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());
1085 SDValue DAGCombiner::visit(SDNode *N) {
1086 switch (N->getOpcode()) {
1088 case ISD::TokenFactor: return visitTokenFactor(N);
1089 case ISD::MERGE_VALUES: return visitMERGE_VALUES(N);
1090 case ISD::ADD: return visitADD(N);
1091 case ISD::SUB: return visitSUB(N);
1092 case ISD::ADDC: return visitADDC(N);
1093 case ISD::SUBC: return visitSUBC(N);
1094 case ISD::ADDE: return visitADDE(N);
1095 case ISD::SUBE: return visitSUBE(N);
1096 case ISD::MUL: return visitMUL(N);
1097 case ISD::SDIV: return visitSDIV(N);
1098 case ISD::UDIV: return visitUDIV(N);
1099 case ISD::SREM: return visitSREM(N);
1100 case ISD::UREM: return visitUREM(N);
1101 case ISD::MULHU: return visitMULHU(N);
1102 case ISD::MULHS: return visitMULHS(N);
1103 case ISD::SMUL_LOHI: return visitSMUL_LOHI(N);
1104 case ISD::UMUL_LOHI: return visitUMUL_LOHI(N);
1105 case ISD::SMULO: return visitSMULO(N);
1106 case ISD::UMULO: return visitUMULO(N);
1107 case ISD::SDIVREM: return visitSDIVREM(N);
1108 case ISD::UDIVREM: return visitUDIVREM(N);
1109 case ISD::AND: return visitAND(N);
1110 case ISD::OR: return visitOR(N);
1111 case ISD::XOR: return visitXOR(N);
1112 case ISD::SHL: return visitSHL(N);
1113 case ISD::SRA: return visitSRA(N);
1114 case ISD::SRL: return visitSRL(N);
1115 case ISD::CTLZ: return visitCTLZ(N);
1116 case ISD::CTLZ_ZERO_UNDEF: return visitCTLZ_ZERO_UNDEF(N);
1117 case ISD::CTTZ: return visitCTTZ(N);
1118 case ISD::CTTZ_ZERO_UNDEF: return visitCTTZ_ZERO_UNDEF(N);
1119 case ISD::CTPOP: return visitCTPOP(N);
1120 case ISD::SELECT: return visitSELECT(N);
1121 case ISD::SELECT_CC: return visitSELECT_CC(N);
1122 case ISD::SETCC: return visitSETCC(N);
1123 case ISD::SIGN_EXTEND: return visitSIGN_EXTEND(N);
1124 case ISD::ZERO_EXTEND: return visitZERO_EXTEND(N);
1125 case ISD::ANY_EXTEND: return visitANY_EXTEND(N);
1126 case ISD::SIGN_EXTEND_INREG: return visitSIGN_EXTEND_INREG(N);
1127 case ISD::TRUNCATE: return visitTRUNCATE(N);
1128 case ISD::BITCAST: return visitBITCAST(N);
1129 case ISD::BUILD_PAIR: return visitBUILD_PAIR(N);
1130 case ISD::FADD: return visitFADD(N);
1131 case ISD::FSUB: return visitFSUB(N);
1132 case ISD::FMUL: return visitFMUL(N);
1133 case ISD::FDIV: return visitFDIV(N);
1134 case ISD::FREM: return visitFREM(N);
1135 case ISD::FCOPYSIGN: return visitFCOPYSIGN(N);
1136 case ISD::SINT_TO_FP: return visitSINT_TO_FP(N);
1137 case ISD::UINT_TO_FP: return visitUINT_TO_FP(N);
1138 case ISD::FP_TO_SINT: return visitFP_TO_SINT(N);
1139 case ISD::FP_TO_UINT: return visitFP_TO_UINT(N);
1140 case ISD::FP_ROUND: return visitFP_ROUND(N);
1141 case ISD::FP_ROUND_INREG: return visitFP_ROUND_INREG(N);
1142 case ISD::FP_EXTEND: return visitFP_EXTEND(N);
1143 case ISD::FNEG: return visitFNEG(N);
1144 case ISD::FABS: return visitFABS(N);
1145 case ISD::BRCOND: return visitBRCOND(N);
1146 case ISD::BR_CC: return visitBR_CC(N);
1147 case ISD::LOAD: return visitLOAD(N);
1148 case ISD::STORE: return visitSTORE(N);
1149 case ISD::INSERT_VECTOR_ELT: return visitINSERT_VECTOR_ELT(N);
1150 case ISD::EXTRACT_VECTOR_ELT: return visitEXTRACT_VECTOR_ELT(N);
1151 case ISD::BUILD_VECTOR: return visitBUILD_VECTOR(N);
1152 case ISD::CONCAT_VECTORS: return visitCONCAT_VECTORS(N);
1153 case ISD::EXTRACT_SUBVECTOR: return visitEXTRACT_SUBVECTOR(N);
1154 case ISD::VECTOR_SHUFFLE: return visitVECTOR_SHUFFLE(N);
1155 case ISD::MEMBARRIER: return visitMEMBARRIER(N);
1160 SDValue DAGCombiner::combine(SDNode *N) {
1161 SDValue RV = visit(N);
1163 // If nothing happened, try a target-specific DAG combine.
1164 if (RV.getNode() == 0) {
1165 assert(N->getOpcode() != ISD::DELETED_NODE &&
1166 "Node was deleted but visit returned NULL!");
1168 if (N->getOpcode() >= ISD::BUILTIN_OP_END ||
1169 TLI.hasTargetDAGCombine((ISD::NodeType)N->getOpcode())) {
1171 // Expose the DAG combiner to the target combiner impls.
1172 TargetLowering::DAGCombinerInfo
1173 DagCombineInfo(DAG, !LegalTypes, !LegalOperations, false, this);
1175 RV = TLI.PerformDAGCombine(N, DagCombineInfo);
1179 // If nothing happened still, try promoting the operation.
1180 if (RV.getNode() == 0) {
1181 switch (N->getOpcode()) {
1189 RV = PromoteIntBinOp(SDValue(N, 0));
1194 RV = PromoteIntShiftOp(SDValue(N, 0));
1196 case ISD::SIGN_EXTEND:
1197 case ISD::ZERO_EXTEND:
1198 case ISD::ANY_EXTEND:
1199 RV = PromoteExtend(SDValue(N, 0));
1202 if (PromoteLoad(SDValue(N, 0)))
1208 // If N is a commutative binary node, try commuting it to enable more
1210 if (RV.getNode() == 0 &&
1211 SelectionDAG::isCommutativeBinOp(N->getOpcode()) &&
1212 N->getNumValues() == 1) {
1213 SDValue N0 = N->getOperand(0);
1214 SDValue N1 = N->getOperand(1);
1216 // Constant operands are canonicalized to RHS.
1217 if (isa<ConstantSDNode>(N0) || !isa<ConstantSDNode>(N1)) {
1218 SDValue Ops[] = { N1, N0 };
1219 SDNode *CSENode = DAG.getNodeIfExists(N->getOpcode(), N->getVTList(),
1222 return SDValue(CSENode, 0);
1229 /// getInputChainForNode - Given a node, return its input chain if it has one,
1230 /// otherwise return a null sd operand.
1231 static SDValue getInputChainForNode(SDNode *N) {
1232 if (unsigned NumOps = N->getNumOperands()) {
1233 if (N->getOperand(0).getValueType() == MVT::Other)
1234 return N->getOperand(0);
1235 else if (N->getOperand(NumOps-1).getValueType() == MVT::Other)
1236 return N->getOperand(NumOps-1);
1237 for (unsigned i = 1; i < NumOps-1; ++i)
1238 if (N->getOperand(i).getValueType() == MVT::Other)
1239 return N->getOperand(i);
1244 SDValue DAGCombiner::visitTokenFactor(SDNode *N) {
1245 // If N has two operands, where one has an input chain equal to the other,
1246 // the 'other' chain is redundant.
1247 if (N->getNumOperands() == 2) {
1248 if (getInputChainForNode(N->getOperand(0).getNode()) == N->getOperand(1))
1249 return N->getOperand(0);
1250 if (getInputChainForNode(N->getOperand(1).getNode()) == N->getOperand(0))
1251 return N->getOperand(1);
1254 SmallVector<SDNode *, 8> TFs; // List of token factors to visit.
1255 SmallVector<SDValue, 8> Ops; // Ops for replacing token factor.
1256 SmallPtrSet<SDNode*, 16> SeenOps;
1257 bool Changed = false; // If we should replace this token factor.
1259 // Start out with this token factor.
1262 // Iterate through token factors. The TFs grows when new token factors are
1264 for (unsigned i = 0; i < TFs.size(); ++i) {
1265 SDNode *TF = TFs[i];
1267 // Check each of the operands.
1268 for (unsigned i = 0, ie = TF->getNumOperands(); i != ie; ++i) {
1269 SDValue Op = TF->getOperand(i);
1271 switch (Op.getOpcode()) {
1272 case ISD::EntryToken:
1273 // Entry tokens don't need to be added to the list. They are
1278 case ISD::TokenFactor:
1279 if (Op.hasOneUse() &&
1280 std::find(TFs.begin(), TFs.end(), Op.getNode()) == TFs.end()) {
1281 // Queue up for processing.
1282 TFs.push_back(Op.getNode());
1283 // Clean up in case the token factor is removed.
1284 AddToWorkList(Op.getNode());
1291 // Only add if it isn't already in the list.
1292 if (SeenOps.insert(Op.getNode()))
1303 // If we've change things around then replace token factor.
1306 // The entry token is the only possible outcome.
1307 Result = DAG.getEntryNode();
1309 // New and improved token factor.
1310 Result = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
1311 MVT::Other, &Ops[0], Ops.size());
1314 // Don't add users to work list.
1315 return CombineTo(N, Result, false);
1321 /// MERGE_VALUES can always be eliminated.
1322 SDValue DAGCombiner::visitMERGE_VALUES(SDNode *N) {
1323 WorkListRemover DeadNodes(*this);
1324 // Replacing results may cause a different MERGE_VALUES to suddenly
1325 // be CSE'd with N, and carry its uses with it. Iterate until no
1326 // uses remain, to ensure that the node can be safely deleted.
1328 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
1329 DAG.ReplaceAllUsesOfValueWith(SDValue(N, i), N->getOperand(i),
1331 } while (!N->use_empty());
1332 removeFromWorkList(N);
1334 return SDValue(N, 0); // Return N so it doesn't get rechecked!
1338 SDValue combineShlAddConstant(DebugLoc DL, SDValue N0, SDValue N1,
1339 SelectionDAG &DAG) {
1340 EVT VT = N0.getValueType();
1341 SDValue N00 = N0.getOperand(0);
1342 SDValue N01 = N0.getOperand(1);
1343 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N01);
1345 if (N01C && N00.getOpcode() == ISD::ADD && N00.getNode()->hasOneUse() &&
1346 isa<ConstantSDNode>(N00.getOperand(1))) {
1347 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1348 N0 = DAG.getNode(ISD::ADD, N0.getDebugLoc(), VT,
1349 DAG.getNode(ISD::SHL, N00.getDebugLoc(), VT,
1350 N00.getOperand(0), N01),
1351 DAG.getNode(ISD::SHL, N01.getDebugLoc(), VT,
1352 N00.getOperand(1), N01));
1353 return DAG.getNode(ISD::ADD, DL, VT, N0, N1);
1359 SDValue DAGCombiner::visitADD(SDNode *N) {
1360 SDValue N0 = N->getOperand(0);
1361 SDValue N1 = N->getOperand(1);
1362 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1363 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1364 EVT VT = N0.getValueType();
1367 if (VT.isVector()) {
1368 SDValue FoldedVOp = SimplifyVBinOp(N);
1369 if (FoldedVOp.getNode()) return FoldedVOp;
1372 // fold (add x, undef) -> undef
1373 if (N0.getOpcode() == ISD::UNDEF)
1375 if (N1.getOpcode() == ISD::UNDEF)
1377 // fold (add c1, c2) -> c1+c2
1379 return DAG.FoldConstantArithmetic(ISD::ADD, VT, N0C, N1C);
1380 // canonicalize constant to RHS
1382 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N1, N0);
1383 // fold (add x, 0) -> x
1384 if (N1C && N1C->isNullValue())
1386 // fold (add Sym, c) -> Sym+c
1387 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1388 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA) && N1C &&
1389 GA->getOpcode() == ISD::GlobalAddress)
1390 return DAG.getGlobalAddress(GA->getGlobal(), N1C->getDebugLoc(), VT,
1392 (uint64_t)N1C->getSExtValue());
1393 // fold ((c1-A)+c2) -> (c1+c2)-A
1394 if (N1C && N0.getOpcode() == ISD::SUB)
1395 if (ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getOperand(0)))
1396 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1397 DAG.getConstant(N1C->getAPIntValue()+
1398 N0C->getAPIntValue(), VT),
1401 SDValue RADD = ReassociateOps(ISD::ADD, N->getDebugLoc(), N0, N1);
1402 if (RADD.getNode() != 0)
1404 // fold ((0-A) + B) -> B-A
1405 if (N0.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N0.getOperand(0)) &&
1406 cast<ConstantSDNode>(N0.getOperand(0))->isNullValue())
1407 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1, N0.getOperand(1));
1408 // fold (A + (0-B)) -> A-B
1409 if (N1.getOpcode() == ISD::SUB && isa<ConstantSDNode>(N1.getOperand(0)) &&
1410 cast<ConstantSDNode>(N1.getOperand(0))->isNullValue())
1411 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, N1.getOperand(1));
1412 // fold (A+(B-A)) -> B
1413 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(1))
1414 return N1.getOperand(0);
1415 // fold ((B-A)+A) -> B
1416 if (N0.getOpcode() == ISD::SUB && N1 == N0.getOperand(1))
1417 return N0.getOperand(0);
1418 // fold (A+(B-(A+C))) to (B-C)
1419 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1420 N0 == N1.getOperand(1).getOperand(0))
1421 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1.getOperand(0),
1422 N1.getOperand(1).getOperand(1));
1423 // fold (A+(B-(C+A))) to (B-C)
1424 if (N1.getOpcode() == ISD::SUB && N1.getOperand(1).getOpcode() == ISD::ADD &&
1425 N0 == N1.getOperand(1).getOperand(1))
1426 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1.getOperand(0),
1427 N1.getOperand(1).getOperand(0));
1428 // fold (A+((B-A)+or-C)) to (B+or-C)
1429 if ((N1.getOpcode() == ISD::SUB || N1.getOpcode() == ISD::ADD) &&
1430 N1.getOperand(0).getOpcode() == ISD::SUB &&
1431 N0 == N1.getOperand(0).getOperand(1))
1432 return DAG.getNode(N1.getOpcode(), N->getDebugLoc(), VT,
1433 N1.getOperand(0).getOperand(0), N1.getOperand(1));
1435 // fold (A-B)+(C-D) to (A+C)-(B+D) when A or C is constant
1436 if (N0.getOpcode() == ISD::SUB && N1.getOpcode() == ISD::SUB) {
1437 SDValue N00 = N0.getOperand(0);
1438 SDValue N01 = N0.getOperand(1);
1439 SDValue N10 = N1.getOperand(0);
1440 SDValue N11 = N1.getOperand(1);
1442 if (isa<ConstantSDNode>(N00) || isa<ConstantSDNode>(N10))
1443 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1444 DAG.getNode(ISD::ADD, N0.getDebugLoc(), VT, N00, N10),
1445 DAG.getNode(ISD::ADD, N1.getDebugLoc(), VT, N01, N11));
1448 if (!VT.isVector() && SimplifyDemandedBits(SDValue(N, 0)))
1449 return SDValue(N, 0);
1451 // fold (a+b) -> (a|b) iff a and b share no bits.
1452 if (VT.isInteger() && !VT.isVector()) {
1453 APInt LHSZero, LHSOne;
1454 APInt RHSZero, RHSOne;
1455 DAG.ComputeMaskedBits(N0, LHSZero, LHSOne);
1457 if (LHSZero.getBoolValue()) {
1458 DAG.ComputeMaskedBits(N1, RHSZero, RHSOne);
1460 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1461 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1462 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1463 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N1);
1467 // fold (add (shl (add x, c1), c2), ) -> (add (add (shl x, c2), c1<<c2), )
1468 if (N0.getOpcode() == ISD::SHL && N0.getNode()->hasOneUse()) {
1469 SDValue Result = combineShlAddConstant(N->getDebugLoc(), N0, N1, DAG);
1470 if (Result.getNode()) return Result;
1472 if (N1.getOpcode() == ISD::SHL && N1.getNode()->hasOneUse()) {
1473 SDValue Result = combineShlAddConstant(N->getDebugLoc(), N1, N0, DAG);
1474 if (Result.getNode()) return Result;
1477 // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
1478 if (N1.getOpcode() == ISD::SHL &&
1479 N1.getOperand(0).getOpcode() == ISD::SUB)
1480 if (ConstantSDNode *C =
1481 dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
1482 if (C->getAPIntValue() == 0)
1483 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0,
1484 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1485 N1.getOperand(0).getOperand(1),
1487 if (N0.getOpcode() == ISD::SHL &&
1488 N0.getOperand(0).getOpcode() == ISD::SUB)
1489 if (ConstantSDNode *C =
1490 dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
1491 if (C->getAPIntValue() == 0)
1492 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1,
1493 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1494 N0.getOperand(0).getOperand(1),
1497 if (N1.getOpcode() == ISD::AND) {
1498 SDValue AndOp0 = N1.getOperand(0);
1499 ConstantSDNode *AndOp1 = dyn_cast<ConstantSDNode>(N1->getOperand(1));
1500 unsigned NumSignBits = DAG.ComputeNumSignBits(AndOp0);
1501 unsigned DestBits = VT.getScalarType().getSizeInBits();
1503 // (add z, (and (sbbl x, x), 1)) -> (sub z, (sbbl x, x))
1504 // and similar xforms where the inner op is either ~0 or 0.
1505 if (NumSignBits == DestBits && AndOp1 && AndOp1->isOne()) {
1506 DebugLoc DL = N->getDebugLoc();
1507 return DAG.getNode(ISD::SUB, DL, VT, N->getOperand(0), AndOp0);
1511 // add (sext i1), X -> sub X, (zext i1)
1512 if (N0.getOpcode() == ISD::SIGN_EXTEND &&
1513 N0.getOperand(0).getValueType() == MVT::i1 &&
1514 !TLI.isOperationLegal(ISD::SIGN_EXTEND, MVT::i1)) {
1515 DebugLoc DL = N->getDebugLoc();
1516 SDValue ZExt = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0));
1517 return DAG.getNode(ISD::SUB, DL, VT, N1, ZExt);
1523 SDValue DAGCombiner::visitADDC(SDNode *N) {
1524 SDValue N0 = N->getOperand(0);
1525 SDValue N1 = N->getOperand(1);
1526 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1527 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1528 EVT VT = N0.getValueType();
1530 // If the flag result is dead, turn this into an ADD.
1531 if (!N->hasAnyUseOfValue(1))
1532 return CombineTo(N, DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0, N1),
1533 DAG.getNode(ISD::CARRY_FALSE,
1534 N->getDebugLoc(), MVT::Glue));
1536 // canonicalize constant to RHS.
1538 return DAG.getNode(ISD::ADDC, N->getDebugLoc(), N->getVTList(), N1, N0);
1540 // fold (addc x, 0) -> x + no carry out
1541 if (N1C && N1C->isNullValue())
1542 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE,
1543 N->getDebugLoc(), MVT::Glue));
1545 // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits.
1546 APInt LHSZero, LHSOne;
1547 APInt RHSZero, RHSOne;
1548 DAG.ComputeMaskedBits(N0, LHSZero, LHSOne);
1550 if (LHSZero.getBoolValue()) {
1551 DAG.ComputeMaskedBits(N1, RHSZero, RHSOne);
1553 // If all possibly-set bits on the LHS are clear on the RHS, return an OR.
1554 // If all possibly-set bits on the RHS are clear on the LHS, return an OR.
1555 if ((RHSZero & ~LHSZero) == ~LHSZero || (LHSZero & ~RHSZero) == ~RHSZero)
1556 return CombineTo(N, DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N1),
1557 DAG.getNode(ISD::CARRY_FALSE,
1558 N->getDebugLoc(), MVT::Glue));
1564 SDValue DAGCombiner::visitADDE(SDNode *N) {
1565 SDValue N0 = N->getOperand(0);
1566 SDValue N1 = N->getOperand(1);
1567 SDValue CarryIn = N->getOperand(2);
1568 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1569 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1571 // canonicalize constant to RHS
1573 return DAG.getNode(ISD::ADDE, N->getDebugLoc(), N->getVTList(),
1576 // fold (adde x, y, false) -> (addc x, y)
1577 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1578 return DAG.getNode(ISD::ADDC, N->getDebugLoc(), N->getVTList(), N0, N1);
1583 // Since it may not be valid to emit a fold to zero for vector initializers
1584 // check if we can before folding.
1585 static SDValue tryFoldToZero(DebugLoc DL, const TargetLowering &TLI, EVT VT,
1586 SelectionDAG &DAG, bool LegalOperations) {
1587 if (!VT.isVector()) {
1588 return DAG.getConstant(0, VT);
1590 if (!LegalOperations || TLI.isOperationLegal(ISD::BUILD_VECTOR, VT)) {
1591 // Produce a vector of zeros.
1592 SDValue El = DAG.getConstant(0, VT.getVectorElementType());
1593 std::vector<SDValue> Ops(VT.getVectorNumElements(), El);
1594 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT,
1595 &Ops[0], Ops.size());
1600 SDValue DAGCombiner::visitSUB(SDNode *N) {
1601 SDValue N0 = N->getOperand(0);
1602 SDValue N1 = N->getOperand(1);
1603 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1604 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1605 ConstantSDNode *N1C1 = N1.getOpcode() != ISD::ADD ? 0 :
1606 dyn_cast<ConstantSDNode>(N1.getOperand(1).getNode());
1607 EVT VT = N0.getValueType();
1610 if (VT.isVector()) {
1611 SDValue FoldedVOp = SimplifyVBinOp(N);
1612 if (FoldedVOp.getNode()) return FoldedVOp;
1615 // fold (sub x, x) -> 0
1616 // FIXME: Refactor this and xor and other similar operations together.
1618 return tryFoldToZero(N->getDebugLoc(), TLI, VT, DAG, LegalOperations);
1619 // fold (sub c1, c2) -> c1-c2
1621 return DAG.FoldConstantArithmetic(ISD::SUB, VT, N0C, N1C);
1622 // fold (sub x, c) -> (add x, -c)
1624 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0,
1625 DAG.getConstant(-N1C->getAPIntValue(), VT));
1626 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1)
1627 if (N0C && N0C->isAllOnesValue())
1628 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0);
1629 // fold A-(A-B) -> B
1630 if (N1.getOpcode() == ISD::SUB && N0 == N1.getOperand(0))
1631 return N1.getOperand(1);
1632 // fold (A+B)-A -> B
1633 if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1)
1634 return N0.getOperand(1);
1635 // fold (A+B)-B -> A
1636 if (N0.getOpcode() == ISD::ADD && N0.getOperand(1) == N1)
1637 return N0.getOperand(0);
1638 // fold C2-(A+C1) -> (C2-C1)-A
1639 if (N1.getOpcode() == ISD::ADD && N0C && N1C1) {
1640 SDValue NewC = DAG.getConstant((N0C->getAPIntValue() - N1C1->getAPIntValue()), VT);
1641 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, NewC,
1644 // fold ((A+(B+or-C))-B) -> A+or-C
1645 if (N0.getOpcode() == ISD::ADD &&
1646 (N0.getOperand(1).getOpcode() == ISD::SUB ||
1647 N0.getOperand(1).getOpcode() == ISD::ADD) &&
1648 N0.getOperand(1).getOperand(0) == N1)
1649 return DAG.getNode(N0.getOperand(1).getOpcode(), N->getDebugLoc(), VT,
1650 N0.getOperand(0), N0.getOperand(1).getOperand(1));
1651 // fold ((A+(C+B))-B) -> A+C
1652 if (N0.getOpcode() == ISD::ADD &&
1653 N0.getOperand(1).getOpcode() == ISD::ADD &&
1654 N0.getOperand(1).getOperand(1) == N1)
1655 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT,
1656 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1657 // fold ((A-(B-C))-C) -> A-B
1658 if (N0.getOpcode() == ISD::SUB &&
1659 N0.getOperand(1).getOpcode() == ISD::SUB &&
1660 N0.getOperand(1).getOperand(1) == N1)
1661 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1662 N0.getOperand(0), N0.getOperand(1).getOperand(0));
1664 // If either operand of a sub is undef, the result is undef
1665 if (N0.getOpcode() == ISD::UNDEF)
1667 if (N1.getOpcode() == ISD::UNDEF)
1670 // If the relocation model supports it, consider symbol offsets.
1671 if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(N0))
1672 if (!LegalOperations && TLI.isOffsetFoldingLegal(GA)) {
1673 // fold (sub Sym, c) -> Sym-c
1674 if (N1C && GA->getOpcode() == ISD::GlobalAddress)
1675 return DAG.getGlobalAddress(GA->getGlobal(), N1C->getDebugLoc(), VT,
1677 (uint64_t)N1C->getSExtValue());
1678 // fold (sub Sym+c1, Sym+c2) -> c1-c2
1679 if (GlobalAddressSDNode *GB = dyn_cast<GlobalAddressSDNode>(N1))
1680 if (GA->getGlobal() == GB->getGlobal())
1681 return DAG.getConstant((uint64_t)GA->getOffset() - GB->getOffset(),
1688 SDValue DAGCombiner::visitSUBC(SDNode *N) {
1689 SDValue N0 = N->getOperand(0);
1690 SDValue N1 = N->getOperand(1);
1691 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1692 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1693 EVT VT = N0.getValueType();
1695 // If the flag result is dead, turn this into an SUB.
1696 if (!N->hasAnyUseOfValue(1))
1697 return CombineTo(N, DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, N1),
1698 DAG.getNode(ISD::CARRY_FALSE, N->getDebugLoc(),
1701 // fold (subc x, x) -> 0 + no borrow
1703 return CombineTo(N, DAG.getConstant(0, VT),
1704 DAG.getNode(ISD::CARRY_FALSE, N->getDebugLoc(),
1707 // fold (subc x, 0) -> x + no borrow
1708 if (N1C && N1C->isNullValue())
1709 return CombineTo(N, N0, DAG.getNode(ISD::CARRY_FALSE, N->getDebugLoc(),
1712 // Canonicalize (sub -1, x) -> ~x, i.e. (xor x, -1) + no borrow
1713 if (N0C && N0C->isAllOnesValue())
1714 return CombineTo(N, DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0),
1715 DAG.getNode(ISD::CARRY_FALSE, N->getDebugLoc(),
1721 SDValue DAGCombiner::visitSUBE(SDNode *N) {
1722 SDValue N0 = N->getOperand(0);
1723 SDValue N1 = N->getOperand(1);
1724 SDValue CarryIn = N->getOperand(2);
1726 // fold (sube x, y, false) -> (subc x, y)
1727 if (CarryIn.getOpcode() == ISD::CARRY_FALSE)
1728 return DAG.getNode(ISD::SUBC, N->getDebugLoc(), N->getVTList(), N0, N1);
1733 SDValue DAGCombiner::visitMUL(SDNode *N) {
1734 SDValue N0 = N->getOperand(0);
1735 SDValue N1 = N->getOperand(1);
1736 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1737 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1738 EVT VT = N0.getValueType();
1741 if (VT.isVector()) {
1742 SDValue FoldedVOp = SimplifyVBinOp(N);
1743 if (FoldedVOp.getNode()) return FoldedVOp;
1746 // fold (mul x, undef) -> 0
1747 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
1748 return DAG.getConstant(0, VT);
1749 // fold (mul c1, c2) -> c1*c2
1751 return DAG.FoldConstantArithmetic(ISD::MUL, VT, N0C, N1C);
1752 // canonicalize constant to RHS
1754 return DAG.getNode(ISD::MUL, N->getDebugLoc(), VT, N1, N0);
1755 // fold (mul x, 0) -> 0
1756 if (N1C && N1C->isNullValue())
1758 // fold (mul x, -1) -> 0-x
1759 if (N1C && N1C->isAllOnesValue())
1760 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1761 DAG.getConstant(0, VT), N0);
1762 // fold (mul x, (1 << c)) -> x << c
1763 if (N1C && N1C->getAPIntValue().isPowerOf2())
1764 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
1765 DAG.getConstant(N1C->getAPIntValue().logBase2(),
1766 getShiftAmountTy(N0.getValueType())));
1767 // fold (mul x, -(1 << c)) -> -(x << c) or (-x) << c
1768 if (N1C && (-N1C->getAPIntValue()).isPowerOf2()) {
1769 unsigned Log2Val = (-N1C->getAPIntValue()).logBase2();
1770 // FIXME: If the input is something that is easily negated (e.g. a
1771 // single-use add), we should put the negate there.
1772 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1773 DAG.getConstant(0, VT),
1774 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
1775 DAG.getConstant(Log2Val,
1776 getShiftAmountTy(N0.getValueType()))));
1778 // (mul (shl X, c1), c2) -> (mul X, c2 << c1)
1779 if (N1C && N0.getOpcode() == ISD::SHL &&
1780 isa<ConstantSDNode>(N0.getOperand(1))) {
1781 SDValue C3 = DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1782 N1, N0.getOperand(1));
1783 AddToWorkList(C3.getNode());
1784 return DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1785 N0.getOperand(0), C3);
1788 // Change (mul (shl X, C), Y) -> (shl (mul X, Y), C) when the shift has one
1791 SDValue Sh(0,0), Y(0,0);
1792 // Check for both (mul (shl X, C), Y) and (mul Y, (shl X, C)).
1793 if (N0.getOpcode() == ISD::SHL && isa<ConstantSDNode>(N0.getOperand(1)) &&
1794 N0.getNode()->hasOneUse()) {
1796 } else if (N1.getOpcode() == ISD::SHL &&
1797 isa<ConstantSDNode>(N1.getOperand(1)) &&
1798 N1.getNode()->hasOneUse()) {
1803 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1804 Sh.getOperand(0), Y);
1805 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
1806 Mul, Sh.getOperand(1));
1810 // fold (mul (add x, c1), c2) -> (add (mul x, c2), c1*c2)
1811 if (N1C && N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse() &&
1812 isa<ConstantSDNode>(N0.getOperand(1)))
1813 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT,
1814 DAG.getNode(ISD::MUL, N0.getDebugLoc(), VT,
1815 N0.getOperand(0), N1),
1816 DAG.getNode(ISD::MUL, N1.getDebugLoc(), VT,
1817 N0.getOperand(1), N1));
1820 SDValue RMUL = ReassociateOps(ISD::MUL, N->getDebugLoc(), N0, N1);
1821 if (RMUL.getNode() != 0)
1827 SDValue DAGCombiner::visitSDIV(SDNode *N) {
1828 SDValue N0 = N->getOperand(0);
1829 SDValue N1 = N->getOperand(1);
1830 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1831 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1832 EVT VT = N->getValueType(0);
1835 if (VT.isVector()) {
1836 SDValue FoldedVOp = SimplifyVBinOp(N);
1837 if (FoldedVOp.getNode()) return FoldedVOp;
1840 // fold (sdiv c1, c2) -> c1/c2
1841 if (N0C && N1C && !N1C->isNullValue())
1842 return DAG.FoldConstantArithmetic(ISD::SDIV, VT, N0C, N1C);
1843 // fold (sdiv X, 1) -> X
1844 if (N1C && N1C->getAPIntValue() == 1LL)
1846 // fold (sdiv X, -1) -> 0-X
1847 if (N1C && N1C->isAllOnesValue())
1848 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1849 DAG.getConstant(0, VT), N0);
1850 // If we know the sign bits of both operands are zero, strength reduce to a
1851 // udiv instead. Handles (X&15) /s 4 -> X&15 >> 2
1852 if (!VT.isVector()) {
1853 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
1854 return DAG.getNode(ISD::UDIV, N->getDebugLoc(), N1.getValueType(),
1857 // fold (sdiv X, pow2) -> simple ops after legalize
1858 if (N1C && !N1C->isNullValue() &&
1859 (N1C->getAPIntValue().isPowerOf2() ||
1860 (-N1C->getAPIntValue()).isPowerOf2())) {
1861 // If dividing by powers of two is cheap, then don't perform the following
1863 if (TLI.isPow2DivCheap())
1866 unsigned lg2 = N1C->getAPIntValue().countTrailingZeros();
1868 // Splat the sign bit into the register
1869 SDValue SGN = DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0,
1870 DAG.getConstant(VT.getSizeInBits()-1,
1871 getShiftAmountTy(N0.getValueType())));
1872 AddToWorkList(SGN.getNode());
1874 // Add (N0 < 0) ? abs2 - 1 : 0;
1875 SDValue SRL = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, SGN,
1876 DAG.getConstant(VT.getSizeInBits() - lg2,
1877 getShiftAmountTy(SGN.getValueType())));
1878 SDValue ADD = DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0, SRL);
1879 AddToWorkList(SRL.getNode());
1880 AddToWorkList(ADD.getNode()); // Divide by pow2
1881 SDValue SRA = DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, ADD,
1882 DAG.getConstant(lg2, getShiftAmountTy(ADD.getValueType())));
1884 // If we're dividing by a positive value, we're done. Otherwise, we must
1885 // negate the result.
1886 if (N1C->getAPIntValue().isNonNegative())
1889 AddToWorkList(SRA.getNode());
1890 return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
1891 DAG.getConstant(0, VT), SRA);
1894 // if integer divide is expensive and we satisfy the requirements, emit an
1895 // alternate sequence.
1896 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
1897 SDValue Op = BuildSDIV(N);
1898 if (Op.getNode()) return Op;
1902 if (N0.getOpcode() == ISD::UNDEF)
1903 return DAG.getConstant(0, VT);
1904 // X / undef -> undef
1905 if (N1.getOpcode() == ISD::UNDEF)
1911 SDValue DAGCombiner::visitUDIV(SDNode *N) {
1912 SDValue N0 = N->getOperand(0);
1913 SDValue N1 = N->getOperand(1);
1914 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0.getNode());
1915 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
1916 EVT VT = N->getValueType(0);
1919 if (VT.isVector()) {
1920 SDValue FoldedVOp = SimplifyVBinOp(N);
1921 if (FoldedVOp.getNode()) return FoldedVOp;
1924 // fold (udiv c1, c2) -> c1/c2
1925 if (N0C && N1C && !N1C->isNullValue())
1926 return DAG.FoldConstantArithmetic(ISD::UDIV, VT, N0C, N1C);
1927 // fold (udiv x, (1 << c)) -> x >>u c
1928 if (N1C && N1C->getAPIntValue().isPowerOf2())
1929 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0,
1930 DAG.getConstant(N1C->getAPIntValue().logBase2(),
1931 getShiftAmountTy(N0.getValueType())));
1932 // fold (udiv x, (shl c, y)) -> x >>u (log2(c)+y) iff c is power of 2
1933 if (N1.getOpcode() == ISD::SHL) {
1934 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
1935 if (SHC->getAPIntValue().isPowerOf2()) {
1936 EVT ADDVT = N1.getOperand(1).getValueType();
1937 SDValue Add = DAG.getNode(ISD::ADD, N->getDebugLoc(), ADDVT,
1939 DAG.getConstant(SHC->getAPIntValue()
1942 AddToWorkList(Add.getNode());
1943 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0, Add);
1947 // fold (udiv x, c) -> alternate
1948 if (N1C && !N1C->isNullValue() && !TLI.isIntDivCheap()) {
1949 SDValue Op = BuildUDIV(N);
1950 if (Op.getNode()) return Op;
1954 if (N0.getOpcode() == ISD::UNDEF)
1955 return DAG.getConstant(0, VT);
1956 // X / undef -> undef
1957 if (N1.getOpcode() == ISD::UNDEF)
1963 SDValue DAGCombiner::visitSREM(SDNode *N) {
1964 SDValue N0 = N->getOperand(0);
1965 SDValue N1 = N->getOperand(1);
1966 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
1967 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
1968 EVT VT = N->getValueType(0);
1970 // fold (srem c1, c2) -> c1%c2
1971 if (N0C && N1C && !N1C->isNullValue())
1972 return DAG.FoldConstantArithmetic(ISD::SREM, VT, N0C, N1C);
1973 // If we know the sign bits of both operands are zero, strength reduce to a
1974 // urem instead. Handles (X & 0x0FFFFFFF) %s 16 -> X&15
1975 if (!VT.isVector()) {
1976 if (DAG.SignBitIsZero(N1) && DAG.SignBitIsZero(N0))
1977 return DAG.getNode(ISD::UREM, N->getDebugLoc(), VT, N0, N1);
1980 // If X/C can be simplified by the division-by-constant logic, lower
1981 // X%C to the equivalent of X-X/C*C.
1982 if (N1C && !N1C->isNullValue()) {
1983 SDValue Div = DAG.getNode(ISD::SDIV, N->getDebugLoc(), VT, N0, N1);
1984 AddToWorkList(Div.getNode());
1985 SDValue OptimizedDiv = combine(Div.getNode());
1986 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
1987 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
1989 SDValue Sub = DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, Mul);
1990 AddToWorkList(Mul.getNode());
1996 if (N0.getOpcode() == ISD::UNDEF)
1997 return DAG.getConstant(0, VT);
1998 // X % undef -> undef
1999 if (N1.getOpcode() == ISD::UNDEF)
2005 SDValue DAGCombiner::visitUREM(SDNode *N) {
2006 SDValue N0 = N->getOperand(0);
2007 SDValue N1 = N->getOperand(1);
2008 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2009 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2010 EVT VT = N->getValueType(0);
2012 // fold (urem c1, c2) -> c1%c2
2013 if (N0C && N1C && !N1C->isNullValue())
2014 return DAG.FoldConstantArithmetic(ISD::UREM, VT, N0C, N1C);
2015 // fold (urem x, pow2) -> (and x, pow2-1)
2016 if (N1C && !N1C->isNullValue() && N1C->getAPIntValue().isPowerOf2())
2017 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0,
2018 DAG.getConstant(N1C->getAPIntValue()-1,VT));
2019 // fold (urem x, (shl pow2, y)) -> (and x, (add (shl pow2, y), -1))
2020 if (N1.getOpcode() == ISD::SHL) {
2021 if (ConstantSDNode *SHC = dyn_cast<ConstantSDNode>(N1.getOperand(0))) {
2022 if (SHC->getAPIntValue().isPowerOf2()) {
2024 DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N1,
2025 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()),
2027 AddToWorkList(Add.getNode());
2028 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, Add);
2033 // If X/C can be simplified by the division-by-constant logic, lower
2034 // X%C to the equivalent of X-X/C*C.
2035 if (N1C && !N1C->isNullValue()) {
2036 SDValue Div = DAG.getNode(ISD::UDIV, N->getDebugLoc(), VT, N0, N1);
2037 AddToWorkList(Div.getNode());
2038 SDValue OptimizedDiv = combine(Div.getNode());
2039 if (OptimizedDiv.getNode() && OptimizedDiv.getNode() != Div.getNode()) {
2040 SDValue Mul = DAG.getNode(ISD::MUL, N->getDebugLoc(), VT,
2042 SDValue Sub = DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0, Mul);
2043 AddToWorkList(Mul.getNode());
2049 if (N0.getOpcode() == ISD::UNDEF)
2050 return DAG.getConstant(0, VT);
2051 // X % undef -> undef
2052 if (N1.getOpcode() == ISD::UNDEF)
2058 SDValue DAGCombiner::visitMULHS(SDNode *N) {
2059 SDValue N0 = N->getOperand(0);
2060 SDValue N1 = N->getOperand(1);
2061 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2062 EVT VT = N->getValueType(0);
2063 DebugLoc DL = N->getDebugLoc();
2065 // fold (mulhs x, 0) -> 0
2066 if (N1C && N1C->isNullValue())
2068 // fold (mulhs x, 1) -> (sra x, size(x)-1)
2069 if (N1C && N1C->getAPIntValue() == 1)
2070 return DAG.getNode(ISD::SRA, N->getDebugLoc(), N0.getValueType(), N0,
2071 DAG.getConstant(N0.getValueType().getSizeInBits() - 1,
2072 getShiftAmountTy(N0.getValueType())));
2073 // fold (mulhs x, undef) -> 0
2074 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2075 return DAG.getConstant(0, VT);
2077 // If the type twice as wide is legal, transform the mulhs to a wider multiply
2079 if (VT.isSimple() && !VT.isVector()) {
2080 MVT Simple = VT.getSimpleVT();
2081 unsigned SimpleSize = Simple.getSizeInBits();
2082 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2083 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2084 N0 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N0);
2085 N1 = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N1);
2086 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2087 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2088 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2089 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2096 SDValue DAGCombiner::visitMULHU(SDNode *N) {
2097 SDValue N0 = N->getOperand(0);
2098 SDValue N1 = N->getOperand(1);
2099 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2100 EVT VT = N->getValueType(0);
2101 DebugLoc DL = N->getDebugLoc();
2103 // fold (mulhu x, 0) -> 0
2104 if (N1C && N1C->isNullValue())
2106 // fold (mulhu x, 1) -> 0
2107 if (N1C && N1C->getAPIntValue() == 1)
2108 return DAG.getConstant(0, N0.getValueType());
2109 // fold (mulhu x, undef) -> 0
2110 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2111 return DAG.getConstant(0, VT);
2113 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2115 if (VT.isSimple() && !VT.isVector()) {
2116 MVT Simple = VT.getSimpleVT();
2117 unsigned SimpleSize = Simple.getSizeInBits();
2118 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2119 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2120 N0 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N0);
2121 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N1);
2122 N1 = DAG.getNode(ISD::MUL, DL, NewVT, N0, N1);
2123 N1 = DAG.getNode(ISD::SRL, DL, NewVT, N1,
2124 DAG.getConstant(SimpleSize, getShiftAmountTy(N1.getValueType())));
2125 return DAG.getNode(ISD::TRUNCATE, DL, VT, N1);
2132 /// SimplifyNodeWithTwoResults - Perform optimizations common to nodes that
2133 /// compute two values. LoOp and HiOp give the opcodes for the two computations
2134 /// that are being performed. Return true if a simplification was made.
2136 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
2138 // If the high half is not needed, just compute the low half.
2139 bool HiExists = N->hasAnyUseOfValue(1);
2141 (!LegalOperations ||
2142 TLI.isOperationLegal(LoOp, N->getValueType(0)))) {
2143 SDValue Res = DAG.getNode(LoOp, N->getDebugLoc(), N->getValueType(0),
2144 N->op_begin(), N->getNumOperands());
2145 return CombineTo(N, Res, Res);
2148 // If the low half is not needed, just compute the high half.
2149 bool LoExists = N->hasAnyUseOfValue(0);
2151 (!LegalOperations ||
2152 TLI.isOperationLegal(HiOp, N->getValueType(1)))) {
2153 SDValue Res = DAG.getNode(HiOp, N->getDebugLoc(), N->getValueType(1),
2154 N->op_begin(), N->getNumOperands());
2155 return CombineTo(N, Res, Res);
2158 // If both halves are used, return as it is.
2159 if (LoExists && HiExists)
2162 // If the two computed results can be simplified separately, separate them.
2164 SDValue Lo = DAG.getNode(LoOp, N->getDebugLoc(), N->getValueType(0),
2165 N->op_begin(), N->getNumOperands());
2166 AddToWorkList(Lo.getNode());
2167 SDValue LoOpt = combine(Lo.getNode());
2168 if (LoOpt.getNode() && LoOpt.getNode() != Lo.getNode() &&
2169 (!LegalOperations ||
2170 TLI.isOperationLegal(LoOpt.getOpcode(), LoOpt.getValueType())))
2171 return CombineTo(N, LoOpt, LoOpt);
2175 SDValue Hi = DAG.getNode(HiOp, N->getDebugLoc(), N->getValueType(1),
2176 N->op_begin(), N->getNumOperands());
2177 AddToWorkList(Hi.getNode());
2178 SDValue HiOpt = combine(Hi.getNode());
2179 if (HiOpt.getNode() && HiOpt != Hi &&
2180 (!LegalOperations ||
2181 TLI.isOperationLegal(HiOpt.getOpcode(), HiOpt.getValueType())))
2182 return CombineTo(N, HiOpt, HiOpt);
2188 SDValue DAGCombiner::visitSMUL_LOHI(SDNode *N) {
2189 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHS);
2190 if (Res.getNode()) return Res;
2192 EVT VT = N->getValueType(0);
2193 DebugLoc DL = N->getDebugLoc();
2195 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2197 if (VT.isSimple() && !VT.isVector()) {
2198 MVT Simple = VT.getSimpleVT();
2199 unsigned SimpleSize = Simple.getSizeInBits();
2200 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2201 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2202 SDValue Lo = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(0));
2203 SDValue Hi = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, N->getOperand(1));
2204 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2205 // Compute the high part as N1.
2206 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2207 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2208 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2209 // Compute the low part as N0.
2210 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2211 return CombineTo(N, Lo, Hi);
2218 SDValue DAGCombiner::visitUMUL_LOHI(SDNode *N) {
2219 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::MUL, ISD::MULHU);
2220 if (Res.getNode()) return Res;
2222 EVT VT = N->getValueType(0);
2223 DebugLoc DL = N->getDebugLoc();
2225 // If the type twice as wide is legal, transform the mulhu to a wider multiply
2227 if (VT.isSimple() && !VT.isVector()) {
2228 MVT Simple = VT.getSimpleVT();
2229 unsigned SimpleSize = Simple.getSizeInBits();
2230 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), SimpleSize*2);
2231 if (TLI.isOperationLegal(ISD::MUL, NewVT)) {
2232 SDValue Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(0));
2233 SDValue Hi = DAG.getNode(ISD::ZERO_EXTEND, DL, NewVT, N->getOperand(1));
2234 Lo = DAG.getNode(ISD::MUL, DL, NewVT, Lo, Hi);
2235 // Compute the high part as N1.
2236 Hi = DAG.getNode(ISD::SRL, DL, NewVT, Lo,
2237 DAG.getConstant(SimpleSize, getShiftAmountTy(Lo.getValueType())));
2238 Hi = DAG.getNode(ISD::TRUNCATE, DL, VT, Hi);
2239 // Compute the low part as N0.
2240 Lo = DAG.getNode(ISD::TRUNCATE, DL, VT, Lo);
2241 return CombineTo(N, Lo, Hi);
2248 SDValue DAGCombiner::visitSMULO(SDNode *N) {
2249 // (smulo x, 2) -> (saddo x, x)
2250 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2251 if (C2->getAPIntValue() == 2)
2252 return DAG.getNode(ISD::SADDO, N->getDebugLoc(), N->getVTList(),
2253 N->getOperand(0), N->getOperand(0));
2258 SDValue DAGCombiner::visitUMULO(SDNode *N) {
2259 // (umulo x, 2) -> (uaddo x, x)
2260 if (ConstantSDNode *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1)))
2261 if (C2->getAPIntValue() == 2)
2262 return DAG.getNode(ISD::UADDO, N->getDebugLoc(), N->getVTList(),
2263 N->getOperand(0), N->getOperand(0));
2268 SDValue DAGCombiner::visitSDIVREM(SDNode *N) {
2269 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::SDIV, ISD::SREM);
2270 if (Res.getNode()) return Res;
2275 SDValue DAGCombiner::visitUDIVREM(SDNode *N) {
2276 SDValue Res = SimplifyNodeWithTwoResults(N, ISD::UDIV, ISD::UREM);
2277 if (Res.getNode()) return Res;
2282 /// SimplifyBinOpWithSameOpcodeHands - If this is a binary operator with
2283 /// two operands of the same opcode, try to simplify it.
2284 SDValue DAGCombiner::SimplifyBinOpWithSameOpcodeHands(SDNode *N) {
2285 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
2286 EVT VT = N0.getValueType();
2287 assert(N0.getOpcode() == N1.getOpcode() && "Bad input!");
2289 // Bail early if none of these transforms apply.
2290 if (N0.getNode()->getNumOperands() == 0) return SDValue();
2292 // For each of OP in AND/OR/XOR:
2293 // fold (OP (zext x), (zext y)) -> (zext (OP x, y))
2294 // fold (OP (sext x), (sext y)) -> (sext (OP x, y))
2295 // fold (OP (aext x), (aext y)) -> (aext (OP x, y))
2296 // fold (OP (trunc x), (trunc y)) -> (trunc (OP x, y)) (if trunc isn't free)
2298 // do not sink logical op inside of a vector extend, since it may combine
2300 EVT Op0VT = N0.getOperand(0).getValueType();
2301 if ((N0.getOpcode() == ISD::ZERO_EXTEND ||
2302 N0.getOpcode() == ISD::SIGN_EXTEND ||
2303 // Avoid infinite looping with PromoteIntBinOp.
2304 (N0.getOpcode() == ISD::ANY_EXTEND &&
2305 (!LegalTypes || TLI.isTypeDesirableForOp(N->getOpcode(), Op0VT))) ||
2306 (N0.getOpcode() == ISD::TRUNCATE &&
2307 (!TLI.isZExtFree(VT, Op0VT) ||
2308 !TLI.isTruncateFree(Op0VT, VT)) &&
2309 TLI.isTypeLegal(Op0VT))) &&
2311 Op0VT == N1.getOperand(0).getValueType() &&
2312 (!LegalOperations || TLI.isOperationLegal(N->getOpcode(), Op0VT))) {
2313 SDValue ORNode = DAG.getNode(N->getOpcode(), N0.getDebugLoc(),
2314 N0.getOperand(0).getValueType(),
2315 N0.getOperand(0), N1.getOperand(0));
2316 AddToWorkList(ORNode.getNode());
2317 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, ORNode);
2320 // For each of OP in SHL/SRL/SRA/AND...
2321 // fold (and (OP x, z), (OP y, z)) -> (OP (and x, y), z)
2322 // fold (or (OP x, z), (OP y, z)) -> (OP (or x, y), z)
2323 // fold (xor (OP x, z), (OP y, z)) -> (OP (xor x, y), z)
2324 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL ||
2325 N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::AND) &&
2326 N0.getOperand(1) == N1.getOperand(1)) {
2327 SDValue ORNode = DAG.getNode(N->getOpcode(), N0.getDebugLoc(),
2328 N0.getOperand(0).getValueType(),
2329 N0.getOperand(0), N1.getOperand(0));
2330 AddToWorkList(ORNode.getNode());
2331 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
2332 ORNode, N0.getOperand(1));
2335 // Simplify xor/and/or (bitcast(A), bitcast(B)) -> bitcast(op (A,B))
2336 // Only perform this optimization after type legalization and before
2337 // LegalizeVectorOprs. LegalizeVectorOprs promotes vector operations by
2338 // adding bitcasts. For example (xor v4i32) is promoted to (v2i64), and
2339 // we don't want to undo this promotion.
2340 // We also handle SCALAR_TO_VECTOR because xor/or/and operations are cheaper
2342 if ((N0.getOpcode() == ISD::BITCAST || N0.getOpcode() == ISD::SCALAR_TO_VECTOR)
2343 && Level == AfterLegalizeVectorOps) {
2344 SDValue In0 = N0.getOperand(0);
2345 SDValue In1 = N1.getOperand(0);
2346 EVT In0Ty = In0.getValueType();
2347 EVT In1Ty = In1.getValueType();
2348 // If both incoming values are integers, and the original types are the same.
2349 if (In0Ty.isInteger() && In1Ty.isInteger() && In0Ty == In1Ty) {
2350 SDValue Op = DAG.getNode(N->getOpcode(), N->getDebugLoc(), In0Ty, In0, In1);
2351 SDValue BC = DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, Op);
2352 AddToWorkList(Op.getNode());
2357 // Xor/and/or are indifferent to the swizzle operation (shuffle of one value).
2358 // Simplify xor/and/or (shuff(A), shuff(B)) -> shuff(op (A,B))
2359 // If both shuffles use the same mask, and both shuffle within a single
2360 // vector, then it is worthwhile to move the swizzle after the operation.
2361 // The type-legalizer generates this pattern when loading illegal
2362 // vector types from memory. In many cases this allows additional shuffle
2364 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
2365 N0.getOperand(1).getOpcode() == ISD::UNDEF &&
2366 N1.getOperand(1).getOpcode() == ISD::UNDEF) {
2367 ShuffleVectorSDNode *SVN0 = cast<ShuffleVectorSDNode>(N0);
2368 ShuffleVectorSDNode *SVN1 = cast<ShuffleVectorSDNode>(N1);
2370 assert(N0.getOperand(0).getValueType() == N1.getOperand(1).getValueType() &&
2371 "Inputs to shuffles are not the same type");
2373 unsigned NumElts = VT.getVectorNumElements();
2375 // Check that both shuffles use the same mask. The masks are known to be of
2376 // the same length because the result vector type is the same.
2377 bool SameMask = true;
2378 for (unsigned i = 0; i != NumElts; ++i) {
2379 int Idx0 = SVN0->getMaskElt(i);
2380 int Idx1 = SVN1->getMaskElt(i);
2388 SDValue Op = DAG.getNode(N->getOpcode(), N->getDebugLoc(), VT,
2389 N0.getOperand(0), N1.getOperand(0));
2390 AddToWorkList(Op.getNode());
2391 return DAG.getVectorShuffle(VT, N->getDebugLoc(), Op,
2392 DAG.getUNDEF(VT), &SVN0->getMask()[0]);
2399 SDValue DAGCombiner::visitAND(SDNode *N) {
2400 SDValue N0 = N->getOperand(0);
2401 SDValue N1 = N->getOperand(1);
2402 SDValue LL, LR, RL, RR, CC0, CC1;
2403 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2404 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2405 EVT VT = N1.getValueType();
2406 unsigned BitWidth = VT.getScalarType().getSizeInBits();
2409 if (VT.isVector()) {
2410 SDValue FoldedVOp = SimplifyVBinOp(N);
2411 if (FoldedVOp.getNode()) return FoldedVOp;
2414 // fold (and x, undef) -> 0
2415 if (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)
2416 return DAG.getConstant(0, VT);
2417 // fold (and c1, c2) -> c1&c2
2419 return DAG.FoldConstantArithmetic(ISD::AND, VT, N0C, N1C);
2420 // canonicalize constant to RHS
2422 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N1, N0);
2423 // fold (and x, -1) -> x
2424 if (N1C && N1C->isAllOnesValue())
2426 // if (and x, c) is known to be zero, return 0
2427 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
2428 APInt::getAllOnesValue(BitWidth)))
2429 return DAG.getConstant(0, VT);
2431 SDValue RAND = ReassociateOps(ISD::AND, N->getDebugLoc(), N0, N1);
2432 if (RAND.getNode() != 0)
2434 // fold (and (or x, C), D) -> D if (C & D) == D
2435 if (N1C && N0.getOpcode() == ISD::OR)
2436 if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
2437 if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue())
2439 // fold (and (any_ext V), c) -> (zero_ext V) if 'and' only clears top bits.
2440 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
2441 SDValue N0Op0 = N0.getOperand(0);
2442 APInt Mask = ~N1C->getAPIntValue();
2443 Mask = Mask.trunc(N0Op0.getValueSizeInBits());
2444 if (DAG.MaskedValueIsZero(N0Op0, Mask)) {
2445 SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(),
2446 N0.getValueType(), N0Op0);
2448 // Replace uses of the AND with uses of the Zero extend node.
2451 // We actually want to replace all uses of the any_extend with the
2452 // zero_extend, to avoid duplicating things. This will later cause this
2453 // AND to be folded.
2454 CombineTo(N0.getNode(), Zext);
2455 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2458 // similarly fold (and (X (load ([non_ext|any_ext|zero_ext] V))), c) ->
2459 // (X (load ([non_ext|zero_ext] V))) if 'and' only clears top bits which must
2460 // already be zero by virtue of the width of the base type of the load.
2462 // the 'X' node here can either be nothing or an extract_vector_elt to catch
2464 if ((N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
2465 N0.getOperand(0).getOpcode() == ISD::LOAD) ||
2466 N0.getOpcode() == ISD::LOAD) {
2467 LoadSDNode *Load = cast<LoadSDNode>( (N0.getOpcode() == ISD::LOAD) ?
2468 N0 : N0.getOperand(0) );
2470 // Get the constant (if applicable) the zero'th operand is being ANDed with.
2471 // This can be a pure constant or a vector splat, in which case we treat the
2472 // vector as a scalar and use the splat value.
2473 APInt Constant = APInt::getNullValue(1);
2474 if (const ConstantSDNode *C = dyn_cast<ConstantSDNode>(N1)) {
2475 Constant = C->getAPIntValue();
2476 } else if (BuildVectorSDNode *Vector = dyn_cast<BuildVectorSDNode>(N1)) {
2477 APInt SplatValue, SplatUndef;
2478 unsigned SplatBitSize;
2480 bool IsSplat = Vector->isConstantSplat(SplatValue, SplatUndef,
2481 SplatBitSize, HasAnyUndefs);
2483 // Undef bits can contribute to a possible optimisation if set, so
2485 SplatValue |= SplatUndef;
2487 // The splat value may be something like "0x00FFFFFF", which means 0 for
2488 // the first vector value and FF for the rest, repeating. We need a mask
2489 // that will apply equally to all members of the vector, so AND all the
2490 // lanes of the constant together.
2491 EVT VT = Vector->getValueType(0);
2492 unsigned BitWidth = VT.getVectorElementType().getSizeInBits();
2493 Constant = APInt::getAllOnesValue(BitWidth);
2494 for (unsigned i = 0, n = VT.getVectorNumElements(); i < n; ++i)
2495 Constant &= SplatValue.lshr(i*BitWidth).zextOrTrunc(BitWidth);
2499 // If we want to change an EXTLOAD to a ZEXTLOAD, ensure a ZEXTLOAD is
2500 // actually legal and isn't going to get expanded, else this is a false
2502 bool CanZextLoadProfitably = TLI.isLoadExtLegal(ISD::ZEXTLOAD,
2503 Load->getMemoryVT());
2505 // Resize the constant to the same size as the original memory access before
2506 // extension. If it is still the AllOnesValue then this AND is completely
2509 Constant.zextOrTrunc(Load->getMemoryVT().getScalarType().getSizeInBits());
2512 switch (Load->getExtensionType()) {
2513 default: B = false; break;
2514 case ISD::EXTLOAD: B = CanZextLoadProfitably; break;
2516 case ISD::NON_EXTLOAD: B = true; break;
2519 if (B && Constant.isAllOnesValue()) {
2520 // If the load type was an EXTLOAD, convert to ZEXTLOAD in order to
2521 // preserve semantics once we get rid of the AND.
2522 SDValue NewLoad(Load, 0);
2523 if (Load->getExtensionType() == ISD::EXTLOAD) {
2524 NewLoad = DAG.getLoad(Load->getAddressingMode(), ISD::ZEXTLOAD,
2525 Load->getValueType(0), Load->getDebugLoc(),
2526 Load->getChain(), Load->getBasePtr(),
2527 Load->getOffset(), Load->getMemoryVT(),
2528 Load->getMemOperand());
2529 // Replace uses of the EXTLOAD with the new ZEXTLOAD.
2530 CombineTo(Load, NewLoad.getValue(0), NewLoad.getValue(1));
2533 // Fold the AND away, taking care not to fold to the old load node if we
2535 CombineTo(N, (N0.getNode() == Load) ? NewLoad : N0);
2537 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2540 // fold (and (setcc x), (setcc y)) -> (setcc (and x, y))
2541 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
2542 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
2543 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
2545 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
2546 LL.getValueType().isInteger()) {
2547 // fold (and (seteq X, 0), (seteq Y, 0)) -> (seteq (or X, Y), 0)
2548 if (cast<ConstantSDNode>(LR)->isNullValue() && Op1 == ISD::SETEQ) {
2549 SDValue ORNode = DAG.getNode(ISD::OR, N0.getDebugLoc(),
2550 LR.getValueType(), LL, RL);
2551 AddToWorkList(ORNode.getNode());
2552 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
2554 // fold (and (seteq X, -1), (seteq Y, -1)) -> (seteq (and X, Y), -1)
2555 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETEQ) {
2556 SDValue ANDNode = DAG.getNode(ISD::AND, N0.getDebugLoc(),
2557 LR.getValueType(), LL, RL);
2558 AddToWorkList(ANDNode.getNode());
2559 return DAG.getSetCC(N->getDebugLoc(), VT, ANDNode, LR, Op1);
2561 // fold (and (setgt X, -1), (setgt Y, -1)) -> (setgt (or X, Y), -1)
2562 if (cast<ConstantSDNode>(LR)->isAllOnesValue() && Op1 == ISD::SETGT) {
2563 SDValue ORNode = DAG.getNode(ISD::OR, N0.getDebugLoc(),
2564 LR.getValueType(), LL, RL);
2565 AddToWorkList(ORNode.getNode());
2566 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
2569 // canonicalize equivalent to ll == rl
2570 if (LL == RR && LR == RL) {
2571 Op1 = ISD::getSetCCSwappedOperands(Op1);
2574 if (LL == RL && LR == RR) {
2575 bool isInteger = LL.getValueType().isInteger();
2576 ISD::CondCode Result = ISD::getSetCCAndOperation(Op0, Op1, isInteger);
2577 if (Result != ISD::SETCC_INVALID &&
2578 (!LegalOperations || TLI.isCondCodeLegal(Result, LL.getValueType())))
2579 return DAG.getSetCC(N->getDebugLoc(), N0.getValueType(),
2584 // Simplify: (and (op x...), (op y...)) -> (op (and x, y))
2585 if (N0.getOpcode() == N1.getOpcode()) {
2586 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
2587 if (Tmp.getNode()) return Tmp;
2590 // fold (and (sign_extend_inreg x, i16 to i32), 1) -> (and x, 1)
2591 // fold (and (sra)) -> (and (srl)) when possible.
2592 if (!VT.isVector() &&
2593 SimplifyDemandedBits(SDValue(N, 0)))
2594 return SDValue(N, 0);
2596 // fold (zext_inreg (extload x)) -> (zextload x)
2597 if (ISD::isEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode())) {
2598 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2599 EVT MemVT = LN0->getMemoryVT();
2600 // If we zero all the possible extended bits, then we can turn this into
2601 // a zextload if we are running before legalize or the operation is legal.
2602 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2603 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2604 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2605 ((!LegalOperations && !LN0->isVolatile()) ||
2606 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2607 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
2608 LN0->getChain(), LN0->getBasePtr(),
2609 LN0->getPointerInfo(), MemVT,
2610 LN0->isVolatile(), LN0->isNonTemporal(),
2611 LN0->getAlignment());
2613 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2614 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2617 // fold (zext_inreg (sextload x)) -> (zextload x) iff load has one use
2618 if (ISD::isSEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
2620 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
2621 EVT MemVT = LN0->getMemoryVT();
2622 // If we zero all the possible extended bits, then we can turn this into
2623 // a zextload if we are running before legalize or the operation is legal.
2624 unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits();
2625 if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth,
2626 BitWidth - MemVT.getScalarType().getSizeInBits())) &&
2627 ((!LegalOperations && !LN0->isVolatile()) ||
2628 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) {
2629 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT,
2631 LN0->getBasePtr(), LN0->getPointerInfo(),
2633 LN0->isVolatile(), LN0->isNonTemporal(),
2634 LN0->getAlignment());
2636 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
2637 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2641 // fold (and (load x), 255) -> (zextload x, i8)
2642 // fold (and (extload x, i16), 255) -> (zextload x, i8)
2643 // fold (and (any_ext (extload x, i16)), 255) -> (zextload x, i8)
2644 if (N1C && (N0.getOpcode() == ISD::LOAD ||
2645 (N0.getOpcode() == ISD::ANY_EXTEND &&
2646 N0.getOperand(0).getOpcode() == ISD::LOAD))) {
2647 bool HasAnyExt = N0.getOpcode() == ISD::ANY_EXTEND;
2648 LoadSDNode *LN0 = HasAnyExt
2649 ? cast<LoadSDNode>(N0.getOperand(0))
2650 : cast<LoadSDNode>(N0);
2651 if (LN0->getExtensionType() != ISD::SEXTLOAD &&
2652 LN0->isUnindexed() && N0.hasOneUse() && LN0->hasOneUse()) {
2653 uint32_t ActiveBits = N1C->getAPIntValue().getActiveBits();
2654 if (ActiveBits > 0 && APIntOps::isMask(ActiveBits, N1C->getAPIntValue())){
2655 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), ActiveBits);
2656 EVT LoadedVT = LN0->getMemoryVT();
2658 if (ExtVT == LoadedVT &&
2659 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2660 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2663 DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy,
2664 LN0->getChain(), LN0->getBasePtr(),
2665 LN0->getPointerInfo(),
2666 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2667 LN0->getAlignment());
2669 CombineTo(LN0, NewLoad, NewLoad.getValue(1));
2670 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2673 // Do not change the width of a volatile load.
2674 // Do not generate loads of non-round integer types since these can
2675 // be expensive (and would be wrong if the type is not byte sized).
2676 if (!LN0->isVolatile() && LoadedVT.bitsGT(ExtVT) && ExtVT.isRound() &&
2677 (!LegalOperations || TLI.isLoadExtLegal(ISD::ZEXTLOAD, ExtVT))) {
2678 EVT PtrType = LN0->getOperand(1).getValueType();
2680 unsigned Alignment = LN0->getAlignment();
2681 SDValue NewPtr = LN0->getBasePtr();
2683 // For big endian targets, we need to add an offset to the pointer
2684 // to load the correct bytes. For little endian systems, we merely
2685 // need to read fewer bytes from the same pointer.
2686 if (TLI.isBigEndian()) {
2687 unsigned LVTStoreBytes = LoadedVT.getStoreSize();
2688 unsigned EVTStoreBytes = ExtVT.getStoreSize();
2689 unsigned PtrOff = LVTStoreBytes - EVTStoreBytes;
2690 NewPtr = DAG.getNode(ISD::ADD, LN0->getDebugLoc(), PtrType,
2691 NewPtr, DAG.getConstant(PtrOff, PtrType));
2692 Alignment = MinAlign(Alignment, PtrOff);
2695 AddToWorkList(NewPtr.getNode());
2697 EVT LoadResultTy = HasAnyExt ? LN0->getValueType(0) : VT;
2699 DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy,
2700 LN0->getChain(), NewPtr,
2701 LN0->getPointerInfo(),
2702 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
2705 CombineTo(LN0, Load, Load.getValue(1));
2706 return SDValue(N, 0); // Return N so it doesn't get rechecked!
2715 /// MatchBSwapHWord - Match (a >> 8) | (a << 8) as (bswap a) >> 16
2717 SDValue DAGCombiner::MatchBSwapHWordLow(SDNode *N, SDValue N0, SDValue N1,
2718 bool DemandHighBits) {
2719 if (!LegalOperations)
2722 EVT VT = N->getValueType(0);
2723 if (VT != MVT::i64 && VT != MVT::i32 && VT != MVT::i16)
2725 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
2728 // Recognize (and (shl a, 8), 0xff), (and (srl a, 8), 0xff00)
2729 bool LookPassAnd0 = false;
2730 bool LookPassAnd1 = false;
2731 if (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::SRL)
2733 if (N1.getOpcode() == ISD::AND && N1.getOperand(0).getOpcode() == ISD::SHL)
2735 if (N0.getOpcode() == ISD::AND) {
2736 if (!N0.getNode()->hasOneUse())
2738 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2739 if (!N01C || N01C->getZExtValue() != 0xFF00)
2741 N0 = N0.getOperand(0);
2742 LookPassAnd0 = true;
2745 if (N1.getOpcode() == ISD::AND) {
2746 if (!N1.getNode()->hasOneUse())
2748 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
2749 if (!N11C || N11C->getZExtValue() != 0xFF)
2751 N1 = N1.getOperand(0);
2752 LookPassAnd1 = true;
2755 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
2757 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
2759 if (!N0.getNode()->hasOneUse() ||
2760 !N1.getNode()->hasOneUse())
2763 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2764 ConstantSDNode *N11C = dyn_cast<ConstantSDNode>(N1.getOperand(1));
2767 if (N01C->getZExtValue() != 8 || N11C->getZExtValue() != 8)
2770 // Look for (shl (and a, 0xff), 8), (srl (and a, 0xff00), 8)
2771 SDValue N00 = N0->getOperand(0);
2772 if (!LookPassAnd0 && N00.getOpcode() == ISD::AND) {
2773 if (!N00.getNode()->hasOneUse())
2775 ConstantSDNode *N001C = dyn_cast<ConstantSDNode>(N00.getOperand(1));
2776 if (!N001C || N001C->getZExtValue() != 0xFF)
2778 N00 = N00.getOperand(0);
2779 LookPassAnd0 = true;
2782 SDValue N10 = N1->getOperand(0);
2783 if (!LookPassAnd1 && N10.getOpcode() == ISD::AND) {
2784 if (!N10.getNode()->hasOneUse())
2786 ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N10.getOperand(1));
2787 if (!N101C || N101C->getZExtValue() != 0xFF00)
2789 N10 = N10.getOperand(0);
2790 LookPassAnd1 = true;
2796 // Make sure everything beyond the low halfword is zero since the SRL 16
2797 // will clear the top bits.
2798 unsigned OpSizeInBits = VT.getSizeInBits();
2799 if (DemandHighBits && OpSizeInBits > 16 &&
2800 (!LookPassAnd0 || !LookPassAnd1) &&
2801 !DAG.MaskedValueIsZero(N10, APInt::getHighBitsSet(OpSizeInBits, 16)))
2804 SDValue Res = DAG.getNode(ISD::BSWAP, N->getDebugLoc(), VT, N00);
2805 if (OpSizeInBits > 16)
2806 Res = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, Res,
2807 DAG.getConstant(OpSizeInBits-16, getShiftAmountTy(VT)));
2811 /// isBSwapHWordElement - Return true if the specified node is an element
2812 /// that makes up a 32-bit packed halfword byteswap. i.e.
2813 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
2814 static bool isBSwapHWordElement(SDValue N, SmallVector<SDNode*,4> &Parts) {
2815 if (!N.getNode()->hasOneUse())
2818 unsigned Opc = N.getOpcode();
2819 if (Opc != ISD::AND && Opc != ISD::SHL && Opc != ISD::SRL)
2822 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2827 switch (N1C->getZExtValue()) {
2830 case 0xFF: Num = 0; break;
2831 case 0xFF00: Num = 1; break;
2832 case 0xFF0000: Num = 2; break;
2833 case 0xFF000000: Num = 3; break;
2836 // Look for (x & 0xff) << 8 as well as ((x << 8) & 0xff00).
2837 SDValue N0 = N.getOperand(0);
2838 if (Opc == ISD::AND) {
2839 if (Num == 0 || Num == 2) {
2841 // (x >> 8) & 0xff0000
2842 if (N0.getOpcode() != ISD::SRL)
2844 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2845 if (!C || C->getZExtValue() != 8)
2848 // (x << 8) & 0xff00
2849 // (x << 8) & 0xff000000
2850 if (N0.getOpcode() != ISD::SHL)
2852 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
2853 if (!C || C->getZExtValue() != 8)
2856 } else if (Opc == ISD::SHL) {
2858 // (x & 0xff0000) << 8
2859 if (Num != 0 && Num != 2)
2861 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2862 if (!C || C->getZExtValue() != 8)
2864 } else { // Opc == ISD::SRL
2865 // (x & 0xff00) >> 8
2866 // (x & 0xff000000) >> 8
2867 if (Num != 1 && Num != 3)
2869 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N.getOperand(1));
2870 if (!C || C->getZExtValue() != 8)
2877 Parts[Num] = N0.getOperand(0).getNode();
2881 /// MatchBSwapHWord - Match a 32-bit packed halfword bswap. That is
2882 /// ((x&0xff)<<8)|((x&0xff00)>>8)|((x&0x00ff0000)<<8)|((x&0xff000000)>>8)
2883 /// => (rotl (bswap x), 16)
2884 SDValue DAGCombiner::MatchBSwapHWord(SDNode *N, SDValue N0, SDValue N1) {
2885 if (!LegalOperations)
2888 EVT VT = N->getValueType(0);
2891 if (!TLI.isOperationLegal(ISD::BSWAP, VT))
2894 SmallVector<SDNode*,4> Parts(4, (SDNode*)0);
2896 // (or (or (and), (and)), (or (and), (and)))
2897 // (or (or (or (and), (and)), (and)), (and))
2898 if (N0.getOpcode() != ISD::OR)
2900 SDValue N00 = N0.getOperand(0);
2901 SDValue N01 = N0.getOperand(1);
2903 if (N1.getOpcode() == ISD::OR) {
2904 // (or (or (and), (and)), (or (and), (and)))
2905 SDValue N000 = N00.getOperand(0);
2906 if (!isBSwapHWordElement(N000, Parts))
2909 SDValue N001 = N00.getOperand(1);
2910 if (!isBSwapHWordElement(N001, Parts))
2912 SDValue N010 = N01.getOperand(0);
2913 if (!isBSwapHWordElement(N010, Parts))
2915 SDValue N011 = N01.getOperand(1);
2916 if (!isBSwapHWordElement(N011, Parts))
2919 // (or (or (or (and), (and)), (and)), (and))
2920 if (!isBSwapHWordElement(N1, Parts))
2922 if (!isBSwapHWordElement(N01, Parts))
2924 if (N00.getOpcode() != ISD::OR)
2926 SDValue N000 = N00.getOperand(0);
2927 if (!isBSwapHWordElement(N000, Parts))
2929 SDValue N001 = N00.getOperand(1);
2930 if (!isBSwapHWordElement(N001, Parts))
2934 // Make sure the parts are all coming from the same node.
2935 if (Parts[0] != Parts[1] || Parts[0] != Parts[2] || Parts[0] != Parts[3])
2938 SDValue BSwap = DAG.getNode(ISD::BSWAP, N->getDebugLoc(), VT,
2939 SDValue(Parts[0],0));
2941 // Result of the bswap should be rotated by 16. If it's not legal, than
2942 // do (x << 16) | (x >> 16).
2943 SDValue ShAmt = DAG.getConstant(16, getShiftAmountTy(VT));
2944 if (TLI.isOperationLegalOrCustom(ISD::ROTL, VT))
2945 return DAG.getNode(ISD::ROTL, N->getDebugLoc(), VT, BSwap, ShAmt);
2946 else if (TLI.isOperationLegalOrCustom(ISD::ROTR, VT))
2947 return DAG.getNode(ISD::ROTR, N->getDebugLoc(), VT, BSwap, ShAmt);
2948 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT,
2949 DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, BSwap, ShAmt),
2950 DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, BSwap, ShAmt));
2953 SDValue DAGCombiner::visitOR(SDNode *N) {
2954 SDValue N0 = N->getOperand(0);
2955 SDValue N1 = N->getOperand(1);
2956 SDValue LL, LR, RL, RR, CC0, CC1;
2957 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
2958 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
2959 EVT VT = N1.getValueType();
2962 if (VT.isVector()) {
2963 SDValue FoldedVOp = SimplifyVBinOp(N);
2964 if (FoldedVOp.getNode()) return FoldedVOp;
2967 // fold (or x, undef) -> -1
2968 if (!LegalOperations &&
2969 (N0.getOpcode() == ISD::UNDEF || N1.getOpcode() == ISD::UNDEF)) {
2970 EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT;
2971 return DAG.getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT);
2973 // fold (or c1, c2) -> c1|c2
2975 return DAG.FoldConstantArithmetic(ISD::OR, VT, N0C, N1C);
2976 // canonicalize constant to RHS
2978 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N1, N0);
2979 // fold (or x, 0) -> x
2980 if (N1C && N1C->isNullValue())
2982 // fold (or x, -1) -> -1
2983 if (N1C && N1C->isAllOnesValue())
2985 // fold (or x, c) -> c iff (x & ~c) == 0
2986 if (N1C && DAG.MaskedValueIsZero(N0, ~N1C->getAPIntValue()))
2989 // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
2990 SDValue BSwap = MatchBSwapHWord(N, N0, N1);
2991 if (BSwap.getNode() != 0)
2993 BSwap = MatchBSwapHWordLow(N, N0, N1);
2994 if (BSwap.getNode() != 0)
2998 SDValue ROR = ReassociateOps(ISD::OR, N->getDebugLoc(), N0, N1);
2999 if (ROR.getNode() != 0)
3001 // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
3002 // iff (c1 & c2) == 0.
3003 if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
3004 isa<ConstantSDNode>(N0.getOperand(1))) {
3005 ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
3006 if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0)
3007 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
3008 DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
3009 N0.getOperand(0), N1),
3010 DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1));
3012 // fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
3013 if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
3014 ISD::CondCode Op0 = cast<CondCodeSDNode>(CC0)->get();
3015 ISD::CondCode Op1 = cast<CondCodeSDNode>(CC1)->get();
3017 if (LR == RR && isa<ConstantSDNode>(LR) && Op0 == Op1 &&
3018 LL.getValueType().isInteger()) {
3019 // fold (or (setne X, 0), (setne Y, 0)) -> (setne (or X, Y), 0)
3020 // fold (or (setlt X, 0), (setlt Y, 0)) -> (setne (or X, Y), 0)
3021 if (cast<ConstantSDNode>(LR)->isNullValue() &&
3022 (Op1 == ISD::SETNE || Op1 == ISD::SETLT)) {
3023 SDValue ORNode = DAG.getNode(ISD::OR, LR.getDebugLoc(),
3024 LR.getValueType(), LL, RL);
3025 AddToWorkList(ORNode.getNode());
3026 return DAG.getSetCC(N->getDebugLoc(), VT, ORNode, LR, Op1);
3028 // fold (or (setne X, -1), (setne Y, -1)) -> (setne (and X, Y), -1)
3029 // fold (or (setgt X, -1), (setgt Y -1)) -> (setgt (and X, Y), -1)
3030 if (cast<ConstantSDNode>(LR)->isAllOnesValue() &&
3031 (Op1 == ISD::SETNE || Op1 == ISD::SETGT)) {
3032 SDValue ANDNode = DAG.getNode(ISD::AND, LR.getDebugLoc(),
3033 LR.getValueType(), LL, RL);
3034 AddToWorkList(ANDNode.getNode());
3035 return DAG.getSetCC(N->getDebugLoc(), VT, ANDNode, LR, Op1);
3038 // canonicalize equivalent to ll == rl
3039 if (LL == RR && LR == RL) {
3040 Op1 = ISD::getSetCCSwappedOperands(Op1);
3043 if (LL == RL && LR == RR) {
3044 bool isInteger = LL.getValueType().isInteger();
3045 ISD::CondCode Result = ISD::getSetCCOrOperation(Op0, Op1, isInteger);
3046 if (Result != ISD::SETCC_INVALID &&
3047 (!LegalOperations || TLI.isCondCodeLegal(Result, LL.getValueType())))
3048 return DAG.getSetCC(N->getDebugLoc(), N0.getValueType(),
3053 // Simplify: (or (op x...), (op y...)) -> (op (or x, y))
3054 if (N0.getOpcode() == N1.getOpcode()) {
3055 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3056 if (Tmp.getNode()) return Tmp;
3059 // (or (and X, C1), (and Y, C2)) -> (and (or X, Y), C3) if possible.
3060 if (N0.getOpcode() == ISD::AND &&
3061 N1.getOpcode() == ISD::AND &&
3062 N0.getOperand(1).getOpcode() == ISD::Constant &&
3063 N1.getOperand(1).getOpcode() == ISD::Constant &&
3064 // Don't increase # computations.
3065 (N0.getNode()->hasOneUse() || N1.getNode()->hasOneUse())) {
3066 // We can only do this xform if we know that bits from X that are set in C2
3067 // but not in C1 are already zero. Likewise for Y.
3068 const APInt &LHSMask =
3069 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
3070 const APInt &RHSMask =
3071 cast<ConstantSDNode>(N1.getOperand(1))->getAPIntValue();
3073 if (DAG.MaskedValueIsZero(N0.getOperand(0), RHSMask&~LHSMask) &&
3074 DAG.MaskedValueIsZero(N1.getOperand(0), LHSMask&~RHSMask)) {
3075 SDValue X = DAG.getNode(ISD::OR, N0.getDebugLoc(), VT,
3076 N0.getOperand(0), N1.getOperand(0));
3077 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, X,
3078 DAG.getConstant(LHSMask | RHSMask, VT));
3082 // See if this is some rotate idiom.
3083 if (SDNode *Rot = MatchRotate(N0, N1, N->getDebugLoc()))
3084 return SDValue(Rot, 0);
3086 // Simplify the operands using demanded-bits information.
3087 if (!VT.isVector() &&
3088 SimplifyDemandedBits(SDValue(N, 0)))
3089 return SDValue(N, 0);
3094 /// MatchRotateHalf - Match "(X shl/srl V1) & V2" where V2 may not be present.
3095 static bool MatchRotateHalf(SDValue Op, SDValue &Shift, SDValue &Mask) {
3096 if (Op.getOpcode() == ISD::AND) {
3097 if (isa<ConstantSDNode>(Op.getOperand(1))) {
3098 Mask = Op.getOperand(1);
3099 Op = Op.getOperand(0);
3105 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SHL) {
3113 // MatchRotate - Handle an 'or' of two operands. If this is one of the many
3114 // idioms for rotate, and if the target supports rotation instructions, generate
3116 SDNode *DAGCombiner::MatchRotate(SDValue LHS, SDValue RHS, DebugLoc DL) {
3117 // Must be a legal type. Expanded 'n promoted things won't work with rotates.
3118 EVT VT = LHS.getValueType();
3119 if (!TLI.isTypeLegal(VT)) return 0;
3121 // The target must have at least one rotate flavor.
3122 bool HasROTL = TLI.isOperationLegalOrCustom(ISD::ROTL, VT);
3123 bool HasROTR = TLI.isOperationLegalOrCustom(ISD::ROTR, VT);
3124 if (!HasROTL && !HasROTR) return 0;
3126 // Match "(X shl/srl V1) & V2" where V2 may not be present.
3127 SDValue LHSShift; // The shift.
3128 SDValue LHSMask; // AND value if any.
3129 if (!MatchRotateHalf(LHS, LHSShift, LHSMask))
3130 return 0; // Not part of a rotate.
3132 SDValue RHSShift; // The shift.
3133 SDValue RHSMask; // AND value if any.
3134 if (!MatchRotateHalf(RHS, RHSShift, RHSMask))
3135 return 0; // Not part of a rotate.
3137 if (LHSShift.getOperand(0) != RHSShift.getOperand(0))
3138 return 0; // Not shifting the same value.
3140 if (LHSShift.getOpcode() == RHSShift.getOpcode())
3141 return 0; // Shifts must disagree.
3143 // Canonicalize shl to left side in a shl/srl pair.
3144 if (RHSShift.getOpcode() == ISD::SHL) {
3145 std::swap(LHS, RHS);
3146 std::swap(LHSShift, RHSShift);
3147 std::swap(LHSMask , RHSMask );
3150 unsigned OpSizeInBits = VT.getSizeInBits();
3151 SDValue LHSShiftArg = LHSShift.getOperand(0);
3152 SDValue LHSShiftAmt = LHSShift.getOperand(1);
3153 SDValue RHSShiftAmt = RHSShift.getOperand(1);
3155 // fold (or (shl x, C1), (srl x, C2)) -> (rotl x, C1)
3156 // fold (or (shl x, C1), (srl x, C2)) -> (rotr x, C2)
3157 if (LHSShiftAmt.getOpcode() == ISD::Constant &&
3158 RHSShiftAmt.getOpcode() == ISD::Constant) {
3159 uint64_t LShVal = cast<ConstantSDNode>(LHSShiftAmt)->getZExtValue();
3160 uint64_t RShVal = cast<ConstantSDNode>(RHSShiftAmt)->getZExtValue();
3161 if ((LShVal + RShVal) != OpSizeInBits)
3166 Rot = DAG.getNode(ISD::ROTL, DL, VT, LHSShiftArg, LHSShiftAmt);
3168 Rot = DAG.getNode(ISD::ROTR, DL, VT, LHSShiftArg, RHSShiftAmt);
3170 // If there is an AND of either shifted operand, apply it to the result.
3171 if (LHSMask.getNode() || RHSMask.getNode()) {
3172 APInt Mask = APInt::getAllOnesValue(OpSizeInBits);
3174 if (LHSMask.getNode()) {
3175 APInt RHSBits = APInt::getLowBitsSet(OpSizeInBits, LShVal);
3176 Mask &= cast<ConstantSDNode>(LHSMask)->getAPIntValue() | RHSBits;
3178 if (RHSMask.getNode()) {
3179 APInt LHSBits = APInt::getHighBitsSet(OpSizeInBits, RShVal);
3180 Mask &= cast<ConstantSDNode>(RHSMask)->getAPIntValue() | LHSBits;
3183 Rot = DAG.getNode(ISD::AND, DL, VT, Rot, DAG.getConstant(Mask, VT));
3186 return Rot.getNode();
3189 // If there is a mask here, and we have a variable shift, we can't be sure
3190 // that we're masking out the right stuff.
3191 if (LHSMask.getNode() || RHSMask.getNode())
3194 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotl x, y)
3195 // fold (or (shl x, y), (srl x, (sub 32, y))) -> (rotr x, (sub 32, y))
3196 if (RHSShiftAmt.getOpcode() == ISD::SUB &&
3197 LHSShiftAmt == RHSShiftAmt.getOperand(1)) {
3198 if (ConstantSDNode *SUBC =
3199 dyn_cast<ConstantSDNode>(RHSShiftAmt.getOperand(0))) {
3200 if (SUBC->getAPIntValue() == OpSizeInBits) {
3202 return DAG.getNode(ISD::ROTL, DL, VT,
3203 LHSShiftArg, LHSShiftAmt).getNode();
3205 return DAG.getNode(ISD::ROTR, DL, VT,
3206 LHSShiftArg, RHSShiftAmt).getNode();
3211 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotr x, y)
3212 // fold (or (shl x, (sub 32, y)), (srl x, r)) -> (rotl x, (sub 32, y))
3213 if (LHSShiftAmt.getOpcode() == ISD::SUB &&
3214 RHSShiftAmt == LHSShiftAmt.getOperand(1)) {
3215 if (ConstantSDNode *SUBC =
3216 dyn_cast<ConstantSDNode>(LHSShiftAmt.getOperand(0))) {
3217 if (SUBC->getAPIntValue() == OpSizeInBits) {
3219 return DAG.getNode(ISD::ROTR, DL, VT,
3220 LHSShiftArg, RHSShiftAmt).getNode();
3222 return DAG.getNode(ISD::ROTL, DL, VT,
3223 LHSShiftArg, LHSShiftAmt).getNode();
3228 // Look for sign/zext/any-extended or truncate cases:
3229 if ((LHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND
3230 || LHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND
3231 || LHSShiftAmt.getOpcode() == ISD::ANY_EXTEND
3232 || LHSShiftAmt.getOpcode() == ISD::TRUNCATE) &&
3233 (RHSShiftAmt.getOpcode() == ISD::SIGN_EXTEND
3234 || RHSShiftAmt.getOpcode() == ISD::ZERO_EXTEND
3235 || RHSShiftAmt.getOpcode() == ISD::ANY_EXTEND
3236 || RHSShiftAmt.getOpcode() == ISD::TRUNCATE)) {
3237 SDValue LExtOp0 = LHSShiftAmt.getOperand(0);
3238 SDValue RExtOp0 = RHSShiftAmt.getOperand(0);
3239 if (RExtOp0.getOpcode() == ISD::SUB &&
3240 RExtOp0.getOperand(1) == LExtOp0) {
3241 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
3243 // fold (or (shl x, (*ext y)), (srl x, (*ext (sub 32, y)))) ->
3244 // (rotr x, (sub 32, y))
3245 if (ConstantSDNode *SUBC =
3246 dyn_cast<ConstantSDNode>(RExtOp0.getOperand(0))) {
3247 if (SUBC->getAPIntValue() == OpSizeInBits) {
3248 return DAG.getNode(HasROTL ? ISD::ROTL : ISD::ROTR, DL, VT,
3250 HasROTL ? LHSShiftAmt : RHSShiftAmt).getNode();
3253 } else if (LExtOp0.getOpcode() == ISD::SUB &&
3254 RExtOp0 == LExtOp0.getOperand(1)) {
3255 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
3257 // fold (or (shl x, (*ext (sub 32, y))), (srl x, (*ext y))) ->
3258 // (rotl x, (sub 32, y))
3259 if (ConstantSDNode *SUBC =
3260 dyn_cast<ConstantSDNode>(LExtOp0.getOperand(0))) {
3261 if (SUBC->getAPIntValue() == OpSizeInBits) {
3262 return DAG.getNode(HasROTR ? ISD::ROTR : ISD::ROTL, DL, VT,
3264 HasROTR ? RHSShiftAmt : LHSShiftAmt).getNode();
3273 SDValue DAGCombiner::visitXOR(SDNode *N) {
3274 SDValue N0 = N->getOperand(0);
3275 SDValue N1 = N->getOperand(1);
3276 SDValue LHS, RHS, CC;
3277 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3278 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3279 EVT VT = N0.getValueType();
3282 if (VT.isVector()) {
3283 SDValue FoldedVOp = SimplifyVBinOp(N);
3284 if (FoldedVOp.getNode()) return FoldedVOp;
3287 // fold (xor undef, undef) -> 0. This is a common idiom (misuse).
3288 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
3289 return DAG.getConstant(0, VT);
3290 // fold (xor x, undef) -> undef
3291 if (N0.getOpcode() == ISD::UNDEF)
3293 if (N1.getOpcode() == ISD::UNDEF)
3295 // fold (xor c1, c2) -> c1^c2
3297 return DAG.FoldConstantArithmetic(ISD::XOR, VT, N0C, N1C);
3298 // canonicalize constant to RHS
3300 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N1, N0);
3301 // fold (xor x, 0) -> x
3302 if (N1C && N1C->isNullValue())
3305 SDValue RXOR = ReassociateOps(ISD::XOR, N->getDebugLoc(), N0, N1);
3306 if (RXOR.getNode() != 0)
3309 // fold !(x cc y) -> (x !cc y)
3310 if (N1C && N1C->getAPIntValue() == 1 && isSetCCEquivalent(N0, LHS, RHS, CC)) {
3311 bool isInt = LHS.getValueType().isInteger();
3312 ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
3315 if (!LegalOperations || TLI.isCondCodeLegal(NotCC, LHS.getValueType())) {
3316 switch (N0.getOpcode()) {
3318 llvm_unreachable("Unhandled SetCC Equivalent!");
3320 return DAG.getSetCC(N->getDebugLoc(), VT, LHS, RHS, NotCC);
3321 case ISD::SELECT_CC:
3322 return DAG.getSelectCC(N->getDebugLoc(), LHS, RHS, N0.getOperand(2),
3323 N0.getOperand(3), NotCC);
3328 // fold (not (zext (setcc x, y))) -> (zext (not (setcc x, y)))
3329 if (N1C && N1C->getAPIntValue() == 1 && N0.getOpcode() == ISD::ZERO_EXTEND &&
3330 N0.getNode()->hasOneUse() &&
3331 isSetCCEquivalent(N0.getOperand(0), LHS, RHS, CC)){
3332 SDValue V = N0.getOperand(0);
3333 V = DAG.getNode(ISD::XOR, N0.getDebugLoc(), V.getValueType(), V,
3334 DAG.getConstant(1, V.getValueType()));
3335 AddToWorkList(V.getNode());
3336 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, V);
3339 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are setcc
3340 if (N1C && N1C->getAPIntValue() == 1 && VT == MVT::i1 &&
3341 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3342 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3343 if (isOneUseSetCC(RHS) || isOneUseSetCC(LHS)) {
3344 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3345 LHS = DAG.getNode(ISD::XOR, LHS.getDebugLoc(), VT, LHS, N1); // LHS = ~LHS
3346 RHS = DAG.getNode(ISD::XOR, RHS.getDebugLoc(), VT, RHS, N1); // RHS = ~RHS
3347 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3348 return DAG.getNode(NewOpcode, N->getDebugLoc(), VT, LHS, RHS);
3351 // fold (not (or x, y)) -> (and (not x), (not y)) iff x or y are constants
3352 if (N1C && N1C->isAllOnesValue() &&
3353 (N0.getOpcode() == ISD::OR || N0.getOpcode() == ISD::AND)) {
3354 SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
3355 if (isa<ConstantSDNode>(RHS) || isa<ConstantSDNode>(LHS)) {
3356 unsigned NewOpcode = N0.getOpcode() == ISD::AND ? ISD::OR : ISD::AND;
3357 LHS = DAG.getNode(ISD::XOR, LHS.getDebugLoc(), VT, LHS, N1); // LHS = ~LHS
3358 RHS = DAG.getNode(ISD::XOR, RHS.getDebugLoc(), VT, RHS, N1); // RHS = ~RHS
3359 AddToWorkList(LHS.getNode()); AddToWorkList(RHS.getNode());
3360 return DAG.getNode(NewOpcode, N->getDebugLoc(), VT, LHS, RHS);
3363 // fold (xor (xor x, c1), c2) -> (xor x, (xor c1, c2))
3364 if (N1C && N0.getOpcode() == ISD::XOR) {
3365 ConstantSDNode *N00C = dyn_cast<ConstantSDNode>(N0.getOperand(0));
3366 ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3368 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N0.getOperand(1),
3369 DAG.getConstant(N1C->getAPIntValue() ^
3370 N00C->getAPIntValue(), VT));
3372 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT, N0.getOperand(0),
3373 DAG.getConstant(N1C->getAPIntValue() ^
3374 N01C->getAPIntValue(), VT));
3376 // fold (xor x, x) -> 0
3378 return tryFoldToZero(N->getDebugLoc(), TLI, VT, DAG, LegalOperations);
3380 // Simplify: xor (op x...), (op y...) -> (op (xor x, y))
3381 if (N0.getOpcode() == N1.getOpcode()) {
3382 SDValue Tmp = SimplifyBinOpWithSameOpcodeHands(N);
3383 if (Tmp.getNode()) return Tmp;
3386 // Simplify the expression using non-local knowledge.
3387 if (!VT.isVector() &&
3388 SimplifyDemandedBits(SDValue(N, 0)))
3389 return SDValue(N, 0);
3394 /// visitShiftByConstant - Handle transforms common to the three shifts, when
3395 /// the shift amount is a constant.
3396 SDValue DAGCombiner::visitShiftByConstant(SDNode *N, unsigned Amt) {
3397 SDNode *LHS = N->getOperand(0).getNode();
3398 if (!LHS->hasOneUse()) return SDValue();
3400 // We want to pull some binops through shifts, so that we have (and (shift))
3401 // instead of (shift (and)), likewise for add, or, xor, etc. This sort of
3402 // thing happens with address calculations, so it's important to canonicalize
3404 bool HighBitSet = false; // Can we transform this if the high bit is set?
3406 switch (LHS->getOpcode()) {
3407 default: return SDValue();
3410 HighBitSet = false; // We can only transform sra if the high bit is clear.
3413 HighBitSet = true; // We can only transform sra if the high bit is set.
3416 if (N->getOpcode() != ISD::SHL)
3417 return SDValue(); // only shl(add) not sr[al](add).
3418 HighBitSet = false; // We can only transform sra if the high bit is clear.
3422 // We require the RHS of the binop to be a constant as well.
3423 ConstantSDNode *BinOpCst = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
3424 if (!BinOpCst) return SDValue();
3426 // FIXME: disable this unless the input to the binop is a shift by a constant.
3427 // If it is not a shift, it pessimizes some common cases like:
3429 // void foo(int *X, int i) { X[i & 1235] = 1; }
3430 // int bar(int *X, int i) { return X[i & 255]; }
3431 SDNode *BinOpLHSVal = LHS->getOperand(0).getNode();
3432 if ((BinOpLHSVal->getOpcode() != ISD::SHL &&
3433 BinOpLHSVal->getOpcode() != ISD::SRA &&
3434 BinOpLHSVal->getOpcode() != ISD::SRL) ||
3435 !isa<ConstantSDNode>(BinOpLHSVal->getOperand(1)))
3438 EVT VT = N->getValueType(0);
3440 // If this is a signed shift right, and the high bit is modified by the
3441 // logical operation, do not perform the transformation. The highBitSet
3442 // boolean indicates the value of the high bit of the constant which would
3443 // cause it to be modified for this operation.
3444 if (N->getOpcode() == ISD::SRA) {
3445 bool BinOpRHSSignSet = BinOpCst->getAPIntValue().isNegative();
3446 if (BinOpRHSSignSet != HighBitSet)
3450 // Fold the constants, shifting the binop RHS by the shift amount.
3451 SDValue NewRHS = DAG.getNode(N->getOpcode(), LHS->getOperand(1).getDebugLoc(),
3453 LHS->getOperand(1), N->getOperand(1));
3455 // Create the new shift.
3456 SDValue NewShift = DAG.getNode(N->getOpcode(),
3457 LHS->getOperand(0).getDebugLoc(),
3458 VT, LHS->getOperand(0), N->getOperand(1));
3460 // Create the new binop.
3461 return DAG.getNode(LHS->getOpcode(), N->getDebugLoc(), VT, NewShift, NewRHS);
3464 SDValue DAGCombiner::visitSHL(SDNode *N) {
3465 SDValue N0 = N->getOperand(0);
3466 SDValue N1 = N->getOperand(1);
3467 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3468 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3469 EVT VT = N0.getValueType();
3470 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3472 // fold (shl c1, c2) -> c1<<c2
3474 return DAG.FoldConstantArithmetic(ISD::SHL, VT, N0C, N1C);
3475 // fold (shl 0, x) -> 0
3476 if (N0C && N0C->isNullValue())
3478 // fold (shl x, c >= size(x)) -> undef
3479 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3480 return DAG.getUNDEF(VT);
3481 // fold (shl x, 0) -> x
3482 if (N1C && N1C->isNullValue())
3484 // fold (shl undef, x) -> 0
3485 if (N0.getOpcode() == ISD::UNDEF)
3486 return DAG.getConstant(0, VT);
3487 // if (shl x, c) is known to be zero, return 0
3488 if (DAG.MaskedValueIsZero(SDValue(N, 0),
3489 APInt::getAllOnesValue(OpSizeInBits)))
3490 return DAG.getConstant(0, VT);
3491 // fold (shl x, (trunc (and y, c))) -> (shl x, (and (trunc y), (trunc c))).
3492 if (N1.getOpcode() == ISD::TRUNCATE &&
3493 N1.getOperand(0).getOpcode() == ISD::AND &&
3494 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3495 SDValue N101 = N1.getOperand(0).getOperand(1);
3496 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3497 EVT TruncVT = N1.getValueType();
3498 SDValue N100 = N1.getOperand(0).getOperand(0);
3499 APInt TruncC = N101C->getAPIntValue();
3500 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
3501 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0,
3502 DAG.getNode(ISD::AND, N->getDebugLoc(), TruncVT,
3503 DAG.getNode(ISD::TRUNCATE,
3506 DAG.getConstant(TruncC, TruncVT)));
3510 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3511 return SDValue(N, 0);
3513 // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
3514 if (N1C && N0.getOpcode() == ISD::SHL &&
3515 N0.getOperand(1).getOpcode() == ISD::Constant) {
3516 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3517 uint64_t c2 = N1C->getZExtValue();
3518 if (c1 + c2 >= OpSizeInBits)
3519 return DAG.getConstant(0, VT);
3520 return DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0.getOperand(0),
3521 DAG.getConstant(c1 + c2, N1.getValueType()));
3524 // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
3525 // For this to be valid, the second form must not preserve any of the bits
3526 // that are shifted out by the inner shift in the first form. This means
3527 // the outer shift size must be >= the number of bits added by the ext.
3528 // As a corollary, we don't care what kind of ext it is.
3529 if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
3530 N0.getOpcode() == ISD::ANY_EXTEND ||
3531 N0.getOpcode() == ISD::SIGN_EXTEND) &&
3532 N0.getOperand(0).getOpcode() == ISD::SHL &&
3533 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
3535 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
3536 uint64_t c2 = N1C->getZExtValue();
3537 EVT InnerShiftVT = N0.getOperand(0).getValueType();
3538 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
3539 if (c2 >= OpSizeInBits - InnerShiftSize) {
3540 if (c1 + c2 >= OpSizeInBits)
3541 return DAG.getConstant(0, VT);
3542 return DAG.getNode(ISD::SHL, N0->getDebugLoc(), VT,
3543 DAG.getNode(N0.getOpcode(), N0->getDebugLoc(), VT,
3544 N0.getOperand(0)->getOperand(0)),
3545 DAG.getConstant(c1 + c2, N1.getValueType()));
3549 // fold (shl (srl x, c1), c2) -> (and (shl x, (sub c2, c1), MASK) or
3550 // (and (srl x, (sub c1, c2), MASK)
3551 // Only fold this if the inner shift has no other uses -- if it does, folding
3552 // this will increase the total number of instructions.
3553 if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse() &&
3554 N0.getOperand(1).getOpcode() == ISD::Constant) {
3555 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3556 if (c1 < VT.getSizeInBits()) {
3557 uint64_t c2 = N1C->getZExtValue();
3558 APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
3559 VT.getSizeInBits() - c1);
3562 Mask = Mask.shl(c2-c1);
3563 Shift = DAG.getNode(ISD::SHL, N->getDebugLoc(), VT, N0.getOperand(0),
3564 DAG.getConstant(c2-c1, N1.getValueType()));
3566 Mask = Mask.lshr(c1-c2);
3567 Shift = DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0),
3568 DAG.getConstant(c1-c2, N1.getValueType()));
3570 return DAG.getNode(ISD::AND, N0.getDebugLoc(), VT, Shift,
3571 DAG.getConstant(Mask, VT));
3574 // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
3575 if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
3576 SDValue HiBitsMask =
3577 DAG.getConstant(APInt::getHighBitsSet(VT.getSizeInBits(),
3578 VT.getSizeInBits() -
3579 N1C->getZExtValue()),
3581 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
3586 SDValue NewSHL = visitShiftByConstant(N, N1C->getZExtValue());
3587 if (NewSHL.getNode())
3594 SDValue DAGCombiner::visitSRA(SDNode *N) {
3595 SDValue N0 = N->getOperand(0);
3596 SDValue N1 = N->getOperand(1);
3597 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3598 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3599 EVT VT = N0.getValueType();
3600 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3602 // fold (sra c1, c2) -> (sra c1, c2)
3604 return DAG.FoldConstantArithmetic(ISD::SRA, VT, N0C, N1C);
3605 // fold (sra 0, x) -> 0
3606 if (N0C && N0C->isNullValue())
3608 // fold (sra -1, x) -> -1
3609 if (N0C && N0C->isAllOnesValue())
3611 // fold (sra x, (setge c, size(x))) -> undef
3612 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3613 return DAG.getUNDEF(VT);
3614 // fold (sra x, 0) -> x
3615 if (N1C && N1C->isNullValue())
3617 // fold (sra (shl x, c1), c1) -> sext_inreg for some c1 and target supports
3619 if (N1C && N0.getOpcode() == ISD::SHL && N1 == N0.getOperand(1)) {
3620 unsigned LowBits = OpSizeInBits - (unsigned)N1C->getZExtValue();
3621 EVT ExtVT = EVT::getIntegerVT(*DAG.getContext(), LowBits);
3623 ExtVT = EVT::getVectorVT(*DAG.getContext(),
3624 ExtVT, VT.getVectorNumElements());
3625 if ((!LegalOperations ||
3626 TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG, ExtVT)))
3627 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
3628 N0.getOperand(0), DAG.getValueType(ExtVT));
3631 // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
3632 if (N1C && N0.getOpcode() == ISD::SRA) {
3633 if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
3634 unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
3635 if (Sum >= OpSizeInBits) Sum = OpSizeInBits-1;
3636 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0.getOperand(0),
3637 DAG.getConstant(Sum, N1C->getValueType(0)));
3641 // fold (sra (shl X, m), (sub result_size, n))
3642 // -> (sign_extend (trunc (shl X, (sub (sub result_size, n), m)))) for
3643 // result_size - n != m.
3644 // If truncate is free for the target sext(shl) is likely to result in better
3646 if (N0.getOpcode() == ISD::SHL) {
3647 // Get the two constanst of the shifts, CN0 = m, CN = n.
3648 const ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
3650 // Determine what the truncate's result bitsize and type would be.
3652 EVT::getIntegerVT(*DAG.getContext(),
3653 OpSizeInBits - N1C->getZExtValue());
3654 // Determine the residual right-shift amount.
3655 signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
3657 // If the shift is not a no-op (in which case this should be just a sign
3658 // extend already), the truncated to type is legal, sign_extend is legal
3659 // on that type, and the truncate to that type is both legal and free,
3660 // perform the transform.
3661 if ((ShiftAmt > 0) &&
3662 TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND, TruncVT) &&
3663 TLI.isOperationLegalOrCustom(ISD::TRUNCATE, VT) &&
3664 TLI.isTruncateFree(VT, TruncVT)) {
3666 SDValue Amt = DAG.getConstant(ShiftAmt,
3667 getShiftAmountTy(N0.getOperand(0).getValueType()));
3668 SDValue Shift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), VT,
3669 N0.getOperand(0), Amt);
3670 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), TruncVT,
3672 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(),
3673 N->getValueType(0), Trunc);
3678 // fold (sra x, (trunc (and y, c))) -> (sra x, (and (trunc y), (trunc c))).
3679 if (N1.getOpcode() == ISD::TRUNCATE &&
3680 N1.getOperand(0).getOpcode() == ISD::AND &&
3681 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3682 SDValue N101 = N1.getOperand(0).getOperand(1);
3683 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3684 EVT TruncVT = N1.getValueType();
3685 SDValue N100 = N1.getOperand(0).getOperand(0);
3686 APInt TruncC = N101C->getAPIntValue();
3687 TruncC = TruncC.trunc(TruncVT.getScalarType().getSizeInBits());
3688 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT, N0,
3689 DAG.getNode(ISD::AND, N->getDebugLoc(),
3691 DAG.getNode(ISD::TRUNCATE,
3694 DAG.getConstant(TruncC, TruncVT)));
3698 // fold (sra (trunc (sr x, c1)), c2) -> (trunc (sra x, c1+c2))
3699 // if c1 is equal to the number of bits the trunc removes
3700 if (N0.getOpcode() == ISD::TRUNCATE &&
3701 (N0.getOperand(0).getOpcode() == ISD::SRL ||
3702 N0.getOperand(0).getOpcode() == ISD::SRA) &&
3703 N0.getOperand(0).hasOneUse() &&
3704 N0.getOperand(0).getOperand(1).hasOneUse() &&
3705 N1C && isa<ConstantSDNode>(N0.getOperand(0).getOperand(1))) {
3706 EVT LargeVT = N0.getOperand(0).getValueType();
3707 ConstantSDNode *LargeShiftAmt =
3708 cast<ConstantSDNode>(N0.getOperand(0).getOperand(1));
3710 if (LargeVT.getScalarType().getSizeInBits() - OpSizeInBits ==
3711 LargeShiftAmt->getZExtValue()) {
3713 DAG.getConstant(LargeShiftAmt->getZExtValue() + N1C->getZExtValue(),
3714 getShiftAmountTy(N0.getOperand(0).getOperand(0).getValueType()));
3715 SDValue SRA = DAG.getNode(ISD::SRA, N->getDebugLoc(), LargeVT,
3716 N0.getOperand(0).getOperand(0), Amt);
3717 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, SRA);
3721 // Simplify, based on bits shifted out of the LHS.
3722 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3723 return SDValue(N, 0);
3726 // If the sign bit is known to be zero, switch this to a SRL.
3727 if (DAG.SignBitIsZero(N0))
3728 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0, N1);
3731 SDValue NewSRA = visitShiftByConstant(N, N1C->getZExtValue());
3732 if (NewSRA.getNode())
3739 SDValue DAGCombiner::visitSRL(SDNode *N) {
3740 SDValue N0 = N->getOperand(0);
3741 SDValue N1 = N->getOperand(1);
3742 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3743 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3744 EVT VT = N0.getValueType();
3745 unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
3747 // fold (srl c1, c2) -> c1 >>u c2
3749 return DAG.FoldConstantArithmetic(ISD::SRL, VT, N0C, N1C);
3750 // fold (srl 0, x) -> 0
3751 if (N0C && N0C->isNullValue())
3753 // fold (srl x, c >= size(x)) -> undef
3754 if (N1C && N1C->getZExtValue() >= OpSizeInBits)
3755 return DAG.getUNDEF(VT);
3756 // fold (srl x, 0) -> x
3757 if (N1C && N1C->isNullValue())
3759 // if (srl x, c) is known to be zero, return 0
3760 if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
3761 APInt::getAllOnesValue(OpSizeInBits)))
3762 return DAG.getConstant(0, VT);
3764 // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
3765 if (N1C && N0.getOpcode() == ISD::SRL &&
3766 N0.getOperand(1).getOpcode() == ISD::Constant) {
3767 uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
3768 uint64_t c2 = N1C->getZExtValue();
3769 if (c1 + c2 >= OpSizeInBits)
3770 return DAG.getConstant(0, VT);
3771 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0),
3772 DAG.getConstant(c1 + c2, N1.getValueType()));
3775 // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
3776 if (N1C && N0.getOpcode() == ISD::TRUNCATE &&
3777 N0.getOperand(0).getOpcode() == ISD::SRL &&
3778 isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
3780 cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
3781 uint64_t c2 = N1C->getZExtValue();
3782 EVT InnerShiftVT = N0.getOperand(0).getValueType();
3783 EVT ShiftCountVT = N0.getOperand(0)->getOperand(1).getValueType();
3784 uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
3785 // This is only valid if the OpSizeInBits + c1 = size of inner shift.
3786 if (c1 + OpSizeInBits == InnerShiftSize) {
3787 if (c1 + c2 >= InnerShiftSize)
3788 return DAG.getConstant(0, VT);
3789 return DAG.getNode(ISD::TRUNCATE, N0->getDebugLoc(), VT,
3790 DAG.getNode(ISD::SRL, N0->getDebugLoc(), InnerShiftVT,
3791 N0.getOperand(0)->getOperand(0),
3792 DAG.getConstant(c1 + c2, ShiftCountVT)));
3796 // fold (srl (shl x, c), c) -> (and x, cst2)
3797 if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1 &&
3798 N0.getValueSizeInBits() <= 64) {
3799 uint64_t ShAmt = N1C->getZExtValue()+64-N0.getValueSizeInBits();
3800 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0.getOperand(0),
3801 DAG.getConstant(~0ULL >> ShAmt, VT));
3805 // fold (srl (anyextend x), c) -> (anyextend (srl x, c))
3806 if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
3807 // Shifting in all undef bits?
3808 EVT SmallVT = N0.getOperand(0).getValueType();
3809 if (N1C->getZExtValue() >= SmallVT.getSizeInBits())
3810 return DAG.getUNDEF(VT);
3812 if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
3813 uint64_t ShiftAmt = N1C->getZExtValue();
3814 SDValue SmallShift = DAG.getNode(ISD::SRL, N0.getDebugLoc(), SmallVT,
3816 DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
3817 AddToWorkList(SmallShift.getNode());
3818 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, SmallShift);
3822 // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
3823 // bit, which is unmodified by sra.
3824 if (N1C && N1C->getZExtValue() + 1 == VT.getSizeInBits()) {
3825 if (N0.getOpcode() == ISD::SRA)
3826 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0.getOperand(0), N1);
3829 // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
3830 if (N1C && N0.getOpcode() == ISD::CTLZ &&
3831 N1C->getAPIntValue() == Log2_32(VT.getSizeInBits())) {
3832 APInt KnownZero, KnownOne;
3833 DAG.ComputeMaskedBits(N0.getOperand(0), KnownZero, KnownOne);
3835 // If any of the input bits are KnownOne, then the input couldn't be all
3836 // zeros, thus the result of the srl will always be zero.
3837 if (KnownOne.getBoolValue()) return DAG.getConstant(0, VT);
3839 // If all of the bits input the to ctlz node are known to be zero, then
3840 // the result of the ctlz is "32" and the result of the shift is one.
3841 APInt UnknownBits = ~KnownZero;
3842 if (UnknownBits == 0) return DAG.getConstant(1, VT);
3844 // Otherwise, check to see if there is exactly one bit input to the ctlz.
3845 if ((UnknownBits & (UnknownBits - 1)) == 0) {
3846 // Okay, we know that only that the single bit specified by UnknownBits
3847 // could be set on input to the CTLZ node. If this bit is set, the SRL
3848 // will return 0, if it is clear, it returns 1. Change the CTLZ/SRL pair
3849 // to an SRL/XOR pair, which is likely to simplify more.
3850 unsigned ShAmt = UnknownBits.countTrailingZeros();
3851 SDValue Op = N0.getOperand(0);
3854 Op = DAG.getNode(ISD::SRL, N0.getDebugLoc(), VT, Op,
3855 DAG.getConstant(ShAmt, getShiftAmountTy(Op.getValueType())));
3856 AddToWorkList(Op.getNode());
3859 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
3860 Op, DAG.getConstant(1, VT));
3864 // fold (srl x, (trunc (and y, c))) -> (srl x, (and (trunc y), (trunc c))).
3865 if (N1.getOpcode() == ISD::TRUNCATE &&
3866 N1.getOperand(0).getOpcode() == ISD::AND &&
3867 N1.hasOneUse() && N1.getOperand(0).hasOneUse()) {
3868 SDValue N101 = N1.getOperand(0).getOperand(1);
3869 if (ConstantSDNode *N101C = dyn_cast<ConstantSDNode>(N101)) {
3870 EVT TruncVT = N1.getValueType();
3871 SDValue N100 = N1.getOperand(0).getOperand(0);
3872 APInt TruncC = N101C->getAPIntValue();
3873 TruncC = TruncC.trunc(TruncVT.getSizeInBits());
3874 return DAG.getNode(ISD::SRL, N->getDebugLoc(), VT, N0,
3875 DAG.getNode(ISD::AND, N->getDebugLoc(),
3877 DAG.getNode(ISD::TRUNCATE,
3880 DAG.getConstant(TruncC, TruncVT)));
3884 // fold operands of srl based on knowledge that the low bits are not
3886 if (N1C && SimplifyDemandedBits(SDValue(N, 0)))
3887 return SDValue(N, 0);
3890 SDValue NewSRL = visitShiftByConstant(N, N1C->getZExtValue());
3891 if (NewSRL.getNode())
3895 // Attempt to convert a srl of a load into a narrower zero-extending load.
3896 SDValue NarrowLoad = ReduceLoadWidth(N);
3897 if (NarrowLoad.getNode())
3900 // Here is a common situation. We want to optimize:
3903 // %b = and i32 %a, 2
3904 // %c = srl i32 %b, 1
3905 // brcond i32 %c ...
3911 // %c = setcc eq %b, 0
3914 // However when after the source operand of SRL is optimized into AND, the SRL
3915 // itself may not be optimized further. Look for it and add the BRCOND into
3917 if (N->hasOneUse()) {
3918 SDNode *Use = *N->use_begin();
3919 if (Use->getOpcode() == ISD::BRCOND)
3921 else if (Use->getOpcode() == ISD::TRUNCATE && Use->hasOneUse()) {
3922 // Also look pass the truncate.
3923 Use = *Use->use_begin();
3924 if (Use->getOpcode() == ISD::BRCOND)
3932 SDValue DAGCombiner::visitCTLZ(SDNode *N) {
3933 SDValue N0 = N->getOperand(0);
3934 EVT VT = N->getValueType(0);
3936 // fold (ctlz c1) -> c2
3937 if (isa<ConstantSDNode>(N0))
3938 return DAG.getNode(ISD::CTLZ, N->getDebugLoc(), VT, N0);
3942 SDValue DAGCombiner::visitCTLZ_ZERO_UNDEF(SDNode *N) {
3943 SDValue N0 = N->getOperand(0);
3944 EVT VT = N->getValueType(0);
3946 // fold (ctlz_zero_undef c1) -> c2
3947 if (isa<ConstantSDNode>(N0))
3948 return DAG.getNode(ISD::CTLZ_ZERO_UNDEF, N->getDebugLoc(), VT, N0);
3952 SDValue DAGCombiner::visitCTTZ(SDNode *N) {
3953 SDValue N0 = N->getOperand(0);
3954 EVT VT = N->getValueType(0);
3956 // fold (cttz c1) -> c2
3957 if (isa<ConstantSDNode>(N0))
3958 return DAG.getNode(ISD::CTTZ, N->getDebugLoc(), VT, N0);
3962 SDValue DAGCombiner::visitCTTZ_ZERO_UNDEF(SDNode *N) {
3963 SDValue N0 = N->getOperand(0);
3964 EVT VT = N->getValueType(0);
3966 // fold (cttz_zero_undef c1) -> c2
3967 if (isa<ConstantSDNode>(N0))
3968 return DAG.getNode(ISD::CTTZ_ZERO_UNDEF, N->getDebugLoc(), VT, N0);
3972 SDValue DAGCombiner::visitCTPOP(SDNode *N) {
3973 SDValue N0 = N->getOperand(0);
3974 EVT VT = N->getValueType(0);
3976 // fold (ctpop c1) -> c2
3977 if (isa<ConstantSDNode>(N0))
3978 return DAG.getNode(ISD::CTPOP, N->getDebugLoc(), VT, N0);
3982 SDValue DAGCombiner::visitSELECT(SDNode *N) {
3983 SDValue N0 = N->getOperand(0);
3984 SDValue N1 = N->getOperand(1);
3985 SDValue N2 = N->getOperand(2);
3986 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
3987 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
3988 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2);
3989 EVT VT = N->getValueType(0);
3990 EVT VT0 = N0.getValueType();
3992 // fold (select C, X, X) -> X
3995 // fold (select true, X, Y) -> X
3996 if (N0C && !N0C->isNullValue())
3998 // fold (select false, X, Y) -> Y
3999 if (N0C && N0C->isNullValue())
4001 // fold (select C, 1, X) -> (or C, X)
4002 if (VT == MVT::i1 && N1C && N1C->getAPIntValue() == 1)
4003 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N2);
4004 // fold (select C, 0, 1) -> (xor C, 1)
4005 if (VT.isInteger() &&
4008 TLI.getBooleanContents(false) == TargetLowering::ZeroOrOneBooleanContent)) &&
4009 N1C && N2C && N1C->isNullValue() && N2C->getAPIntValue() == 1) {
4012 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT0,
4013 N0, DAG.getConstant(1, VT0));
4014 XORNode = DAG.getNode(ISD::XOR, N0.getDebugLoc(), VT0,
4015 N0, DAG.getConstant(1, VT0));
4016 AddToWorkList(XORNode.getNode());
4018 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, XORNode);
4019 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, XORNode);
4021 // fold (select C, 0, X) -> (and (not C), X)
4022 if (VT == VT0 && VT == MVT::i1 && N1C && N1C->isNullValue()) {
4023 SDValue NOTNode = DAG.getNOT(N0.getDebugLoc(), N0, VT);
4024 AddToWorkList(NOTNode.getNode());
4025 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, NOTNode, N2);
4027 // fold (select C, X, 1) -> (or (not C), X)
4028 if (VT == VT0 && VT == MVT::i1 && N2C && N2C->getAPIntValue() == 1) {
4029 SDValue NOTNode = DAG.getNOT(N0.getDebugLoc(), N0, VT);
4030 AddToWorkList(NOTNode.getNode());
4031 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, NOTNode, N1);
4033 // fold (select C, X, 0) -> (and C, X)
4034 if (VT == MVT::i1 && N2C && N2C->isNullValue())
4035 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, N1);
4036 // fold (select X, X, Y) -> (or X, Y)
4037 // fold (select X, 1, Y) -> (or X, Y)
4038 if (VT == MVT::i1 && (N0 == N1 || (N1C && N1C->getAPIntValue() == 1)))
4039 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, N0, N2);
4040 // fold (select X, Y, X) -> (and X, Y)
4041 // fold (select X, Y, 0) -> (and X, Y)
4042 if (VT == MVT::i1 && (N0 == N2 || (N2C && N2C->getAPIntValue() == 0)))
4043 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, N0, N1);
4045 // If we can fold this based on the true/false value, do so.
4046 if (SimplifySelectOps(N, N1, N2))
4047 return SDValue(N, 0); // Don't revisit N.
4049 // fold selects based on a setcc into other things, such as min/max/abs
4050 if (N0.getOpcode() == ISD::SETCC) {
4052 // Check against MVT::Other for SELECT_CC, which is a workaround for targets
4053 // having to say they don't support SELECT_CC on every type the DAG knows
4054 // about, since there is no way to mark an opcode illegal at all value types
4055 if (TLI.isOperationLegalOrCustom(ISD::SELECT_CC, MVT::Other) &&
4056 TLI.isOperationLegalOrCustom(ISD::SELECT_CC, VT))
4057 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), VT,
4058 N0.getOperand(0), N0.getOperand(1),
4059 N1, N2, N0.getOperand(2));
4060 return SimplifySelect(N->getDebugLoc(), N0, N1, N2);
4066 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
4067 SDValue N0 = N->getOperand(0);
4068 SDValue N1 = N->getOperand(1);
4069 SDValue N2 = N->getOperand(2);
4070 SDValue N3 = N->getOperand(3);
4071 SDValue N4 = N->getOperand(4);
4072 ISD::CondCode CC = cast<CondCodeSDNode>(N4)->get();
4074 // fold select_cc lhs, rhs, x, x, cc -> x
4078 // Determine if the condition we're dealing with is constant
4079 SDValue SCC = SimplifySetCC(TLI.getSetCCResultType(N0.getValueType()),
4080 N0, N1, CC, N->getDebugLoc(), false);
4081 if (SCC.getNode()) AddToWorkList(SCC.getNode());
4083 if (ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode())) {
4084 if (!SCCC->isNullValue())
4085 return N2; // cond always true -> true val
4087 return N3; // cond always false -> false val
4090 // Fold to a simpler select_cc
4091 if (SCC.getNode() && SCC.getOpcode() == ISD::SETCC)
4092 return DAG.getNode(ISD::SELECT_CC, N->getDebugLoc(), N2.getValueType(),
4093 SCC.getOperand(0), SCC.getOperand(1), N2, N3,
4096 // If we can fold this based on the true/false value, do so.
4097 if (SimplifySelectOps(N, N2, N3))
4098 return SDValue(N, 0); // Don't revisit N.
4100 // fold select_cc into other things, such as min/max/abs
4101 return SimplifySelectCC(N->getDebugLoc(), N0, N1, N2, N3, CC);
4104 SDValue DAGCombiner::visitSETCC(SDNode *N) {
4105 return SimplifySetCC(N->getValueType(0), N->getOperand(0), N->getOperand(1),
4106 cast<CondCodeSDNode>(N->getOperand(2))->get(),
4110 // ExtendUsesToFormExtLoad - Trying to extend uses of a load to enable this:
4111 // "fold ({s|z|a}ext (load x)) -> ({s|z|a}ext (truncate ({s|z|a}extload x)))"
4112 // transformation. Returns true if extension are possible and the above
4113 // mentioned transformation is profitable.
4114 static bool ExtendUsesToFormExtLoad(SDNode *N, SDValue N0,
4116 SmallVector<SDNode*, 4> &ExtendNodes,
4117 const TargetLowering &TLI) {
4118 bool HasCopyToRegUses = false;
4119 bool isTruncFree = TLI.isTruncateFree(N->getValueType(0), N0.getValueType());
4120 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
4121 UE = N0.getNode()->use_end();
4126 if (UI.getUse().getResNo() != N0.getResNo())
4128 // FIXME: Only extend SETCC N, N and SETCC N, c for now.
4129 if (ExtOpc != ISD::ANY_EXTEND && User->getOpcode() == ISD::SETCC) {
4130 ISD::CondCode CC = cast<CondCodeSDNode>(User->getOperand(2))->get();
4131 if (ExtOpc == ISD::ZERO_EXTEND && ISD::isSignedIntSetCC(CC))
4132 // Sign bits will be lost after a zext.
4135 for (unsigned i = 0; i != 2; ++i) {
4136 SDValue UseOp = User->getOperand(i);
4139 if (!isa<ConstantSDNode>(UseOp))
4144 ExtendNodes.push_back(User);
4147 // If truncates aren't free and there are users we can't
4148 // extend, it isn't worthwhile.
4151 // Remember if this value is live-out.
4152 if (User->getOpcode() == ISD::CopyToReg)
4153 HasCopyToRegUses = true;
4156 if (HasCopyToRegUses) {
4157 bool BothLiveOut = false;
4158 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
4160 SDUse &Use = UI.getUse();
4161 if (Use.getResNo() == 0 && Use.getUser()->getOpcode() == ISD::CopyToReg) {
4167 // Both unextended and extended values are live out. There had better be
4168 // a good reason for the transformation.
4169 return ExtendNodes.size();
4174 void DAGCombiner::ExtendSetCCUses(SmallVector<SDNode*, 4> SetCCs,
4175 SDValue Trunc, SDValue ExtLoad, DebugLoc DL,
4176 ISD::NodeType ExtType) {
4177 // Extend SetCC uses if necessary.
4178 for (unsigned i = 0, e = SetCCs.size(); i != e; ++i) {
4179 SDNode *SetCC = SetCCs[i];
4180 SmallVector<SDValue, 4> Ops;
4182 for (unsigned j = 0; j != 2; ++j) {
4183 SDValue SOp = SetCC->getOperand(j);
4185 Ops.push_back(ExtLoad);
4187 Ops.push_back(DAG.getNode(ExtType, DL, ExtLoad->getValueType(0), SOp));
4190 Ops.push_back(SetCC->getOperand(2));
4191 CombineTo(SetCC, DAG.getNode(ISD::SETCC, DL, SetCC->getValueType(0),
4192 &Ops[0], Ops.size()));
4196 SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
4197 SDValue N0 = N->getOperand(0);
4198 EVT VT = N->getValueType(0);
4200 // fold (sext c1) -> c1
4201 if (isa<ConstantSDNode>(N0))
4202 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N0);
4204 // fold (sext (sext x)) -> (sext x)
4205 // fold (sext (aext x)) -> (sext x)
4206 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4207 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT,
4210 if (N0.getOpcode() == ISD::TRUNCATE) {
4211 // fold (sext (truncate (load x))) -> (sext (smaller load x))
4212 // fold (sext (truncate (srl (load x), c))) -> (sext (smaller load (x+c/n)))
4213 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4214 if (NarrowLoad.getNode()) {
4215 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4216 if (NarrowLoad.getNode() != N0.getNode()) {
4217 CombineTo(N0.getNode(), NarrowLoad);
4218 // CombineTo deleted the truncate, if needed, but not what's under it.
4221 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4224 // See if the value being truncated is already sign extended. If so, just
4225 // eliminate the trunc/sext pair.
4226 SDValue Op = N0.getOperand(0);
4227 unsigned OpBits = Op.getValueType().getScalarType().getSizeInBits();
4228 unsigned MidBits = N0.getValueType().getScalarType().getSizeInBits();
4229 unsigned DestBits = VT.getScalarType().getSizeInBits();
4230 unsigned NumSignBits = DAG.ComputeNumSignBits(Op);
4232 if (OpBits == DestBits) {
4233 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
4234 // bits, it is already ready.
4235 if (NumSignBits > DestBits-MidBits)
4237 } else if (OpBits < DestBits) {
4238 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
4239 // bits, just sext from i32.
4240 if (NumSignBits > OpBits-MidBits)
4241 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, Op);
4243 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
4244 // bits, just truncate to i32.
4245 if (NumSignBits > OpBits-MidBits)
4246 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4249 // fold (sext (truncate x)) -> (sextinreg x).
4250 if (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND_INREG,
4251 N0.getValueType())) {
4252 if (OpBits < DestBits)
4253 Op = DAG.getNode(ISD::ANY_EXTEND, N0.getDebugLoc(), VT, Op);
4254 else if (OpBits > DestBits)
4255 Op = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), VT, Op);
4256 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT, Op,
4257 DAG.getValueType(N0.getValueType()));
4261 // fold (sext (load x)) -> (sext (truncate (sextload x)))
4262 // None of the supported targets knows how to perform load and sign extend
4263 // on vectors in one instruction. We only perform this transformation on
4265 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4266 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4267 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()))) {
4268 bool DoXform = true;
4269 SmallVector<SDNode*, 4> SetCCs;
4270 if (!N0.hasOneUse())
4271 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::SIGN_EXTEND, SetCCs, TLI);
4273 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4274 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
4276 LN0->getBasePtr(), LN0->getPointerInfo(),
4278 LN0->isVolatile(), LN0->isNonTemporal(),
4279 LN0->getAlignment());
4280 CombineTo(N, ExtLoad);
4281 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4282 N0.getValueType(), ExtLoad);
4283 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4284 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4286 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4290 // fold (sext (sextload x)) -> (sext (truncate (sextload x)))
4291 // fold (sext ( extload x)) -> (sext (truncate (sextload x)))
4292 if ((ISD::isSEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
4293 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
4294 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4295 EVT MemVT = LN0->getMemoryVT();
4296 if ((!LegalOperations && !LN0->isVolatile()) ||
4297 TLI.isLoadExtLegal(ISD::SEXTLOAD, MemVT)) {
4298 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
4300 LN0->getBasePtr(), LN0->getPointerInfo(),
4302 LN0->isVolatile(), LN0->isNonTemporal(),
4303 LN0->getAlignment());
4304 CombineTo(N, ExtLoad);
4305 CombineTo(N0.getNode(),
4306 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4307 N0.getValueType(), ExtLoad),
4308 ExtLoad.getValue(1));
4309 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4313 // fold (sext (and/or/xor (load x), cst)) ->
4314 // (and/or/xor (sextload x), (sext cst))
4315 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
4316 N0.getOpcode() == ISD::XOR) &&
4317 isa<LoadSDNode>(N0.getOperand(0)) &&
4318 N0.getOperand(1).getOpcode() == ISD::Constant &&
4319 TLI.isLoadExtLegal(ISD::SEXTLOAD, N0.getValueType()) &&
4320 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
4321 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
4322 if (LN0->getExtensionType() != ISD::ZEXTLOAD) {
4323 bool DoXform = true;
4324 SmallVector<SDNode*, 4> SetCCs;
4325 if (!N0.hasOneUse())
4326 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::SIGN_EXTEND,
4329 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, LN0->getDebugLoc(), VT,
4330 LN0->getChain(), LN0->getBasePtr(),
4331 LN0->getPointerInfo(),
4334 LN0->isNonTemporal(),
4335 LN0->getAlignment());
4336 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4337 Mask = Mask.sext(VT.getSizeInBits());
4338 SDValue And = DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
4339 ExtLoad, DAG.getConstant(Mask, VT));
4340 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
4341 N0.getOperand(0).getDebugLoc(),
4342 N0.getOperand(0).getValueType(), ExtLoad);
4344 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
4345 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4347 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4352 if (N0.getOpcode() == ISD::SETCC) {
4353 // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
4354 // Only do this before legalize for now.
4355 if (VT.isVector() && !LegalOperations) {
4356 EVT N0VT = N0.getOperand(0).getValueType();
4357 // On some architectures (such as SSE/NEON/etc) the SETCC result type is
4358 // of the same size as the compared operands. Only optimize sext(setcc())
4359 // if this is the case.
4360 EVT SVT = TLI.getSetCCResultType(N0VT);
4362 // We know that the # elements of the results is the same as the
4363 // # elements of the compare (and the # elements of the compare result
4364 // for that matter). Check to see that they are the same size. If so,
4365 // we know that the element size of the sext'd result matches the
4366 // element size of the compare operands.
4367 if (VT.getSizeInBits() == SVT.getSizeInBits())
4368 return DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4370 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4371 // If the desired elements are smaller or larger than the source
4372 // elements we can use a matching integer vector type and then
4373 // truncate/sign extend
4375 EVT MatchingElementType =
4376 EVT::getIntegerVT(*DAG.getContext(),
4377 N0VT.getScalarType().getSizeInBits());
4378 EVT MatchingVectorType =
4379 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4380 N0VT.getVectorNumElements());
4382 if (SVT == MatchingVectorType) {
4383 SDValue VsetCC = DAG.getSetCC(N->getDebugLoc(), MatchingVectorType,
4384 N0.getOperand(0), N0.getOperand(1),
4385 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4386 return DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT);
4391 // sext(setcc x, y, cc) -> (select_cc x, y, -1, 0, cc)
4392 unsigned ElementWidth = VT.getScalarType().getSizeInBits();
4394 DAG.getConstant(APInt::getAllOnesValue(ElementWidth), VT);
4396 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4397 NegOne, DAG.getConstant(0, VT),
4398 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4399 if (SCC.getNode()) return SCC;
4400 if (!LegalOperations ||
4401 TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(VT)))
4402 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT,
4403 DAG.getSetCC(N->getDebugLoc(),
4404 TLI.getSetCCResultType(VT),
4405 N0.getOperand(0), N0.getOperand(1),
4406 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
4407 NegOne, DAG.getConstant(0, VT));
4410 // fold (sext x) -> (zext x) if the sign bit is known zero.
4411 if ((!LegalOperations || TLI.isOperationLegal(ISD::ZERO_EXTEND, VT)) &&
4412 DAG.SignBitIsZero(N0))
4413 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
4418 // isTruncateOf - If N is a truncate of some other value, return true, record
4419 // the value being truncated in Op and which of Op's bits are zero in KnownZero.
4420 // This function computes KnownZero to avoid a duplicated call to
4421 // ComputeMaskedBits in the caller.
4422 static bool isTruncateOf(SelectionDAG &DAG, SDValue N, SDValue &Op,
4425 if (N->getOpcode() == ISD::TRUNCATE) {
4426 Op = N->getOperand(0);
4427 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
4431 if (N->getOpcode() != ISD::SETCC || N->getValueType(0) != MVT::i1 ||
4432 cast<CondCodeSDNode>(N->getOperand(2))->get() != ISD::SETNE)
4435 SDValue Op0 = N->getOperand(0);
4436 SDValue Op1 = N->getOperand(1);
4437 assert(Op0.getValueType() == Op1.getValueType());
4439 ConstantSDNode *COp0 = dyn_cast<ConstantSDNode>(Op0);
4440 ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
4441 if (COp0 && COp0->isNullValue())
4443 else if (COp1 && COp1->isNullValue())
4448 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
4450 if (!(KnownZero | APInt(Op.getValueSizeInBits(), 1)).isAllOnesValue())
4456 SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) {
4457 SDValue N0 = N->getOperand(0);
4458 EVT VT = N->getValueType(0);
4460 // fold (zext c1) -> c1
4461 if (isa<ConstantSDNode>(N0))
4462 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, N0);
4463 // fold (zext (zext x)) -> (zext x)
4464 // fold (zext (aext x)) -> (zext x)
4465 if (N0.getOpcode() == ISD::ZERO_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND)
4466 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT,
4469 // fold (zext (truncate x)) -> (zext x) or
4470 // (zext (truncate x)) -> (truncate x)
4471 // This is valid when the truncated bits of x are already zero.
4472 // FIXME: We should extend this to work for vectors too.
4475 if (!VT.isVector() && isTruncateOf(DAG, N0, Op, KnownZero)) {
4476 APInt TruncatedBits =
4477 (Op.getValueSizeInBits() == N0.getValueSizeInBits()) ?
4478 APInt(Op.getValueSizeInBits(), 0) :
4479 APInt::getBitsSet(Op.getValueSizeInBits(),
4480 N0.getValueSizeInBits(),
4481 std::min(Op.getValueSizeInBits(),
4482 VT.getSizeInBits()));
4483 if (TruncatedBits == (KnownZero & TruncatedBits)) {
4484 if (VT.bitsGT(Op.getValueType()))
4485 return DAG.getNode(ISD::ZERO_EXTEND, N->getDebugLoc(), VT, Op);
4486 if (VT.bitsLT(Op.getValueType()))
4487 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4493 // fold (zext (truncate (load x))) -> (zext (smaller load x))
4494 // fold (zext (truncate (srl (load x), c))) -> (zext (small load (x+c/n)))
4495 if (N0.getOpcode() == ISD::TRUNCATE) {
4496 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4497 if (NarrowLoad.getNode()) {
4498 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4499 if (NarrowLoad.getNode() != N0.getNode()) {
4500 CombineTo(N0.getNode(), NarrowLoad);
4501 // CombineTo deleted the truncate, if needed, but not what's under it.
4504 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4508 // fold (zext (truncate x)) -> (and x, mask)
4509 if (N0.getOpcode() == ISD::TRUNCATE &&
4510 (!LegalOperations || TLI.isOperationLegal(ISD::AND, VT))) {
4512 // fold (zext (truncate (load x))) -> (zext (smaller load x))
4513 // fold (zext (truncate (srl (load x), c))) -> (zext (smaller load (x+c/n)))
4514 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4515 if (NarrowLoad.getNode()) {
4516 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4517 if (NarrowLoad.getNode() != N0.getNode()) {
4518 CombineTo(N0.getNode(), NarrowLoad);
4519 // CombineTo deleted the truncate, if needed, but not what's under it.
4522 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4525 SDValue Op = N0.getOperand(0);
4526 if (Op.getValueType().bitsLT(VT)) {
4527 Op = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, Op);
4528 } else if (Op.getValueType().bitsGT(VT)) {
4529 Op = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Op);
4531 return DAG.getZeroExtendInReg(Op, N->getDebugLoc(),
4532 N0.getValueType().getScalarType());
4535 // Fold (zext (and (trunc x), cst)) -> (and x, cst),
4536 // if either of the casts is not free.
4537 if (N0.getOpcode() == ISD::AND &&
4538 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
4539 N0.getOperand(1).getOpcode() == ISD::Constant &&
4540 (!TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
4541 N0.getValueType()) ||
4542 !TLI.isZExtFree(N0.getValueType(), VT))) {
4543 SDValue X = N0.getOperand(0).getOperand(0);
4544 if (X.getValueType().bitsLT(VT)) {
4545 X = DAG.getNode(ISD::ANY_EXTEND, X.getDebugLoc(), VT, X);
4546 } else if (X.getValueType().bitsGT(VT)) {
4547 X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
4549 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4550 Mask = Mask.zext(VT.getSizeInBits());
4551 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4552 X, DAG.getConstant(Mask, VT));
4555 // fold (zext (load x)) -> (zext (truncate (zextload x)))
4556 // None of the supported targets knows how to perform load and vector_zext
4557 // on vectors in one instruction. We only perform this transformation on
4559 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4560 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4561 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()))) {
4562 bool DoXform = true;
4563 SmallVector<SDNode*, 4> SetCCs;
4564 if (!N0.hasOneUse())
4565 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ZERO_EXTEND, SetCCs, TLI);
4567 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4568 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N->getDebugLoc(), VT,
4570 LN0->getBasePtr(), LN0->getPointerInfo(),
4572 LN0->isVolatile(), LN0->isNonTemporal(),
4573 LN0->getAlignment());
4574 CombineTo(N, ExtLoad);
4575 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4576 N0.getValueType(), ExtLoad);
4577 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4579 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4581 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4585 // fold (zext (and/or/xor (load x), cst)) ->
4586 // (and/or/xor (zextload x), (zext cst))
4587 if ((N0.getOpcode() == ISD::AND || N0.getOpcode() == ISD::OR ||
4588 N0.getOpcode() == ISD::XOR) &&
4589 isa<LoadSDNode>(N0.getOperand(0)) &&
4590 N0.getOperand(1).getOpcode() == ISD::Constant &&
4591 TLI.isLoadExtLegal(ISD::ZEXTLOAD, N0.getValueType()) &&
4592 (!LegalOperations && TLI.isOperationLegal(N0.getOpcode(), VT))) {
4593 LoadSDNode *LN0 = cast<LoadSDNode>(N0.getOperand(0));
4594 if (LN0->getExtensionType() != ISD::SEXTLOAD) {
4595 bool DoXform = true;
4596 SmallVector<SDNode*, 4> SetCCs;
4597 if (!N0.hasOneUse())
4598 DoXform = ExtendUsesToFormExtLoad(N, N0.getOperand(0), ISD::ZERO_EXTEND,
4601 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), VT,
4602 LN0->getChain(), LN0->getBasePtr(),
4603 LN0->getPointerInfo(),
4606 LN0->isNonTemporal(),
4607 LN0->getAlignment());
4608 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4609 Mask = Mask.zext(VT.getSizeInBits());
4610 SDValue And = DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
4611 ExtLoad, DAG.getConstant(Mask, VT));
4612 SDValue Trunc = DAG.getNode(ISD::TRUNCATE,
4613 N0.getOperand(0).getDebugLoc(),
4614 N0.getOperand(0).getValueType(), ExtLoad);
4616 CombineTo(N0.getOperand(0).getNode(), Trunc, ExtLoad.getValue(1));
4617 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4619 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4624 // fold (zext (zextload x)) -> (zext (truncate (zextload x)))
4625 // fold (zext ( extload x)) -> (zext (truncate (zextload x)))
4626 if ((ISD::isZEXTLoad(N0.getNode()) || ISD::isEXTLoad(N0.getNode())) &&
4627 ISD::isUNINDEXEDLoad(N0.getNode()) && N0.hasOneUse()) {
4628 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4629 EVT MemVT = LN0->getMemoryVT();
4630 if ((!LegalOperations && !LN0->isVolatile()) ||
4631 TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT)) {
4632 SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N->getDebugLoc(), VT,
4634 LN0->getBasePtr(), LN0->getPointerInfo(),
4636 LN0->isVolatile(), LN0->isNonTemporal(),
4637 LN0->getAlignment());
4638 CombineTo(N, ExtLoad);
4639 CombineTo(N0.getNode(),
4640 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), N0.getValueType(),
4642 ExtLoad.getValue(1));
4643 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4647 if (N0.getOpcode() == ISD::SETCC) {
4648 if (!LegalOperations && VT.isVector()) {
4649 // zext(setcc) -> (and (vsetcc), (1, 1, ...) for vectors.
4650 // Only do this before legalize for now.
4651 EVT N0VT = N0.getOperand(0).getValueType();
4652 EVT EltVT = VT.getVectorElementType();
4653 SmallVector<SDValue,8> OneOps(VT.getVectorNumElements(),
4654 DAG.getConstant(1, EltVT));
4655 if (VT.getSizeInBits() == N0VT.getSizeInBits())
4656 // We know that the # elements of the results is the same as the
4657 // # elements of the compare (and the # elements of the compare result
4658 // for that matter). Check to see that they are the same size. If so,
4659 // we know that the element size of the sext'd result matches the
4660 // element size of the compare operands.
4661 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4662 DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4664 cast<CondCodeSDNode>(N0.getOperand(2))->get()),
4665 DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
4666 &OneOps[0], OneOps.size()));
4668 // If the desired elements are smaller or larger than the source
4669 // elements we can use a matching integer vector type and then
4670 // truncate/sign extend
4671 EVT MatchingElementType =
4672 EVT::getIntegerVT(*DAG.getContext(),
4673 N0VT.getScalarType().getSizeInBits());
4674 EVT MatchingVectorType =
4675 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4676 N0VT.getVectorNumElements());
4678 DAG.getSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
4680 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4681 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4682 DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT),
4683 DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), VT,
4684 &OneOps[0], OneOps.size()));
4687 // zext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
4689 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4690 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
4691 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4692 if (SCC.getNode()) return SCC;
4695 // (zext (shl (zext x), cst)) -> (shl (zext x), cst)
4696 if ((N0.getOpcode() == ISD::SHL || N0.getOpcode() == ISD::SRL) &&
4697 isa<ConstantSDNode>(N0.getOperand(1)) &&
4698 N0.getOperand(0).getOpcode() == ISD::ZERO_EXTEND &&
4700 SDValue ShAmt = N0.getOperand(1);
4701 unsigned ShAmtVal = cast<ConstantSDNode>(ShAmt)->getZExtValue();
4702 if (N0.getOpcode() == ISD::SHL) {
4703 SDValue InnerZExt = N0.getOperand(0);
4704 // If the original shl may be shifting out bits, do not perform this
4706 unsigned KnownZeroBits = InnerZExt.getValueType().getSizeInBits() -
4707 InnerZExt.getOperand(0).getValueType().getSizeInBits();
4708 if (ShAmtVal > KnownZeroBits)
4712 DebugLoc DL = N->getDebugLoc();
4714 // Ensure that the shift amount is wide enough for the shifted value.
4715 if (VT.getSizeInBits() >= 256)
4716 ShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, ShAmt);
4718 return DAG.getNode(N0.getOpcode(), DL, VT,
4719 DAG.getNode(ISD::ZERO_EXTEND, DL, VT, N0.getOperand(0)),
4726 SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) {
4727 SDValue N0 = N->getOperand(0);
4728 EVT VT = N->getValueType(0);
4730 // fold (aext c1) -> c1
4731 if (isa<ConstantSDNode>(N0))
4732 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, N0);
4733 // fold (aext (aext x)) -> (aext x)
4734 // fold (aext (zext x)) -> (zext x)
4735 // fold (aext (sext x)) -> (sext x)
4736 if (N0.getOpcode() == ISD::ANY_EXTEND ||
4737 N0.getOpcode() == ISD::ZERO_EXTEND ||
4738 N0.getOpcode() == ISD::SIGN_EXTEND)
4739 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT, N0.getOperand(0));
4741 // fold (aext (truncate (load x))) -> (aext (smaller load x))
4742 // fold (aext (truncate (srl (load x), c))) -> (aext (small load (x+c/n)))
4743 if (N0.getOpcode() == ISD::TRUNCATE) {
4744 SDValue NarrowLoad = ReduceLoadWidth(N0.getNode());
4745 if (NarrowLoad.getNode()) {
4746 SDNode* oye = N0.getNode()->getOperand(0).getNode();
4747 if (NarrowLoad.getNode() != N0.getNode()) {
4748 CombineTo(N0.getNode(), NarrowLoad);
4749 // CombineTo deleted the truncate, if needed, but not what's under it.
4752 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4756 // fold (aext (truncate x))
4757 if (N0.getOpcode() == ISD::TRUNCATE) {
4758 SDValue TruncOp = N0.getOperand(0);
4759 if (TruncOp.getValueType() == VT)
4760 return TruncOp; // x iff x size == zext size.
4761 if (TruncOp.getValueType().bitsGT(VT))
4762 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, TruncOp);
4763 return DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, TruncOp);
4766 // Fold (aext (and (trunc x), cst)) -> (and x, cst)
4767 // if the trunc is not free.
4768 if (N0.getOpcode() == ISD::AND &&
4769 N0.getOperand(0).getOpcode() == ISD::TRUNCATE &&
4770 N0.getOperand(1).getOpcode() == ISD::Constant &&
4771 !TLI.isTruncateFree(N0.getOperand(0).getOperand(0).getValueType(),
4772 N0.getValueType())) {
4773 SDValue X = N0.getOperand(0).getOperand(0);
4774 if (X.getValueType().bitsLT(VT)) {
4775 X = DAG.getNode(ISD::ANY_EXTEND, N->getDebugLoc(), VT, X);
4776 } else if (X.getValueType().bitsGT(VT)) {
4777 X = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, X);
4779 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
4780 Mask = Mask.zext(VT.getSizeInBits());
4781 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
4782 X, DAG.getConstant(Mask, VT));
4785 // fold (aext (load x)) -> (aext (truncate (extload x)))
4786 // None of the supported targets knows how to perform load and any_ext
4787 // on vectors in one instruction. We only perform this transformation on
4789 if (ISD::isNON_EXTLoad(N0.getNode()) && !VT.isVector() &&
4790 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
4791 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
4792 bool DoXform = true;
4793 SmallVector<SDNode*, 4> SetCCs;
4794 if (!N0.hasOneUse())
4795 DoXform = ExtendUsesToFormExtLoad(N, N0, ISD::ANY_EXTEND, SetCCs, TLI);
4797 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4798 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
4800 LN0->getBasePtr(), LN0->getPointerInfo(),
4802 LN0->isVolatile(), LN0->isNonTemporal(),
4803 LN0->getAlignment());
4804 CombineTo(N, ExtLoad);
4805 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4806 N0.getValueType(), ExtLoad);
4807 CombineTo(N0.getNode(), Trunc, ExtLoad.getValue(1));
4808 ExtendSetCCUses(SetCCs, Trunc, ExtLoad, N->getDebugLoc(),
4810 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4814 // fold (aext (zextload x)) -> (aext (truncate (zextload x)))
4815 // fold (aext (sextload x)) -> (aext (truncate (sextload x)))
4816 // fold (aext ( extload x)) -> (aext (truncate (extload x)))
4817 if (N0.getOpcode() == ISD::LOAD &&
4818 !ISD::isNON_EXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
4820 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
4821 EVT MemVT = LN0->getMemoryVT();
4822 SDValue ExtLoad = DAG.getExtLoad(LN0->getExtensionType(), N->getDebugLoc(),
4823 VT, LN0->getChain(), LN0->getBasePtr(),
4824 LN0->getPointerInfo(), MemVT,
4825 LN0->isVolatile(), LN0->isNonTemporal(),
4826 LN0->getAlignment());
4827 CombineTo(N, ExtLoad);
4828 CombineTo(N0.getNode(),
4829 DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(),
4830 N0.getValueType(), ExtLoad),
4831 ExtLoad.getValue(1));
4832 return SDValue(N, 0); // Return N so it doesn't get rechecked!
4835 if (N0.getOpcode() == ISD::SETCC) {
4836 // aext(setcc) -> sext_in_reg(vsetcc) for vectors.
4837 // Only do this before legalize for now.
4838 if (VT.isVector() && !LegalOperations) {
4839 EVT N0VT = N0.getOperand(0).getValueType();
4840 // We know that the # elements of the results is the same as the
4841 // # elements of the compare (and the # elements of the compare result
4842 // for that matter). Check to see that they are the same size. If so,
4843 // we know that the element size of the sext'd result matches the
4844 // element size of the compare operands.
4845 if (VT.getSizeInBits() == N0VT.getSizeInBits())
4846 return DAG.getSetCC(N->getDebugLoc(), VT, N0.getOperand(0),
4848 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4849 // If the desired elements are smaller or larger than the source
4850 // elements we can use a matching integer vector type and then
4851 // truncate/sign extend
4853 EVT MatchingElementType =
4854 EVT::getIntegerVT(*DAG.getContext(),
4855 N0VT.getScalarType().getSizeInBits());
4856 EVT MatchingVectorType =
4857 EVT::getVectorVT(*DAG.getContext(), MatchingElementType,
4858 N0VT.getVectorNumElements());
4860 DAG.getSetCC(N->getDebugLoc(), MatchingVectorType, N0.getOperand(0),
4862 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4863 return DAG.getSExtOrTrunc(VsetCC, N->getDebugLoc(), VT);
4867 // aext(setcc x,y,cc) -> select_cc x, y, 1, 0, cc
4869 SimplifySelectCC(N->getDebugLoc(), N0.getOperand(0), N0.getOperand(1),
4870 DAG.getConstant(1, VT), DAG.getConstant(0, VT),
4871 cast<CondCodeSDNode>(N0.getOperand(2))->get(), true);
4879 /// GetDemandedBits - See if the specified operand can be simplified with the
4880 /// knowledge that only the bits specified by Mask are used. If so, return the
4881 /// simpler operand, otherwise return a null SDValue.
4882 SDValue DAGCombiner::GetDemandedBits(SDValue V, const APInt &Mask) {
4883 switch (V.getOpcode()) {
4885 case ISD::Constant: {
4886 const ConstantSDNode *CV = cast<ConstantSDNode>(V.getNode());
4887 assert(CV != 0 && "Const value should be ConstSDNode.");
4888 const APInt &CVal = CV->getAPIntValue();
4889 APInt NewVal = CVal & Mask;
4890 if (NewVal != CVal) {
4891 return DAG.getConstant(NewVal, V.getValueType());
4897 // If the LHS or RHS don't contribute bits to the or, drop them.
4898 if (DAG.MaskedValueIsZero(V.getOperand(0), Mask))
4899 return V.getOperand(1);
4900 if (DAG.MaskedValueIsZero(V.getOperand(1), Mask))
4901 return V.getOperand(0);
4904 // Only look at single-use SRLs.
4905 if (!V.getNode()->hasOneUse())
4907 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
4908 // See if we can recursively simplify the LHS.
4909 unsigned Amt = RHSC->getZExtValue();
4911 // Watch out for shift count overflow though.
4912 if (Amt >= Mask.getBitWidth()) break;
4913 APInt NewMask = Mask << Amt;
4914 SDValue SimplifyLHS = GetDemandedBits(V.getOperand(0), NewMask);
4915 if (SimplifyLHS.getNode())
4916 return DAG.getNode(ISD::SRL, V.getDebugLoc(), V.getValueType(),
4917 SimplifyLHS, V.getOperand(1));
4923 /// ReduceLoadWidth - If the result of a wider load is shifted to right of N
4924 /// bits and then truncated to a narrower type and where N is a multiple
4925 /// of number of bits of the narrower type, transform it to a narrower load
4926 /// from address + N / num of bits of new type. If the result is to be
4927 /// extended, also fold the extension to form a extending load.
4928 SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) {
4929 unsigned Opc = N->getOpcode();
4931 ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
4932 SDValue N0 = N->getOperand(0);
4933 EVT VT = N->getValueType(0);
4936 // This transformation isn't valid for vector loads.
4940 // Special case: SIGN_EXTEND_INREG is basically truncating to ExtVT then
4942 if (Opc == ISD::SIGN_EXTEND_INREG) {
4943 ExtType = ISD::SEXTLOAD;
4944 ExtVT = cast<VTSDNode>(N->getOperand(1))->getVT();
4945 } else if (Opc == ISD::SRL) {
4946 // Another special-case: SRL is basically zero-extending a narrower value.
4947 ExtType = ISD::ZEXTLOAD;
4949 ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1));
4950 if (!N01) return SDValue();
4951 ExtVT = EVT::getIntegerVT(*DAG.getContext(),
4952 VT.getSizeInBits() - N01->getZExtValue());
4954 if (LegalOperations && !TLI.isLoadExtLegal(ExtType, ExtVT))
4957 unsigned EVTBits = ExtVT.getSizeInBits();
4959 // Do not generate loads of non-round integer types since these can
4960 // be expensive (and would be wrong if the type is not byte sized).
4961 if (!ExtVT.isRound())
4965 if (N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
4966 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
4967 ShAmt = N01->getZExtValue();
4968 // Is the shift amount a multiple of size of VT?
4969 if ((ShAmt & (EVTBits-1)) == 0) {
4970 N0 = N0.getOperand(0);
4971 // Is the load width a multiple of size of VT?
4972 if ((N0.getValueType().getSizeInBits() & (EVTBits-1)) != 0)
4976 // At this point, we must have a load or else we can't do the transform.
4977 if (!isa<LoadSDNode>(N0)) return SDValue();
4979 // If the shift amount is larger than the input type then we're not
4980 // accessing any of the loaded bytes. If the load was a zextload/extload
4981 // then the result of the shift+trunc is zero/undef (handled elsewhere).
4982 // If the load was a sextload then the result is a splat of the sign bit
4983 // of the extended byte. This is not worth optimizing for.
4984 if (ShAmt >= cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits())
4989 // If the load is shifted left (and the result isn't shifted back right),
4990 // we can fold the truncate through the shift.
4991 unsigned ShLeftAmt = 0;
4992 if (ShAmt == 0 && N0.getOpcode() == ISD::SHL && N0.hasOneUse() &&
4993 ExtVT == VT && TLI.isNarrowingProfitable(N0.getValueType(), VT)) {
4994 if (ConstantSDNode *N01 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
4995 ShLeftAmt = N01->getZExtValue();
4996 N0 = N0.getOperand(0);
5000 // If we haven't found a load, we can't narrow it. Don't transform one with
5001 // multiple uses, this would require adding a new load.
5002 if (!isa<LoadSDNode>(N0) || !N0.hasOneUse() ||
5003 // Don't change the width of a volatile load.
5004 cast<LoadSDNode>(N0)->isVolatile())
5007 // Verify that we are actually reducing a load width here.
5008 if (cast<LoadSDNode>(N0)->getMemoryVT().getSizeInBits() < EVTBits)
5011 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5012 EVT PtrType = N0.getOperand(1).getValueType();
5014 // For big endian targets, we need to adjust the offset to the pointer to
5015 // load the correct bytes.
5016 if (TLI.isBigEndian()) {
5017 unsigned LVTStoreBits = LN0->getMemoryVT().getStoreSizeInBits();
5018 unsigned EVTStoreBits = ExtVT.getStoreSizeInBits();
5019 ShAmt = LVTStoreBits - EVTStoreBits - ShAmt;
5022 uint64_t PtrOff = ShAmt / 8;
5023 unsigned NewAlign = MinAlign(LN0->getAlignment(), PtrOff);
5024 SDValue NewPtr = DAG.getNode(ISD::ADD, LN0->getDebugLoc(),
5025 PtrType, LN0->getBasePtr(),
5026 DAG.getConstant(PtrOff, PtrType));
5027 AddToWorkList(NewPtr.getNode());
5030 if (ExtType == ISD::NON_EXTLOAD)
5031 Load = DAG.getLoad(VT, N0.getDebugLoc(), LN0->getChain(), NewPtr,
5032 LN0->getPointerInfo().getWithOffset(PtrOff),
5033 LN0->isVolatile(), LN0->isNonTemporal(),
5034 LN0->isInvariant(), NewAlign);
5036 Load = DAG.getExtLoad(ExtType, N0.getDebugLoc(), VT, LN0->getChain(),NewPtr,
5037 LN0->getPointerInfo().getWithOffset(PtrOff),
5038 ExtVT, LN0->isVolatile(), LN0->isNonTemporal(),
5041 // Replace the old load's chain with the new load's chain.
5042 WorkListRemover DeadNodes(*this);
5043 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Load.getValue(1),
5046 // Shift the result left, if we've swallowed a left shift.
5047 SDValue Result = Load;
5048 if (ShLeftAmt != 0) {
5049 EVT ShImmTy = getShiftAmountTy(Result.getValueType());
5050 if (!isUIntN(ShImmTy.getSizeInBits(), ShLeftAmt))
5052 Result = DAG.getNode(ISD::SHL, N0.getDebugLoc(), VT,
5053 Result, DAG.getConstant(ShLeftAmt, ShImmTy));
5056 // Return the new loaded value.
5060 SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
5061 SDValue N0 = N->getOperand(0);
5062 SDValue N1 = N->getOperand(1);
5063 EVT VT = N->getValueType(0);
5064 EVT EVT = cast<VTSDNode>(N1)->getVT();
5065 unsigned VTBits = VT.getScalarType().getSizeInBits();
5066 unsigned EVTBits = EVT.getScalarType().getSizeInBits();
5068 // fold (sext_in_reg c1) -> c1
5069 if (isa<ConstantSDNode>(N0) || N0.getOpcode() == ISD::UNDEF)
5070 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT, N0, N1);
5072 // If the input is already sign extended, just drop the extension.
5073 if (DAG.ComputeNumSignBits(N0) >= VTBits-EVTBits+1)
5076 // fold (sext_in_reg (sext_in_reg x, VT2), VT1) -> (sext_in_reg x, minVT) pt2
5077 if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG &&
5078 EVT.bitsLT(cast<VTSDNode>(N0.getOperand(1))->getVT())) {
5079 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
5080 N0.getOperand(0), N1);
5083 // fold (sext_in_reg (sext x)) -> (sext x)
5084 // fold (sext_in_reg (aext x)) -> (sext x)
5085 // if x is small enough.
5086 if (N0.getOpcode() == ISD::SIGN_EXTEND || N0.getOpcode() == ISD::ANY_EXTEND) {
5087 SDValue N00 = N0.getOperand(0);
5088 if (N00.getValueType().getScalarType().getSizeInBits() <= EVTBits &&
5089 (!LegalOperations || TLI.isOperationLegal(ISD::SIGN_EXTEND, VT)))
5090 return DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, N00, N1);
5093 // fold (sext_in_reg x) -> (zext_in_reg x) if the sign bit is known zero.
5094 if (DAG.MaskedValueIsZero(N0, APInt::getBitsSet(VTBits, EVTBits-1, EVTBits)))
5095 return DAG.getZeroExtendInReg(N0, N->getDebugLoc(), EVT);
5097 // fold operands of sext_in_reg based on knowledge that the top bits are not
5099 if (SimplifyDemandedBits(SDValue(N, 0)))
5100 return SDValue(N, 0);
5102 // fold (sext_in_reg (load x)) -> (smaller sextload x)
5103 // fold (sext_in_reg (srl (load x), c)) -> (smaller sextload (x+c/evtbits))
5104 SDValue NarrowLoad = ReduceLoadWidth(N);
5105 if (NarrowLoad.getNode())
5108 // fold (sext_in_reg (srl X, 24), i8) -> (sra X, 24)
5109 // fold (sext_in_reg (srl X, 23), i8) -> (sra X, 23) iff possible.
5110 // We already fold "(sext_in_reg (srl X, 25), i8) -> srl X, 25" above.
5111 if (N0.getOpcode() == ISD::SRL) {
5112 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1)))
5113 if (ShAmt->getZExtValue()+EVTBits <= VTBits) {
5114 // We can turn this into an SRA iff the input to the SRL is already sign
5116 unsigned InSignBits = DAG.ComputeNumSignBits(N0.getOperand(0));
5117 if (VTBits-(ShAmt->getZExtValue()+EVTBits) < InSignBits)
5118 return DAG.getNode(ISD::SRA, N->getDebugLoc(), VT,
5119 N0.getOperand(0), N0.getOperand(1));
5123 // fold (sext_inreg (extload x)) -> (sextload x)
5124 if (ISD::isEXTLoad(N0.getNode()) &&
5125 ISD::isUNINDEXEDLoad(N0.getNode()) &&
5126 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5127 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5128 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5129 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5130 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
5132 LN0->getBasePtr(), LN0->getPointerInfo(),
5134 LN0->isVolatile(), LN0->isNonTemporal(),
5135 LN0->getAlignment());
5136 CombineTo(N, ExtLoad);
5137 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5138 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5140 // fold (sext_inreg (zextload x)) -> (sextload x) iff load has one use
5141 if (ISD::isZEXTLoad(N0.getNode()) && ISD::isUNINDEXEDLoad(N0.getNode()) &&
5143 EVT == cast<LoadSDNode>(N0)->getMemoryVT() &&
5144 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
5145 TLI.isLoadExtLegal(ISD::SEXTLOAD, EVT))) {
5146 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5147 SDValue ExtLoad = DAG.getExtLoad(ISD::SEXTLOAD, N->getDebugLoc(), VT,
5149 LN0->getBasePtr(), LN0->getPointerInfo(),
5151 LN0->isVolatile(), LN0->isNonTemporal(),
5152 LN0->getAlignment());
5153 CombineTo(N, ExtLoad);
5154 CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1));
5155 return SDValue(N, 0); // Return N so it doesn't get rechecked!
5158 // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16))
5159 if (EVTBits <= 16 && N0.getOpcode() == ISD::OR) {
5160 SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0),
5161 N0.getOperand(1), false);
5162 if (BSwap.getNode() != 0)
5163 return DAG.getNode(ISD::SIGN_EXTEND_INREG, N->getDebugLoc(), VT,
5170 SDValue DAGCombiner::visitTRUNCATE(SDNode *N) {
5171 SDValue N0 = N->getOperand(0);
5172 EVT VT = N->getValueType(0);
5173 bool isLE = TLI.isLittleEndian();
5176 if (N0.getValueType() == N->getValueType(0))
5178 // fold (truncate c1) -> c1
5179 if (isa<ConstantSDNode>(N0))
5180 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0);
5181 // fold (truncate (truncate x)) -> (truncate x)
5182 if (N0.getOpcode() == ISD::TRUNCATE)
5183 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0.getOperand(0));
5184 // fold (truncate (ext x)) -> (ext x) or (truncate x) or x
5185 if (N0.getOpcode() == ISD::ZERO_EXTEND ||
5186 N0.getOpcode() == ISD::SIGN_EXTEND ||
5187 N0.getOpcode() == ISD::ANY_EXTEND) {
5188 if (N0.getOperand(0).getValueType().bitsLT(VT))
5189 // if the source is smaller than the dest, we still need an extend
5190 return DAG.getNode(N0.getOpcode(), N->getDebugLoc(), VT,
5192 else if (N0.getOperand(0).getValueType().bitsGT(VT))
5193 // if the source is larger than the dest, than we just need the truncate
5194 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, N0.getOperand(0));
5196 // if the source and dest are the same type, we can drop both the extend
5197 // and the truncate.
5198 return N0.getOperand(0);
5201 // Fold extract-and-trunc into a narrow extract. For example:
5202 // i64 x = EXTRACT_VECTOR_ELT(v2i64 val, i32 1)
5203 // i32 y = TRUNCATE(i64 x)
5205 // v16i8 b = BITCAST (v2i64 val)
5206 // i8 x = EXTRACT_VECTOR_ELT(v16i8 b, i32 8)
5208 // Note: We only run this optimization after type legalization (which often
5209 // creates this pattern) and before operation legalization after which
5210 // we need to be more careful about the vector instructions that we generate.
5211 if (N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5212 LegalTypes && !LegalOperations && N0->hasOneUse()) {
5214 EVT VecTy = N0.getOperand(0).getValueType();
5215 EVT ExTy = N0.getValueType();
5216 EVT TrTy = N->getValueType(0);
5218 unsigned NumElem = VecTy.getVectorNumElements();
5219 unsigned SizeRatio = ExTy.getSizeInBits()/TrTy.getSizeInBits();
5221 EVT NVT = EVT::getVectorVT(*DAG.getContext(), TrTy, SizeRatio * NumElem);
5222 assert(NVT.getSizeInBits() == VecTy.getSizeInBits() && "Invalid Size");
5224 SDValue EltNo = N0->getOperand(1);
5225 if (isa<ConstantSDNode>(EltNo) && isTypeLegal(NVT)) {
5226 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5228 int Index = isLE ? (Elt*SizeRatio) : (Elt*SizeRatio + (SizeRatio-1));
5230 SDValue V = DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
5231 NVT, N0.getOperand(0));
5233 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT,
5234 N->getDebugLoc(), TrTy, V,
5235 DAG.getConstant(Index, MVT::i32));
5239 // See if we can simplify the input to this truncate through knowledge that
5240 // only the low bits are being used.
5241 // For example "trunc (or (shl x, 8), y)" // -> trunc y
5242 // Currently we only perform this optimization on scalars because vectors
5243 // may have different active low bits.
5244 if (!VT.isVector()) {
5246 GetDemandedBits(N0, APInt::getLowBitsSet(N0.getValueSizeInBits(),
5247 VT.getSizeInBits()));
5248 if (Shorter.getNode())
5249 return DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, Shorter);
5251 // fold (truncate (load x)) -> (smaller load x)
5252 // fold (truncate (srl (load x), c)) -> (smaller load (x+c/evtbits))
5253 if (!LegalTypes || TLI.isTypeDesirableForOp(N0.getOpcode(), VT)) {
5254 SDValue Reduced = ReduceLoadWidth(N);
5255 if (Reduced.getNode())
5259 // Simplify the operands using demanded-bits information.
5260 if (!VT.isVector() &&
5261 SimplifyDemandedBits(SDValue(N, 0)))
5262 return SDValue(N, 0);
5267 static SDNode *getBuildPairElt(SDNode *N, unsigned i) {
5268 SDValue Elt = N->getOperand(i);
5269 if (Elt.getOpcode() != ISD::MERGE_VALUES)
5270 return Elt.getNode();
5271 return Elt.getOperand(Elt.getResNo()).getNode();
5274 /// CombineConsecutiveLoads - build_pair (load, load) -> load
5275 /// if load locations are consecutive.
5276 SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) {
5277 assert(N->getOpcode() == ISD::BUILD_PAIR);
5279 LoadSDNode *LD1 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 0));
5280 LoadSDNode *LD2 = dyn_cast<LoadSDNode>(getBuildPairElt(N, 1));
5281 if (!LD1 || !LD2 || !ISD::isNON_EXTLoad(LD1) || !LD1->hasOneUse() ||
5282 LD1->getPointerInfo().getAddrSpace() !=
5283 LD2->getPointerInfo().getAddrSpace())
5285 EVT LD1VT = LD1->getValueType(0);
5287 if (ISD::isNON_EXTLoad(LD2) &&
5289 // If both are volatile this would reduce the number of volatile loads.
5290 // If one is volatile it might be ok, but play conservative and bail out.
5291 !LD1->isVolatile() &&
5292 !LD2->isVolatile() &&
5293 DAG.isConsecutiveLoad(LD2, LD1, LD1VT.getSizeInBits()/8, 1)) {
5294 unsigned Align = LD1->getAlignment();
5295 unsigned NewAlign = TLI.getTargetData()->
5296 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
5298 if (NewAlign <= Align &&
5299 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT)))
5300 return DAG.getLoad(VT, N->getDebugLoc(), LD1->getChain(),
5301 LD1->getBasePtr(), LD1->getPointerInfo(),
5302 false, false, false, Align);
5308 SDValue DAGCombiner::visitBITCAST(SDNode *N) {
5309 SDValue N0 = N->getOperand(0);
5310 EVT VT = N->getValueType(0);
5312 // If the input is a BUILD_VECTOR with all constant elements, fold this now.
5313 // Only do this before legalize, since afterward the target may be depending
5314 // on the bitconvert.
5315 // First check to see if this is all constant.
5317 N0.getOpcode() == ISD::BUILD_VECTOR && N0.getNode()->hasOneUse() &&
5319 bool isSimple = true;
5320 for (unsigned i = 0, e = N0.getNumOperands(); i != e; ++i)
5321 if (N0.getOperand(i).getOpcode() != ISD::UNDEF &&
5322 N0.getOperand(i).getOpcode() != ISD::Constant &&
5323 N0.getOperand(i).getOpcode() != ISD::ConstantFP) {
5328 EVT DestEltVT = N->getValueType(0).getVectorElementType();
5329 assert(!DestEltVT.isVector() &&
5330 "Element type of vector ValueType must not be vector!");
5332 return ConstantFoldBITCASTofBUILD_VECTOR(N0.getNode(), DestEltVT);
5335 // If the input is a constant, let getNode fold it.
5336 if (isa<ConstantSDNode>(N0) || isa<ConstantFPSDNode>(N0)) {
5337 SDValue Res = DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, N0);
5338 if (Res.getNode() != N) {
5339 if (!LegalOperations ||
5340 TLI.isOperationLegal(Res.getNode()->getOpcode(), VT))
5343 // Folding it resulted in an illegal node, and it's too late to
5344 // do that. Clean up the old node and forego the transformation.
5345 // Ideally this won't happen very often, because instcombine
5346 // and the earlier dagcombine runs (where illegal nodes are
5347 // permitted) should have folded most of them already.
5348 DAG.DeleteNode(Res.getNode());
5352 // (conv (conv x, t1), t2) -> (conv x, t2)
5353 if (N0.getOpcode() == ISD::BITCAST)
5354 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT,
5357 // fold (conv (load x)) -> (load (conv*)x)
5358 // If the resultant load doesn't need a higher alignment than the original!
5359 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
5360 // Do not change the width of a volatile load.
5361 !cast<LoadSDNode>(N0)->isVolatile() &&
5362 (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT))) {
5363 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
5364 unsigned Align = TLI.getTargetData()->
5365 getABITypeAlignment(VT.getTypeForEVT(*DAG.getContext()));
5366 unsigned OrigAlign = LN0->getAlignment();
5368 if (Align <= OrigAlign) {
5369 SDValue Load = DAG.getLoad(VT, N->getDebugLoc(), LN0->getChain(),
5370 LN0->getBasePtr(), LN0->getPointerInfo(),
5371 LN0->isVolatile(), LN0->isNonTemporal(),
5372 LN0->isInvariant(), OrigAlign);
5374 CombineTo(N0.getNode(),
5375 DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5376 N0.getValueType(), Load),
5382 // fold (bitconvert (fneg x)) -> (xor (bitconvert x), signbit)
5383 // fold (bitconvert (fabs x)) -> (and (bitconvert x), (not signbit))
5384 // This often reduces constant pool loads.
5385 if (((N0.getOpcode() == ISD::FNEG && !TLI.isFNegFree(VT)) ||
5386 (N0.getOpcode() == ISD::FABS && !TLI.isFAbsFree(VT))) &&
5387 N0.getNode()->hasOneUse() && VT.isInteger() && !VT.isVector()) {
5388 SDValue NewConv = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(), VT,
5390 AddToWorkList(NewConv.getNode());
5392 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
5393 if (N0.getOpcode() == ISD::FNEG)
5394 return DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
5395 NewConv, DAG.getConstant(SignBit, VT));
5396 assert(N0.getOpcode() == ISD::FABS);
5397 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
5398 NewConv, DAG.getConstant(~SignBit, VT));
5401 // fold (bitconvert (fcopysign cst, x)) ->
5402 // (or (and (bitconvert x), sign), (and cst, (not sign)))
5403 // Note that we don't handle (copysign x, cst) because this can always be
5404 // folded to an fneg or fabs.
5405 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse() &&
5406 isa<ConstantFPSDNode>(N0.getOperand(0)) &&
5407 VT.isInteger() && !VT.isVector()) {
5408 unsigned OrigXWidth = N0.getOperand(1).getValueType().getSizeInBits();
5409 EVT IntXVT = EVT::getIntegerVT(*DAG.getContext(), OrigXWidth);
5410 if (isTypeLegal(IntXVT)) {
5411 SDValue X = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5412 IntXVT, N0.getOperand(1));
5413 AddToWorkList(X.getNode());
5415 // If X has a different width than the result/lhs, sext it or truncate it.
5416 unsigned VTWidth = VT.getSizeInBits();
5417 if (OrigXWidth < VTWidth) {
5418 X = DAG.getNode(ISD::SIGN_EXTEND, N->getDebugLoc(), VT, X);
5419 AddToWorkList(X.getNode());
5420 } else if (OrigXWidth > VTWidth) {
5421 // To get the sign bit in the right place, we have to shift it right
5422 // before truncating.
5423 X = DAG.getNode(ISD::SRL, X.getDebugLoc(),
5424 X.getValueType(), X,
5425 DAG.getConstant(OrigXWidth-VTWidth, X.getValueType()));
5426 AddToWorkList(X.getNode());
5427 X = DAG.getNode(ISD::TRUNCATE, X.getDebugLoc(), VT, X);
5428 AddToWorkList(X.getNode());
5431 APInt SignBit = APInt::getSignBit(VT.getSizeInBits());
5432 X = DAG.getNode(ISD::AND, X.getDebugLoc(), VT,
5433 X, DAG.getConstant(SignBit, VT));
5434 AddToWorkList(X.getNode());
5436 SDValue Cst = DAG.getNode(ISD::BITCAST, N0.getDebugLoc(),
5437 VT, N0.getOperand(0));
5438 Cst = DAG.getNode(ISD::AND, Cst.getDebugLoc(), VT,
5439 Cst, DAG.getConstant(~SignBit, VT));
5440 AddToWorkList(Cst.getNode());
5442 return DAG.getNode(ISD::OR, N->getDebugLoc(), VT, X, Cst);
5446 // bitconvert(build_pair(ld, ld)) -> ld iff load locations are consecutive.
5447 if (N0.getOpcode() == ISD::BUILD_PAIR) {
5448 SDValue CombineLD = CombineConsecutiveLoads(N0.getNode(), VT);
5449 if (CombineLD.getNode())
5456 SDValue DAGCombiner::visitBUILD_PAIR(SDNode *N) {
5457 EVT VT = N->getValueType(0);
5458 return CombineConsecutiveLoads(N, VT);
5461 /// ConstantFoldBITCASTofBUILD_VECTOR - We know that BV is a build_vector
5462 /// node with Constant, ConstantFP or Undef operands. DstEltVT indicates the
5463 /// destination element value type.
5464 SDValue DAGCombiner::
5465 ConstantFoldBITCASTofBUILD_VECTOR(SDNode *BV, EVT DstEltVT) {
5466 EVT SrcEltVT = BV->getValueType(0).getVectorElementType();
5468 // If this is already the right type, we're done.
5469 if (SrcEltVT == DstEltVT) return SDValue(BV, 0);
5471 unsigned SrcBitSize = SrcEltVT.getSizeInBits();
5472 unsigned DstBitSize = DstEltVT.getSizeInBits();
5474 // If this is a conversion of N elements of one type to N elements of another
5475 // type, convert each element. This handles FP<->INT cases.
5476 if (SrcBitSize == DstBitSize) {
5477 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
5478 BV->getValueType(0).getVectorNumElements());
5480 // Due to the FP element handling below calling this routine recursively,
5481 // we can end up with a scalar-to-vector node here.
5482 if (BV->getOpcode() == ISD::SCALAR_TO_VECTOR)
5483 return DAG.getNode(ISD::SCALAR_TO_VECTOR, BV->getDebugLoc(), VT,
5484 DAG.getNode(ISD::BITCAST, BV->getDebugLoc(),
5485 DstEltVT, BV->getOperand(0)));
5487 SmallVector<SDValue, 8> Ops;
5488 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
5489 SDValue Op = BV->getOperand(i);
5490 // If the vector element type is not legal, the BUILD_VECTOR operands
5491 // are promoted and implicitly truncated. Make that explicit here.
5492 if (Op.getValueType() != SrcEltVT)
5493 Op = DAG.getNode(ISD::TRUNCATE, BV->getDebugLoc(), SrcEltVT, Op);
5494 Ops.push_back(DAG.getNode(ISD::BITCAST, BV->getDebugLoc(),
5496 AddToWorkList(Ops.back().getNode());
5498 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5499 &Ops[0], Ops.size());
5502 // Otherwise, we're growing or shrinking the elements. To avoid having to
5503 // handle annoying details of growing/shrinking FP values, we convert them to
5505 if (SrcEltVT.isFloatingPoint()) {
5506 // Convert the input float vector to a int vector where the elements are the
5508 assert((SrcEltVT == MVT::f32 || SrcEltVT == MVT::f64) && "Unknown FP VT!");
5509 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), SrcEltVT.getSizeInBits());
5510 BV = ConstantFoldBITCASTofBUILD_VECTOR(BV, IntVT).getNode();
5514 // Now we know the input is an integer vector. If the output is a FP type,
5515 // convert to integer first, then to FP of the right size.
5516 if (DstEltVT.isFloatingPoint()) {
5517 assert((DstEltVT == MVT::f32 || DstEltVT == MVT::f64) && "Unknown FP VT!");
5518 EVT TmpVT = EVT::getIntegerVT(*DAG.getContext(), DstEltVT.getSizeInBits());
5519 SDNode *Tmp = ConstantFoldBITCASTofBUILD_VECTOR(BV, TmpVT).getNode();
5521 // Next, convert to FP elements of the same size.
5522 return ConstantFoldBITCASTofBUILD_VECTOR(Tmp, DstEltVT);
5525 // Okay, we know the src/dst types are both integers of differing types.
5526 // Handling growing first.
5527 assert(SrcEltVT.isInteger() && DstEltVT.isInteger());
5528 if (SrcBitSize < DstBitSize) {
5529 unsigned NumInputsPerOutput = DstBitSize/SrcBitSize;
5531 SmallVector<SDValue, 8> Ops;
5532 for (unsigned i = 0, e = BV->getNumOperands(); i != e;
5533 i += NumInputsPerOutput) {
5534 bool isLE = TLI.isLittleEndian();
5535 APInt NewBits = APInt(DstBitSize, 0);
5536 bool EltIsUndef = true;
5537 for (unsigned j = 0; j != NumInputsPerOutput; ++j) {
5538 // Shift the previously computed bits over.
5539 NewBits <<= SrcBitSize;
5540 SDValue Op = BV->getOperand(i+ (isLE ? (NumInputsPerOutput-j-1) : j));
5541 if (Op.getOpcode() == ISD::UNDEF) continue;
5544 NewBits |= cast<ConstantSDNode>(Op)->getAPIntValue().
5545 zextOrTrunc(SrcBitSize).zext(DstBitSize);
5549 Ops.push_back(DAG.getUNDEF(DstEltVT));
5551 Ops.push_back(DAG.getConstant(NewBits, DstEltVT));
5554 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT, Ops.size());
5555 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5556 &Ops[0], Ops.size());
5559 // Finally, this must be the case where we are shrinking elements: each input
5560 // turns into multiple outputs.
5561 bool isS2V = ISD::isScalarToVector(BV);
5562 unsigned NumOutputsPerInput = SrcBitSize/DstBitSize;
5563 EVT VT = EVT::getVectorVT(*DAG.getContext(), DstEltVT,
5564 NumOutputsPerInput*BV->getNumOperands());
5565 SmallVector<SDValue, 8> Ops;
5567 for (unsigned i = 0, e = BV->getNumOperands(); i != e; ++i) {
5568 if (BV->getOperand(i).getOpcode() == ISD::UNDEF) {
5569 for (unsigned j = 0; j != NumOutputsPerInput; ++j)
5570 Ops.push_back(DAG.getUNDEF(DstEltVT));
5574 APInt OpVal = cast<ConstantSDNode>(BV->getOperand(i))->
5575 getAPIntValue().zextOrTrunc(SrcBitSize);
5577 for (unsigned j = 0; j != NumOutputsPerInput; ++j) {
5578 APInt ThisVal = OpVal.trunc(DstBitSize);
5579 Ops.push_back(DAG.getConstant(ThisVal, DstEltVT));
5580 if (isS2V && i == 0 && j == 0 && ThisVal.zext(SrcBitSize) == OpVal)
5581 // Simply turn this into a SCALAR_TO_VECTOR of the new type.
5582 return DAG.getNode(ISD::SCALAR_TO_VECTOR, BV->getDebugLoc(), VT,
5584 OpVal = OpVal.lshr(DstBitSize);
5587 // For big endian targets, swap the order of the pieces of each element.
5588 if (TLI.isBigEndian())
5589 std::reverse(Ops.end()-NumOutputsPerInput, Ops.end());
5592 return DAG.getNode(ISD::BUILD_VECTOR, BV->getDebugLoc(), VT,
5593 &Ops[0], Ops.size());
5596 SDValue DAGCombiner::visitFADD(SDNode *N) {
5597 SDValue N0 = N->getOperand(0);
5598 SDValue N1 = N->getOperand(1);
5599 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5600 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5601 EVT VT = N->getValueType(0);
5604 if (VT.isVector()) {
5605 SDValue FoldedVOp = SimplifyVBinOp(N);
5606 if (FoldedVOp.getNode()) return FoldedVOp;
5609 // fold (fadd c1, c2) -> (fadd c1, c2)
5610 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5611 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N1);
5612 // canonicalize constant to RHS
5613 if (N0CFP && !N1CFP)
5614 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N1, N0);
5615 // fold (fadd A, 0) -> A
5616 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
5617 N1CFP->getValueAPF().isZero())
5619 // fold (fadd A, (fneg B)) -> (fsub A, B)
5620 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
5621 isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
5622 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N0,
5623 GetNegatedExpression(N1, DAG, LegalOperations));
5624 // fold (fadd (fneg A), B) -> (fsub B, A)
5625 if ((!LegalOperations || TLI.isOperationLegalOrCustom(ISD::FSUB, VT)) &&
5626 isNegatibleForFree(N0, LegalOperations, TLI, &DAG.getTarget().Options) == 2)
5627 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N1,
5628 GetNegatedExpression(N0, DAG, LegalOperations));
5630 // If allowed, fold (fadd (fadd x, c1), c2) -> (fadd x, (fadd c1, c2))
5631 if (DAG.getTarget().Options.UnsafeFPMath && N1CFP &&
5632 N0.getOpcode() == ISD::FADD && N0.getNode()->hasOneUse() &&
5633 isa<ConstantFPSDNode>(N0.getOperand(1)))
5634 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0.getOperand(0),
5635 DAG.getNode(ISD::FADD, N->getDebugLoc(), VT,
5636 N0.getOperand(1), N1));
5641 SDValue DAGCombiner::visitFSUB(SDNode *N) {
5642 SDValue N0 = N->getOperand(0);
5643 SDValue N1 = N->getOperand(1);
5644 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5645 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5646 EVT VT = N->getValueType(0);
5649 if (VT.isVector()) {
5650 SDValue FoldedVOp = SimplifyVBinOp(N);
5651 if (FoldedVOp.getNode()) return FoldedVOp;
5654 // fold (fsub c1, c2) -> c1-c2
5655 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5656 return DAG.getNode(ISD::FSUB, N->getDebugLoc(), VT, N0, N1);
5657 // fold (fsub A, 0) -> A
5658 if (DAG.getTarget().Options.UnsafeFPMath &&
5659 N1CFP && N1CFP->getValueAPF().isZero())
5661 // fold (fsub 0, B) -> -B
5662 if (DAG.getTarget().Options.UnsafeFPMath &&
5663 N0CFP && N0CFP->getValueAPF().isZero()) {
5664 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
5665 return GetNegatedExpression(N1, DAG, LegalOperations);
5666 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
5667 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT, N1);
5669 // fold (fsub A, (fneg B)) -> (fadd A, B)
5670 if (isNegatibleForFree(N1, LegalOperations, TLI, &DAG.getTarget().Options))
5671 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0,
5672 GetNegatedExpression(N1, DAG, LegalOperations));
5674 // If 'unsafe math' is enabled, fold
5675 // (fsub x, (fadd x, y)) -> (fneg y) &
5676 // (fsub x, (fadd y, x)) -> (fneg y)
5677 if (DAG.getTarget().Options.UnsafeFPMath) {
5678 if (N1.getOpcode() == ISD::FADD) {
5679 SDValue N10 = N1->getOperand(0);
5680 SDValue N11 = N1->getOperand(1);
5682 if (N10 == N0 && isNegatibleForFree(N11, LegalOperations, TLI,
5683 &DAG.getTarget().Options))
5684 return GetNegatedExpression(N11, DAG, LegalOperations);
5685 else if (N11 == N0 && isNegatibleForFree(N10, LegalOperations, TLI,
5686 &DAG.getTarget().Options))
5687 return GetNegatedExpression(N10, DAG, LegalOperations);
5694 SDValue DAGCombiner::visitFMUL(SDNode *N) {
5695 SDValue N0 = N->getOperand(0);
5696 SDValue N1 = N->getOperand(1);
5697 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5698 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5699 EVT VT = N->getValueType(0);
5700 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5703 if (VT.isVector()) {
5704 SDValue FoldedVOp = SimplifyVBinOp(N);
5705 if (FoldedVOp.getNode()) return FoldedVOp;
5708 // fold (fmul c1, c2) -> c1*c2
5709 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5710 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0, N1);
5711 // canonicalize constant to RHS
5712 if (N0CFP && !N1CFP)
5713 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N1, N0);
5714 // fold (fmul A, 0) -> 0
5715 if (DAG.getTarget().Options.UnsafeFPMath &&
5716 N1CFP && N1CFP->getValueAPF().isZero())
5718 // fold (fmul A, 0) -> 0, vector edition.
5719 if (DAG.getTarget().Options.UnsafeFPMath &&
5720 ISD::isBuildVectorAllZeros(N1.getNode()))
5722 // fold (fmul X, 2.0) -> (fadd X, X)
5723 if (N1CFP && N1CFP->isExactlyValue(+2.0))
5724 return DAG.getNode(ISD::FADD, N->getDebugLoc(), VT, N0, N0);
5725 // fold (fmul X, -1.0) -> (fneg X)
5726 if (N1CFP && N1CFP->isExactlyValue(-1.0))
5727 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
5728 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT, N0);
5730 // fold (fmul (fneg X), (fneg Y)) -> (fmul X, Y)
5731 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
5732 &DAG.getTarget().Options)) {
5733 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
5734 &DAG.getTarget().Options)) {
5735 // Both can be negated for free, check to see if at least one is cheaper
5737 if (LHSNeg == 2 || RHSNeg == 2)
5738 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5739 GetNegatedExpression(N0, DAG, LegalOperations),
5740 GetNegatedExpression(N1, DAG, LegalOperations));
5744 // If allowed, fold (fmul (fmul x, c1), c2) -> (fmul x, (fmul c1, c2))
5745 if (DAG.getTarget().Options.UnsafeFPMath &&
5746 N1CFP && N0.getOpcode() == ISD::FMUL &&
5747 N0.getNode()->hasOneUse() && isa<ConstantFPSDNode>(N0.getOperand(1)))
5748 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0.getOperand(0),
5749 DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT,
5750 N0.getOperand(1), N1));
5755 SDValue DAGCombiner::visitFDIV(SDNode *N) {
5756 SDValue N0 = N->getOperand(0);
5757 SDValue N1 = N->getOperand(1);
5758 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5759 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5760 EVT VT = N->getValueType(0);
5761 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5764 if (VT.isVector()) {
5765 SDValue FoldedVOp = SimplifyVBinOp(N);
5766 if (FoldedVOp.getNode()) return FoldedVOp;
5769 // fold (fdiv c1, c2) -> c1/c2
5770 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5771 return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT, N0, N1);
5773 // fold (fdiv X, c2) -> fmul X, 1/c2 if losing precision is acceptable.
5774 if (N1CFP && VT != MVT::ppcf128 && DAG.getTarget().Options.UnsafeFPMath) {
5775 // Compute the reciprocal 1.0 / c2.
5776 APFloat N1APF = N1CFP->getValueAPF();
5777 APFloat Recip(N1APF.getSemantics(), 1); // 1.0
5778 APFloat::opStatus st = Recip.divide(N1APF, APFloat::rmNearestTiesToEven);
5779 // Only do the transform if the reciprocal is a legal fp immediate that
5780 // isn't too nasty (eg NaN, denormal, ...).
5781 if ((st == APFloat::opOK || st == APFloat::opInexact) && // Not too nasty
5782 (!LegalOperations ||
5783 // FIXME: custom lowering of ConstantFP might fail (see e.g. ARM
5784 // backend)... we should handle this gracefully after Legalize.
5785 // TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT) ||
5786 TLI.isOperationLegal(llvm::ISD::ConstantFP, VT) ||
5787 TLI.isFPImmLegal(Recip, VT)))
5788 return DAG.getNode(ISD::FMUL, N->getDebugLoc(), VT, N0,
5789 DAG.getConstantFP(Recip, VT));
5792 // (fdiv (fneg X), (fneg Y)) -> (fdiv X, Y)
5793 if (char LHSNeg = isNegatibleForFree(N0, LegalOperations, TLI,
5794 &DAG.getTarget().Options)) {
5795 if (char RHSNeg = isNegatibleForFree(N1, LegalOperations, TLI,
5796 &DAG.getTarget().Options)) {
5797 // Both can be negated for free, check to see if at least one is cheaper
5799 if (LHSNeg == 2 || RHSNeg == 2)
5800 return DAG.getNode(ISD::FDIV, N->getDebugLoc(), VT,
5801 GetNegatedExpression(N0, DAG, LegalOperations),
5802 GetNegatedExpression(N1, DAG, LegalOperations));
5809 SDValue DAGCombiner::visitFREM(SDNode *N) {
5810 SDValue N0 = N->getOperand(0);
5811 SDValue N1 = N->getOperand(1);
5812 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5813 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5814 EVT VT = N->getValueType(0);
5816 // fold (frem c1, c2) -> fmod(c1,c2)
5817 if (N0CFP && N1CFP && VT != MVT::ppcf128)
5818 return DAG.getNode(ISD::FREM, N->getDebugLoc(), VT, N0, N1);
5823 SDValue DAGCombiner::visitFCOPYSIGN(SDNode *N) {
5824 SDValue N0 = N->getOperand(0);
5825 SDValue N1 = N->getOperand(1);
5826 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5827 ConstantFPSDNode *N1CFP = dyn_cast<ConstantFPSDNode>(N1);
5828 EVT VT = N->getValueType(0);
5830 if (N0CFP && N1CFP && VT != MVT::ppcf128) // Constant fold
5831 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT, N0, N1);
5834 const APFloat& V = N1CFP->getValueAPF();
5835 // copysign(x, c1) -> fabs(x) iff ispos(c1)
5836 // copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
5837 if (!V.isNegative()) {
5838 if (!LegalOperations || TLI.isOperationLegal(ISD::FABS, VT))
5839 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
5841 if (!LegalOperations || TLI.isOperationLegal(ISD::FNEG, VT))
5842 return DAG.getNode(ISD::FNEG, N->getDebugLoc(), VT,
5843 DAG.getNode(ISD::FABS, N0.getDebugLoc(), VT, N0));
5847 // copysign(fabs(x), y) -> copysign(x, y)
5848 // copysign(fneg(x), y) -> copysign(x, y)
5849 // copysign(copysign(x,z), y) -> copysign(x, y)
5850 if (N0.getOpcode() == ISD::FABS || N0.getOpcode() == ISD::FNEG ||
5851 N0.getOpcode() == ISD::FCOPYSIGN)
5852 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
5853 N0.getOperand(0), N1);
5855 // copysign(x, abs(y)) -> abs(x)
5856 if (N1.getOpcode() == ISD::FABS)
5857 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
5859 // copysign(x, copysign(y,z)) -> copysign(x, z)
5860 if (N1.getOpcode() == ISD::FCOPYSIGN)
5861 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
5862 N0, N1.getOperand(1));
5864 // copysign(x, fp_extend(y)) -> copysign(x, y)
5865 // copysign(x, fp_round(y)) -> copysign(x, y)
5866 if (N1.getOpcode() == ISD::FP_EXTEND || N1.getOpcode() == ISD::FP_ROUND)
5867 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
5868 N0, N1.getOperand(0));
5873 SDValue DAGCombiner::visitSINT_TO_FP(SDNode *N) {
5874 SDValue N0 = N->getOperand(0);
5875 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
5876 EVT VT = N->getValueType(0);
5877 EVT OpVT = N0.getValueType();
5879 // fold (sint_to_fp c1) -> c1fp
5880 if (N0C && OpVT != MVT::ppcf128 &&
5881 // ...but only if the target supports immediate floating-point values
5882 (!LegalOperations ||
5883 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
5884 return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
5886 // If the input is a legal type, and SINT_TO_FP is not legal on this target,
5887 // but UINT_TO_FP is legal on this target, try to convert.
5888 if (!TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT) &&
5889 TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT)) {
5890 // If the sign bit is known to be zero, we can change this to UINT_TO_FP.
5891 if (DAG.SignBitIsZero(N0))
5892 return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
5898 SDValue DAGCombiner::visitUINT_TO_FP(SDNode *N) {
5899 SDValue N0 = N->getOperand(0);
5900 ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
5901 EVT VT = N->getValueType(0);
5902 EVT OpVT = N0.getValueType();
5904 // fold (uint_to_fp c1) -> c1fp
5905 if (N0C && OpVT != MVT::ppcf128 &&
5906 // ...but only if the target supports immediate floating-point values
5907 (!LegalOperations ||
5908 TLI.isOperationLegalOrCustom(llvm::ISD::ConstantFP, VT)))
5909 return DAG.getNode(ISD::UINT_TO_FP, N->getDebugLoc(), VT, N0);
5911 // If the input is a legal type, and UINT_TO_FP is not legal on this target,
5912 // but SINT_TO_FP is legal on this target, try to convert.
5913 if (!TLI.isOperationLegalOrCustom(ISD::UINT_TO_FP, OpVT) &&
5914 TLI.isOperationLegalOrCustom(ISD::SINT_TO_FP, OpVT)) {
5915 // If the sign bit is known to be zero, we can change this to SINT_TO_FP.
5916 if (DAG.SignBitIsZero(N0))
5917 return DAG.getNode(ISD::SINT_TO_FP, N->getDebugLoc(), VT, N0);
5923 SDValue DAGCombiner::visitFP_TO_SINT(SDNode *N) {
5924 SDValue N0 = N->getOperand(0);
5925 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5926 EVT VT = N->getValueType(0);
5928 // fold (fp_to_sint c1fp) -> c1
5930 return DAG.getNode(ISD::FP_TO_SINT, N->getDebugLoc(), VT, N0);
5935 SDValue DAGCombiner::visitFP_TO_UINT(SDNode *N) {
5936 SDValue N0 = N->getOperand(0);
5937 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5938 EVT VT = N->getValueType(0);
5940 // fold (fp_to_uint c1fp) -> c1
5941 if (N0CFP && VT != MVT::ppcf128)
5942 return DAG.getNode(ISD::FP_TO_UINT, N->getDebugLoc(), VT, N0);
5947 SDValue DAGCombiner::visitFP_ROUND(SDNode *N) {
5948 SDValue N0 = N->getOperand(0);
5949 SDValue N1 = N->getOperand(1);
5950 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5951 EVT VT = N->getValueType(0);
5953 // fold (fp_round c1fp) -> c1fp
5954 if (N0CFP && N0.getValueType() != MVT::ppcf128)
5955 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT, N0, N1);
5957 // fold (fp_round (fp_extend x)) -> x
5958 if (N0.getOpcode() == ISD::FP_EXTEND && VT == N0.getOperand(0).getValueType())
5959 return N0.getOperand(0);
5961 // fold (fp_round (fp_round x)) -> (fp_round x)
5962 if (N0.getOpcode() == ISD::FP_ROUND) {
5963 // This is a value preserving truncation if both round's are.
5964 bool IsTrunc = N->getConstantOperandVal(1) == 1 &&
5965 N0.getNode()->getConstantOperandVal(1) == 1;
5966 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT, N0.getOperand(0),
5967 DAG.getIntPtrConstant(IsTrunc));
5970 // fold (fp_round (copysign X, Y)) -> (copysign (fp_round X), Y)
5971 if (N0.getOpcode() == ISD::FCOPYSIGN && N0.getNode()->hasOneUse()) {
5972 SDValue Tmp = DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(), VT,
5973 N0.getOperand(0), N1);
5974 AddToWorkList(Tmp.getNode());
5975 return DAG.getNode(ISD::FCOPYSIGN, N->getDebugLoc(), VT,
5976 Tmp, N0.getOperand(1));
5982 SDValue DAGCombiner::visitFP_ROUND_INREG(SDNode *N) {
5983 SDValue N0 = N->getOperand(0);
5984 EVT VT = N->getValueType(0);
5985 EVT EVT = cast<VTSDNode>(N->getOperand(1))->getVT();
5986 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
5988 // fold (fp_round_inreg c1fp) -> c1fp
5989 if (N0CFP && isTypeLegal(EVT)) {
5990 SDValue Round = DAG.getConstantFP(*N0CFP->getConstantFPValue(), EVT);
5991 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, Round);
5997 SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) {
5998 SDValue N0 = N->getOperand(0);
5999 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6000 EVT VT = N->getValueType(0);
6002 // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
6003 if (N->hasOneUse() &&
6004 N->use_begin()->getOpcode() == ISD::FP_ROUND)
6007 // fold (fp_extend c1fp) -> c1fp
6008 if (N0CFP && VT != MVT::ppcf128)
6009 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, N0);
6011 // Turn fp_extend(fp_round(X, 1)) -> x since the fp_round doesn't affect the
6013 if (N0.getOpcode() == ISD::FP_ROUND
6014 && N0.getNode()->getConstantOperandVal(1) == 1) {
6015 SDValue In = N0.getOperand(0);
6016 if (In.getValueType() == VT) return In;
6017 if (VT.bitsLT(In.getValueType()))
6018 return DAG.getNode(ISD::FP_ROUND, N->getDebugLoc(), VT,
6019 In, N0.getOperand(1));
6020 return DAG.getNode(ISD::FP_EXTEND, N->getDebugLoc(), VT, In);
6023 // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
6024 if (ISD::isNON_EXTLoad(N0.getNode()) && N0.hasOneUse() &&
6025 ((!LegalOperations && !cast<LoadSDNode>(N0)->isVolatile()) ||
6026 TLI.isLoadExtLegal(ISD::EXTLOAD, N0.getValueType()))) {
6027 LoadSDNode *LN0 = cast<LoadSDNode>(N0);
6028 SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, N->getDebugLoc(), VT,
6030 LN0->getBasePtr(), LN0->getPointerInfo(),
6032 LN0->isVolatile(), LN0->isNonTemporal(),
6033 LN0->getAlignment());
6034 CombineTo(N, ExtLoad);
6035 CombineTo(N0.getNode(),
6036 DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(),
6037 N0.getValueType(), ExtLoad, DAG.getIntPtrConstant(1)),
6038 ExtLoad.getValue(1));
6039 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6045 SDValue DAGCombiner::visitFNEG(SDNode *N) {
6046 SDValue N0 = N->getOperand(0);
6047 EVT VT = N->getValueType(0);
6049 if (isNegatibleForFree(N0, LegalOperations, DAG.getTargetLoweringInfo(),
6050 &DAG.getTarget().Options))
6051 return GetNegatedExpression(N0, DAG, LegalOperations);
6053 // Transform fneg(bitconvert(x)) -> bitconvert(x^sign) to avoid loading
6054 // constant pool values.
6055 if (!TLI.isFNegFree(VT) && N0.getOpcode() == ISD::BITCAST &&
6057 N0.getNode()->hasOneUse() &&
6058 N0.getOperand(0).getValueType().isInteger()) {
6059 SDValue Int = N0.getOperand(0);
6060 EVT IntVT = Int.getValueType();
6061 if (IntVT.isInteger() && !IntVT.isVector()) {
6062 Int = DAG.getNode(ISD::XOR, N0.getDebugLoc(), IntVT, Int,
6063 DAG.getConstant(APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
6064 AddToWorkList(Int.getNode());
6065 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
6073 SDValue DAGCombiner::visitFABS(SDNode *N) {
6074 SDValue N0 = N->getOperand(0);
6075 ConstantFPSDNode *N0CFP = dyn_cast<ConstantFPSDNode>(N0);
6076 EVT VT = N->getValueType(0);
6078 // fold (fabs c1) -> fabs(c1)
6079 if (N0CFP && VT != MVT::ppcf128)
6080 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0);
6081 // fold (fabs (fabs x)) -> (fabs x)
6082 if (N0.getOpcode() == ISD::FABS)
6083 return N->getOperand(0);
6084 // fold (fabs (fneg x)) -> (fabs x)
6085 // fold (fabs (fcopysign x, y)) -> (fabs x)
6086 if (N0.getOpcode() == ISD::FNEG || N0.getOpcode() == ISD::FCOPYSIGN)
6087 return DAG.getNode(ISD::FABS, N->getDebugLoc(), VT, N0.getOperand(0));
6089 // Transform fabs(bitconvert(x)) -> bitconvert(x&~sign) to avoid loading
6090 // constant pool values.
6091 if (!TLI.isFAbsFree(VT) &&
6092 N0.getOpcode() == ISD::BITCAST && N0.getNode()->hasOneUse() &&
6093 N0.getOperand(0).getValueType().isInteger() &&
6094 !N0.getOperand(0).getValueType().isVector()) {
6095 SDValue Int = N0.getOperand(0);
6096 EVT IntVT = Int.getValueType();
6097 if (IntVT.isInteger() && !IntVT.isVector()) {
6098 Int = DAG.getNode(ISD::AND, N0.getDebugLoc(), IntVT, Int,
6099 DAG.getConstant(~APInt::getSignBit(IntVT.getSizeInBits()), IntVT));
6100 AddToWorkList(Int.getNode());
6101 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(),
6102 N->getValueType(0), Int);
6109 SDValue DAGCombiner::visitBRCOND(SDNode *N) {
6110 SDValue Chain = N->getOperand(0);
6111 SDValue N1 = N->getOperand(1);
6112 SDValue N2 = N->getOperand(2);
6114 // If N is a constant we could fold this into a fallthrough or unconditional
6115 // branch. However that doesn't happen very often in normal code, because
6116 // Instcombine/SimplifyCFG should have handled the available opportunities.
6117 // If we did this folding here, it would be necessary to update the
6118 // MachineBasicBlock CFG, which is awkward.
6120 // fold a brcond with a setcc condition into a BR_CC node if BR_CC is legal
6122 if (N1.getOpcode() == ISD::SETCC &&
6123 TLI.isOperationLegalOrCustom(ISD::BR_CC, MVT::Other)) {
6124 return DAG.getNode(ISD::BR_CC, N->getDebugLoc(), MVT::Other,
6125 Chain, N1.getOperand(2),
6126 N1.getOperand(0), N1.getOperand(1), N2);
6129 if ((N1.hasOneUse() && N1.getOpcode() == ISD::SRL) ||
6130 ((N1.getOpcode() == ISD::TRUNCATE && N1.hasOneUse()) &&
6131 (N1.getOperand(0).hasOneUse() &&
6132 N1.getOperand(0).getOpcode() == ISD::SRL))) {
6134 if (N1.getOpcode() == ISD::TRUNCATE) {
6135 // Look pass the truncate.
6136 Trunc = N1.getNode();
6137 N1 = N1.getOperand(0);
6140 // Match this pattern so that we can generate simpler code:
6143 // %b = and i32 %a, 2
6144 // %c = srl i32 %b, 1
6145 // brcond i32 %c ...
6150 // %b = and i32 %a, 2
6151 // %c = setcc eq %b, 0
6154 // This applies only when the AND constant value has one bit set and the
6155 // SRL constant is equal to the log2 of the AND constant. The back-end is
6156 // smart enough to convert the result into a TEST/JMP sequence.
6157 SDValue Op0 = N1.getOperand(0);
6158 SDValue Op1 = N1.getOperand(1);
6160 if (Op0.getOpcode() == ISD::AND &&
6161 Op1.getOpcode() == ISD::Constant) {
6162 SDValue AndOp1 = Op0.getOperand(1);
6164 if (AndOp1.getOpcode() == ISD::Constant) {
6165 const APInt &AndConst = cast<ConstantSDNode>(AndOp1)->getAPIntValue();
6167 if (AndConst.isPowerOf2() &&
6168 cast<ConstantSDNode>(Op1)->getAPIntValue()==AndConst.logBase2()) {
6170 DAG.getSetCC(N->getDebugLoc(),
6171 TLI.getSetCCResultType(Op0.getValueType()),
6172 Op0, DAG.getConstant(0, Op0.getValueType()),
6175 SDValue NewBRCond = DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
6176 MVT::Other, Chain, SetCC, N2);
6177 // Don't add the new BRCond into the worklist or else SimplifySelectCC
6178 // will convert it back to (X & C1) >> C2.
6179 CombineTo(N, NewBRCond, false);
6180 // Truncate is dead.
6182 removeFromWorkList(Trunc);
6183 DAG.DeleteNode(Trunc);
6185 // Replace the uses of SRL with SETCC
6186 WorkListRemover DeadNodes(*this);
6187 DAG.ReplaceAllUsesOfValueWith(N1, SetCC, &DeadNodes);
6188 removeFromWorkList(N1.getNode());
6189 DAG.DeleteNode(N1.getNode());
6190 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6196 // Restore N1 if the above transformation doesn't match.
6197 N1 = N->getOperand(1);
6200 // Transform br(xor(x, y)) -> br(x != y)
6201 // Transform br(xor(xor(x,y), 1)) -> br (x == y)
6202 if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) {
6203 SDNode *TheXor = N1.getNode();
6204 SDValue Op0 = TheXor->getOperand(0);
6205 SDValue Op1 = TheXor->getOperand(1);
6206 if (Op0.getOpcode() == Op1.getOpcode()) {
6207 // Avoid missing important xor optimizations.
6208 SDValue Tmp = visitXOR(TheXor);
6209 if (Tmp.getNode() && Tmp.getNode() != TheXor) {
6210 DEBUG(dbgs() << "\nReplacing.8 ";
6212 dbgs() << "\nWith: ";
6213 Tmp.getNode()->dump(&DAG);
6215 WorkListRemover DeadNodes(*this);
6216 DAG.ReplaceAllUsesOfValueWith(N1, Tmp, &DeadNodes);
6217 removeFromWorkList(TheXor);
6218 DAG.DeleteNode(TheXor);
6219 return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
6220 MVT::Other, Chain, Tmp, N2);
6224 if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) {
6226 if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0))
6227 if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() &&
6228 Op0.getOpcode() == ISD::XOR) {
6229 TheXor = Op0.getNode();
6233 EVT SetCCVT = N1.getValueType();
6235 SetCCVT = TLI.getSetCCResultType(SetCCVT);
6236 SDValue SetCC = DAG.getSetCC(TheXor->getDebugLoc(),
6239 Equal ? ISD::SETEQ : ISD::SETNE);
6240 // Replace the uses of XOR with SETCC
6241 WorkListRemover DeadNodes(*this);
6242 DAG.ReplaceAllUsesOfValueWith(N1, SetCC, &DeadNodes);
6243 removeFromWorkList(N1.getNode());
6244 DAG.DeleteNode(N1.getNode());
6245 return DAG.getNode(ISD::BRCOND, N->getDebugLoc(),
6246 MVT::Other, Chain, SetCC, N2);
6253 // Operand List for BR_CC: Chain, CondCC, CondLHS, CondRHS, DestBB.
6255 SDValue DAGCombiner::visitBR_CC(SDNode *N) {
6256 CondCodeSDNode *CC = cast<CondCodeSDNode>(N->getOperand(1));
6257 SDValue CondLHS = N->getOperand(2), CondRHS = N->getOperand(3);
6259 // If N is a constant we could fold this into a fallthrough or unconditional
6260 // branch. However that doesn't happen very often in normal code, because
6261 // Instcombine/SimplifyCFG should have handled the available opportunities.
6262 // If we did this folding here, it would be necessary to update the
6263 // MachineBasicBlock CFG, which is awkward.
6265 // Use SimplifySetCC to simplify SETCC's.
6266 SDValue Simp = SimplifySetCC(TLI.getSetCCResultType(CondLHS.getValueType()),
6267 CondLHS, CondRHS, CC->get(), N->getDebugLoc(),
6269 if (Simp.getNode()) AddToWorkList(Simp.getNode());
6271 // fold to a simpler setcc
6272 if (Simp.getNode() && Simp.getOpcode() == ISD::SETCC)
6273 return DAG.getNode(ISD::BR_CC, N->getDebugLoc(), MVT::Other,
6274 N->getOperand(0), Simp.getOperand(2),
6275 Simp.getOperand(0), Simp.getOperand(1),
6281 /// canFoldInAddressingMode - Return true if 'Use' is a load or a store that
6282 /// uses N as its base pointer and that N may be folded in the load / store
6283 /// addressing mode.
6284 static bool canFoldInAddressingMode(SDNode *N, SDNode *Use,
6286 const TargetLowering &TLI) {
6288 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Use)) {
6289 if (LD->isIndexed() || LD->getBasePtr().getNode() != N)
6291 VT = Use->getValueType(0);
6292 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Use)) {
6293 if (ST->isIndexed() || ST->getBasePtr().getNode() != N)
6295 VT = ST->getValue().getValueType();
6299 TargetLowering::AddrMode AM;
6300 if (N->getOpcode() == ISD::ADD) {
6301 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
6304 AM.BaseOffs = Offset->getSExtValue();
6308 } else if (N->getOpcode() == ISD::SUB) {
6309 ConstantSDNode *Offset = dyn_cast<ConstantSDNode>(N->getOperand(1));
6312 AM.BaseOffs = -Offset->getSExtValue();
6319 return TLI.isLegalAddressingMode(AM, VT.getTypeForEVT(*DAG.getContext()));
6322 /// CombineToPreIndexedLoadStore - Try turning a load / store into a
6323 /// pre-indexed load / store when the base pointer is an add or subtract
6324 /// and it has other uses besides the load / store. After the
6325 /// transformation, the new indexed load / store has effectively folded
6326 /// the add / subtract in and all of its other uses are redirected to the
6327 /// new load / store.
6328 bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
6329 if (Level < AfterLegalizeDAG)
6335 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6336 if (LD->isIndexed())
6338 VT = LD->getMemoryVT();
6339 if (!TLI.isIndexedLoadLegal(ISD::PRE_INC, VT) &&
6340 !TLI.isIndexedLoadLegal(ISD::PRE_DEC, VT))
6342 Ptr = LD->getBasePtr();
6343 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6344 if (ST->isIndexed())
6346 VT = ST->getMemoryVT();
6347 if (!TLI.isIndexedStoreLegal(ISD::PRE_INC, VT) &&
6348 !TLI.isIndexedStoreLegal(ISD::PRE_DEC, VT))
6350 Ptr = ST->getBasePtr();
6356 // If the pointer is not an add/sub, or if it doesn't have multiple uses, bail
6357 // out. There is no reason to make this a preinc/predec.
6358 if ((Ptr.getOpcode() != ISD::ADD && Ptr.getOpcode() != ISD::SUB) ||
6359 Ptr.getNode()->hasOneUse())
6362 // Ask the target to do addressing mode selection.
6365 ISD::MemIndexedMode AM = ISD::UNINDEXED;
6366 if (!TLI.getPreIndexedAddressParts(N, BasePtr, Offset, AM, DAG))
6368 // Don't create a indexed load / store with zero offset.
6369 if (isa<ConstantSDNode>(Offset) &&
6370 cast<ConstantSDNode>(Offset)->isNullValue())
6373 // Try turning it into a pre-indexed load / store except when:
6374 // 1) The new base ptr is a frame index.
6375 // 2) If N is a store and the new base ptr is either the same as or is a
6376 // predecessor of the value being stored.
6377 // 3) Another use of old base ptr is a predecessor of N. If ptr is folded
6378 // that would create a cycle.
6379 // 4) All uses are load / store ops that use it as old base ptr.
6381 // Check #1. Preinc'ing a frame index would require copying the stack pointer
6382 // (plus the implicit offset) to a register to preinc anyway.
6383 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
6388 SDValue Val = cast<StoreSDNode>(N)->getValue();
6389 if (Val == BasePtr || BasePtr.getNode()->isPredecessorOf(Val.getNode()))
6393 // Now check for #3 and #4.
6394 bool RealUse = false;
6396 // Caches for hasPredecessorHelper
6397 SmallPtrSet<const SDNode *, 32> Visited;
6398 SmallVector<const SDNode *, 16> Worklist;
6400 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
6401 E = Ptr.getNode()->use_end(); I != E; ++I) {
6405 if (N->hasPredecessorHelper(Use, Visited, Worklist))
6408 // If Ptr may be folded in addressing mode of other use, then it's
6409 // not profitable to do this transformation.
6410 if (!canFoldInAddressingMode(Ptr.getNode(), Use, DAG, TLI))
6419 Result = DAG.getIndexedLoad(SDValue(N,0), N->getDebugLoc(),
6420 BasePtr, Offset, AM);
6422 Result = DAG.getIndexedStore(SDValue(N,0), N->getDebugLoc(),
6423 BasePtr, Offset, AM);
6426 DEBUG(dbgs() << "\nReplacing.4 ";
6428 dbgs() << "\nWith: ";
6429 Result.getNode()->dump(&DAG);
6431 WorkListRemover DeadNodes(*this);
6433 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0),
6435 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2),
6438 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1),
6442 // Finally, since the node is now dead, remove it from the graph.
6445 // Replace the uses of Ptr with uses of the updated base value.
6446 DAG.ReplaceAllUsesOfValueWith(Ptr, Result.getValue(isLoad ? 1 : 0),
6448 removeFromWorkList(Ptr.getNode());
6449 DAG.DeleteNode(Ptr.getNode());
6454 /// CombineToPostIndexedLoadStore - Try to combine a load / store with a
6455 /// add / sub of the base pointer node into a post-indexed load / store.
6456 /// The transformation folded the add / subtract into the new indexed
6457 /// load / store effectively and all of its uses are redirected to the
6458 /// new load / store.
6459 bool DAGCombiner::CombineToPostIndexedLoadStore(SDNode *N) {
6460 if (Level < AfterLegalizeDAG)
6466 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
6467 if (LD->isIndexed())
6469 VT = LD->getMemoryVT();
6470 if (!TLI.isIndexedLoadLegal(ISD::POST_INC, VT) &&
6471 !TLI.isIndexedLoadLegal(ISD::POST_DEC, VT))
6473 Ptr = LD->getBasePtr();
6474 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
6475 if (ST->isIndexed())
6477 VT = ST->getMemoryVT();
6478 if (!TLI.isIndexedStoreLegal(ISD::POST_INC, VT) &&
6479 !TLI.isIndexedStoreLegal(ISD::POST_DEC, VT))
6481 Ptr = ST->getBasePtr();
6487 if (Ptr.getNode()->hasOneUse())
6490 for (SDNode::use_iterator I = Ptr.getNode()->use_begin(),
6491 E = Ptr.getNode()->use_end(); I != E; ++I) {
6494 (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB))
6499 ISD::MemIndexedMode AM = ISD::UNINDEXED;
6500 if (TLI.getPostIndexedAddressParts(N, Op, BasePtr, Offset, AM, DAG)) {
6501 // Don't create a indexed load / store with zero offset.
6502 if (isa<ConstantSDNode>(Offset) &&
6503 cast<ConstantSDNode>(Offset)->isNullValue())
6506 // Try turning it into a post-indexed load / store except when
6507 // 1) All uses are load / store ops that use it as base ptr (and
6508 // it may be folded as addressing mmode).
6509 // 2) Op must be independent of N, i.e. Op is neither a predecessor
6510 // nor a successor of N. Otherwise, if Op is folded that would
6513 if (isa<FrameIndexSDNode>(BasePtr) || isa<RegisterSDNode>(BasePtr))
6517 bool TryNext = false;
6518 for (SDNode::use_iterator II = BasePtr.getNode()->use_begin(),
6519 EE = BasePtr.getNode()->use_end(); II != EE; ++II) {
6521 if (Use == Ptr.getNode())
6524 // If all the uses are load / store addresses, then don't do the
6526 if (Use->getOpcode() == ISD::ADD || Use->getOpcode() == ISD::SUB){
6527 bool RealUse = false;
6528 for (SDNode::use_iterator III = Use->use_begin(),
6529 EEE = Use->use_end(); III != EEE; ++III) {
6530 SDNode *UseUse = *III;
6531 if (!canFoldInAddressingMode(Use, UseUse, DAG, TLI))
6546 if (!Op->isPredecessorOf(N) && !N->isPredecessorOf(Op)) {
6547 SDValue Result = isLoad
6548 ? DAG.getIndexedLoad(SDValue(N,0), N->getDebugLoc(),
6549 BasePtr, Offset, AM)
6550 : DAG.getIndexedStore(SDValue(N,0), N->getDebugLoc(),
6551 BasePtr, Offset, AM);
6554 DEBUG(dbgs() << "\nReplacing.5 ";
6556 dbgs() << "\nWith: ";
6557 Result.getNode()->dump(&DAG);
6559 WorkListRemover DeadNodes(*this);
6561 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(0),
6563 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Result.getValue(2),
6566 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Result.getValue(1),
6570 // Finally, since the node is now dead, remove it from the graph.
6573 // Replace the uses of Use with uses of the updated base value.
6574 DAG.ReplaceAllUsesOfValueWith(SDValue(Op, 0),
6575 Result.getValue(isLoad ? 1 : 0),
6577 removeFromWorkList(Op);
6587 SDValue DAGCombiner::visitLOAD(SDNode *N) {
6588 LoadSDNode *LD = cast<LoadSDNode>(N);
6589 SDValue Chain = LD->getChain();
6590 SDValue Ptr = LD->getBasePtr();
6592 // If load is not volatile and there are no uses of the loaded value (and
6593 // the updated indexed value in case of indexed loads), change uses of the
6594 // chain value into uses of the chain input (i.e. delete the dead load).
6595 if (!LD->isVolatile()) {
6596 if (N->getValueType(1) == MVT::Other) {
6598 if (!N->hasAnyUseOfValue(0)) {
6599 // It's not safe to use the two value CombineTo variant here. e.g.
6600 // v1, chain2 = load chain1, loc
6601 // v2, chain3 = load chain2, loc
6603 // Now we replace use of chain2 with chain1. This makes the second load
6604 // isomorphic to the one we are deleting, and thus makes this load live.
6605 DEBUG(dbgs() << "\nReplacing.6 ";
6607 dbgs() << "\nWith chain: ";
6608 Chain.getNode()->dump(&DAG);
6610 WorkListRemover DeadNodes(*this);
6611 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), Chain, &DeadNodes);
6613 if (N->use_empty()) {
6614 removeFromWorkList(N);
6618 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6622 assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?");
6623 if (!N->hasAnyUseOfValue(0) && !N->hasAnyUseOfValue(1)) {
6624 SDValue Undef = DAG.getUNDEF(N->getValueType(0));
6625 DEBUG(dbgs() << "\nReplacing.7 ";
6627 dbgs() << "\nWith: ";
6628 Undef.getNode()->dump(&DAG);
6629 dbgs() << " and 2 other values\n");
6630 WorkListRemover DeadNodes(*this);
6631 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Undef, &DeadNodes);
6632 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1),
6633 DAG.getUNDEF(N->getValueType(1)),
6635 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 2), Chain, &DeadNodes);
6636 removeFromWorkList(N);
6638 return SDValue(N, 0); // Return N so it doesn't get rechecked!
6643 // If this load is directly stored, replace the load value with the stored
6645 // TODO: Handle store large -> read small portion.
6646 // TODO: Handle TRUNCSTORE/LOADEXT
6647 if (ISD::isNormalLoad(N) && !LD->isVolatile()) {
6648 if (ISD::isNON_TRUNCStore(Chain.getNode())) {
6649 StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
6650 if (PrevST->getBasePtr() == Ptr &&
6651 PrevST->getValue().getValueType() == N->getValueType(0))
6652 return CombineTo(N, Chain.getOperand(1), Chain);
6656 // Try to infer better alignment information than the load already has.
6657 if (OptLevel != CodeGenOpt::None && LD->isUnindexed()) {
6658 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
6659 if (Align > LD->getAlignment())
6660 return DAG.getExtLoad(LD->getExtensionType(), N->getDebugLoc(),
6661 LD->getValueType(0),
6662 Chain, Ptr, LD->getPointerInfo(),
6664 LD->isVolatile(), LD->isNonTemporal(), Align);
6669 // Walk up chain skipping non-aliasing memory nodes.
6670 SDValue BetterChain = FindBetterChain(N, Chain);
6672 // If there is a better chain.
6673 if (Chain != BetterChain) {
6676 // Replace the chain to void dependency.
6677 if (LD->getExtensionType() == ISD::NON_EXTLOAD) {
6678 ReplLoad = DAG.getLoad(N->getValueType(0), LD->getDebugLoc(),
6679 BetterChain, Ptr, LD->getPointerInfo(),
6680 LD->isVolatile(), LD->isNonTemporal(),
6681 LD->isInvariant(), LD->getAlignment());
6683 ReplLoad = DAG.getExtLoad(LD->getExtensionType(), LD->getDebugLoc(),
6684 LD->getValueType(0),
6685 BetterChain, Ptr, LD->getPointerInfo(),
6688 LD->isNonTemporal(),
6689 LD->getAlignment());
6692 // Create token factor to keep old chain connected.
6693 SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
6694 MVT::Other, Chain, ReplLoad.getValue(1));
6696 // Make sure the new and old chains are cleaned up.
6697 AddToWorkList(Token.getNode());
6699 // Replace uses with load result and token factor. Don't add users
6701 return CombineTo(N, ReplLoad.getValue(0), Token, false);
6705 // Try transforming N to an indexed load.
6706 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
6707 return SDValue(N, 0);
6712 /// CheckForMaskedLoad - Check to see if V is (and load (ptr), imm), where the
6713 /// load is having specific bytes cleared out. If so, return the byte size
6714 /// being masked out and the shift amount.
6715 static std::pair<unsigned, unsigned>
6716 CheckForMaskedLoad(SDValue V, SDValue Ptr, SDValue Chain) {
6717 std::pair<unsigned, unsigned> Result(0, 0);
6719 // Check for the structure we're looking for.
6720 if (V->getOpcode() != ISD::AND ||
6721 !isa<ConstantSDNode>(V->getOperand(1)) ||
6722 !ISD::isNormalLoad(V->getOperand(0).getNode()))
6725 // Check the chain and pointer.
6726 LoadSDNode *LD = cast<LoadSDNode>(V->getOperand(0));
6727 if (LD->getBasePtr() != Ptr) return Result; // Not from same pointer.
6729 // The store should be chained directly to the load or be an operand of a
6731 if (LD == Chain.getNode())
6733 else if (Chain->getOpcode() != ISD::TokenFactor)
6734 return Result; // Fail.
6737 for (unsigned i = 0, e = Chain->getNumOperands(); i != e; ++i)
6738 if (Chain->getOperand(i).getNode() == LD) {
6742 if (!isOk) return Result;
6745 // This only handles simple types.
6746 if (V.getValueType() != MVT::i16 &&
6747 V.getValueType() != MVT::i32 &&
6748 V.getValueType() != MVT::i64)
6751 // Check the constant mask. Invert it so that the bits being masked out are
6752 // 0 and the bits being kept are 1. Use getSExtValue so that leading bits
6753 // follow the sign bit for uniformity.
6754 uint64_t NotMask = ~cast<ConstantSDNode>(V->getOperand(1))->getSExtValue();
6755 unsigned NotMaskLZ = CountLeadingZeros_64(NotMask);
6756 if (NotMaskLZ & 7) return Result; // Must be multiple of a byte.
6757 unsigned NotMaskTZ = CountTrailingZeros_64(NotMask);
6758 if (NotMaskTZ & 7) return Result; // Must be multiple of a byte.
6759 if (NotMaskLZ == 64) return Result; // All zero mask.
6761 // See if we have a continuous run of bits. If so, we have 0*1+0*
6762 if (CountTrailingOnes_64(NotMask >> NotMaskTZ)+NotMaskTZ+NotMaskLZ != 64)
6765 // Adjust NotMaskLZ down to be from the actual size of the int instead of i64.
6766 if (V.getValueType() != MVT::i64 && NotMaskLZ)
6767 NotMaskLZ -= 64-V.getValueSizeInBits();
6769 unsigned MaskedBytes = (V.getValueSizeInBits()-NotMaskLZ-NotMaskTZ)/8;
6770 switch (MaskedBytes) {
6774 default: return Result; // All one mask, or 5-byte mask.
6777 // Verify that the first bit starts at a multiple of mask so that the access
6778 // is aligned the same as the access width.
6779 if (NotMaskTZ && NotMaskTZ/8 % MaskedBytes) return Result;
6781 Result.first = MaskedBytes;
6782 Result.second = NotMaskTZ/8;
6787 /// ShrinkLoadReplaceStoreWithStore - Check to see if IVal is something that
6788 /// provides a value as specified by MaskInfo. If so, replace the specified
6789 /// store with a narrower store of truncated IVal.
6791 ShrinkLoadReplaceStoreWithStore(const std::pair<unsigned, unsigned> &MaskInfo,
6792 SDValue IVal, StoreSDNode *St,
6794 unsigned NumBytes = MaskInfo.first;
6795 unsigned ByteShift = MaskInfo.second;
6796 SelectionDAG &DAG = DC->getDAG();
6798 // Check to see if IVal is all zeros in the part being masked in by the 'or'
6799 // that uses this. If not, this is not a replacement.
6800 APInt Mask = ~APInt::getBitsSet(IVal.getValueSizeInBits(),
6801 ByteShift*8, (ByteShift+NumBytes)*8);
6802 if (!DAG.MaskedValueIsZero(IVal, Mask)) return 0;
6804 // Check that it is legal on the target to do this. It is legal if the new
6805 // VT we're shrinking to (i8/i16/i32) is legal or we're still before type
6807 MVT VT = MVT::getIntegerVT(NumBytes*8);
6808 if (!DC->isTypeLegal(VT))
6811 // Okay, we can do this! Replace the 'St' store with a store of IVal that is
6812 // shifted by ByteShift and truncated down to NumBytes.
6814 IVal = DAG.getNode(ISD::SRL, IVal->getDebugLoc(), IVal.getValueType(), IVal,
6815 DAG.getConstant(ByteShift*8,
6816 DC->getShiftAmountTy(IVal.getValueType())));
6818 // Figure out the offset for the store and the alignment of the access.
6820 unsigned NewAlign = St->getAlignment();
6822 if (DAG.getTargetLoweringInfo().isLittleEndian())
6823 StOffset = ByteShift;
6825 StOffset = IVal.getValueType().getStoreSize() - ByteShift - NumBytes;
6827 SDValue Ptr = St->getBasePtr();
6829 Ptr = DAG.getNode(ISD::ADD, IVal->getDebugLoc(), Ptr.getValueType(),
6830 Ptr, DAG.getConstant(StOffset, Ptr.getValueType()));
6831 NewAlign = MinAlign(NewAlign, StOffset);
6834 // Truncate down to the new size.
6835 IVal = DAG.getNode(ISD::TRUNCATE, IVal->getDebugLoc(), VT, IVal);
6838 return DAG.getStore(St->getChain(), St->getDebugLoc(), IVal, Ptr,
6839 St->getPointerInfo().getWithOffset(StOffset),
6840 false, false, NewAlign).getNode();
6844 /// ReduceLoadOpStoreWidth - Look for sequence of load / op / store where op is
6845 /// one of 'or', 'xor', and 'and' of immediates. If 'op' is only touching some
6846 /// of the loaded bits, try narrowing the load and store if it would end up
6847 /// being a win for performance or code size.
6848 SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) {
6849 StoreSDNode *ST = cast<StoreSDNode>(N);
6850 if (ST->isVolatile())
6853 SDValue Chain = ST->getChain();
6854 SDValue Value = ST->getValue();
6855 SDValue Ptr = ST->getBasePtr();
6856 EVT VT = Value.getValueType();
6858 if (ST->isTruncatingStore() || VT.isVector() || !Value.hasOneUse())
6861 unsigned Opc = Value.getOpcode();
6863 // If this is "store (or X, Y), P" and X is "(and (load P), cst)", where cst
6864 // is a byte mask indicating a consecutive number of bytes, check to see if
6865 // Y is known to provide just those bytes. If so, we try to replace the
6866 // load + replace + store sequence with a single (narrower) store, which makes
6868 if (Opc == ISD::OR) {
6869 std::pair<unsigned, unsigned> MaskedLoad;
6870 MaskedLoad = CheckForMaskedLoad(Value.getOperand(0), Ptr, Chain);
6871 if (MaskedLoad.first)
6872 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
6873 Value.getOperand(1), ST,this))
6874 return SDValue(NewST, 0);
6876 // Or is commutative, so try swapping X and Y.
6877 MaskedLoad = CheckForMaskedLoad(Value.getOperand(1), Ptr, Chain);
6878 if (MaskedLoad.first)
6879 if (SDNode *NewST = ShrinkLoadReplaceStoreWithStore(MaskedLoad,
6880 Value.getOperand(0), ST,this))
6881 return SDValue(NewST, 0);
6884 if ((Opc != ISD::OR && Opc != ISD::XOR && Opc != ISD::AND) ||
6885 Value.getOperand(1).getOpcode() != ISD::Constant)
6888 SDValue N0 = Value.getOperand(0);
6889 if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
6890 Chain == SDValue(N0.getNode(), 1)) {
6891 LoadSDNode *LD = cast<LoadSDNode>(N0);
6892 if (LD->getBasePtr() != Ptr ||
6893 LD->getPointerInfo().getAddrSpace() !=
6894 ST->getPointerInfo().getAddrSpace())
6897 // Find the type to narrow it the load / op / store to.
6898 SDValue N1 = Value.getOperand(1);
6899 unsigned BitWidth = N1.getValueSizeInBits();
6900 APInt Imm = cast<ConstantSDNode>(N1)->getAPIntValue();
6901 if (Opc == ISD::AND)
6902 Imm ^= APInt::getAllOnesValue(BitWidth);
6903 if (Imm == 0 || Imm.isAllOnesValue())
6905 unsigned ShAmt = Imm.countTrailingZeros();
6906 unsigned MSB = BitWidth - Imm.countLeadingZeros() - 1;
6907 unsigned NewBW = NextPowerOf2(MSB - ShAmt);
6908 EVT NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
6909 while (NewBW < BitWidth &&
6910 !(TLI.isOperationLegalOrCustom(Opc, NewVT) &&
6911 TLI.isNarrowingProfitable(VT, NewVT))) {
6912 NewBW = NextPowerOf2(NewBW);
6913 NewVT = EVT::getIntegerVT(*DAG.getContext(), NewBW);
6915 if (NewBW >= BitWidth)
6918 // If the lsb changed does not start at the type bitwidth boundary,
6919 // start at the previous one.
6921 ShAmt = (((ShAmt + NewBW - 1) / NewBW) * NewBW) - NewBW;
6922 APInt Mask = APInt::getBitsSet(BitWidth, ShAmt, ShAmt + NewBW);
6923 if ((Imm & Mask) == Imm) {
6924 APInt NewImm = (Imm & Mask).lshr(ShAmt).trunc(NewBW);
6925 if (Opc == ISD::AND)
6926 NewImm ^= APInt::getAllOnesValue(NewBW);
6927 uint64_t PtrOff = ShAmt / 8;
6928 // For big endian targets, we need to adjust the offset to the pointer to
6929 // load the correct bytes.
6930 if (TLI.isBigEndian())
6931 PtrOff = (BitWidth + 7 - NewBW) / 8 - PtrOff;
6933 unsigned NewAlign = MinAlign(LD->getAlignment(), PtrOff);
6934 Type *NewVTTy = NewVT.getTypeForEVT(*DAG.getContext());
6935 if (NewAlign < TLI.getTargetData()->getABITypeAlignment(NewVTTy))
6938 SDValue NewPtr = DAG.getNode(ISD::ADD, LD->getDebugLoc(),
6939 Ptr.getValueType(), Ptr,
6940 DAG.getConstant(PtrOff, Ptr.getValueType()));
6941 SDValue NewLD = DAG.getLoad(NewVT, N0.getDebugLoc(),
6942 LD->getChain(), NewPtr,
6943 LD->getPointerInfo().getWithOffset(PtrOff),
6944 LD->isVolatile(), LD->isNonTemporal(),
6945 LD->isInvariant(), NewAlign);
6946 SDValue NewVal = DAG.getNode(Opc, Value.getDebugLoc(), NewVT, NewLD,
6947 DAG.getConstant(NewImm, NewVT));
6948 SDValue NewST = DAG.getStore(Chain, N->getDebugLoc(),
6950 ST->getPointerInfo().getWithOffset(PtrOff),
6951 false, false, NewAlign);
6953 AddToWorkList(NewPtr.getNode());
6954 AddToWorkList(NewLD.getNode());
6955 AddToWorkList(NewVal.getNode());
6956 WorkListRemover DeadNodes(*this);
6957 DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), NewLD.getValue(1),
6967 /// TransformFPLoadStorePair - For a given floating point load / store pair,
6968 /// if the load value isn't used by any other operations, then consider
6969 /// transforming the pair to integer load / store operations if the target
6970 /// deems the transformation profitable.
6971 SDValue DAGCombiner::TransformFPLoadStorePair(SDNode *N) {
6972 StoreSDNode *ST = cast<StoreSDNode>(N);
6973 SDValue Chain = ST->getChain();
6974 SDValue Value = ST->getValue();
6975 if (ISD::isNormalStore(ST) && ISD::isNormalLoad(Value.getNode()) &&
6976 Value.hasOneUse() &&
6977 Chain == SDValue(Value.getNode(), 1)) {
6978 LoadSDNode *LD = cast<LoadSDNode>(Value);
6979 EVT VT = LD->getMemoryVT();
6980 if (!VT.isFloatingPoint() ||
6981 VT != ST->getMemoryVT() ||
6982 LD->isNonTemporal() ||
6983 ST->isNonTemporal() ||
6984 LD->getPointerInfo().getAddrSpace() != 0 ||
6985 ST->getPointerInfo().getAddrSpace() != 0)
6988 EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), VT.getSizeInBits());
6989 if (!TLI.isOperationLegal(ISD::LOAD, IntVT) ||
6990 !TLI.isOperationLegal(ISD::STORE, IntVT) ||
6991 !TLI.isDesirableToTransformToIntegerOp(ISD::LOAD, VT) ||
6992 !TLI.isDesirableToTransformToIntegerOp(ISD::STORE, VT))
6995 unsigned LDAlign = LD->getAlignment();
6996 unsigned STAlign = ST->getAlignment();
6997 Type *IntVTTy = IntVT.getTypeForEVT(*DAG.getContext());
6998 unsigned ABIAlign = TLI.getTargetData()->getABITypeAlignment(IntVTTy);
6999 if (LDAlign < ABIAlign || STAlign < ABIAlign)
7002 SDValue NewLD = DAG.getLoad(IntVT, Value.getDebugLoc(),
7003 LD->getChain(), LD->getBasePtr(),
7004 LD->getPointerInfo(),
7005 false, false, false, LDAlign);
7007 SDValue NewST = DAG.getStore(NewLD.getValue(1), N->getDebugLoc(),
7008 NewLD, ST->getBasePtr(),
7009 ST->getPointerInfo(),
7010 false, false, STAlign);
7012 AddToWorkList(NewLD.getNode());
7013 AddToWorkList(NewST.getNode());
7014 WorkListRemover DeadNodes(*this);
7015 DAG.ReplaceAllUsesOfValueWith(Value.getValue(1), NewLD.getValue(1),
7024 SDValue DAGCombiner::visitSTORE(SDNode *N) {
7025 StoreSDNode *ST = cast<StoreSDNode>(N);
7026 SDValue Chain = ST->getChain();
7027 SDValue Value = ST->getValue();
7028 SDValue Ptr = ST->getBasePtr();
7030 // If this is a store of a bit convert, store the input value if the
7031 // resultant store does not need a higher alignment than the original.
7032 if (Value.getOpcode() == ISD::BITCAST && !ST->isTruncatingStore() &&
7033 ST->isUnindexed()) {
7034 unsigned OrigAlign = ST->getAlignment();
7035 EVT SVT = Value.getOperand(0).getValueType();
7036 unsigned Align = TLI.getTargetData()->
7037 getABITypeAlignment(SVT.getTypeForEVT(*DAG.getContext()));
7038 if (Align <= OrigAlign &&
7039 ((!LegalOperations && !ST->isVolatile()) ||
7040 TLI.isOperationLegalOrCustom(ISD::STORE, SVT)))
7041 return DAG.getStore(Chain, N->getDebugLoc(), Value.getOperand(0),
7042 Ptr, ST->getPointerInfo(), ST->isVolatile(),
7043 ST->isNonTemporal(), OrigAlign);
7046 // Turn 'store undef, Ptr' -> nothing.
7047 if (Value.getOpcode() == ISD::UNDEF && ST->isUnindexed())
7050 // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr'
7051 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Value)) {
7052 // NOTE: If the original store is volatile, this transform must not increase
7053 // the number of stores. For example, on x86-32 an f64 can be stored in one
7054 // processor operation but an i64 (which is not legal) requires two. So the
7055 // transform should not be done in this case.
7056 if (Value.getOpcode() != ISD::TargetConstantFP) {
7058 switch (CFP->getValueType(0).getSimpleVT().SimpleTy) {
7059 default: llvm_unreachable("Unknown FP type");
7060 case MVT::f80: // We don't do this for these yet.
7065 if ((isTypeLegal(MVT::i32) && !LegalOperations && !ST->isVolatile()) ||
7066 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
7067 Tmp = DAG.getConstant((uint32_t)CFP->getValueAPF().
7068 bitcastToAPInt().getZExtValue(), MVT::i32);
7069 return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
7070 Ptr, ST->getPointerInfo(), ST->isVolatile(),
7071 ST->isNonTemporal(), ST->getAlignment());
7075 if ((TLI.isTypeLegal(MVT::i64) && !LegalOperations &&
7076 !ST->isVolatile()) ||
7077 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i64)) {
7078 Tmp = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt().
7079 getZExtValue(), MVT::i64);
7080 return DAG.getStore(Chain, N->getDebugLoc(), Tmp,
7081 Ptr, ST->getPointerInfo(), ST->isVolatile(),
7082 ST->isNonTemporal(), ST->getAlignment());
7085 if (!ST->isVolatile() &&
7086 TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) {
7087 // Many FP stores are not made apparent until after legalize, e.g. for
7088 // argument passing. Since this is so common, custom legalize the
7089 // 64-bit integer store into two 32-bit stores.
7090 uint64_t Val = CFP->getValueAPF().bitcastToAPInt().getZExtValue();
7091 SDValue Lo = DAG.getConstant(Val & 0xFFFFFFFF, MVT::i32);
7092 SDValue Hi = DAG.getConstant(Val >> 32, MVT::i32);
7093 if (TLI.isBigEndian()) std::swap(Lo, Hi);
7095 unsigned Alignment = ST->getAlignment();
7096 bool isVolatile = ST->isVolatile();
7097 bool isNonTemporal = ST->isNonTemporal();
7099 SDValue St0 = DAG.getStore(Chain, ST->getDebugLoc(), Lo,
7100 Ptr, ST->getPointerInfo(),
7101 isVolatile, isNonTemporal,
7102 ST->getAlignment());
7103 Ptr = DAG.getNode(ISD::ADD, N->getDebugLoc(), Ptr.getValueType(), Ptr,
7104 DAG.getConstant(4, Ptr.getValueType()));
7105 Alignment = MinAlign(Alignment, 4U);
7106 SDValue St1 = DAG.getStore(Chain, ST->getDebugLoc(), Hi,
7107 Ptr, ST->getPointerInfo().getWithOffset(4),
7108 isVolatile, isNonTemporal,
7110 return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
7119 // Try to infer better alignment information than the store already has.
7120 if (OptLevel != CodeGenOpt::None && ST->isUnindexed()) {
7121 if (unsigned Align = DAG.InferPtrAlignment(Ptr)) {
7122 if (Align > ST->getAlignment())
7123 return DAG.getTruncStore(Chain, N->getDebugLoc(), Value,
7124 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
7125 ST->isVolatile(), ST->isNonTemporal(), Align);
7129 // Try transforming a pair floating point load / store ops to integer
7130 // load / store ops.
7131 SDValue NewST = TransformFPLoadStorePair(N);
7132 if (NewST.getNode())
7136 // Walk up chain skipping non-aliasing memory nodes.
7137 SDValue BetterChain = FindBetterChain(N, Chain);
7139 // If there is a better chain.
7140 if (Chain != BetterChain) {
7143 // Replace the chain to avoid dependency.
7144 if (ST->isTruncatingStore()) {
7145 ReplStore = DAG.getTruncStore(BetterChain, N->getDebugLoc(), Value, Ptr,
7146 ST->getPointerInfo(),
7147 ST->getMemoryVT(), ST->isVolatile(),
7148 ST->isNonTemporal(), ST->getAlignment());
7150 ReplStore = DAG.getStore(BetterChain, N->getDebugLoc(), Value, Ptr,
7151 ST->getPointerInfo(),
7152 ST->isVolatile(), ST->isNonTemporal(),
7153 ST->getAlignment());
7156 // Create token to keep both nodes around.
7157 SDValue Token = DAG.getNode(ISD::TokenFactor, N->getDebugLoc(),
7158 MVT::Other, Chain, ReplStore);
7160 // Make sure the new and old chains are cleaned up.
7161 AddToWorkList(Token.getNode());
7163 // Don't add users to work list.
7164 return CombineTo(N, Token, false);
7168 // Try transforming N to an indexed store.
7169 if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
7170 return SDValue(N, 0);
7172 // FIXME: is there such a thing as a truncating indexed store?
7173 if (ST->isTruncatingStore() && ST->isUnindexed() &&
7174 Value.getValueType().isInteger()) {
7175 // See if we can simplify the input to this truncstore with knowledge that
7176 // only the low bits are being used. For example:
7177 // "truncstore (or (shl x, 8), y), i8" -> "truncstore y, i8"
7179 GetDemandedBits(Value,
7180 APInt::getLowBitsSet(
7181 Value.getValueType().getScalarType().getSizeInBits(),
7182 ST->getMemoryVT().getScalarType().getSizeInBits()));
7183 AddToWorkList(Value.getNode());
7184 if (Shorter.getNode())
7185 return DAG.getTruncStore(Chain, N->getDebugLoc(), Shorter,
7186 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
7187 ST->isVolatile(), ST->isNonTemporal(),
7188 ST->getAlignment());
7190 // Otherwise, see if we can simplify the operation with
7191 // SimplifyDemandedBits, which only works if the value has a single use.
7192 if (SimplifyDemandedBits(Value,
7193 APInt::getLowBitsSet(
7194 Value.getValueType().getScalarType().getSizeInBits(),
7195 ST->getMemoryVT().getScalarType().getSizeInBits())))
7196 return SDValue(N, 0);
7199 // If this is a load followed by a store to the same location, then the store
7201 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Value)) {
7202 if (Ld->getBasePtr() == Ptr && ST->getMemoryVT() == Ld->getMemoryVT() &&
7203 ST->isUnindexed() && !ST->isVolatile() &&
7204 // There can't be any side effects between the load and store, such as
7206 Chain.reachesChainWithoutSideEffects(SDValue(Ld, 1))) {
7207 // The store is dead, remove it.
7212 // If this is an FP_ROUND or TRUNC followed by a store, fold this into a
7213 // truncating store. We can do this even if this is already a truncstore.
7214 if ((Value.getOpcode() == ISD::FP_ROUND || Value.getOpcode() == ISD::TRUNCATE)
7215 && Value.getNode()->hasOneUse() && ST->isUnindexed() &&
7216 TLI.isTruncStoreLegal(Value.getOperand(0).getValueType(),
7217 ST->getMemoryVT())) {
7218 return DAG.getTruncStore(Chain, N->getDebugLoc(), Value.getOperand(0),
7219 Ptr, ST->getPointerInfo(), ST->getMemoryVT(),
7220 ST->isVolatile(), ST->isNonTemporal(),
7221 ST->getAlignment());
7224 return ReduceLoadOpStoreWidth(N);
7227 SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
7228 SDValue InVec = N->getOperand(0);
7229 SDValue InVal = N->getOperand(1);
7230 SDValue EltNo = N->getOperand(2);
7231 DebugLoc dl = N->getDebugLoc();
7233 // If the inserted element is an UNDEF, just use the input vector.
7234 if (InVal.getOpcode() == ISD::UNDEF)
7237 EVT VT = InVec.getValueType();
7239 // If we can't generate a legal BUILD_VECTOR, exit
7240 if (LegalOperations && !TLI.isOperationLegal(ISD::BUILD_VECTOR, VT))
7243 // Check that we know which element is being inserted
7244 if (!isa<ConstantSDNode>(EltNo))
7246 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
7248 // Check that the operand is a BUILD_VECTOR (or UNDEF, which can essentially
7249 // be converted to a BUILD_VECTOR). Fill in the Ops vector with the
7251 SmallVector<SDValue, 8> Ops;
7252 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
7253 Ops.append(InVec.getNode()->op_begin(),
7254 InVec.getNode()->op_end());
7255 } else if (InVec.getOpcode() == ISD::UNDEF) {
7256 unsigned NElts = VT.getVectorNumElements();
7257 Ops.append(NElts, DAG.getUNDEF(InVal.getValueType()));
7262 // Insert the element
7263 if (Elt < Ops.size()) {
7264 // All the operands of BUILD_VECTOR must have the same type;
7265 // we enforce that here.
7266 EVT OpVT = Ops[0].getValueType();
7267 if (InVal.getValueType() != OpVT)
7268 InVal = OpVT.bitsGT(InVal.getValueType()) ?
7269 DAG.getNode(ISD::ANY_EXTEND, dl, OpVT, InVal) :
7270 DAG.getNode(ISD::TRUNCATE, dl, OpVT, InVal);
7274 // Return the new vector
7275 return DAG.getNode(ISD::BUILD_VECTOR, dl,
7276 VT, &Ops[0], Ops.size());
7279 SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
7280 // (vextract (scalar_to_vector val, 0) -> val
7281 SDValue InVec = N->getOperand(0);
7282 EVT VT = InVec.getValueType();
7283 EVT NVT = N->getValueType(0);
7285 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
7286 // Check if the result type doesn't match the inserted element type. A
7287 // SCALAR_TO_VECTOR may truncate the inserted element and the
7288 // EXTRACT_VECTOR_ELT may widen the extracted vector.
7289 SDValue InOp = InVec.getOperand(0);
7290 if (InOp.getValueType() != NVT) {
7291 assert(InOp.getValueType().isInteger() && NVT.isInteger());
7292 return DAG.getSExtOrTrunc(InOp, InVec.getDebugLoc(), NVT);
7297 SDValue EltNo = N->getOperand(1);
7298 bool ConstEltNo = isa<ConstantSDNode>(EltNo);
7300 // Transform: (EXTRACT_VECTOR_ELT( VECTOR_SHUFFLE )) -> EXTRACT_VECTOR_ELT.
7301 // We only perform this optimization before the op legalization phase because
7302 // we may introduce new vector instructions which are not backed by TD patterns.
7303 // For example on AVX, extracting elements from a wide vector without using
7304 // extract_subvector.
7305 if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE
7306 && ConstEltNo && !LegalOperations) {
7307 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
7308 int NumElem = VT.getVectorNumElements();
7309 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(InVec);
7310 // Find the new index to extract from.
7311 int OrigElt = SVOp->getMaskElt(Elt);
7313 // Extracting an undef index is undef.
7315 return DAG.getUNDEF(NVT);
7317 // Select the right vector half to extract from.
7318 if (OrigElt < NumElem) {
7319 InVec = InVec->getOperand(0);
7321 InVec = InVec->getOperand(1);
7325 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, N->getDebugLoc(), NVT,
7326 InVec, DAG.getConstant(OrigElt, MVT::i32));
7329 // Perform only after legalization to ensure build_vector / vector_shuffle
7330 // optimizations have already been done.
7331 if (!LegalOperations) return SDValue();
7333 // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
7334 // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
7335 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
7338 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
7339 bool NewLoad = false;
7340 bool BCNumEltsChanged = false;
7341 EVT ExtVT = VT.getVectorElementType();
7344 // If the result of load has to be truncated, then it's not necessarily
7346 if (NVT.bitsLT(LVT) && !TLI.isTruncateFree(LVT, NVT))
7349 if (InVec.getOpcode() == ISD::BITCAST) {
7350 // Don't duplicate a load with other uses.
7351 if (!InVec.hasOneUse())
7354 EVT BCVT = InVec.getOperand(0).getValueType();
7355 if (!BCVT.isVector() || ExtVT.bitsGT(BCVT.getVectorElementType()))
7357 if (VT.getVectorNumElements() != BCVT.getVectorNumElements())
7358 BCNumEltsChanged = true;
7359 InVec = InVec.getOperand(0);
7360 ExtVT = BCVT.getVectorElementType();
7364 LoadSDNode *LN0 = NULL;
7365 const ShuffleVectorSDNode *SVN = NULL;
7366 if (ISD::isNormalLoad(InVec.getNode())) {
7367 LN0 = cast<LoadSDNode>(InVec);
7368 } else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
7369 InVec.getOperand(0).getValueType() == ExtVT &&
7370 ISD::isNormalLoad(InVec.getOperand(0).getNode())) {
7371 // Don't duplicate a load with other uses.
7372 if (!InVec.hasOneUse())
7375 LN0 = cast<LoadSDNode>(InVec.getOperand(0));
7376 } else if ((SVN = dyn_cast<ShuffleVectorSDNode>(InVec))) {
7377 // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
7379 // (load $addr+1*size)
7381 // Don't duplicate a load with other uses.
7382 if (!InVec.hasOneUse())
7385 // If the bit convert changed the number of elements, it is unsafe
7386 // to examine the mask.
7387 if (BCNumEltsChanged)
7390 // Select the input vector, guarding against out of range extract vector.
7391 unsigned NumElems = VT.getVectorNumElements();
7392 int Idx = (Elt > (int)NumElems) ? -1 : SVN->getMaskElt(Elt);
7393 InVec = (Idx < (int)NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
7395 if (InVec.getOpcode() == ISD::BITCAST) {
7396 // Don't duplicate a load with other uses.
7397 if (!InVec.hasOneUse())
7400 InVec = InVec.getOperand(0);
7402 if (ISD::isNormalLoad(InVec.getNode())) {
7403 LN0 = cast<LoadSDNode>(InVec);
7404 Elt = (Idx < (int)NumElems) ? Idx : Idx - (int)NumElems;
7408 // Make sure we found a non-volatile load and the extractelement is
7410 if (!LN0 || !LN0->hasNUsesOfValue(1,0) || LN0->isVolatile())
7413 // If Idx was -1 above, Elt is going to be -1, so just return undef.
7415 return DAG.getUNDEF(LVT);
7417 unsigned Align = LN0->getAlignment();
7419 // Check the resultant load doesn't need a higher alignment than the
7423 ->getABITypeAlignment(LVT.getTypeForEVT(*DAG.getContext()));
7425 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, LVT))
7431 SDValue NewPtr = LN0->getBasePtr();
7432 unsigned PtrOff = 0;
7435 PtrOff = LVT.getSizeInBits() * Elt / 8;
7436 EVT PtrType = NewPtr.getValueType();
7437 if (TLI.isBigEndian())
7438 PtrOff = VT.getSizeInBits() / 8 - PtrOff;
7439 NewPtr = DAG.getNode(ISD::ADD, N->getDebugLoc(), PtrType, NewPtr,
7440 DAG.getConstant(PtrOff, PtrType));
7443 // The replacement we need to do here is a little tricky: we need to
7444 // replace an extractelement of a load with a load.
7445 // Use ReplaceAllUsesOfValuesWith to do the replacement.
7446 // Note that this replacement assumes that the extractvalue is the only
7447 // use of the load; that's okay because we don't want to perform this
7448 // transformation in other cases anyway.
7451 if (NVT.bitsGT(LVT)) {
7452 // If the result type of vextract is wider than the load, then issue an
7453 // extending load instead.
7454 ISD::LoadExtType ExtType = TLI.isLoadExtLegal(ISD::ZEXTLOAD, LVT)
7455 ? ISD::ZEXTLOAD : ISD::EXTLOAD;
7456 Load = DAG.getExtLoad(ExtType, N->getDebugLoc(), NVT, LN0->getChain(),
7457 NewPtr, LN0->getPointerInfo().getWithOffset(PtrOff),
7458 LVT, LN0->isVolatile(), LN0->isNonTemporal(),Align);
7459 Chain = Load.getValue(1);
7461 Load = DAG.getLoad(LVT, N->getDebugLoc(), LN0->getChain(), NewPtr,
7462 LN0->getPointerInfo().getWithOffset(PtrOff),
7463 LN0->isVolatile(), LN0->isNonTemporal(),
7464 LN0->isInvariant(), Align);
7465 Chain = Load.getValue(1);
7466 if (NVT.bitsLT(LVT))
7467 Load = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), NVT, Load);
7469 Load = DAG.getNode(ISD::BITCAST, N->getDebugLoc(), NVT, Load);
7471 WorkListRemover DeadNodes(*this);
7472 SDValue From[] = { SDValue(N, 0), SDValue(LN0,1) };
7473 SDValue To[] = { Load, Chain };
7474 DAG.ReplaceAllUsesOfValuesWith(From, To, 2, &DeadNodes);
7475 // Since we're explcitly calling ReplaceAllUses, add the new node to the
7476 // worklist explicitly as well.
7477 AddToWorkList(Load.getNode());
7478 AddUsersToWorkList(Load.getNode()); // Add users too
7479 // Make sure to revisit this node to clean it up; it will usually be dead.
7481 return SDValue(N, 0);
7487 SDValue DAGCombiner::visitBUILD_VECTOR(SDNode *N) {
7488 unsigned NumInScalars = N->getNumOperands();
7489 DebugLoc dl = N->getDebugLoc();
7490 EVT VT = N->getValueType(0);
7491 // Check to see if this is a BUILD_VECTOR of a bunch of values
7492 // which come from any_extend or zero_extend nodes. If so, we can create
7493 // a new BUILD_VECTOR using bit-casts which may enable other BUILD_VECTOR
7494 // optimizations. We do not handle sign-extend because we can't fill the sign
7496 EVT SourceType = MVT::Other;
7497 bool AllAnyExt = true;
7498 bool AllUndef = true;
7499 for (unsigned i = 0; i != NumInScalars; ++i) {
7500 SDValue In = N->getOperand(i);
7501 // Ignore undef inputs.
7502 if (In.getOpcode() == ISD::UNDEF) continue;
7505 bool AnyExt = In.getOpcode() == ISD::ANY_EXTEND;
7506 bool ZeroExt = In.getOpcode() == ISD::ZERO_EXTEND;
7508 // Abort if the element is not an extension.
7509 if (!ZeroExt && !AnyExt) {
7510 SourceType = MVT::Other;
7514 // The input is a ZeroExt or AnyExt. Check the original type.
7515 EVT InTy = In.getOperand(0).getValueType();
7517 // Check that all of the widened source types are the same.
7518 if (SourceType == MVT::Other)
7521 else if (InTy != SourceType) {
7522 // Multiple income types. Abort.
7523 SourceType = MVT::Other;
7527 // Check if all of the extends are ANY_EXTENDs.
7528 AllAnyExt &= AnyExt;
7532 return DAG.getUNDEF(VT);
7534 // In order to have valid types, all of the inputs must be extended from the
7535 // same source type and all of the inputs must be any or zero extend.
7536 // Scalar sizes must be a power of two.
7537 EVT OutScalarTy = N->getValueType(0).getScalarType();
7538 bool ValidTypes = SourceType != MVT::Other &&
7539 isPowerOf2_32(OutScalarTy.getSizeInBits()) &&
7540 isPowerOf2_32(SourceType.getSizeInBits());
7542 // We perform this optimization post type-legalization because
7543 // the type-legalizer often scalarizes integer-promoted vectors.
7544 // Performing this optimization before may create bit-casts which
7545 // will be type-legalized to complex code sequences.
7546 // We perform this optimization only before the operation legalizer because we
7547 // may introduce illegal operations.
7548 // Create a new simpler BUILD_VECTOR sequence which other optimizations can
7549 // turn into a single shuffle instruction.
7550 if ((Level == AfterLegalizeVectorOps || Level == AfterLegalizeTypes) &&
7552 bool isLE = TLI.isLittleEndian();
7553 unsigned ElemRatio = OutScalarTy.getSizeInBits()/SourceType.getSizeInBits();
7554 assert(ElemRatio > 1 && "Invalid element size ratio");
7555 SDValue Filler = AllAnyExt ? DAG.getUNDEF(SourceType):
7556 DAG.getConstant(0, SourceType);
7558 unsigned NewBVElems = ElemRatio * N->getValueType(0).getVectorNumElements();
7559 SmallVector<SDValue, 8> Ops(NewBVElems, Filler);
7561 // Populate the new build_vector
7562 for (unsigned i=0; i < N->getNumOperands(); ++i) {
7563 SDValue Cast = N->getOperand(i);
7564 assert((Cast.getOpcode() == ISD::ANY_EXTEND ||
7565 Cast.getOpcode() == ISD::ZERO_EXTEND ||
7566 Cast.getOpcode() == ISD::UNDEF) && "Invalid cast opcode");
7568 if (Cast.getOpcode() == ISD::UNDEF)
7569 In = DAG.getUNDEF(SourceType);
7571 In = Cast->getOperand(0);
7572 unsigned Index = isLE ? (i * ElemRatio) :
7573 (i * ElemRatio + (ElemRatio - 1));
7575 assert(Index < Ops.size() && "Invalid index");
7579 // The type of the new BUILD_VECTOR node.
7580 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), SourceType, NewBVElems);
7581 assert(VecVT.getSizeInBits() == N->getValueType(0).getSizeInBits() &&
7582 "Invalid vector size");
7583 // Check if the new vector type is legal.
7584 if (!isTypeLegal(VecVT)) return SDValue();
7586 // Make the new BUILD_VECTOR.
7587 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
7588 VecVT, &Ops[0], Ops.size());
7590 // The new BUILD_VECTOR node has the potential to be further optimized.
7591 AddToWorkList(BV.getNode());
7592 // Bitcast to the desired type.
7593 return DAG.getNode(ISD::BITCAST, dl, N->getValueType(0), BV);
7596 // Check to see if this is a BUILD_VECTOR of a bunch of EXTRACT_VECTOR_ELT
7597 // operations. If so, and if the EXTRACT_VECTOR_ELT vector inputs come from
7598 // at most two distinct vectors, turn this into a shuffle node.
7600 // May only combine to shuffle after legalize if shuffle is legal.
7601 if (LegalOperations &&
7602 !TLI.isOperationLegalOrCustom(ISD::VECTOR_SHUFFLE, VT))
7605 SDValue VecIn1, VecIn2;
7606 for (unsigned i = 0; i != NumInScalars; ++i) {
7607 // Ignore undef inputs.
7608 if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
7610 // If this input is something other than a EXTRACT_VECTOR_ELT with a
7611 // constant index, bail out.
7612 if (N->getOperand(i).getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
7613 !isa<ConstantSDNode>(N->getOperand(i).getOperand(1))) {
7614 VecIn1 = VecIn2 = SDValue(0, 0);
7618 // We allow up to two distinct input vectors.
7619 SDValue ExtractedFromVec = N->getOperand(i).getOperand(0);
7620 if (ExtractedFromVec == VecIn1 || ExtractedFromVec == VecIn2)
7623 if (VecIn1.getNode() == 0) {
7624 VecIn1 = ExtractedFromVec;
7625 } else if (VecIn2.getNode() == 0) {
7626 VecIn2 = ExtractedFromVec;
7629 VecIn1 = VecIn2 = SDValue(0, 0);
7634 // If everything is good, we can make a shuffle operation.
7635 if (VecIn1.getNode()) {
7636 SmallVector<int, 8> Mask;
7637 for (unsigned i = 0; i != NumInScalars; ++i) {
7638 if (N->getOperand(i).getOpcode() == ISD::UNDEF) {
7643 // If extracting from the first vector, just use the index directly.
7644 SDValue Extract = N->getOperand(i);
7645 SDValue ExtVal = Extract.getOperand(1);
7646 if (Extract.getOperand(0) == VecIn1) {
7647 unsigned ExtIndex = cast<ConstantSDNode>(ExtVal)->getZExtValue();
7648 if (ExtIndex > VT.getVectorNumElements())
7651 Mask.push_back(ExtIndex);
7655 // Otherwise, use InIdx + VecSize
7656 unsigned Idx = cast<ConstantSDNode>(ExtVal)->getZExtValue();
7657 Mask.push_back(Idx+NumInScalars);
7660 // We can't generate a shuffle node with mismatched input and output types.
7661 // Attempt to transform a single input vector to the correct type.
7662 if ((VT != VecIn1.getValueType())) {
7663 // We don't support shuffeling between TWO values of different types.
7664 if (VecIn2.getNode() != 0)
7667 // We only support widening of vectors which are half the size of the
7668 // output registers. For example XMM->YMM widening on X86 with AVX.
7669 if (VecIn1.getValueType().getSizeInBits()*2 != VT.getSizeInBits())
7672 // Widen the input vector by adding undef values.
7673 VecIn1 = DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT,
7674 VecIn1, DAG.getUNDEF(VecIn1.getValueType()));
7677 // If VecIn2 is unused then change it to undef.
7678 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
7680 // Check that we were able to transform all incoming values to the same type.
7681 if (VecIn2.getValueType() != VecIn1.getValueType() ||
7682 VecIn1.getValueType() != VT)
7685 // Only type-legal BUILD_VECTOR nodes are converted to shuffle nodes.
7686 if (!isTypeLegal(VT))
7689 // Return the new VECTOR_SHUFFLE node.
7693 return DAG.getVectorShuffle(VT, N->getDebugLoc(), Ops[0], Ops[1], &Mask[0]);
7699 SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
7700 // TODO: Check to see if this is a CONCAT_VECTORS of a bunch of
7701 // EXTRACT_SUBVECTOR operations. If so, and if the EXTRACT_SUBVECTOR vector
7702 // inputs come from at most two distinct vectors, turn this into a shuffle
7705 // If we only have one input vector, we don't need to do any concatenation.
7706 if (N->getNumOperands() == 1)
7707 return N->getOperand(0);
7712 SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
7713 EVT NVT = N->getValueType(0);
7714 SDValue V = N->getOperand(0);
7716 if (V->getOpcode() == ISD::INSERT_SUBVECTOR) {
7717 // Handle only simple case where vector being inserted and vector
7718 // being extracted are of same type, and are half size of larger vectors.
7719 EVT BigVT = V->getOperand(0).getValueType();
7720 EVT SmallVT = V->getOperand(1).getValueType();
7721 if (NVT != SmallVT || NVT.getSizeInBits()*2 != BigVT.getSizeInBits())
7724 // Only handle cases where both indexes are constants with the same type.
7725 ConstantSDNode *InsIdx = dyn_cast<ConstantSDNode>(N->getOperand(1));
7726 ConstantSDNode *ExtIdx = dyn_cast<ConstantSDNode>(V->getOperand(2));
7728 if (InsIdx && ExtIdx &&
7729 InsIdx->getValueType(0).getSizeInBits() <= 64 &&
7730 ExtIdx->getValueType(0).getSizeInBits() <= 64) {
7732 // (extract_subvec (insert_subvec V1, V2, InsIdx), ExtIdx)
7734 // indices are equal => V1
7735 // otherwise => (extract_subvec V1, ExtIdx)
7736 if (InsIdx->getZExtValue() == ExtIdx->getZExtValue())
7737 return V->getOperand(1);
7738 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, N->getDebugLoc(), NVT,
7739 V->getOperand(0), N->getOperand(1));
7746 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
7747 EVT VT = N->getValueType(0);
7748 unsigned NumElts = VT.getVectorNumElements();
7750 SDValue N0 = N->getOperand(0);
7751 SDValue N1 = N->getOperand(1);
7753 assert(N0.getValueType() == VT && "Vector shuffle must be normalized in DAG");
7755 // Canonicalize shuffle undef, undef -> undef
7756 if (N0.getOpcode() == ISD::UNDEF && N1.getOpcode() == ISD::UNDEF)
7757 return DAG.getUNDEF(VT);
7759 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
7761 // Canonicalize shuffle v, v -> v, undef
7763 SmallVector<int, 8> NewMask;
7764 for (unsigned i = 0; i != NumElts; ++i) {
7765 int Idx = SVN->getMaskElt(i);
7766 if (Idx >= (int)NumElts) Idx -= NumElts;
7767 NewMask.push_back(Idx);
7769 return DAG.getVectorShuffle(VT, N->getDebugLoc(), N0, DAG.getUNDEF(VT),
7773 // Canonicalize shuffle undef, v -> v, undef. Commute the shuffle mask.
7774 if (N0.getOpcode() == ISD::UNDEF) {
7775 SmallVector<int, 8> NewMask;
7776 for (unsigned i = 0; i != NumElts; ++i) {
7777 int Idx = SVN->getMaskElt(i);
7779 if (Idx < (int)NumElts)
7784 NewMask.push_back(Idx);
7786 return DAG.getVectorShuffle(VT, N->getDebugLoc(), N1, DAG.getUNDEF(VT),
7790 // Remove references to rhs if it is undef
7791 if (N1.getOpcode() == ISD::UNDEF) {
7792 bool Changed = false;
7793 SmallVector<int, 8> NewMask;
7794 for (unsigned i = 0; i != NumElts; ++i) {
7795 int Idx = SVN->getMaskElt(i);
7796 if (Idx >= (int)NumElts) {
7800 NewMask.push_back(Idx);
7803 return DAG.getVectorShuffle(VT, N->getDebugLoc(), N0, N1, &NewMask[0]);
7806 // If it is a splat, check if the argument vector is another splat or a
7807 // build_vector with all scalar elements the same.
7808 if (SVN->isSplat() && SVN->getSplatIndex() < (int)NumElts) {
7809 SDNode *V = N0.getNode();
7811 // If this is a bit convert that changes the element type of the vector but
7812 // not the number of vector elements, look through it. Be careful not to
7813 // look though conversions that change things like v4f32 to v2f64.
7814 if (V->getOpcode() == ISD::BITCAST) {
7815 SDValue ConvInput = V->getOperand(0);
7816 if (ConvInput.getValueType().isVector() &&
7817 ConvInput.getValueType().getVectorNumElements() == NumElts)
7818 V = ConvInput.getNode();
7821 if (V->getOpcode() == ISD::BUILD_VECTOR) {
7822 assert(V->getNumOperands() == NumElts &&
7823 "BUILD_VECTOR has wrong number of operands");
7825 bool AllSame = true;
7826 for (unsigned i = 0; i != NumElts; ++i) {
7827 if (V->getOperand(i).getOpcode() != ISD::UNDEF) {
7828 Base = V->getOperand(i);
7832 // Splat of <u, u, u, u>, return <u, u, u, u>
7833 if (!Base.getNode())
7835 for (unsigned i = 0; i != NumElts; ++i) {
7836 if (V->getOperand(i) != Base) {
7841 // Splat of <x, x, x, x>, return <x, x, x, x>
7847 // If this shuffle node is simply a swizzle of another shuffle node,
7848 // and it reverses the swizzle of the previous shuffle then we can
7849 // optimize shuffle(shuffle(x, undef), undef) -> x.
7850 if (N0.getOpcode() == ISD::VECTOR_SHUFFLE && Level < AfterLegalizeDAG &&
7851 N1.getOpcode() == ISD::UNDEF) {
7853 ShuffleVectorSDNode *OtherSV = cast<ShuffleVectorSDNode>(N0);
7855 // Shuffle nodes can only reverse shuffles with a single non-undef value.
7856 if (N0.getOperand(1).getOpcode() != ISD::UNDEF)
7859 // The incoming shuffle must be of the same type as the result of the
7861 assert(OtherSV->getOperand(0).getValueType() == VT &&
7862 "Shuffle types don't match");
7864 for (unsigned i = 0; i != NumElts; ++i) {
7865 int Idx = SVN->getMaskElt(i);
7866 assert(Idx < (int)NumElts && "Index references undef operand");
7867 // Next, this index comes from the first value, which is the incoming
7868 // shuffle. Adopt the incoming index.
7870 Idx = OtherSV->getMaskElt(Idx);
7872 // The combined shuffle must map each index to itself.
7873 if (Idx >= 0 && (unsigned)Idx != i)
7877 return OtherSV->getOperand(0);
7883 SDValue DAGCombiner::visitMEMBARRIER(SDNode* N) {
7884 if (!TLI.getShouldFoldAtomicFences())
7887 SDValue atomic = N->getOperand(0);
7888 switch (atomic.getOpcode()) {
7889 case ISD::ATOMIC_CMP_SWAP:
7890 case ISD::ATOMIC_SWAP:
7891 case ISD::ATOMIC_LOAD_ADD:
7892 case ISD::ATOMIC_LOAD_SUB:
7893 case ISD::ATOMIC_LOAD_AND:
7894 case ISD::ATOMIC_LOAD_OR:
7895 case ISD::ATOMIC_LOAD_XOR:
7896 case ISD::ATOMIC_LOAD_NAND:
7897 case ISD::ATOMIC_LOAD_MIN:
7898 case ISD::ATOMIC_LOAD_MAX:
7899 case ISD::ATOMIC_LOAD_UMIN:
7900 case ISD::ATOMIC_LOAD_UMAX:
7906 SDValue fence = atomic.getOperand(0);
7907 if (fence.getOpcode() != ISD::MEMBARRIER)
7910 switch (atomic.getOpcode()) {
7911 case ISD::ATOMIC_CMP_SWAP:
7912 return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
7913 fence.getOperand(0),
7914 atomic.getOperand(1), atomic.getOperand(2),
7915 atomic.getOperand(3)), atomic.getResNo());
7916 case ISD::ATOMIC_SWAP:
7917 case ISD::ATOMIC_LOAD_ADD:
7918 case ISD::ATOMIC_LOAD_SUB:
7919 case ISD::ATOMIC_LOAD_AND:
7920 case ISD::ATOMIC_LOAD_OR:
7921 case ISD::ATOMIC_LOAD_XOR:
7922 case ISD::ATOMIC_LOAD_NAND:
7923 case ISD::ATOMIC_LOAD_MIN:
7924 case ISD::ATOMIC_LOAD_MAX:
7925 case ISD::ATOMIC_LOAD_UMIN:
7926 case ISD::ATOMIC_LOAD_UMAX:
7927 return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
7928 fence.getOperand(0),
7929 atomic.getOperand(1), atomic.getOperand(2)),
7936 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
7937 /// an AND to a vector_shuffle with the destination vector and a zero vector.
7938 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
7939 /// vector_shuffle V, Zero, <0, 4, 2, 4>
7940 SDValue DAGCombiner::XformToShuffleWithZero(SDNode *N) {
7941 EVT VT = N->getValueType(0);
7942 DebugLoc dl = N->getDebugLoc();
7943 SDValue LHS = N->getOperand(0);
7944 SDValue RHS = N->getOperand(1);
7945 if (N->getOpcode() == ISD::AND) {
7946 if (RHS.getOpcode() == ISD::BITCAST)
7947 RHS = RHS.getOperand(0);
7948 if (RHS.getOpcode() == ISD::BUILD_VECTOR) {
7949 SmallVector<int, 8> Indices;
7950 unsigned NumElts = RHS.getNumOperands();
7951 for (unsigned i = 0; i != NumElts; ++i) {
7952 SDValue Elt = RHS.getOperand(i);
7953 if (!isa<ConstantSDNode>(Elt))
7956 if (cast<ConstantSDNode>(Elt)->isAllOnesValue())
7957 Indices.push_back(i);
7958 else if (cast<ConstantSDNode>(Elt)->isNullValue())
7959 Indices.push_back(NumElts);
7964 // Let's see if the target supports this vector_shuffle.
7965 EVT RVT = RHS.getValueType();
7966 if (!TLI.isVectorClearMaskLegal(Indices, RVT))
7969 // Return the new VECTOR_SHUFFLE node.
7970 EVT EltVT = RVT.getVectorElementType();
7971 SmallVector<SDValue,8> ZeroOps(RVT.getVectorNumElements(),
7972 DAG.getConstant(0, EltVT));
7973 SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
7974 RVT, &ZeroOps[0], ZeroOps.size());
7975 LHS = DAG.getNode(ISD::BITCAST, dl, RVT, LHS);
7976 SDValue Shuf = DAG.getVectorShuffle(RVT, dl, LHS, Zero, &Indices[0]);
7977 return DAG.getNode(ISD::BITCAST, dl, VT, Shuf);
7984 /// SimplifyVBinOp - Visit a binary vector operation, like ADD.
7985 SDValue DAGCombiner::SimplifyVBinOp(SDNode *N) {
7986 // After legalize, the target may be depending on adds and other
7987 // binary ops to provide legal ways to construct constants or other
7988 // things. Simplifying them may result in a loss of legality.
7989 if (LegalOperations) return SDValue();
7991 assert(N->getValueType(0).isVector() &&
7992 "SimplifyVBinOp only works on vectors!");
7994 SDValue LHS = N->getOperand(0);
7995 SDValue RHS = N->getOperand(1);
7996 SDValue Shuffle = XformToShuffleWithZero(N);
7997 if (Shuffle.getNode()) return Shuffle;
7999 // If the LHS and RHS are BUILD_VECTOR nodes, see if we can constant fold
8001 if (LHS.getOpcode() == ISD::BUILD_VECTOR &&
8002 RHS.getOpcode() == ISD::BUILD_VECTOR) {
8003 SmallVector<SDValue, 8> Ops;
8004 for (unsigned i = 0, e = LHS.getNumOperands(); i != e; ++i) {
8005 SDValue LHSOp = LHS.getOperand(i);
8006 SDValue RHSOp = RHS.getOperand(i);
8007 // If these two elements can't be folded, bail out.
8008 if ((LHSOp.getOpcode() != ISD::UNDEF &&
8009 LHSOp.getOpcode() != ISD::Constant &&
8010 LHSOp.getOpcode() != ISD::ConstantFP) ||
8011 (RHSOp.getOpcode() != ISD::UNDEF &&
8012 RHSOp.getOpcode() != ISD::Constant &&
8013 RHSOp.getOpcode() != ISD::ConstantFP))
8016 // Can't fold divide by zero.
8017 if (N->getOpcode() == ISD::SDIV || N->getOpcode() == ISD::UDIV ||
8018 N->getOpcode() == ISD::FDIV) {
8019 if ((RHSOp.getOpcode() == ISD::Constant &&
8020 cast<ConstantSDNode>(RHSOp.getNode())->isNullValue()) ||
8021 (RHSOp.getOpcode() == ISD::ConstantFP &&
8022 cast<ConstantFPSDNode>(RHSOp.getNode())->getValueAPF().isZero()))
8026 EVT VT = LHSOp.getValueType();
8027 EVT RVT = RHSOp.getValueType();
8029 // Integer BUILD_VECTOR operands may have types larger than the element
8030 // size (e.g., when the element type is not legal). Prior to type
8031 // legalization, the types may not match between the two BUILD_VECTORS.
8032 // Truncate one of the operands to make them match.
8033 if (RVT.getSizeInBits() > VT.getSizeInBits()) {
8034 RHSOp = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), VT, RHSOp);
8036 LHSOp = DAG.getNode(ISD::TRUNCATE, N->getDebugLoc(), RVT, LHSOp);
8040 SDValue FoldOp = DAG.getNode(N->getOpcode(), LHS.getDebugLoc(), VT,
8042 if (FoldOp.getOpcode() != ISD::UNDEF &&
8043 FoldOp.getOpcode() != ISD::Constant &&
8044 FoldOp.getOpcode() != ISD::ConstantFP)
8046 Ops.push_back(FoldOp);
8047 AddToWorkList(FoldOp.getNode());
8050 if (Ops.size() == LHS.getNumOperands())
8051 return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
8052 LHS.getValueType(), &Ops[0], Ops.size());
8058 SDValue DAGCombiner::SimplifySelect(DebugLoc DL, SDValue N0,
8059 SDValue N1, SDValue N2){
8060 assert(N0.getOpcode() ==ISD::SETCC && "First argument must be a SetCC node!");
8062 SDValue SCC = SimplifySelectCC(DL, N0.getOperand(0), N0.getOperand(1), N1, N2,
8063 cast<CondCodeSDNode>(N0.getOperand(2))->get());
8065 // If we got a simplified select_cc node back from SimplifySelectCC, then
8066 // break it down into a new SETCC node, and a new SELECT node, and then return
8067 // the SELECT node, since we were called with a SELECT node.
8068 if (SCC.getNode()) {
8069 // Check to see if we got a select_cc back (to turn into setcc/select).
8070 // Otherwise, just return whatever node we got back, like fabs.
8071 if (SCC.getOpcode() == ISD::SELECT_CC) {
8072 SDValue SETCC = DAG.getNode(ISD::SETCC, N0.getDebugLoc(),
8074 SCC.getOperand(0), SCC.getOperand(1),
8076 AddToWorkList(SETCC.getNode());
8077 return DAG.getNode(ISD::SELECT, SCC.getDebugLoc(), SCC.getValueType(),
8078 SCC.getOperand(2), SCC.getOperand(3), SETCC);
8086 /// SimplifySelectOps - Given a SELECT or a SELECT_CC node, where LHS and RHS
8087 /// are the two values being selected between, see if we can simplify the
8088 /// select. Callers of this should assume that TheSelect is deleted if this
8089 /// returns true. As such, they should return the appropriate thing (e.g. the
8090 /// node) back to the top-level of the DAG combiner loop to avoid it being
8092 bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS,
8095 // Cannot simplify select with vector condition
8096 if (TheSelect->getOperand(0).getValueType().isVector()) return false;
8098 // If this is a select from two identical things, try to pull the operation
8099 // through the select.
8100 if (LHS.getOpcode() != RHS.getOpcode() ||
8101 !LHS.hasOneUse() || !RHS.hasOneUse())
8104 // If this is a load and the token chain is identical, replace the select
8105 // of two loads with a load through a select of the address to load from.
8106 // This triggers in things like "select bool X, 10.0, 123.0" after the FP
8107 // constants have been dropped into the constant pool.
8108 if (LHS.getOpcode() == ISD::LOAD) {
8109 LoadSDNode *LLD = cast<LoadSDNode>(LHS);
8110 LoadSDNode *RLD = cast<LoadSDNode>(RHS);
8112 // Token chains must be identical.
8113 if (LHS.getOperand(0) != RHS.getOperand(0) ||
8114 // Do not let this transformation reduce the number of volatile loads.
8115 LLD->isVolatile() || RLD->isVolatile() ||
8116 // If this is an EXTLOAD, the VT's must match.
8117 LLD->getMemoryVT() != RLD->getMemoryVT() ||
8118 // If this is an EXTLOAD, the kind of extension must match.
8119 (LLD->getExtensionType() != RLD->getExtensionType() &&
8120 // The only exception is if one of the extensions is anyext.
8121 LLD->getExtensionType() != ISD::EXTLOAD &&
8122 RLD->getExtensionType() != ISD::EXTLOAD) ||
8123 // FIXME: this discards src value information. This is
8124 // over-conservative. It would be beneficial to be able to remember
8125 // both potential memory locations. Since we are discarding
8126 // src value info, don't do the transformation if the memory
8127 // locations are not in the default address space.
8128 LLD->getPointerInfo().getAddrSpace() != 0 ||
8129 RLD->getPointerInfo().getAddrSpace() != 0)
8132 // Check that the select condition doesn't reach either load. If so,
8133 // folding this will induce a cycle into the DAG. If not, this is safe to
8134 // xform, so create a select of the addresses.
8136 if (TheSelect->getOpcode() == ISD::SELECT) {
8137 SDNode *CondNode = TheSelect->getOperand(0).getNode();
8138 if ((LLD->hasAnyUseOfValue(1) && LLD->isPredecessorOf(CondNode)) ||
8139 (RLD->hasAnyUseOfValue(1) && RLD->isPredecessorOf(CondNode)))
8141 Addr = DAG.getNode(ISD::SELECT, TheSelect->getDebugLoc(),
8142 LLD->getBasePtr().getValueType(),
8143 TheSelect->getOperand(0), LLD->getBasePtr(),
8145 } else { // Otherwise SELECT_CC
8146 SDNode *CondLHS = TheSelect->getOperand(0).getNode();
8147 SDNode *CondRHS = TheSelect->getOperand(1).getNode();
8149 if ((LLD->hasAnyUseOfValue(1) &&
8150 (LLD->isPredecessorOf(CondLHS) || LLD->isPredecessorOf(CondRHS))) ||
8151 (RLD->hasAnyUseOfValue(1) &&
8152 (RLD->isPredecessorOf(CondLHS) || RLD->isPredecessorOf(CondRHS))))
8155 Addr = DAG.getNode(ISD::SELECT_CC, TheSelect->getDebugLoc(),
8156 LLD->getBasePtr().getValueType(),
8157 TheSelect->getOperand(0),
8158 TheSelect->getOperand(1),
8159 LLD->getBasePtr(), RLD->getBasePtr(),
8160 TheSelect->getOperand(4));
8164 if (LLD->getExtensionType() == ISD::NON_EXTLOAD) {
8165 Load = DAG.getLoad(TheSelect->getValueType(0),
8166 TheSelect->getDebugLoc(),
8167 // FIXME: Discards pointer info.
8168 LLD->getChain(), Addr, MachinePointerInfo(),
8169 LLD->isVolatile(), LLD->isNonTemporal(),
8170 LLD->isInvariant(), LLD->getAlignment());
8172 Load = DAG.getExtLoad(LLD->getExtensionType() == ISD::EXTLOAD ?
8173 RLD->getExtensionType() : LLD->getExtensionType(),
8174 TheSelect->getDebugLoc(),
8175 TheSelect->getValueType(0),
8176 // FIXME: Discards pointer info.
8177 LLD->getChain(), Addr, MachinePointerInfo(),
8178 LLD->getMemoryVT(), LLD->isVolatile(),
8179 LLD->isNonTemporal(), LLD->getAlignment());
8182 // Users of the select now use the result of the load.
8183 CombineTo(TheSelect, Load);
8185 // Users of the old loads now use the new load's chain. We know the
8186 // old-load value is dead now.
8187 CombineTo(LHS.getNode(), Load.getValue(0), Load.getValue(1));
8188 CombineTo(RHS.getNode(), Load.getValue(0), Load.getValue(1));
8195 /// SimplifySelectCC - Simplify an expression of the form (N0 cond N1) ? N2 : N3
8196 /// where 'cond' is the comparison specified by CC.
8197 SDValue DAGCombiner::SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1,
8198 SDValue N2, SDValue N3,
8199 ISD::CondCode CC, bool NotExtCompare) {
8200 // (x ? y : y) -> y.
8201 if (N2 == N3) return N2;
8203 EVT VT = N2.getValueType();
8204 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode());
8205 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(N2.getNode());
8206 ConstantSDNode *N3C = dyn_cast<ConstantSDNode>(N3.getNode());
8208 // Determine if the condition we're dealing with is constant
8209 SDValue SCC = SimplifySetCC(TLI.getSetCCResultType(N0.getValueType()),
8210 N0, N1, CC, DL, false);
8211 if (SCC.getNode()) AddToWorkList(SCC.getNode());
8212 ConstantSDNode *SCCC = dyn_cast_or_null<ConstantSDNode>(SCC.getNode());
8214 // fold select_cc true, x, y -> x
8215 if (SCCC && !SCCC->isNullValue())
8217 // fold select_cc false, x, y -> y
8218 if (SCCC && SCCC->isNullValue())
8221 // Check to see if we can simplify the select into an fabs node
8222 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1)) {
8223 // Allow either -0.0 or 0.0
8224 if (CFP->getValueAPF().isZero()) {
8225 // select (setg[te] X, +/-0.0), X, fneg(X) -> fabs
8226 if ((CC == ISD::SETGE || CC == ISD::SETGT) &&
8227 N0 == N2 && N3.getOpcode() == ISD::FNEG &&
8228 N2 == N3.getOperand(0))
8229 return DAG.getNode(ISD::FABS, DL, VT, N0);
8231 // select (setl[te] X, +/-0.0), fneg(X), X -> fabs
8232 if ((CC == ISD::SETLT || CC == ISD::SETLE) &&
8233 N0 == N3 && N2.getOpcode() == ISD::FNEG &&
8234 N2.getOperand(0) == N3)
8235 return DAG.getNode(ISD::FABS, DL, VT, N3);
8239 // Turn "(a cond b) ? 1.0f : 2.0f" into "load (tmp + ((a cond b) ? 0 : 4)"
8240 // where "tmp" is a constant pool entry containing an array with 1.0 and 2.0
8241 // in it. This is a win when the constant is not otherwise available because
8242 // it replaces two constant pool loads with one. We only do this if the FP
8243 // type is known to be legal, because if it isn't, then we are before legalize
8244 // types an we want the other legalization to happen first (e.g. to avoid
8245 // messing with soft float) and if the ConstantFP is not legal, because if
8246 // it is legal, we may not need to store the FP constant in a constant pool.
8247 if (ConstantFPSDNode *TV = dyn_cast<ConstantFPSDNode>(N2))
8248 if (ConstantFPSDNode *FV = dyn_cast<ConstantFPSDNode>(N3)) {
8249 if (TLI.isTypeLegal(N2.getValueType()) &&
8250 (TLI.getOperationAction(ISD::ConstantFP, N2.getValueType()) !=
8251 TargetLowering::Legal) &&
8252 // If both constants have multiple uses, then we won't need to do an
8253 // extra load, they are likely around in registers for other users.
8254 (TV->hasOneUse() || FV->hasOneUse())) {
8255 Constant *Elts[] = {
8256 const_cast<ConstantFP*>(FV->getConstantFPValue()),
8257 const_cast<ConstantFP*>(TV->getConstantFPValue())
8259 Type *FPTy = Elts[0]->getType();
8260 const TargetData &TD = *TLI.getTargetData();
8262 // Create a ConstantArray of the two constants.
8263 Constant *CA = ConstantArray::get(ArrayType::get(FPTy, 2), Elts);
8264 SDValue CPIdx = DAG.getConstantPool(CA, TLI.getPointerTy(),
8265 TD.getPrefTypeAlignment(FPTy));
8266 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
8268 // Get the offsets to the 0 and 1 element of the array so that we can
8269 // select between them.
8270 SDValue Zero = DAG.getIntPtrConstant(0);
8271 unsigned EltSize = (unsigned)TD.getTypeAllocSize(Elts[0]->getType());
8272 SDValue One = DAG.getIntPtrConstant(EltSize);
8274 SDValue Cond = DAG.getSetCC(DL,
8275 TLI.getSetCCResultType(N0.getValueType()),
8277 AddToWorkList(Cond.getNode());
8278 SDValue CstOffset = DAG.getNode(ISD::SELECT, DL, Zero.getValueType(),
8280 AddToWorkList(CstOffset.getNode());
8281 CPIdx = DAG.getNode(ISD::ADD, DL, TLI.getPointerTy(), CPIdx,
8283 AddToWorkList(CPIdx.getNode());
8284 return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx,
8285 MachinePointerInfo::getConstantPool(), false,
8286 false, false, Alignment);
8291 // Check to see if we can perform the "gzip trick", transforming
8292 // (select_cc setlt X, 0, A, 0) -> (and (sra X, (sub size(X), 1), A)
8293 if (N1C && N3C && N3C->isNullValue() && CC == ISD::SETLT &&
8294 (N1C->isNullValue() || // (a < 0) ? b : 0
8295 (N1C->getAPIntValue() == 1 && N0 == N2))) { // (a < 1) ? a : 0
8296 EVT XType = N0.getValueType();
8297 EVT AType = N2.getValueType();
8298 if (XType.bitsGE(AType)) {
8299 // and (sra X, size(X)-1, A) -> "and (srl X, C2), A" iff A is a
8300 // single-bit constant.
8301 if (N2C && ((N2C->getAPIntValue() & (N2C->getAPIntValue()-1)) == 0)) {
8302 unsigned ShCtV = N2C->getAPIntValue().logBase2();
8303 ShCtV = XType.getSizeInBits()-ShCtV-1;
8304 SDValue ShCt = DAG.getConstant(ShCtV,
8305 getShiftAmountTy(N0.getValueType()));
8306 SDValue Shift = DAG.getNode(ISD::SRL, N0.getDebugLoc(),
8308 AddToWorkList(Shift.getNode());
8310 if (XType.bitsGT(AType)) {
8311 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
8312 AddToWorkList(Shift.getNode());
8315 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
8318 SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(),
8320 DAG.getConstant(XType.getSizeInBits()-1,
8321 getShiftAmountTy(N0.getValueType())));
8322 AddToWorkList(Shift.getNode());
8324 if (XType.bitsGT(AType)) {
8325 Shift = DAG.getNode(ISD::TRUNCATE, DL, AType, Shift);
8326 AddToWorkList(Shift.getNode());
8329 return DAG.getNode(ISD::AND, DL, AType, Shift, N2);
8333 // fold (select_cc seteq (and x, y), 0, 0, A) -> (and (shr (shl x)) A)
8334 // where y is has a single bit set.
8335 // A plaintext description would be, we can turn the SELECT_CC into an AND
8336 // when the condition can be materialized as an all-ones register. Any
8337 // single bit-test can be materialized as an all-ones register with
8338 // shift-left and shift-right-arith.
8339 if (CC == ISD::SETEQ && N0->getOpcode() == ISD::AND &&
8340 N0->getValueType(0) == VT &&
8341 N1C && N1C->isNullValue() &&
8342 N2C && N2C->isNullValue()) {
8343 SDValue AndLHS = N0->getOperand(0);
8344 ConstantSDNode *ConstAndRHS = dyn_cast<ConstantSDNode>(N0->getOperand(1));
8345 if (ConstAndRHS && ConstAndRHS->getAPIntValue().countPopulation() == 1) {
8346 // Shift the tested bit over the sign bit.
8347 APInt AndMask = ConstAndRHS->getAPIntValue();
8349 DAG.getConstant(AndMask.countLeadingZeros(),
8350 getShiftAmountTy(AndLHS.getValueType()));
8351 SDValue Shl = DAG.getNode(ISD::SHL, N0.getDebugLoc(), VT, AndLHS, ShlAmt);
8353 // Now arithmetic right shift it all the way over, so the result is either
8354 // all-ones, or zero.
8356 DAG.getConstant(AndMask.getBitWidth()-1,
8357 getShiftAmountTy(Shl.getValueType()));
8358 SDValue Shr = DAG.getNode(ISD::SRA, N0.getDebugLoc(), VT, Shl, ShrAmt);
8360 return DAG.getNode(ISD::AND, DL, VT, Shr, N3);
8364 // fold select C, 16, 0 -> shl C, 4
8365 if (N2C && N3C && N3C->isNullValue() && N2C->getAPIntValue().isPowerOf2() &&
8366 TLI.getBooleanContents(N0.getValueType().isVector()) ==
8367 TargetLowering::ZeroOrOneBooleanContent) {
8369 // If the caller doesn't want us to simplify this into a zext of a compare,
8371 if (NotExtCompare && N2C->getAPIntValue() == 1)
8374 // Get a SetCC of the condition
8375 // FIXME: Should probably make sure that setcc is legal if we ever have a
8376 // target where it isn't.
8378 // cast from setcc result type to select result type
8380 SCC = DAG.getSetCC(DL, TLI.getSetCCResultType(N0.getValueType()),
8382 if (N2.getValueType().bitsLT(SCC.getValueType()))
8383 Temp = DAG.getZeroExtendInReg(SCC, N2.getDebugLoc(), N2.getValueType());
8385 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getDebugLoc(),
8386 N2.getValueType(), SCC);
8388 SCC = DAG.getSetCC(N0.getDebugLoc(), MVT::i1, N0, N1, CC);
8389 Temp = DAG.getNode(ISD::ZERO_EXTEND, N2.getDebugLoc(),
8390 N2.getValueType(), SCC);
8393 AddToWorkList(SCC.getNode());
8394 AddToWorkList(Temp.getNode());
8396 if (N2C->getAPIntValue() == 1)
8399 // shl setcc result by log2 n2c
8400 return DAG.getNode(ISD::SHL, DL, N2.getValueType(), Temp,
8401 DAG.getConstant(N2C->getAPIntValue().logBase2(),
8402 getShiftAmountTy(Temp.getValueType())));
8405 // Check to see if this is the equivalent of setcc
8406 // FIXME: Turn all of these into setcc if setcc if setcc is legal
8407 // otherwise, go ahead with the folds.
8408 if (0 && N3C && N3C->isNullValue() && N2C && (N2C->getAPIntValue() == 1ULL)) {
8409 EVT XType = N0.getValueType();
8410 if (!LegalOperations ||
8411 TLI.isOperationLegal(ISD::SETCC, TLI.getSetCCResultType(XType))) {
8412 SDValue Res = DAG.getSetCC(DL, TLI.getSetCCResultType(XType), N0, N1, CC);
8413 if (Res.getValueType() != VT)
8414 Res = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, Res);
8418 // fold (seteq X, 0) -> (srl (ctlz X, log2(size(X))))
8419 if (N1C && N1C->isNullValue() && CC == ISD::SETEQ &&
8420 (!LegalOperations ||
8421 TLI.isOperationLegal(ISD::CTLZ, XType))) {
8422 SDValue Ctlz = DAG.getNode(ISD::CTLZ, N0.getDebugLoc(), XType, N0);
8423 return DAG.getNode(ISD::SRL, DL, XType, Ctlz,
8424 DAG.getConstant(Log2_32(XType.getSizeInBits()),
8425 getShiftAmountTy(Ctlz.getValueType())));
8427 // fold (setgt X, 0) -> (srl (and (-X, ~X), size(X)-1))
8428 if (N1C && N1C->isNullValue() && CC == ISD::SETGT) {
8429 SDValue NegN0 = DAG.getNode(ISD::SUB, N0.getDebugLoc(),
8430 XType, DAG.getConstant(0, XType), N0);
8431 SDValue NotN0 = DAG.getNOT(N0.getDebugLoc(), N0, XType);
8432 return DAG.getNode(ISD::SRL, DL, XType,
8433 DAG.getNode(ISD::AND, DL, XType, NegN0, NotN0),
8434 DAG.getConstant(XType.getSizeInBits()-1,
8435 getShiftAmountTy(XType)));
8437 // fold (setgt X, -1) -> (xor (srl (X, size(X)-1), 1))
8438 if (N1C && N1C->isAllOnesValue() && CC == ISD::SETGT) {
8439 SDValue Sign = DAG.getNode(ISD::SRL, N0.getDebugLoc(), XType, N0,
8440 DAG.getConstant(XType.getSizeInBits()-1,
8441 getShiftAmountTy(N0.getValueType())));
8442 return DAG.getNode(ISD::XOR, DL, XType, Sign, DAG.getConstant(1, XType));
8446 // Check to see if this is an integer abs.
8447 // select_cc setg[te] X, 0, X, -X ->
8448 // select_cc setgt X, -1, X, -X ->
8449 // select_cc setl[te] X, 0, -X, X ->
8450 // select_cc setlt X, 1, -X, X ->
8451 // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
8453 ConstantSDNode *SubC = NULL;
8454 if (((N1C->isNullValue() && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
8455 (N1C->isAllOnesValue() && CC == ISD::SETGT)) &&
8456 N0 == N2 && N3.getOpcode() == ISD::SUB && N0 == N3.getOperand(1))
8457 SubC = dyn_cast<ConstantSDNode>(N3.getOperand(0));
8458 else if (((N1C->isNullValue() && (CC == ISD::SETLT || CC == ISD::SETLE)) ||
8459 (N1C->isOne() && CC == ISD::SETLT)) &&
8460 N0 == N3 && N2.getOpcode() == ISD::SUB && N0 == N2.getOperand(1))
8461 SubC = dyn_cast<ConstantSDNode>(N2.getOperand(0));
8463 EVT XType = N0.getValueType();
8464 if (SubC && SubC->isNullValue() && XType.isInteger()) {
8465 SDValue Shift = DAG.getNode(ISD::SRA, N0.getDebugLoc(), XType,
8467 DAG.getConstant(XType.getSizeInBits()-1,
8468 getShiftAmountTy(N0.getValueType())));
8469 SDValue Add = DAG.getNode(ISD::ADD, N0.getDebugLoc(),
8471 AddToWorkList(Shift.getNode());
8472 AddToWorkList(Add.getNode());
8473 return DAG.getNode(ISD::XOR, DL, XType, Add, Shift);
8480 /// SimplifySetCC - This is a stub for TargetLowering::SimplifySetCC.
8481 SDValue DAGCombiner::SimplifySetCC(EVT VT, SDValue N0,
8482 SDValue N1, ISD::CondCode Cond,
8483 DebugLoc DL, bool foldBooleans) {
8484 TargetLowering::DAGCombinerInfo
8485 DagCombineInfo(DAG, !LegalTypes, !LegalOperations, false, this);
8486 return TLI.SimplifySetCC(VT, N0, N1, Cond, foldBooleans, DagCombineInfo, DL);
8489 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant,
8490 /// return a DAG expression to select that will generate the same value by
8491 /// multiplying by a magic number. See:
8492 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
8493 SDValue DAGCombiner::BuildSDIV(SDNode *N) {
8494 std::vector<SDNode*> Built;
8495 SDValue S = TLI.BuildSDIV(N, DAG, LegalOperations, &Built);
8497 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
8503 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant,
8504 /// return a DAG expression to select that will generate the same value by
8505 /// multiplying by a magic number. See:
8506 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html>
8507 SDValue DAGCombiner::BuildUDIV(SDNode *N) {
8508 std::vector<SDNode*> Built;
8509 SDValue S = TLI.BuildUDIV(N, DAG, LegalOperations, &Built);
8511 for (std::vector<SDNode*>::iterator ii = Built.begin(), ee = Built.end();
8517 /// FindBaseOffset - Return true if base is a frame index, which is known not
8518 // to alias with anything but itself. Provides base object and offset as
8520 static bool FindBaseOffset(SDValue Ptr, SDValue &Base, int64_t &Offset,
8521 const GlobalValue *&GV, void *&CV) {
8522 // Assume it is a primitive operation.
8523 Base = Ptr; Offset = 0; GV = 0; CV = 0;
8525 // If it's an adding a simple constant then integrate the offset.
8526 if (Base.getOpcode() == ISD::ADD) {
8527 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Base.getOperand(1))) {
8528 Base = Base.getOperand(0);
8529 Offset += C->getZExtValue();
8533 // Return the underlying GlobalValue, and update the Offset. Return false
8534 // for GlobalAddressSDNode since the same GlobalAddress may be represented
8535 // by multiple nodes with different offsets.
8536 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Base)) {
8537 GV = G->getGlobal();
8538 Offset += G->getOffset();
8542 // Return the underlying Constant value, and update the Offset. Return false
8543 // for ConstantSDNodes since the same constant pool entry may be represented
8544 // by multiple nodes with different offsets.
8545 if (ConstantPoolSDNode *C = dyn_cast<ConstantPoolSDNode>(Base)) {
8546 CV = C->isMachineConstantPoolEntry() ? (void *)C->getMachineCPVal()
8547 : (void *)C->getConstVal();
8548 Offset += C->getOffset();
8551 // If it's any of the following then it can't alias with anything but itself.
8552 return isa<FrameIndexSDNode>(Base);
8555 /// isAlias - Return true if there is any possibility that the two addresses
8557 bool DAGCombiner::isAlias(SDValue Ptr1, int64_t Size1,
8558 const Value *SrcValue1, int SrcValueOffset1,
8559 unsigned SrcValueAlign1,
8560 const MDNode *TBAAInfo1,
8561 SDValue Ptr2, int64_t Size2,
8562 const Value *SrcValue2, int SrcValueOffset2,
8563 unsigned SrcValueAlign2,
8564 const MDNode *TBAAInfo2) const {
8565 // If they are the same then they must be aliases.
8566 if (Ptr1 == Ptr2) return true;
8568 // Gather base node and offset information.
8569 SDValue Base1, Base2;
8570 int64_t Offset1, Offset2;
8571 const GlobalValue *GV1, *GV2;
8573 bool isFrameIndex1 = FindBaseOffset(Ptr1, Base1, Offset1, GV1, CV1);
8574 bool isFrameIndex2 = FindBaseOffset(Ptr2, Base2, Offset2, GV2, CV2);
8576 // If they have a same base address then check to see if they overlap.
8577 if (Base1 == Base2 || (GV1 && (GV1 == GV2)) || (CV1 && (CV1 == CV2)))
8578 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
8580 // It is possible for different frame indices to alias each other, mostly
8581 // when tail call optimization reuses return address slots for arguments.
8582 // To catch this case, look up the actual index of frame indices to compute
8583 // the real alias relationship.
8584 if (isFrameIndex1 && isFrameIndex2) {
8585 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8586 Offset1 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base1)->getIndex());
8587 Offset2 += MFI->getObjectOffset(cast<FrameIndexSDNode>(Base2)->getIndex());
8588 return !((Offset1 + Size1) <= Offset2 || (Offset2 + Size2) <= Offset1);
8591 // Otherwise, if we know what the bases are, and they aren't identical, then
8592 // we know they cannot alias.
8593 if ((isFrameIndex1 || CV1 || GV1) && (isFrameIndex2 || CV2 || GV2))
8596 // If we know required SrcValue1 and SrcValue2 have relatively large alignment
8597 // compared to the size and offset of the access, we may be able to prove they
8598 // do not alias. This check is conservative for now to catch cases created by
8599 // splitting vector types.
8600 if ((SrcValueAlign1 == SrcValueAlign2) &&
8601 (SrcValueOffset1 != SrcValueOffset2) &&
8602 (Size1 == Size2) && (SrcValueAlign1 > Size1)) {
8603 int64_t OffAlign1 = SrcValueOffset1 % SrcValueAlign1;
8604 int64_t OffAlign2 = SrcValueOffset2 % SrcValueAlign1;
8606 // There is no overlap between these relatively aligned accesses of similar
8607 // size, return no alias.
8608 if ((OffAlign1 + Size1) <= OffAlign2 || (OffAlign2 + Size2) <= OffAlign1)
8612 if (CombinerGlobalAA) {
8613 // Use alias analysis information.
8614 int64_t MinOffset = std::min(SrcValueOffset1, SrcValueOffset2);
8615 int64_t Overlap1 = Size1 + SrcValueOffset1 - MinOffset;
8616 int64_t Overlap2 = Size2 + SrcValueOffset2 - MinOffset;
8617 AliasAnalysis::AliasResult AAResult =
8618 AA.alias(AliasAnalysis::Location(SrcValue1, Overlap1, TBAAInfo1),
8619 AliasAnalysis::Location(SrcValue2, Overlap2, TBAAInfo2));
8620 if (AAResult == AliasAnalysis::NoAlias)
8624 // Otherwise we have to assume they alias.
8628 /// FindAliasInfo - Extracts the relevant alias information from the memory
8629 /// node. Returns true if the operand was a load.
8630 bool DAGCombiner::FindAliasInfo(SDNode *N,
8631 SDValue &Ptr, int64_t &Size,
8632 const Value *&SrcValue,
8633 int &SrcValueOffset,
8634 unsigned &SrcValueAlign,
8635 const MDNode *&TBAAInfo) const {
8636 LSBaseSDNode *LS = cast<LSBaseSDNode>(N);
8638 Ptr = LS->getBasePtr();
8639 Size = LS->getMemoryVT().getSizeInBits() >> 3;
8640 SrcValue = LS->getSrcValue();
8641 SrcValueOffset = LS->getSrcValueOffset();
8642 SrcValueAlign = LS->getOriginalAlignment();
8643 TBAAInfo = LS->getTBAAInfo();
8644 return isa<LoadSDNode>(LS);
8647 /// GatherAllAliases - Walk up chain skipping non-aliasing memory nodes,
8648 /// looking for aliasing nodes and adding them to the Aliases vector.
8649 void DAGCombiner::GatherAllAliases(SDNode *N, SDValue OriginalChain,
8650 SmallVector<SDValue, 8> &Aliases) {
8651 SmallVector<SDValue, 8> Chains; // List of chains to visit.
8652 SmallPtrSet<SDNode *, 16> Visited; // Visited node set.
8654 // Get alias information for node.
8657 const Value *SrcValue;
8659 unsigned SrcValueAlign;
8660 const MDNode *SrcTBAAInfo;
8661 bool IsLoad = FindAliasInfo(N, Ptr, Size, SrcValue, SrcValueOffset,
8662 SrcValueAlign, SrcTBAAInfo);
8665 Chains.push_back(OriginalChain);
8668 // Look at each chain and determine if it is an alias. If so, add it to the
8669 // aliases list. If not, then continue up the chain looking for the next
8671 while (!Chains.empty()) {
8672 SDValue Chain = Chains.back();
8675 // For TokenFactor nodes, look at each operand and only continue up the
8676 // chain until we find two aliases. If we've seen two aliases, assume we'll
8677 // find more and revert to original chain since the xform is unlikely to be
8680 // FIXME: The depth check could be made to return the last non-aliasing
8681 // chain we found before we hit a tokenfactor rather than the original
8683 if (Depth > 6 || Aliases.size() == 2) {
8685 Aliases.push_back(OriginalChain);
8689 // Don't bother if we've been before.
8690 if (!Visited.insert(Chain.getNode()))
8693 switch (Chain.getOpcode()) {
8694 case ISD::EntryToken:
8695 // Entry token is ideal chain operand, but handled in FindBetterChain.
8700 // Get alias information for Chain.
8703 const Value *OpSrcValue;
8704 int OpSrcValueOffset;
8705 unsigned OpSrcValueAlign;
8706 const MDNode *OpSrcTBAAInfo;
8707 bool IsOpLoad = FindAliasInfo(Chain.getNode(), OpPtr, OpSize,
8708 OpSrcValue, OpSrcValueOffset,
8712 // If chain is alias then stop here.
8713 if (!(IsLoad && IsOpLoad) &&
8714 isAlias(Ptr, Size, SrcValue, SrcValueOffset, SrcValueAlign,
8716 OpPtr, OpSize, OpSrcValue, OpSrcValueOffset,
8717 OpSrcValueAlign, OpSrcTBAAInfo)) {
8718 Aliases.push_back(Chain);
8720 // Look further up the chain.
8721 Chains.push_back(Chain.getOperand(0));
8727 case ISD::TokenFactor:
8728 // We have to check each of the operands of the token factor for "small"
8729 // token factors, so we queue them up. Adding the operands to the queue
8730 // (stack) in reverse order maintains the original order and increases the
8731 // likelihood that getNode will find a matching token factor (CSE.)
8732 if (Chain.getNumOperands() > 16) {
8733 Aliases.push_back(Chain);
8736 for (unsigned n = Chain.getNumOperands(); n;)
8737 Chains.push_back(Chain.getOperand(--n));
8742 // For all other instructions we will just have to take what we can get.
8743 Aliases.push_back(Chain);
8749 /// FindBetterChain - Walk up chain skipping non-aliasing memory nodes, looking
8750 /// for a better chain (aliasing node.)
8751 SDValue DAGCombiner::FindBetterChain(SDNode *N, SDValue OldChain) {
8752 SmallVector<SDValue, 8> Aliases; // Ops for replacing token factor.
8754 // Accumulate all the aliases to this node.
8755 GatherAllAliases(N, OldChain, Aliases);
8757 // If no operands then chain to entry token.
8758 if (Aliases.size() == 0)
8759 return DAG.getEntryNode();
8761 // If a single operand then chain to it. We don't need to revisit it.
8762 if (Aliases.size() == 1)
8765 // Construct a custom tailored token factor.
8766 return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other,
8767 &Aliases[0], Aliases.size());
8770 // SelectionDAG::Combine - This is the entry point for the file.
8772 void SelectionDAG::Combine(CombineLevel Level, AliasAnalysis &AA,
8773 CodeGenOpt::Level OptLevel) {
8774 /// run - This is the main entry point to this class.
8776 DAGCombiner(*this, AA, OptLevel).Run(Level);