListRemover DeadNodes(*this);
1049 if (N->getNumValues() == RV.getNode()->getNumValues())
1050 DAG.ReplaceAllUsesWith(N, RV.getNode(), &DeadNodes);
1052 assert(N->getValueType(0) == RV.getValueType() &&
1053 N->getNumValues() == 1 && "Type mismatch");
1055 DAG.ReplaceAllUsesWith(N, &OpV, &DeadNodes);
1058 // Push the new node and any users onto the worklist
1059 AddToWorkList(RV.getNode());
1060 AddUsersToWorkList(RV.getNode());
1062 // Add any uses of the old node to the worklist in case this node is the
1063 // last one that uses them. They may become dead after this node is
1065 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1066 AddToWorkList(N->getOperand(i).getNode());
1068 // Finally, if the node is now dead, remove it from the graph. The node
1069 // may not be dead if the replacement process recursively simplified to
1070 // something else needing this node.
1071 if (N->use_empty()) {
1072 // Nodes can be reintroduced into the worklist. Make sure we do not
1073 // process a node that has been replaced.
1074 removeFromWorkList(N);
1076 // Finally, since the node is now dead, remove it from the graph.
1081 // If the root changed (e.g. it was a dead load, update the root).
1082 DAG.setRoot(Dummy.getValue());
1083 DAG.RemoveDeadNodes();
1086 SDValue DAGCombiner::visit(SDNode *N) {
1087 switch (N->getOpcode()) {
1089 case ISD::TokenFactor: return visitTokenFactor(N);
1090 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1091 case ISD::ADD: return visitADD(N);
1092 case ISD::SUB: return visitSUB(N);
1093 case ISD::ADDC: return visitADDC(N);
1094 case ISD::SUBC: return visitSUBC(N);
1095 case ISD::ADDE: return visitADDE(N);
1096 case ISD::SUBE: return visitSUBE(N);
1097 case ISD::MUL: return visitMUL(N);
1098 case ISD::SDIV: return visitSDIV(N);
1099 case ISD::UDIV: return visitUDIV(N);
1100 case ISD::SREM: return visitSREM(N);
1101 case ISD::UREM: return visitUREM(N);
1102 case ISD::MULHU: return visitMULHU(N);
1103 case ISD::MULHS: return visitMULHS(N);
1104 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1105 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1106 case ISD::SMULO: return visitSMULO(N);
1107 case ISD::UMULO: return visitUMULO(N);
1108 case ISD::SDIVREM: return visitSDIVREM(N);
1109 case ISD::UDIVREM: return visitUDIVREM(N);
1110 case ISD::AND: return visitAND(N);
1111 case ISD::OR: return visitOR(N);
1112 case ISD::XOR: return visitXOR(N);
1113 case ISD::SHL: return visitSHL(N);
1114 case ISD::SRA: return visitSRA(N);
1115 case ISD::SRL: return visitSRL(N);
1116 case ISD::CTLZ: return visitCTLZ(N);
1117 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1118 case ISD::CTTZ: return visitCTTZ(N);
1119 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1120 case ISD::CTPOP: return visitCTPOP(N);
1121 case ISD::SELECT: return visitSELECT(N);
1122 case ISD::SELECT_CC: return visitSELECT_CC(N);
1123 case ISD::SETCC: return visitSETCC(N);
1124 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1125 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1126 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1127 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1128 case ISD::TRUNCATE: return visitTRUNCATE(N);
1129 case ISD::BITCAST: return visitBITCAST(N);
1130 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1131 case ISD::FADD: return visitFADD(N);
1132 case ISD::FSUB: return visitFSUB(N);
1133 case ISD::FMUL: return visitFMUL(N);
1134 case ISD::FDIV: return visitFDIV(N);
1135 case ISD::FREM: return visitFREM(N);
1136 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1137 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1138 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1139 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1140 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1141 case ISD::FP_ROUND: return visitFP_ROUND(N);
1142 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1143 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1144 case ISD::FNEG: return visitFNEG(N);
1145 case ISD::FABS: return visitFABS(N);
1146 case ISD::BRCOND: return visitBRCOND(N);
1147 case ISD::BR_CC: return visitBR_CC(N);
1148 case ISD::LOAD: return visitLOAD(N);
1149 case ISD::STORE: return visitSTORE(N);
1150 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1151 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1152 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1153 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1154 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1155 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1156 case ISD::MEMBARRIER: return visitMEMBARRIER(N);
1161 SDValue DAGCombiner::combine(SDNode *N) {
1162 SDValue RV = visit(N);
1164 // If nothing happened, try a target-specific DAG combine.
1165 if (RV.getNode() == 0) {
1166 assert(N->getOpcode() != ISD::DELETED_NODE &&
1167 "Node was deleted but visit returned NULL!");
1169 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1170 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1172 // Expose the DAG combiner to the target combiner impls.
1173 TargetLowering::DAGCombinerInfo
1174 DagCombineInfo(DAG, !LegalTypes, !LegalOperations, false, this);
1176 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1180 // If nothing happened still, try promoting the operation.
1181 if (RV.getNode() == 0) {
1182 switch (N->getOpcode()) {
1190 RV = PromoteIntBinOp(SDValue(N, 0));
1195 RV = PromoteIntShiftOp(SDValue(N, 0));
1197 case ISD::SIGN_EXTEND:
1198 case ISD::ZERO_EXTEND:
1199 case ISD::ANY_EXTEND:
1200 RV = PromoteExtend(SDValue(N, 0));
1203 if (PromoteLoad(SDValue(N, 0)))
1209 // If N is a commutative binary node, try commuting it to enable more
1211 if (RV.getNode() == 0 &&
1212 SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1213 N->getNumValues() == 1) {
1214 SDValue N0 = N->getOperand(0);
1215 SDValue N1 = N->getOperand(1);
1217 // Constant operands are canonicalized to RHS.
1218 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1219 SDValue Ops[] = { N1, N0 };
1220 SDNode *CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(),
1223 return SDValue(CSENode, 0);
1230 /// getInputChainForNode - Given a node, return its input chain if it has one,
1231 /// otherwise return a null sd operand.
1232 static SDValue getInputChainForNode(SDNode *N) {
1233 if (unsigned NumOps = N->getNumOperands()) {
1234 if (N->getOperand(0).getValueType() == MVT::Other)
1235 return N->getOperand(0);
1236 else if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1237 return N->getOperand(NumOps-1);
1238 for (unsigned i = 1; i < NumOps-1; ++i)
1239 if (N->getOperand(i).getValueType() == MVT::Other)
1240 return N->getOperand(i);
1245 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1246 // If N has two operands, where one has an input chain equal to the other,
1247 // the 'other' chain is redundant.
1248 if (N->getNumOperands() == 2) {
1249 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1250 return N->getOperand(0);
1251 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1252 return N->getOperand(1);
1255 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1256 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1257 SmallPtrSet<SDNode*, 16> SeenOps;
1258 bool Changed = false; // If we should replace this token factor.
1260 // Start out with this token factor.
1263 // Iterate through token factors. The TFs grows when new token factors are
1265 for (unsigned i = 0; i < TFs.size(); ++i) {
1266 SDNode *TF = TFs[i];
1268 // Check each of the operands.
1269 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1270 SDValue Op = TF->getOperand(i);
1272 switch (Op.getOpcode()) {
1273 case ISD::EntryToken:
1274 // Entry tokens don't need to be added to the list. They are
1279 case ISD::TokenFactor:
1280 if (Op.hasOneUse() &&
1281 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1282 // Queue up for processing.
1283 TFs.push_back(Op.getNode());
1284 // Clean up in case the token factor is removed.
1285 AddToWorkList(Op.getNode());
1292 // Only add if it isn't already in the list.
1293 if (SeenOps.insert(Op.getNode()))
1304 // If we've change things around then replace token factor.
1307 // The entry token is the only possible outcome.
1308 Result = DAG.getEntryNode();
1310 // New and improved token factor.
1311 Result = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
1312 MVT::Other, &Ops[0], Ops.size());
1315 // Don't add users to work list.
1316 return CombineTo(N, Result, false);
1322 /// MERGE_VALUES can always be eliminated.
1323 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1324 WorkListRemover DeadNodes(*this);
1325 // Replacing results may cause a different MERGE_VALUES to suddenly
1326 // be CSE'd with N, and carry its uses with it. Iterate until no
1327 // uses remain, to ensure that the node can be safely deleted.
1329 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1330 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i),
1332 } while (!N->use_empty());
1333 removeFromWorkList(N);
1335 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1339 SDValue combineShlAddConstant(DebugLoc DL, SDValue N0, SDValue N1,
1340 SelectionDAG &DAG) {
1341 EVT VT = N0.getValueType();
1342 SDValue N00 = N0.getOperand(0);
1343 SDValue N01 = N0.getOperand(1);
1344 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N01);
1346 if (N01C && N00.getOpcode() == ISD::ADD && N00.getNode()->hasOneUse() &&
1347 isa<ConstantSDNode>(N00.getOperand(1))) {
1348 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1349 N0 = DAG.getNode(ISD::ADD, N0.getDebugLoc(), VT,
1350 DAG.getNode(ISD::SHL, N00.getDebugLoc(), VT,
1351 N00.getOperand(0), N01),
1352 DAG.getNode(ISD::SHL, N01.getDebugLoc(), VT,
1353 N00.getOperand(1), N01));
1354 return DAG.getNode(ISD::ADD, DL, VT, N0, N1);
1360 SDValue DAGCombiner::visitADD(SDNode *N) {
1361 SDValue N0 = N->getOperand(0);
1362 SDValue N1 = N->getOperand(1);
1363 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1364 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1365 EVT VT = N0.getValueType();
1368 if (VT.isVector()) {
1369 SDValue FoldedVOp = SimplifyVBinOp(N);
1370 if (FoldedVOp.getNode()) return FoldedVOp;
1373 // fold (add x, undef) -> undef
1374 if (N0.getOpcode() == ISD::UNDEF)
1376 if (N1.getOpcode() == ISD::UNDEF)
1378 // fold (add c1, c2) -> c1+c2
1380 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1381 // canonicalize constant to RHS
1383 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N1, N0);
1384 // fold (add x, 0) -> x
1385 if (N1C && N1C->isNullValue())
1387 // fold (add Sym, c) -> Sym+c
1388 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1389 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1390 GA->getOpcode() == ISD::GlobalAddress)
1391 return DAG.getGlobalAddress(GA->getGlobal(), N1C->getDebugLoc(), VT,
1393 (uint64_t)N1C->getSExtValue());
1394 // fold ((c1-A)+c2) -> (c1+c2)-A
1395 if (N1C && N0.getOpcode() == ISD::SUB)
1396 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1397 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1398 DAG.getConstant(N1C->getAPIntValue()+
1399 N0C->getAPIntValue(), VT),
1402 SDValue RADD = ReassociateOps(ISD::ADD, N->getDebugLoc(), N0, N1);
1403 if (RADD.getNode() != 0)
1405 // fold ((0-A) + B) -> B-A
1406 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1407 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1408 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1, N0.getOperand(1));
1409 // fold (A + (0-B)) -> A-B
1410 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1411 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1412 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, N1.getOperand(1));
1413 // fold (A+(B-A)) -> B
1414 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1415 return N1.getOperand(0);
1416 // fold ((B-A)+A) -> B
1417 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1418 return N0.getOperand(0);
1419 // fold (A+(B-(A+C))) to (B-C)
1420 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1421 N0 == N1.getOperand(1).getOperand(0))
1422 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1.getOperand(0),
1423 N1.getOperand(1).getOperand(1));
1424 // fold (A+(B-(C+A))) to (B-C)
1425 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1426 N0 == N1.getOperand(1).getOperand(1))
1427 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1.getOperand(0),
1428 N1.getOperand(1).getOperand(0));
1429 // fold (A+((B-A)+or-C)) to (B+or-C)
1430 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1431 N1.getOperand(0).getOpcode() == ISD::SUB &&
1432 N0 == N1.getOperand(0).getOperand(1))
1433 return DAG.getNode(N1.getOpcode(), N->getDebugLoc(), VT,
1434 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1436 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1437 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1438 SDValue N00 = N0.getOperand(0);
1439 SDValue N01 = N0.getOperand(1);
1440 SDValue N10 = N1.getOperand(0);
1441 SDValue N11 = N1.getOperand(1);
1443 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1444 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1445 DAG.getNode(ISD::ADD, N0.getDebugLoc(), VT, N00, N10),
1446 DAG.getNode(ISD::ADD, N1.getDebugLoc(), VT, N01, N11));
1449 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1450 return SDValue(N, 0);
1452 // fold (a+b) -> (a|b) iff a and b share no bits.
1453 if (VT.isInteger() && !VT.isVector()) {
1454 APInt LHSZero, LHSOne;
1455 APInt RHSZero, RHSOne;
1456 DAG.ComputeMaskedBits(N0, LHSZero, LHSOne);
1458 if (LHSZero.getBoolValue()) {
1459 DAG.ComputeMaskedBits(N1, RHSZero, RHSOne);
1461 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1462 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1463 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1464 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N1);
1468 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1469 if (N0.getOpcode() == ISD::SHL && N0.getNode()->hasOneUse()) {
1470 SDValue Result = combineShlAddConstant(N->getDebugLoc(), N0, N1, DAG);
1471 if (Result.getNode()) return Result;
1473 if (N1.getOpcode() == ISD::SHL && N1.getNode()->hasOneUse()) {
1474 SDValue Result = combineShlAddConstant(N->getDebugLoc(), N1, N0, DAG);
1475 if (Result.getNode()) return Result;
1478 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1479 if (N1.getOpcode() == ISD::SHL &&
1480 N1.getOperand(0).getOpcode() == ISD::SUB)
1481 if (ConstantSDNode *C =
1482 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1483 if (C->getAPIntValue() == 0)
1484 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0,
1485 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1486 N1.getOperand(0).getOperand(1),
1488 if (N0.getOpcode() == ISD::SHL &&
1489 N0.getOperand(0).getOpcode() == ISD::SUB)
1490 if (ConstantSDNode *C =
1491 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1492 if (C->getAPIntValue() == 0)
1493 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1,
1494 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1495 N0.getOperand(0).getOperand(1),
1498 if (N1.getOpcode() == ISD::AND) {
1499 SDValue AndOp0 = N1.getOperand(0);
1500 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1501 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1502 unsigned DestBits = VT.getScalarType().getSizeInBits();
1504 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1505 // and similar xforms where the inner op is either ~0 or 0.
1506 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1507 DebugLoc DL = N->getDebugLoc();
1508 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1512 // add (sext i1), X -> sub X, (zext i1)
1513 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1514 N0.getOperand(0).getValueType() == MVT::i1 &&
1515 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1516 DebugLoc DL = N->getDebugLoc();
1517 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1518 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1524 SDValue DAGCombiner::visitADDC(SDNode *N) {
1525 SDValue N0 = N->getOperand(0);
1526 SDValue N1 = N->getOperand(1);
1527 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1528 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1529 EVT VT = N0.getValueType();
1531 // If the flag result is dead, turn this into an ADD.
1532 if (!N->hasAnyUseOfValue(1))
1533 return CombineTo(N, DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0, N1),
1534 DAG.getNode(ISD::CARRY_FALSE,
1535 N->getDebugLoc(), MVT::Glue));
1537 // canonicalize constant to RHS.
1539 return DAG.getNode(ISD::ADDC, N->getDebugLoc(), N->getVTList(), N1, N0);
1541 // fold (addc x, 0) -> x + no carry out
1542 if (N1C && N1C->isNullValue())
1543 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1544 N->getDebugLoc(), MVT::Glue));
1546 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1547 APInt LHSZero, LHSOne;
1548 APInt RHSZero, RHSOne;
1549 DAG.ComputeMaskedBits(N0, LHSZero, LHSOne);
1551 if (LHSZero.getBoolValue()) {
1552 DAG.ComputeMaskedBits(N1, RHSZero, RHSOne);
1554 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1555 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1556 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1557 return CombineTo(N, DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N1),
1558 DAG.getNode(ISD::CARRY_FALSE,
1559 N->getDebugLoc(), MVT::Glue));
1565 SDValue DAGCombiner::visitADDE(SDNode *N) {
1566 SDValue N0 = N->getOperand(0);
1567 SDValue N1 = N->getOperand(1);
1568 SDValue CarryIn = N->getOperand(2);
1569 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1570 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1572 // canonicalize constant to RHS
1574 return DAG.getNode(ISD::ADDE, N->getDebugLoc(), N->getVTList(),
1577 // fold (adde x, y, false) -> (addc x, y)
1578 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1579 return DAG.getNode(ISD::ADDC, N->getDebugLoc(), N->getVTList(), N0, N1);
1584 // Since it may not be valid to emit a fold to zero for vector initializers
1585 // check if we can before folding.
1586 static SDValue tryFoldToZero(DebugLoc DL, const TargetLowering &TLI, EVT VT,
1587 SelectionDAG &DAG, bool LegalOperations) {
1588 if (!VT.isVector()) {
1589 return DAG.getConstant(0, VT);
1591 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) {
1592 // Produce a vector of zeros.
1593 SDValue El = DAG.getConstant(0, VT.getVectorElementType());
1594 std::vector<SDValue> Ops(VT.getVectorNumElements(), El);
1595 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT,
1596 &Ops[0], Ops.size());
1601 SDValue DAGCombiner::visitSUB(SDNode *N) {
1602 SDValue N0 = N->getOperand(0);
1603 SDValue N1 = N->getOperand(1);
1604 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1605 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1606 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? 0 :
1607 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1608 EVT VT = N0.getValueType();
1611 if (VT.isVector()) {
1612 SDValue FoldedVOp = SimplifyVBinOp(N);
1613 if (FoldedVOp.getNode()) return FoldedVOp;
1616 // fold (sub x, x) -> 0
1617 // FIXME: Refactor this and xor and other similar operations together.
1619 return tryFoldToZero(N->getDebugLoc(), TLI, VT, DAG, LegalOperations);
1620 // fold (sub c1, c2) -> c1-c2
1622 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1623 // fold (sub x, c) -> (add x, -c)
1625 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0,
1626 DAG.getConstant(-N1C->getAPIntValue(), VT));
1627 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1628 if (N0C && N0C->isAllOnesValue())
1629 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0);
1630 // fold A-(A-B) -> B
1631 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1632 return N1.getOperand(1);
1633 // fold (A+B)-A -> B
1634 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1635 return N0.getOperand(1);
1636 // fold (A+B)-B -> A
1637 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1638 return N0.getOperand(0);
1639 // fold C2-(A+C1) -> (C2-C1)-A
1640 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1641 SDValue NewC = DAG.getConstant((N0C->getAPIntValue() - N1C1->getAPIntValue()), VT);
1642 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, NewC,
1645 // fold ((A+(B+or-C))-B) -> A+or-C
1646 if (N0.getOpcode() == ISD::ADD &&
1647 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1648 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1649 N0.getOperand(1).getOperand(0) == N1)
1650 return DAG.getNode(N0.getOperand(1).getOpcode(), N->getDebugLoc(), VT,
1651 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1652 // fold ((A+(C+B))-B) -> A+C
1653 if (N0.getOpcode() == ISD::ADD &&
1654 N0.getOperand(1).getOpcode() == ISD::ADD &&
1655 N0.getOperand(1).getOperand(1) == N1)
1656 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT,
1657 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1658 // fold ((A-(B-C))-C) -> A-B
1659 if (N0.getOpcode() == ISD::SUB &&
1660 N0.getOperand(1).getOpcode() == ISD::SUB &&
1661 N0.getOperand(1).getOperand(1) == N1)
1662 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1663 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1665 // If either operand of a sub is undef, the result is undef
1666 if (N0.getOpcode() == ISD::UNDEF)
1668 if (N1.getOpcode() == ISD::UNDEF)
1671 // If the relocation model supports it, consider symbol offsets.
1672 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1673 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1674 // fold (sub Sym, c) -> Sym-c
1675 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1676 return DAG.getGlobalAddress(GA->getGlobal(), N1C->getDebugLoc(), VT,
1678 (uint64_t)N1C->getSExtValue());
1679 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1680 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1681 if (GA->getGlobal() == GB->getGlobal())
1682 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1689 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1690 SDValue N0 = N->getOperand(0);
1691 SDValue N1 = N->getOperand(1);
1692 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1693 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1694 EVT VT = N0.getValueType();
1696 // If the flag result is dead, turn this into an SUB.
1697 if (!N->hasAnyUseOfValue(1))
1698 return CombineTo(N, DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, N1),
1699 DAG.getNode(ISD::CARRY_FALSE, N->getDebugLoc(),
1702 // fold (subc x, x) -> 0 + no borrow
1704 return CombineTo(N, DAG.getConstant(0, VT),
1705 DAG.getNode(ISD::CARRY_FALSE, N->getDebugLoc(),
1708 // fold (subc x, 0) -> x + no borrow
1709 if (N1C && N1C->isNullValue())
1710 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, N->getDebugLoc(),
1713 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1714 if (N0C && N0C->isAllOnesValue())
1715 return CombineTo(N, DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0),
1716 DAG.getNode(ISD::CARRY_FALSE, N->getDebugLoc(),
1722 SDValue DAGCombiner::visitSUBE(SDNode *N) {
1723 SDValue N0 = N->getOperand(0);
1724 SDValue N1 = N->getOperand(1);
1725 SDValue CarryIn = N->getOperand(2);
1727 // fold (sube x, y, false) -> (subc x, y)
1728 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1729 return DAG.getNode(ISD::SUBC, N->getDebugLoc(), N->getVTList(), N0, N1);
1734 SDValue DAGCombiner::visitMUL(SDNode *N) {
1735 SDValue N0 = N->getOperand(0);
1736 SDValue N1 = N->getOperand(1);
1737 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1738 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1739 EVT VT = N0.getValueType();
1742 if (VT.isVector()) {
1743 SDValue FoldedVOp = SimplifyVBinOp(N);
1744 if (FoldedVOp.getNode()) return FoldedVOp;
1747 // fold (mul x, undef) -> 0
1748 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1749 return DAG.getConstant(0, VT);
1750 // fold (mul c1, c2) -> c1*c2
1752 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0C, N1C);
1753 // canonicalize constant to RHS
1755 return DAG.getNode(ISD::MUL, N->getDebugLoc(), VT, N1, N0);
1756 // fold (mul x, 0) -> 0
1757 if (N1C && N1C->isNullValue())
1759 // fold (mul x, -1) -> 0-x
1760 if (N1C && N1C->isAllOnesValue())
1761 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1762 DAG.getConstant(0, VT), N0);
1763 // fold (mul x, (1 << c)) -> x << c
1764 if (N1C && N1C->getAPIntValue().isPowerOf2())
1765 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
1766 DAG.getConstant(N1C->getAPIntValue().logBase2(),
1767 getShiftAmountTy(N0.getValueType())));
1768 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
1769 if (N1C && (-N1C->getAPIntValue()).isPowerOf2()) {
1770 unsigned Log2Val = (-N1C->getAPIntValue()).logBase2();
1771 // FIXME: If the input is something that is easily negated (e.g. a
1772 // single-use add), we should put the negate there.
1773 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1774 DAG.getConstant(0, VT),
1775 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
1776 DAG.getConstant(Log2Val,
1777 getShiftAmountTy(N0.getValueType()))));
1779 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
1780 if (N1C && N0.getOpcode() == ISD::SHL &&
1781 isa<ConstantSDNode>(N0.getOperand(1))) {
1782 SDValue C3 = DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1783 N1, N0.getOperand(1));
1784 AddToWorkList(C3.getNode());
1785 return DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1786 N0.getOperand(0), C3);
1789 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
1792 SDValue Sh(0,0), Y(0,0);
1793 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
1794 if (N0.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N0.getOperand(1)) &&
1795 N0.getNode()->hasOneUse()) {
1797 } else if (N1.getOpcode() == ISD::SHL &&
1798 isa<ConstantSDNode>(N1.getOperand(1)) &&
1799 N1.getNode()->hasOneUse()) {
1804 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1805 Sh.getOperand(0), Y);
1806 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1807 Mul, Sh.getOperand(1));
1811 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
1812 if (N1C && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
1813 isa<ConstantSDNode>(N0.getOperand(1)))
1814 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT,
1815 DAG.getNode(ISD::MUL, N0.getDebugLoc(), VT,
1816 N0.getOperand(0), N1),
1817 DAG.getNode(ISD::MUL, N1.getDebugLoc(), VT,
1818 N0.getOperand(1), N1));
1821 SDValue RMUL = ReassociateOps(ISD::MUL, N->getDebugLoc(), N0, N1);
1822 if (RMUL.getNode() != 0)
1828 SDValue DAGCombiner::visitSDIV(SDNode *N) {
1829 SDValue N0 = N->getOperand(0);
1830 SDValue N1 = N->getOperand(1);
1831 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1832 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1833 EVT VT = N->getValueType(0);
1836 if (VT.isVector()) {
1837 SDValue FoldedVOp = SimplifyVBinOp(N);
1838 if (FoldedVOp.getNode()) return FoldedVOp;
1841 // fold (sdiv c1, c2) -> c1/c2
1842 if (N0C && N1C && !N1C->isNullValue())
1843 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
1844 // fold (sdiv X, 1) -> X
1845 if (N1C && N1C->getAPIntValue() == 1LL)
1847 // fold (sdiv X, -1) -> 0-X
1848 if (N1C && N1C->isAllOnesValue())
1849 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1850 DAG.getConstant(0, VT), N0);
1851 // If we know the sign bits of both operands are zero, strength reduce to a
1852 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
1853 if (!VT.isVector()) {
1854 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
1855 return DAG.getNode(ISD::UDIV, N->getDebugLoc(), N1.getValueType(),
1858 // fold (sdiv X, pow2) -> simple ops after legalize
1859 if (N1C && !N1C->isNullValue() &&
1860 (N1C->getAPIntValue().isPowerOf2() ||
1861 (-N1C->getAPIntValue()).isPowerOf2())) {
1862 // If dividing by powers of two is cheap, then don't perform the following
1864 if (TLI.isPow2DivCheap())
1867 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
1869 // Splat the sign bit into the register
1870 SDValue SGN = DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0,
1871 DAG.getConstant(VT.getSizeInBits()-1,
1872 getShiftAmountTy(N0.getValueType())));
1873 AddToWorkList(SGN.getNode());
1875 // Add (N0 < 0) ? abs2 - 1 : 0;
1876 SDValue SRL = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, SGN,
1877 DAG.getConstant(VT.getSizeInBits() - lg2,
1878 getShiftAmountTy(SGN.getValueType())));
1879 SDValue ADD = DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0, SRL);
1880 AddToWorkList(SRL.getNode());
1881 AddToWorkList(ADD.getNode()); // Divide by pow2
1882 SDValue SRA = DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, ADD,
1883 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType())));
1885 // If we're dividing by a positive value, we're done. Otherwise, we must
1886 // negate the result.
1887 if (N1C->getAPIntValue().isNonNegative())
1890 AddToWorkList(SRA.getNode());
1891 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1892 DAG.getConstant(0, VT), SRA);
1895 // if integer divide is expensive and we satisfy the requirements, emit an
1896 // alternate sequence.
1897 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
1898 SDValue Op = BuildSDIV(N);
1899 if (Op.getNode()) return Op;
1903 if (N0.getOpcode() == ISD::UNDEF)
1904 return DAG.getConstant(0, VT);
1905 // X / undef -> undef
1906 if (N1.getOpcode() == ISD::UNDEF)
1912 SDValue DAGCombiner::visitUDIV(SDNode *N) {
1913 SDValue N0 = N->getOperand(0);
1914 SDValue N1 = N->getOperand(1);
1915 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1916 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1917 EVT VT = N->getValueType(0);
1920 if (VT.isVector()) {
1921 SDValue FoldedVOp = SimplifyVBinOp(N);
1922 if (FoldedVOp.getNode()) return FoldedVOp;
1925 // fold (udiv c1, c2) -> c1/c2
1926 if (N0C && N1C && !N1C->isNullValue())
1927 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
1928 // fold (udiv x, (1 << c)) -> x >>u c
1929 if (N1C && N1C->getAPIntValue().isPowerOf2())
1930 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0,
1931 DAG.getConstant(N1C->getAPIntValue().logBase2(),
1932 getShiftAmountTy(N0.getValueType())));
1933 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
1934 if (N1.getOpcode() == ISD::SHL) {
1935 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
1936 if (SHC->getAPIntValue().isPowerOf2()) {
1937 EVT ADDVT = N1.getOperand(1).getValueType();
1938 SDValue Add = DAG.getNode(ISD::ADD, N->getDebugLoc(), ADDVT,
1940 DAG.getConstant(SHC->getAPIntValue()
1943 AddToWorkList(Add.getNode());
1944 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0, Add);
1948 // fold (udiv x, c) -> alternate
1949 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
1950 SDValue Op = BuildUDIV(N);
1951 if (Op.getNode()) return Op;
1955 if (N0.getOpcode() == ISD::UNDEF)
1956 return DAG.getConstant(0, VT);
1957 // X / undef -> undef
1958 if (N1.getOpcode() == ISD::UNDEF)
1964 SDValue DAGCombiner::visitSREM(SDNode *N) {
1965 SDValue N0 = N->getOperand(0);
1966 SDValue N1 = N->getOperand(1);
1967 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1968 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1969 EVT VT = N->getValueType(0);
1971 // fold (srem c1, c2) -> c1%c2
1972 if (N0C && N1C && !N1C->isNullValue())
1973 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
1974 // If we know the sign bits of both operands are zero, strength reduce to a
1975 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
1976 if (!VT.isVector()) {
1977 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
1978 return DAG.getNode(ISD::UREM, N->getDebugLoc(), VT, N0, N1);
1981 // If X/C can be simplified by the division-by-constant logic, lower
1982 // X%C to the equivalent of X-X/C*C.
1983 if (N1C && !N1C->isNullValue()) {
1984 SDValue Div = DAG.getNode(ISD::SDIV, N->getDebugLoc(), VT, N0, N1);
1985 AddToWorkList(Div.getNode());
1986 SDValue OptimizedDiv = combine(Div.getNode());
1987 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
1988 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1990 SDValue Sub = DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, Mul);
1991 AddToWorkList(Mul.getNode());
1997 if (N0.getOpcode() == ISD::UNDEF)
1998 return DAG.getConstant(0, VT);
1999 // X % undef -> undef
2000 if (N1.getOpcode() == ISD::UNDEF)
2006 SDValue DAGCombiner::visitUREM(SDNode *N) {
2007 SDValue N0 = N->getOperand(0);
2008 SDValue N1 = N->getOperand(1);
2009 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2010 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2011 EVT VT = N->getValueType(0);
2013 // fold (urem c1, c2) -> c1%c2
2014 if (N0C && N1C && !N1C->isNullValue())
2015 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
2016 // fold (urem x, pow2) -> (and x, pow2-1)
2017 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
2018 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0,
2019 DAG.getConstant(N1C->getAPIntValue()-1,VT));
2020 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2021 if (N1.getOpcode() == ISD::SHL) {
2022 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2023 if (SHC->getAPIntValue().isPowerOf2()) {
2025 DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N1,
2026 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
2028 AddToWorkList(Add.getNode());
2029 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, Add);
2034 // If X/C can be simplified by the division-by-constant logic, lower
2035 // X%C to the equivalent of X-X/C*C.
2036 if (N1C && !N1C->isNullValue()) {
2037 SDValue Div = DAG.getNode(ISD::UDIV, N->getDebugLoc(), VT, N0, N1);
2038 AddToWorkList(Div.getNode());
2039 SDValue OptimizedDiv = combine(Div.getNode());
2040 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2041 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
2043 SDValue Sub = DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, Mul);
2044 AddToWorkList(Mul.getNode());
2050 if (N0.getOpcode() == ISD::UNDEF)
2051 return DAG.getConstant(0, VT);
2052 // X % undef -> undef
2053 if (N1.getOpcode() == ISD::UNDEF)
2059 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2060 SDValue N0 = N->getOperand(0);
2061 SDValue N1 = N->getOperand(1);
2062 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2063 EVT VT = N->getValueType(0);
2064 DebugLoc DL = N->getDebugLoc();
2066 // fold (mulhs x, 0) -> 0
2067 if (N1C && N1C->isNullValue())
2069 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2070 if (N1C && N1C->getAPIntValue() == 1)
2071 return DAG.getNode(ISD::SRA, N->getDebugLoc(), N0.getValueType(), N0,
2072 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2073 getShiftAmountTy(N0.getValueType())));
2074 // fold (mulhs x, undef) -> 0
2075 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2076 return DAG.getConstant(0, VT);
2078 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2080 if (VT.isSimple() && !VT.isVector()) {
2081 MVT Simple = VT.getSimpleVT();
2082 unsigned SimpleSize = Simple.getSizeInBits();
2083 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2084 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2085 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2086 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2087 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2088 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2089 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2090 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2097 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2098 SDValue N0 = N->getOperand(0);
2099 SDValue N1 = N->getOperand(1);
2100 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2101 EVT VT = N->getValueType(0);
2102 DebugLoc DL = N->getDebugLoc();
2104 // fold (mulhu x, 0) -> 0
2105 if (N1C && N1C->isNullValue())
2107 // fold (mulhu x, 1) -> 0
2108 if (N1C && N1C->getAPIntValue() == 1)
2109 return DAG.getConstant(0, N0.getValueType());
2110 // fold (mulhu x, undef) -> 0
2111 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2112 return DAG.getConstant(0, VT);
2114 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2116 if (VT.isSimple() && !VT.isVector()) {
2117 MVT Simple = VT.getSimpleVT();
2118 unsigned SimpleSize = Simple.getSizeInBits();
2119 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2120 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2121 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2122 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2123 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2124 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2125 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2126 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2133 /// SimplifyNodeWithTwoResults - Perform optimizations common to nodes that
2134 /// compute two values. LoOp and HiOp give the opcodes for the two computations
2135 /// that are being performed. Return true if a simplification was made.
2137 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2139 // If the high half is not needed, just compute the low half.
2140 bool HiExists = N->hasAnyUseOfValue(1);
2142 (!LegalOperations ||
2143 TLI.isOperationLegal(LoOp, N->getValueType(0)))) {
2144 SDValue Res = DAG.getNode(LoOp, N->getDebugLoc(), N->getValueType(0),
2145 N->op_begin(), N->getNumOperands());
2146 return CombineTo(N, Res, Res);
2149 // If the low half is not needed, just compute the high half.
2150 bool LoExists = N->hasAnyUseOfValue(0);
2152 (!LegalOperations ||
2153 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2154 SDValue Res = DAG.getNode(HiOp, N->getDebugLoc(), N->getValueType(1),
2155 N->op_begin(), N->getNumOperands());
2156 return CombineTo(N, Res, Res);
2159 // If both halves are used, return as it is.
2160 if (LoExists && HiExists)
2163 // If the two computed results can be simplified separately, separate them.
2165 SDValue Lo = DAG.getNode(LoOp, N->getDebugLoc(), N->getValueType(0),
2166 N->op_begin(), N->getNumOperands());
2167 AddToWorkList(Lo.getNode());
2168 SDValue LoOpt = combine(Lo.getNode());
2169 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2170 (!LegalOperations ||
2171 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2172 return CombineTo(N, LoOpt, LoOpt);
2176 SDValue Hi = DAG.getNode(HiOp, N->getDebugLoc(), N->getValueType(1),
2177 N->op_begin(), N->getNumOperands());
2178 AddToWorkList(Hi.getNode());
2179 SDValue HiOpt = combine(Hi.getNode());
2180 if (HiOpt.getNode() && HiOpt != Hi &&
2181 (!LegalOperations ||
2182 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2183 return CombineTo(N, HiOpt, HiOpt);
2189 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2190 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2191 if (Res.getNode()) return Res;
2193 EVT VT = N->getValueType(0);
2194 DebugLoc DL = N->getDebugLoc();
2196 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2198 if (VT.isSimple() && !VT.isVector()) {
2199 MVT Simple = VT.getSimpleVT();
2200 unsigned SimpleSize = Simple.getSizeInBits();
2201 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2202 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2203 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2204 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2205 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2206 // Compute the high part as N1.
2207 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2208 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2209 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2210 // Compute the low part as N0.
2211 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2212 return CombineTo(N, Lo, Hi);
2219 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2220 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2221 if (Res.getNode()) return Res;
2223 EVT VT = N->getValueType(0);
2224 DebugLoc DL = N->getDebugLoc();
2226 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2228 if (VT.isSimple() && !VT.isVector()) {
2229 MVT Simple = VT.getSimpleVT();
2230 unsigned SimpleSize = Simple.getSizeInBits();
2231 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2232 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2233 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2234 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2235 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2236 // Compute the high part as N1.
2237 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2238 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2239 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2240 // Compute the low part as N0.
2241 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2242 return CombineTo(N, Lo, Hi);
2249 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2250 // (smulo x, 2) -> (saddo x, x)
2251 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2252 if (C2->getAPIntValue() == 2)
2253 return DAG.getNode(ISD::SADDO, N->getDebugLoc(), N->getVTList(),
2254 N->getOperand(0), N->getOperand(0));
2259 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2260 // (umulo x, 2) -> (uaddo x, x)
2261 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2262 if (C2->getAPIntValue() == 2)
2263 return DAG.getNode(ISD::UADDO, N->getDebugLoc(), N->getVTList(),
2264 N->getOperand(0), N->getOperand(0));
2269 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2270 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2271 if (Res.getNode()) return Res;
2276 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2277 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2278 if (Res.getNode()) return Res;
2283 /// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with
2284 /// two operands of the same opcode, try to simplify it.
2285 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2286 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2287 EVT VT = N0.getValueType();
2288 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2290 // Bail early if none of these transforms apply.
2291 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2293 // For each of OP in AND/OR/XOR:
2294 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2295 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2296 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2297 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2299 // do not sink logical op inside of a vector extend, since it may combine
2301 EVT Op0VT = N0.getOperand(0).getValueType();
2302 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2303 N0.getOpcode() == ISD::SIGN_EXTEND ||
2304 // Avoid infinite looping with PromoteIntBinOp.
2305 (N0.getOpcode() == ISD::ANY_EXTEND &&
2306 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2307 (N0.getOpcode() == ISD::TRUNCATE &&
2308 (!TLI.isZExtFree(VT, Op0VT) ||
2309 !TLI.isTruncateFree(Op0VT, VT)) &&
2310 TLI.isTypeLegal(Op0VT))) &&
2312 Op0VT == N1.getOperand(0).getValueType() &&
2313 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2314 SDValue ORNode = DAG.getNode(N->getOpcode(), N0.getDebugLoc(),
2315 N0.getOperand(0).getValueType(),
2316 N0.getOperand(0), N1.getOperand(0));
2317 AddToWorkList(ORNode.getNode());
2318 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, ORNode);
2321 // For each of OP in SHL/SRL/SRA/AND...
2322 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2323 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2324 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2325 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2326 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2327 N0.getOperand(1) == N1.getOperand(1)) {
2328 SDValue ORNode = DAG.getNode(N->getOpcode(), N0.getDebugLoc(),
2329 N0.getOperand(0).getValueType(),
2330 N0.getOperand(0), N1.getOperand(0));
2331 AddToWorkList(ORNode.getNode());
2332 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
2333 ORNode, N0.getOperand(1));
2336 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2337 // Only perform this optimization after type legalization and before
2338 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2339 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2340 // we don't want to undo this promotion.
2341 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2343 if ((N0.getOpcode() == ISD::BITCAST || N0.getOpcode() == ISD::SCALAR_TO_VECTOR)
2344 && Level == AfterLegalizeVectorOps) {
2345 SDValue In0 = N0.getOperand(0);
2346 SDValue In1 = N1.getOperand(0);
2347 EVT In0Ty = In0.getValueType();
2348 EVT In1Ty = In1.getValueType();
2349 // If both incoming values are integers, and the original types are the same.
2350 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2351 SDValue Op = DAG.getNode(N->getOpcode(), N->getDebugLoc(), In0Ty, In0, In1);
2352 SDValue BC = DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, Op);
2353 AddToWorkList(Op.getNode());
2358 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2359 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2360 // If both shuffles use the same mask, and both shuffle within a single
2361 // vector, then it is worthwhile to move the swizzle after the operation.
2362 // The type-legalizer generates this pattern when loading illegal
2363 // vector types from memory. In many cases this allows additional shuffle
2365 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
2366 N0.getOperand(1).getOpcode() == ISD::UNDEF &&
2367 N1.getOperand(1).getOpcode() == ISD::UNDEF) {
2368 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2369 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2371 assert(N0.getOperand(0).getValueType() == N1.getOperand(1).getValueType() &&
2372 "Inputs to shuffles are not the same type");
2374 unsigned NumElts = VT.getVectorNumElements();
2376 // Check that both shuffles use the same mask. The masks are known to be of
2377 // the same length because the result vector type is the same.
2378 bool SameMask = true;
2379 for (unsigned i = 0; i != NumElts; ++i) {
2380 int Idx0 = SVN0->getMaskElt(i);
2381 int Idx1 = SVN1->getMaskElt(i);
2389 SDValue Op = DAG.getNode(N->getOpcode(), N->getDebugLoc(), VT,
2390 N0.getOperand(0), N1.getOperand(0));
2391 AddToWorkList(Op.getNode());
2392 return DAG.getVectorShuffle(VT, N->getDebugLoc(), Op,
2393 DAG.getUNDEF(VT), &SVN0->getMask()[0]);
2400 SDValue DAGCombiner::visitAND(SDNode *N) {
2401 SDValue N0 = N->getOperand(0);
2402 SDValue N1 = N->getOperand(1);
2403 SDValue LL, LR, RL, RR, CC0, CC1;
2404 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2405 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2406 EVT VT = N1.getValueType();
2407 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2410 if (VT.isVector()) {
2411 SDValue FoldedVOp = SimplifyVBinOp(N);
2412 if (FoldedVOp.getNode()) return FoldedVOp;
2415 // fold (and x, undef) -> 0
2416 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2417 return DAG.getConstant(0, VT);
2418 // fold (and c1, c2) -> c1&c2
2420 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2421 // canonicalize constant to RHS
2423 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N1, N0);
2424 // fold (and x, -1) -> x
2425 if (N1C && N1C->isAllOnesValue())
2427 // if (and x, c) is known to be zero, return 0
2428 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2429 APInt::getAllOnesValue(BitWidth)))
2430 return DAG.getConstant(0, VT);
2432 SDValue RAND = ReassociateOps(ISD::AND, N->getDebugLoc(), N0, N1);
2433 if (RAND.getNode() != 0)
2435 // fold (and (or x, C), D) -> D if (C & D) == D
2436 if (N1C && N0.getOpcode() == ISD::OR)
2437 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2438 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2440 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2441 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2442 SDValue N0Op0 = N0.getOperand(0);
2443 APInt Mask = ~N1C->getAPIntValue();
2444 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2445 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2446 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(),
2447 N0.getValueType(), N0Op0);
2449 // Replace uses of the AND with uses of the Zero extend node.
2452 // We actually want to replace all uses of the any_extend with the
2453 // zero_extend, to avoid duplicating things. This will later cause this
2454 // AND to be folded.
2455 CombineTo(N0.getNode(), Zext);
2456 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2459 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2460 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2461 // already be zero by virtue of the width of the base type of the load.
2463 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2465 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2466 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2467 N0.getOpcode() == ISD::LOAD) {
2468 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2469 N0 : N0.getOperand(0) );
2471 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2472 // This can be a pure constant or a vector splat, in which case we treat the
2473 // vector as a scalar and use the splat value.
2474 APInt Constant = APInt::getNullValue(1);
2475 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2476 Constant = C->getAPIntValue();
2477 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2478 APInt SplatValue, SplatUndef;
2479 unsigned SplatBitSize;
2481 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2482 SplatBitSize, HasAnyUndefs);
2484 // Undef bits can contribute to a possible optimisation if set, so
2486 SplatValue |= SplatUndef;
2488 // The splat value may be something like "0x00FFFFFF", which means 0 for
2489 // the first vector value and FF for the rest, repeating. We need a mask
2490 // that will apply equally to all members of the vector, so AND all the
2491 // lanes of the constant together.
2492 EVT VT = Vector->getValueType(0);
2493 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2494 Constant = APInt::getAllOnesValue(BitWidth);
2495 for (unsigned i = 0, n = VT.getVectorNumElements(); i < n; ++i)
2496 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
2500 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
2501 // actually legal and isn't going to get expanded, else this is a false
2503 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
2504 Load->getMemoryVT());
2506 // Resize the constant to the same size as the original memory access before
2507 // extension. If it is still the AllOnesValue then this AND is completely
2510 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
2513 switch (Load->getExtensionType()) {
2514 default: B = false; break;
2515 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
2517 case ISD::NON_EXTLOAD: B = true; break;
2520 if (B && Constant.isAllOnesValue()) {
2521 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
2522 // preserve semantics once we get rid of the AND.
2523 SDValue NewLoad(Load, 0);
2524 if (Load->getExtensionType() == ISD::EXTLOAD) {
2525 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
2526 Load->getValueType(0), Load->getDebugLoc(),
2527 Load->getChain(), Load->getBasePtr(),
2528 Load->getOffset(), Load->getMemoryVT(),
2529 Load->getMemOperand());
2530 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
2531 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
2534 // Fold the AND away, taking care not to fold to the old load node if we
2536 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
2538 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2541 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2542 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2543 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2544 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2546 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2547 LL.getValueType().isInteger()) {
2548 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2549 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
2550 SDValue ORNode = DAG.getNode(ISD::OR, N0.getDebugLoc(),
2551 LR.getValueType(), LL, RL);
2552 AddToWorkList(ORNode.getNode());
2553 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
2555 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2556 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2557 SDValue ANDNode = DAG.getNode(ISD::AND, N0.getDebugLoc(),
2558 LR.getValueType(), LL, RL);
2559 AddToWorkList(ANDNode.getNode());
2560 return DAG.getSetCC(N->getDebugLoc(), VT, ANDNode, LR, Op1);
2562 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2563 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2564 SDValue ORNode = DAG.getNode(ISD::OR, N0.getDebugLoc(),
2565 LR.getValueType(), LL, RL);
2566 AddToWorkList(ORNode.getNode());
2567 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
2570 // canonicalize equivalent to ll == rl
2571 if (LL == RR && LR == RL) {
2572 Op1 = ISD::getSetCCSwappedOperands(Op1);
2575 if (LL == RL && LR == RR) {
2576 bool isInteger = LL.getValueType().isInteger();
2577 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2578 if (Result != ISD::SETCC_INVALID &&
2579 (!LegalOperations || TLI.isCondCodeLegal(Result, LL.getValueType())))
2580 return DAG.getSetCC(N->getDebugLoc(), N0.getValueType(),
2585 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
2586 if (N0.getOpcode() == N1.getOpcode()) {
2587 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2588 if (Tmp.getNode()) return Tmp;
2591 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
2592 // fold (and (sra)) -> (and (srl)) when possible.
2593 if (!VT.isVector() &&
2594 SimplifyDemandedBits(SDValue(N, 0)))
2595 return SDValue(N, 0);
2597 // fold (zext_inreg (extload x)) -> (zextload x)
2598 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
2599 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2600 EVT MemVT = LN0->getMemoryVT();
2601 // If we zero all the possible extended bits, then we can turn this into
2602 // a zextload if we are running before legalize or the operation is legal.
2603 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2604 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2605 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2606 ((!LegalOperations && !LN0->isVolatile()) ||
2607 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2608 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
2609 LN0->getChain(), LN0->getBasePtr(),
2610 LN0->getPointerInfo(), MemVT,
2611 LN0->isVolatile(), LN0->isNonTemporal(),
2612 LN0->getAlignment());
2614 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2615 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2618 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
2619 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
2621 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2622 EVT MemVT = LN0->getMemoryVT();
2623 // If we zero all the possible extended bits, then we can turn this into
2624 // a zextload if we are running before legalize or the operation is legal.
2625 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2626 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2627 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2628 ((!LegalOperations && !LN0->isVolatile()) ||
2629 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2630 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
2632 LN0->getBasePtr(), LN0->getPointerInfo(),
2634 LN0->isVolatile(), LN0->isNonTemporal(),
2635 LN0->getAlignment());
2637 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2638 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2642 // fold (and (load x), 255) -> (zextload x, i8)
2643 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2644 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2645 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2646 (N0.getOpcode() == ISD::ANY_EXTEND &&
2647 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2648 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2649 LoadSDNode *LN0 = HasAnyExt
2650 ? cast<LoadSDNode>(N0.getOperand(0))
2651 : cast<LoadSDNode>(N0);
2652 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2653 LN0->isUnindexed() && N0.hasOneUse() && LN0->hasOneUse()) {
2654 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2655 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2656 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2657 EVT LoadedVT = LN0->getMemoryVT();
2659 if (ExtVT == LoadedVT &&
2660 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2661 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2664 DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy,
2665 LN0->getChain(), LN0->getBasePtr(),
2666 LN0->getPointerInfo(),
2667 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2668 LN0->getAlignment());
2670 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
2671 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2674 // Do not change the width of a volatile load.
2675 // Do not generate loads of non-round integer types since these can
2676 // be expensive (and would be wrong if the type is not byte sized).
2677 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
2678 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2679 EVT PtrType = LN0->getOperand(1).getValueType();
2681 unsigned Alignment = LN0->getAlignment();
2682 SDValue NewPtr = LN0->getBasePtr();
2684 // For big endian targets, we need to add an offset to the pointer
2685 // to load the correct bytes. For little endian systems, we merely
2686 // need to read fewer bytes from the same pointer.
2687 if (TLI.isBigEndian()) {
2688 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
2689 unsigned EVTStoreBytes = ExtVT.getStoreSize();
2690 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
2691 NewPtr = DAG.getNode(ISD::ADD, LN0->getDebugLoc(), PtrType,
2692 NewPtr, DAG.getConstant(PtrOff, PtrType));
2693 Alignment = MinAlign(Alignment, PtrOff);
2696 AddToWorkList(NewPtr.getNode());
2698 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2700 DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy,
2701 LN0->getChain(), NewPtr,
2702 LN0->getPointerInfo(),
2703 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2706 CombineTo(LN0, Load, Load.getValue(1));
2707 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2716 /// MatchBSwapHWord - Match (a >> 8) | (a << 8) as (bswap a) >> 16
2718 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
2719 bool DemandHighBits) {
2720 if (!LegalOperations)
2723 EVT VT = N->getValueType(0);
2724 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
2726 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
2729 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
2730 bool LookPassAnd0 = false;
2731 bool LookPassAnd1 = false;
2732 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
2734 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
2736 if (N0.getOpcode() == ISD::AND) {
2737 if (!N0.getNode()->hasOneUse())
2739 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2740 if (!N01C || N01C->getZExtValue() != 0xFF00)
2742 N0 = N0.getOperand(0);
2743 LookPassAnd0 = true;
2746 if (N1.getOpcode() == ISD::AND) {
2747 if (!N1.getNode()->hasOneUse())
2749 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
2750 if (!N11C || N11C->getZExtValue() != 0xFF)
2752 N1 = N1.getOperand(0);
2753 LookPassAnd1 = true;
2756 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
2758 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
2760 if (!N0.getNode()->hasOneUse() ||
2761 !N1.getNode()->hasOneUse())
2764 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2765 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
2768 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
2771 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
2772 SDValue N00 = N0->getOperand(0);
2773 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
2774 if (!N00.getNode()->hasOneUse())
2776 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
2777 if (!N001C || N001C->getZExtValue() != 0xFF)
2779 N00 = N00.getOperand(0);
2780 LookPassAnd0 = true;
2783 SDValue N10 = N1->getOperand(0);
2784 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
2785 if (!N10.getNode()->hasOneUse())
2787 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
2788 if (!N101C || N101C->getZExtValue() != 0xFF00)
2790 N10 = N10.getOperand(0);
2791 LookPassAnd1 = true;
2797 // Make sure everything beyond the low halfword is zero since the SRL 16
2798 // will clear the top bits.
2799 unsigned OpSizeInBits = VT.getSizeInBits();
2800 if (DemandHighBits && OpSizeInBits > 16 &&
2801 (!LookPassAnd0 || !LookPassAnd1) &&
2802 !DAG.MaskedValueIsZero(N10, APInt::getHighBitsSet(OpSizeInBits, 16)))
2805 SDValue Res = DAG.getNode(ISD::BSWAP, N->getDebugLoc(), VT, N00);
2806 if (OpSizeInBits > 16)
2807 Res = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, Res,
2808 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
2812 /// isBSwapHWordElement - Return true if the specified node is an element
2813 /// that makes up a 32-bit packed halfword byteswap. i.e.
2814 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
2815 static bool isBSwapHWordElement(SDValue N, SmallVector<SDNode*,4> &Parts) {
2816 if (!N.getNode()->hasOneUse())
2819 unsigned Opc = N.getOpcode();
2820 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
2823 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2828 switch (N1C->getZExtValue()) {
2831 case 0xFF: Num = 0; break;
2832 case 0xFF00: Num = 1; break;
2833 case 0xFF0000: Num = 2; break;
2834 case 0xFF000000: Num = 3; break;
2837 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
2838 SDValue N0 = N.getOperand(0);
2839 if (Opc == ISD::AND) {
2840 if (Num == 0 || Num == 2) {
2842 // (x >> 8) & 0xff0000
2843 if (N0.getOpcode() != ISD::SRL)
2845 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2846 if (!C || C->getZExtValue() != 8)
2849 // (x << 8) & 0xff00
2850 // (x << 8) & 0xff000000
2851 if (N0.getOpcode() != ISD::SHL)
2853 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2854 if (!C || C->getZExtValue() != 8)
2857 } else if (Opc == ISD::SHL) {
2859 // (x & 0xff0000) << 8
2860 if (Num != 0 && Num != 2)
2862 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2863 if (!C || C->getZExtValue() != 8)
2865 } else { // Opc == ISD::SRL
2866 // (x & 0xff00) >> 8
2867 // (x & 0xff000000) >> 8
2868 if (Num != 1 && Num != 3)
2870 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2871 if (!C || C->getZExtValue() != 8)
2878 Parts[Num] = N0.getOperand(0).getNode();
2882 /// MatchBSwapHWord - Match a 32-bit packed halfword bswap. That is
2883 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
2884 /// => (rotl (bswap x), 16)
2885 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
2886 if (!LegalOperations)
2889 EVT VT = N->getValueType(0);
2892 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
2895 SmallVector<SDNode*,4> Parts(4, (SDNode*)0);
2897 // (or (or (and), (and)), (or (and), (and)))
2898 // (or (or (or (and), (and)), (and)), (and))
2899 if (N0.getOpcode() != ISD::OR)
2901 SDValue N00 = N0.getOperand(0);
2902 SDValue N01 = N0.getOperand(1);
2904 if (N1.getOpcode() == ISD::OR) {
2905 // (or (or (and), (and)), (or (and), (and)))
2906 SDValue N000 = N00.getOperand(0);
2907 if (!isBSwapHWordElement(N000, Parts))
2910 SDValue N001 = N00.getOperand(1);
2911 if (!isBSwapHWordElement(N001, Parts))
2913 SDValue N010 = N01.getOperand(0);
2914 if (!isBSwapHWordElement(N010, Parts))
2916 SDValue N011 = N01.getOperand(1);
2917 if (!isBSwapHWordElement(N011, Parts))
2920 // (or (or (or (and), (and)), (and)), (and))
2921 if (!isBSwapHWordElement(N1, Parts))
2923 if (!isBSwapHWordElement(N01, Parts))
2925 if (N00.getOpcode() != ISD::OR)
2927 SDValue N000 = N00.getOperand(0);
2928 if (!isBSwapHWordElement(N000, Parts))
2930 SDValue N001 = N00.getOperand(1);
2931 if (!isBSwapHWordElement(N001, Parts))
2935 // Make sure the parts are all coming from the same node.
2936 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
2939 SDValue BSwap = DAG.getNode(ISD::BSWAP, N->getDebugLoc(), VT,
2940 SDValue(Parts[0],0));
2942 // Result of the bswap should be rotated by 16. If it's not legal, than
2943 // do (x << 16) | (x >> 16).
2944 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
2945 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
2946 return DAG.getNode(ISD::ROTL, N->getDebugLoc(), VT, BSwap, ShAmt);
2947 else if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
2948 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, BSwap, ShAmt);
2949 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT,
2950 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, BSwap, ShAmt),
2951 DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, BSwap, ShAmt));
2954 SDValue DAGCombiner::visitOR(SDNode *N) {
2955 SDValue N0 = N->getOperand(0);
2956 SDValue N1 = N->getOperand(1);
2957 SDValue LL, LR, RL, RR, CC0, CC1;
2958 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2959 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2960 EVT VT = N1.getValueType();
2963 if (VT.isVector()) {
2964 SDValue FoldedVOp = SimplifyVBinOp(N);
2965 if (FoldedVOp.getNode()) return FoldedVOp;
2968 // fold (or x, undef) -> -1
2969 if (!LegalOperations &&
2970 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
2971 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
2972 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
2974 // fold (or c1, c2) -> c1|c2
2976 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
2977 // canonicalize constant to RHS
2979 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N1, N0);
2980 // fold (or x, 0) -> x
2981 if (N1C && N1C->isNullValue())
2983 // fold (or x, -1) -> -1
2984 if (N1C && N1C->isAllOnesValue())
2986 // fold (or x, c) -> c iff (x & ~c) == 0
2987 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
2990 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
2991 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
2992 if (BSwap.getNode() != 0)
2994 BSwap = MatchBSwapHWordLow(N, N0, N1);
2995 if (BSwap.getNode() != 0)
2999 SDValue ROR = ReassociateOps(ISD::OR, N->getDebugLoc(), N0, N1);
3000 if (ROR.getNode() != 0)
3002 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3003 // iff (c1 & c2) == 0.
3004 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3005 isa<ConstantSDNode>(N0.getOperand(1))) {
3006 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3007 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0)
3008 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
3009 DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
3010 N0.getOperand(0), N1),
3011 DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1));
3013 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3014 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3015 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3016 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3018 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3019 LL.getValueType().isInteger()) {
3020 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3021 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3022 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3023 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3024 SDValue ORNode = DAG.getNode(ISD::OR, LR.getDebugLoc(),
3025 LR.getValueType(), LL, RL);
3026 AddToWorkList(ORNode.getNode());
3027 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
3029 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3030 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3031 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3032 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3033 SDValue ANDNode = DAG.getNode(ISD::AND, LR.getDebugLoc(),
3034 LR.getValueType(), LL, RL);
3035 AddToWorkList(ANDNode.getNode());
3036 return DAG.getSetCC(N->getDebugLoc(), VT, ANDNode, LR, Op1);
3039 // canonicalize equivalent to ll == rl
3040 if (LL == RR && LR == RL) {
3041 Op1 = ISD::getSetCCSwappedOperands(Op1);
3044 if (LL == RL && LR == RR) {
3045 bool isInteger = LL.getValueType().isInteger();
3046 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3047 if (Result != ISD::SETCC_INVALID &&
3048 (!LegalOperations || TLI.isCondCodeLegal(Result, LL.getValueType())))
3049 return DAG.getSetCC(N->getDebugLoc(), N0.getValueType(),
3054 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3055 if (N0.getOpcode() == N1.getOpcode()) {
3056 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3057 if (Tmp.getNode()) return Tmp;
3060 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3061 if (N0.getOpcode() == ISD::AND &&
3062 N1.getOpcode() == ISD::AND &&
3063 N0.getOperand(1).getOpcode() == ISD::Constant &&
3064 N1.getOperand(1).getOpcode() == ISD::Constant &&
3065 // Don't increase # computations.
3066 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3067 // We can only do this xform if we know that bits from X that are set in C2
3068 // but not in C1 are already zero. Likewise for Y.
3069 const APInt &LHSMask =
3070 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3071 const APInt &RHSMask =
3072 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3074 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3075 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3076 SDValue X = DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
3077 N0.getOperand(0), N1.getOperand(0));
3078 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, X,
3079 DAG.getConstant(LHSMask | RHSMask, VT));
3083 // See if this is some rotate idiom.
3084 if (SDNode *Rot = MatchRotate(N0, N1, N->getDebugLoc()))
3085 return SDValue(Rot, 0);
3087 // Simplify the operands using demanded-bits information.
3088 if (!VT.isVector() &&
3089 SimplifyDemandedBits(SDValue(N, 0)))
3090 return SDValue(N, 0);
3095 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
3096 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3097 if (Op.getOpcode() == ISD::AND) {
3098 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3099 Mask = Op.getOperand(1);
3100 Op = Op.getOperand(0);
3106 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3114 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3115 // idioms for rotate, and if the target supports rotation instructions, generate
3117 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, DebugLoc DL) {
3118 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3119 EVT VT = LHS.getValueType();
3120 if (!TLI.isTypeLegal(VT)) return 0;
3122 // The target must have at least one rotate flavor.
3123 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3124 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3125 if (!HasROTL && !HasROTR) return 0;
3127 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3128 SDValue LHSShift; // The shift.
3129 SDValue LHSMask; // AND value if any.
3130 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3131 return 0; // Not part of a rotate.
3133 SDValue RHSShift; // The shift.
3134 SDValue RHSMask; // AND value if any.
3135 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3136 return 0; // Not part of a rotate.
3138 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3139 return 0; // Not shifting the same value.
3141 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3142 return 0; // Shifts must disagree.
3144 // Canonicalize shl to left side in a shl/srl pair.
3145 if (RHSShift.getOpcode() == ISD::SHL) {
3146 std::swap(LHS, RHS);
3147 std::swap(LHSShift, RHSShift);
3148 std::swap(LHSMask , RHSMask );
3151 unsigned OpSizeInBits = VT.getSizeInBits();
3152 SDValue LHSShiftArg = LHSShift.getOperand(0);
3153 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3154 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3156 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3157 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3158 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3159 RHSShiftAmt.getOpcode() == ISD::Constant) {
3160 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3161 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3162 if ((LShVal + RShVal) != OpSizeInBits)
3167 Rot = DAG.getNode(ISD::ROTL, DL, VT, LHSShiftArg, LHSShiftAmt);
3169 Rot = DAG.getNode(ISD::ROTR, DL, VT, LHSShiftArg, RHSShiftAmt);
3171 // If there is an AND of either shifted operand, apply it to the result.
3172 if (LHSMask.getNode() || RHSMask.getNode()) {
3173 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3175 if (LHSMask.getNode()) {
3176 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3177 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3179 if (RHSMask.getNode()) {
3180 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3181 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3184 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3187 return Rot.getNode();
3190 // If there is a mask here, and we have a variable shift, we can't be sure
3191 // that we're masking out the right stuff.
3192 if (LHSMask.getNode() || RHSMask.getNode())
3195 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y)
3196 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotr x, (sub 32, y))
3197 if (RHSShiftAmt.getOpcode() == ISD::SUB &&
3198 LHSShiftAmt == RHSShiftAmt.getOperand(1)) {
3199 if (ConstantSDNode *SUBC =
3200 dyn_cast<ConstantSDNode>(RHSShiftAmt.getOperand(0))) {
3201 if (SUBC->getAPIntValue() == OpSizeInBits) {
3203 return DAG.getNode(ISD::ROTL, DL, VT,
3204 LHSShiftArg, LHSShiftAmt).getNode();
3206 return DAG.getNode(ISD::ROTR, DL, VT,
3207 LHSShiftArg, RHSShiftAmt).getNode();
3212 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y)
3213 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotl x, (sub 32, y))
3214 if (LHSShiftAmt.getOpcode() == ISD::SUB &&
3215 RHSShiftAmt == LHSShiftAmt.getOperand(1)) {
3216 if (ConstantSDNode *SUBC =
3217 dyn_cast<ConstantSDNode>(LHSShiftAmt.getOperand(0))) {
3218 if (SUBC->getAPIntValue() == OpSizeInBits) {
3220 return DAG.getNode(ISD::ROTR, DL, VT,
3221 LHSShiftArg, RHSShiftAmt).getNode();
3223 return DAG.getNode(ISD::ROTL, DL, VT,
3224 LHSShiftArg, LHSShiftAmt).getNode();
3229 // Look for sign/zext/any-extended or truncate cases:
3230 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND
3231 || LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND
3232 || LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND
3233 || LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3234 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND
3235 || RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND
3236 || RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND
3237 || RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3238 SDValue LExtOp0 = LHSShiftAmt.getOperand(0);
3239 SDValue RExtOp0 = RHSShiftAmt.getOperand(0);
3240 if (RExtOp0.getOpcode() == ISD::SUB &&
3241 RExtOp0.getOperand(1) == LExtOp0) {
3242 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
3244 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
3245 // (rotr x, (sub 32, y))
3246 if (ConstantSDNode *SUBC =
3247 dyn_cast<ConstantSDNode>(RExtOp0.getOperand(0))) {
3248 if (SUBC->getAPIntValue() == OpSizeInBits) {
3249 return DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3251 HasROTL ? LHSShiftAmt : RHSShiftAmt).getNode();
3254 } else if (LExtOp0.getOpcode() == ISD::SUB &&
3255 RExtOp0 == LExtOp0.getOperand(1)) {
3256 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
3258 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
3259 // (rotl x, (sub 32, y))
3260 if (ConstantSDNode *SUBC =
3261 dyn_cast<ConstantSDNode>(LExtOp0.getOperand(0))) {
3262 if (SUBC->getAPIntValue() == OpSizeInBits) {
3263 return DAG.getNode(HasROTR ? ISD::ROTR : ISD::ROTL, DL, VT,
3265 HasROTR ? RHSShiftAmt : LHSShiftAmt).getNode();
3274 SDValue DAGCombiner::visitXOR(SDNode *N) {
3275 SDValue N0 = N->getOperand(0);
3276 SDValue N1 = N->getOperand(1);
3277 SDValue LHS, RHS, CC;
3278 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3279 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3280 EVT VT = N0.getValueType();
3283 if (VT.isVector()) {
3284 SDValue FoldedVOp = SimplifyVBinOp(N);
3285 if (FoldedVOp.getNode()) return FoldedVOp;
3288 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3289 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3290 return DAG.getConstant(0, VT);
3291 // fold (xor x, undef) -> undef
3292 if (N0.getOpcode() == ISD::UNDEF)
3294 if (N1.getOpcode() == ISD::UNDEF)
3296 // fold (xor c1, c2) -> c1^c2
3298 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3299 // canonicalize constant to RHS
3301 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0);
3302 // fold (xor x, 0) -> x
3303 if (N1C && N1C->isNullValue())
3306 SDValue RXOR = ReassociateOps(ISD::XOR, N->getDebugLoc(), N0, N1);
3307 if (RXOR.getNode() != 0)
3310 // fold !(x cc y) -> (x !cc y)
3311 if (N1C && N1C->getAPIntValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3312 bool isInt = LHS.getValueType().isInteger();
3313 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3316 if (!LegalOperations || TLI.isCondCodeLegal(NotCC, LHS.getValueType())) {
3317 switch (N0.getOpcode()) {
3319 llvm_unreachable("Unhandled SetCC Equivalent!");
3321 return DAG.getSetCC(N->getDebugLoc(), VT, LHS, RHS, NotCC);
3322 case ISD::SELECT_CC:
3323 return DAG.getSelectCC(N->getDebugLoc(), LHS, RHS, N0.getOperand(2),
3324 N0.getOperand(3), NotCC);
3329 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3330 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3331 N0.getNode()->hasOneUse() &&
3332 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3333 SDValue V = N0.getOperand(0);
3334 V = DAG.getNode(ISD::XOR, N0.getDebugLoc(), V.getValueType(), V,
3335 DAG.getConstant(1, V.getValueType()));
3336 AddToWorkList(V.getNode());
3337 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, V);
3340 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3341 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3342 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3343 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3344 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3345 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3346 LHS = DAG.getNode(ISD::XOR, LHS.getDebugLoc(), VT, LHS, N1); // LHS = ~LHS
3347 RHS = DAG.getNode(ISD::XOR, RHS.getDebugLoc(), VT, RHS, N1); // RHS = ~RHS
3348 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3349 return DAG.getNode(NewOpcode, N->getDebugLoc(), VT, LHS, RHS);
3352 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3353 if (N1C && N1C->isAllOnesValue() &&
3354 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3355 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3356 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3357 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3358 LHS = DAG.getNode(ISD::XOR, LHS.getDebugLoc(), VT, LHS, N1); // LHS = ~LHS
3359 RHS = DAG.getNode(ISD::XOR, RHS.getDebugLoc(), VT, RHS, N1); // RHS = ~RHS
3360 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3361 return DAG.getNode(NewOpcode, N->getDebugLoc(), VT, LHS, RHS);
3364 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3365 if (N1C && N0.getOpcode() == ISD::XOR) {
3366 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3367 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3369 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N0.getOperand(1),
3370 DAG.getConstant(N1C->getAPIntValue() ^
3371 N00C->getAPIntValue(), VT));
3373 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N0.getOperand(0),
3374 DAG.getConstant(N1C->getAPIntValue() ^
3375 N01C->getAPIntValue(), VT));
3377 // fold (xor x, x) -> 0
3379 return tryFoldToZero(N->getDebugLoc(), TLI, VT, DAG, LegalOperations);
3381 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
3382 if (N0.getOpcode() == N1.getOpcode()) {
3383 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3384 if (Tmp.getNode()) return Tmp;
3387 // Simplify the expression using non-local knowledge.
3388 if (!VT.isVector() &&
3389 SimplifyDemandedBits(SDValue(N, 0)))
3390 return SDValue(N, 0);
3395 /// visitShiftByConstant - Handle transforms common to the three shifts, when
3396 /// the shift amount is a constant.
3397 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, unsigned Amt) {
3398 SDNode *LHS = N->getOperand(0).getNode();
3399 if (!LHS->hasOneUse()) return SDValue();
3401 // We want to pull some binops through shifts, so that we have (and (shift))
3402 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
3403 // thing happens with address calculations, so it's important to canonicalize
3405 bool HighBitSet = false; // Can we transform this if the high bit is set?
3407 switch (LHS->getOpcode()) {
3408 default: return SDValue();
3411 HighBitSet = false; // We can only transform sra if the high bit is clear.
3414 HighBitSet = true; // We can only transform sra if the high bit is set.
3417 if (N->getOpcode() != ISD::SHL)
3418 return SDValue(); // only shl(add) not sr[al](add).
3419 HighBitSet = false; // We can only transform sra if the high bit is clear.
3423 // We require the RHS of the binop to be a constant as well.
3424 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
3425 if (!BinOpCst) return SDValue();
3427 // FIXME: disable this unless the input to the binop is a shift by a constant.
3428 // If it is not a shift, it pessimizes some common cases like:
3430 // void foo(int *X, int i) { X[i & 1235] = 1; }
3431 // int bar(int *X, int i) { return X[i & 255]; }
3432 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
3433 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
3434 BinOpLHSVal->getOpcode() != ISD::SRA &&
3435 BinOpLHSVal->getOpcode() != ISD::SRL) ||
3436 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
3439 EVT VT = N->getValueType(0);
3441 // If this is a signed shift right, and the high bit is modified by the
3442 // logical operation, do not perform the transformation. The highBitSet
3443 // boolean indicates the value of the high bit of the constant which would
3444 // cause it to be modified for this operation.
3445 if (N->getOpcode() == ISD::SRA) {
3446 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
3447 if (BinOpRHSSignSet != HighBitSet)
3451 // Fold the constants, shifting the binop RHS by the shift amount.
3452 SDValue NewRHS = DAG.getNode(N->getOpcode(), LHS->getOperand(1).getDebugLoc(),
3454 LHS->getOperand(1), N->getOperand(1));
3456 // Create the new shift.
3457 SDValue NewShift = DAG.getNode(N->getOpcode(),
3458 LHS->getOperand(0).getDebugLoc(),
3459 VT, LHS->getOperand(0), N->getOperand(1));
3461 // Create the new binop.
3462 return DAG.getNode(LHS->getOpcode(), N->getDebugLoc(), VT, NewShift, NewRHS);
3465 SDValue DAGCombiner::visitSHL(SDNode *N) {
3466 SDValue N0 = N->getOperand(0);
3467 SDValue N1 = N->getOperand(1);
3468 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3469 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3470 EVT VT = N0.getValueType();
3471 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3473 // fold (shl c1, c2) -> c1<<c2
3475 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
3476 // fold (shl 0, x) -> 0
3477 if (N0C && N0C->isNullValue())
3479 // fold (shl x, c >= size(x)) -> undef
3480 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3481 return DAG.getUNDEF(VT);
3482 // fold (shl x, 0) -> x
3483 if (N1C && N1C->isNullValue())
3485 // fold (shl undef, x) -> 0
3486 if (N0.getOpcode() == ISD::UNDEF)
3487 return DAG.getConstant(0, VT);
3488 // if (shl x, c) is known to be zero, return 0
3489 if (DAG.MaskedValueIsZero(SDValue(N, 0),
3490 APInt::getAllOnesValue(OpSizeInBits)))
3491 return DAG.getConstant(0, VT);
3492 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
3493 if (N1.getOpcode() == ISD::TRUNCATE &&
3494 N1.getOperand(0).getOpcode() == ISD::AND &&
3495 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3496 SDValue N101 = N1.getOperand(0).getOperand(1);
3497 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3498 EVT TruncVT = N1.getValueType();
3499 SDValue N100 = N1.getOperand(0).getOperand(0);
3500 APInt TruncC = N101C->getAPIntValue();
3501 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
3502 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
3503 DAG.getNode(ISD::AND, N->getDebugLoc(), TruncVT,
3504 DAG.getNode(ISD::TRUNCATE,
3507 DAG.getConstant(TruncC, TruncVT)));
3511 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3512 return SDValue(N, 0);
3514 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
3515 if (N1C && N0.getOpcode() == ISD::SHL &&
3516 N0.getOperand(1).getOpcode() == ISD::Constant) {
3517 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3518 uint64_t c2 = N1C->getZExtValue();
3519 if (c1 + c2 >= OpSizeInBits)
3520 return DAG.getConstant(0, VT);
3521 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0.getOperand(0),
3522 DAG.getConstant(c1 + c2, N1.getValueType()));
3525 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
3526 // For this to be valid, the second form must not preserve any of the bits
3527 // that are shifted out by the inner shift in the first form. This means
3528 // the outer shift size must be >= the number of bits added by the ext.
3529 // As a corollary, we don't care what kind of ext it is.
3530 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
3531 N0.getOpcode() == ISD::ANY_EXTEND ||
3532 N0.getOpcode() == ISD::SIGN_EXTEND) &&
3533 N0.getOperand(0).getOpcode() == ISD::SHL &&
3534 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
3536 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
3537 uint64_t c2 = N1C->getZExtValue();
3538 EVT InnerShiftVT = N0.getOperand(0).getValueType();
3539 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
3540 if (c2 >= OpSizeInBits - InnerShiftSize) {
3541 if (c1 + c2 >= OpSizeInBits)
3542 return DAG.getConstant(0, VT);
3543 return DAG.getNode(ISD::SHL, N0->getDebugLoc(), VT,
3544 DAG.getNode(N0.getOpcode(), N0->getDebugLoc(), VT,
3545 N0.getOperand(0)->getOperand(0)),
3546 DAG.getConstant(c1 + c2, N1.getValueType()));
3550 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
3551 // (and (srl x, (sub c1, c2), MASK)
3552 // Only fold this if the inner shift has no other uses -- if it does, folding
3553 // this will increase the total number of instructions.
3554 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse() &&
3555 N0.getOperand(1).getOpcode() == ISD::Constant) {
3556 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3557 if (c1 < VT.getSizeInBits()) {
3558 uint64_t c2 = N1C->getZExtValue();
3559 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
3560 VT.getSizeInBits() - c1);
3563 Mask = Mask.shl(c2-c1);
3564 Shift = DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0.getOperand(0),
3565 DAG.getConstant(c2-c1, N1.getValueType()));
3567 Mask = Mask.lshr(c1-c2);
3568 Shift = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0),
3569 DAG.getConstant(c1-c2, N1.getValueType()));
3571 return DAG.getNode(ISD::AND, N0.getDebugLoc(), VT, Shift,
3572 DAG.getConstant(Mask, VT));
3575 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
3576 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
3577 SDValue HiBitsMask =
3578 DAG.getConstant(APInt::getHighBitsSet(VT.getSizeInBits(),
3579 VT.getSizeInBits() -
3580 N1C->getZExtValue()),
3582 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
3587 SDValue NewSHL = visitShiftByConstant(N, N1C->getZExtValue());
3588 if (NewSHL.getNode())
3595 SDValue DAGCombiner::visitSRA(SDNode *N) {
3596 SDValue N0 = N->getOperand(0);
3597 SDValue N1 = N->getOperand(1);
3598 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3599 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3600 EVT VT = N0.getValueType();
3601 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3603 // fold (sra c1, c2) -> (sra c1, c2)
3605 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
3606 // fold (sra 0, x) -> 0
3607 if (N0C && N0C->isNullValue())
3609 // fold (sra -1, x) -> -1
3610 if (N0C && N0C->isAllOnesValue())
3612 // fold (sra x, (setge c, size(x))) -> undef
3613 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3614 return DAG.getUNDEF(VT);
3615 // fold (sra x, 0) -> x
3616 if (N1C && N1C->isNullValue())
3618 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
3620 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
3621 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
3622 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
3624 ExtVT = EVT::getVectorVT(*DAG.getContext(),
3625 ExtVT, VT.getVectorNumElements());
3626 if ((!LegalOperations ||
3627 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
3628 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
3629 N0.getOperand(0), DAG.getValueType(ExtVT));
3632 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
3633 if (N1C && N0.getOpcode() == ISD::SRA) {
3634 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3635 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
3636 if (Sum >= OpSizeInBits) Sum = OpSizeInBits-1;
3637 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0.getOperand(0),
3638 DAG.getConstant(Sum, N1C->getValueType(0)));
3642 // fold (sra (shl X, m), (sub result_size, n))
3643 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
3644 // result_size - n != m.
3645 // If truncate is free for the target sext(shl) is likely to result in better
3647 if (N0.getOpcode() == ISD::SHL) {
3648 // Get the two constanst of the shifts, CN0 = m, CN = n.
3649 const ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3651 // Determine what the truncate's result bitsize and type would be.
3653 EVT::getIntegerVT(*DAG.getContext(),
3654 OpSizeInBits - N1C->getZExtValue());
3655 // Determine the residual right-shift amount.
3656 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
3658 // If the shift is not a no-op (in which case this should be just a sign
3659 // extend already), the truncated to type is legal, sign_extend is legal
3660 // on that type, and the truncate to that type is both legal and free,
3661 // perform the transform.
3662 if ((ShiftAmt > 0) &&
3663 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
3664 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
3665 TLI.isTruncateFree(VT, TruncVT)) {
3667 SDValue Amt = DAG.getConstant(ShiftAmt,
3668 getShiftAmountTy(N0.getOperand(0).getValueType()));
3669 SDValue Shift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), VT,
3670 N0.getOperand(0), Amt);
3671 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), TruncVT,
3673 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(),
3674 N->getValueType(0), Trunc);
3679 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
3680 if (N1.getOpcode() == ISD::TRUNCATE &&
3681 N1.getOperand(0).getOpcode() == ISD::AND &&
3682 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3683 SDValue N101 = N1.getOperand(0).getOperand(1);
3684 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3685 EVT TruncVT = N1.getValueType();
3686 SDValue N100 = N1.getOperand(0).getOperand(0);
3687 APInt TruncC = N101C->getAPIntValue();
3688 TruncC = TruncC.trunc(TruncVT.getScalarType().getSizeInBits());
3689 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0,
3690 DAG.getNode(ISD::AND, N->getDebugLoc(),
3692 DAG.getNode(ISD::TRUNCATE,
3695 DAG.getConstant(TruncC, TruncVT)));
3699 // fold (sra (trunc (sr x, c1)), c2) -> (trunc (sra x, c1+c2))
3700 // if c1 is equal to the number of bits the trunc removes
3701 if (N0.getOpcode() == ISD::TRUNCATE &&
3702 (N0.getOperand(0).getOpcode() == ISD::SRL ||
3703 N0.getOperand(0).getOpcode() == ISD::SRA) &&
3704 N0.getOperand(0).hasOneUse() &&
3705 N0.getOperand(0).getOperand(1).hasOneUse() &&
3706 N1C && isa<ConstantSDNode>(N0.getOperand(0).getOperand(1))) {
3707 EVT LargeVT = N0.getOperand(0).getValueType();
3708 ConstantSDNode *LargeShiftAmt =
3709 cast<ConstantSDNode>(N0.getOperand(0).getOperand(1));
3711 if (LargeVT.getScalarType().getSizeInBits() - OpSizeInBits ==
3712 LargeShiftAmt->getZExtValue()) {
3714 DAG.getConstant(LargeShiftAmt->getZExtValue() + N1C->getZExtValue(),
3715 getShiftAmountTy(N0.getOperand(0).getOperand(0).getValueType()));
3716 SDValue SRA = DAG.getNode(ISD::SRA, N->getDebugLoc(), LargeVT,
3717 N0.getOperand(0).getOperand(0), Amt);
3718 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, SRA);
3722 // Simplify, based on bits shifted out of the LHS.
3723 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3724 return SDValue(N, 0);
3727 // If the sign bit is known to be zero, switch this to a SRL.
3728 if (DAG.SignBitIsZero(N0))
3729 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0, N1);
3732 SDValue NewSRA = visitShiftByConstant(N, N1C->getZExtValue());
3733 if (NewSRA.getNode())
3740 SDValue DAGCombiner::visitSRL(SDNode *N) {
3741 SDValue N0 = N->getOperand(0);
3742 SDValue N1 = N->getOperand(1);
3743 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3744 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3745 EVT VT = N0.getValueType();
3746 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3748 // fold (srl c1, c2) -> c1 >>u c2
3750 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
3751 // fold (srl 0, x) -> 0
3752 if (N0C && N0C->isNullValue())
3754 // fold (srl x, c >= size(x)) -> undef
3755 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3756 return DAG.getUNDEF(VT);
3757 // fold (srl x, 0) -> x
3758 if (N1C && N1C->isNullValue())
3760 // if (srl x, c) is known to be zero, return 0
3761 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
3762 APInt::getAllOnesValue(OpSizeInBits)))
3763 return DAG.getConstant(0, VT);
3765 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
3766 if (N1C && N0.getOpcode() == ISD::SRL &&
3767 N0.getOperand(1).getOpcode() == ISD::Constant) {
3768 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3769 uint64_t c2 = N1C->getZExtValue();
3770 if (c1 + c2 >= OpSizeInBits)
3771 return DAG.getConstant(0, VT);
3772 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0),
3773 DAG.getConstant(c1 + c2, N1.getValueType()));
3776 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
3777 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
3778 N0.getOperand(0).getOpcode() == ISD::SRL &&
3779 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
3781 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
3782 uint64_t c2 = N1C->getZExtValue();
3783 EVT InnerShiftVT = N0.getOperand(0).getValueType();
3784 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
3785 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
3786 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
3787 if (c1 + OpSizeInBits == InnerShiftSize) {
3788 if (c1 + c2 >= InnerShiftSize)
3789 return DAG.getConstant(0, VT);
3790 return DAG.getNode(ISD::TRUNCATE, N0->getDebugLoc(), VT,
3791 DAG.getNode(ISD::SRL, N0->getDebugLoc(), InnerShiftVT,
3792 N0.getOperand(0)->getOperand(0),
3793 DAG.getConstant(c1 + c2, ShiftCountVT)));
3797 // fold (srl (shl x, c), c) -> (and x, cst2)
3798 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1 &&
3799 N0.getValueSizeInBits() <= 64) {
3800 uint64_t ShAmt = N1C->getZExtValue()+64-N0.getValueSizeInBits();
3801 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
3802 DAG.getConstant(~0ULL >> ShAmt, VT));
3806 // fold (srl (anyextend x), c) -> (anyextend (srl x, c))
3807 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
3808 // Shifting in all undef bits?
3809 EVT SmallVT = N0.getOperand(0).getValueType();
3810 if (N1C->getZExtValue() >= SmallVT.getSizeInBits())
3811 return DAG.getUNDEF(VT);
3813 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
3814 uint64_t ShiftAmt = N1C->getZExtValue();
3815 SDValue SmallShift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), SmallVT,
3817 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
3818 AddToWorkList(SmallShift.getNode());
3819 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, SmallShift);
3823 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
3824 // bit, which is unmodified by sra.
3825 if (N1C && N1C->getZExtValue() + 1 == VT.getSizeInBits()) {
3826 if (N0.getOpcode() == ISD::SRA)
3827 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0), N1);
3830 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
3831 if (N1C && N0.getOpcode() == ISD::CTLZ &&
3832 N1C->getAPIntValue() == Log2_32(VT.getSizeInBits())) {
3833 APInt KnownZero, KnownOne;
3834 DAG.ComputeMaskedBits(N0.getOperand(0), KnownZero, KnownOne);
3836 // If any of the input bits are KnownOne, then the input couldn't be all
3837 // zeros, thus the result of the srl will always be zero.
3838 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
3840 // If all of the bits input the to ctlz node are known to be zero, then
3841 // the result of the ctlz is "32" and the result of the shift is one.
3842 APInt UnknownBits = ~KnownZero;
3843 if (UnknownBits == 0) return DAG.getConstant(1, VT);
3845 // Otherwise, check to see if there is exactly one bit input to the ctlz.
3846 if ((UnknownBits & (UnknownBits - 1)) == 0) {
3847 // Okay, we know that only that the single bit specified by UnknownBits
3848 // could be set on input to the CTLZ node. If this bit is set, the SRL
3849 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
3850 // to an SRL/XOR pair, which is likely to simplify more.
3851 unsigned ShAmt = UnknownBits.countTrailingZeros();
3852 SDValue Op = N0.getOperand(0);
3855 Op = DAG.getNode(ISD::SRL, N0.getDebugLoc(), VT, Op,
3856 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
3857 AddToWorkList(Op.getNode());
3860 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
3861 Op, DAG.getConstant(1, VT));
3865 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
3866 if (N1.getOpcode() == ISD::TRUNCATE &&
3867 N1.getOperand(0).getOpcode() == ISD::AND &&
3868 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3869 SDValue N101 = N1.getOperand(0).getOperand(1);
3870 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3871 EVT TruncVT = N1.getValueType();
3872 SDValue N100 = N1.getOperand(0).getOperand(0);
3873 APInt TruncC = N101C->getAPIntValue();
3874 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
3875 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0,
3876 DAG.getNode(ISD::AND, N->getDebugLoc(),
3878 DAG.getNode(ISD::TRUNCATE,
3881 DAG.getConstant(TruncC, TruncVT)));
3885 // fold operands of srl based on knowledge that the low bits are not
3887 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3888 return SDValue(N, 0);
3891 SDValue NewSRL = visitShiftByConstant(N, N1C->getZExtValue());
3892 if (NewSRL.getNode())
3896 // Attempt to convert a srl of a load into a narrower zero-extending load.
3897 SDValue NarrowLoad = ReduceLoadWidth(N);
3898 if (NarrowLoad.getNode())
3901 // Here is a common situation. We want to optimize:
3904 // %b = and i32 %a, 2
3905 // %c = srl i32 %b, 1
3906 // brcond i32 %c ...
3912 // %c = setcc eq %b, 0
3915 // However when after the source operand of SRL is optimized into AND, the SRL
3916 // itself may not be optimized further. Look for it and add the BRCOND into
3918 if (N->hasOneUse()) {
3919 SDNode *Use = *N->use_begin();
3920 if (Use->getOpcode() == ISD::BRCOND)
3922 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
3923 // Also look pass the truncate.
3924 Use = *Use->use_begin();
3925 if (Use->getOpcode() == ISD::BRCOND)
3933 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
3934 SDValue N0 = N->getOperand(0);
3935 EVT VT = N->getValueType(0);
3937 // fold (ctlz c1) -> c2
3938 if (isa<ConstantSDNode>(N0))
3939 return DAG.getNode(ISD::CTLZ, N->getDebugLoc(), VT, N0);
3943 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
3944 SDValue N0 = N->getOperand(0);
3945 EVT VT = N->getValueType(0);
3947 // fold (ctlz_zero_undef c1) -> c2
3948 if (isa<ConstantSDNode>(N0))
3949 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, N->getDebugLoc(), VT, N0);
3953 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
3954 SDValue N0 = N->getOperand(0);
3955 EVT VT = N->getValueType(0);
3957 // fold (cttz c1) -> c2
3958 if (isa<ConstantSDNode>(N0))
3959 return DAG.getNode(ISD::CTTZ, N->getDebugLoc(), VT, N0);
3963 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
3964 SDValue N0 = N->getOperand(0);
3965 EVT VT = N->getValueType(0);
3967 // fold (cttz_zero_undef c1) -> c2
3968 if (isa<ConstantSDNode>(N0))
3969 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, N->getDebugLoc(), VT, N0);
3973 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
3974 SDValue N0 = N->getOperand(0);
3975 EVT VT = N->getValueType(0);
3977 // fold (ctpop c1) -> c2
3978 if (isa<ConstantSDNode>(N0))
3979 return DAG.getNode(ISD::CTPOP, N->getDebugLoc(), VT, N0);
3983 SDValue DAGCombiner::visitSELECT(SDNode *N) {
3984 SDValue N0 = N->getOperand(0);
3985 SDValue N1 = N->getOperand(1);
3986 SDValue N2 = N->getOperand(2);
3987 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3988 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3989 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
3990 EVT VT = N->getValueType(0);
3991 EVT VT0 = N0.getValueType();
3993 // fold (select C, X, X) -> X
3996 // fold (select true, X, Y) -> X
3997 if (N0C && !N0C->isNullValue())
3999 // fold (select false, X, Y) -> Y
4000 if (N0C && N0C->isNullValue())
4002 // fold (select C, 1, X) -> (or C, X)
4003 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4004 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N2);
4005 // fold (select C, 0, 1) -> (xor C, 1)
4006 if (VT.isInteger() &&
4009 TLI.getBooleanContents(false) == TargetLowering::ZeroOrOneBooleanContent)) &&
4010 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4013 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT0,
4014 N0, DAG.getConstant(1, VT0));
4015 XORNode = DAG.getNode(ISD::XOR, N0.getDebugLoc(), VT0,
4016 N0, DAG.getConstant(1, VT0));
4017 AddToWorkList(XORNode.getNode());
4019 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, XORNode);
4020 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, XORNode);
4022 // fold (select C, 0, X) -> (and (not C), X)
4023 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4024 SDValue NOTNode = DAG.getNOT(N0.getDebugLoc(), N0, VT);
4025 AddToWorkList(NOTNode.getNode());
4026 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, NOTNode, N2);
4028 // fold (select C, X, 1) -> (or (not C), X)
4029 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4030 SDValue NOTNode = DAG.getNOT(N0.getDebugLoc(), N0, VT);
4031 AddToWorkList(NOTNode.getNode());
4032 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, NOTNode, N1);
4034 // fold (select C, X, 0) -> (and C, X)
4035 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4036 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, N1);
4037 // fold (select X, X, Y) -> (or X, Y)
4038 // fold (select X, 1, Y) -> (or X, Y)
4039 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4040 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N2);
4041 // fold (select X, Y, X) -> (and X, Y)
4042 // fold (select X, Y, 0) -> (and X, Y)
4043 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4044 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, N1);
4046 // If we can fold this based on the true/false value, do so.
4047 if (SimplifySelectOps(N, N1, N2))
4048 return SDValue(N, 0); // Don't revisit N.
4050 // fold selects based on a setcc into other things, such as min/max/abs
4051 if (N0.getOpcode() == ISD::SETCC) {
4053 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
4054 // having to say they don't support SELECT_CC on every type the DAG knows
4055 // about, since there is no way to mark an opcode illegal at all value types
4056 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other) &&
4057 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT))
4058 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT,
4059 N0.getOperand(0), N0.getOperand(1),
4060 N1, N2, N0.getOperand(2));
4061 return SimplifySelect(N->getDebugLoc(), N0, N1, N2);
4067 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
4068 SDValue N0 = N->getOperand(0);
4069 SDValue N1 = N->getOperand(1);
4070 SDValue N2 = N->getOperand(2);
4071 SDValue N3 = N->getOperand(3);
4072 SDValue N4 = N->getOperand(4);
4073 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
4075 // fold select_cc lhs, rhs, x, x, cc -> x
4079 // Determine if the condition we're dealing with is constant
4080 SDValue SCC = SimplifySetCC(TLI.getSetCCResultType(N0.getValueType()),
4081 N0, N1, CC, N->getDebugLoc(), false);
4082 if (SCC.getNode()) AddToWorkList(SCC.getNode());
4084 if (ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode())) {
4085 if (!SCCC->isNullValue())
4086 return N2; // cond always true -> true val
4088 return N3; // cond always false -> false val
4091 // Fold to a simpler select_cc
4092 if (SCC.getNode() && SCC.getOpcode() == ISD::SETCC)
4093 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), N2.getValueType(),
4094 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
4097 // If we can fold this based on the true/false value, do so.
4098 if (SimplifySelectOps(N, N2, N3))
4099 return SDValue(N, 0); // Don't revisit N.
4101 // fold select_cc into other things, such as min/max/abs
4102 return SimplifySelectCC(N->getDebugLoc(), N0, N1, N2, N3, CC);
4105 SDValue DAGCombiner::visitSETCC(SDNode *N) {
4106 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
4107 cast<CondCodeSDNode>(N->getOperand(2))->get(),
4111 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
4112 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
4113 // transformation. Returns true if extension are possible and the above
4114 // mentioned transformation is profitable.
4115 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
4117 SmallVector<SDNode*, 4> &ExtendNodes,
4118 const TargetLowering &TLI) {
4119 bool HasCopyToRegUses = false;
4120 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
4121 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
4122 UE = N0.getNode()->use_end();
4127 if (UI.getUse().getResNo() != N0.getResNo())
4129 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
4130 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
4131 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
4132 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
4133 // Sign bits will be lost after a zext.
4136 for (unsigned i = 0; i != 2; ++i) {
4137 SDValue UseOp = User->getOperand(i);
4140 if (!isa<ConstantSDNode>(UseOp))
4145 ExtendNodes.push_back(User);
4148 // If truncates aren't free and there are users we can't
4149 // extend, it isn't worthwhile.
4152 // Remember if this value is live-out.
4153 if (User->getOpcode() == ISD::CopyToReg)
4154 HasCopyToRegUses = true;
4157 if (HasCopyToRegUses) {
4158 bool BothLiveOut = false;
4159 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
4161 SDUse &Use = UI.getUse();
4162 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
4168 // Both unextended and extended values are live out. There had better be
4169 // a good reason for the transformation.
4170 return ExtendNodes.size();
4175 void DAGCombiner::ExtendSetCCUses(SmallVector<SDNode*, 4> SetCCs,
4176 SDValue Trunc, SDValue ExtLoad, DebugLoc DL,
4177 ISD::NodeType ExtType) {
4178 // Extend SetCC uses if necessary.
4179 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
4180 SDNode *SetCC = SetCCs[i];
4181 SmallVector<SDValue, 4> Ops;
4183 for (unsigned j = 0; j != 2; ++j) {
4184 SDValue SOp = SetCC->getOperand(j);
4186 Ops.push_back(ExtLoad);
4188 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
4191 Ops.push_back(SetCC->getOperand(2));
4192 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0),
4193 &Ops[0], Ops.size()));
4197 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
4198 SDValue N0 = N->getOperand(0);
4199 EVT VT = N->getValueType(0);
4201 // fold (sext c1) -> c1
4202 if (isa<ConstantSDNode>(N0))
4203 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N0);
4205 // fold (sext (sext x)) -> (sext x)
4206 // fold (sext (aext x)) -> (sext x)
4207 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4208 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT,
4211 if (N0.getOpcode() == ISD::TRUNCATE) {
4212 // fold (sext (truncate (load x))) -> (sext (smaller load x))
4213 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
4214 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4215 if (NarrowLoad.getNode()) {
4216 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4217 if (NarrowLoad.getNode() != N0.getNode()) {
4218 CombineTo(N0.getNode(), NarrowLoad);
4219 // CombineTo deleted the truncate, if needed, but not what's under it.
4222 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4225 // See if the value being truncated is already sign extended. If so, just
4226 // eliminate the trunc/sext pair.
4227 SDValue Op = N0.getOperand(0);
4228 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
4229 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
4230 unsigned DestBits = VT.getScalarType().getSizeInBits();
4231 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
4233 if (OpBits == DestBits) {
4234 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
4235 // bits, it is already ready.
4236 if (NumSignBits > DestBits-MidBits)
4238 } else if (OpBits < DestBits) {
4239 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
4240 // bits, just sext from i32.
4241 if (NumSignBits > OpBits-MidBits)
4242 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, Op);
4244 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
4245 // bits, just truncate to i32.
4246 if (NumSignBits > OpBits-MidBits)
4247 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4250 // fold (sext (truncate x)) -> (sextinreg x).
4251 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
4252 N0.getValueType())) {
4253 if (OpBits < DestBits)
4254 Op = DAG.getNode(ISD::ANY_EXTEND, N0.getDebugLoc(), VT, Op);
4255 else if (OpBits > DestBits)
4256 Op = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), VT, Op);
4257 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT, Op,
4258 DAG.getValueType(N0.getValueType()));
4262 // fold (sext (load x)) -> (sext (truncate (sextload x)))
4263 // None of the supported targets knows how to perform load and sign extend
4264 // on vectors in one instruction. We only perform this transformation on
4266 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4267 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4268 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) {
4269 bool DoXform = true;
4270 SmallVector<SDNode*, 4> SetCCs;
4271 if (!N0.hasOneUse())
4272 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
4274 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4275 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
4277 LN0->getBasePtr(), LN0->getPointerInfo(),
4279 LN0->isVolatile(), LN0->isNonTemporal(),
4280 LN0->getAlignment());
4281 CombineTo(N, ExtLoad);
4282 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4283 N0.getValueType(), ExtLoad);
4284 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4285 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4287 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4291 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
4292 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
4293 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
4294 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
4295 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4296 EVT MemVT = LN0->getMemoryVT();
4297 if ((!LegalOperations && !LN0->isVolatile()) ||
4298 TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) {
4299 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
4301 LN0->getBasePtr(), LN0->getPointerInfo(),
4303 LN0->isVolatile(), LN0->isNonTemporal(),
4304 LN0->getAlignment());
4305 CombineTo(N, ExtLoad);
4306 CombineTo(N0.getNode(),
4307 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4308 N0.getValueType(), ExtLoad),
4309 ExtLoad.getValue(1));
4310 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4314 // fold (sext (and/or/xor (load x), cst)) ->
4315 // (and/or/xor (sextload x), (sext cst))
4316 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
4317 N0.getOpcode() == ISD::XOR) &&
4318 isa<LoadSDNode>(N0.getOperand(0)) &&
4319 N0.getOperand(1).getOpcode() == ISD::Constant &&
4320 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()) &&
4321 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
4322 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
4323 if (LN0->getExtensionType() != ISD::ZEXTLOAD) {
4324 bool DoXform = true;
4325 SmallVector<SDNode*, 4> SetCCs;
4326 if (!N0.hasOneUse())
4327 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
4330 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, LN0->getDebugLoc(), VT,
4331 LN0->getChain(), LN0->getBasePtr(),
4332 LN0->getPointerInfo(),
4335 LN0->isNonTemporal(),
4336 LN0->getAlignment());
4337 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4338 Mask = Mask.sext(VT.getSizeInBits());
4339 SDValue And = DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
4340 ExtLoad, DAG.getConstant(Mask, VT));
4341 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
4342 N0.getOperand(0).getDebugLoc(),
4343 N0.getOperand(0).getValueType(), ExtLoad);
4345 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
4346 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4348 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4353 if (N0.getOpcode() == ISD::SETCC) {
4354 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
4355 // Only do this before legalize for now.
4356 if (VT.isVector() && !LegalOperations) {
4357 EVT N0VT = N0.getOperand(0).getValueType();
4358 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
4359 // of the same size as the compared operands. Only optimize sext(setcc())
4360 // if this is the case.
4361 EVT SVT = TLI.getSetCCResultType(N0VT);
4363 // We know that the # elements of the results is the same as the
4364 // # elements of the compare (and the # elements of the compare result
4365 // for that matter). Check to see that they are the same size. If so,
4366 // we know that the element size of the sext'd result matches the
4367 // element size of the compare operands.
4368 if (VT.getSizeInBits() == SVT.getSizeInBits())
4369 return DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4371 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4372 // If the desired elements are smaller or larger than the source
4373 // elements we can use a matching integer vector type and then
4374 // truncate/sign extend
4376 EVT MatchingElementType =
4377 EVT::getIntegerVT(*DAG.getContext(),
4378 N0VT.getScalarType().getSizeInBits());
4379 EVT MatchingVectorType =
4380 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4381 N0VT.getVectorNumElements());
4383 if (SVT == MatchingVectorType) {
4384 SDValue VsetCC = DAG.getSetCC(N->getDebugLoc(), MatchingVectorType,
4385 N0.getOperand(0), N0.getOperand(1),
4386 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4387 return DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT);
4392 // sext(setcc x, y, cc) -> (select_cc x, y, -1, 0, cc)
4393 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
4395 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
4397 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4398 NegOne, DAG.getConstant(0, VT),
4399 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4400 if (SCC.getNode()) return SCC;
4401 if (!LegalOperations ||
4402 TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(VT)))
4403 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4404 DAG.getSetCC(N->getDebugLoc(),
4405 TLI.getSetCCResultType(VT),
4406 N0.getOperand(0), N0.getOperand(1),
4407 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
4408 NegOne, DAG.getConstant(0, VT));
4411 // fold (sext x) -> (zext x) if the sign bit is known zero.
4412 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
4413 DAG.SignBitIsZero(N0))
4414 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
4419 // isTruncateOf - If N is a truncate of some other value, return true, record
4420 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
4421 // This function computes KnownZero to avoid a duplicated call to
4422 // ComputeMaskedBits in the caller.
4423 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
4426 if (N->getOpcode() == ISD::TRUNCATE) {
4427 Op = N->getOperand(0);
4428 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
4432 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
4433 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
4436 SDValue Op0 = N->getOperand(0);
4437 SDValue Op1 = N->getOperand(1);
4438 assert(Op0.getValueType() == Op1.getValueType());
4440 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
4441 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
4442 if (COp0 && COp0->isNullValue())
4444 else if (COp1 && COp1->isNullValue())
4449 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
4451 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
4457 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
4458 SDValue N0 = N->getOperand(0);
4459 EVT VT = N->getValueType(0);
4461 // fold (zext c1) -> c1
4462 if (isa<ConstantSDNode>(N0))
4463 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
4464 // fold (zext (zext x)) -> (zext x)
4465 // fold (zext (aext x)) -> (zext x)
4466 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4467 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT,
4470 // fold (zext (truncate x)) -> (zext x) or
4471 // (zext (truncate x)) -> (truncate x)
4472 // This is valid when the truncated bits of x are already zero.
4473 // FIXME: We should extend this to work for vectors too.
4476 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
4477 APInt TruncatedBits =
4478 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
4479 APInt(Op.getValueSizeInBits(), 0) :
4480 APInt::getBitsSet(Op.getValueSizeInBits(),
4481 N0.getValueSizeInBits(),
4482 std::min(Op.getValueSizeInBits(),
4483 VT.getSizeInBits()));
4484 if (TruncatedBits == (KnownZero & TruncatedBits)) {
4485 if (VT.bitsGT(Op.getValueType()))
4486 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, Op);
4487 if (VT.bitsLT(Op.getValueType()))
4488 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4494 // fold (zext (truncate (load x))) -> (zext (smaller load x))
4495 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
4496 if (N0.getOpcode() == ISD::TRUNCATE) {
4497 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4498 if (NarrowLoad.getNode()) {
4499 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4500 if (NarrowLoad.getNode() != N0.getNode()) {
4501 CombineTo(N0.getNode(), NarrowLoad);
4502 // CombineTo deleted the truncate, if needed, but not what's under it.
4505 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4509 // fold (zext (truncate x)) -> (and x, mask)
4510 if (N0.getOpcode() == ISD::TRUNCATE &&
4511 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
4513 // fold (zext (truncate (load x))) -> (zext (smaller load x))
4514 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
4515 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4516 if (NarrowLoad.getNode()) {
4517 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4518 if (NarrowLoad.getNode() != N0.getNode()) {
4519 CombineTo(N0.getNode(), NarrowLoad);
4520 // CombineTo deleted the truncate, if needed, but not what's under it.
4523 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4526 SDValue Op = N0.getOperand(0);
4527 if (Op.getValueType().bitsLT(VT)) {
4528 Op = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, Op);
4529 } else if (Op.getValueType().bitsGT(VT)) {
4530 Op = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4532 return DAG.getZeroExtendInReg(Op, N->getDebugLoc(),
4533 N0.getValueType().getScalarType());
4536 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
4537 // if either of the casts is not free.
4538 if (N0.getOpcode() == ISD::AND &&
4539 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
4540 N0.getOperand(1).getOpcode() == ISD::Constant &&
4541 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
4542 N0.getValueType()) ||
4543 !TLI.isZExtFree(N0.getValueType(), VT))) {
4544 SDValue X = N0.getOperand(0).getOperand(0);
4545 if (X.getValueType().bitsLT(VT)) {
4546 X = DAG.getNode(ISD::ANY_EXTEND, X.getDebugLoc(), VT, X);
4547 } else if (X.getValueType().bitsGT(VT)) {
4548 X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
4550 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4551 Mask = Mask.zext(VT.getSizeInBits());
4552 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4553 X, DAG.getConstant(Mask, VT));
4556 // fold (zext (load x)) -> (zext (truncate (zextload x)))
4557 // None of the supported targets knows how to perform load and vector_zext
4558 // on vectors in one instruction. We only perform this transformation on
4560 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4561 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4562 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
4563 bool DoXform = true;
4564 SmallVector<SDNode*, 4> SetCCs;
4565 if (!N0.hasOneUse())
4566 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
4568 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4569 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N->getDebugLoc(), VT,
4571 LN0->getBasePtr(), LN0->getPointerInfo(),
4573 LN0->isVolatile(), LN0->isNonTemporal(),
4574 LN0->getAlignment());
4575 CombineTo(N, ExtLoad);
4576 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4577 N0.getValueType(), ExtLoad);
4578 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4580 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4582 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4586 // fold (zext (and/or/xor (load x), cst)) ->
4587 // (and/or/xor (zextload x), (zext cst))
4588 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
4589 N0.getOpcode() == ISD::XOR) &&
4590 isa<LoadSDNode>(N0.getOperand(0)) &&
4591 N0.getOperand(1).getOpcode() == ISD::Constant &&
4592 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()) &&
4593 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
4594 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
4595 if (LN0->getExtensionType() != ISD::SEXTLOAD) {
4596 bool DoXform = true;
4597 SmallVector<SDNode*, 4> SetCCs;
4598 if (!N0.hasOneUse())
4599 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
4602 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), VT,
4603 LN0->getChain(), LN0->getBasePtr(),
4604 LN0->getPointerInfo(),
4607 LN0->isNonTemporal(),
4608 LN0->getAlignment());
4609 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4610 Mask = Mask.zext(VT.getSizeInBits());
4611 SDValue And = DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
4612 ExtLoad, DAG.getConstant(Mask, VT));
4613 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
4614 N0.getOperand(0).getDebugLoc(),
4615 N0.getOperand(0).getValueType(), ExtLoad);
4617 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
4618 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4620 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4625 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
4626 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
4627 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
4628 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
4629 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4630 EVT MemVT = LN0->getMemoryVT();
4631 if ((!LegalOperations && !LN0->isVolatile()) ||
4632 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) {
4633 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N->getDebugLoc(), VT,
4635 LN0->getBasePtr(), LN0->getPointerInfo(),
4637 LN0->isVolatile(), LN0->isNonTemporal(),
4638 LN0->getAlignment());
4639 CombineTo(N, ExtLoad);
4640 CombineTo(N0.getNode(),
4641 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), N0.getValueType(),
4643 ExtLoad.getValue(1));
4644 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4648 if (N0.getOpcode() == ISD::SETCC) {
4649 if (!LegalOperations && VT.isVector()) {
4650 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
4651 // Only do this before legalize for now.
4652 EVT N0VT = N0.getOperand(0).getValueType();
4653 EVT EltVT = VT.getVectorElementType();
4654 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
4655 DAG.getConstant(1, EltVT));
4656 if (VT.getSizeInBits() == N0VT.getSizeInBits())
4657 // We know that the # elements of the results is the same as the
4658 // # elements of the compare (and the # elements of the compare result
4659 // for that matter). Check to see that they are the same size. If so,
4660 // we know that the element size of the sext'd result matches the
4661 // element size of the compare operands.
4662 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4663 DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4665 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
4666 DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
4667 &OneOps[0], OneOps.size()));
4669 // If the desired elements are smaller or larger than the source
4670 // elements we can use a matching integer vector type and then
4671 // truncate/sign extend
4672 EVT MatchingElementType =
4673 EVT::getIntegerVT(*DAG.getContext(),
4674 N0VT.getScalarType().getSizeInBits());
4675 EVT MatchingVectorType =
4676 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4677 N0VT.getVectorNumElements());
4679 DAG.getSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
4681 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4682 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4683 DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT),
4684 DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
4685 &OneOps[0], OneOps.size()));
4688 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
4690 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4691 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
4692 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4693 if (SCC.getNode()) return SCC;
4696 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
4697 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
4698 isa<ConstantSDNode>(N0.getOperand(1)) &&
4699 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
4701 SDValue ShAmt = N0.getOperand(1);
4702 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
4703 if (N0.getOpcode() == ISD::SHL) {
4704 SDValue InnerZExt = N0.getOperand(0);
4705 // If the original shl may be shifting out bits, do not perform this
4707 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
4708 InnerZExt.getOperand(0).getValueType().getSizeInBits();
4709 if (ShAmtVal > KnownZeroBits)
4713 DebugLoc DL = N->getDebugLoc();
4715 // Ensure that the shift amount is wide enough for the shifted value.
4716 if (VT.getSizeInBits() >= 256)
4717 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
4719 return DAG.getNode(N0.getOpcode(), DL, VT,
4720 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
4727 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
4728 SDValue N0 = N->getOperand(0);
4729 EVT VT = N->getValueType(0);
4731 // fold (aext c1) -> c1
4732 if (isa<ConstantSDNode>(N0))
4733 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, N0);
4734 // fold (aext (aext x)) -> (aext x)
4735 // fold (aext (zext x)) -> (zext x)
4736 // fold (aext (sext x)) -> (sext x)
4737 if (N0.getOpcode() == ISD::ANY_EXTEND ||
4738 N0.getOpcode() == ISD::ZERO_EXTEND ||
4739 N0.getOpcode() == ISD::SIGN_EXTEND)
4740 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, N0.getOperand(0));
4742 // fold (aext (truncate (load x))) -> (aext (smaller load x))
4743 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
4744 if (N0.getOpcode() == ISD::TRUNCATE) {
4745 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4746 if (NarrowLoad.getNode()) {
4747 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4748 if (NarrowLoad.getNode() != N0.getNode()) {
4749 CombineTo(N0.getNode(), NarrowLoad);
4750 // CombineTo deleted the truncate, if needed, but not what's under it.
4753 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4757 // fold (aext (truncate x))
4758 if (N0.getOpcode() == ISD::TRUNCATE) {
4759 SDValue TruncOp = N0.getOperand(0);
4760 if (TruncOp.getValueType() == VT)
4761 return TruncOp; // x iff x size == zext size.
4762 if (TruncOp.getValueType().bitsGT(VT))
4763 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, TruncOp);
4764 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, TruncOp);
4767 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
4768 // if the trunc is not free.
4769 if (N0.getOpcode() == ISD::AND &&
4770 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
4771 N0.getOperand(1).getOpcode() == ISD::Constant &&
4772 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
4773 N0.getValueType())) {
4774 SDValue X = N0.getOperand(0).getOperand(0);
4775 if (X.getValueType().bitsLT(VT)) {
4776 X = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, X);
4777 } else if (X.getValueType().bitsGT(VT)) {
4778 X = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, X);
4780 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4781 Mask = Mask.zext(VT.getSizeInBits());
4782 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4783 X, DAG.getConstant(Mask, VT));
4786 // fold (aext (load x)) -> (aext (truncate (extload x)))
4787 // None of the supported targets knows how to perform load and any_ext
4788 // on vectors in one instruction. We only perform this transformation on
4790 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4791 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4792 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
4793 bool DoXform = true;
4794 SmallVector<SDNode*, 4> SetCCs;
4795 if (!N0.hasOneUse())
4796 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
4798 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4799 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
4801 LN0->getBasePtr(), LN0->getPointerInfo(),
4803 LN0->isVolatile(), LN0->isNonTemporal(),
4804 LN0->getAlignment());
4805 CombineTo(N, ExtLoad);
4806 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4807 N0.getValueType(), ExtLoad);
4808 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4809 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4811 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4815 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
4816 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
4817 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
4818 if (N0.getOpcode() == ISD::LOAD &&
4819 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
4821 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4822 EVT MemVT = LN0->getMemoryVT();
4823 SDValue ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), N->getDebugLoc(),
4824 VT, LN0->getChain(), LN0->getBasePtr(),
4825 LN0->getPointerInfo(), MemVT,
4826 LN0->isVolatile(), LN0->isNonTemporal(),
4827 LN0->getAlignment());
4828 CombineTo(N, ExtLoad);
4829 CombineTo(N0.getNode(),
4830 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4831 N0.getValueType(), ExtLoad),
4832 ExtLoad.getValue(1));
4833 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4836 if (N0.getOpcode() == ISD::SETCC) {
4837 // aext(setcc) -> sext_in_reg(vsetcc) for vectors.
4838 // Only do this before legalize for now.
4839 if (VT.isVector() && !LegalOperations) {
4840 EVT N0VT = N0.getOperand(0).getValueType();
4841 // We know that the # elements of the results is the same as the
4842 // # elements of the compare (and the # elements of the compare result
4843 // for that matter). Check to see that they are the same size. If so,
4844 // we know that the element size of the sext'd result matches the
4845 // element size of the compare operands.
4846 if (VT.getSizeInBits() == N0VT.getSizeInBits())
4847 return DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4849 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4850 // If the desired elements are smaller or larger than the source
4851 // elements we can use a matching integer vector type and then
4852 // truncate/sign extend
4854 EVT MatchingElementType =
4855 EVT::getIntegerVT(*DAG.getContext(),
4856 N0VT.getScalarType().getSizeInBits());
4857 EVT MatchingVectorType =
4858 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4859 N0VT.getVectorNumElements());
4861 DAG.getSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
4863 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4864 return DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT);
4868 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
4870 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4871 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
4872 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4880 /// GetDemandedBits - See if the specified operand can be simplified with the
4881 /// knowledge that only the bits specified by Mask are used. If so, return the
4882 /// simpler operand, otherwise return a null SDValue.
4883 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
4884 switch (V.getOpcode()) {
4886 case ISD::Constant: {
4887 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
4888 assert(CV != 0 && "Const value should be ConstSDNode.");
4889 const APInt &CVal = CV->getAPIntValue();
4890 APInt NewVal = CVal & Mask;
4891 if (NewVal != CVal) {
4892 return DAG.getConstant(NewVal, V.getValueType());
4898 // If the LHS or RHS don't contribute bits to the or, drop them.
4899 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
4900 return V.getOperand(1);
4901 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
4902 return V.getOperand(0);
4905 // Only look at single-use SRLs.
4906 if (!V.getNode()->hasOneUse())
4908 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
4909 // See if we can recursively simplify the LHS.
4910 unsigned Amt = RHSC->getZExtValue();
4912 // Watch out for shift count overflow though.
4913 if (Amt >= Mask.getBitWidth()) break;
4914 APInt NewMask = Mask << Amt;
4915 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
4916 if (SimplifyLHS.getNode())
4917 return DAG.getNode(ISD::SRL, V.getDebugLoc(), V.getValueType(),
4918 SimplifyLHS, V.getOperand(1));
4924 /// ReduceLoadWidth - If the result of a wider load is shifted to right of N
4925 /// bits and then truncated to a narrower type and where N is a multiple
4926 /// of number of bits of the narrower type, transform it to a narrower load
4927 /// from address + N / num of bits of new type. If the result is to be
4928 /// extended, also fold the extension to form a extending load.
4929 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
4930 unsigned Opc = N->getOpcode();
4932 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
4933 SDValue N0 = N->getOperand(0);
4934 EVT VT = N->getValueType(0);
4937 // This transformation isn't valid for vector loads.
4941 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
4943 if (Opc == ISD::SIGN_EXTEND_INREG) {
4944 ExtType = ISD::SEXTLOAD;
4945 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
4946 } else if (Opc == ISD::SRL) {
4947 // Another special-case: SRL is basically zero-extending a narrower value.
4948 ExtType = ISD::ZEXTLOAD;
4950 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4951 if (!N01) return SDValue();
4952 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
4953 VT.getSizeInBits() - N01->getZExtValue());
4955 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, ExtVT))
4958 unsigned EVTBits = ExtVT.getSizeInBits();
4960 // Do not generate loads of non-round integer types since these can
4961 // be expensive (and would be wrong if the type is not byte sized).
4962 if (!ExtVT.isRound())
4966 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
4967 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
4968 ShAmt = N01->getZExtValue();
4969 // Is the shift amount a multiple of size of VT?
4970 if ((ShAmt & (EVTBits-1)) == 0) {
4971 N0 = N0.getOperand(0);
4972 // Is the load width a multiple of size of VT?
4973 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
4977 // At this point, we must have a load or else we can't do the transform.
4978 if (!isa<LoadSDNode>(N0)) return SDValue();
4980 // If the shift amount is larger than the input type then we're not
4981 // accessing any of the loaded bytes. If the load was a zextload/extload
4982 // then the result of the shift+trunc is zero/undef (handled elsewhere).
4983 // If the load was a sextload then the result is a splat of the sign bit
4984 // of the extended byte. This is not worth optimizing for.
4985 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
4990 // If the load is shifted left (and the result isn't shifted back right),
4991 // we can fold the truncate through the shift.
4992 unsigned ShLeftAmt = 0;
4993 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
4994 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
4995 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
4996 ShLeftAmt = N01->getZExtValue();
4997 N0 = N0.getOperand(0);
5001 // If we haven't found a load, we can't narrow it. Don't transform one with
5002 // multiple uses, this would require adding a new load.
5003 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse() ||
5004 // Don't change the width of a volatile load.
5005 cast<LoadSDNode>(N0)->isVolatile())
5008 // Verify that we are actually reducing a load width here.
5009 if (cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits() < EVTBits)
5012 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5013 EVT PtrType = N0.getOperand(1).getValueType();
5015 // For big endian targets, we need to adjust the offset to the pointer to
5016 // load the correct bytes.
5017 if (TLI.isBigEndian()) {
5018 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
5019 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
5020 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
5023 uint64_t PtrOff = ShAmt / 8;
5024 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
5025 SDValue NewPtr = DAG.getNode(ISD::ADD, LN0->getDebugLoc(),
5026 PtrType, LN0->getBasePtr(),
5027 DAG.getConstant(PtrOff, PtrType));
5028 AddToWorkList(NewPtr.getNode());
5031 if (ExtType == ISD::NON_EXTLOAD)
5032 Load = DAG.getLoad(VT, N0.getDebugLoc(), LN0->getChain(), NewPtr,
5033 LN0->getPointerInfo().getWithOffset(PtrOff),
5034 LN0->isVolatile(), LN0->isNonTemporal(),
5035 LN0->isInvariant(), NewAlign);
5037 Load = DAG.getExtLoad(ExtType, N0.getDebugLoc(), VT, LN0->getChain(),NewPtr,
5038 LN0->getPointerInfo().getWithOffset(PtrOff),
5039 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
5042 // Replace the old load's chain with the new load's chain.
5043 WorkListRemover DeadNodes(*this);
5044 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1),
5047 // Shift the result left, if we've swallowed a left shift.
5048 SDValue Result = Load;
5049 if (ShLeftAmt != 0) {
5050 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
5051 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
5053 Result = DAG.getNode(ISD::SHL, N0.getDebugLoc(), VT,
5054 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
5057 // Return the new loaded value.
5061 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
5062 SDValue N0 = N->getOperand(0);
5063 SDValue N1 = N->getOperand(1);
5064 EVT VT = N->getValueType(0);
5065 EVT EVT = cast<VTSDNode>(N1)->getVT();
5066 unsigned VTBits = VT.getScalarType().getSizeInBits();
5067 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
5069 // fold (sext_in_reg c1) -> c1
5070 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
5071 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT, N0, N1);
5073 // If the input is already sign extended, just drop the extension.
5074 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
5077 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
5078 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
5079 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT())) {
5080 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
5081 N0.getOperand(0), N1);
5084 // fold (sext_in_reg (sext x)) -> (sext x)
5085 // fold (sext_in_reg (aext x)) -> (sext x)
5086 // if x is small enough.
5087 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
5088 SDValue N00 = N0.getOperand(0);
5089 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
5090 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
5091 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N00, N1);
5094 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
5095 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
5096 return DAG.getZeroExtendInReg(N0, N->getDebugLoc(), EVT);
5098 // fold operands of sext_in_reg based on knowledge that the top bits are not
5100 if (SimplifyDemandedBits(SDValue(N, 0)))
5101 return SDValue(N, 0);
5103 // fold (sext_in_reg (load x)) -> (smaller sextload x)
5104 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
5105 SDValue NarrowLoad = ReduceLoadWidth(N);
5106 if (NarrowLoad.getNode())
5109 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
5110 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
5111 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
5112 if (N0.getOpcode() == ISD::SRL) {
5113 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
5114 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
5115 // We can turn this into an SRA iff the input to the SRL is already sign
5117 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
5118 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
5119 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT,
5120 N0.getOperand(0), N0.getOperand(1));
5124 // fold (sext_inreg (extload x)) -> (sextload x)
5125 if (ISD::isEXTLoad(N0.getNode()) &&
5126 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5127 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5128 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5129 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5130 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5131 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
5133 LN0->getBasePtr(), LN0->getPointerInfo(),
5135 LN0->isVolatile(), LN0->isNonTemporal(),
5136 LN0->getAlignment());
5137 CombineTo(N, ExtLoad);
5138 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5139 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5141 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
5142 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5144 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5145 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5146 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5147 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5148 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
5150 LN0->getBasePtr(), LN0->getPointerInfo(),
5152 LN0->isVolatile(), LN0->isNonTemporal(),
5153 LN0->getAlignment());
5154 CombineTo(N, ExtLoad);
5155 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5156 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5159 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
5160 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
5161 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
5162 N0.getOperand(1), false);
5163 if (BSwap.getNode() != 0)
5164 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
5171 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
5172 SDValue N0 = N->getOperand(0);
5173 EVT VT = N->getValueType(0);
5174 bool isLE = TLI.isLittleEndian();
5177 if (N0.getValueType() == N->getValueType(0))
5179 // fold (truncate c1) -> c1
5180 if (isa<ConstantSDNode>(N0))
5181 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0);
5182 // fold (truncate (truncate x)) -> (truncate x)
5183 if (N0.getOpcode() == ISD::TRUNCATE)
5184 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0.getOperand(0));
5185 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
5186 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
5187 N0.getOpcode() == ISD::SIGN_EXTEND ||
5188 N0.getOpcode() == ISD::ANY_EXTEND) {
5189 if (N0.getOperand(0).getValueType().bitsLT(VT))
5190 // if the source is smaller than the dest, we still need an extend
5191 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
5193 else if (N0.getOperand(0).getValueType().bitsGT(VT))
5194 // if the source is larger than the dest, than we just need the truncate
5195 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0.getOperand(0));
5197 // if the source and dest are the same type, we can drop both the extend
5198 // and the truncate.
5199 return N0.getOperand(0);
5202 // Fold extract-and-trunc into a narrow extract. For example:
5203 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
5204 // i32 y = TRUNCATE(i64 x)
5206 // v16i8 b = BITCAST (v2i64 val)
5207 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
5209 // Note: We only run this optimization after type legalization (which often
5210 // creates this pattern) and before operation legalization after which
5211 // we need to be more careful about the vector instructions that we generate.
5212 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5213 LegalTypes && !LegalOperations && N0->hasOneUse()) {
5215 EVT VecTy = N0.getOperand(0).getValueType();
5216 EVT ExTy = N0.getValueType();
5217 EVT TrTy = N->getValueType(0);
5219 unsigned NumElem = VecTy.getVectorNumElements();
5220 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
5222 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
5223 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
5225 SDValue EltNo = N0->getOperand(1);
5226 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
5227 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5229 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
5231 SDValue V = DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
5232 NVT, N0.getOperand(0));
5234 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
5235 N->getDebugLoc(), TrTy, V,
5236 DAG.getConstant(Index, MVT::i32));
5240 // See if we can simplify the input to this truncate through knowledge that
5241 // only the low bits are being used.
5242 // For example "trunc (or (shl x, 8), y)" // -> trunc y
5243 // Currently we only perform this optimization on scalars because vectors
5244 // may have different active low bits.
5245 if (!VT.isVector()) {
5247 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
5248 VT.getSizeInBits()));
5249 if (Shorter.getNode())
5250 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Shorter);
5252 // fold (truncate (load x)) -> (smaller load x)
5253 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
5254 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
5255 SDValue Reduced = ReduceLoadWidth(N);
5256 if (Reduced.getNode())
5260 // Simplify the operands using demanded-bits information.
5261 if (!VT.isVector() &&
5262 SimplifyDemandedBits(SDValue(N, 0)))
5263 return SDValue(N, 0);
5268 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
5269 SDValue Elt = N->getOperand(i);
5270 if (Elt.getOpcode() != ISD::MERGE_VALUES)
5271 return Elt.getNode();
5272 return Elt.getOperand(Elt.getResNo()).getNode();
5275 /// CombineConsecutiveLoads - build_pair (load, load) -> load
5276 /// if load locations are consecutive.
5277 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
5278 assert(N->getOpcode() == ISD::BUILD_PAIR);
5280 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
5281 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
5282 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
5283 LD1->getPointerInfo().getAddrSpace() !=
5284 LD2->getPointerInfo().getAddrSpace())
5286 EVT LD1VT = LD1->getValueType(0);
5288 if (ISD::isNON_EXTLoad(LD2) &&
5290 // If both are volatile this would reduce the number of volatile loads.
5291 // If one is volatile it might be ok, but play conservative and bail out.
5292 !LD1->isVolatile() &&
5293 !LD2->isVolatile() &&
5294 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
5295 unsigned Align = LD1->getAlignment();
5296 unsigned NewAlign = TLI.getTargetData()->
5297 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
5299 if (NewAlign <= Align &&
5300 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
5301 return DAG.getLoad(VT, N->getDebugLoc(), LD1->getChain(),
5302 LD1->getBasePtr(), LD1->getPointerInfo(),
5303 false, false, false, Align);
5309 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
5310 SDValue N0 = N->getOperand(0);
5311 EVT VT = N->getValueType(0);
5313 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
5314 // Only do this before legalize, since afterward the target may be depending
5315 // on the bitconvert.
5316 // First check to see if this is all constant.
5318 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
5320 bool isSimple = true;
5321 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i)
5322 if (N0.getOperand(i).getOpcode() != ISD::UNDEF &&
5323 N0.getOperand(i).getOpcode() != ISD::Constant &&
5324 N0.getOperand(i).getOpcode() != ISD::ConstantFP) {
5329 EVT DestEltVT = N->getValueType(0).getVectorElementType();
5330 assert(!DestEltVT.isVector() &&
5331 "Element type of vector ValueType must not be vector!");
5333 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
5336 // If the input is a constant, let getNode fold it.
5337 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
5338 SDValue Res = DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, N0);
5339 if (Res.getNode() != N) {
5340 if (!LegalOperations ||
5341 TLI.isOperationLegal(Res.getNode()->getOpcode(), VT))
5344 // Folding it resulted in an illegal node, and it's too late to
5345 // do that. Clean up the old node and forego the transformation.
5346 // Ideally this won't happen very often, because instcombine
5347 // and the earlier dagcombine runs (where illegal nodes are
5348 // permitted) should have folded most of them already.
5349 DAG.DeleteNode(Res.getNode());
5353 // (conv (conv x, t1), t2) -> (conv x, t2)
5354 if (N0.getOpcode() == ISD::BITCAST)
5355 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT,
5358 // fold (conv (load x)) -> (load (conv*)x)
5359 // If the resultant load doesn't need a higher alignment than the original!
5360 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
5361 // Do not change the width of a volatile load.
5362 !cast<LoadSDNode>(N0)->isVolatile() &&
5363 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT))) {
5364 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5365 unsigned Align = TLI.getTargetData()->
5366 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
5367 unsigned OrigAlign = LN0->getAlignment();
5369 if (Align <= OrigAlign) {
5370 SDValue Load = DAG.getLoad(VT, N->getDebugLoc(), LN0->getChain(),
5371 LN0->getBasePtr(), LN0->getPointerInfo(),
5372 LN0->isVolatile(), LN0->isNonTemporal(),
5373 LN0->isInvariant(), OrigAlign);
5375 CombineTo(N0.getNode(),
5376 DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5377 N0.getValueType(), Load),
5383 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
5384 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
5385 // This often reduces constant pool loads.
5386 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(VT)) ||
5387 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(VT))) &&
5388 N0.getNode()->hasOneUse() && VT.isInteger() && !VT.isVector()) {
5389 SDValue NewConv = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(), VT,
5391 AddToWorkList(NewConv.getNode());
5393 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
5394 if (N0.getOpcode() == ISD::FNEG)
5395 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
5396 NewConv, DAG.getConstant(SignBit, VT));
5397 assert(N0.getOpcode() == ISD::FABS);
5398 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
5399 NewConv, DAG.getConstant(~SignBit, VT));
5402 // fold (bitconvert (fcopysign cst, x)) ->
5403 // (or (and (bitconvert x), sign), (and cst, (not sign)))
5404 // Note that we don't handle (copysign x, cst) because this can always be
5405 // folded to an fneg or fabs.
5406 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
5407 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
5408 VT.isInteger() && !VT.isVector()) {
5409 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
5410 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
5411 if (isTypeLegal(IntXVT)) {
5412 SDValue X = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5413 IntXVT, N0.getOperand(1));
5414 AddToWorkList(X.getNode());
5416 // If X has a different width than the result/lhs, sext it or truncate it.
5417 unsigned VTWidth = VT.getSizeInBits();
5418 if (OrigXWidth < VTWidth) {
5419 X = DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, X);
5420 AddToWorkList(X.getNode());
5421 } else if (OrigXWidth > VTWidth) {
5422 // To get the sign bit in the right place, we have to shift it right
5423 // before truncating.
5424 X = DAG.getNode(ISD::SRL, X.getDebugLoc(),
5425 X.getValueType(), X,
5426 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
5427 AddToWorkList(X.getNode());
5428 X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
5429 AddToWorkList(X.getNode());
5432 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
5433 X = DAG.getNode(ISD::AND, X.getDebugLoc(), VT,
5434 X, DAG.getConstant(SignBit, VT));
5435 AddToWorkList(X.getNode());
5437 SDValue Cst = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5438 VT, N0.getOperand(0));
5439 Cst = DAG.getNode(ISD::AND, Cst.getDebugLoc(), VT,
5440 Cst, DAG.getConstant(~SignBit, VT));
5441 AddToWorkList(Cst.getNode());
5443 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, X, Cst);
5447 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
5448 if (N0.getOpcode() == ISD::BUILD_PAIR) {
5449 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
5450 if (CombineLD.getNode())
5457 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
5458 EVT VT = N->getValueType(0);
5459 return CombineConsecutiveLoads(N, VT);
5462 /// ConstantFoldBITCASTofBUILD_VECTOR - We know that BV is a build_vector
5463 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
5464 /// destination element value type.
5465 SDValue DAGCombiner::
5466 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
5467 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
5469 // If this is already the right type, we're done.
5470 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
5472 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
5473 unsigned DstBitSize = DstEltVT.getSizeInBits();
5475 // If this is a conversion of N elements of one type to N elements of another
5476 // type, convert each element. This handles FP<->INT cases.
5477 if (SrcBitSize == DstBitSize) {
5478 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
5479 BV->getValueType(0).getVectorNumElements());
5481 // Due to the FP element handling below calling this routine recursively,
5482 // we can end up with a scalar-to-vector node here.
5483 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
5484 return DAG.getNode(ISD::SCALAR_TO_VECTOR, BV->getDebugLoc(), VT,
5485 DAG.getNode(ISD::BITCAST, BV->getDebugLoc(),
5486 DstEltVT, BV->getOperand(0)));
5488 SmallVector<SDValue, 8> Ops;
5489 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
5490 SDValue Op = BV->getOperand(i);
5491 // If the vector element type is not legal, the BUILD_VECTOR operands
5492 // are promoted and implicitly truncated. Make that explicit here.
5493 if (Op.getValueType() != SrcEltVT)
5494 Op = DAG.getNode(ISD::TRUNCATE, BV->getDebugLoc(), SrcEltVT, Op);
5495 Ops.push_back(DAG.getNode(ISD::BITCAST, BV->getDebugLoc(),
5497 AddToWorkList(Ops.back().getNode());
5499 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5500 &Ops[0], Ops.size());
5503 // Otherwise, we're growing or shrinking the elements. To avoid having to
5504 // handle annoying details of growing/shrinking FP values, we convert them to
5506 if (SrcEltVT.isFloatingPoint()) {
5507 // Convert the input float vector to a int vector where the elements are the
5509 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
5510 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
5511 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
5515 // Now we know the input is an integer vector. If the output is a FP type,
5516 // convert to integer first, then to FP of the right size.
5517 if (DstEltVT.isFloatingPoint()) {
5518 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
5519 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
5520 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
5522 // Next, convert to FP elements of the same size.
5523 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
5526 // Okay, we know the src/dst types are both integers of differing types.
5527 // Handling growing first.
5528 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
5529 if (SrcBitSize < DstBitSize) {
5530 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
5532 SmallVector<SDValue, 8> Ops;
5533 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
5534 i += NumInputsPerOutput) {
5535 bool isLE = TLI.isLittleEndian();
5536 APInt NewBits = APInt(DstBitSize, 0);
5537 bool EltIsUndef = true;
5538 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
5539 // Shift the previously computed bits over.
5540 NewBits <<= SrcBitSize;
5541 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
5542 if (Op.getOpcode() == ISD::UNDEF) continue;
5545 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
5546 zextOrTrunc(SrcBitSize).zext(DstBitSize);
5550 Ops.push_back(DAG.getUNDEF(DstEltVT));
5552 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
5555 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
5556 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5557 &Ops[0], Ops.size());
5560 // Finally, this must be the case where we are shrinking elements: each input
5561 // turns into multiple outputs.
5562 bool isS2V = ISD::isScalarToVector(BV);
5563 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
5564 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
5565 NumOutputsPerInput*BV->getNumOperands());
5566 SmallVector<SDValue, 8> Ops;
5568 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
5569 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
5570 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
5571 Ops.push_back(DAG.getUNDEF(DstEltVT));
5575 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
5576 getAPIntValue().zextOrTrunc(SrcBitSize);
5578 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
5579 APInt ThisVal = OpVal.trunc(DstBitSize);
5580 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
5581 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
5582 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
5583 return DAG.getNode(ISD::SCALAR_TO_VECTOR, BV->getDebugLoc(), VT,
5585 OpVal = OpVal.lshr(DstBitSize);
5588 // For big endian targets, swap the order of the pieces of each element.
5589 if (TLI.isBigEndian())
5590 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
5593 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5594 &Ops[0], Ops.size());
5597 SDValue DAGCombiner::visitFADD(SDNode *N) {
5598 SDValue N0 = N->getOperand(0);
5599 SDValue N1 = N->getOperand(1);
5600 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5601 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5602 EVT VT = N->getValueType(0);
5605 if (VT.isVector()) {
5606 SDValue FoldedVOp = SimplifyVBinOp(N);
5607 if (FoldedVOp.getNode()) return FoldedVOp;
5610 // fold (fadd c1, c2) -> (fadd c1, c2)
5611 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5612 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N1);
5613 // canonicalize constant to RHS
5614 if (N0CFP && !N1CFP)
5615 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N1, N0);
5616 // fold (fadd A, 0) -> A
5617 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
5618 N1CFP->getValueAPF().isZero())
5620 // fold (fadd A, (fneg B)) -> (fsub A, B)
5621 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
5622 isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
5623 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N0,
5624 GetNegatedExpression(N1, DAG, LegalOperations));
5625 // fold (fadd (fneg A), B) -> (fsub B, A)
5626 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
5627 isNegatibleForFree(N0, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
5628 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N1,
5629 GetNegatedExpression(N0, DAG, LegalOperations));
5631 // If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
5632 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
5633 N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
5634 isa<ConstantFPSDNode>(N0.getOperand(1)))
5635 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0.getOperand(0),
5636 DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5637 N0.getOperand(1), N1));
5642 SDValue DAGCombiner::visitFSUB(SDNode *N) {
5643 SDValue N0 = N->getOperand(0);
5644 SDValue N1 = N->getOperand(1);
5645 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5646 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5647 EVT VT = N->getValueType(0);
5650 if (VT.isVector()) {
5651 SDValue FoldedVOp = SimplifyVBinOp(N);
5652 if (FoldedVOp.getNode()) return FoldedVOp;
5655 // fold (fsub c1, c2) -> c1-c2
5656 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5657 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N0, N1);
5658 // fold (fsub A, 0) -> A
5659 if (DAG.getTarget().Options.UnsafeFPMath &&
5660 N1CFP && N1CFP->getValueAPF().isZero())
5662 // fold (fsub 0, B) -> -B
5663 if (DAG.getTarget().Options.UnsafeFPMath &&
5664 N0CFP && N0CFP->getValueAPF().isZero()) {
5665 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
5666 return GetNegatedExpression(N1, DAG, LegalOperations);
5667 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
5668 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT, N1);
5670 // fold (fsub A, (fneg B)) -> (fadd A, B)
5671 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
5672 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0,
5673 GetNegatedExpression(N1, DAG, LegalOperations));
5675 // If 'unsafe math' is enabled, fold
5676 // (fsub x, (fadd x, y)) -> (fneg y) &
5677 // (fsub x, (fadd y, x)) -> (fneg y)
5678 if (DAG.getTarget().Options.UnsafeFPMath) {
5679 if (N1.getOpcode() == ISD::FADD) {
5680 SDValue N10 = N1->getOperand(0);
5681 SDValue N11 = N1->getOperand(1);
5683 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI,
5684 &DAG.getTarget().Options))
5685 return GetNegatedExpression(N11, DAG, LegalOperations);
5686 else if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI,
5687 &DAG.getTarget().Options))
5688 return GetNegatedExpression(N10, DAG, LegalOperations);
5695 SDValue DAGCombiner::visitFMUL(SDNode *N) {
5696 SDValue N0 = N->getOperand(0);
5697 SDValue N1 = N->getOperand(1);
5698 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5699 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5700 EVT VT = N->getValueType(0);
5701 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5704 if (VT.isVector()) {
5705 SDValue FoldedVOp = SimplifyVBinOp(N);
5706 if (FoldedVOp.getNode()) return FoldedVOp;
5709 // fold (fmul c1, c2) -> c1*c2
5710 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5711 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0, N1);
5712 // canonicalize constant to RHS
5713 if (N0CFP && !N1CFP)
5714 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N1, N0);
5715 // fold (fmul A, 0) -> 0
5716 if (DAG.getTarget().Options.UnsafeFPMath &&
5717 N1CFP && N1CFP->getValueAPF().isZero())
5719 // fold (fmul A, 0) -> 0, vector edition.
5720 if (DAG.getTarget().Options.UnsafeFPMath &&
5721 ISD::isBuildVectorAllZeros(N1.getNode()))
5723 // fold (fmul X, 2.0) -> (fadd X, X)
5724 if (N1CFP && N1CFP->isExactlyValue(+2.0))
5725 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N0);
5726 // fold (fmul X, -1.0) -> (fneg X)
5727 if (N1CFP && N1CFP->isExactlyValue(-1.0))
5728 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
5729 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT, N0);
5731 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
5732 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
5733 &DAG.getTarget().Options)) {
5734 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
5735 &DAG.getTarget().Options)) {
5736 // Both can be negated for free, check to see if at least one is cheaper
5738 if (LHSNeg == 2 || RHSNeg == 2)
5739 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5740 GetNegatedExpression(N0, DAG, LegalOperations),
5741 GetNegatedExpression(N1, DAG, LegalOperations));
5745 // If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
5746 if (DAG.getTarget().Options.UnsafeFPMath &&
5747 N1CFP && N0.getOpcode() == ISD::FMUL &&
5748 N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
5749 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0.getOperand(0),
5750 DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5751 N0.getOperand(1), N1));
5756 SDValue DAGCombiner::visitFDIV(SDNode *N) {
5757 SDValue N0 = N->getOperand(0);
5758 SDValue N1 = N->getOperand(1);
5759 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5760 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5761 EVT VT = N->getValueType(0);
5762 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5765 if (VT.isVector()) {
5766 SDValue FoldedVOp = SimplifyVBinOp(N);
5767 if (FoldedVOp.getNode()) return FoldedVOp;
5770 // fold (fdiv c1, c2) -> c1/c2
5771 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5772 return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT, N0, N1);
5774 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
5775 if (N1CFP && VT != MVT::ppcf128 && DAG.getTarget().Options.UnsafeFPMath) {
5776 // Compute the reciprocal 1.0 / c2.
5777 APFloat N1APF = N1CFP->getValueAPF();
5778 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
5779 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
5780 // Only do the transform if the reciprocal is a legal fp immediate that
5781 // isn't too nasty (eg NaN, denormal, ...).
5782 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
5783 (!LegalOperations ||
5784 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
5785 // backend)... we should handle this gracefully after Legalize.
5786 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
5787 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
5788 TLI.isFPImmLegal(Recip, VT)))
5789 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0,
5790 DAG.getConstantFP(Recip, VT));
5793 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
5794 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
5795 &DAG.getTarget().Options)) {
5796 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
5797 &DAG.getTarget().Options)) {
5798 // Both can be negated for free, check to see if at least one is cheaper
5800 if (LHSNeg == 2 || RHSNeg == 2)
5801 return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT,
5802 GetNegatedExpression(N0, DAG, LegalOperations),
5803 GetNegatedExpression(N1, DAG, LegalOperations));
5810 SDValue DAGCombiner::visitFREM(SDNode *N) {
5811 SDValue N0 = N->getOperand(0);
5812 SDValue N1 = N->getOperand(1);
5813 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5814 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5815 EVT VT = N->getValueType(0);
5817 // fold (frem c1, c2) -> fmod(c1,c2)
5818 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5819 return DAG.getNode(ISD::FREM, N->getDebugLoc(), VT, N0, N1);
5824 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
5825 SDValue N0 = N->getOperand(0);
5826 SDValue N1 = N->getOperand(1);
5827 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5828 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5829 EVT VT = N->getValueType(0);
5831 if (N0CFP && N1CFP && VT != MVT::ppcf128) // Constant fold
5832 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT, N0, N1);
5835 const APFloat& V = N1CFP->getValueAPF();
5836 // copysign(x, c1) -> fabs(x) iff ispos(c1)
5837 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
5838 if (!V.isNegative()) {
5839 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
5840 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
5842 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
5843 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT,
5844 DAG.getNode(ISD::FABS, N0.getDebugLoc(), VT, N0));
5848 // copysign(fabs(x), y) -> copysign(x, y)
5849 // copysign(fneg(x), y) -> copysign(x, y)
5850 // copysign(copysign(x,z), y) -> copysign(x, y)
5851 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
5852 N0.getOpcode() == ISD::FCOPYSIGN)
5853 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
5854 N0.getOperand(0), N1);
5856 // copysign(x, abs(y)) -> abs(x)
5857 if (N1.getOpcode() == ISD::FABS)
5858 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
5860 // copysign(x, copysign(y,z)) -> copysign(x, z)
5861 if (N1.getOpcode() == ISD::FCOPYSIGN)
5862 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
5863 N0, N1.getOperand(1));
5865 // copysign(x, fp_extend(y)) -> copysign(x, y)
5866 // copysign(x, fp_round(y)) -> copysign(x, y)
5867 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
5868 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
5869 N0, N1.getOperand(0));
5874 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
5875 SDValue N0 = N->getOperand(0);
5876 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
5877 EVT VT = N->getValueType(0);
5878 EVT OpVT = N0.getValueType();
5880 // fold (sint_to_fp c1) -> c1fp
5881 if (N0C && OpVT != MVT::ppcf128 &&
5882 // ...but only if the target supports immediate floating-point values
5883 (!LegalOperations ||
5884 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
5885 return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
5887 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
5888 // but UINT_TO_FP is legal on this target, try to convert.
5889 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
5890 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
5891 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
5892 if (DAG.SignBitIsZero(N0))
5893 return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
5899 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
5900 SDValue N0 = N->getOperand(0);
5901 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
5902 EVT VT = N->getValueType(0);
5903 EVT OpVT = N0.getValueType();
5905 // fold (uint_to_fp c1) -> c1fp
5906 if (N0C && OpVT != MVT::ppcf128 &&
5907 // ...but only if the target supports immediate floating-point values
5908 (!LegalOperations ||
5909 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
5910 return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
5912 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
5913 // but SINT_TO_FP is legal on this target, try to convert.
5914 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
5915 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
5916 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
5917 if (DAG.SignBitIsZero(N0))
5918 return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
5924 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
5925 SDValue N0 = N->getOperand(0);
5926 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5927 EVT VT = N->getValueType(0);
5929 // fold (fp_to_sint c1fp) -> c1
5931 return DAG.getNode(ISD::FP_TO_SINT, N->getDebugLoc(), VT, N0);
5936 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
5937 SDValue N0 = N->getOperand(0);
5938 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5939 EVT VT = N->getValueType(0);
5941 // fold (fp_to_uint c1fp) -> c1
5942 if (N0CFP && VT != MVT::ppcf128)
5943 return DAG.getNode(ISD::FP_TO_UINT, N->getDebugLoc(), VT, N0);
5948 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
5949 SDValue N0 = N->getOperand(0);
5950 SDValue N1 = N->getOperand(1);
5951 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5952 EVT VT = N->getValueType(0);
5954 // fold (fp_round c1fp) -> c1fp
5955 if (N0CFP && N0.getValueType() != MVT::ppcf128)
5956 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT, N0, N1);
5958 // fold (fp_round (fp_extend x)) -> x
5959 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
5960 return N0.getOperand(0);
5962 // fold (fp_round (fp_round x)) -> (fp_round x)
5963 if (N0.getOpcode() == ISD::FP_ROUND) {
5964 // This is a value preserving truncation if both round's are.
5965 bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
5966 N0.getNode()->getConstantOperandVal(1) == 1;
5967 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT, N0.getOperand(0),
5968 DAG.getIntPtrConstant(IsTrunc));
5971 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
5972 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
5973 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(), VT,
5974 N0.getOperand(0), N1);
5975 AddToWorkList(Tmp.getNode());
5976 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
5977 Tmp, N0.getOperand(1));
5983 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
5984 SDValue N0 = N->getOperand(0);
5985 EVT VT = N->getValueType(0);
5986 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
5987 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5989 // fold (fp_round_inreg c1fp) -> c1fp
5990 if (N0CFP && isTypeLegal(EVT)) {
5991 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
5992 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, Round);
5998 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
5999 SDValue N0 = N->getOperand(0);
6000 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6001 EVT VT = N->getValueType(0);
6003 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
6004 if (N->hasOneUse() &&
6005 N->use_begin()->getOpcode() == ISD::FP_ROUND)
6008 // fold (fp_extend c1fp) -> c1fp
6009 if (N0CFP && VT != MVT::ppcf128)
6010 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, N0);
6012 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
6014 if (N0.getOpcode() == ISD::FP_ROUND
6015 && N0.getNode()->getConstantOperandVal(1) == 1) {
6016 SDValue In = N0.getOperand(0);
6017 if (In.getValueType() == VT) return In;
6018 if (VT.bitsLT(In.getValueType()))
6019 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT,
6020 In, N0.getOperand(1));
6021 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, In);
6024 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
6025 if (ISD::isNON_EXTLoad(N0.getNode()) && N0.hasOneUse() &&
6026 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6027 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
6028 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6029 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
6031 LN0->getBasePtr(), LN0->getPointerInfo(),
6033 LN0->isVolatile(), LN0->isNonTemporal(),
6034 LN0->getAlignment());
6035 CombineTo(N, ExtLoad);
6036 CombineTo(N0.getNode(),
6037 DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(),
6038 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
6039 ExtLoad.getValue(1));
6040 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6046 SDValue DAGCombiner::visitFNEG(SDNode *N) {
6047 SDValue N0 = N->getOperand(0);
6048 EVT VT = N->getValueType(0);
6050 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
6051 &DAG.getTarget().Options))
6052 return GetNegatedExpression(N0, DAG, LegalOperations);
6054 // Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading
6055 // constant pool values.
6056 if (!TLI.isFNegFree(VT) && N0.getOpcode() == ISD::BITCAST &&
6058 N0.getNode()->hasOneUse() &&
6059 N0.getOperand(0).getValueType().isInteger()) {
6060 SDValue Int = N0.getOperand(0);
6061 EVT IntVT = Int.getValueType();
6062 if (IntVT.isInteger() && !IntVT.isVector()) {
6063 Int = DAG.getNode(ISD::XOR, N0.getDebugLoc(), IntVT, Int,
6064 DAG.getConstant(APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
6065 AddToWorkList(Int.getNode());
6066 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
6074 SDValue DAGCombiner::visitFABS(SDNode *N) {
6075 SDValue N0 = N->getOperand(0);
6076 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6077 EVT VT = N->getValueType(0);
6079 // fold (fabs c1) -> fabs(c1)
6080 if (N0CFP && VT != MVT::ppcf128)
6081 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
6082 // fold (fabs (fabs x)) -> (fabs x)
6083 if (N0.getOpcode() == ISD::FABS)
6084 return N->getOperand(0);
6085 // fold (fabs (fneg x)) -> (fabs x)
6086 // fold (fabs (fcopysign x, y)) -> (fabs x)
6087 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
6088 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0.getOperand(0));
6090 // Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading
6091 // constant pool values.
6092 if (!TLI.isFAbsFree(VT) &&
6093 N0.getOpcode() == ISD::BITCAST && N0.getNode()->hasOneUse() &&
6094 N0.getOperand(0).getValueType().isInteger() &&
6095 !N0.getOperand(0).getValueType().isVector()) {
6096 SDValue Int = N0.getOperand(0);
6097 EVT IntVT = Int.getValueType();
6098 if (IntVT.isInteger() && !IntVT.isVector()) {
6099 Int = DAG.getNode(ISD::AND, N0.getDebugLoc(), IntVT, Int,
6100 DAG.getConstant(~APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
6101 AddToWorkList(Int.getNode());
6102 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
6103 N->getValueType(0), Int);
6110 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
6111 SDValue Chain = N->getOperand(0);
6112 SDValue N1 = N->getOperand(1);
6113 SDValue N2 = N->getOperand(2);
6115 // If N is a constant we could fold this into a fallthrough or unconditional
6116 // branch. However that doesn't happen very often in normal code, because
6117 // Instcombine/SimplifyCFG should have handled the available opportunities.
6118 // If we did this folding here, it would be necessary to update the
6119 // MachineBasicBlock CFG, which is awkward.
6121 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
6123 if (N1.getOpcode() == ISD::SETCC &&
6124 TLI.isOperationLegalOrCustom(ISD::BR_CC, MVT::Other)) {
6125 return DAG.getNode(ISD::BR_CC, N->getDebugLoc(), MVT::Other,
6126 Chain, N1.getOperand(2),
6127 N1.getOperand(0), N1.getOperand(1), N2);
6130 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
6131 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
6132 (N1.getOperand(0).hasOneUse() &&
6133 N1.getOperand(0).getOpcode() == ISD::SRL))) {
6135 if (N1.getOpcode() == ISD::TRUNCATE) {
6136 // Look pass the truncate.
6137 Trunc = N1.getNode();
6138 N1 = N1.getOperand(0);
6141 // Match this pattern so that we can generate simpler code:
6144 // %b = and i32 %a, 2
6145 // %c = srl i32 %b, 1
6146 // brcond i32 %c ...
6151 // %b = and i32 %a, 2
6152 // %c = setcc eq %b, 0
6155 // This applies only when the AND constant value has one bit set and the
6156 // SRL constant is equal to the log2 of the AND constant. The back-end is
6157 // smart enough to convert the result into a TEST/JMP sequence.
6158 SDValue Op0 = N1.getOperand(0);
6159 SDValue Op1 = N1.getOperand(1);
6161 if (Op0.getOpcode() == ISD::AND &&
6162 Op1.getOpcode() == ISD::Constant) {
6163 SDValue AndOp1 = Op0.getOperand(1);
6165 if (AndOp1.getOpcode() == ISD::Constant) {
6166 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
6168 if (AndConst.isPowerOf2() &&
6169 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
6171 DAG.getSetCC(N->getDebugLoc(),
6172 TLI.getSetCCResultType(Op0.getValueType()),
6173 Op0, DAG.getConstant(0, Op0.getValueType()),
6176 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
6177 MVT::Other, Chain, SetCC, N2);
6178 // Don't add the new BRCond into the worklist or else SimplifySelectCC
6179 // will convert it back to (X & C1) >> C2.
6180 CombineTo(N, NewBRCond, false);
6181 // Truncate is dead.
6183 removeFromWorkList(Trunc);
6184 DAG.DeleteNode(Trunc);
6186 // Replace the uses of SRL with SETCC
6187 WorkListRemover DeadNodes(*this);
6188 DAG.ReplaceAllUsesOfValueWith(N1, SetCC, &DeadNodes);
6189 removeFromWorkList(N1.getNode());
6190 DAG.DeleteNode(N1.getNode());
6191 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6197 // Restore N1 if the above transformation doesn't match.
6198 N1 = N->getOperand(1);
6201 // Transform br(xor(x, y)) -> br(x != y)
6202 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
6203 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
6204 SDNode *TheXor = N1.getNode();
6205 SDValue Op0 = TheXor->getOperand(0);
6206 SDValue Op1 = TheXor->getOperand(1);
6207 if (Op0.getOpcode() == Op1.getOpcode()) {
6208 // Avoid missing important xor optimizations.
6209 SDValue Tmp = visitXOR(TheXor);
6210 if (Tmp.getNode() && Tmp.getNode() != TheXor) {
6211 DEBUG(dbgs() << "\nReplacing.8 ";
6213 dbgs() << "\nWith: ";
6214 Tmp.getNode()->dump(&DAG);
6216 WorkListRemover DeadNodes(*this);
6217 DAG.ReplaceAllUsesOfValueWith(N1, Tmp, &DeadNodes);
6218 removeFromWorkList(TheXor);
6219 DAG.DeleteNode(TheXor);
6220 return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
6221 MVT::Other, Chain, Tmp, N2);
6225 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
6227 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
6228 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
6229 Op0.getOpcode() == ISD::XOR) {
6230 TheXor = Op0.getNode();
6234 EVT SetCCVT = N1.getValueType();
6236 SetCCVT = TLI.getSetCCResultType(SetCCVT);
6237 SDValue SetCC = DAG.getSetCC(TheXor->getDebugLoc(),
6240 Equal ? ISD::SETEQ : ISD::SETNE);
6241 // Replace the uses of XOR with SETCC
6242 WorkListRemover DeadNodes(*this);
6243 DAG.ReplaceAllUsesOfValueWith(N1, SetCC, &DeadNodes);
6244 removeFromWorkList(N1.getNode());
6245 DAG.DeleteNode(N1.getNode());
6246 return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
6247 MVT::Other, Chain, SetCC, N2);
6254 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
6256 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
6257 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
6258 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
6260 // If N is a constant we could fold this into a fallthrough or unconditional
6261 // branch. However that doesn't happen very often in normal code, because
6262 // Instcombine/SimplifyCFG should have handled the available opportunities.
6263 // If we did this folding here, it would be necessary to update the
6264 // MachineBasicBlock CFG, which is awkward.
6266 // Use SimplifySetCC to simplify SETCC's.
6267 SDValue Simp = SimplifySetCC(TLI.getSetCCResultType(CondLHS.getValueType()),
6268 CondLHS, CondRHS, CC->get(), N->getDebugLoc(),
6270 if (Simp.getNode()) AddToWorkList(Simp.getNode());
6272 // fold to a simpler setcc
6273 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
6274 return DAG.getNode(ISD::BR_CC, N->getDebugLoc(), MVT::Other,
6275 N->getOperand(0), Simp.getOperand(2),
6276 Simp.getOperand(0), Simp.getOperand(1),
6282 /// canFoldInAddressingMode - Return true if 'Use' is a load or a store that
6283 /// uses N as its base pointer and that N may be folded in the load / store
6284 /// addressing mode.
6285 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
6287 const TargetLowering &TLI) {
6289 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
6290 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
6292 VT = Use->getValueType(0);
6293 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
6294 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
6296 VT = ST->getValue().getValueType();
6300 TargetLowering::AddrMode AM;
6301 if (N->getOpcode() == ISD::ADD) {
6302 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
6305 AM.BaseOffs = Offset->getSExtValue();
6309 } else if (N->getOpcode() == ISD::SUB) {
6310 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
6313 AM.BaseOffs = -Offset->getSExtValue();
6320 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
6323 /// CombineToPreIndexedLoadStore - Try turning a load / store into a
6324 /// pre-indexed load / store when the base pointer is an add or subtract
6325 /// and it has other uses besides the load / store. After the
6326 /// transformation, the new indexed load / store has effectively folded
6327 /// the add / subtract in and all of its other uses are redirected to the
6328 /// new load / store.
6329 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
6330 if (Level < AfterLegalizeDAG)
6336 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6337 if (LD->isIndexed())
6339 VT = LD->getMemoryVT();
6340 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
6341 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
6343 Ptr = LD->getBasePtr();
6344 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6345 if (ST->isIndexed())
6347 VT = ST->getMemoryVT();
6348 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
6349 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
6351 Ptr = ST->getBasePtr();
6357 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
6358 // out. There is no reason to make this a preinc/predec.
6359 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
6360 Ptr.getNode()->hasOneUse())
6363 // Ask the target to do addressing mode selection.
6366 ISD::MemIndexedMode AM = ISD::UNINDEXED;
6367 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
6369 // Don't create a indexed load / store with zero offset.
6370 if (isa<ConstantSDNode>(Offset) &&
6371 cast<ConstantSDNode>(Offset)->isNullValue())
6374 // Try turning it into a pre-indexed load / store except when:
6375 // 1) The new base ptr is a frame index.
6376 // 2) If N is a store and the new base ptr is either the same as or is a
6377 // predecessor of the value being stored.
6378 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
6379 // that would create a cycle.
6380 // 4) All uses are load / store ops that use it as old base ptr.
6382 // Check #1. Preinc'ing a frame index would require copying the stack pointer
6383 // (plus the implicit offset) to a register to preinc anyway.
6384 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
6389 SDValue Val = cast<StoreSDNode>(N)->getValue();
6390 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
6394 // Now check for #3 and #4.
6395 bool RealUse = false;
6397 // Caches for hasPredecessorHelper
6398 SmallPtrSet<const SDNode *, 32> Visited;
6399 SmallVector<const SDNode *, 16> Worklist;
6401 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
6402 E = Ptr.getNode()->use_end(); I != E; ++I) {
6406 if (N->hasPredecessorHelper(Use, Visited, Worklist))
6409 // If Ptr may be folded in addressing mode of other use, then it's
6410 // not profitable to do this transformation.
6411 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
6420 Result = DAG.getIndexedLoad(SDValue(N,0), N->getDebugLoc(),
6421 BasePtr, Offset, AM);
6423 Result = DAG.getIndexedStore(SDValue(N,0), N->getDebugLoc(),
6424 BasePtr, Offset, AM);
6427 DEBUG(dbgs() << "\nReplacing.4 ";
6429 dbgs() << "\nWith: ";
6430 Result.getNode()->dump(&DAG);
6432 WorkListRemover DeadNodes(*this);
6434 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0),
6436 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2),
6439 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1),
6443 // Finally, since the node is now dead, remove it from the graph.
6446 // Replace the uses of Ptr with uses of the updated base value.
6447 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0),
6449 removeFromWorkList(Ptr.getNode());
6450 DAG.DeleteNode(Ptr.getNode());
6455 /// CombineToPostIndexedLoadStore - Try to combine a load / store with a
6456 /// add / sub of the base pointer node into a post-indexed load / store.
6457 /// The transformation folded the add / subtract into the new indexed
6458 /// load / store effectively and all of its uses are redirected to the
6459 /// new load / store.
6460 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
6461 if (Level < AfterLegalizeDAG)
6467 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6468 if (LD->isIndexed())
6470 VT = LD->getMemoryVT();
6471 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
6472 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
6474 Ptr = LD->getBasePtr();
6475 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6476 if (ST->isIndexed())
6478 VT = ST->getMemoryVT();
6479 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
6480 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
6482 Ptr = ST->getBasePtr();
6488 if (Ptr.getNode()->hasOneUse())
6491 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
6492 E = Ptr.getNode()->use_end(); I != E; ++I) {
6495 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
6500 ISD::MemIndexedMode AM = ISD::UNINDEXED;
6501 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
6502 // Don't create a indexed load / store with zero offset.
6503 if (isa<ConstantSDNode>(Offset) &&
6504 cast<ConstantSDNode>(Offset)->isNullValue())
6507 // Try turning it into a post-indexed load / store except when
6508 // 1) All uses are load / store ops that use it as base ptr (and
6509 // it may be folded as addressing mmode).
6510 // 2) Op must be independent of N, i.e. Op is neither a predecessor
6511 // nor a successor of N. Otherwise, if Op is folded that would
6514 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
6518 bool TryNext = false;
6519 for (SDNode::use_iterator II = BasePtr.getNode()->use_begin(),
6520 EE = BasePtr.getNode()->use_end(); II != EE; ++II) {
6522 if (Use == Ptr.getNode())
6525 // If all the uses are load / store addresses, then don't do the
6527 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
6528 bool RealUse = false;
6529 for (SDNode::use_iterator III = Use->use_begin(),
6530 EEE = Use->use_end(); III != EEE; ++III) {
6531 SDNode *UseUse = *III;
6532 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
6547 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
6548 SDValue Result = isLoad
6549 ? DAG.getIndexedLoad(SDValue(N,0), N->getDebugLoc(),
6550 BasePtr, Offset, AM)
6551 : DAG.getIndexedStore(SDValue(N,0), N->getDebugLoc(),
6552 BasePtr, Offset, AM);
6555 DEBUG(dbgs() << "\nReplacing.5 ";
6557 dbgs() << "\nWith: ";
6558 Result.getNode()->dump(&DAG);
6560 WorkListRemover DeadNodes(*this);
6562 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0),
6564 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2),
6567 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1),
6571 // Finally, since the node is now dead, remove it from the graph.
6574 // Replace the uses of Use with uses of the updated base value.
6575 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
6576 Result.getValue(isLoad ? 1 : 0),
6578 removeFromWorkList(Op);
6588 SDValue DAGCombiner::visitLOAD(SDNode *N) {
6589 LoadSDNode *LD = cast<LoadSDNode>(N);
6590 SDValue Chain = LD->getChain();
6591 SDValue Ptr = LD->getBasePtr();
6593 // If load is not volatile and there are no uses of the loaded value (and
6594 // the updated indexed value in case of indexed loads), change uses of the
6595 // chain value into uses of the chain input (i.e. delete the dead load).
6596 if (!LD->isVolatile()) {
6597 if (N->getValueType(1) == MVT::Other) {
6599 if (!N->hasAnyUseOfValue(0)) {
6600 // It's not safe to use the two value CombineTo variant here. e.g.
6601 // v1, chain2 = load chain1, loc
6602 // v2, chain3 = load chain2, loc
6604 // Now we replace use of chain2 with chain1. This makes the second load
6605 // isomorphic to the one we are deleting, and thus makes this load live.
6606 DEBUG(dbgs() << "\nReplacing.6 ";
6608 dbgs() << "\nWith chain: ";
6609 Chain.getNode()->dump(&DAG);
6611 WorkListRemover DeadNodes(*this);
6612 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain, &DeadNodes);
6614 if (N->use_empty()) {
6615 removeFromWorkList(N);
6619 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6623 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
6624 if (!N->hasAnyUseOfValue(0) && !N->hasAnyUseOfValue(1)) {
6625 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
6626 DEBUG(dbgs() << "\nReplacing.7 ";
6628 dbgs() << "\nWith: ";
6629 Undef.getNode()->dump(&DAG);
6630 dbgs() << " and 2 other values\n");
6631 WorkListRemover DeadNodes(*this);
6632 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef, &DeadNodes);
6633 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1),
6634 DAG.getUNDEF(N->getValueType(1)),
6636 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain, &DeadNodes);
6637 removeFromWorkList(N);
6639 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6644 // If this load is directly stored, replace the load value with the stored
6646 // TODO: Handle store large -> read small portion.
6647 // TODO: Handle TRUNCSTORE/LOADEXT
6648 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
6649 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
6650 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
6651 if (PrevST->getBasePtr() == Ptr &&
6652 PrevST->getValue().getValueType() == N->getValueType(0))
6653 return CombineTo(N, Chain.getOperand(1), Chain);
6657 // Try to infer better alignment information than the load already has.
6658 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
6659 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
6660 if (Align > LD->getAlignment())
6661 return DAG.getExtLoad(LD->getExtensionType(), N->getDebugLoc(),
6662 LD->getValueType(0),
6663 Chain, Ptr, LD->getPointerInfo(),
6665 LD->isVolatile(), LD->isNonTemporal(), Align);
6670 // Walk up chain skipping non-aliasing memory nodes.
6671 SDValue BetterChain = FindBetterChain(N, Chain);
6673 // If there is a better chain.
6674 if (Chain != BetterChain) {
6677 // Replace the chain to void dependency.
6678 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
6679 ReplLoad = DAG.getLoad(N->getValueType(0), LD->getDebugLoc(),
6680 BetterChain, Ptr, LD->getPointerInfo(),
6681 LD->isVolatile(), LD->isNonTemporal(),
6682 LD->isInvariant(), LD->getAlignment());
6684 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), LD->getDebugLoc(),
6685 LD->getValueType(0),
6686 BetterChain, Ptr, LD->getPointerInfo(),
6689 LD->isNonTemporal(),
6690 LD->getAlignment());
6693 // Create token factor to keep old chain connected.
6694 SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
6695 MVT::Other, Chain, ReplLoad.getValue(1));
6697 // Make sure the new and old chains are cleaned up.
6698 AddToWorkList(Token.getNode());
6700 // Replace uses with load result and token factor. Don't add users
6702 return CombineTo(N, ReplLoad.getValue(0), Token, false);
6706 // Try transforming N to an indexed load.
6707 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
6708 return SDValue(N, 0);
6713 /// CheckForMaskedLoad - Check to see if V is (and load (ptr), imm), where the
6714 /// load is having specific bytes cleared out. If so, return the byte size
6715 /// being masked out and the shift amount.
6716 static std::pair<unsigned, unsigned>
6717 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
6718 std::pair<unsigned, unsigned> Result(0, 0);
6720 // Check for the structure we're looking for.
6721 if (V->getOpcode() != ISD::AND ||
6722 !isa<ConstantSDNode>(V->getOperand(1)) ||
6723 !ISD::isNormalLoad(V->getOperand(0).getNode()))
6726 // Check the chain and pointer.
6727 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
6728 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
6730 // The store should be chained directly to the load or be an operand of a
6732 if (LD == Chain.getNode())
6734 else if (Chain->getOpcode() != ISD::TokenFactor)
6735 return Result; // Fail.
6738 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
6739 if (Chain->getOperand(i).getNode() == LD) {
6743 if (!isOk) return Result;
6746 // This only handles simple types.
6747 if (V.getValueType() != MVT::i16 &&
6748 V.getValueType() != MVT::i32 &&
6749 V.getValueType() != MVT::i64)
6752 // Check the constant mask. Invert it so that the bits being masked out are
6753 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
6754 // follow the sign bit for uniformity.
6755 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
6756 unsigned NotMaskLZ = CountLeadingZeros_64(NotMask);
6757 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
6758 unsigned NotMaskTZ = CountTrailingZeros_64(NotMask);
6759 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
6760 if (NotMaskLZ == 64) return Result; // All zero mask.
6762 // See if we have a continuous run of bits. If so, we have 0*1+0*
6763 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
6766 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
6767 if (V.getValueType() != MVT::i64 && NotMaskLZ)
6768 NotMaskLZ -= 64-V.getValueSizeInBits();
6770 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
6771 switch (MaskedBytes) {
6775 default: return Result; // All one mask, or 5-byte mask.
6778 // Verify that the first bit starts at a multiple of mask so that the access
6779 // is aligned the same as the access width.
6780 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
6782 Result.first = MaskedBytes;
6783 Result.second = NotMaskTZ/8;
6788 /// ShrinkLoadReplaceStoreWithStore - Check to see if IVal is something that
6789 /// provides a value as specified by MaskInfo. If so, replace the specified
6790 /// store with a narrower store of truncated IVal.
6792 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
6793 SDValue IVal, StoreSDNode *St,
6795 unsigned NumBytes = MaskInfo.first;
6796 unsigned ByteShift = MaskInfo.second;
6797 SelectionDAG &DAG = DC->getDAG();
6799 // Check to see if IVal is all zeros in the part being masked in by the 'or'
6800 // that uses this. If not, this is not a replacement.
6801 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
6802 ByteShift*8, (ByteShift+NumBytes)*8);
6803 if (!DAG.MaskedValueIsZero(IVal, Mask)) return 0;
6805 // Check that it is legal on the target to do this. It is legal if the new
6806 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
6808 MVT VT = MVT::getIntegerVT(NumBytes*8);
6809 if (!DC->isTypeLegal(VT))
6812 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
6813 // shifted by ByteShift and truncated down to NumBytes.
6815 IVal = DAG.getNode(ISD::SRL, IVal->getDebugLoc(), IVal.getValueType(), IVal,
6816 DAG.getConstant(ByteShift*8,
6817 DC->getShiftAmountTy(IVal.getValueType())));
6819 // Figure out the offset for the store and the alignment of the access.
6821 unsigned NewAlign = St->getAlignment();
6823 if (DAG.getTargetLoweringInfo().isLittleEndian())
6824 StOffset = ByteShift;
6826 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
6828 SDValue Ptr = St->getBasePtr();
6830 Ptr = DAG.getNode(ISD::ADD, IVal->getDebugLoc(), Ptr.getValueType(),
6831 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
6832 NewAlign = MinAlign(NewAlign, StOffset);
6835 // Truncate down to the new size.
6836 IVal = DAG.getNode(ISD::TRUNCATE, IVal->getDebugLoc(), VT, IVal);
6839 return DAG.getStore(St->getChain(), St->getDebugLoc(), IVal, Ptr,
6840 St->getPointerInfo().getWithOffset(StOffset),
6841 false, false, NewAlign).getNode();
6845 /// ReduceLoadOpStoreWidth - Look for sequence of load / op / store where op is
6846 /// one of 'or', 'xor', and 'and' of immediates. If 'op' is only touching some
6847 /// of the loaded bits, try narrowing the load and store if it would end up
6848 /// being a win for performance or code size.
6849 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
6850 StoreSDNode *ST = cast<StoreSDNode>(N);
6851 if (ST->isVolatile())
6854 SDValue Chain = ST->getChain();
6855 SDValue Value = ST->getValue();
6856 SDValue Ptr = ST->getBasePtr();
6857 EVT VT = Value.getValueType();
6859 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
6862 unsigned Opc = Value.getOpcode();
6864 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
6865 // is a byte mask indicating a consecutive number of bytes, check to see if
6866 // Y is known to provide just those bytes. If so, we try to replace the
6867 // load + replace + store sequence with a single (narrower) store, which makes
6869 if (Opc == ISD::OR) {
6870 std::pair<unsigned, unsigned> MaskedLoad;
6871 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
6872 if (MaskedLoad.first)
6873 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
6874 Value.getOperand(1), ST,this))
6875 return SDValue(NewST, 0);
6877 // Or is commutative, so try swapping X and Y.
6878 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
6879 if (MaskedLoad.first)
6880 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
6881 Value.getOperand(0), ST,this))
6882 return SDValue(NewST, 0);
6885 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
6886 Value.getOperand(1).getOpcode() != ISD::Constant)
6889 SDValue N0 = Value.getOperand(0);
6890 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6891 Chain == SDValue(N0.getNode(), 1)) {
6892 LoadSDNode *LD = cast<LoadSDNode>(N0);
6893 if (LD->getBasePtr() != Ptr ||
6894 LD->getPointerInfo().getAddrSpace() !=
6895 ST->getPointerInfo().getAddrSpace())
6898 // Find the type to narrow it the load / op / store to.
6899 SDValue N1 = Value.getOperand(1);
6900 unsigned BitWidth = N1.getValueSizeInBits();
6901 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
6902 if (Opc == ISD::AND)
6903 Imm ^= APInt::getAllOnesValue(BitWidth);
6904 if (Imm == 0 || Imm.isAllOnesValue())
6906 unsigned ShAmt = Imm.countTrailingZeros();
6907 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
6908 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
6909 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
6910 while (NewBW < BitWidth &&
6911 !(TLI.isOperationLegalOrCustom(Opc, NewVT) &&
6912 TLI.isNarrowingProfitable(VT, NewVT))) {
6913 NewBW = NextPowerOf2(NewBW);
6914 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
6916 if (NewBW >= BitWidth)
6919 // If the lsb changed does not start at the type bitwidth boundary,
6920 // start at the previous one.
6922 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
6923 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt, ShAmt + NewBW);
6924 if ((Imm & Mask) == Imm) {
6925 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
6926 if (Opc == ISD::AND)
6927 NewImm ^= APInt::getAllOnesValue(NewBW);
6928 uint64_t PtrOff = ShAmt / 8;
6929 // For big endian targets, we need to adjust the offset to the pointer to
6930 // load the correct bytes.
6931 if (TLI.isBigEndian())
6932 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
6934 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
6935 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
6936 if (NewAlign < TLI.getTargetData()->getABITypeAlignment(NewVTTy))
6939 SDValue NewPtr = DAG.getNode(ISD::ADD, LD->getDebugLoc(),
6940 Ptr.getValueType(), Ptr,
6941 DAG.getConstant(PtrOff, Ptr.getValueType()));
6942 SDValue NewLD = DAG.getLoad(NewVT, N0.getDebugLoc(),
6943 LD->getChain(), NewPtr,
6944 LD->getPointerInfo().getWithOffset(PtrOff),
6945 LD->isVolatile(), LD->isNonTemporal(),
6946 LD->isInvariant(), NewAlign);
6947 SDValue NewVal = DAG.getNode(Opc, Value.getDebugLoc(), NewVT, NewLD,
6948 DAG.getConstant(NewImm, NewVT));
6949 SDValue NewST = DAG.getStore(Chain, N->getDebugLoc(),
6951 ST->getPointerInfo().getWithOffset(PtrOff),
6952 false, false, NewAlign);
6954 AddToWorkList(NewPtr.getNode());
6955 AddToWorkList(NewLD.getNode());
6956 AddToWorkList(NewVal.getNode());
6957 WorkListRemover DeadNodes(*this);
6958 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1),
6968 /// TransformFPLoadStorePair - For a given floating point load / store pair,
6969 /// if the load value isn't used by any other operations, then consider
6970 /// transforming the pair to integer load / store operations if the target
6971 /// deems the transformation profitable.
6972 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
6973 StoreSDNode *ST = cast<StoreSDNode>(N);
6974 SDValue Chain = ST->getChain();
6975 SDValue Value = ST->getValue();
6976 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
6977 Value.hasOneUse() &&
6978 Chain == SDValue(Value.getNode(), 1)) {
6979 LoadSDNode *LD = cast<LoadSDNode>(Value);
6980 EVT VT = LD->getMemoryVT();
6981 if (!VT.isFloatingPoint() ||
6982 VT != ST->getMemoryVT() ||
6983 LD->isNonTemporal() ||
6984 ST->isNonTemporal() ||
6985 LD->getPointerInfo().getAddrSpace() != 0 ||
6986 ST->getPointerInfo().getAddrSpace() != 0)
6989 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
6990 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
6991 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
6992 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
6993 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
6996 unsigned LDAlign = LD->getAlignment();
6997 unsigned STAlign = ST->getAlignment();
6998 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
6999 unsigned ABIAlign = TLI.getTargetData()->getABITypeAlignment(IntVTTy);
7000 if (LDAlign < ABIAlign || STAlign < ABIAlign)
7003 SDValue NewLD = DAG.getLoad(IntVT, Value.getDebugLoc(),
7004 LD->getChain(), LD->getBasePtr(),
7005 LD->getPointerInfo(),
7006 false, false, false, LDAlign);
7008 SDValue NewST = DAG.getStore(NewLD.getValue(1), N->getDebugLoc(),
7009 NewLD, ST->getBasePtr(),
7010 ST->getPointerInfo(),
7011 false, false, STAlign);
7013 AddToWorkList(NewLD.getNode());
7014 AddToWorkList(NewST.getNode());
7015 WorkListRemover DeadNodes(*this);
7016 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1),
7025 SDValue DAGCombiner::visitSTORE(SDNode *N) {
7026 StoreSDNode *ST = cast<StoreSDNode>(N);
7027 SDValue Chain = ST->getChain();
7028 SDValue Value = ST->getValue();
7029 SDValue Ptr = ST->getBasePtr();
7031 // If this is a store of a bit convert, store the input value if the
7032 // resultant store does not need a higher alignment than the original.
7033 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
7034 ST->isUnindexed()) {
7035 unsigned OrigAlign = ST->getAlignment();
7036 EVT SVT = Value.getOperand(0).getValueType();
7037 unsigned Align = TLI.getTargetData()->
7038 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
7039 if (Align <= OrigAlign &&
7040 ((!LegalOperations && !ST->isVolatile()) ||
7041 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
7042 return DAG.getStore(Chain, N->getDebugLoc(), Value.getOperand(0),
7043 Ptr, ST->getPointerInfo(), ST->isVolatile(),
7044 ST->isNonTemporal(), OrigAlign);
7047 // Turn 'store undef, Ptr' -> nothing.
7048 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
7051 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
7052 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
7053 // NOTE: If the original store is volatile, this transform must not increase
7054 // the number of stores. For example, on x86-32 an f64 can be stored in one
7055 // processor operation but an i64 (which is not legal) requires two. So the
7056 // transform should not be done in this case.
7057 if (Value.getOpcode() != ISD::TargetConstantFP) {
7059 switch (CFP->getValueType(0).getSimpleVT().SimpleTy) {
7060 default: llvm_unreachable("Unknown FP type");
7061 case MVT::f80: // We don't do this for these yet.
7066 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
7067 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
7068 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
7069 bitcastToAPInt().getZExtValue(), MVT::i32);
7070 return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
7071 Ptr, ST->getPointerInfo(), ST->isVolatile(),
7072 ST->isNonTemporal(), ST->getAlignment());
7076 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
7077 !ST->isVolatile()) ||
7078 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
7079 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
7080 getZExtValue(), MVT::i64);
7081 return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
7082 Ptr, ST->getPointerInfo(), ST->isVolatile(),
7083 ST->isNonTemporal(), ST->getAlignment());
7086 if (!ST->isVolatile() &&
7087 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
7088 // Many FP stores are not made apparent until after legalize, e.g. for
7089 // argument passing. Since this is so common, custom legalize the
7090 // 64-bit integer store into two 32-bit stores.
7091 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
7092 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
7093 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
7094 if (TLI.isBigEndian()) std::swap(Lo, Hi);
7096 unsigned Alignment = ST->getAlignment();
7097 bool isVolatile = ST->isVolatile();
7098 bool isNonTemporal = ST->isNonTemporal();
7100 SDValue St0 = DAG.getStore(Chain, ST->getDebugLoc(), Lo,
7101 Ptr, ST->getPointerInfo(),
7102 isVolatile, isNonTemporal,
7103 ST->getAlignment());
7104 Ptr = DAG.getNode(ISD::ADD, N->getDebugLoc(), Ptr.getValueType(), Ptr,
7105 DAG.getConstant(4, Ptr.getValueType()));
7106 Alignment = MinAlign(Alignment, 4U);
7107 SDValue St1 = DAG.getStore(Chain, ST->getDebugLoc(), Hi,
7108 Ptr, ST->getPointerInfo().getWithOffset(4),
7109 isVolatile, isNonTemporal,
7111 return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
7120 // Try to infer better alignment information than the store already has.
7121 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
7122 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
7123 if (Align > ST->getAlignment())
7124 return DAG.getTruncStore(Chain, N->getDebugLoc(), Value,
7125 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
7126 ST->isVolatile(), ST->isNonTemporal(), Align);
7130 // Try transforming a pair floating point load / store ops to integer
7131 // load / store ops.
7132 SDValue NewST = TransformFPLoadStorePair(N);
7133 if (NewST.getNode())
7137 // Walk up chain skipping non-aliasing memory nodes.
7138 SDValue BetterChain = FindBetterChain(N, Chain);
7140 // If there is a better chain.
7141 if (Chain != BetterChain) {
7144 // Replace the chain to avoid dependency.
7145 if (ST->isTruncatingStore()) {
7146 ReplStore = DAG.getTruncStore(BetterChain, N->getDebugLoc(), Value, Ptr,
7147 ST->getPointerInfo(),
7148 ST->getMemoryVT(), ST->isVolatile(),
7149 ST->isNonTemporal(), ST->getAlignment());
7151 ReplStore = DAG.getStore(BetterChain, N->getDebugLoc(), Value, Ptr,
7152 ST->getPointerInfo(),
7153 ST->isVolatile(), ST->isNonTemporal(),
7154 ST->getAlignment());
7157 // Create token to keep both nodes around.
7158 SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
7159 MVT::Other, Chain, ReplStore);
7161 // Make sure the new and old chains are cleaned up.
7162 AddToWorkList(Token.getNode());
7164 // Don't add users to work list.
7165 return CombineTo(N, Token, false);
7169 // Try transforming N to an indexed store.
7170 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
7171 return SDValue(N, 0);
7173 // FIXME: is there such a thing as a truncating indexed store?
7174 if (ST->isTruncatingStore() && ST->isUnindexed() &&
7175 Value.getValueType().isInteger()) {
7176 // See if we can simplify the input to this truncstore with knowledge that
7177 // only the low bits are being used. For example:
7178 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
7180 GetDemandedBits(Value,
7181 APInt::getLowBitsSet(
7182 Value.getValueType().getScalarType().getSizeInBits(),
7183 ST->getMemoryVT().getScalarType().getSizeInBits()));
7184 AddToWorkList(Value.getNode());
7185 if (Shorter.getNode())
7186 return DAG.getTruncStore(Chain, N->getDebugLoc(), Shorter,
7187 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
7188 ST->isVolatile(), ST->isNonTemporal(),
7189 ST->getAlignment());
7191 // Otherwise, see if we can simplify the operation with
7192 // SimplifyDemandedBits, which only works if the value has a single use.
7193 if (SimplifyDemandedBits(Value,
7194 APInt::getLowBitsSet(
7195 Value.getValueType().getScalarType().getSizeInBits(),
7196 ST->getMemoryVT().getScalarType().getSizeInBits())))
7197 return SDValue(N, 0);
7200 // If this is a load followed by a store to the same location, then the store
7202 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
7203 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
7204 ST->isUnindexed() && !ST->isVolatile() &&
7205 // There can't be any side effects between the load and store, such as
7207 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
7208 // The store is dead, remove it.
7213 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
7214 // truncating store. We can do this even if this is already a truncstore.
7215 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
7216 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
7217 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
7218 ST->getMemoryVT())) {
7219 return DAG.getTruncStore(Chain, N->getDebugLoc(), Value.getOperand(0),
7220 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
7221 ST->isVolatile(), ST->isNonTemporal(),
7222 ST->getAlignment());
7225 return ReduceLoadOpStoreWidth(N);
7228 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
7229 SDValue InVec = N->getOperand(0);
7230 SDValue InVal = N->getOperand(1);
7231 SDValue EltNo = N->getOperand(2);
7232 DebugLoc dl = N->getDebugLoc();
7234 // If the inserted element is an UNDEF, just use the input vector.
7235 if (InVal.getOpcode() == ISD::UNDEF)
7238 EVT VT = InVec.getValueType();
7240 // If we can't generate a legal BUILD_VECTOR, exit
7241 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
7244 // Check that we know which element is being inserted
7245 if (!isa<ConstantSDNode>(EltNo))
7247 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
7249 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
7250 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
7252 SmallVector<SDValue, 8> Ops;
7253 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
7254 Ops.append(InVec.getNode()->op_begin(),
7255 InVec.getNode()->op_end());
7256 } else if (InVec.getOpcode() == ISD::UNDEF) {
7257 unsigned NElts = VT.getVectorNumElements();
7258 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
7263 // Insert the element
7264 if (Elt < Ops.size()) {
7265 // All the operands of BUILD_VECTOR must have the same type;
7266 // we enforce that here.
7267 EVT OpVT = Ops[0].getValueType();
7268 if (InVal.getValueType() != OpVT)
7269 InVal = OpVT.bitsGT(InVal.getValueType()) ?
7270 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
7271 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
7275 // Return the new vector
7276 return DAG.getNode(ISD::BUILD_VECTOR, dl,
7277 VT, &Ops[0], Ops.size());
7280 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
7281 // (vextract (scalar_to_vector val, 0) -> val
7282 SDValue InVec = N->getOperand(0);
7283 EVT VT = InVec.getValueType();
7284 EVT NVT = N->getValueType(0);
7286 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
7287 // Check if the result type doesn't match the inserted element type. A
7288 // SCALAR_TO_VECTOR may truncate the inserted element and the
7289 // EXTRACT_VECTOR_ELT may widen the extracted vector.
7290 SDValue InOp = InVec.getOperand(0);
7291 if (InOp.getValueType() != NVT) {
7292 assert(InOp.getValueType().isInteger() && NVT.isInteger());
7293 return DAG.getSExtOrTrunc(InOp, InVec.getDebugLoc(), NVT);
7298 SDValue EltNo = N->getOperand(1);
7299 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
7301 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
7302 // We only perform this optimization before the op legalization phase because
7303 // we may introduce new vector instructions which are not backed by TD patterns.
7304 // For example on AVX, extracting elements from a wide vector without using
7305 // extract_subvector.
7306 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
7307 && ConstEltNo && !LegalOperations) {
7308 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
7309 int NumElem = VT.getVectorNumElements();
7310 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
7311 // Find the new index to extract from.
7312 int OrigElt = SVOp->getMaskElt(Elt);
7314 // Extracting an undef index is undef.
7316 return DAG.getUNDEF(NVT);
7318 // Select the right vector half to extract from.
7319 if (OrigElt < NumElem) {
7320 InVec = InVec->getOperand(0);
7322 InVec = InVec->getOperand(1);
7326 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, N->getDebugLoc(), NVT,
7327 InVec, DAG.getConstant(OrigElt, MVT::i32));
7330 // Perform only after legalization to ensure build_vector / vector_shuffle
7331 // optimizations have already been done.
7332 if (!LegalOperations) return SDValue();
7334 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
7335 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
7336 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
7339 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
7340 bool NewLoad = false;
7341 bool BCNumEltsChanged = false;
7342 EVT ExtVT = VT.getVectorElementType();
7345 // If the result of load has to be truncated, then it's not necessarily
7347 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
7350 if (InVec.getOpcode() == ISD::BITCAST) {
7351 // Don't duplicate a load with other uses.
7352 if (!InVec.hasOneUse())
7355 EVT BCVT = InVec.getOperand(0).getValueType();
7356 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
7358 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
7359 BCNumEltsChanged = true;
7360 InVec = InVec.getOperand(0);
7361 ExtVT = BCVT.getVectorElementType();
7365 LoadSDNode *LN0 = NULL;
7366 const ShuffleVectorSDNode *SVN = NULL;
7367 if (ISD::isNormalLoad(InVec.getNode())) {
7368 LN0 = cast<LoadSDNode>(InVec);
7369 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
7370 InVec.getOperand(0).getValueType() == ExtVT &&
7371 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
7372 // Don't duplicate a load with other uses.
7373 if (!InVec.hasOneUse())
7376 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
7377 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
7378 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
7380 // (load $addr+1*size)
7382 // Don't duplicate a load with other uses.
7383 if (!InVec.hasOneUse())
7386 // If the bit convert changed the number of elements, it is unsafe
7387 // to examine the mask.
7388 if (BCNumEltsChanged)
7391 // Select the input vector, guarding against out of range extract vector.
7392 unsigned NumElems = VT.getVectorNumElements();
7393 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
7394 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
7396 if (InVec.getOpcode() == ISD::BITCAST) {
7397 // Don't duplicate a load with other uses.
7398 if (!InVec.hasOneUse())
7401 InVec = InVec.getOperand(0);
7403 if (ISD::isNormalLoad(InVec.getNode())) {
7404 LN0 = cast<LoadSDNode>(InVec);
7405 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
7409 // Make sure we found a non-volatile load and the extractelement is
7411 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
7414 // If Idx was -1 above, Elt is going to be -1, so just return undef.
7416 return DAG.getUNDEF(LVT);
7418 unsigned Align = LN0->getAlignment();
7420 // Check the resultant load doesn't need a higher alignment than the
7424 ->getABITypeAlignment(LVT.getTypeForEVT(*DAG.getContext()));
7426 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, LVT))
7432 SDValue NewPtr = LN0->getBasePtr();
7433 unsigned PtrOff = 0;
7436 PtrOff = LVT.getSizeInBits() * Elt / 8;
7437 EVT PtrType = NewPtr.getValueType();
7438 if (TLI.isBigEndian())
7439 PtrOff = VT.getSizeInBits() / 8 - PtrOff;
7440 NewPtr = DAG.getNode(ISD::ADD, N->getDebugLoc(), PtrType, NewPtr,
7441 DAG.getConstant(PtrOff, PtrType));
7444 // The replacement we need to do here is a little tricky: we need to
7445 // replace an extractelement of a load with a load.
7446 // Use ReplaceAllUsesOfValuesWith to do the replacement.
7447 // Note that this replacement assumes that the extractvalue is the only
7448 // use of the load; that's okay because we don't want to perform this
7449 // transformation in other cases anyway.
7452 if (NVT.bitsGT(LVT)) {
7453 // If the result type of vextract is wider than the load, then issue an
7454 // extending load instead.
7455 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, LVT)
7456 ? ISD::ZEXTLOAD : ISD::EXTLOAD;
7457 Load = DAG.getExtLoad(ExtType, N->getDebugLoc(), NVT, LN0->getChain(),
7458 NewPtr, LN0->getPointerInfo().getWithOffset(PtrOff),
7459 LVT, LN0->isVolatile(), LN0->isNonTemporal(),Align);
7460 Chain = Load.getValue(1);
7462 Load = DAG.getLoad(LVT, N->getDebugLoc(), LN0->getChain(), NewPtr,
7463 LN0->getPointerInfo().getWithOffset(PtrOff),
7464 LN0->isVolatile(), LN0->isNonTemporal(),
7465 LN0->isInvariant(), Align);
7466 Chain = Load.getValue(1);
7467 if (NVT.bitsLT(LVT))
7468 Load = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), NVT, Load);
7470 Load = DAG.getNode(ISD::BITCAST, N->getDebugLoc(), NVT, Load);
7472 WorkListRemover DeadNodes(*this);
7473 SDValue From[] = { SDValue(N, 0), SDValue(LN0,1) };
7474 SDValue To[] = { Load, Chain };
7475 DAG.ReplaceAllUsesOfValuesWith(From, To, 2, &DeadNodes);
7476 // Since we're explcitly calling ReplaceAllUses, add the new node to the
7477 // worklist explicitly as well.
7478 AddToWorkList(Load.getNode());
7479 AddUsersToWorkList(Load.getNode()); // Add users too
7480 // Make sure to revisit this node to clean it up; it will usually be dead.
7482 return SDValue(N, 0);
7488 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
7489 unsigned NumInScalars = N->getNumOperands();
7490 DebugLoc dl = N->getDebugLoc();
7491 EVT VT = N->getValueType(0);
7492 // Check to see if this is a BUILD_VECTOR of a bunch of values
7493 // which come from any_extend or zero_extend nodes. If so, we can create
7494 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
7495 // optimizations. We do not handle sign-extend because we can't fill the sign
7497 EVT SourceType = MVT::Other;
7498 bool AllAnyExt = true;
7499 bool AllUndef = true;
7500 for (unsigned i = 0; i != NumInScalars; ++i) {
7501 SDValue In = N->getOperand(i);
7502 // Ignore undef inputs.
7503 if (In.getOpcode() == ISD::UNDEF) continue;
7506 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
7507 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
7509 // Abort if the element is not an extension.
7510 if (!ZeroExt && !AnyExt) {
7511 SourceType = MVT::Other;
7515 // The input is a ZeroExt or AnyExt. Check the original type.
7516 EVT InTy = In.getOperand(0).getValueType();
7518 // Check that all of the widened source types are the same.
7519 if (SourceType == MVT::Other)
7522 else if (InTy != SourceType) {
7523 // Multiple income types. Abort.
7524 SourceType = MVT::Other;
7528 // Check if all of the extends are ANY_EXTENDs.
7529 AllAnyExt &= AnyExt;
7533 return DAG.getUNDEF(VT);
7535 // In order to have valid types, all of the inputs must be extended from the
7536 // same source type and all of the inputs must be any or zero extend.
7537 // Scalar sizes must be a power of two.
7538 EVT OutScalarTy = N->getValueType(0).getScalarType();
7539 bool ValidTypes = SourceType != MVT::Other &&
7540 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
7541 isPowerOf2_32(SourceType.getSizeInBits());
7543 // We perform this optimization post type-legalization because
7544 // the type-legalizer often scalarizes integer-promoted vectors.
7545 // Performing this optimization before may create bit-casts which
7546 // will be type-legalized to complex code sequences.
7547 // We perform this optimization only before the operation legalizer because we
7548 // may introduce illegal operations.
7549 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
7550 // turn into a single shuffle instruction.
7551 if ((Level == AfterLegalizeVectorOps || Level == AfterLegalizeTypes) &&
7553 bool isLE = TLI.isLittleEndian();
7554 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
7555 assert(ElemRatio > 1 && "Invalid element size ratio");
7556 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
7557 DAG.getConstant(0, SourceType);
7559 unsigned NewBVElems = ElemRatio * N->getValueType(0).getVectorNumElements();
7560 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
7562 // Populate the new build_vector
7563 for (unsigned i=0; i < N->getNumOperands(); ++i) {
7564 SDValue Cast = N->getOperand(i);
7565 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
7566 Cast.getOpcode() == ISD::ZERO_EXTEND ||
7567 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
7569 if (Cast.getOpcode() == ISD::UNDEF)
7570 In = DAG.getUNDEF(SourceType);
7572 In = Cast->getOperand(0);
7573 unsigned Index = isLE ? (i * ElemRatio) :
7574 (i * ElemRatio + (ElemRatio - 1));
7576 assert(Index < Ops.size() && "Invalid index");
7580 // The type of the new BUILD_VECTOR node.
7581 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
7582 assert(VecVT.getSizeInBits() == N->getValueType(0).getSizeInBits() &&
7583 "Invalid vector size");
7584 // Check if the new vector type is legal.
7585 if (!isTypeLegal(VecVT)) return SDValue();
7587 // Make the new BUILD_VECTOR.
7588 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
7589 VecVT, &Ops[0], Ops.size());
7591 // The new BUILD_VECTOR node has the potential to be further optimized.
7592 AddToWorkList(BV.getNode());
7593 // Bitcast to the desired type.
7594 return DAG.getNode(ISD::BITCAST, dl, N->getValueType(0), BV);
7597 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
7598 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
7599 // at most two distinct vectors, turn this into a shuffle node.
7601 // May only combine to shuffle after legalize if shuffle is legal.
7602 if (LegalOperations &&
7603 !TLI.isOperationLegalOrCustom(ISD::VECTOR_SHUFFLE, VT))
7606 SDValue VecIn1, VecIn2;
7607 for (unsigned i = 0; i != NumInScalars; ++i) {
7608 // Ignore undef inputs.
7609 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
7611 // If this input is something other than a EXTRACT_VECTOR_ELT with a
7612 // constant index, bail out.
7613 if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
7614 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
7615 VecIn1 = VecIn2 = SDValue(0, 0);
7619 // We allow up to two distinct input vectors.
7620 SDValue ExtractedFromVec = N->getOperand(i).getOperand(0);
7621 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
7624 if (VecIn1.getNode() == 0) {
7625 VecIn1 = ExtractedFromVec;
7626 } else if (VecIn2.getNode() == 0) {
7627 VecIn2 = ExtractedFromVec;
7630 VecIn1 = VecIn2 = SDValue(0, 0);
7635 // If everything is good, we can make a shuffle operation.
7636 if (VecIn1.getNode()) {
7637 SmallVector<int, 8> Mask;
7638 for (unsigned i = 0; i != NumInScalars; ++i) {
7639 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
7644 // If extracting from the first vector, just use the index directly.
7645 SDValue Extract = N->getOperand(i);
7646 SDValue ExtVal = Extract.getOperand(1);
7647 if (Extract.getOperand(0) == VecIn1) {
7648 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
7649 if (ExtIndex > VT.getVectorNumElements())
7652 Mask.push_back(ExtIndex);
7656 // Otherwise, use InIdx + VecSize
7657 unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue();
7658 Mask.push_back(Idx+NumInScalars);
7661 // We can't generate a shuffle node with mismatched input and output types.
7662 // Attempt to transform a single input vector to the correct type.
7663 if ((VT != VecIn1.getValueType())) {
7664 // We don't support shuffeling between TWO values of different types.
7665 if (VecIn2.getNode() != 0)
7668 // We only support widening of vectors which are half the size of the
7669 // output registers. For example XMM->YMM widening on X86 with AVX.
7670 if (VecIn1.getValueType().getSizeInBits()*2 != VT.getSizeInBits())
7673 // Widen the input vector by adding undef values.
7674 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
7675 VecIn1, DAG.getUNDEF(VecIn1.getValueType()));
7678 // If VecIn2 is unused then change it to undef.
7679 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
7681 // Check that we were able to transform all incoming values to the same type.
7682 if (VecIn2.getValueType() != VecIn1.getValueType() ||
7683 VecIn1.getValueType() != VT)
7686 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
7687 if (!isTypeLegal(VT))
7690 // Return the new VECTOR_SHUFFLE node.
7694 return DAG.getVectorShuffle(VT, N->getDebugLoc(), Ops[0], Ops[1], &Mask[0]);
7700 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
7701 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
7702 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
7703 // inputs come from at most two distinct vectors, turn this into a shuffle
7706 // If we only have one input vector, we don't need to do any concatenation.
7707 if (N->getNumOperands() == 1)
7708 return N->getOperand(0);
7713 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
7714 EVT NVT = N->getValueType(0);
7715 SDValue V = N->getOperand(0);
7717 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
7718 // Handle only simple case where vector being inserted and vector
7719 // being extracted are of same type, and are half size of larger vectors.
7720 EVT BigVT = V->getOperand(0).getValueType();
7721 EVT SmallVT = V->getOperand(1).getValueType();
7722 if (NVT != SmallVT || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
7725 // Only handle cases where both indexes are constants with the same type.
7726 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
7727 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
7729 if (InsIdx && ExtIdx &&
7730 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
7731 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
7733 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
7735 // indices are equal => V1
7736 // otherwise => (extract_subvec V1, ExtIdx)
7737 if (InsIdx->getZExtValue() == ExtIdx->getZExtValue())
7738 return V->getOperand(1);
7739 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, N->getDebugLoc(), NVT,
7740 V->getOperand(0), N->getOperand(1));
7747 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
7748 EVT VT = N->getValueType(0);
7749 unsigned NumElts = VT.getVectorNumElements();
7751 SDValue N0 = N->getOperand(0);
7752 SDValue N1 = N->getOperand(1);
7754 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
7756 // Canonicalize shuffle undef, undef -> undef
7757 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
7758 return DAG.getUNDEF(VT);
7760 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7762 // Canonicalize shuffle v, v -> v, undef
7764 SmallVector<int, 8> NewMask;
7765 for (unsigned i = 0; i != NumElts; ++i) {
7766 int Idx = SVN->getMaskElt(i);
7767 if (Idx >= (int)NumElts) Idx -= NumElts;
7768 NewMask.push_back(Idx);
7770 return DAG.getVectorShuffle(VT, N->getDebugLoc(), N0, DAG.getUNDEF(VT),
7774 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
7775 if (N0.getOpcode() == ISD::UNDEF) {
7776 SmallVector<int, 8> NewMask;
7777 for (unsigned i = 0; i != NumElts; ++i) {
7778 int Idx = SVN->getMaskElt(i);
7780 if (Idx < (int)NumElts)
7785 NewMask.push_back(Idx);
7787 return DAG.getVectorShuffle(VT, N->getDebugLoc(), N1, DAG.getUNDEF(VT),
7791 // Remove references to rhs if it is undef
7792 if (N1.getOpcode() == ISD::UNDEF) {
7793 bool Changed = false;
7794 SmallVector<int, 8> NewMask;
7795 for (unsigned i = 0; i != NumElts; ++i) {
7796 int Idx = SVN->getMaskElt(i);
7797 if (Idx >= (int)NumElts) {
7801 NewMask.push_back(Idx);
7804 return DAG.getVectorShuffle(VT, N->getDebugLoc(), N0, N1, &NewMask[0]);
7807 // If it is a splat, check if the argument vector is another splat or a
7808 // build_vector with all scalar elements the same.
7809 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
7810 SDNode *V = N0.getNode();
7812 // If this is a bit convert that changes the element type of the vector but
7813 // not the number of vector elements, look through it. Be careful not to
7814 // look though conversions that change things like v4f32 to v2f64.
7815 if (V->getOpcode() == ISD::BITCAST) {
7816 SDValue ConvInput = V->getOperand(0);
7817 if (ConvInput.getValueType().isVector() &&
7818 ConvInput.getValueType().getVectorNumElements() == NumElts)
7819 V = ConvInput.getNode();
7822 if (V->getOpcode() == ISD::BUILD_VECTOR) {
7823 assert(V->getNumOperands() == NumElts &&
7824 "BUILD_VECTOR has wrong number of operands");
7826 bool AllSame = true;
7827 for (unsigned i = 0; i != NumElts; ++i) {
7828 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
7829 Base = V->getOperand(i);
7833 // Splat of <u, u, u, u>, return <u, u, u, u>
7834 if (!Base.getNode())
7836 for (unsigned i = 0; i != NumElts; ++i) {
7837 if (V->getOperand(i) != Base) {
7842 // Splat of <x, x, x, x>, return <x, x, x, x>
7848 // If this shuffle node is simply a swizzle of another shuffle node,
7849 // and it reverses the swizzle of the previous shuffle then we can
7850 // optimize shuffle(shuffle(x, undef), undef) -> x.
7851 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
7852 N1.getOpcode() == ISD::UNDEF) {
7854 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
7856 // Shuffle nodes can only reverse shuffles with a single non-undef value.
7857 if (N0.getOperand(1).getOpcode() != ISD::UNDEF)
7860 // The incoming shuffle must be of the same type as the result of the
7862 assert(OtherSV->getOperand(0).getValueType() == VT &&
7863 "Shuffle types don't match");
7865 for (unsigned i = 0; i != NumElts; ++i) {
7866 int Idx = SVN->getMaskElt(i);
7867 assert(Idx < (int)NumElts && "Index references undef operand");
7868 // Next, this index comes from the first value, which is the incoming
7869 // shuffle. Adopt the incoming index.
7871 Idx = OtherSV->getMaskElt(Idx);
7873 // The combined shuffle must map each index to itself.
7874 if (Idx >= 0 && (unsigned)Idx != i)
7878 return OtherSV->getOperand(0);
7884 SDValue DAGCombiner::visitMEMBARRIER(SDNode* N) {
7885 if (!TLI.getShouldFoldAtomicFences())
7888 SDValue atomic = N->getOperand(0);
7889 switch (atomic.getOpcode()) {
7890 case ISD::ATOMIC_CMP_SWAP:
7891 case ISD::ATOMIC_SWAP:
7892 case ISD::ATOMIC_LOAD_ADD:
7893 case ISD::ATOMIC_LOAD_SUB:
7894 case ISD::ATOMIC_LOAD_AND:
7895 case ISD::ATOMIC_LOAD_OR:
7896 case ISD::ATOMIC_LOAD_XOR:
7897 case ISD::ATOMIC_LOAD_NAND:
7898 case ISD::ATOMIC_LOAD_MIN:
7899 case ISD::ATOMIC_LOAD_MAX:
7900 case ISD::ATOMIC_LOAD_UMIN:
7901 case ISD::ATOMIC_LOAD_UMAX:
7907 SDValue fence = atomic.getOperand(0);
7908 if (fence.getOpcode() != ISD::MEMBARRIER)
7911 switch (atomic.getOpcode()) {
7912 case ISD::ATOMIC_CMP_SWAP:
7913 return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
7914 fence.getOperand(0),
7915 atomic.getOperand(1), atomic.getOperand(2),
7916 atomic.getOperand(3)), atomic.getResNo());
7917 case ISD::ATOMIC_SWAP:
7918 case ISD::ATOMIC_LOAD_ADD:
7919 case ISD::ATOMIC_LOAD_SUB:
7920 case ISD::ATOMIC_LOAD_AND:
7921 case ISD::ATOMIC_LOAD_OR:
7922 case ISD::ATOMIC_LOAD_XOR:
7923 case ISD::ATOMIC_LOAD_NAND:
7924 case ISD::ATOMIC_LOAD_MIN:
7925 case ISD::ATOMIC_LOAD_MAX:
7926 case ISD::ATOMIC_LOAD_UMIN:
7927 case ISD::ATOMIC_LOAD_UMAX:
7928 return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
7929 fence.getOperand(0),
7930 atomic.getOperand(1), atomic.getOperand(2)),
7937 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
7938 /// an AND to a vector_shuffle with the destination vector and a zero vector.
7939 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
7940 /// vector_shuffle V, Zero, <0, 4, 2, 4>
7941 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
7942 EVT VT = N->getValueType(0);
7943 DebugLoc dl = N->getDebugLoc();
7944 SDValue LHS = N->getOperand(0);
7945 SDValue RHS = N->getOperand(1);
7946 if (N->getOpcode() == ISD::AND) {
7947 if (RHS.getOpcode() == ISD::BITCAST)
7948 RHS = RHS.getOperand(0);
7949 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
7950 SmallVector<int, 8> Indices;
7951 unsigned NumElts = RHS.getNumOperands();
7952 for (unsigned i = 0; i != NumElts; ++i) {
7953 SDValue Elt = RHS.getOperand(i);
7954 if (!isa<ConstantSDNode>(Elt))
7957 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
7958 Indices.push_back(i);
7959 else if (cast<ConstantSDNode>(Elt)->isNullValue())
7960 Indices.push_back(NumElts);
7965 // Let's see if the target supports this vector_shuffle.
7966 EVT RVT = RHS.getValueType();
7967 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
7970 // Return the new VECTOR_SHUFFLE node.
7971 EVT EltVT = RVT.getVectorElementType();
7972 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
7973 DAG.getConstant(0, EltVT));
7974 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
7975 RVT, &ZeroOps[0], ZeroOps.size());
7976 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
7977 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
7978 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
7985 /// SimplifyVBinOp - Visit a binary vector operation, like ADD.
7986 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
7987 // After legalize, the target may be depending on adds and other
7988 // binary ops to provide legal ways to construct constants or other
7989 // things. Simplifying them may result in a loss of legality.
7990 if (LegalOperations) return SDValue();
7992 assert(N->getValueType(0).isVector() &&
7993 "SimplifyVBinOp only works on vectors!");
7995 SDValue LHS = N->getOperand(0);
7996 SDValue RHS = N->getOperand(1);
7997 SDValue Shuffle = XformToShuffleWithZero(N);
7998 if (Shuffle.getNode()) return Shuffle;
8000 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
8002 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
8003 RHS.getOpcode() == ISD::BUILD_VECTOR) {
8004 SmallVector<SDValue, 8> Ops;
8005 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
8006 SDValue LHSOp = LHS.getOperand(i);
8007 SDValue RHSOp = RHS.getOperand(i);
8008 // If these two elements can't be folded, bail out.
8009 if ((LHSOp.getOpcode() != ISD::UNDEF &&
8010 LHSOp.getOpcode() != ISD::Constant &&
8011 LHSOp.getOpcode() != ISD::ConstantFP) ||
8012 (RHSOp.getOpcode() != ISD::UNDEF &&
8013 RHSOp.getOpcode() != ISD::Constant &&
8014 RHSOp.getOpcode() != ISD::ConstantFP))
8017 // Can't fold divide by zero.
8018 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
8019 N->getOpcode() == ISD::FDIV) {
8020 if ((RHSOp.getOpcode() == ISD::Constant &&
8021 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
8022 (RHSOp.getOpcode() == ISD::ConstantFP &&
8023 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
8027 EVT VT = LHSOp.getValueType();
8028 EVT RVT = RHSOp.getValueType();
8030 // Integer BUILD_VECTOR operands may have types larger than the element
8031 // size (e.g., when the element type is not legal). Prior to type
8032 // legalization, the types may not match between the two BUILD_VECTORS.
8033 // Truncate one of the operands to make them match.
8034 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
8035 RHSOp = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, RHSOp);
8037 LHSOp = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), RVT, LHSOp);
8041 SDValue FoldOp = DAG.getNode(N->getOpcode(), LHS.getDebugLoc(), VT,
8043 if (FoldOp.getOpcode() != ISD::UNDEF &&
8044 FoldOp.getOpcode() != ISD::Constant &&
8045 FoldOp.getOpcode() != ISD::ConstantFP)
8047 Ops.push_back(FoldOp);
8048 AddToWorkList(FoldOp.getNode());
8051 if (Ops.size() == LHS.getNumOperands())
8052 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
8053 LHS.getValueType(), &Ops[0], Ops.size());
8059 SDValue DAGCombiner::SimplifySelect(DebugLoc DL, SDValue N0,
8060 SDValue N1, SDValue N2){
8061 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
8063 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
8064 cast<CondCodeSDNode>(N0.getOperand(2))->get());
8066 // If we got a simplified select_cc node back from SimplifySelectCC, then
8067 // break it down into a new SETCC node, and a new SELECT node, and then return
8068 // the SELECT node, since we were called with a SELECT node.
8069 if (SCC.getNode()) {
8070 // Check to see if we got a select_cc back (to turn into setcc/select).
8071 // Otherwise, just return whatever node we got back, like fabs.
8072 if (SCC.getOpcode() == ISD::SELECT_CC) {
8073 SDValue SETCC = DAG.getNode(ISD::SETCC, N0.getDebugLoc(),
8075 SCC.getOperand(0), SCC.getOperand(1),
8077 AddToWorkList(SETCC.getNode());
8078 return DAG.getNode(ISD::SELECT, SCC.getDebugLoc(), SCC.getValueType(),
8079 SCC.getOperand(2), SCC.getOperand(3), SETCC);
8087 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
8088 /// are the two values being selected between, see if we can simplify the
8089 /// select. Callers of this should assume that TheSelect is deleted if this
8090 /// returns true. As such, they should return the appropriate thing (e.g. the
8091 /// node) back to the top-level of the DAG combiner loop to avoid it being
8093 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
8096 // Cannot simplify select with vector condition
8097 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
8099 // If this is a select from two identical things, try to pull the operation
8100 // through the select.
8101 if (LHS.getOpcode() != RHS.getOpcode() ||
8102 !LHS.hasOneUse() || !RHS.hasOneUse())
8105 // If this is a load and the token chain is identical, replace the select
8106 // of two loads with a load through a select of the address to load from.
8107 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
8108 // constants have been dropped into the constant pool.
8109 if (LHS.getOpcode() == ISD::LOAD) {
8110 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
8111 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
8113 // Token chains must be identical.
8114 if (LHS.getOperand(0) != RHS.getOperand(0) ||
8115 // Do not let this transformation reduce the number of volatile loads.
8116 LLD->isVolatile() || RLD->isVolatile() ||
8117 // If this is an EXTLOAD, the VT's must match.
8118 LLD->getMemoryVT() != RLD->getMemoryVT() ||
8119 // If this is an EXTLOAD, the kind of extension must match.
8120 (LLD->getExtensionType() != RLD->getExtensionType() &&
8121 // The only exception is if one of the extensions is anyext.
8122 LLD->getExtensionType() != ISD::EXTLOAD &&
8123 RLD->getExtensionType() != ISD::EXTLOAD) ||
8124 // FIXME: this discards src value information. This is
8125 // over-conservative. It would be beneficial to be able to remember
8126 // both potential memory locations. Since we are discarding
8127 // src value info, don't do the transformation if the memory
8128 // locations are not in the default address space.
8129 LLD->getPointerInfo().getAddrSpace() != 0 ||
8130 RLD->getPointerInfo().getAddrSpace() != 0)
8133 // Check that the select condition doesn't reach either load. If so,
8134 // folding this will induce a cycle into the DAG. If not, this is safe to
8135 // xform, so create a select of the addresses.
8137 if (TheSelect->getOpcode() == ISD::SELECT) {
8138 SDNode *CondNode = TheSelect->getOperand(0).getNode();
8139 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
8140 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
8142 Addr = DAG.getNode(ISD::SELECT, TheSelect->getDebugLoc(),
8143 LLD->getBasePtr().getValueType(),
8144 TheSelect->getOperand(0), LLD->getBasePtr(),
8146 } else { // Otherwise SELECT_CC
8147 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
8148 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
8150 if ((LLD->hasAnyUseOfValue(1) &&
8151 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
8152 (RLD->hasAnyUseOfValue(1) &&
8153 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
8156 Addr = DAG.getNode(ISD::SELECT_CC, TheSelect->getDebugLoc(),
8157 LLD->getBasePtr().getValueType(),
8158 TheSelect->getOperand(0),
8159 TheSelect->getOperand(1),
8160 LLD->getBasePtr(), RLD->getBasePtr(),
8161 TheSelect->getOperand(4));
8165 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
8166 Load = DAG.getLoad(TheSelect->getValueType(0),
8167 TheSelect->getDebugLoc(),
8168 // FIXME: Discards pointer info.
8169 LLD->getChain(), Addr, MachinePointerInfo(),
8170 LLD->isVolatile(), LLD->isNonTemporal(),
8171 LLD->isInvariant(), LLD->getAlignment());
8173 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
8174 RLD->getExtensionType() : LLD->getExtensionType(),
8175 TheSelect->getDebugLoc(),
8176 TheSelect->getValueType(0),
8177 // FIXME: Discards pointer info.
8178 LLD->getChain(), Addr, MachinePointerInfo(),
8179 LLD->getMemoryVT(), LLD->isVolatile(),
8180 LLD->isNonTemporal(), LLD->getAlignment());
8183 // Users of the select now use the result of the load.
8184 CombineTo(TheSelect, Load);
8186 // Users of the old loads now use the new load's chain. We know the
8187 // old-load value is dead now.
8188 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
8189 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
8196 /// SimplifySelectCC - Simplify an expression of the form (N0 cond N1) ? N2 : N3
8197 /// where 'cond' is the comparison specified by CC.
8198 SDValue DAGCombiner::SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1,
8199 SDValue N2, SDValue N3,
8200 ISD::CondCode CC, bool NotExtCompare) {
8201 // (x ? y : y) -> y.
8202 if (N2 == N3) return N2;
8204 EVT VT = N2.getValueType();
8205 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
8206 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
8207 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
8209 // Determine if the condition we're dealing with is constant
8210 SDValue SCC = SimplifySetCC(TLI.getSetCCResultType(N0.getValueType()),
8211 N0, N1, CC, DL, false);
8212 if (SCC.getNode()) AddToWorkList(SCC.getNode());
8213 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
8215 // fold select_cc true, x, y -> x
8216 if (SCCC && !SCCC->isNullValue())
8218 // fold select_cc false, x, y -> y
8219 if (SCCC && SCCC->isNullValue())
8222 // Check to see if we can simplify the select into an fabs node
8223 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
8224 // Allow either -0.0 or 0.0
8225 if (CFP->getValueAPF().isZero()) {
8226 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
8227 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
8228 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
8229 N2 == N3.getOperand(0))
8230 return DAG.getNode(ISD::FABS, DL, VT, N0);
8232 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
8233 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
8234 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
8235 N2.getOperand(0) == N3)
8236 return DAG.getNode(ISD::FABS, DL, VT, N3);
8240 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
8241 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
8242 // in it. This is a win when the constant is not otherwise available because
8243 // it replaces two constant pool loads with one. We only do this if the FP
8244 // type is known to be legal, because if it isn't, then we are before legalize
8245 // types an we want the other legalization to happen first (e.g. to avoid
8246 // messing with soft float) and if the ConstantFP is not legal, because if
8247 // it is legal, we may not need to store the FP constant in a constant pool.
8248 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
8249 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
8250 if (TLI.isTypeLegal(N2.getValueType()) &&
8251 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
8252 TargetLowering::Legal) &&
8253 // If both constants have multiple uses, then we won't need to do an
8254 // extra load, they are likely around in registers for other users.
8255 (TV->hasOneUse() || FV->hasOneUse())) {
8256 Constant *Elts[] = {
8257 const_cast<ConstantFP*>(FV->getConstantFPValue()),
8258 const_cast<ConstantFP*>(TV->getConstantFPValue())
8260 Type *FPTy = Elts[0]->getType();
8261 const TargetData &TD = *TLI.getTargetData();
8263 // Create a ConstantArray of the two constants.
8264 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
8265 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
8266 TD.getPrefTypeAlignment(FPTy));
8267 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
8269 // Get the offsets to the 0 and 1 element of the array so that we can
8270 // select between them.
8271 SDValue Zero = DAG.getIntPtrConstant(0);
8272 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
8273 SDValue One = DAG.getIntPtrConstant(EltSize);
8275 SDValue Cond = DAG.getSetCC(DL,
8276 TLI.getSetCCResultType(N0.getValueType()),
8278 AddToWorkList(Cond.getNode());
8279 SDValue CstOffset = DAG.getNode(ISD::SELECT, DL, Zero.getValueType(),
8281 AddToWorkList(CstOffset.getNode());
8282 CPIdx = DAG.getNode(ISD::ADD, DL, TLI.getPointerTy(), CPIdx,
8284 AddToWorkList(CPIdx.getNode());
8285 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
8286 MachinePointerInfo::getConstantPool(), false,
8287 false, false, Alignment);
8292 // Check to see if we can perform the "gzip trick", transforming
8293 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
8294 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
8295 (N1C->isNullValue() || // (a < 0) ? b : 0
8296 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
8297 EVT XType = N0.getValueType();
8298 EVT AType = N2.getValueType();
8299 if (XType.bitsGE(AType)) {
8300 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
8301 // single-bit constant.
8302 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
8303 unsigned ShCtV = N2C->getAPIntValue().logBase2();
8304 ShCtV = XType.getSizeInBits()-ShCtV-1;
8305 SDValue ShCt = DAG.getConstant(ShCtV,
8306 getShiftAmountTy(N0.getValueType()));
8307 SDValue Shift = DAG.getNode(ISD::SRL, N0.getDebugLoc(),
8309 AddToWorkList(Shift.getNode());
8311 if (XType.bitsGT(AType)) {
8312 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
8313 AddToWorkList(Shift.getNode());
8316 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
8319 SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(),
8321 DAG.getConstant(XType.getSizeInBits()-1,
8322 getShiftAmountTy(N0.getValueType())));
8323 AddToWorkList(Shift.getNode());
8325 if (XType.bitsGT(AType)) {
8326 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
8327 AddToWorkList(Shift.getNode());
8330 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
8334 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
8335 // where y is has a single bit set.
8336 // A plaintext description would be, we can turn the SELECT_CC into an AND
8337 // when the condition can be materialized as an all-ones register. Any
8338 // single bit-test can be materialized as an all-ones register with
8339 // shift-left and shift-right-arith.
8340 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
8341 N0->getValueType(0) == VT &&
8342 N1C && N1C->isNullValue() &&
8343 N2C && N2C->isNullValue()) {
8344 SDValue AndLHS = N0->getOperand(0);
8345 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
8346 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
8347 // Shift the tested bit over the sign bit.
8348 APInt AndMask = ConstAndRHS->getAPIntValue();
8350 DAG.getConstant(AndMask.countLeadingZeros(),
8351 getShiftAmountTy(AndLHS.getValueType()));
8352 SDValue Shl = DAG.getNode(ISD::SHL, N0.getDebugLoc(), VT, AndLHS, ShlAmt);
8354 // Now arithmetic right shift it all the way over, so the result is either
8355 // all-ones, or zero.
8357 DAG.getConstant(AndMask.getBitWidth()-1,
8358 getShiftAmountTy(Shl.getValueType()));
8359 SDValue Shr = DAG.getNode(ISD::SRA, N0.getDebugLoc(), VT, Shl, ShrAmt);
8361 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
8365 // fold select C, 16, 0 -> shl C, 4
8366 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
8367 TLI.getBooleanContents(N0.getValueType().isVector()) ==
8368 TargetLowering::ZeroOrOneBooleanContent) {
8370 // If the caller doesn't want us to simplify this into a zext of a compare,
8372 if (NotExtCompare && N2C->getAPIntValue() == 1)
8375 // Get a SetCC of the condition
8376 // FIXME: Should probably make sure that setcc is legal if we ever have a
8377 // target where it isn't.
8379 // cast from setcc result type to select result type
8381 SCC = DAG.getSetCC(DL, TLI.getSetCCResultType(N0.getValueType()),
8383 if (N2.getValueType().bitsLT(SCC.getValueType()))
8384 Temp = DAG.getZeroExtendInReg(SCC, N2.getDebugLoc(), N2.getValueType());
8386 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getDebugLoc(),
8387 N2.getValueType(), SCC);
8389 SCC = DAG.getSetCC(N0.getDebugLoc(), MVT::i1, N0, N1, CC);
8390 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getDebugLoc(),
8391 N2.getValueType(), SCC);
8394 AddToWorkList(SCC.getNode());
8395 AddToWorkList(Temp.getNode());
8397 if (N2C->getAPIntValue() == 1)
8400 // shl setcc result by log2 n2c
8401 return DAG.getNode(ISD::SHL, DL, N2.getValueType(), Temp,
8402 DAG.getConstant(N2C->getAPIntValue().logBase2(),
8403 getShiftAmountTy(Temp.getValueType())));
8406 // Check to see if this is the equivalent of setcc
8407 // FIXME: Turn all of these into setcc if setcc if setcc is legal
8408 // otherwise, go ahead with the folds.
8409 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
8410 EVT XType = N0.getValueType();
8411 if (!LegalOperations ||
8412 TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(XType))) {
8413 SDValue Res = DAG.getSetCC(DL, TLI.getSetCCResultType(XType), N0, N1, CC);
8414 if (Res.getValueType() != VT)
8415 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
8419 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
8420 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
8421 (!LegalOperations ||
8422 TLI.isOperationLegal(ISD::CTLZ, XType))) {
8423 SDValue Ctlz = DAG.getNode(ISD::CTLZ, N0.getDebugLoc(), XType, N0);
8424 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
8425 DAG.getConstant(Log2_32(XType.getSizeInBits()),
8426 getShiftAmountTy(Ctlz.getValueType())));
8428 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
8429 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
8430 SDValue NegN0 = DAG.getNode(ISD::SUB, N0.getDebugLoc(),
8431 XType, DAG.getConstant(0, XType), N0);
8432 SDValue NotN0 = DAG.getNOT(N0.getDebugLoc(), N0, XType);
8433 return DAG.getNode(ISD::SRL, DL, XType,
8434 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
8435 DAG.getConstant(XType.getSizeInBits()-1,
8436 getShiftAmountTy(XType)));
8438 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
8439 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
8440 SDValue Sign = DAG.getNode(ISD::SRL, N0.getDebugLoc(), XType, N0,
8441 DAG.getConstant(XType.getSizeInBits()-1,
8442 getShiftAmountTy(N0.getValueType())));
8443 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
8447 // Check to see if this is an integer abs.
8448 // select_cc setg[te] X, 0, X, -X ->
8449 // select_cc setgt X, -1, X, -X ->
8450 // select_cc setl[te] X, 0, -X, X ->
8451 // select_cc setlt X, 1, -X, X ->
8452 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
8454 ConstantSDNode *SubC = NULL;
8455 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
8456 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
8457 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
8458 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
8459 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
8460 (N1C->isOne() && CC == ISD::SETLT)) &&
8461 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
8462 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
8464 EVT XType = N0.getValueType();
8465 if (SubC && SubC->isNullValue() && XType.isInteger()) {
8466 SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(), XType,
8468 DAG.getConstant(XType.getSizeInBits()-1,
8469 getShiftAmountTy(N0.getValueType())));
8470 SDValue Add = DAG.getNode(ISD::ADD, N0.getDebugLoc(),
8472 AddToWorkList(Shift.getNode());
8473 AddToWorkList(Add.getNode());
8474 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
8481 /// SimplifySetCC - This is a stub for TargetLowering::SimplifySetCC.
8482 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
8483 SDValue N1, ISD::CondCode Cond,
8484 DebugLoc DL, bool foldBooleans) {
8485 TargetLowering::DAGCombinerInfo
8486 DagCombineInfo(DAG, !LegalTypes, !LegalOperations, false, this);
8487 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
8490 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
8491 /// return a DAG expression to select that will generate the same value by
8492 /// multiplying by a magic number. See:
8493 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
8494 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
8495 std::vector<SDNode*> Built;
8496 SDValue S = TLI.BuildSDIV(N, DAG, LegalOperations, &Built);
8498 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
8504 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
8505 /// return a DAG expression to select that will generate the same value by
8506 /// multiplying by a magic number. See:
8507 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
8508 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
8509 std::vector<SDNode*> Built;
8510 SDValue S = TLI.BuildUDIV(N, DAG, LegalOperations, &Built);
8512 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
8518 /// FindBaseOffset - Return true if base is a frame index, which is known not
8519 // to alias with anything but itself. Provides base object and offset as
8521 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
8522 const GlobalValue *&GV, void *&CV) {
8523 // Assume it is a primitive operation.
8524 Base = Ptr; Offset = 0; GV = 0; CV = 0;
8526 // If it's an adding a simple constant then integrate the offset.
8527 if (Base.getOpcode() == ISD::ADD) {
8528 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
8529 Base = Base.getOperand(0);
8530 Offset += C->getZExtValue();
8534 // Return the underlying GlobalValue, and update the Offset. Return false
8535 // for GlobalAddressSDNode since the same GlobalAddress may be represented
8536 // by multiple nodes with different offsets.
8537 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
8538 GV = G->getGlobal();
8539 Offset += G->getOffset();
8543 // Return the underlying Constant value, and update the Offset. Return false
8544 // for ConstantSDNodes since the same constant pool entry may be represented
8545 // by multiple nodes with different offsets.
8546 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
8547 CV = C->isMachineConstantPoolEntry() ? (void *)C->getMachineCPVal()
8548 : (void *)C->getConstVal();
8549 Offset += C->getOffset();
8552 // If it's any of the following then it can't alias with anything but itself.
8553 return isa<FrameIndexSDNode>(Base);
8556 /// isAlias - Return true if there is any possibility that the two addresses
8558 bool DAGCombiner::isAlias(SDValue Ptr1, int64_t Size1,
8559 const Value *SrcValue1, int SrcValueOffset1,
8560 unsigned SrcValueAlign1,
8561 const MDNode *TBAAInfo1,
8562 SDValue Ptr2, int64_t Size2,
8563 const Value *SrcValue2, int SrcValueOffset2,
8564 unsigned SrcValueAlign2,
8565 const MDNode *TBAAInfo2) const {
8566 // If they are the same then they must be aliases.
8567 if (Ptr1 == Ptr2) return true;
8569 // Gather base node and offset information.
8570 SDValue Base1, Base2;
8571 int64_t Offset1, Offset2;
8572 const GlobalValue *GV1, *GV2;
8574 bool isFrameIndex1 = FindBaseOffset(Ptr1, Base1, Offset1, GV1, CV1);
8575 bool isFrameIndex2 = FindBaseOffset(Ptr2, Base2, Offset2, GV2, CV2);
8577 // If they have a same base address then check to see if they overlap.
8578 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
8579 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
8581 // It is possible for different frame indices to alias each other, mostly
8582 // when tail call optimization reuses return address slots for arguments.
8583 // To catch this case, look up the actual index of frame indices to compute
8584 // the real alias relationship.
8585 if (isFrameIndex1 && isFrameIndex2) {
8586 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8587 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
8588 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
8589 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
8592 // Otherwise, if we know what the bases are, and they aren't identical, then
8593 // we know they cannot alias.
8594 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
8597 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
8598 // compared to the size and offset of the access, we may be able to prove they
8599 // do not alias. This check is conservative for now to catch cases created by
8600 // splitting vector types.
8601 if ((SrcValueAlign1 == SrcValueAlign2) &&
8602 (SrcValueOffset1 != SrcValueOffset2) &&
8603 (Size1 == Size2) && (SrcValueAlign1 > Size1)) {
8604 int64_t OffAlign1 = SrcValueOffset1 % SrcValueAlign1;
8605 int64_t OffAlign2 = SrcValueOffset2 % SrcValueAlign1;
8607 // There is no overlap between these relatively aligned accesses of similar
8608 // size, return no alias.
8609 if ((OffAlign1 + Size1) <= OffAlign2 || (OffAlign2 + Size2) <= OffAlign1)
8613 if (CombinerGlobalAA) {
8614 // Use alias analysis information.
8615 int64_t MinOffset = std::min(SrcValueOffset1, SrcValueOffset2);
8616 int64_t Overlap1 = Size1 + SrcValueOffset1 - MinOffset;
8617 int64_t Overlap2 = Size2 + SrcValueOffset2 - MinOffset;
8618 AliasAnalysis::AliasResult AAResult =
8619 AA.alias(AliasAnalysis::Location(SrcValue1, Overlap1, TBAAInfo1),
8620 AliasAnalysis::Location(SrcValue2, Overlap2, TBAAInfo2));
8621 if (AAResult == AliasAnalysis::NoAlias)
8625 // Otherwise we have to assume they alias.
8629 /// FindAliasInfo - Extracts the relevant alias information from the memory
8630 /// node. Returns true if the operand was a load.
8631 bool DAGCombiner::FindAliasInfo(SDNode *N,
8632 SDValue &Ptr, int64_t &Size,
8633 const Value *&SrcValue,
8634 int &SrcValueOffset,
8635 unsigned &SrcValueAlign,
8636 const MDNode *&TBAAInfo) const {
8637 LSBaseSDNode *LS = cast<LSBaseSDNode>(N);
8639 Ptr = LS->getBasePtr();
8640 Size = LS->getMemoryVT().getSizeInBits() >> 3;
8641 SrcValue = LS->getSrcValue();
8642 SrcValueOffset = LS->getSrcValueOffset();
8643 SrcValueAlign = LS->getOriginalAlignment();
8644 TBAAInfo = LS->getTBAAInfo();
8645 return isa<LoadSDNode>(LS);
8648 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
8649 /// looking for aliasing nodes and adding them to the Aliases vector.
8650 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
8651 SmallVector<SDValue, 8> &Aliases) {
8652 SmallVector<SDValue, 8> Chains; // List of chains to visit.
8653 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
8655 // Get alias information for node.
8658 const Value *SrcValue;
8660 unsigned SrcValueAlign;
8661 const MDNode *SrcTBAAInfo;
8662 bool IsLoad = FindAliasInfo(N, Ptr, Size, SrcValue, SrcValueOffset,
8663 SrcValueAlign, SrcTBAAInfo);
8666 Chains.push_back(OriginalChain);
8669 // Look at each chain and determine if it is an alias. If so, add it to the
8670 // aliases list. If not, then continue up the chain looking for the next
8672 while (!Chains.empty()) {
8673 SDValue Chain = Chains.back();
8676 // For TokenFactor nodes, look at each operand and only continue up the
8677 // chain until we find two aliases. If we've seen two aliases, assume we'll
8678 // find more and revert to original chain since the xform is unlikely to be
8681 // FIXME: The depth check could be made to return the last non-aliasing
8682 // chain we found before we hit a tokenfactor rather than the original
8684 if (Depth > 6 || Aliases.size() == 2) {
8686 Aliases.push_back(OriginalChain);
8690 // Don't bother if we've been before.
8691 if (!Visited.insert(Chain.getNode()))
8694 switch (Chain.getOpcode()) {
8695 case ISD::EntryToken:
8696 // Entry token is ideal chain operand, but handled in FindBetterChain.
8701 // Get alias information for Chain.
8704 const Value *OpSrcValue;
8705 int OpSrcValueOffset;
8706 unsigned OpSrcValueAlign;
8707 const MDNode *OpSrcTBAAInfo;
8708 bool IsOpLoad = FindAliasInfo(Chain.getNode(), OpPtr, OpSize,
8709 OpSrcValue, OpSrcValueOffset,
8713 // If chain is alias then stop here.
8714 if (!(IsLoad && IsOpLoad) &&
8715 isAlias(Ptr, Size, SrcValue, SrcValueOffset, SrcValueAlign,
8717 OpPtr, OpSize, OpSrcValue, OpSrcValueOffset,
8718 OpSrcValueAlign, OpSrcTBAAInfo)) {
8719 Aliases.push_back(Chain);
8721 // Look further up the chain.
8722 Chains.push_back(Chain.getOperand(0));
8728 case ISD::TokenFactor:
8729 // We have to check each of the operands of the token factor for "small"
8730 // token factors, so we queue them up. Adding the operands to the queue
8731 // (stack) in reverse order maintains the original order and increases the
8732 // likelihood that getNode will find a matching token factor (CSE.)
8733 if (Chain.getNumOperands() > 16) {
8734 Aliases.push_back(Chain);
8737 for (unsigned n = Chain.getNumOperands(); n;)
8738 Chains.push_back(Chain.getOperand(--n));
8743 // For all other instructions we will just have to take what we can get.
8744 Aliases.push_back(Chain);
8750 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
8751 /// for a better chain (aliasing node.)
8752 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
8753 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
8755 // Accumulate all the aliases to this node.
8756 GatherAllAliases(N, OldChain, Aliases);
8758 // If no operands then chain to entry token.
8759 if (Aliases.size() == 0)
8760 return DAG.getEntryNode();
8762 // If a single operand then chain to it. We don't need to revisit it.
8763 if (Aliases.size() == 1)
8766 // Construct a custom tailored token factor.
8767 return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
8768 &Aliases[0], Aliases.size());
8771 // SelectionDAG::Combine - This is the entry point for the file.
8773 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
8774 CodeGenOpt::Level OptLevel) {
8775 /// run - This is the main entry point to this class.
8777 DAGCombiner(*this, AA, OptLevel).Run(Level);